diff --git a/CMakeLists.txt b/CMakeLists.txt index 4921226ec1c90a969fa1cfc383823820500c7757..4783095194dc9c6409dc31c95588f46c9bee7c61 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -86,6 +86,14 @@ if(ANDROID OR IOS) "Disable MKLDNN when cross-compiling for Android and iOS" FORCE) set(WITH_MKLML OFF CACHE STRING "Disable MKLML package when cross-compiling for Android and iOS" FORCE) + + # Compile PaddlePaddle mobile inference library + if (NOT WITH_C_API) + set(WITH_C_API ON CACHE STRING + "Always compile the C_API when cross-compiling for Android and iOS" FORCE) + endif() + set(MOBILE_INFERENCE ON) + add_definitions(-DPADDLE_MOBILE_INFERENCE) endif() set(THIRD_PARTY_PATH "${CMAKE_BINARY_DIR}/third_party" CACHE STRING @@ -160,9 +168,11 @@ endif(USE_NNPACK) add_subdirectory(proto) -# "add_subdirectory(go)" should be placed after the following loine, -# because it depends on paddle/optimizer. -add_subdirectory(paddle/optimizer) +if(NOT MOBILE_INFERENCE) + # "add_subdirectory(go)" should be placed after the following loine, + # because it depends on paddle/optimizer. + add_subdirectory(paddle/optimizer) +endif() # "add_subdirectory(paddle)" and "add_subdirectory(python)" should be # placed after this block, because they depends on it. diff --git a/README.md b/README.md index b9793c3eab5d40c28f01cc67ad607b97261b3235..db0fbd88b250cdc2a3cc77521cc1c2cea77c6e87 100644 --- a/README.md +++ b/README.md @@ -51,19 +51,19 @@ Please refer to our [release announcement](https://github.com/PaddlePaddle/Paddl - **Connected to Products** In addition, PaddlePaddle is also designed to be easily deployable. At Baidu, - PaddlePaddle has been deployed into products or service with a vast number + PaddlePaddle has been deployed into products and services with a vast number of users, including ad click-through rate (CTR) prediction, large-scale image classification, optical character recognition(OCR), search ranking, computer virus detection, recommendation, etc. It is widely utilized in products at - Baidu and it has achieved a significant impact. We hope you can also exploit - the capability of PaddlePaddle to make a huge impact for your product. + Baidu and it has achieved a significant impact. We hope you can also explore + the capability of PaddlePaddle to make an impact on your product. ## Installation It is recommended to check out the [Docker installation guide](http://doc.paddlepaddle.org/develop/doc/getstarted/build_and_install/docker_install_en.html) before looking into the -[build from source guide](http://doc.paddlepaddle.org/develop/doc/getstarted/build_and_install/build_from_source_en.html) +[build from source guide](http://doc.paddlepaddle.org/develop/doc/getstarted/build_and_install/build_from_source_en.html). ## Documentation @@ -72,7 +72,7 @@ We provide [English](http://doc.paddlepaddle.org/develop/doc/) and - [Deep Learning 101](http://book.paddlepaddle.org/index.html) - You might want to start from this online interactive book that can run in Jupyter Notebook. + You might want to start from this online interactive book that can run in a Jupyter Notebook. - [Distributed Training](http://doc.paddlepaddle.org/develop/doc/howto/usage/cluster/cluster_train_en.html) diff --git a/benchmark/paddle/image/run_mkldnn.sh b/benchmark/paddle/image/run_mkldnn.sh index 5b0a0373448e5b81ff0718db3465a4694690ec37..e31fec1cd850157d90ddcab2d559d52381ecd317 100755 --- a/benchmark/paddle/image/run_mkldnn.sh +++ b/benchmark/paddle/image/run_mkldnn.sh @@ -1,23 +1,24 @@ set -e -unset OMP_NUM_THREADS MKL_NUM_THREADS -export OMP_DYNAMIC="FALSE" -export KMP_AFFINITY="granularity=fine,compact,0,0" - function train() { + unset OMP_NUM_THREADS MKL_NUM_THREADS + export OMP_DYNAMIC="FALSE" + export KMP_AFFINITY="granularity=fine,compact,0,0" topology=$1 bs=$2 use_mkldnn=$3 if [ $3 == "True" ]; then - use_mkldnn=$3 thread=1 log="logs/${topology}-mkldnn-${bs}.log" elif [ $3 == "False" ]; then - use_mkldnn=$3 thread=`nproc` + # each trainer_count use only 1 core to avoid conflict + export OMP_NUM_THREADS=1 + export MKL_NUM_THREADS=1 log="logs/${topology}-${thread}mklml-${bs}.log" else echo "Wrong input $3, use True or False." + exit 0 fi args="batch_size=${bs}" config="${topology}.py" @@ -39,8 +40,7 @@ if [ ! -d "logs" ]; then mkdir logs fi -#========= mkldnn =========# -# vgg +#========== mkldnn ==========# train vgg 64 True train vgg 128 True train vgg 256 True diff --git a/cmake/configure.cmake b/cmake/configure.cmake index 51c3b918cc4ef4cf6c8052ccc14028a872309fcf..c1c93e17fd82ea048ba27b127b1527d9a8c9da41 100644 --- a/cmake/configure.cmake +++ b/cmake/configure.cmake @@ -49,11 +49,12 @@ if(NOT WITH_GOLANG) endif(NOT WITH_GOLANG) if(NOT WITH_GPU) - add_definitions(-DPADDLE_ONLY_CPU) add_definitions(-DHPPL_STUB_FUNC) list(APPEND CMAKE_CXX_SOURCE_FILE_EXTENSIONS cu) else() + add_definitions(-DPADDLE_WITH_CUDA) + FIND_PACKAGE(CUDA REQUIRED) if(${CUDA_VERSION_MAJOR} VERSION_LESS 7) diff --git a/cmake/util.cmake b/cmake/util.cmake index d1aee3e170a2d143ac06b438725e907e96f041c8..117ab7f49cdf4a568cd203b2b17767643d0b2d50 100644 --- a/cmake/util.cmake +++ b/cmake/util.cmake @@ -73,25 +73,43 @@ function(link_paddle_exe TARGET_NAME) generate_rdma_links() endif() - target_circle_link_libraries(${TARGET_NAME} - ARCHIVE_START - paddle_gserver - paddle_function - ARCHIVE_END - paddle_pserver - paddle_trainer_lib - paddle_network - paddle_math - paddle_utils - paddle_parameter - paddle_proto - paddle_cuda - paddle_optimizer - ${EXTERNAL_LIBS} - ${CMAKE_THREAD_LIBS_INIT} - ${CMAKE_DL_LIBS} - ${RDMA_LD_FLAGS} - ${RDMA_LIBS}) + if(MOBILE_INFERENCE) + target_circle_link_libraries(${TARGET_NAME} + ARCHIVE_START + paddle_gserver + paddle_function + ARCHIVE_END + paddle_math + paddle_utils + paddle_parameter + paddle_proto + paddle_cuda + ${EXTERNAL_LIBS} + ${CMAKE_THREAD_LIBS_INIT} + ${CMAKE_DL_LIBS} + ${RDMA_LD_FLAGS} + ${RDMA_LIBS}) + else() + target_circle_link_libraries(${TARGET_NAME} + ARCHIVE_START + paddle_gserver + paddle_function + ARCHIVE_END + paddle_pserver + paddle_trainer_lib + paddle_network + paddle_math + paddle_utils + paddle_parameter + paddle_proto + paddle_cuda + paddle_optimizer + ${EXTERNAL_LIBS} + ${CMAKE_THREAD_LIBS_INIT} + ${CMAKE_DL_LIBS} + ${RDMA_LD_FLAGS} + ${RDMA_LIBS}) + endif() if(ANDROID) target_link_libraries(${TARGET_NAME} log) diff --git a/doc/api/v1/index_cn.rst b/doc/api/v1/index_cn.rst index 3718cd73a2003b8ef6c406a9bd51dc68e76402dc..cf146dc088e3905a751ff55c26fd82ef0ba02c89 100644 --- a/doc/api/v1/index_cn.rst +++ b/doc/api/v1/index_cn.rst @@ -21,7 +21,7 @@ Model Config API trainer_config_helpers/optimizers.rst trainer_config_helpers/data_sources.rst trainer_config_helpers/layers.rst - trainer_config_helpers/activations.rst + trainer_config_helpers/activations.rst trainer_config_helpers/poolings.rst trainer_config_helpers/networks.rst trainer_config_helpers/evaluators.rst diff --git a/doc/api/v2/config/layer.rst b/doc/api/v2/config/layer.rst index c94627a72806fa2eca77c79da24f7f3ca18f0259..d4e9d53e5c0955912a594fe8cd9cd41a4080a2d2 100644 --- a/doc/api/v2/config/layer.rst +++ b/doc/api/v2/config/layer.rst @@ -345,6 +345,11 @@ clip .. autoclass:: paddle.v2.layer.clip :noindex: +resize +------ +.. autoclass:: paddle.v2.layer.resize + :noindex: + slope_intercept --------------- .. autoclass:: paddle.v2.layer.slope_intercept diff --git a/doc/design/block.md b/doc/design/block.md index be8800122035984df281692fc40009c397565046..4d5dd4ba95a686d18b2339c69f0316c340681909 100644 --- a/doc/design/block.md +++ b/doc/design/block.md @@ -55,17 +55,23 @@ Let us consolidate the discussion by presenting some examples. The following C++ programs shows how blocks are used with the `if-else` structure: ```c++ +namespace pd = paddle; + int x = 10; -int y = 20; -int out; +int y = 1; +int z = 10; bool cond = false; +int o1, o2; if (cond) { int z = x + y; - out = softmax(z); + o1 = z; + o2 = pd::layer::softmax(z); } else { - int z = fc(x); - out = z; + int d = pd::layer::fc(z); + o1 = d; + o2 = d+1; } + ``` An equivalent PaddlePaddle program from the design doc of the [IfElseOp operator](./if_else_op.md) is as follows: @@ -73,57 +79,55 @@ An equivalent PaddlePaddle program from the design doc of the [IfElseOp operator ```python import paddle as pd -x = var(10) -y = var(20) -cond = var(false) -ie = pd.create_ifelseop(inputs=[x], output_num=1) +x = minibatch([10, 20, 30]) # shape=[None, 1] +y = var(1) # shape=[1], value=1 +z = minibatch([10, 20, 30]) # shape=[None, 1] +cond = larger_than(x, 15) # [false, true, true] + +ie = pd.ifelse() with ie.true_block(): - x = ie.inputs(true, 0) - z = operator.add(x, y) - ie.set_output(true, 0, operator.softmax(z)) + d = pd.layer.add_scalar(x, y) + ie.output(d, pd.layer.softmax(d)) with ie.false_block(): - x = ie.inputs(false, 0) - z = layer.fc(x) - ie.set_output(true, 0, operator.softmax(z)) -out = b(cond) + d = pd.layer.fc(z) + ie.output(d, d+1) +o1, o2 = ie(cond) ``` -In both examples, the left branch computes `softmax(x+y)` and the right branch computes `fc(x)`. +In both examples, the left branch computes `x+y` and `softmax(x+y)`, the right branch computes `x+1` and `fc(x)`. A difference is that variables in the C++ program contain scalar values, whereas those in the PaddlePaddle programs are mini-batches of instances. The `ie.input(true, 0)` invocation returns instances in the 0-th input, `x`, that corresponds to true values in `cond` as the local variable `x`, where `ie.input(false, 0)` returns instances corresponding to false values. + ### Blocks with `for` and `RNNOp` The following RNN model from the [RNN design doc](./rnn.md) ```python -x = sequence([10, 20, 30]) -m = var(0) -W = tensor() -U = tensor() - -rnn = create_rnn(inputs=[input]) -with rnn.stepnet() as net: - x = net.set_inputs(0) - h = net.add_memory(init=m) - fc_out = pd.matmul(W, x) - hidden_out = pd.matmul(U, h.pre(n=1)) - sum = pd.add_two(fc_out, hidden_out) - act = pd.sigmoid(sum) - h.update(act) # update memory with act - net.set_outputs(0, act, hidden_out) # two outputs - +x = sequence([10, 20, 30]) # shape=[None, 1] +m = var(0) # shape=[1] +W = var(0.314, param=true) # shape=[1] +U = var(0.375, param=true) # shape=[1] + +rnn = pd.rnn() +with rnn.step(): + h = rnn.memory(init = m) + hh = rnn.previous_memory(h) + a = layer.fc(W, x) + b = layer.fc(U, hh) + s = pd.add(a, b) + act = pd.sigmoid(s) + rnn.update_memory(h, act) + rnn.output(a, b) o1, o2 = rnn() -print o1, o2 ``` - has its equivalent C++ program as follows ```c++ int* x = {10, 20, 30}; -int m = 0; -int W = some_value(); -int U = some_other_value(); +int* m = {0}; +int* W = {0.314}; +int* U = {0.375}; int mem[sizeof(x) / sizeof(x[0]) + 1]; int o1[sizeof(x) / sizeof(x[0]) + 1]; @@ -131,20 +135,16 @@ int o2[sizeof(x) / sizeof(x[0]) + 1]; for (int i = 1; i <= sizeof(x)/sizeof(x[0]); ++i) { int x = x[i-1]; if (i == 1) mem[0] = m; - int fc_out = W * x; - int hidden_out = Y * mem[i-1]; - int sum = fc_out + hidden_out; + int a = W * x; + int b = Y * mem[i-1]; + int s = fc_out + hidden_out; int act = sigmoid(sum); mem[i] = act; o1[i] = act; o2[i] = hidden_out; } - -print_array(o1); -print_array(o2); ``` - ## Compilation and Execution Like TensorFlow programs, a PaddlePaddle program is written in Python. The first part describes a neural network as a protobuf message, and the rest part executes the message for training or inference. @@ -210,11 +210,11 @@ a = pd.Varaible(shape=[20, 20]) b = pd.fc(a, params=["fc.w", "fc.b"]) rnn = pd.create_rnn() -with rnn.stepnet() as net: - x = net.set_inputs(a) +with rnn.stepnet() + x = a.as_step_input() # reuse fc's parameter fc_without_b = pd.get_variable("fc.w") - net.set_outputs(fc_without_b) + rnn.output(fc_without_b) out = rnn() ``` diff --git a/doc/design/if_else_op.md b/doc/design/if_else_op.md index 954a19c0733358c235eae3cffe134c23dac94c95..26d140f06db4ecefa86be015eaa731ffddc6910c 100644 --- a/doc/design/if_else_op.md +++ b/doc/design/if_else_op.md @@ -1,41 +1,51 @@ -IfOp should have only one branch. An IfOp operator takes a `cond` variable whose value must be a vector of N boolean elements. Its return value has N instances. If cond[i] == True, input instance input[i] will go through true_block() and generate output[i]; otherwise it will produce output from false_bloack(). +# The `IfElse` Operator -```python -import paddle as pd +PaddlePaddle's `IfElse` operator differs from TensorFlow's: -x = var() -y = var() -cond = var() -default_value = var() -b = pd.create_ifelseop(inputs=[x], output_num=1) -with b.true_block(): - x = b.inputs(0) - z = operator.add(x, y) - b.set_output(0, operator.softmax(z)) - -with b.false_block(): - x = b.inputs(0) - z = layer.fc(x) - b.set_output(0, operator.softmax(z)) - -out = b(cond) -``` +- the TensorFlow version takes a scalar boolean value as the condition so that the whole mini-batch goes to either the true or the false branch, whereas +- the PaddlePaddle version takes a vector of boolean value as the condition, and instances corresponding to true values go to the true branch, those corresponding to false values go to the false branch. + +## Example + +The following PaddlePaddle program shows the usage of the IfElse operator: -If only true_block is set in an IfElseOp, a special case is that we can have a default value for false as: ```python import paddle as pd -x = var() -y = var() -cond = var() -default_value = var() -b = pd.create_ifelseop(inputs=[x], output_num=1, default_value) - -with b.true_block(): - x = b.inputs(0) - z = operator.add(x, y) - b.set_output(0, operator.softmax(z)) +x = minibatch([10, 20, 30]) # shape=[None, 1] +y = var(1) # shape=[1], value=1 +z = minibatch([10, 20, 30]) # shape=[None, 1] +cond = larger_than(x, 15) # [false, true, true] + +ie = pd.ifelse() +with ie.true_block(): + d = pd.layer.add(x, y) + ie.output(d, pd.layer.softmax(d)) +with ie.false_block(): + d = pd.layer.fc(z) + ie.output(d, d+1) +o1, o2 = ie(cond) +``` -out = b(cond) +A challenge to implement the `IfElse` operator is to infer those variables to be split, or, say, to identify the variable of the mini-batch or those derived from the mini-batch. + +An equivalent C++ program is as follows: + +```c++ +namespace pd = paddle; + +int x = 10; +int y = 1; +int z = 10; +bool cond = false; +int o1, o2; +if (cond) { + int d = x + y; + o1 = z; + o2 = pd::layer::softmax(z); +} else { + int d = pd::layer::fc(z); + o1 = d; + o2 = d+1; +} ``` -where default_value is a list of vars for `cond` == False. diff --git a/doc/design/program.md b/doc/design/program.md index fb8f86ac07af403c9fee015f2a3adbfaa3c6d631..bd2456787c4e336d357a65255a8274a7c9e465cc 100644 --- a/doc/design/program.md +++ b/doc/design/program.md @@ -1,8 +1,10 @@ -# Design Doc: ProgramDesc +# Design Doc: PaddlePaddle Programs -The basic structure of a PaddlePaddle program is some nested blocks, as a C++ or Java program. +## Compile and Execution + +A PaddlePaddle program consists of two parts -- the first generates a `ProgramDesc` protobuf message that describes the program, and the second runs this message using a C++ class `Executor`. -As described in [graph.md](./graph.md), the first five lines of the following PaddlePaddle program +A simple example PaddlePaddle program can be found in [graph.md](./graph.md): ```python x = layer.data("images") @@ -13,36 +15,112 @@ optimize(cost) train(cost, reader=mnist.train()) ``` -generates, or compiles, a PaddelPaddle program, which is represented by the following protobuf message: +The first five lines of the following PaddlePaddle program generates, or, compiles, the `ProgramDesc` message. The last line runs it. -```protobuf -message ProgramDesc { - repeated BlockDesc blocks = 1; +## Programs and Blocks + +The basic structure of a PaddlePaddle program is some nested blocks, as a C++ or Java program. + +- program: some nested blocks +- [block](./block.md): + - some local variable definitions, and + - a sequence of operators + +The concept of block comes from usual programs. For example, the following C++ program has three blocks: + +```c++ +int main() { // block 0 + int i = 0; + if (i < 10) { // block 1 + for (int j = 0; j < 10; j++) { // block 2 + } + } + return 0; } +``` + +The following PaddlePaddle program has three blocks: + +```python +import paddle as pd // block 0 + +x = minibatch([10, 20, 30]) # shape=[None, 1] +y = var(1) # shape=[1], value=1 +z = minibatch([10, 20, 30]) # shape=[None, 1] +cond = larger_than(x, 15) # [false, true, true] +ie = pd.ifelse() +with ie.true_block(): // block 1 + d = pd.layer.add_scalar(x, y) + ie.output(d, pd.layer.softmax(d)) +with ie.false_block(): // block 2 + d = pd.layer.fc(z) + ie.output(d, d+1) +o1, o2 = ie(cond) +``` + +## `BlockDesc` and `ProgramDesc` + +All protobuf messages are defined in `framework.proto`. + +`BlockDesc` is straight-forward -- it includes local variable definitions, `vars`, and a sequence of operators, `ops`. + +```protobuf message BlockDesc { required int32 parent = 1; repeated VarDesc vars = 2; repeated OpDesc ops = 3; } +``` + +The parent ID indicates the parent block so that operators in a block can refer to variables defined locally and also those defined in their ancestor blocks. + +All hierarchical blocks in a program are flattened and stored in an array. The block ID is the index of the block in this array. + +```protobuf +message ProgramDesc { + repeated BlockDesc blocks = 1; +} +``` + + +### Global Block +The global block is the first one in the above array. + +## Operators that Use Blocks + +In the above example, the operator `IfElseOp` has two blocks -- the true branch and the false branch. + +The definition of `OpDesc` shows that an operator could have some attributes: + +```protobuf message OpDesc { AttrDesc attrs = 1; ... } +``` + +and an attribute could be of type block, which is, in fact, a block ID as described above: +``` message AttrDesc { - required AttrType type = 1; + required string name = 1; - // index into ProgramDesc::blocks when type==BLOCK - optional int32 block = 2; + enum AttrType { + INT = 1, + STRING = 2, + ... + BLOCK = ... + } + required AttrType type = 2; + + optional int32 block = 10; // when type == BLOCK ... } ``` -When each of the first five lines runs, related Python function, e.g., `layer.fc`, calls C++ InferShape functions. This InferShape function needs to access the properties of VarDesc's accessed by the current OpDesc. These VarDesc's might not be defined in the current block, but in some ancestor blocks. This requires that we can trace the parent of a block. - -A nested block is often an attribute of an operator, most likely, an IfElseOp or a WhileOp. In above solution, all blocks are in `ProgramDesc::blocks`, this implicitly assigns a zero-based ID to each block -- the index of the block in `ProgramDesc::blocks`. So that `AttrDesc::block` could be an integer block ID. +## InferShape With this design, the InferShape function should take the following parameters: diff --git a/doc/design/python_api.md b/doc/design/python_api.md new file mode 100644 index 0000000000000000000000000000000000000000..6213da65c8c5931bc16e42574b8628b676424873 --- /dev/null +++ b/doc/design/python_api.md @@ -0,0 +1,216 @@ +# Design Doc: Python API + +Due to the refactorization of the PaddlePaddle core, we need Python classes to construct corresponding protobuf messages that describe a DL program. + +| Python classes | Protobuf messages | +| --- | --- | +| Program | ProgramDesc | +| Block | BlockDesc | +| Operator | OpDesc | +| Variable | VarDesc | + +Please be aware that these Python classes need to maintain some construction-time information, which are not part of the protobuf messages. + +## Core Concepts + +### Program + +A `ProgramDesc` describes a [DL program](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/program.md), which is composed of an array of `BlockDesc`s. The `BlockDesc`s in a `ProgramDesc` can have a tree-like hierarchical structure. However, the `ProgramDesc` onlys stores a flattened array of `BlockDesc`s. A `BlockDesc` refers to its parent block by its index in the array. For example, operators in the step block of an RNN operator need to be able to access variables in its ancestor blocks. + +Whenever we create a block, we need to set its parent block to the current block, hence the Python class `Program` needs to maintain a data member `current_block`. + +```python +class Program(objects): + def __init__(self): + self.proto = core.NewProgram() # a C++ ProgramDesc pointer. + self.blocks = vector() + self.blocks.append(Block(self, -1)) # the global block + self.current_block = 0 # initialized to the global block + + def global_block(): + return self.blocks[0] + + def current_block(): + return self.get_block(self.current_block) + + def rollback(): + self.current_block = self.current_block().parent_idx + + def create_block(): + new_block_idx = len(self.block) + self.blocks.append(Block(self, self.current_block)) + self.current_block = new_block_idx + return current_block() +``` + +`Program` is an accessor to the protobuf message `ProgramDesc`, which is created in C++ space, because the InferShape function is in C++, which manipulates `VarDesc` messages, which are in turn members of `BlockDesc`, which is a member of `ProgramDesc`. + +`Program` creates the first block as the global block in its constructor. All parameters and their initializer operators are in the global block. + +### Block + +A [Block](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/block.md) includes + +1. a map from variable names to an instance of the Python `Variable` class, and +1. a list of `Operator` instances. + +```python +class Block(objects): + def __init__(self, program, parent_idx): + self.proto = core.NewBlock(program.proto) + self.program = program + self.vars = map() + self.ops = vector() + self.parent_idx = parent_idx + + def create_var(self, ...): + return Variable(self, ...) + + def _create_global_var(self, ...): + program.global_block().create_var(...) + + def create_parameter(self, name, ...): + # Parameter is a subclass of variable. See Parameter section for details. + self.vars[name] = Parameter(self._create_global_var(...), ...) + return self.vars[name] + + def append_operator(self, ...): + self.ops.append(Operator(self, ...)) + + def prepend_operator(self, ...): # Parameter's ctor prepands initialize operators. + self.ops.prepend(Operator(self, ...)) +``` + +`create_parameter` is necessary because parameters are global variables, defined in the global block, but can be created in some sub-blocks. For example, an FC layer in the step block of an RNN operator. + +`prepend_operator` is necessary because the constructor of `Parameter` needs to create the initialize (or load) operator of the parameter, and would like to put it in the *preamble* of the global block. + +### Operator + +The `Operator` class fills in the `OpDesc` message and calls the C++ function `InferShape` to infer the output shapes from the input shapes. + +```python +class Operator(object): + def __init__(self, + block, # Block + type, # string + inputs, # dict + outputs,# dict + attrs # dict + ): + self.proto = core.NewOpDesc(block.proto, type, inputs, outputs, attrs) + core.infer_shape(self.proto, inputs, outputs) + + def type(self): + return self.proto.type() +``` + +`Operator` creates the `OpDesc` message in C++ space, so that it can call the `InferShape` function, which is in C++. + +### Variable + +Operators take Variables as its inputs and outputs. + +```python +class Variable(object): + def __init__(self, + block=None, # Block + name=None, # string + shape, # tuple + dtype="float32", # string + lod_level=None # int + ): + if name is None: + name = unique_name_generator() + self.name = name + self.block = block + self.proto = core.NewVarDesc(block.proto, name, shape, lod_level) + self.writer = None +``` + +Please be aware of `self.writer`, that tracks operator who creates the variable. It possible that there are more than one operators who write a variable, but in Python space, each write to a variable is represented by a Variable class. This is guaranteed by the fact that **`core.NewVarDesc` must NOT create a new `VarDesc` message if its name already exists in the specified block**. + +### Parameter + +A parameter is a global variable with an initializer (or load) operator. + +```python +class Parameter(Variable): + def __init__(self, + block=None, # Block + name=None, # string + shape, # tuple + dtype="float32", # string + lod_level=None # int + trainable, # bool + initialize_op_attrs, + optimize_op_attrs): + super(Parameter, self).__init__(block, name, shape, dtype, lod_level) + self.trainable = trainable + self.optimize_op_attrs = optimize_op_attrs + block.prepend(Operator(block, # Block + initialize_op_attrs['type'], # string + None, # no inputs + self, # output is the parameter + initialize_op_attrs) +``` + +When users create a parameter, they can call + +```python +program.create_parameter( + ..., + init_attr={ + type: "uniform_random", + min: -1.0, + max: 1.0, + }) +) +``` + +In above example, `init_attr.type` names an initialize operator. It can also name the load operator + +```python +init_attr={ + type: "load", + filename: "something.numpy", +} +``` + +`optimize_op_attrs` is not in the `VarDesc` message, but kept in the Python instance, as it will be used in the Python space when creating the optimize operator's `OpDesc`, and will be in the `OpDesc` message. + +## Layer Functions + +A layer is a Python function that creates some operators and variables. Layers simplify the work of application programmers. + +### Data Layer + +```python +def data_layer(name, type, column_name): + block = the_current_program.glolal_block() + var = block.create_global_var( + name=name, + shape=[None] + type.dims(), + dtype=type.dtype) + block.prepend_operator(block, + type="Feed", + inputs = None, + outputs = [var], + {column_name: column_name}) + return var +``` + +The input to the feed operator is a special variable in the global scope, which is the output of [Python readers](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/reader/README.md). + +### FC Layer + +```python +def fc_layer(input, size, ...): + block = program.current_block() + w = block.create_parameter(...) + b = block.create_parameter(...) + out = block.create_var() + op = block.append_operator("FC", X=input, W=w, b=b, out=out) + out.writer = op + return out +``` diff --git a/doc/design/refactor/distributed_architecture.md b/doc/design/refactor/distributed_architecture.md new file mode 100644 index 0000000000000000000000000000000000000000..ac7e98ccf1aadbb973a4801fde842375cf63448c --- /dev/null +++ b/doc/design/refactor/distributed_architecture.md @@ -0,0 +1,222 @@ +# Design Doc: Distributed Training Architecture + +## Abstract + +PaddlePaddle v0.10.0 uses the "trainer-parameter server" +architecture. We run multiple replicated instances of trainers (runs +the same code written by the user) and parameter servers for +distributed training. This architecture served us well, but has some +limitations: + +1. Need to write special code to handle tasks which should only be run + by a single trainer. E.g., initializing model and saving model. + +2. Model parallelism is hard: need to write if-else branches conditioned + on the trainer ID to partition model onto each trainer, and manually + write the inter-model-shard communication code. + +3. The user can not directly specify the parameter update rule: need + to modify the parameter server C++ code and compile a new + binary. This adds complication for researchers: A lot of extra + effort is required. Besides, the training job submission program + may not allow running arbitrary binaries. + +This design doc discusses PaddlePaddle's new distributed training +architecture that addresses the above limitations. + +## Analysis + +We will assume the user writes the trainer program by Python, the same +analysis holds if the trainer program is written in C++. + +### Limitation 1 + +If we look at the Python code that the user writes, there are two +kinds of functionalities: + +- The training logic such as load / save model and print log. +- The neural network definition such as the definition of the data + layer, the fully connected layer, the cost function and the + optimizer. + +When we training with PaddlePaddle v0.10.0 distributedly, multiple +replicated Python instances are running on different nodes: both the +training logic and the neural network computation is replicated. + +The tasks that should only run once all belong to the training logic, +if we only replicate the neural network computation, but do **not** +replicate the training logic, the limitation could be solved. + +### Limitation 2 + +Model parallelism means running a single model on multiple nodes by +partitioning the model onto different nodes and managing the +inter-model-shard communications. + +PaddlePaddle should be able to modify the nerual network computation +definition to support model parallelism automatically. However, the +computation is only specified in Python code, and PaddlePaddle can not +modify Python code. + +Just like compiler uses a intermediate representation (IR) so that +programmer does not need to manually optimize their code in most of +the cases - the compiler will optimize the IR: + + + +We can have our own IR too: PaddlePaddle can support model parallel by +converting the IR so the user no longer need to manually do it in +Python: + + + +The IR for PaddlePaddle after refactor is called `Block`, it specifies +the computation dependency graph and the variables used in the +computation. + +### Limitation 3 + +The user can not directly specify the parameter update rule for the +parameter server because the parameter server does not use the same +computation definition as the trainer. Instead, the update rule is +baked in the parameter server. The user can not specify the update +rule in the same way of specifying the trainer computation. + +This could be fixed by making the parameter server run the same +computation definition as the trainer. For a detailed explanation, +please +see +[Design Doc: Operation Graph Based Parameter Server](./dist_train.md) + +## Distributed Training Architecture + +The new distributed training architecture can address the above +limitations. Below is the illustration: + + + +The architecture includes major components: *PaddlePaddle Python*, +*PaddlePaddle converter* and *PaddlePaddle runtime*: + +### PaddlePaddle Python + +PaddlePaddle Python is the Python library that user's Python trainer +invoke to build the neural network topology, start training, etc. + +```Python +paddle.init() +input = paddle.op.recordIO("/home/data/mnist.recordio") # file stored on the cluster +img, label = input[0], input[1] +hidden = paddle.layer.fc(input=img, size=200, act=paddle.activation.Tanh()) +prediction = paddle.layer.fc(input=img, size=10, act=paddle.activation.Softmax()) +cost = paddle.layer.classification_cost(input=prediction, label=label) +optimizer = paddle.optimizer.SGD(cost, learning_rate=0.01) +session = paddle.session.NewRemote(num_trainer=3, num_ps=2, GPU_per_trainer=1) +for i in range(1000): + _, cost_val = session.eval(targets=[cost, optimizer]) + print cost_val +``` + +The code above is a typical Python trainer code, the neural network +topology is built using helper functions such as +`paddle.layer.fc`. The training is done by calling `session.eval` +iteratively. + +#### session.eval + +As shown in the graph, `session.eval` sends the IR and the evaluation +inputs/targets to the PaddlePaddle cluster for evaluation. The +targets can be any variable in the computation graph. When the target +is the `optimizer` variable, the neural network will be optimized +once. When the target is the `cost` variable, `session.eval` returns +the cost value. + +The Python `session` is a wrapper of the C++ `Session` class. For more +information about `Session`, please +see [Design Doc: Session](./session.md). + +### PaddlePaddle Converter + +PaddlePaddle converter automatically converts the IR in the request +(IR and evaluation inputs/targets) from PaddlePaddle Python to new +partitioned IRs and dispatch the new IRs and evaluation inputs/targets +to different PaddlePaddle runtimes. Below are the steps: + +1. Add `feed` OP that feeds the eval inputs, and `fetch` OP that + fetches the eval targets to the IR. + +1. Extract a new computation (sub)graph with `feed` and `fetch` OP as + the boundary. The runtime does not need to run the OP that is not + dependent by the `fetch` OP. + +1. Optimizes the computation graph. + +1. Place the OPs in the graph onto different devices on different + PaddlePaddle runtime according to a placement algorithm and device + constraint specified by the user. + +1. Partition the graph according to runtime boundaries and add `send` / + `recv` OP pair on the runtime boundaries. + +1. Dispatch the partitioned graph to different PaddlePaddle runtimes. + +1. PaddlePaddle runtimes with the `fetch` OP reports evaluation + results back to the converter, the convert reports the evaluation + results back to the PaddlePaddle Python. + +The output IRs will be cached to optimize the conversion latency. + + +#### Placement Algorithm + +Our first implementation will only support "trainer-parameter server" +placement: the parameters, initializers, and optimizers are placed on +the PaddlePaddle runtimes with the parameter server role. And +everything else will be placed on the PaddlePaddle runtimes with the +trainer role. This has the same functionality of our +"trainer-parameter server" architecture of PaddlePaddle v0.10.0, but +is more general and flexible. + +In the future, we will implement the general placement algorithm, +which makes placements according to the input IR, and a model of +device computation time and device communication time. Model +parallelism requires the general placement algorithm. + + +### PaddlePaddle Runtime + +The PaddlePaddle runtime owns multiple devices (e.g., CPUs, GPUs) and +runs the IR. The runtime does not need to do OP placement since it's +already done by the converter. + + +### Local Training Architecture + +The local training architecture will be the same as the distributed +training architecture, the differences are everything runs locally, +and there is just one PaddlePaddle runtime: + + + + +### Training Data + +In PaddlePaddle v0.10.0, training data is typically read +with [data reader](../reader/README.md) from Python. This approach is +no longer efficient when training distributedly since the Python +process no longer runs on the same node with the trainer processes, +the Python reader will need to read from the distributed filesystem +(assuming it has the access) and send to the trainers, doubling the +network traffic. + +When doing distributed training, the user can still use Python data +reader: the training data are sent with `session.eval`. However should +be used for debugging purpose only. The users are encouraged to use +the read data OPs. + + +## References: + +[1] [TensorFlow: Large-Scale Machine Learning on Heterogeneous Distributed Systems](https://static.googleusercontent.com/media/research.google.com/en//pubs/archive/45166.pdf) + +[2] [TensorFlow: A System for Large-Scale Machine Learning](https://www.usenix.org/system/files/conference/osdi16/osdi16-abadi.pdf) diff --git a/doc/design/ops/dist_train.md b/doc/design/refactor/parameter_server.md similarity index 100% rename from doc/design/ops/dist_train.md rename to doc/design/refactor/parameter_server.md diff --git a/doc/design/refactor/session.md b/doc/design/refactor/session.md new file mode 100644 index 0000000000000000000000000000000000000000..1d9a26683c14f54e3b5fe41675cd03b5620646b8 --- /dev/null +++ b/doc/design/refactor/session.md @@ -0,0 +1,180 @@ +# Design Doc: Session + +## Abstract + +The *session* object encapsulates the environment in which the +computation graph is executed. + +We will have the *local* session and *remote* session, they offer the +same [interface](#interface). The local session encapsulates the local +runtime environment and the remote session encapsulates the cluster +runtime environment. + +The local runtime environment contains: + +1. computation devices (i.e., CPU, GPU) handles, and +1. the [scope](../scope.md) which holds all variables. + +The remote runtime environment contains: + +1. computation devices (i.e., CPU and GPU on node 0, 1) in a cluster, + and +1. the distributed [scope](../scope.md) in a cluster which holds all + variables. + +The user can create a remote session on Paddle Cloud and evaluate the +computation graph with it. In this way, the user can control the +remote computation resource in a cluster from his local computer. + + +## Background + +The current design has an implicit global session in which +`paddle.eval()` is executed. The pain point is: + +Since the user is not able to explicitly switch between runtime +environments, the user cannot run a topology in two independent +environments. + +For example, in reinforcement learning, the user may want to have a +stale model for inference and a fresh model for training, and only +replace the stale model with the fresh model periodically. + +Furthermore, we have no concept that encapsulates a remote environment +that executes a computation graph. + +We need the session object to address above issues. + + +## Session + +A session is an object that owns the runtime environment. All +computations are executed through `session.eval()`. + + +### Interface + +```python +eval( + targets, + feed_dict=None, +) +``` + +Evaluates the target Operations or Variables in `targets`. + +- *targets*: the evaluation targets. Can be a single Operation or + Variable, or a list with the Operations or Variables as + elements. The value returned by `eval()` has the same shape as the + `target` argument. + + The PaddlePaddle program is represented by + the [ProgramDesc](../design/program.md), `eval()` will infer the + ProgramDesc from the given targets and run the PaddlePaddle + program. Please + see + [this graph](./distributed_architecture.md#local-training-architecture) for + the detailed illustration for the local session + and + [this graph](./distributed_architecture.md#distributed-training-architecture) for + the detailed illustration for the remote session. + +- *feed_dict*: a dictionary that contains the tensors which override + the edges of the computation graph. + + feed_dict not only can provide the input data, it can override any + OP's input as well: + + ```python + a = pd.constant(2.0, name="a") + b = pd.variable(name="b") + c = pd.mul(a,b) + sess.eval(targets=c, feed_dict={"b":3.0}) # returns 6.0 + ``` + +```python +close() +``` + +Closes the session and releases the scope that the session owns. + + +### Create a Local Session + +```python +session( + devices=None +) +``` + +Creates a new session. One session owns one global scope, so creating +multiple sessions will create different scopes. + +- *devices*: a single `string` or a list of `string` of device names, + the corresponding devices will be the computation devices for + `eval()`. If not specified, all available devices (e.g., all GPUs) + will be used. The user doesn't need to specify the CPU device since + it will be always used. Multiple sessions can use the same device. + + +#### Example + +```Python +a = paddle.constant(1.0) +b = paddle.constant(2.0) +c = a + b +sess = paddle.session(devices=["gpu:0", "gpu:1", "fpga:0"]) +sess.eval(c) +sess.close() +``` + +### Create a Remote Session + +```python +create_cloud_job( + name, + num_trainer, + mem_per_trainer, + gpu_per_trainer, + cpu_per_trainer, + num_ps, + mem_per_ps, + cpu_per_ps, +) +``` + +Creates a Paddle Cloud job. Fails if the job name exists. + +```python +get_cloud_job( + name +) +``` + +Gets a Paddle Cloud job. + +```python +remote_session( + job +) +``` + +- *job*: the Paddle Cloud job. + +#### Example + +```Python +reader = paddle.reader.recordio("/pfs/home/peter/mnist-train-*") # data stored on Paddle Cloud +image = reader.column(0) +label = reader.column(1) +fc1 = paddle.op.fc(image, size=256, act="sigmoid") +fc2 = paddle.op.fc(fc1, size=10, act="softmax") +cost = paddle.op.cross_entropy(fc2, label) +opt = paddle.optimizer.sgd(cost) + +job = paddle.create_cloud_job("test", 3, "1G", 1, 1, 2, "1G", 1) +sess = paddle.remote_ession(job) +for i in range(1000): + sess.eval(opt) +sess.close() +``` diff --git a/doc/design/refactor/src/compiler.graffle b/doc/design/refactor/src/compiler.graffle new file mode 100644 index 0000000000000000000000000000000000000000..8cc678fea3c820103e7ce81f7a5d625d6c1d92de Binary files /dev/null and b/doc/design/refactor/src/compiler.graffle differ diff --git a/doc/design/refactor/src/compiler.png b/doc/design/refactor/src/compiler.png new file mode 100644 index 0000000000000000000000000000000000000000..65d34f841afce9756def07dd8ecb9ca44e658bfe Binary files /dev/null and b/doc/design/refactor/src/compiler.png differ diff --git a/doc/design/ops/src/dist-graph.graffle b/doc/design/refactor/src/dist-graph.graffle similarity index 100% rename from doc/design/ops/src/dist-graph.graffle rename to doc/design/refactor/src/dist-graph.graffle diff --git a/doc/design/ops/src/dist-graph.png b/doc/design/refactor/src/dist-graph.png similarity index 100% rename from doc/design/ops/src/dist-graph.png rename to doc/design/refactor/src/dist-graph.png diff --git a/doc/design/refactor/src/distributed_architecture.graffle b/doc/design/refactor/src/distributed_architecture.graffle new file mode 100644 index 0000000000000000000000000000000000000000..f8496e57326c38de7468eb452a7713291d57653c Binary files /dev/null and b/doc/design/refactor/src/distributed_architecture.graffle differ diff --git a/doc/design/refactor/src/distributed_architecture.png b/doc/design/refactor/src/distributed_architecture.png new file mode 100644 index 0000000000000000000000000000000000000000..410c4510c6aab301dec95e6427fe80ac24e105fe Binary files /dev/null and b/doc/design/refactor/src/distributed_architecture.png differ diff --git a/doc/design/ops/src/local-graph.graffle b/doc/design/refactor/src/local-graph.graffle similarity index 100% rename from doc/design/ops/src/local-graph.graffle rename to doc/design/refactor/src/local-graph.graffle diff --git a/doc/design/ops/src/local-graph.png b/doc/design/refactor/src/local-graph.png similarity index 100% rename from doc/design/ops/src/local-graph.png rename to doc/design/refactor/src/local-graph.png diff --git a/doc/design/refactor/src/local_architecture.graffle b/doc/design/refactor/src/local_architecture.graffle new file mode 100644 index 0000000000000000000000000000000000000000..cc7783c45381f25ded0b898649322c81418ad317 Binary files /dev/null and b/doc/design/refactor/src/local_architecture.graffle differ diff --git a/doc/design/refactor/src/local_architecture.png b/doc/design/refactor/src/local_architecture.png new file mode 100644 index 0000000000000000000000000000000000000000..4b999538b7825c805292ee28b5e3256d5543bd09 Binary files /dev/null and b/doc/design/refactor/src/local_architecture.png differ diff --git a/doc/design/refactor/src/paddle-compile.graffle b/doc/design/refactor/src/paddle-compile.graffle new file mode 100644 index 0000000000000000000000000000000000000000..a6348cc3dbcaca923c6e794681b2edb85cb9f8f6 Binary files /dev/null and b/doc/design/refactor/src/paddle-compile.graffle differ diff --git a/doc/design/refactor/src/paddle-compile.png b/doc/design/refactor/src/paddle-compile.png new file mode 100644 index 0000000000000000000000000000000000000000..e0f13d551ac41afaec627a57dea79356464bf0bf Binary files /dev/null and b/doc/design/refactor/src/paddle-compile.png differ diff --git a/doc/design/refactorization.md b/doc/design/refactorization.md index ad801ca421ca31c84b0a6b0a18d1d625c87e0de5..629422e7743af666b42fd69fbff442ce15bef596 100644 --- a/doc/design/refactorization.md +++ b/doc/design/refactorization.md @@ -1,40 +1,40 @@ # Design Doc: Refactorization Overview -The goal of refactorizaiton include: +The goals of refactoring include: -1. Make it easy for external contributors to write new elementory computaiton operations. -1. Make the codebase clean and readable. -1. Introduce a new design of computation representation -- a computation graph of operators and variables. -1. The graph representation helps implementing auto-scalable and auto fault recoverable distributed computing. +1. Making it easy for external contributors to write new elementary computation operations. +1. Making the codebase clean and readable. +1. Designing a new computation representation -- a computation graph of operators and variables. +1. Implementing auto-scalability and auto fault recoverable distributed computing with the help of computation graphs. ## Computation Graphs -1. PaddlePaddle represent the computation, training and inference of DL models, by computation graphs. +1. PaddlePaddle represents the computation, training and inference of Deep Learning models, by computation graphs. - 1. Please dig into [computation graphs](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/graph.md) for a solid example. + 1. Please refer to [computation graphs](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/graph.md) for a concrete example. -1. Users write Python programs to describe the graphs and run it (locally or remotely). +1. Users write Python programs to describe the graphs and run them (locally or remotely). 1. A graph is composed of *variables* and *operators*. -1. The description of graphs must be able to be serialized/deserialized, so it +1. The description of graphs must be capable of being serialized/deserialized, so that: - 1. could to be sent to the cloud for distributed execution, and - 1. be sent to clients for mobile or enterprise deployment. + 1. It can to be sent to the cloud for distributed execution, and + 1. It can be sent to clients for mobile or enterprise deployment. -1. The Python program do +1. The Python program does the following steps - 1. *compilation*: runs a Python program to generate a protobuf message representation of the graph and send it to + 1. *compilation*: run a Python program to generate a protobuf message representation of the graph and send it to 1. the C++ library `libpaddle.so` for local execution, 1. the master process of a distributed training job for training, or 1. the server process of a Kubernetes serving job for distributed serving. - 1. *execution*: according to the protobuf message, constructs instances of class `Variable` and `OperatorBase`, and run them. + 1. *execution*: execute the graph by constructing instances of class [`Variable`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/variable.h#L24) and [`OperatorBase`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/operator.h#L70), according to the protobuf message. -## Description and Realization +## Description and Realization of Computation Graph -At compile time, the Python program generates protobuf message representation of the graph, or the description of the graph. +At compile time, the Python program generates a protobuf message representation of the graph, or the description of the graph. -At runtime, the C++ program realizes the graph and run it. +At runtime, the C++ program realizes the graph and runs it. | | Representation (protobuf messages) | Realization (C++ class objects) | |---|---|---| @@ -42,30 +42,31 @@ At runtime, the C++ program realizes the graph and run it. |Operation|[OpDesc](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/framework.proto#L35)|[Operator](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/operator.h#L64)| |Block|BlockDesc|Block| -The word *graph* is exchangable with *block* in this document. A graph represent computation steps and local variables as a C++/Java program block, or a pair of { and }. +The word *graph* is interchangeable with *block* in this document. A graph represents computation steps and local variables similar to a C++/Java program block, or a pair of parentheses(`{` and `}`). ## Compilation and Execution -1. Run an applicaton Python program to describe the graph. In particular, +1. Run an application Python program to describe the graph. In particular, the Python application program does the following: - 1. create VarDesc to represent local/intermediate variables, - 1. create operators and set attributes, - 1. validate attribute values, - 1. inference the type and the shape of variables, - 1. plan for memory-reuse for variables, - 1. generate backward and optimization part of the Graph. - 1. possiblly split the graph for distributed training. + 1. Create `VarDesc` to represent local/intermediate variables, + 1. Create operators and set attributes, + 1. Validate attribute values, + 1. Infer the type and the shape of variables, + 1. Plan memory-reuse for variables, + 1. Generate the backward graph + 1. Optimize the computation graph. + 1. Potentially, split the graph for distributed training. -1. The invocation of `train` or `infer` in the application Python program: +1. The invocation of `train` or [`infer`](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/inference.py#L108) methods in the application Python program does the following: - 1. create a new Scope instance in the [scope hierarchy](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/scope.md) for each run of a block, + 1. Create a new Scope instance in the [scope hierarchy](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/scope.md) for each run of a block, 1. realize local variables defined in the BlockDesc message in the new scope, 1. a scope is similar to the stack frame in programming languages, - 1. create an instance of class `Block`, in which, + 1. Create an instance of class `Block`, in which, 1. realize operators in the BlockDesc message, - 1. run the Block by calling + 1. Run the Block by calling 1. `Block::Eval(vector* targets)` for forward and backward computations, or 1. `Block::Eval(vector* targets)` for optimization. @@ -76,14 +77,14 @@ The word *graph* is exchangable with *block* in this document. A graph represen Compile Time -> IR -> Runtime ``` -### Benefit +### Benefits of IR - Optimization ```text Compile Time -> IR -> Optimized IR -> Runtime ``` -- Send automatically partitioned IR to different nodes. - - Automatic data parallel +- Automatically send partitioned IR to different nodes. + - Automatic Data Parallelism ```text Compile Time |-> Single GPU IR @@ -92,7 +93,7 @@ Compile Time -> IR -> Runtime |-> Node-1 (runs trainer-IR-1) |-> Node-2 (runs pserver-IR) ``` - - Automatic model parallel (planned for future) + - Automatic Model Parallelism (planned for future) --- @@ -105,10 +106,10 @@ Compile Time -> IR -> Runtime # Operator ![class_diagram](http://api.paddlepaddle.org/graphviz?dot=https://gist.githubusercontent.com/reyoung/53df507f6749762675dff3e7ce53372f/raw/dd598e8f1976f5759f58af5e5ef94738a6b2e661/op.dot) -* `Operator` is the fundamental building block as the user interface. - * Operator stores input/output variable name, and attributes. - * The `InferShape` interface is used to infer output variable shapes by its input shapes. - * Use `Run` to compute `input variables` to `output variables`. +* `Operator` is the fundamental building block of the user interface. + * Operator stores input/output variable names, and attributes. + * The `InferShape` interface is used to infer the shape of the output variable shapes based on the shapes of the input variables. + * Use `Run` to compute the `output` variables from the `input` variables. --- @@ -126,30 +127,29 @@ Compile Time -> IR -> Runtime # Why separate Kernel and Operator * Separate GPU and CPU code. - * Make Paddle can run without GPU. -* Make one operator (which is user interface) can contain many implementations. - * Same mul op, different FP16, FP32 Kernel. different MKL, eigen kernel. + * Make Paddle capable of running without GPU. +* Make one operator (which is a user interface) and create many implementations. + * For example, same multiplication op can have different implementations kernels such as FP16 kernel, FP32 kernel, MKL, eigen kernel. --- # Libraries for Kernel development * `Eigen::Tensor` contains basic math and element-wise functions. * Note that `Eigen::Tensor` has broadcast implementation. - * Limit number of `tensor.device(dev) = ` in your code. -* `thrust::tranform` and `std::transform`. - * `thrust` has the same API as C++ standard library. Using `transform` can quickly implement a customized elementwise kernel. - * `thrust` has more complex API, like `scan`, `reduce`, `reduce_by_key`. + * Limit the number of `tensor.device(dev) = ` in your code. +* `thrust::transform` and `std::transform`. + * `thrust` has the same API as C++ standard library. Using `transform`, one can quickly implement customized element-wise kernels. + * `thrust` also has more complex APIs, like `scan`, `reduce`, `reduce_by_key`. * Hand-writing `GPUKernel` and `CPU` code - * Do not write `.h`. CPU Kernel should be in `.cc`. GPU kernel should be in `.cu`. (`GCC` cannot compile GPU code.) + * Do not write in header (`.h`) files. CPU Kernel should be in cpp source (`.cc`) and GPU kernels should be in cuda (`.cu`) files. (GCC cannot compile GPU code.) --- -# Operator Register +# Operator Registration -## Why register is necessary? +## Why is registration necessary? We need a method to build mappings between Op type names and Op classes. -## How to do the register? - -Maintain a map, whose key is the type name and value is corresponding Op constructor. +## How is registration implemented? +Maintaining a map, whose key is the type name and the value is the corresponding Op constructor. --- # The Registry Map @@ -169,7 +169,7 @@ Maintain a map, whose key is the type name and value is corresponding Op constru # Related Concepts ### Op_Maker -It's constructor takes `proto` and `checker`. They are compeleted during Op_Maker's construction. ([ScaleOpMaker](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/scale_op.cc#L37)) +It's constructor takes `proto` and `checker`. They are completed during Op_Maker's construction. ([ScaleOpMaker](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/scale_op.cc#L37)) ### Register Macros ```cpp @@ -177,34 +177,34 @@ REGISTER_OP(op_type, op_class, op_maker_class, grad_op_type, grad_op_class) REGISTER_OP_WITHOUT_GRADIENT(op_type, op_class, op_maker_class) ``` -### `USE` Macros -make sure the registration process is executed and linked. +### USE Macros +Make sure the registration process is executed and linked. --- -# Register Process -1. Write Op class, as well as its gradient Op class if there is. -2. Write Op maker class. In the constructor, describe its inputs, outputs, and attributes. -3. Invoke macro `REGISTER_OP`. The macro will - 1. call maker class to complete `proto` and `checker` - 2. with the completed `proto` and `checker`, build a new key-value pair in the `OpInfoMap` +# Registration Process +1. Write an Op class and its gradient Op class, if required. +2. Write an Op maker class. In the constructor of this class, describe the inputs, outputs and attributes of the operator. +3. Invoke the macro `REGISTER_OP`. This macro will + 1. Call maker class to complete the `proto` and the `checker` + 2. Using the completed `proto` and `checker`, it will add a new key-value pair to the `OpInfoMap` -4. Invoke `USE` macro in where the Op is used to make sure it is linked. +4. Invoke the `USE` macro in which the Op is used, to make sure that it is linked. --- # Backward Module (1/2) ### Create Backward Operator -- Mapping from forwarding Op to backward Op +- Mapping from forward Op to backward Op ![backward](https://gist.githubusercontent.com/dzhwinter/a6fbd4623ee76c459f7f94591fd1abf0/raw/61026ab6e518e66bde66a889bc42557a1fccff33/backward.png) --- # Backward Module (2/2) ### Build Backward Network -- **Input** graph of forwarding operators -- **Output** graph of backward operators -- **corner case in construction** - - shared variable => insert `Add` operator - - no gradient => insert `fill_zero_grad` operator - - recursive netOp => call `Backward` recursively +- **Input**: graph of forward operators +- **Output**: graph of backward operators +- **Corner cases in construction** + - Shared Variables => insert an `Add` operator to combine gradients + - No Gradient => insert a `fill_zero_grad` operator + - Recursive NetOp => call `Backward` recursively - RNN Op => recursively call `Backward` on stepnet @@ -213,41 +213,41 @@ make sure the registration process is executed and linked. * `Tensor` is an n-dimension array with type. * Only dims and data pointers are stored in `Tensor`. - * All operators on `Tensor` is written in `Operator` or global functions. - * variable length Tensor design [LoDTensor](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/lod_tensor.md) -* `Variable` is the inputs and outputs of an operator. Not just `Tensor`. - * step_scopes in RNN is a variable and not a tensor. -* `Scope` is where variables store at. - * map - * `Scope` has a hierarchical structure. The local scope can get variable from its parent scope. + * All operations on `Tensor` are written in `Operator` or global functions. + * Variable length Tensor design [LoDTensor](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/lod_tensor.md) +* `Variable` instances are the inputs and the outputs of an operator. Not just `Tensor`. + * `step_scopes` in RNN is a variable and not a tensor. +* `Scope` is where variables are stores. + * map + * `Scope` has a hierarchical structure. The local scope can get variables from its parent scope. --- # Block (in design) -## the difference with original RNNOp -- as an operator is more intuitive than `RNNOp`, -- offers new interface `Eval(targets)` to deduce the minimal block to `Run`, -- fits the compile-time/ runtime separation design. - - during the compilation, `SymbolTable` stores `VarDesc`s and `OpDesc`s and serialize to a `BlockDesc` - - when graph executes, a Block with `BlockDesc` passed in creates `Op` and `Var` then `Run` +## the difference between original RNNOp and Block +- As an operator is more intuitive than `RNNOp`, +- Offers a new interface `Eval(targets)` to deduce the minimal block to `Run`, +- Fits the compile-time/ runtime separation design paradigm. + - During the compilation, `SymbolTable` stores `VarDesc`s and `OpDesc`s and serialize to a `BlockDesc` + - When graph executes, a Block with `BlockDesc` is passed. It then creates `Op` and `Var` instances and then invokes `Run`. --- # Milestone -- take Paddle/books as the main line, the requirement of the models motivates framework refactoring, -- model migration - - framework development gives **priority support** to model migration, for example, +- Take Paddle/books as the main line, the requirement of the models motivates framework refactoring, +- Model migration + - Framework development gives **priority support** to model migration, for example, - the MNIST demo needs a Python interface, - the RNN models require the framework to support `LoDTensor`. - - determine some timelines, - - heavily-relied Ops need to be migrated first, - - different models can be migrated parallelly. -- improve the framework at the same time -- accept imperfection, concentrated on solving the specific problem at the right price. + - Determine some timelines, + - Frequently used Ops need to be migrated first, + - Different models can be migrated in parallel. +- Improve the framework at the same time +- Accept imperfection, concentrate on solving the specific problem at the right price. --- # Control the migration quality -- compare the performance of migrated models with old ones. -- follow google C style -- build the automatic workflow of generating Python/C++ documentations - - the documentation of layers and ops should be written inside the code - - take the documentation quality into account when doing PR - - preview the documentations, read and improve them from users' perspective +- Compare the performance of migrated models with old ones. +- Follow the google C++ style +- Build the automatic workflow of generating Python/C++ documentations. + - The documentation of layers and ops should be written inside the code. + - Take the documentation quality into account when submitting pull requests. + - Preview the documentations, read and improve them from a user's perspective. diff --git a/doc/design/register_grad_op.md b/doc/design/register_grad_op.md new file mode 100644 index 0000000000000000000000000000000000000000..3cf8a59446d244bb3a388b87b14273d9096c839a --- /dev/null +++ b/doc/design/register_grad_op.md @@ -0,0 +1,90 @@ +# Design Doc: Gradient Operators Registration + + +## The Problem Posed + +In our current operator registration mechanism, for each operator, the programmer should register a *gradient operator creator* function, which takes a C++ operator instance, and returns the corresponding gradient instance. + +However, as we decided to separate the *compilation* and *execution* of DL models, we need to reshape the creator to take a protobuf `OpDesc` message, and returns a corresponding message. + +More than that, the new registration mechanism need to support the fact that an operators' gradient computation might be a composition of operators. + +## Current Implementation + +OpInfos store in a association map which key is the operator type. The `grad_op_type` indicate associated gradient operator type. Operator can create gradient operator by `OpInfo::creator_` of gradient. The pseudo code is + +```cpp +struct OpInfo { + std::function creator_; + std::string grad_op_type_; + ... +}; + +map OpInfoMap; + +OperatorBase* CreateGradientOperator(const OperatorBase& op) { + return OpInfoMap.at(op.Type()).creator_(...); +} +``` + +## Proposed Solution + +The mapping relationship between an operator and its gradient operators is a function. The interface of that function is: + +```cpp +// (OpDesc) --> vector +std::function(const OpDescBind&)>; +``` + +The function takes an `OpDescBind` of the forward operator and returns one or many gradient operator descriptions. `OpDescBind` is a C++ wrapper for protobuf message `OpDesc` to manipulate `OpDesc` fast. + +The `GradOpDescMaker` will be registered in `OpInfo`, to replace `grad_op_type_` field. The `OpInfo` should be + +```cpp +struct OpInfo { + std::function>(const OpDescBind&)> grad_op_maker_; + ... +}; +``` + +The `grad_op_maker_ ` is `nullptr` if the operator does not have associated gradient operators. + +We propose a base class called `GradOpDescMakerBase` to let operator developers generate `Gradient Operators` easily. The public interface of that class is + +```cpp +class GradOpDescMakerBase { +public: + GradOpDescMakerBase(const OpDescBind& ); + virtual std::vector> operator()()const = 0; +}; +``` + +We can convert `GradOpDescMakerBase` to `std::function>(const OpDescBind&)>` by + +```cpp +using GradOpMaker = ...; +std::function(const OpDescBind&)> func; +func = [] (const OpDescBind& fwd_op) { + GradOpMaker maker(fwd_op); + return maker(); +}; +``` + +We can write many helper functions since the `GradOpDescMakerBase` is a class now. The basic helper functions get the variables of `Input`, `Output`, `InputGradient` and `OutputGradient` in the forwarding operator. + +We should chagne register macros at the same time. In the current solution, there is no difference between forwarding operators and backward operators. So `REGISTER_OP` just register one operator. If the `REGISTER_OPERATOR ` contains `OpProtoAndCheckerMaker` and `GradOpDescMaker`, we just list them in the same macro. It can be done by a macro contains `__VA_ARGS__`. + +The user interface should be + +```cpp +vector MinusOpGradMaker(OpDesc) {...} +REGISTER_OPERATOR(minus, MinusOp, MinusOpProtoAndCheckerMaker, SumOpGradMaker); +// Developers can still manually implement gradient operator. +REGISTER_OPERATOR(minus_grad, MinusGradOp); +``` + +The interface of current `REGISTER_OP` macro could not be changed. In `REGISTER_OP`, it will invoke `REGISTER_OPERATOR` two times and generate GradOpDescMaker inside. + +```cpp +REGISTER_OP(minus, MinusOp, MinusOpProtoAndCheckerMaker, minus_grad, MinusGradOp); +``` diff --git a/doc/design/tensor_array.md b/doc/design/tensor_array.md new file mode 100644 index 0000000000000000000000000000000000000000..8378e97bf7cfaae54c36b1b92e202b16e4fe1e28 --- /dev/null +++ b/doc/design/tensor_array.md @@ -0,0 +1,271 @@ +# Design for TensorArray +This design doc presents the necessity of a new C++ class `TensorArray`. +In addition to the very simple C++ implementation + +```c++ +class TensorArray { + public: + explicit TensorArray(const LoDTensor&); + explicit TensorArray(size_t size); + + private: + vector values_; +}; +``` + +We also need to expose it to PaddlePaddle's Python API, +because users would want to use it with our very flexible operators `WhileLoop`. +An example for a RNN based on dynamic operators is + +```python +input = pd.data(...) +num_steps = Var(12) + +TensorArray states(size=num_steps) +TensorArray step_inputs(unstack_from=input) +TensorArray step_outputs(size=num_steps) + +W = Tensor(...) +U = Tensor(...) +default_state = some_op() + +step = Var(1) + +wloop = paddle.create_whileloop(loop_vars=[step]) +with wloop.frame(): + wloop.break_if(pd.equal(step, num_steps) + pre_state = states.read(step-1, default_state) + step_input = step_inputs.read(step) + state = pd.sigmoid(pd.matmul(U, pre_state) + pd.matmul(W, step_input)) + states.write(step, state) + step_outputs.write(step, state) # output state + step.update(state+1) + +output = step_outputs.stack() +``` + +## Background +Steps are one of the core concepts of RNN. In each time step of RNN, there should be several input segments, states, and output segments; all these components act like arrays, for example, call `states[step_id]` will get the state in `step_id`th time step. + +An RNN can be implemented with the following pseudocode + +```c++ +Array states; +Array input_segments; +Array output_segments; +Parameter W, U; + +step = 1 +seq_len = 12 +while_loop { + if (step == seq_len) break; + states[step] = sigmoid(W * states[step-1] + U * input_segments[step]); + output_segments[step] = states[step] // take state as output + step++; +} +``` +According to the [RNN roadmap](https://github.com/PaddlePaddle/Paddle/issues/4561), there are several different RNNs that PaddlePaddle will eventually support. + +Currently, the basic RNN implementation supported by PaddlePaddle is the `recurrent_op` which takes tensors as input and splits them into `input_segments`. + + +Since a tensor cannot store variable-length sequences directly, PaddlePaddle implements the tensor with level of details (`LoDTensor` for short). +Segmenting the `LoDTensor` is much more complicated than splitting a tensor, that makes it necessary to refactor the `recurrent_op` with `LoDTensor` segmenting support. + +As the next step in RNN support, `dynamic_recurrent_op` should be introduced to handle inputs with variable-length sequences. + +The implementation is similar to `recurrent_op`. +The key difference is the way **the original input `LoDTensors` and outupts are split to get the `input_segments` and the `output_segments`.** + + +Though it can't be built over `recurrent_op` or `dynamic_recurrent_op` directly, +the logic behind splitting a tensor or a LoD tensor into `input_segments` remains the same. + +## Why `TensorArray` +The logic behind splitting the inputs to segments, states and outputs is similar and can be shared in a seperate module. + +The array of `states`, `input_segments` and `output_segments` would be exposed to users when writing a dynamic RNN model similar to the above pseudo codes. + +So there should be an array-like container, which can store the segments of a tensor or LoD tensor. + +**This container can store an array of tensors and provides several methods to split a tensor or a LoD tensor** . +This is where the notion of `TensorArray` comes from. + +## Introduce TensorArray to uniform all the three RNNs +TensorArray as a new concept is borrowed from TensorFlow, +it is meant to be used with dynamic iteration primitives such as `while_loop` and `map_fn`. + +This concept can be used to support our new design of dynamic operations, and help to refactor some existing variant-sentence-related layers, +such as `recurrent_op`, `RecurrentGradientMachine`. + +In [our design for dynamic RNN](https://github.com/PaddlePaddle/Paddle/pull/4401), +`TensorArray` is used to segment inputs and store states in all time steps. +By providing some methods similar to a C++ array, +the definition of some state-based dynamic models such as RNN can be more natural and highly flexible. + +## Dynamic-operations on TensorArray + +`TensorArray` will be used directly when defining dynamic models, so some operators listed below should be implemented + +```python +# several helper operators for TensorArray +def tensor_array_stack(ta, tensor): + ''' + get a tensor array `ta`, return a packed `tensor`. + ''' + pass + +def tensor_array_unstack(tensor, ta): + ''' + get a `tensor`, unstack it and get a tensor array `ta`. + ''' + pass + +def tensor_array_write(ta, index, tensor, data_shared): + ''' + get a `tensor` and a scalar tensor `index`, write `tensor` into index-th + value of the tensor array `ta`. + `data_shared` is an attribute that specifies whether to copy or reference the tensors. + ''' + pass + +def tensor_array_read(ta, index, tensor): + ''' + get a tensor array `ta`, a scalar tensor `index`, read the index-th value of + `ta` and return as the `tensor`. + ''' + pass + +def tensor_array_size(ta, tensor): + ''' + get a tensor array `ta`, return the size of `ta` and return as the scalar `tensor`. + ''' + pass +``` + +It is trivial for users to use so many low-level operators, so some helper methods should be proposed in python wrapper to make `TensorArray` easier to use, +for example + +```python +class TensorArray: + def __init__(self, name): + self.name = name + self.desc = TensorArrayDesc() + + def stack(self, name=None): + ''' + Pack the values in a `TensorArray` into a tensor with rank one higher + than each tensor in `values`. + `stack` can be used to split tensor into time steps for RNN or whileloop. + + @name: str + the name of the variable to output. + ''' + tensor = NewVar(name) + tensor_array_stack(self.name, tensor) + return tensor + + def unstack(self, input): + ''' + Unpacks the given dimension of a rank-`R` tensor into rank-`(R-1)` tensors. + `unstack` can be used to concatenate all the time steps for RNN or whileloop. + + @input: str + the name of input tensor + ''' + tensor_array_unstack(tensor, self.name) + + def write(self, index, value, data_shared=True): + ''' + Write value into index of the TensorArray. + If `data_shared` is set to True, than the index-th value in TensorArray will + be shared with the tensor passed in. + + @index: str + name of a scalar tensor + @value: str + name of a tensor + @data_shared: bool + ''' + tensor_array_write(self.name, index, value, data_shared) + + def read(self, index, output): + ''' + Read the value at location `index` in the `TensorArray`. + + @index: str + name of a scalar tensor + @output: + name of a output variable + ''' + tensor_array_read(self.name, index, output) + + + def size(self, output): + ''' + Return the number of values. + + @output: str + name of a scalar tensor + ''' + tensor_array_size(self.name, output) +``` + +## LoDTensor-related Supports +The `RecurrentGradientMachine` in Paddle serves as a flexible RNN layer; it takes varience-length sequences as input, and output sequences too. + +Since each step of RNN can only take a tensor-represented batch of data as input, +some preprocess should be taken on the inputs such as sorting the sentences by their length in descending order and cut each word and pack to new batches. + +Such cut-like operations can be embedded into `TensorArray` as general methods called `unpack` and `pack`, +these two operations are similar to `stack` and `unstack` except that they operate on variable-length sequences formated as a LoD tensor rather than a tensor. + +Some definitions are like + +```python +def unpack(level): + ''' + Split LodTensor in some `level` and generate batches, if set `sort_by_length`, + will sort by length. + + Returns: + - a new `TensorArray`, whose values are LodTensors and represents batches + of data. + - an int32 Tensor, which stores the map from the new batch's indices to + original LoDTensor + ''' + pass + +def pack(level, indices_map): + ''' + Recover the original LoD-arranged LoDTensor with the values in a `TensorArray` + and `level` and `indices_map`. + ''' + pass +``` + +With these two methods, a varience-length sentence supported RNN can be implemented like + +```c++ +// input is the varient-length data +LodTensor sentence_input(xxx); +TensorArray ta; +Tensor indice_map; +Tensor boot_state = xxx; // to initialize rnn's first state +TensorArray::unpack(input, 1/*level*/, true/*sort_by_length*/, &ta, &indice_map); +TessorArray step_outputs; +TensorArray states; + +for (int step = 0; step = ta.size(); step++) { + auto state = states.read(step); + // rnnstep is a function which acts like a step of RNN + auto step_input = ta.read(step); + auto step_output = rnnstep(step_input, state); + step_outputs.write(step_output, true/*data_shared*/); +} + +// rnn_output is the final output of an rnn +LoDTensor rnn_output = ta.pack(ta, indice_map); +``` +the code above shows that by embedding the LoDTensor-related preprocess operations into `TensorArray`, +the implementation of a RNN that supports varient-length sentences is far more concise than `RecurrentGradientMachine` because the latter mixes all the codes together, hard to read and extend. diff --git a/doc/faq/build_and_install/index_cn.rst b/doc/faq/build_and_install/index_cn.rst new file mode 100644 index 0000000000000000000000000000000000000000..f1677e216f31d79b53ac29a0afbf6fbb886a0dcd --- /dev/null +++ b/doc/faq/build_and_install/index_cn.rst @@ -0,0 +1,111 @@ +################### +编译安装与单元测试 +################### + +.. contents:: + +1. 运行Docker GPU镜像出现 "CUDA driver version is insufficient" +---------------------------------------------------------------- + +用户在使用PaddlePaddle GPU的Docker镜像的时候，常常出现 `Cuda Error: CUDA driver version is insufficient for CUDA runtime version`, 原因在于没有把机器上CUDA相关的驱动和库映射到容器内部。 +具体的解决方法是： + +.. code-block:: bash + + $ export CUDA_SO="$(\ls usr/lib64/libcuda* | xargs -I{} echo '-v {}:{}') $(\ls /usr/lib64/libnvidia* | xargs -I{} echo '-v {}:{}')" + $ export DEVICES=$(\ls /dev/nvidia* | xargs -I{} echo '--device {}:{}') + $ docker run ${CUDA_SO} ${DEVICES} -it paddledev/paddlepaddle:latest-gpu + +更多关于Docker的安装与使用, 请参考 `PaddlePaddle Docker 文档 `_ 。 + + +2. CMake源码编译, 找到的PythonLibs和PythonInterp版本不一致 +---------------------------------------------------------------- + +这是目前CMake寻找Python的逻辑存在缺陷，如果系统安装了多个Python版本，CMake找到的Python库和Python解释器版本可能有不一致现象，导致编译PaddlePaddle失败。正确的解决方法是， +用户强制指定特定的Python版本，具体操作如下： + + .. code-block:: bash + + cmake .. -DPYTHON_EXECUTABLE= -DPYTHON_LIBRARY= -DPYTHON_INCLUDE_DIR= + +用户需要指定本机上Python的路径：````, ````, ```` + +3. CMake源码编译，Paddle版本号为0.0.0 +-------------------------------------- + +如果运行 :code:`paddle version`, 出现 :code:`PaddlePaddle 0.0.0`；或者运行 :code:`cmake ..`，出现 + +.. code-block:: bash + + CMake Warning at cmake/version.cmake:20 (message): + Cannot add paddle version from git tag + +那么用户需要拉取所有的远程分支到本机，命令为 :code:`git fetch upstream`，然后重新cmake即可。 + +4. paddlepaddle\*.whl is not a supported wheel on this platform. +------------------------------------------------------------------------ + +出现这个问题的主要原因是，没有找到和当前系统匹配的paddlepaddle安装包。最新的paddlepaddle python安装包支持Linux x86_64和MacOS 10.12操作系统，并安装了python 2.7和pip 9.0.1。 + +更新 :code:`pip` 包的方法是\: + +.. code-block:: bash + + pip install --upgrade pip + +如果还不行，可以执行 :code:`python -c "import pip; print(pip.pep425tags.get_supported())"` 获取当前系统支持的python包的后缀， +并对比是否和正在安装的后缀一致。 + +如果系统支持的是 :code:`linux_x86_64` 而安装包是 :code:`manylinux1_x86_64` ，需要升级pip版本到最新； +如果系统支持 :code:`manylinux1_x86_64` 而安装包（本地）是 :code:`linux_x86_64` ，可以重命名这个whl包为 :code:`manylinux1_x86_64` 再安装。 + +5. 编译安装后执行 import paddle.v2 as paddle 报ImportError: No module named v2 +------------------------------------------------------------------------------------------ +先查看一下是否曾经安装过paddle v1版本，有的话需要先卸载： + +pip uninstall py_paddle paddle + +然后安装paddle的python环境, 在build目录下执行 + +pip install python/dist/paddle*.whl && pip install ../paddle/dist/py_paddle*.whl + +6. 遇到“非法指令”或者是“illegal instruction” +-------------------------------------------- + +PaddlePaddle使用avx SIMD指令提高cpu执行效率，因此错误的使用二进制发行版可能会导致这种错误，请选择正确的版本。 + +7. python相关的单元测试都过不了 +-------------------------------- + +如果出现以下python相关的单元测试都过不了的情况： + +.. code-block:: bash + + 24 - test_PyDataProvider (Failed) + 26 - test_RecurrentGradientMachine (Failed) + 27 - test_NetworkCompare (Failed) + 28 - test_PyDataProvider2 (Failed) + 32 - test_Prediction (Failed) + 33 - test_Compare (Failed) + 34 - test_Trainer (Failed) + 35 - test_TrainerOnePass (Failed) + 36 - test_CompareTwoNets (Failed) + 37 - test_CompareTwoOpts (Failed) + 38 - test_CompareSparse (Failed) + 39 - test_recurrent_machine_generation (Failed) + 40 - test_PyDataProviderWrapper (Failed) + 41 - test_config_parser (Failed) + 42 - test_swig_api (Failed) + 43 - layers_test (Failed) + +并且查询PaddlePaddle单元测试的日志，提示： + +.. code-block:: bash + + paddle package is already in your PYTHONPATH. But unittest need a clean environment. + Please uninstall paddle package before start unittest. Try to 'pip uninstall paddle'. + +解决办法是： + +* 卸载PaddlePaddle包 :code:`pip uninstall paddle`, 清理掉老旧的PaddlePaddle安装包，使得单元测试有一个干净的环境。如果PaddlePaddle包已经在python的site-packages里面，单元测试会引用site-packages里面的python包，而不是源码目录里 :code:`/python` 目录下的python包。同时，即便设置 :code:`PYTHONPATH` 到 :code:`/python` 也没用，因为python的搜索路径是优先已经安装的python包。 diff --git a/doc/faq/cluster/index_cn.rst b/doc/faq/cluster/index_cn.rst new file mode 100644 index 0000000000000000000000000000000000000000..e59c1e1a54a0c876d1e6e89f88030de59fb9fc1a --- /dev/null +++ b/doc/faq/cluster/index_cn.rst @@ -0,0 +1,17 @@ +############### +集群训练与预测 +############### + +.. contents:: + +1. 集群多节点训练，日志中保存均为网络通信类错误 +------------------------------------------------ + +集群多节点训练，日志报错为网络通信类错误，比如 :code:`Connection reset by peer` 等。 +此类报错通常是由于某一个节点的错误导致这个节点的训练进程退出，从而引发其他节点无法连接导致，可以参考下面的步骤排查： + +* 从 :code:`train.log` ， :code:`server.log` 找到最早报错的地方，查看是否是其他错误引发的报错（比如FPE，内存不足，磁盘空间不足等）。 + +* 如果发现最早的报错就是网络通信的问题，很有可能是非独占方式执行导致的端口冲突，可以联系OP，看当前MPI集群是否支持resource=full参数提交，如果支持增加此参数提交，并更换job 端口。 + +* 如果当前MPI集群并不支持任务独占模式，可以联系OP是否可以更换集群或升级当前集群。 diff --git a/doc/faq/index_cn.rst b/doc/faq/index_cn.rst index b3ecfba791ead8349ded018a30059b03eacbdacd..9929767cac212237b3e2c3a547ba9a3c9d5f0979 100644 --- a/doc/faq/index_cn.rst +++ b/doc/faq/index_cn.rst @@ -1,514 +1,11 @@ -#################### FAQ -#################### +==== -.. contents:: - -1. 如何减少内存占用 ---------------------------------- - -神经网络的训练本身是一个非常消耗内存和显存的工作，经常会消耗数10GB的内存和数GB的显存。 -PaddlePaddle的内存占用主要分为如下几个方面\: - -* DataProvider缓冲池内存（只针对内存） -* 神经元激活内存（针对内存和显存） -* 参数内存 （针对内存和显存） -* 其他内存杂项 - -其中，其他内存杂项是指PaddlePaddle本身所用的一些内存，包括字符串分配，临时变量等等，暂不考虑在内。 - -减少DataProvider缓冲池内存 -++++++++++++++++++++++++++ - -PyDataProvider使用的是异步加载，同时在内存里直接随即选取数据来做Shuffle。即 - -.. graphviz:: - - digraph { - rankdir=LR; - 数据文件 -> 内存池 -> PaddlePaddle训练 - } - -所以，减小这个内存池即可减小内存占用，同时也可以加速开始训练前数据载入的过程。但是，这 -个内存池实际上决定了shuffle的粒度。所以，如果将这个内存池减小，又要保证数据是随机的， -那么最好将数据文件在每次读取之前做一次shuffle。可能的代码为 - -.. literalinclude:: src/reduce_min_pool_size.py - -这样做可以极大的减少内存占用，并且可能会加速训练过程，详细文档参考 :ref:`api_pydataprovider2` 。 - -神经元激活内存 -++++++++++++++ - -神经网络在训练的时候，会对每一个激活暂存一些数据，如神经元激活值等。 -在反向传递的时候，这些数据会被用来更新参数。这些数据使用的内存主要和两个参数有关系， -一是batch size，另一个是每条序列(Sequence)长度。所以，其实也是和每个mini-batch中包含 -的时间步信息成正比。 - -所以做法可以有两种： - -* 减小batch size。 即在网络配置中 :code:`settings(batch_size=1000)` 设置成一个小一些的值。但是batch size本身是神经网络的超参数，减小batch size可能会对训练结果产生影响。 -* 减小序列的长度，或者直接扔掉非常长的序列。比如，一个数据集大部分序列长度是100-200, - 但是突然有一个10000长的序列，就很容易导致内存超限，特别是在LSTM等RNN中。 - -参数内存 -++++++++ - -PaddlePaddle支持非常多的优化算法(Optimizer)，不同的优化算法需要使用不同大小的内存。 -例如使用 :code:`adadelta` 算法，则需要使用等于权重参数规模大约5倍的内存。举例，如果参数保存下来的模型目录 -文件为 :code:`100M`， 那么该优化算法至少需要 :code:`500M` 的内存。 - -可以考虑使用一些优化算法，例如 :code:`momentum`。 - -2. 如何加速PaddlePaddle的训练速度 ---------------------------------- - -加速PaddlePaddle训练可以考虑从以下几个方面\： - -* 减少数据载入的耗时 -* 加速训练速度 -* 利用分布式训练驾驭更多的计算资源 - -减少数据载入的耗时 -++++++++++++++++++ - -使用\ :code:`pydataprovider`\ 时，可以减少缓存池的大小，同时设置内存缓存功能，即可以极大的加速数据载入流程。 -:code:`DataProvider` 缓存池的减小，和之前减小通过减小缓存池来减小内存占用的原理一致。 - -.. literalinclude:: src/reduce_min_pool_size.py - -同时 :code:`@provider` 接口有一个 :code:`cache` 参数来控制缓存方法，将其设置成 :code:`CacheType.CACHE_PASS_IN_MEM` 的话，会将第一个 :code:`pass` (过完所有训练数据即为一个pass)生成的数据缓存在内存里，在之后的 :code:`pass` 中，不会再从 :code:`python` 端读取数据，而是直接从内存的缓存里读取数据。这也会极大减少数据读入的耗时。 - - -加速训练速度 -++++++++++++ - -PaddlePaddle支持Sparse的训练，sparse训练需要训练特征是 :code:`sparse_binary_vector` 、 :code:`sparse_vector` 、或者 :code:`integer_value` 的任一一种。同时，与这个训练数据交互的Layer，需要将其Parameter设置成 sparse 更新模式，即设置 :code:`sparse_update=True` - -这里使用简单的 :code:`word2vec` 训练语言模型距离，具体使用方法为\: - -使用一个词前两个词和后两个词，来预测这个中间的词。这个任务的DataProvider为\: - -.. literalinclude:: src/word2vec_dataprovider.py - -这个任务的配置为\: - -.. literalinclude:: src/word2vec_config.py - - -利用更多的计算资源 -++++++++++++++++++ - -利用更多的计算资源可以分为一下几个方式来进行\: - -* 单机CPU训练 - - * 使用多线程训练。设置命令行参数 :code:`trainer_count`。 - -* 单机GPU训练 - - * 使用显卡训练。设置命令行参数 :code:`use_gpu`。 - * 使用多块显卡训练。设置命令行参数 :code:`use_gpu` 和 :code:`trainer_count` 。 - -* 多机训练 - - * 请参考 :ref:`cluster_train` 。 - - -3. 遇到“非法指令”或者是“illegal instruction” --------------------------------------------- - -PaddlePaddle使用avx SIMD指令提高cpu执行效率，因此错误的使用二进制发行版可能会导致这种错误，请选择正确的版本。 - -4. 如何选择SGD算法的学习率 --------------------------- - -在采用sgd/async_sgd进行训练时，一个重要的问题是选择正确的learning_rate。如果learning_rate太大，那么训练有可能不收敛，如果learning_rate太小，那么收敛可能很慢，导致训练时间过长。 - -通常做法是从一个比较大的learning_rate开始试，如果不收敛，那减少学习率10倍继续试验，直到训练收敛为止。那么如何判断训练不收敛呢？可以估计出如果模型采用不变的输出最小的cost0是多少。 - -如果训练过程的的cost明显高于这个常数输出的cost，那么我们可以判断为训练不收敛。举一个例子，假如我们是三分类问题，采用multi-class-cross-entropy作为cost，数据中0,1,2三类的比例为 :code:`0.2, 0.5, 0.3` , 那么常数输出所能达到的最小cost是 :code:`-(0.2*log(0.2)+0.5*log(0.5)+0.3*log(0.3))=1.03` 。如果训练一个pass（或者更早）后，cost还大于这个数，那么可以认为训练不收敛，应该降低学习率。 - - -5. 如何初始化参数 ------------------ - -默认情况下，PaddlePaddle使用均值0，标准差为 :math:`\frac{1}{\sqrt{d}}` 来初始化参数。其中 :math:`d` 为参数矩阵的宽度。这种初始化方式在一般情况下不会产生很差的结果。如果用户想要自定义初始化方式，PaddlePaddle目前提供两种参数初始化的方式\: - -* 高斯分布。将 :code:`param_attr` 设置成 :code:`param_attr=ParamAttr(initial_mean=0.0, initial_std=1.0)` -* 均匀分布。将 :code:`param_attr` 设置成 :code:`param_attr=ParamAttr(initial_max=1.0, initial_min=-1.0)` - -比如设置一个全连接层的参数初始化方式和bias初始化方式，可以使用如下代码。 - -.. code-block:: python - - hidden = fc_layer(input=ipt, param_attr=ParamAttr(initial_max=1.0, initial_min=-1.0), - bias_attr=ParamAttr(initial_mean=1.0, initial_std=0.0)) - -上述代码将bias全部初始化为1.0, 同时将参数初始化为 :code:`[1.0, -1.0]` 的均匀分布。 - -6. 如何共享参数 ---------------- - -PaddlePaddle的参数使用名字 :code:`name` 作为参数的ID，相同名字的参数，会共享参数。设置参数的名字，可以使用 :code:`ParamAttr(name="YOUR_PARAM_NAME")` 来设置。更方便的设置方式，是使得要共享的参数使用同样的 :code:`ParamAttr` 对象。 - -简单的全连接网络，参数共享的配置示例为\: - -.. literalinclude:: ../../python/paddle/trainer_config_helpers/tests/configs/shared_fc.py - -这里 :code:`hidden_a` 和 :code:`hidden_b` 使用了同样的parameter和bias。并且softmax层的两个输入也使用了同样的参数 :code:`softmax_param`。 - -7. paddlepaddle\*.whl is not a supported wheel on this platform. ------------------------------------------------------------------------- - -出现这个问题的主要原因是，没有找到和当前系统匹配的paddlepaddle安装包。最新的paddlepaddle python安装包支持Linux x86_64和MacOS 10.12操作系统，并安装了python 2.7和pip 9.0.1。 - -更新 :code:`pip` 包的方法是\: - -.. code-block:: bash - - pip install --upgrade pip - -如果还不行，可以执行 :code:`python -c "import pip; print(pip.pep425tags.get_supported())"` 获取当前系统支持的python包的后缀， -并对比是否和正在安装的后缀一致。 - -如果系统支持的是 :code:`linux_x86_64` 而安装包是 :code:`manylinux1_x86_64` ，需要升级pip版本到最新； -如果系统支持 :code:`manylinux1_x86_64` 而安装包（本地）是 :code:`linux_x86_64` ，可以重命名这个whl包为 :code:`manylinux1_x86_64` 再安装。 - -8. python相关的单元测试都过不了 --------------------------------- - -如果出现以下python相关的单元测试都过不了的情况： - -.. code-block:: bash - - 24 - test_PyDataProvider (Failed) - 26 - test_RecurrentGradientMachine (Failed) - 27 - test_NetworkCompare (Failed) - 28 - test_PyDataProvider2 (Failed) - 32 - test_Prediction (Failed) - 33 - test_Compare (Failed) - 34 - test_Trainer (Failed) - 35 - test_TrainerOnePass (Failed) - 36 - test_CompareTwoNets (Failed) - 37 - test_CompareTwoOpts (Failed) - 38 - test_CompareSparse (Failed) - 39 - test_recurrent_machine_generation (Failed) - 40 - test_PyDataProviderWrapper (Failed) - 41 - test_config_parser (Failed) - 42 - test_swig_api (Failed) - 43 - layers_test (Failed) - -并且查询PaddlePaddle单元测试的日志，提示： - -.. code-block:: bash - - paddle package is already in your PYTHONPATH. But unittest need a clean environment. - Please uninstall paddle package before start unittest. Try to 'pip uninstall paddle'. - -解决办法是： - -* 卸载PaddlePaddle包 :code:`pip uninstall paddle`, 清理掉老旧的PaddlePaddle安装包，使得单元测试有一个干净的环境。如果PaddlePaddle包已经在python的site-packages里面，单元测试会引用site-packages里面的python包，而不是源码目录里 :code:`/python` 目录下的python包。同时，即便设置 :code:`PYTHONPATH` 到 :code:`/python` 也没用，因为python的搜索路径是优先已经安装的python包。 - - -9. 运行Docker GPU镜像出现 "CUDA driver version is insufficient" ----------------------------------------------------------------- - -用户在使用PaddlePaddle GPU的Docker镜像的时候，常常出现 `Cuda Error: CUDA driver version is insufficient for CUDA runtime version`, 原因在于没有把机器上CUDA相关的驱动和库映射到容器内部。 -具体的解决方法是： - -.. code-block:: bash - - $ export CUDA_SO="$(\ls usr/lib64/libcuda* | xargs -I{} echo '-v {}:{}') $(\ls /usr/lib64/libnvidia* | xargs -I{} echo '-v {}:{}')" - $ export DEVICES=$(\ls /dev/nvidia* | xargs -I{} echo '--device {}:{}') - $ docker run ${CUDA_SO} ${DEVICES} -it paddledev/paddlepaddle:latest-gpu - -更多关于Docker的安装与使用, 请参考 `PaddlePaddle Docker 文档 `_ 。 - - -10. CMake源码编译, 找到的PythonLibs和PythonInterp版本不一致 ----------------------------------------------------------------- - -这是目前CMake寻找Python的逻辑存在缺陷，如果系统安装了多个Python版本，CMake找到的Python库和Python解释器版本可能有不一致现象，导致编译PaddlePaddle失败。正确的解决方法是， -用户强制指定特定的Python版本，具体操作如下： - - .. code-block:: bash - - cmake .. -DPYTHON_EXECUTABLE= -DPYTHON_LIBRARY= -DPYTHON_INCLUDE_DIR= - -用户需要指定本机上Python的路径：````, ````, ```` - -11. CMake源码编译，Paddle版本号为0.0.0 --------------------------------------- - -如果运行 :code:`paddle version`, 出现 :code:`PaddlePaddle 0.0.0`；或者运行 :code:`cmake ..`，出现 - -.. code-block:: bash - - CMake Warning at cmake/version.cmake:20 (message): - Cannot add paddle version from git tag - -那么用户需要拉取所有的远程分支到本机，命令为 :code:`git fetch upstream`，然后重新cmake即可。 - -12. A protocol message was rejected because it was too big ----------------------------------------------------------- - -如果在训练NLP相关模型时，出现以下错误： - -.. code-block:: bash - - [libprotobuf ERROR google/protobuf/io/coded_stream.cc:171] A protocol message was rejected because it was too big (more than 67108864 bytes). To increase the limit (or to disable these warnings), see CodedInputStream::SetTotalBytesLimit() in google/protobuf/io/coded_stream.h. - F1205 14:59:50.295174 14703 TrainerConfigHelper.cpp:59] Check failed: m->conf.ParseFromString(configProtoStr) - -可能的原因是：传给dataprovider的某一个args过大，一般是由于直接传递大字典导致的。错误的define_py_data_sources2类似： - -.. code-block:: python - - src_dict = dict() - for line_count, line in enumerate(open(src_dict_path, "r")): - src_dict[line.strip()] = line_count - - define_py_data_sources2( - train_list, - test_list, - module="dataprovider", - obj="process", - args={"src_dict": src_dict}) - -解决方案是：将字典的地址作为args传给dataprovider，然后在dataprovider里面根据该地址加载字典。即define_py_data_sources2应改为： - -.. code-block:: python - - define_py_data_sources2( - train_list, - test_list, - module="dataprovider", - obj="process", - args={"src_dict_path": src_dict_path}) - -完整源码可参考 `seqToseq `_ 示例。 - -13. 如何指定GPU设备 -------------------- - -例如机器上有4块GPU，编号从0开始，指定使用2、3号GPU： - -* 方式1：通过 `CUDA_VISIBLE_DEVICES `_ 环境变量来指定特定的GPU。 - -.. code-block:: bash - - env CUDA_VISIBLE_DEVICES=2,3 paddle train --use_gpu=true --trainer_count=2 - -* 方式2：通过命令行参数 ``--gpu_id`` 指定。 - -.. code-block:: bash - - paddle train --use_gpu=true --trainer_count=2 --gpu_id=2 - - -14. 训练过程中出现 :code:`Floating point exception`, 训练因此退出怎么办? ------------------------------------------------------------------------- - -Paddle二进制在运行时捕获了浮点数异常，只要出现浮点数异常(即训练过程中出现NaN或者Inf)，立刻退出。浮点异常通常的原因是浮点数溢出、除零等问题。 -主要原因包括两个方面: - -* 训练过程中参数或者训练过程中的梯度尺度过大，导致参数累加，乘除等时候，导致了浮点数溢出。 -* 模型一直不收敛，发散到了一个数值特别大的地方。 -* 训练数据有问题，导致参数收敛到了一些奇异的情况。或者输入数据尺度过大，有些特征的取值达到数百万，这时进行矩阵乘法运算就可能导致浮点数溢出。 - -主要的解决办法是减小学习率或者对数据进行归一化处理。 - -15. 编译安装后执行 import paddle.v2 as paddle 报ImportError: No module named v2 ------------------------------------------------------------------------- -先查看一下是否曾经安装过paddle v1版本，有的话需要先卸载： - -pip uninstall py_paddle paddle - -然后安装paddle的python环境, 在build目录下执行 - -pip install python/dist/paddle*.whl && pip install ../paddle/dist/py_paddle*.whl - -16. PaddlePaddle存储的参数格式是什么，如何和明文进行相互转化 ---------------------------------------------------------- - -PaddlePaddle保存的模型参数文件内容由16字节头信息和网络参数两部分组成。头信息中，1~4字节表示PaddlePaddle版本信息，请直接填充0；5~8字节表示每个参数占用的字节数，当保存的网络参数为float类型时为4，double类型时为8；9~16字节表示保存的参数总个数。 - -将PaddlePaddle保存的模型参数还原回明文时，可以使用相应数据类型的 :code:`numpy.array` 加载具体网络参数，此时可以跳过PaddlePaddle模型参数文件的头信息。若在PaddlePaddle编译时，未指定按照double精度编译，默认情况下按照float精度计算，保存的参数也是float类型。这时在使用 :code:`numpy.array` 时，一般设置 :code:`dtype=float32` 。示例如下： - -.. code-block:: python - - def read_parameter(fname, width): - s = open(fname).read() - # skip header - vec = np.fromstring(s[16:], dtype=np.float32) - # width is the size of the corresponding layer - np.savetxt(fname + ".csv", vec.reshape(width, -1), - fmt="%.6f", delimiter=",") - - -将明文参数转化为PaddlePaddle可加载的模型参数时，首先构造头信息，再写入网络参数。下面的代码将随机生成的矩阵转化为可以被PaddlePaddle加载的模型参数。 - -.. code-block:: python - - def gen_rand_param(param_file, width, height, need_trans): - np.random.seed() - header = struct.pack("iil", 0, 4, height * width) - param = np.float32(np.random.rand(height, width)) - with open(param_file, "w") as fparam: - fparam.write(header + param.tostring()) - -17. 如何加载预训练参数 ------------------------------- - -* 对加载预训练参数的层，设置其参数属性 :code:`is_static=True`，使该层的参数在训练过程中保持不变。以embedding层为例，代码如下： - -.. code-block:: python - - emb_para = paddle.attr.Param(name='emb', is_static=True) - paddle.layer.embedding(size=word_dim, input=x, param_attr=emb_para) - - -* 从模型文件将预训练参数载入 :code:`numpy.array`，在创建parameters后，使用 :code:`parameters.set()` 加载预训练参数。PaddlePaddle保存的模型参数文件前16字节为头信息，用户将参数载入 :code:`numpy.array` 时须从第17字节开始。以embedding层为例，代码如下： - -.. code-block:: python - - def load_parameter(file_name, h, w): - with open(file_name, 'rb') as f: - f.read(16) # skip header. - return np.fromfile(f, dtype=np.float32).reshape(h, w) - - parameters = paddle.parameters.create(my_cost) - parameters.set('emb', load_parameter(emb_param_file, 30000, 256)) - -18. 集群多节点训练，日志中保存均为网络通信类错误 ------------------------------- - -集群多节点训练，日志报错为网络通信类错误，比如 :code:`Connection reset by peer` 等。 -此类报错通常是由于某一个节点的错误导致这个节点的训练进程退出，从而引发其他节点无法连接导致，可以参考下面的步骤排查： - -* 从 :code:`train.log` ， :code:`server.log` 找到最早报错的地方，查看是否是其他错误引发的报错（比如FPE，内存不足，磁盘空间不足等）。 - -* 如果发现最早的报错就是网络通信的问题，很有可能是非独占方式执行导致的端口冲突，可以联系OP，看当前MPI集群是否支持resource=full参数提交，如果支持增加此参数提交，并更换job 端口。 - -* 如果当前MPI集群并不支持任务独占模式，可以联系OP是否可以更换集群或升级当前集群。 - -19. PaddlePaddle如何输出多个层 ------------------------------- - -* 将需要输出的层作为 :code:`paddle.inference.Inference()` 接口的 :code:`output_layer` 参数输入，代码如下： - -.. code-block:: python - - inferer = paddle.inference.Inference(output_layer=[layer1, layer2], parameters=parameters) - -* 指定要输出的字段进行输出。以输出 :code:`value` 字段为例，代码如下： - -.. code-block:: python - - out = inferer.infer(input=data_batch, flatten_result=False, field=["value"]) - -这里设置 :code:`flatten_result=False`，得到的输出结果是元素个数等于输出字段数的 :code:`list`，该 :code:`list` 的每个元素是由所有输出层相应字段结果组成的 :code:`list`，每个字段结果的类型是 :code:`numpy.array`。:code:`flatten_result` 的默认值为 :code:`True`，该情况下，PaddlePaddle会分别对每个字段将所有输出层的结果按行进行拼接，如果各输出层该字段 :code:`numpy.array` 结果的相应维数不匹配，程序将不能正常运行。 - -20. :code:`paddle.layer.memory` 的参数 :code:`name` 如何使用 -------------------------------------------------------------- - -* :code:`paddle.layer.memory` 用于获取特定layer上一时间步的输出，该layer是通过参数 :code:`name` 指定，即，:code:`paddle.layer.memory` 会关联参数 :code:`name` 取值相同的layer，并将该layer上一时间步的输出作为自身当前时间步的输出。 - -* PaddlePaddle的所有layer都有唯一的name，用户通过参数 :code:`name` 设定，当用户没有显式设定时，PaddlePaddle会自动设定。而 :code:`paddle.layer.memory` 不是真正的layer，其name由参数 :code:`memory_name` 设定，当用户没有显式设定时，PaddlePaddle会自动设定。:code:`paddle.layer.memory` 的参数 :code:`name` 用于指定其要关联的layer，需要用户显式设定。 - -21. dropout 使用 ------------------ - -* 在PaddlePaddle中使用dropout有两种方式 - - * 在相应layer的 :code:`layer_atter` 设置 :code:`drop_rate`，以 :code:`paddle.layer.fc` 为例，代码如下： - - .. code-block:: python - - fc = paddle.layer.fc(input=input, layer_attr=paddle.attr.ExtraLayerAttribute(drop_rate=0.5)) - - * 使用 :code:`paddle.layer.dropout`，以 :code:`paddle.layer.fc` 为例，代码如下： - - .. code-block:: python - - fc = paddle.layer.fc(input=input) - drop_fc = paddle.layer.dropout(input=fc, dropout_rate=0.5) - -* :code:`paddle.layer.dropout` 实际上使用了 :code:`paddle.layer.add_to`，并在该layer里采用第一种方式设置 :code:`drop_rate` 来使用dropout的。这种方式对内存消耗较大。 - -* PaddlePaddle在激活函数里实现dropout，而不是在layer里实现。 - -* :code:`paddle.layer.lstmemory`、:code:`paddle.layer.grumemory`、:code:`paddle.layer.recurrent` 不是通过一般的方式来实现对输出的激活，所以不能采用第一种方式在这几个layer里设置 :code:`drop_rate` 来使用dropout。若要对这几个layer使用dropout，可采用第二种方式，即使用 :code:`paddle.layer.dropout`。 - -22. 如何设置学习率退火（learning rate annealing） ------------------------------------------------- - -在相应的优化算法里设置learning_rate_schedule及相关参数，以使用Adam算法为例，代码如下： - -.. code-block:: python - - optimizer = paddle.optimizer.Adam( - learning_rate=1e-3, - learning_rate_decay_a=0.5, - learning_rate_decay_b=0.75, - learning_rate_schedule="poly",) - -PaddlePaddle目前支持8种learning_rate_schedule，这8种learning_rate_schedule及其对应学习率计算方式如下： - -* "constant" - - lr = learning_rate - -* "poly" - - lr = learning_rate * pow(1 + learning_rate_decay_a * num_samples_processed, -learning_rate_decay_b) - - 其中，num_samples_processed为已训练样本数，下同。 - -* "caffe_poly" - - lr = learning_rate * pow(1.0 - num_samples_processed / learning_rate_decay_a, learning_rate_decay_b) - -* "exp" - - lr = learning_rate * pow(learning_rate_decay_a, num_samples_processed / learning_rate_decay_b) - -* "discexp" - - lr = learning_rate * pow(learning_rate_decay_a, floor(num_samples_processed / learning_rate_decay_b)) - -* "linear" - - lr = max(learning_rate - learning_rate_decay_a * num_samples_processed, learning_rate_decay_b) - -* "manual" - - 这是一种按已训练样本数分段取值的学习率退火方法。使用该learning_rate_schedule时，用户通过参数 :code:`learning_rate_args` 设置学习率衰减因子分段函数，当前的学习率为所设置 :code:`learning_rate` 与当前的衰减因子的乘积。以使用Adam算法为例，代码如下： - - .. code-block:: python - - optimizer = paddle.optimizer.Adam( - learning_rate=1e-3, - learning_rate_schedule="manual", - learning_rate_args="1000:1.0,2000:0.9,3000:0.8",) - - 在该示例中，当已训练样本数小于等于1000时，学习率为 :code:`1e-3 * 1.0`；当已训练样本数大于1000小于等于2000时，学习率为 :code:`1e-3 * 0.9`；当已训练样本数大于2000时，学习率为 :code:`1e-3 * 0.8`。 - -* "pass_manual" - - 这是一种按已训练pass数分段取值的学习率退火方法。使用该learning_rate_schedule时，用户通过参数 :code:`learning_rate_args` 设置学习率衰减因子分段函数，当前的学习率为所设置 :code:`learning_rate` 与当前的衰减因子的乘积。以使用Adam算法为例，代码如下： - - .. code-block:: python - - optimizer = paddle.optimizer.Adam( - learning_rate=1e-3, - learning_rate_schedule="manual", - learning_rate_args="1:1.0,2:0.9,3:0.8",) - - 在该示例中，当已训练pass数小于等于1时，学习率为 :code:`1e-3 * 1.0`；当已训练pass数大于1小于等于2时，学习率为 :code:`1e-3 * 0.9`；当已训练pass数大于2时，学习率为 :code:`1e-3 * 0.8`。 - -23. 出现 :code:`Duplicated layer name` 错误怎么办 --------------------------------------------------- - -出现该错误的原因一般是用户对不同layer的参数 :code:`name` 设置了相同的取值。遇到该错误时，先找出参数 :code:`name` 取值相同的layer，然后将这些layer的参数 :code:`name` 设置为不同的值。 +.. toctree:: + :maxdepth: 1 + build_and_install/index_cn.rst + model/index_cn.rst + parameter/index_cn.rst + local/index_cn.rst + cluster/index_cn.rst diff --git a/doc/faq/local/index_cn.rst b/doc/faq/local/index_cn.rst new file mode 100644 index 0000000000000000000000000000000000000000..75c4ba028e497e29e9030a86514348726d9c0a80 --- /dev/null +++ b/doc/faq/local/index_cn.rst @@ -0,0 +1,213 @@ +############### +本地训练与预测 +############### + +.. contents:: + +1. 如何减少内存占用 +------------------- + +神经网络的训练本身是一个非常消耗内存和显存的工作，经常会消耗数10GB的内存和数GB的显存。 +PaddlePaddle的内存占用主要分为如下几个方面\: + +* DataProvider缓冲池内存（只针对内存） +* 神经元激活内存（针对内存和显存） +* 参数内存 （针对内存和显存） +* 其他内存杂项 + +其中，其他内存杂项是指PaddlePaddle本身所用的一些内存，包括字符串分配，临时变量等等，暂不考虑在内。 + +减少DataProvider缓冲池内存 +++++++++++++++++++++++++++ + +PyDataProvider使用的是异步加载，同时在内存里直接随即选取数据来做Shuffle。即 + +.. graphviz:: + + digraph { + rankdir=LR; + 数据文件 -> 内存池 -> PaddlePaddle训练 + } + +所以，减小这个内存池即可减小内存占用，同时也可以加速开始训练前数据载入的过程。但是，这 +个内存池实际上决定了shuffle的粒度。所以，如果将这个内存池减小，又要保证数据是随机的， +那么最好将数据文件在每次读取之前做一次shuffle。可能的代码为 + +.. literalinclude:: src/reduce_min_pool_size.py + +这样做可以极大的减少内存占用，并且可能会加速训练过程，详细文档参考 :ref:`api_pydataprovider2` 。 + +神经元激活内存 +++++++++++++++ + +神经网络在训练的时候，会对每一个激活暂存一些数据，如神经元激活值等。 +在反向传递的时候，这些数据会被用来更新参数。这些数据使用的内存主要和两个参数有关系， +一是batch size，另一个是每条序列(Sequence)长度。所以，其实也是和每个mini-batch中包含 +的时间步信息成正比。 + +所以做法可以有两种： + +* 减小batch size。 即在网络配置中 :code:`settings(batch_size=1000)` 设置成一个小一些的值。但是batch size本身是神经网络的超参数，减小batch size可能会对训练结果产生影响。 +* 减小序列的长度，或者直接扔掉非常长的序列。比如，一个数据集大部分序列长度是100-200, + 但是突然有一个10000长的序列，就很容易导致内存超限，特别是在LSTM等RNN中。 + +参数内存 +++++++++ + +PaddlePaddle支持非常多的优化算法(Optimizer)，不同的优化算法需要使用不同大小的内存。 +例如使用 :code:`adadelta` 算法，则需要使用等于权重参数规模大约5倍的内存。举例，如果参数保存下来的模型目录 +文件为 :code:`100M`， 那么该优化算法至少需要 :code:`500M` 的内存。 + +可以考虑使用一些优化算法，例如 :code:`momentum`。 + +2. 如何加速训练速度 +------------------- + +加速PaddlePaddle训练可以考虑从以下几个方面\： + +* 减少数据载入的耗时 +* 加速训练速度 +* 利用分布式训练驾驭更多的计算资源 + +减少数据载入的耗时 +++++++++++++++++++ + +使用\ :code:`pydataprovider`\ 时，可以减少缓存池的大小，同时设置内存缓存功能，即可以极大的加速数据载入流程。 +:code:`DataProvider` 缓存池的减小，和之前减小通过减小缓存池来减小内存占用的原理一致。 + +.. literalinclude:: src/reduce_min_pool_size.py + +同时 :code:`@provider` 接口有一个 :code:`cache` 参数来控制缓存方法，将其设置成 :code:`CacheType.CACHE_PASS_IN_MEM` 的话，会将第一个 :code:`pass` (过完所有训练数据即为一个pass)生成的数据缓存在内存里，在之后的 :code:`pass` 中，不会再从 :code:`python` 端读取数据，而是直接从内存的缓存里读取数据。这也会极大减少数据读入的耗时。 + + +加速训练速度 +++++++++++++ + +PaddlePaddle支持Sparse的训练，sparse训练需要训练特征是 :code:`sparse_binary_vector` 、 :code:`sparse_vector` 、或者 :code:`integer_value` 的任一一种。同时，与这个训练数据交互的Layer，需要将其Parameter设置成 sparse 更新模式，即设置 :code:`sparse_update=True` + +这里使用简单的 :code:`word2vec` 训练语言模型距离，具体使用方法为\: + +使用一个词前两个词和后两个词，来预测这个中间的词。这个任务的DataProvider为\: + +.. literalinclude:: src/word2vec_dataprovider.py + +这个任务的配置为\: + +.. literalinclude:: src/word2vec_config.py + + +利用更多的计算资源 +++++++++++++++++++ + +利用更多的计算资源可以分为一下几个方式来进行\: + +* 单机CPU训练 + + * 使用多线程训练。设置命令行参数 :code:`trainer_count`。 + +* 单机GPU训练 + + * 使用显卡训练。设置命令行参数 :code:`use_gpu`。 + * 使用多块显卡训练。设置命令行参数 :code:`use_gpu` 和 :code:`trainer_count` 。 + +* 多机训练 + + * 请参考 :ref:`cluster_train` 。 + +3. 如何指定GPU设备 +------------------ + +例如机器上有4块GPU，编号从0开始，指定使用2、3号GPU： + +* 方式1：通过 `CUDA_VISIBLE_DEVICES `_ 环境变量来指定特定的GPU。 + +.. code-block:: bash + + env CUDA_VISIBLE_DEVICES=2,3 paddle train --use_gpu=true --trainer_count=2 + +* 方式2：通过命令行参数 ``--gpu_id`` 指定。 + +.. code-block:: bash + + paddle train --use_gpu=true --trainer_count=2 --gpu_id=2 + + +4. 训练过程中出现 :code:`Floating point exception`, 训练因此退出怎么办? +------------------------------------------------------------------------ + +Paddle二进制在运行时捕获了浮点数异常，只要出现浮点数异常(即训练过程中出现NaN或者Inf)，立刻退出。浮点异常通常的原因是浮点数溢出、除零等问题。 +主要原因包括两个方面: + +* 训练过程中参数或者训练过程中的梯度尺度过大，导致参数累加，乘除等时候，导致了浮点数溢出。 +* 模型一直不收敛，发散到了一个数值特别大的地方。 +* 训练数据有问题，导致参数收敛到了一些奇异的情况。或者输入数据尺度过大，有些特征的取值达到数百万，这时进行矩阵乘法运算就可能导致浮点数溢出。 + +这里有两种有效的解决方法： + +1. 设置 :code:`gradient_clipping_threshold` 参数，示例代码如下： + +.. code-block:: python + +optimizer = paddle.optimizer.RMSProp( + learning_rate=1e-3, + gradient_clipping_threshold=10.0, + regularization=paddle.optimizer.L2Regularization(rate=8e-4)) + +具体可以参考 `nmt_without_attention `_ 示例。 + +2. 设置 :code:`error_clipping_threshold` 参数，示例代码如下： + +.. code-block:: python + +decoder_inputs = paddle.layer.fc( + act=paddle.activation.Linear(), + size=decoder_size * 3, + bias_attr=False, + input=[context, current_word], + layer_attr=paddle.attr.ExtraLayerAttribute( + error_clipping_threshold=100.0)) + +完整代码可以参考示例 `machine translation `_ 。 + +两种方法的区别： + +1. 两者都是对梯度的截断，但截断时机不同，前者在 :code:`optimzier` 更新网络参数时应用；后者在激活函数反向计算时被调用； +2. 截断对象不同：前者截断可学习参数的梯度，后者截断回传给前层的梯度; + +除此之外，还可以通过减小学习律或者对数据进行归一化处理来解决这类问题。 + +5. 如何调用 infer 接口输出多个layer的预测结果 +----------------------------------------------- + +* 将需要输出的层作为 :code:`paddle.inference.Inference()` 接口的 :code:`output_layer` 参数输入，代码如下： + +.. code-block:: python + + inferer = paddle.inference.Inference(output_layer=[layer1, layer2], parameters=parameters) + +* 指定要输出的字段进行输出。以输出 :code:`value` 字段为例，代码如下： + +.. code-block:: python + + out = inferer.infer(input=data_batch, field=["value"]) + +需要注意的是： + +* 如果指定了2个layer作为输出层，实际上需要的输出结果是两个矩阵； +* 假设第一个layer的输出A是一个 N1 * M1 的矩阵，第二个 Layer 的输出B是一个 N2 * M2 的矩阵； +* paddle.v2 默认会将A和B 横向拼接，当N1 和 N2 大小不一样时，会报如下的错误： + +.. code-block:: python + + ValueError: all the input array dimensions except for the concatenation axis must match exactly + +多个层的输出矩阵的高度不一致导致拼接失败，这种情况常常发生在： + +* 同时输出序列层和非序列层； +* 多个输出层处理多个不同长度的序列; + +此时可以在调用infer接口时通过设置 :code:`flatten_result=False` , 跳过“拼接”步骤，来解决上面的问题。这时，infer接口的返回值是一个python list: + +* list 中元素的个数等于网络中输出层的个数； +* list 中每个元素是一个layer的输出结果矩阵，类型是numpy的ndarray； +* 每一个layer输出矩阵的高度，在非序列输入时：等于样本数；序列输入时等于：输入序列中元素的总数；宽度等于配置中layer的size； diff --git a/doc/faq/src/reduce_min_pool_size.py b/doc/faq/local/src/reduce_min_pool_size.py similarity index 100% rename from doc/faq/src/reduce_min_pool_size.py rename to doc/faq/local/src/reduce_min_pool_size.py diff --git a/doc/faq/src/word2vec_config.py b/doc/faq/local/src/word2vec_config.py similarity index 100% rename from doc/faq/src/word2vec_config.py rename to doc/faq/local/src/word2vec_config.py diff --git a/doc/faq/src/word2vec_dataprovider.py b/doc/faq/local/src/word2vec_dataprovider.py similarity index 100% rename from doc/faq/src/word2vec_dataprovider.py rename to doc/faq/local/src/word2vec_dataprovider.py diff --git a/doc/faq/model/index_cn.rst b/doc/faq/model/index_cn.rst new file mode 100644 index 0000000000000000000000000000000000000000..b47bbe05bdb39d1ade9434a7e54bf6ca88a91cc9 --- /dev/null +++ b/doc/faq/model/index_cn.rst @@ -0,0 +1,69 @@ +######### +模型配置 +######### + +.. contents:: + +1. 出现 :code:`Duplicated layer name` 错误怎么办 +-------------------------------------------------- + +出现该错误的原因一般是用户对不同layer的参数 :code:`name` 设置了相同的取值。遇到该错误时，先找出参数 :code:`name` 取值相同的layer，然后将这些layer的参数 :code:`name` 设置为不同的值。 + +2. :code:`paddle.layer.memory` 的参数 :code:`name` 如何使用 +------------------------------------------------------------- + +* :code:`paddle.layer.memory` 用于获取特定layer上一时间步的输出，该layer是通过参数 :code:`name` 指定，即，:code:`paddle.layer.memory` 会关联参数 :code:`name` 取值相同的layer，并将该layer上一时间步的输出作为自身当前时间步的输出。 + +* PaddlePaddle的所有layer都有唯一的name，用户通过参数 :code:`name` 设定，当用户没有显式设定时，PaddlePaddle会自动设定。而 :code:`paddle.layer.memory` 不是真正的layer，其name由参数 :code:`memory_name` 设定，当用户没有显式设定时，PaddlePaddle会自动设定。:code:`paddle.layer.memory` 的参数 :code:`name` 用于指定其要关联的layer，需要用户显式设定。 + +3. 两种使用 drop_out 的方法有何区别 +------------------------------------ + +* 在PaddlePaddle中使用dropout有两种方式 + + * 在相应layer的 :code:`layer_atter` 设置 :code:`drop_rate`，以 :code:`paddle.layer.fc` 为例，代码如下： + + .. code-block:: python + + fc = paddle.layer.fc(input=input, layer_attr=paddle.attr.ExtraLayerAttribute(drop_rate=0.5)) + + * 使用 :code:`paddle.layer.dropout`，以 :code:`paddle.layer.fc` 为例，代码如下： + + .. code-block:: python + + fc = paddle.layer.fc(input=input) + drop_fc = paddle.layer.dropout(input=fc, dropout_rate=0.5) + +* :code:`paddle.layer.dropout` 实际上使用了 :code:`paddle.layer.add_to`，并在该layer里采用第一种方式设置 :code:`drop_rate` 来使用dropout的。这种方式对内存消耗较大。 + +* PaddlePaddle在激活函数里实现dropout，而不是在layer里实现。 + +* :code:`paddle.layer.lstmemory`、:code:`paddle.layer.grumemory`、:code:`paddle.layer.recurrent` 不是通过一般的方式来实现对输出的激活，所以不能采用第一种方式在这几个layer里设置 :code:`drop_rate` 来使用dropout。若要对这几个layer使用dropout，可采用第二种方式，即使用 :code:`paddle.layer.dropout`。 + +4. 不同的 recurrent layer 的区别 +---------------------------------- +以LSTM为例，在PaddlePaddle中包含以下 recurrent layer： + +* :code:`paddle.layer.lstmemory` +* :code:`paddle.networks.simple_lstm` +* :code:`paddle.networks.lstmemory_group` +* :code:`paddle.networks.bidirectional_lstm` + +按照具体实现方式可以归纳为2类： + +1. 由 recurrent_group 实现的 recurrent layer： + + * 用户在使用这一类recurrent layer时，可以访问由recurrent unit在一个时间步内计算得到的中间值（例如：hidden states, memory cells等）； + * 上述的 :code:`paddle.networks.lstmemory_group` 是这一类的 recurrent layer ； + +2. 将recurrent layer作为一个整体来实现： + + * 用户在使用这一类recurrent layer，只能访问它们的输出值； + * 上述的 :code:`paddle.networks.lstmemory_group` 、 :code:`paddle.networks.simple_lstm` 和 :code:`paddle.networks.bidirectional_lstm` 属于这一类的实现； + +将recurrent layer作为一个整体来实现， 能够针对CPU和GPU的计算做更多优化， 所以相比于recurrent group的实现方式， 第二类 recurrent layer 计算效率更高。 在实际应用中，如果用户不需要访问LSTM的中间变量，而只需要获得recurrent layer计算的输出，我们建议使用第二类实现。 + +此外，关于LSTM, PaddlePaddle中还包含 :code:`paddle.networks.lstmemory_unit` 这一计算单元： + + * 不同于上述介绍的recurrent layer , :code:`paddle.networks.lstmemory_unit` 定义了LSTM单元在一个时间步内的计算过程，它并不是一个完整的recurrent layer，也不能接收序列数据作为输入； + * :code:`paddle.networks.lstmemory_unit` 只能在recurrent_group中作为step function使用； diff --git a/doc/faq/parameter/index_cn.rst b/doc/faq/parameter/index_cn.rst new file mode 100644 index 0000000000000000000000000000000000000000..c721b623183cc7d8d17e2c9fb1635ea07b8970cc --- /dev/null +++ b/doc/faq/parameter/index_cn.rst @@ -0,0 +1,201 @@ +######### +参数设置 +######### + +.. contents:: + +1. 如何选择SGD算法的学习率 +-------------------------- + +在采用sgd/async_sgd进行训练时，一个重要的问题是选择正确的learning_rate。如果learning_rate太大，那么训练有可能不收敛，如果learning_rate太小，那么收敛可能很慢，导致训练时间过长。 + +通常做法是从一个比较大的learning_rate开始试，如果不收敛，那减少学习率10倍继续试验，直到训练收敛为止。那么如何判断训练不收敛呢？可以估计出如果模型采用不变的输出最小的cost0是多少。 + +如果训练过程的的cost明显高于这个常数输出的cost，那么我们可以判断为训练不收敛。举一个例子，假如我们是三分类问题，采用multi-class-cross-entropy作为cost，数据中0,1,2三类的比例为 :code:`0.2, 0.5, 0.3` , 那么常数输出所能达到的最小cost是 :code:`-(0.2*log(0.2)+0.5*log(0.5)+0.3*log(0.3))=1.03` 。如果训练一个pass（或者更早）后，cost还大于这个数，那么可以认为训练不收敛，应该降低学习率。 + +2. 如何设置学习率退火（learning rate annealing） +------------------------------------------------ + +在相应的优化算法里设置learning_rate_schedule及相关参数，以使用Adam算法为例，代码如下： + +.. code-block:: python + + optimizer = paddle.optimizer.Adam( + learning_rate=1e-3, + learning_rate_decay_a=0.5, + learning_rate_decay_b=0.75, + learning_rate_schedule="poly",) + +PaddlePaddle目前支持8种learning_rate_schedule，这8种learning_rate_schedule及其对应学习率计算方式如下： + +* "constant" + + lr = learning_rate + +* "poly" + + lr = learning_rate * pow(1 + learning_rate_decay_a * num_samples_processed, -learning_rate_decay_b) + + 其中，num_samples_processed为已训练样本数，下同。 + +* "caffe_poly" + + lr = learning_rate * pow(1.0 - num_samples_processed / learning_rate_decay_a, learning_rate_decay_b) + +* "exp" + + lr = learning_rate * pow(learning_rate_decay_a, num_samples_processed / learning_rate_decay_b) + +* "discexp" + + lr = learning_rate * pow(learning_rate_decay_a, floor(num_samples_processed / learning_rate_decay_b)) + +* "linear" + + lr = max(learning_rate - learning_rate_decay_a * num_samples_processed, learning_rate_decay_b) + +* "manual" + + 这是一种按已训练样本数分段取值的学习率退火方法。使用该learning_rate_schedule时，用户通过参数 :code:`learning_rate_args` 设置学习率衰减因子分段函数，当前的学习率为所设置 :code:`learning_rate` 与当前的衰减因子的乘积。以使用Adam算法为例，代码如下： + + .. code-block:: python + + optimizer = paddle.optimizer.Adam( + learning_rate=1e-3, + learning_rate_schedule="manual", + learning_rate_args="1000:1.0,2000:0.9,3000:0.8",) + + 在该示例中，当已训练样本数小于等于1000时，学习率为 :code:`1e-3 * 1.0`；当已训练样本数大于1000小于等于2000时，学习率为 :code:`1e-3 * 0.9`；当已训练样本数大于2000时，学习率为 :code:`1e-3 * 0.8`。 + +* "pass_manual" + + 这是一种按已训练pass数分段取值的学习率退火方法。使用该learning_rate_schedule时，用户通过参数 :code:`learning_rate_args` 设置学习率衰减因子分段函数，当前的学习率为所设置 :code:`learning_rate` 与当前的衰减因子的乘积。以使用Adam算法为例，代码如下： + + .. code-block:: python + + optimizer = paddle.optimizer.Adam( + learning_rate=1e-3, + learning_rate_schedule="manual", + learning_rate_args="1:1.0,2:0.9,3:0.8",) + + 在该示例中，当已训练pass数小于等于1时，学习率为 :code:`1e-3 * 1.0`；当已训练pass数大于1小于等于2时，学习率为 :code:`1e-3 * 0.9`；当已训练pass数大于2时，学习率为 :code:`1e-3 * 0.8`。 + +3. 如何初始化参数 +----------------- + +默认情况下，PaddlePaddle使用均值0，标准差为 :math:`\frac{1}{\sqrt{d}}` 来初始化参数。其中 :math:`d` 为参数矩阵的宽度。这种初始化方式在一般情况下不会产生很差的结果。如果用户想要自定义初始化方式，PaddlePaddle目前提供两种参数初始化的方式\: + +* 高斯分布。将 :code:`param_attr` 设置成 :code:`param_attr=ParamAttr(initial_mean=0.0, initial_std=1.0)` +* 均匀分布。将 :code:`param_attr` 设置成 :code:`param_attr=ParamAttr(initial_max=1.0, initial_min=-1.0)` + +比如设置一个全连接层的参数初始化方式和bias初始化方式，可以使用如下代码。 + +.. code-block:: python + + hidden = fc_layer(input=ipt, param_attr=ParamAttr(initial_max=1.0, initial_min=-1.0), + bias_attr=ParamAttr(initial_mean=1.0, initial_std=0.0)) + +上述代码将bias全部初始化为1.0, 同时将参数初始化为 :code:`[1.0, -1.0]` 的均匀分布。 + +4. 如何共享参数 +--------------- + +PaddlePaddle的参数使用名字 :code:`name` 作为参数的ID，相同名字的参数，会共享参数。设置参数的名字，可以使用 :code:`ParamAttr(name="YOUR_PARAM_NAME")` 来设置。更方便的设置方式，是使得要共享的参数使用同样的 :code:`ParamAttr` 对象。 + +简单的全连接网络，参数共享的配置示例为\: + +.. literalinclude:: ../../python/paddle/trainer_config_helpers/tests/configs/shared_fc.py + +这里 :code:`hidden_a` 和 :code:`hidden_b` 使用了同样的parameter和bias。并且softmax层的两个输入也使用了同样的参数 :code:`softmax_param`。 + +5. 如何加载预训练参数 +------------------------ + +* 对加载预训练参数的层，设置其参数属性 :code:`is_static=True`，使该层的参数在训练过程中保持不变。以embedding层为例，代码如下： + +.. code-block:: python + + emb_para = paddle.attr.Param(name='emb', is_static=True) + paddle.layer.embedding(size=word_dim, input=x, param_attr=emb_para) + + +* 从模型文件将预训练参数载入 :code:`numpy.array`，在创建parameters后，使用 :code:`parameters.set()` 加载预训练参数。PaddlePaddle保存的模型参数文件前16字节为头信息，用户将参数载入 :code:`numpy.array` 时须从第17字节开始。以embedding层为例，代码如下： + +.. code-block:: python + + def load_parameter(file_name, h, w): + with open(file_name, 'rb') as f: + f.read(16) # skip header. + return np.fromfile(f, dtype=np.float32).reshape(h, w) + + parameters = paddle.parameters.create(my_cost) + parameters.set('emb', load_parameter(emb_param_file, 30000, 256)) + +6. 存储的参数格式是什么，如何和明文进行相互转化 +-------------------------------------------------- + +PaddlePaddle保存的模型参数文件内容由16字节头信息和网络参数两部分组成。头信息中，1~4字节表示PaddlePaddle版本信息，请直接填充0；5~8字节表示每个参数占用的字节数，当保存的网络参数为float类型时为4，double类型时为8；9~16字节表示保存的参数总个数。 + +将PaddlePaddle保存的模型参数还原回明文时，可以使用相应数据类型的 :code:`numpy.array` 加载具体网络参数，此时可以跳过PaddlePaddle模型参数文件的头信息。若在PaddlePaddle编译时，未指定按照double精度编译，默认情况下按照float精度计算，保存的参数也是float类型。这时在使用 :code:`numpy.array` 时，一般设置 :code:`dtype=float32` 。示例如下： + +.. code-block:: python + + def read_parameter(fname, width): + s = open(fname).read() + # skip header + vec = np.fromstring(s[16:], dtype=np.float32) + # width is the size of the corresponding layer + np.savetxt(fname + ".csv", vec.reshape(width, -1), + fmt="%.6f", delimiter=",") + + +将明文参数转化为PaddlePaddle可加载的模型参数时，首先构造头信息，再写入网络参数。下面的代码将随机生成的矩阵转化为可以被PaddlePaddle加载的模型参数。 + +.. code-block:: python + + def gen_rand_param(param_file, width, height, need_trans): + np.random.seed() + header = struct.pack("iil", 0, 4, height * width) + param = np.float32(np.random.rand(height, width)) + with open(param_file, "w") as fparam: + fparam.write(header + param.tostring()) + +7. A protocol message was rejected because it was too big +------------------------------------------------------------ + +如果在训练NLP相关模型时，出现以下错误： + +.. code-block:: bash + + [libprotobuf ERROR google/protobuf/io/coded_stream.cc:171] A protocol message was rejected because it was too big (more than 67108864 bytes). To increase the limit (or to disable these warnings), see CodedInputStream::SetTotalBytesLimit() in google/protobuf/io/coded_stream.h. + F1205 14:59:50.295174 14703 TrainerConfigHelper.cpp:59] Check failed: m->conf.ParseFromString(configProtoStr) + +可能的原因是：传给dataprovider的某一个args过大，一般是由于直接传递大字典导致的。错误的define_py_data_sources2类似： + +.. code-block:: python + + src_dict = dict() + for line_count, line in enumerate(open(src_dict_path, "r")): + src_dict[line.strip()] = line_count + + define_py_data_sources2( + train_list, + test_list, + module="dataprovider", + obj="process", + args={"src_dict": src_dict}) + +解决方案是：将字典的地址作为args传给dataprovider，然后在dataprovider里面根据该地址加载字典。即define_py_data_sources2应改为： + +.. code-block:: python + + define_py_data_sources2( + train_list, + test_list, + module="dataprovider", + obj="process", + args={"src_dict_path": src_dict_path}) + +完整源码可参考 `seqToseq `_ 示例。 + + diff --git a/doc/getstarted/build_and_install/docker_install_cn.rst b/doc/getstarted/build_and_install/docker_install_cn.rst index 84e33177740ca1652efc09c8081c2519b4366906..30b144d849bec367cd0197b6082889e011193a9a 100644 --- a/doc/getstarted/build_and_install/docker_install_cn.rst +++ b/doc/getstarted/build_and_install/docker_install_cn.rst @@ -20,7 +20,7 @@ Docker使用入门 docker pull paddlepaddle/paddle:0.10.0 - 来下载Docker镜像，paddlepaddle/paddle是从官方镜像源Dockerhub.com下载的，推荐国内用户使用ocker.paddlepaddle.org/paddle下载。 + 来下载Docker镜像，paddlepaddle/paddle是从官方镜像源Dockerhub.com下载的，推荐国内用户使用docker.paddlepaddle.org/paddle下载。 - *容器*： 如果说一个Docker镜像就是一个程序，那容器就是这个程序运行时产生的“进程”。 实际上，一个容器就是一个操作系统的进程，但是是运行在独立的进程空间，文件系统以及网络之上。 diff --git a/doc/howto/dev/new_op_cn.md b/doc/howto/dev/new_op_cn.md index 264b998f50df016da0741d97d4b26f759ee90900..c823d7e9fcd63dd7719ac1403952b03c2d2f03c0 100644 --- a/doc/howto/dev/new_op_cn.md +++ b/doc/howto/dev/new_op_cn.md @@ -206,7 +206,7 @@ MulOp(const std::string &type, const framework::VariableNameMap &inputs, - `REGISTER_OP` ： 注册`ops::MulOp`类，类型名为`mul`，该类的`ProtoMaker`为`ops::MulOpMaker`，注册`ops::MulOpGrad`，类型名为`mul_grad`。 - `REGISTER_OP_WITHOUT_GRADIENT` ： 用于注册没有反向的Op。 - - `REGISTER_OP_CPU_KERNEL` ：注册`ops::MulKernel`类，并特化模板参数为`paddle::platform::CPUPlace`和`float`类型，同理，注册`ops::MulKernel`类。 + - `REGISTER_OP_CPU_KERNEL` ：注册`ops::MulKernel`类，并特化模板参数为`paddle::platform::CPUPlace`和`float`类型，同理，注册`ops::MulGradKernel`类。 - 在 `.cu`文件中注册GPU Kernel。 @@ -285,41 +285,27 @@ class TestMulGradOp(GradientChecker): 'Y': np.random.random((84, 100)).astype("float32") } - def test_cpu_gpu_compare(self): - self.compare_grad(self.op, self.inputs) - - def test_normal(self): + def test_check_grad_normal(self): # mul op will enlarge the relative error - self.check_grad( - self.op, self.inputs, ["X", "Y"], "Out", max_relative_error=0.5) + self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.5) - def test_ignore_x(self): + def test_check_grad_ingore_x(self): self.check_grad( - self.op, - self.inputs, ["Y"], - "Out", - max_relative_error=0.5, - no_grad_set={"X"}) + ['Y'], 'Out', max_relative_error=0.5, no_grad_set=set("X")) - def test_ignore_y(self): + def test_check_grad_ingore_y(self): self.check_grad( - self.op, - self.inputs, ["X"], - "Out", - max_relative_error=0.5, - no_grad_set={"Y"}) + ['X'], 'Out', max_relative_error=0.5, no_grad_set=set('Y')) ``` 下面解释代码中一些关键的地方: - 调用`create_op("mul")`创建反向Op对应的前向Op。 -- 调用`compare_grad`函数对比CPU、GPU计算结果。 -- `test_normal`中调用`check_grad`使用数值法检测梯度正确性和稳定性。 - - 第一个参数`self.op` : 前向Op。 - - 第二个参数`self.inputs` : 输入词典，词典的Key和`ProtoMaker`定义保持一致。 - - 第三个参数`["X", "Y"]` : 指定对输入变量`X`、`Y`做梯度检测。 - - 第四个参数`"Out"` : 指定前向网络最终的输出目标变量`Out` -- `test_ignore_x`和`test_ignore_y`分支用来测试只需要计算一个输入梯度的情况。 +- `test_check_grad_normal`中调用`check_grad`使用数值法检测梯度正确性和稳定性。 + - 第一个参数`["X", "Y"]` : 指定对输入变量`X`、`Y`做梯度检测。 + - 第二个参数`"Out"` : 指定前向网络最终的输出目标变量`Out`。 + - 第三个参数`max_relative_error`：指定检测梯度时能容忍的最大错误值。 +- `test_check_grad_ingore_x`和`test_check_grad_ingore_y`分支用来测试只需要计算一个输入梯度的情况。 ### 编译和执行单元测试 diff --git a/doc/howto/dev/new_op_en.md b/doc/howto/dev/new_op_en.md index b7aa501db9e5c7378398fad48503f82bff893b60..1e88e1f5b4df710f1b69f0305d8d8a2921c4249a 100644 --- a/doc/howto/dev/new_op_en.md +++ b/doc/howto/dev/new_op_en.md @@ -1,14 +1,17 @@ # How to write a new operator - - [Background](#Background) - - [Implementing C++ Types](#Implementing_C++_Types) - - [Defining ProtoMaker](#Defining_ProtoMaker) - - [Defining Operator](#Defining_Operator) - - [Registering Operator](#Registering_Operator) - - [Compilation](#Compilation) - - [Python Binding](#Python_Binding) - - [Unit Tests](#Unit_Tests) - + - [Background](#background) + - [Implementing C++ Types](#implementing-c++-types) + - [Defining ProtoMaker](#defining-protoMaker) + - [Defining Operator](#defining-operator) + - [Registering Operator](#registering-operator) + - [Compilation](#compilation) + - [Python Binding](#python-binding) + - [Unit Tests](#unit-tests) + - [Testing Forward Operators](#testing-forward-operators) + - [Testing Backward Operators](#testing-backward-operators) + - [Compiling and Running](#compiling-and-running) + - [Remarks](#remarks) ## Background Here are the base types needed. For details, please refer to the design docs. @@ -179,7 +182,7 @@ Note that **different devices (CPU, GPU)share an Op definition; whether or not t `MulOp`'s CPU and GPU share the same `Kernel`. A non-sharing `OpKernel` example can be seen in [`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43). -To ease the writing of `OpKernel` compute, and for reusing code cross-device, `Eigen unsupported Tensor` module is used to implement `Compute` interface. To learn about how the Eigen library is used in PaddlePaddle, please see [usage document](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/dev/use_eigen_cn.md). +To ease the writing of `OpKernel` compute, and for reusing code cross-device, [`Eigen-unsupported Tensor`](https://bitbucket.org/eigen/eigen/src/default/unsupported/Eigen/CXX11/src/Tensor/README.md?fileviewer=file-view-default) module is used to implement `Compute` interface. To learn about how the Eigen library is used in PaddlePaddle, please see [usage document](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/dev/use_eigen_cn.md). This concludes the forward implementation of an operator. Next its operation and kernel need to be registered in a `.cc` file. @@ -202,7 +205,7 @@ The definition of its corresponding backward operator, if applicable, is similar - `REGISTER_OP` registers the `ops::MulOp` class, type named `mul`, its type `ProtoMaker` is `ops::MulOpMaker`, registering `ops::MulOpGrad` as `mul_grad`. - `REGISTER_OP_WITHOUT_GRADIENT` registers an operator without gradient. - - `REGISTER_OP_CPU_KERNEL` registers `ops::MulKernel` class and specialized template types `paddle::platform::CPUPlace` and `float`, which also registers `ops::MulKernel`. + - `REGISTER_OP_CPU_KERNEL` registers `ops::MulKernel` class and specialized template types `paddle::platform::CPUPlace` and `float`, which also registers `ops::MulGradKernel`. - Registering GPU Kernel in `.cu` files @@ -232,4 +235,108 @@ The system will automatically bind to Python and link it to a generated library. ## Unit Tests -Unit tests include comparing a forward operator's implementations on different devices, comparing a backward operator's implementation on different devices, and a scaling test for the backward operator. Here, we introduce the [unit tests for `MulOp`](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/framework/tests/test_mul_op.py). +Unit tests for an operator include + +1. comparing a forward operator's implementations on different devices, + +2. comparing a backward operator's implementation on different devices, and + +3. a scaling test for the backward operator. + +Here, we introduce the [unit tests for `MulOp`](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/framework/tests/test_mul_op.py). + +### Testing Forward Operators + +A forward operator unit test inherits `unittest.TestCase` and defines metaclass `__metaclass__ = OpTestMeta`. More concrete tests are performed in `OpTestMeta`. Testing a forward operator requires the following: + +1. Defining input, output and relevant attributes in `setUp` method. + +2. Generating random input data. + +3. Implementing the same computation logic in a Python script: + + ```python + import unittest + import numpy as np + from gradient_checker import GradientChecker, create_op + from op_test_util import OpTestMeta + + class TestMulOp(unittest.TestCase): + __metaclass__ = OpTestMeta + + def setUp(self): + self.type = "mul" + self.inputs = { + 'X': np.random.random((32, 84)).astype("float32"), + 'Y': np.random.random((84, 100)).astype("float32") + } + self.outputs = {'Out': np.dot(self.inputs['X'], self.inputs['Y'])} + ``` +Get its output, and compare it with the forward operator's own output. + +The code above first loads required packages. In addition, we have + +- `self.type = "mul" ` defines the type that is identical to what the operator's registered type. +- `self.inputs` defines input, with type `numpy.array` and initializes it. +- `self.outputs` defines output and completes the same operator computation in the Python script, and returns its result from the Python script. + +### Testing Backward Operators + +A backward operator unit test inherits `GradientChecker`, which inherits `unittest.TestCase`. As a result, **a backward operator unit test needs to be have the prefix `test_`**. + +```python +class TestMulGradOp(GradientChecker): + def setUp(self): + self.op = create_op("mul") + self.inputs = { + 'X': np.random.random((32, 84)).astype("float32"), + 'Y': np.random.random((84, 100)).astype("float32") + } + + def test_check_grad_normal(self): + # mul op will enlarge the relative error + self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.5) + + def test_check_grad_ingore_x(self): + self.check_grad( + ['Y'], 'Out', max_relative_error=0.5, no_grad_set=set("X")) + + def test_check_grad_ingore_y(self): + self.check_grad( + ['X'], 'Out', max_relative_error=0.5, no_grad_set=set('Y')) +``` + +Some key points in the code above include: + +- `create_op("mul")` creates the backward operator's corresponding forward operator. +- `test_normal` calls `check_grad` to validate scaling tests' correctness and stability through numeric methods. + - The first variable `["X", "Y"]` appoints `X` and `Y` to be scale tested. + - The second variable `"Out"` points to the network's final output target `Out`. + - The third variable `max_relative_error` points to the maximum relative tolerance error during scaling tests. +- `test_check_grad_ingore_x` and `test_check_grad_ingore_y`branches test the cases where there is only one scaling input. + +### Compiling and Running + + +Any new unit testing file of the format `test_*.py` added to the director `python/paddle/v2/framework/tests` is automatically added to the project to compile. + +Note that **unlike the compile test for Ops, running unit tests requires compiling the entire project** and requires compiling with flag `WITH_TESTING` on i.e. `cmake paddle_dir -DWITH_TESTING=ON`. + +After successfully compiling the project, run the following command to run unit tests: + +```bash +make test ARGS="-R test_mul_op -V" +``` + +Or, + +```bash +ctest -R test_mul_op +``` + +## Remarks + +- Every `*_op.h` (if applicable), `*_op.cc`, and `*_op.cu` (if applicable) must be created for a unique Op. Compiling will fail if multiple operators are included per file. +- The type with which an operator is registered needs to be identical to the Op's name. Registering `REGISTER_OP(B, ...)` in `A_op.cc` will cause unit testing failures. +- If the operator does not implement a GPU kernel, please refrain from creating an empty `*_op.cu` file, or else unit tests will fail. +- If multiple operators rely on some shared methods, a file NOT named `*_op.*` can be created to store them, such as `gather.h`. diff --git a/doc/howto/dev/use_eigen_en.md b/doc/howto/dev/use_eigen_en.md new file mode 100644 index 0000000000000000000000000000000000000000..e169106e12f5d62696f1f0e7163562793b32c18c --- /dev/null +++ b/doc/howto/dev/use_eigen_en.md @@ -0,0 +1,146 @@ +## How to use Eigen in Paddle + +Essentially, a neural network is a compute graph. T data needed for the computation is stored in `Tensor`s and its computation procedure is described by `Operator`s. An `Operator` calls the `Compute` interface in its corresponding `OpKernel` and operates on the `Tensor`. + + +### Eigen Tensor Module + +The Eigen Tensor module supports powerful element-wise computation. In addition, a piece of code written using it can be run on both the CPU and the GPU. + +Note that Eigen Tensor is still being actively developed, so its tests are not completely covered and its documentation may be sparse. + +For details on Eigen Tensor module, please see [doc 1](https://github.com/RLovelett/eigen/blob/master/unsupported/Eigen/CXX11/src/Tensor/README.md) and [doc 2](https://bitbucket.org/eigen/eigen/src/default/unsupported/Eigen/CXX11/src/Tensor/README.md). + + +### paddle::framework::Tensor + +Paddle Tensor's is defined in the framework directory with the following interface: + +```cpp +class Tensor { + public: + /*! Return a pointer to mutable memory block. */ + template + inline T* data(); + + /** + * @brief Return a pointer to mutable memory block. + * @note If not exist, then allocation. + */ + template + inline T* mutable_data(platform::Place place); + + /** + * @brief Return a pointer to mutable memory block. + * + * @param[in] dims The dimensions of the memory block. + * @param[in] place The place of the memory block. + * + * @note If not exist, then allocation. + */ + template + inline T* mutable_data(DDim dims, platform::Place place); + + /*! Resize the dimensions of the memory block. */ + inline Tensor& Resize(const DDim& dims); + + /*! Return the dimensions of the memory block. */ + inline const DDim& dims() const; + + private: + /*! holds the memory block if allocated. */ + std::shared_ptr holder_; + + /*! points to dimensions of memory block. */ + DDim dim_; +}; +``` + +`Placeholder` is used to delay memory allocation; that is, we can first define a tensor, using `Resize` to configure its shape, and then call `mutuable_data` to allocate the actual memory. + +```cpp +paddle::framework::Tensor t; +paddle::platform::CPUPlace place; +// set size first +t.Resize({2, 3}); +// allocate memory on CPU later +t.mutable_data(place); +``` + +### paddle::framework::Tensor Usage +`AddOp` demonstrates Tensor's usage. + +- InferShape + +When computing a neural network's compute graph, first call every `Operator`'s `InferShape` method, and use `Resize` to configure the size of the output tensor. + +```cpp +void InferShape(const framework::InferShapeContext &ctx) const override { + PADDLE_ENFORCE_EQ(ctx.Input("X")->dims(), + ctx.Input("Y")->dims(), + "Two input of Add Op's dimension must be same."); + ctx.Output("Out")->Resize(ctx.Input("X")->dims()); +} +``` + + +- Run + +```cpp +void Compute(const framework::ExecutionContext& context) const override { + auto* input0 = context.Input("X"); + auto* input1 = context.Input("Y"); + auto* output = context.Output("Out"); + + output->mutable_data(context.GetPlace()); + + auto x = EigenVector::Flatten(*input0); + auto y = EigenVector::Flatten(*input1); + auto z = EigenVector::Flatten(*output); + + auto place = context.GetEigenDevice(); + + z.device(place) = x + y; +} +``` + + +### paddle::framework::Tensor到EigenTensor的转换 + +As shown above, in actual computation, we need to transform the input and output `Tensor`s into formats Eigen supports. We show some functions in [eigen.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/eigen.h) to implement the transformation from `paddle::framework::Tensor`to `EigenTensor/EigenMatrix/EigenVector/EigenScalar`. + +Using EigenTensor as an example: + +```cpp +Tensor t; +float* p = t.mutable_data(make_ddim({1, 2, 3}), platform::CPUPlace()); +for (int i = 0; i < 1 * 2 * 3; i++) { + p[i] = static_cast(i); +} + +EigenTensor::Type et = EigenTensor::From(t); +``` + +`From` is an interfacing method provided by the EigenTensor template, which implements the transformation from a `paddle::framework::Tensor` object to an EigenTensor. Since `rank` is a template parameter, it needs to be explicitly specified at the time of the transformation. + +In Eigen, tensors with different ranks are different types, with `Vector` bring a rank-1 instance. Note that `EigenVector::From` uses a transformation from an 1-dimensional Paddle tensor to a 1-dimensional Eigen tensor while `EigenVector::Flatten` reshapes a paddle tensor and flattens it into a 1-dimensional Eigen tensor. Both resulting tensors are still typed EigenVector. + +For more transformations, see the [unit tests](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/eigen_test.cc) in the `eigen_test.cc` file. + + + +### Implementing Computation + +While computing, the device interface is needed from the EigenTensors on the left hand side of the assignments. Note that the computation between EigenTensors only changes the data originally inthe Tensor and does not change all the shape information associated with the Tensor. + +```cpp +auto x = EigenVector::Flatten(*input0); +auto y = EigenVector::Flatten(*input1); +auto z = EigenVector::Flatten(*output); +auto place = context.GetEigenDevice(); +z.device(place) = x + y; +``` + +In this code segment, input0/input1/output can be Tensors of arbitrary dimension. We are calling Flatten from EigenVector, transforming a tensor of any dimension into a 1-dimensional EigenVector. After completing computation, input0/input1/output will retain the same shape information, and they can be resized using the `Resize` interface. + +Because the Eigen Tensor module is under-documented, please refer to `OpKernel`'s computation code in TensorFlow's [kernel module documentation](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/core/kernels). diff --git a/paddle/CMakeLists.txt b/paddle/CMakeLists.txt index b435de80a224571d16efdee168541aa301c3f73a..7d2becbdd772747d77890321fce6721d8d17fb30 100644 --- a/paddle/CMakeLists.txt +++ b/paddle/CMakeLists.txt @@ -1,27 +1,32 @@ add_subdirectory(cuda) add_subdirectory(function) add_subdirectory(utils) -add_subdirectory(testing) add_subdirectory(math) -add_subdirectory(parameter) add_subdirectory(gserver) -add_subdirectory(pserver) -add_subdirectory(trainer) -add_subdirectory(scripts) -add_subdirectory(string) - -if(Boost_FOUND) - add_subdirectory(memory) - add_subdirectory(platform) - add_subdirectory(framework) - add_subdirectory(operators) - add_subdirectory(pybind) -endif() +add_subdirectory(parameter) +add_subdirectory(testing) -if(WITH_C_API) +if(MOBILE_INFERENCE) add_subdirectory(capi) -endif() +else() + add_subdirectory(pserver) + add_subdirectory(trainer) + add_subdirectory(string) + add_subdirectory(scripts) + + if(WITH_C_API) + add_subdirectory(capi) + endif() + + if(Boost_FOUND) + add_subdirectory(memory) + add_subdirectory(platform) + add_subdirectory(framework) + add_subdirectory(operators) + add_subdirectory(pybind) + endif() -if(WITH_SWIG_PY) - add_subdirectory(api) + if(WITH_SWIG_PY) + add_subdirectory(api) + endif() endif() diff --git a/paddle/api/Util.cpp b/paddle/api/Util.cpp index d369df5d4e04b4a8d822db0e72a8051150868ce6..11bd05c09d1ecbbcec6b6596c16416c26635a072 100644 --- a/paddle/api/Util.cpp +++ b/paddle/api/Util.cpp @@ -47,7 +47,7 @@ bool isUsingGpu() { return FLAGS_use_gpu; } void setUseGpu(bool useGpu) { FLAGS_use_gpu = useGpu; } bool isGpuVersion() { -#ifdef PADDLE_ONLY_CPU +#ifndef PADDLE_WITH_CUDA return false; #else return true; diff --git a/paddle/capi/CMakeLists.txt b/paddle/capi/CMakeLists.txt index b9bbe58951c643f1b1649858880fbd2ba3a2a7b7..2c458a78c598bf206b30c0c07599ce605af77701 100644 --- a/paddle/capi/CMakeLists.txt +++ b/paddle/capi/CMakeLists.txt @@ -37,9 +37,7 @@ set(PADDLE_CAPI_INFER_LIBS paddle_cuda paddle_function paddle_gserver - paddle_proto - paddle_pserver - paddle_network) + paddle_proto) cc_library(paddle_capi_whole DEPS paddle_capi ${PADDLE_CAPI_INFER_LIBS}) diff --git a/paddle/capi/Matrix.cpp b/paddle/capi/Matrix.cpp index d898ebe2612d749ca261d35139d1cd45bd355eef..4547afaf1dc9af8bc7909a684db766fdd7b159c0 100644 --- a/paddle/capi/Matrix.cpp +++ b/paddle/capi/Matrix.cpp @@ -46,7 +46,7 @@ paddle_error paddle_matrix_set_row(paddle_matrix mat, if (rowID >= ptr->mat->getHeight()) return kPD_OUT_OF_RANGE; paddle::real* buf = ptr->mat->getRowBuf(rowID); size_t width = ptr->mat->getWidth(); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA hl_memcpy(buf, rowArray, sizeof(paddle::real) * width); #else std::copy(rowArray, rowArray + width, buf); diff --git a/paddle/capi/tests/CMakeLists.txt b/paddle/capi/tests/CMakeLists.txt index 8208808b94f54f2ddaf4d426a65b8db562b36aca..bb38ace62808db5ce95a1a57ff465e8edc059213 100644 --- a/paddle/capi/tests/CMakeLists.txt +++ b/paddle/capi/tests/CMakeLists.txt @@ -4,11 +4,12 @@ add_unittest(capi_test_mats test_Vector.cpp target_include_directories(capi_test_mats PUBLIC ${PADDLE_CAPI_INC_PATH}) target_link_libraries(capi_test_mats paddle_capi) - -add_unittest_without_exec(capi_test_gradientMachine test_GradientMachine.cpp) -target_include_directories(capi_test_gradientMachine PUBLIC - ${PADDLE_CAPI_INC_PATH}) -target_link_libraries(capi_test_gradientMachine paddle_capi) -add_test(NAME capi_test_gradientMachine - COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python ${CMAKE_CURRENT_BINARY_DIR}/capi_test_gradientMachine - WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/capi/tests) +if(NOT MOBILE_INFERENCE) + add_unittest_without_exec(capi_test_gradientMachine test_GradientMachine.cpp) + target_include_directories(capi_test_gradientMachine PUBLIC + ${PADDLE_CAPI_INC_PATH}) + target_link_libraries(capi_test_gradientMachine paddle_capi) + add_test(NAME capi_test_gradientMachine + COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python ${CMAKE_CURRENT_BINARY_DIR}/capi_test_gradientMachine + WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/capi/tests) +endif() diff --git a/paddle/framework/CMakeLists.txt b/paddle/framework/CMakeLists.txt index e535f84dba7c2726fbb70fa11ca8e9e2d29b8665..1bf80b3e58df591376b79253c3eaf69355b3397f 100644 --- a/paddle/framework/CMakeLists.txt +++ b/paddle/framework/CMakeLists.txt @@ -19,16 +19,15 @@ cc_test(scope_test SRCS scope_test.cc DEPS scope) proto_library(framework_proto SRCS framework.proto) cc_library(attribute SRCS attribute.cc DEPS framework_proto) +cc_library(proto_desc SRCS var_desc.cc op_desc.cc block_desc.cc program_desc.cc DEPS attribute ddim) cc_library(op_proto_maker SRCS op_proto_maker.cc DEPS framework_proto attribute) cc_test(op_proto_maker_test SRCS op_proto_maker_test.cc DEPS op_proto_maker) -cc_library(op_info SRCS op_info.cc DEPS attribute framework_proto) -cc_library(operator SRCS operator.cc DEPS op_info device_context tensor scope) +cc_library(op_info SRCS op_info.cc DEPS attribute framework_proto proto_desc) +cc_library(operator SRCS operator.cc DEPS op_info device_context tensor scope proto_desc) cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry) -cc_library(grad_op_builder SRCS grad_op_builder.cc DEPS operator) -cc_library(op_registry SRCS op_registry.cc DEPS grad_op_builder op_proto_maker) +cc_library(op_registry SRCS op_registry.cc DEPS op_proto_maker op_info operator) cc_test(op_registry_test SRCS op_registry_test.cc DEPS op_registry) -cc_test(grad_op_builder_test SRCS grad_op_builder_test.cc DEPS grad_op_builder op_registry add_op) py_proto_compile(framework_py_proto SRCS framework.proto) # Generate an empty __init__.py to make framework_py_proto as a valid python module. @@ -42,3 +41,6 @@ add_custom_command(TARGET framework_py_proto POST_BUILD cc_library(backward SRCS backward.cc DEPS net_op) cc_test(backward_test SRCS backward_test.cc DEPS backward recurrent_op device_context) + +cc_library(tensor_array SRCS tensor_array.cc DEPS lod_tensor) +cc_test(tensor_array_test SRCS tensor_array_test.cc DEPS tensor_array place) diff --git a/paddle/framework/attribute.cc b/paddle/framework/attribute.cc index 510dc28c57f642786e7c64d86961c76ac80014a8..d6a2975aaa419406aef7b228e78381dbce78890d 100644 --- a/paddle/framework/attribute.cc +++ b/paddle/framework/attribute.cc @@ -24,6 +24,9 @@ static ProgramDesc* g_program_desc = nullptr; ProgramDesc& GetProgramDesc() { if (g_program_desc == nullptr) { g_program_desc = new ProgramDesc(); + auto root_block = g_program_desc->mutable_blocks()->Add(); + root_block->set_idx(0); + root_block->set_parent_idx(-1); } return *g_program_desc; } diff --git a/paddle/framework/attribute.h b/paddle/framework/attribute.h index 488fa38faf12ee51087643f79295f36bfd33ee22..d13530e3408a54c7ecab87c3bd9e6288e342f9af 100644 --- a/paddle/framework/attribute.h +++ b/paddle/framework/attribute.h @@ -21,20 +21,12 @@ limitations under the License. */ #include #include "paddle/framework/framework.pb.h" +#include "paddle/framework/type_defs.h" #include "paddle/platform/enforce.h" -#include "paddle/platform/variant.h" namespace paddle { namespace framework { -// The order should be as same as framework.proto -typedef boost::variant, - std::vector, std::vector, bool, - std::vector, BlockDesc*> - Attribute; - -typedef std::unordered_map AttributeMap; - ProgramDesc& GetProgramDesc(); template @@ -45,6 +37,21 @@ inline AttrType AttrTypeID() { Attribute GetAttrValue(const OpDesc::Attr& attr_desc); +class AttrReader { + public: + explicit AttrReader(const AttributeMap& attrs) : attrs_(attrs) {} + + template + inline const T& Get(const std::string& name) const { + PADDLE_ENFORCE(attrs_.count(name) != 0, "%s should be in AttributeMap", + name); + return boost::get(attrs_.at(name)); + } + + private: + const AttributeMap& attrs_; +}; + // check whether a value(attribute) fit a certain limit template class GreaterThanChecker { diff --git a/paddle/framework/backward.cc b/paddle/framework/backward.cc index 0ec18de5b8a0e7cebdb91c30d2b45596b02dfa51..0a4688db9c930201c95d4a47658f43cad87fdbca 100644 --- a/paddle/framework/backward.cc +++ b/paddle/framework/backward.cc @@ -13,10 +13,13 @@ limitations under the License. */ #include "paddle/framework/backward.h" +#include "paddle/operators/net_op.h" +#include #include #include +#include "paddle/framework/block_desc.h" #include "paddle/framework/op_registry.h" #include "paddle/operators/net_op.h" #include "paddle/operators/recurrent_op.h" @@ -24,6 +27,35 @@ namespace paddle { namespace framework { +static inline std::unique_ptr CreateGradOp( + const OperatorBase& op) { + OpDescBind op_desc; + op_desc.SetInputMap(op.Inputs()); + op_desc.SetOutputMap(op.Outputs()); + op_desc.SetType(op.Type()); + op_desc.SetAttrMap(op.Attrs()); + auto& info = OpInfoMap::Instance().Get(op.Type()); + auto grad_descs = info.GradOpMaker()(op_desc); + std::vector> grad_ops; + grad_ops.reserve(grad_descs.size()); + std::transform(grad_descs.begin(), grad_descs.end(), + std::back_inserter(grad_ops), + [](const std::unique_ptr& grad_desc) { + return OpRegistry::CreateOp(*grad_desc); + }); + PADDLE_ENFORCE(!grad_ops.empty()); + if (grad_ops.size() == 1) { + return std::move(grad_ops[0]); + } else { + auto net_op = new operators::NetOp(); + for (auto& grad_op : grad_ops) { + net_op->AppendOp(std::move(grad_op)); + } + net_op->CompleteAddOp(); + return std::unique_ptr(net_op); + } +} + template static void ForEachVarName(const Map& names, T callback) { for (auto& name : names) { @@ -141,9 +173,26 @@ static std::unique_ptr BackwardRecursive( net->ops_[op_offset]->Rename(name, dup_outputs.back()); } // collect all the offset to append `add` op for each alias - insert_position.push_back( - {dup_op.back(), OpRegistry::CreateOp("add", {{"X", {dup_outputs}}}, - {{"Out", {name}}}, {})}); + // + // one variable is shared between multiple operators. + // insert add operator one by one, then add it to output + for (size_t output_idx = 0; output_idx < dup_outputs.size() - 1; + ++output_idx) { + auto insert_add_x = dup_outputs[output_idx]; + auto insert_add_y = dup_outputs[output_idx + 1]; + auto insert_add_out = name + "@SHARED@" + std::to_string(output_idx); + // first add op inserted + if (output_idx == dup_outputs.size() - 2) { + insert_add_out = name; + } + if (output_idx != 0) { + insert_add_y = name + "@SHARED@" + std::to_string(output_idx - 1); + } + insert_position.push_back( + {dup_op.back(), + OpRegistry::CreateOp("sum", {{"X", {insert_add_x, insert_add_y}}}, + {{"Out", {insert_add_out}}}, {})}); + } } // make sure the inserted `add` ops follow the BFS order. @@ -154,7 +203,7 @@ static std::unique_ptr BackwardRecursive( net->InsertOp(pos.first + 1, std::move(pos.second)); } } else { - std::unique_ptr grad_op(OpRegistry::CreateGradOp(forwardOp)); + std::unique_ptr grad_op(CreateGradOp(forwardOp)); ForEachVarName(grad_op->Inputs(), [&no_grad_names, &net, &grad_op]( const std::string& grad_input) { @@ -182,7 +231,8 @@ static std::unique_ptr BackwardRecursive( // process recurrent gradient op as a special operator. if (forwardOp.Type() == "recurrent") { - // NOTE clean up cycle call somewhere (RNN's stepnet constains itself), or + // NOTE clean up cycle call somewhere (RNN's stepnet constains itself), + // or // this will result in infinite loop. const auto& rnnop = *static_cast(&forwardOp); @@ -222,5 +272,145 @@ std::unique_ptr Backward( return BackwardRecursive(forwardOp, no_grad_names, uid); } +// ==================================== // + +static bool AllGradInSet(const std::vector& names, + const std::unordered_set& set) { + for (const std::string& name : names) { + if (!set.count(GradVarName(name))) { + return false; + } + } + return true; +} + +std::vector> MakeOpGrad( + const std::unique_ptr& op_desc, + std::unordered_set& no_grad_vars) { + std::vector> grad_op_descs; + // All input gradients of forwarding operator do not need to calculat. + const std::vector& inputs = op_desc->InputArgumentNames(); + if (AllGradInSet(inputs, no_grad_vars)) { + return grad_op_descs; // empty vector + } + // All output gradients of forwarding operator do not need to calculate. + const std::vector& outputs = op_desc->OutputArgumentNames(); + if (AllGradInSet(outputs, no_grad_vars)) { + for (const std::string& name : inputs) { + no_grad_vars.insert(GradVarName(name)); + } + return grad_op_descs; // empty vector + } + + grad_op_descs = OpRegistry::CreateGradOpDescs(op_desc.get()); + + std::list> pending_fill_zeros_ops; + for (auto& desc : grad_op_descs) { + for (const std::string& in_name : desc->InputArgumentNames()) { + if (no_grad_vars.count(in_name)) { + std::string prefix = in_name.substr( + 0, in_name.size() - sizeof(kGradVarSuffix) / sizeof(char) + 1); + std::string new_name = prefix + kZeroVarSuffix; + desc->Rename(in_name, new_name); + std::unique_ptr fill_zeros_op(new OpDescBind( + "fill_zeros_like", {{"X", {prefix}}}, {{"Y", {new_name}}}, {})); + pending_fill_zeros_ops.push_back(std::move(fill_zeros_op)); + } + } + for (const std::string& out_name : desc->OutputArgumentNames()) { + if (no_grad_vars.count(out_name)) { + desc->Rename(out_name, kEmptyVarName); + } + } + } + + for (auto& p : pending_fill_zeros_ops) { + grad_op_descs.insert(grad_op_descs.begin(), std::move(p)); + } + return grad_op_descs; +} + +std::vector> MakeBlockBackward( + ProgramDescBind& program_desc, int block_idx, + std::unordered_set& no_grad_vars) { + BlockDescBind* cur_block = program_desc.Block(block_idx); + std::deque>& op_descs = cur_block->ops_; + std::unordered_map> dup_out_ops; + size_t grad_desc_idx = 0; + std::vector> backward_descs; + for (auto it = op_descs.rbegin(); it != op_descs.rend(); ++it) { + std::vector> op_grads = + MakeOpGrad(*it, no_grad_vars); + + if ((*it)->Type() == "recurrent") { + PADDLE_ENFORCE_EQ( + op_grads.size(), size_t(1), + "rnn_op's gradient process should contain only one op."); + int step_block_idx = (*it)->GetBlockAttr("stop_block"); + auto backward_block_op_descs = + MakeBlockBackward(program_desc, step_block_idx, no_grad_vars); + BlockDescBind* backward_block = program_desc.AppendBlock(*cur_block); + for (auto& ptr : backward_block_op_descs) { + backward_block->ops_.push_back(std::move(ptr)); + } + op_grads[0]->SetBlockAttr("step_block", *backward_block); + } + + for (const auto& desc : op_grads) { + for (const std::string& out_name : desc->OutputArgumentNames()) { + dup_out_ops[out_name].emplace_back(grad_desc_idx); + } + ++grad_desc_idx; + } + std::transform( + op_grads.begin(), op_grads.end(), std::back_inserter(backward_descs), + [](std::unique_ptr& ptr) { return std::move(ptr); }); + } + // Check whether some variables are written more than once + std::list>> pending_sum_ops; + for (const auto& dup : dup_out_ops) { + const std::string& out_name = dup.first; + const std::vector dup_op = dup.second; + if (out_name != kEmptyVarName && dup_op.size() > 1) { + std::vector sum_op_inputs; + for (size_t i = 0; i < dup_op.size(); ++i) { + std::string new_name = out_name + "@RENAME@" + std::to_string(i); + backward_descs[dup_op[i]]->Rename(out_name, new_name); + sum_op_inputs.emplace_back(new_name); + } + std::unique_ptr sum_op(new OpDescBind( + "sum", {{"X", sum_op_inputs}}, {{"Out", {out_name}}}, {})); + pending_sum_ops.push_back({dup_op.back(), std::move(sum_op)}); + } + } + pending_sum_ops.sort( + [](const std::pair>& a, + const std::pair>& b) { + return a.first > b.first; + }); + for (auto& p : pending_sum_ops) { + backward_descs.insert(backward_descs.begin() + p.first + 1, + std::move(p.second)); + } + return backward_descs; +} + +void AppendBackward(ProgramDescBind& program_desc, + const std::unordered_set& no_grad_vars) { + std::unordered_set no_grad_var_names; + no_grad_var_names.reserve(no_grad_vars.size() + 1); + no_grad_var_names.insert(std::string(kEmptyVarName) + kGradVarSuffix); + for (auto& name : no_grad_vars) { + no_grad_var_names.insert(GradVarName(name)); + } + const int root_block_idx = 0; + auto backward_op_descs = + MakeBlockBackward(program_desc, root_block_idx, no_grad_var_names); + auto& forw_op_descs = program_desc.Block(root_block_idx)->ops_; + for (auto& ptr : backward_op_descs) { + forw_op_descs.push_back(std::move(ptr)); + } +} + } // namespace framework } // namespace paddle diff --git a/paddle/framework/backward.h b/paddle/framework/backward.h index 1ecf69881b3126c2904920b9f4b77bfcccc9cf86..7ffe4c28103f9d6a9f179422d1beb86106ef786e 100644 --- a/paddle/framework/backward.h +++ b/paddle/framework/backward.h @@ -13,8 +13,11 @@ limitations under the License. */ #pragma once + #include -#include "operator.h" +#include "paddle/framework/operator.h" +#include "paddle/framework/program_desc.h" + namespace paddle { namespace framework { @@ -23,5 +26,9 @@ namespace framework { extern std::unique_ptr Backward( const OperatorBase& forwardOp, const std::unordered_set& no_grad_vars); + +void AppendBackward(ProgramDescBind& program_desc, + const std::unordered_set& no_grad_vars); + } // namespace framework } // namespace paddle diff --git a/paddle/framework/backward.md b/paddle/framework/backward.md index 0a6d762bc8be5201ac196b4bc6107c06d07a31d7..ac60be572419b62f4beb644ff192d413c35e19bb 100644 --- a/paddle/framework/backward.md +++ b/paddle/framework/backward.md @@ -2,7 +2,7 @@ ## Motivation -In Neural Network, many model is solved by the the backpropagation algorithm(known as BP) at present. Technically it caculates the gradient of the loss function, then distributed back through the networks. Follows the chain rule, so we need a module chains the gradient operators/expressions together with to construct the backward pass. Every forward network needs a backward network to construct the full computation graph, the operator/expression's backward pass will be generated respect to forward pass. +In Neural Network, most models are solved by the backpropagation algorithm(known as **BP**) at present. Technically, BP calculates the gradient of the loss function, then propagates it back through the networks following the chain rule. Hence we need a module that chains the gradient operators/expressions together to construct the backward pass. Every forward network needs a backward network to construct the full computation graph. The operator/expression's backward pass will be generated with respect to the forward pass. ## Implementation @@ -24,9 +24,9 @@ A backward network is built up with several backward operators. Backward operato | **Operator::inputs_** | Inputs | Inputs, Outputs, OutputGradients | | **Operator::outputs_** | Outputs | InputGradients | - In most cases, there is a one-to-one correspondence between the forward and backward operators. These correspondences are recorded by a global hash map(`OpInfoMap`). To follow the philosophy of minimum core and make operators pluggable, the registry mechanism is introduced. + In most cases, there is a one-to-one relation between the forward and backward operators. These relations are recorded by a global hash map(`OpInfoMap`). To follow the philosophy of minimum core and to make operators pluggable, the registry mechanism is introduced. -For example, we have got a `mul_op`, and we can register its information and corresponding backward operator by the following macro: +For example, we have `mul_op`, and we can register its information and corresponding backward operator by the following macro: ```cpp REGISTER_OP(mul, MulOp, MulOpMaker, mul_grad, MulOpGrad); @@ -48,7 +48,7 @@ The function `BuildGradOp` will sequentially execute following processes: 1. Get the `type_` of given forward operator, and then get the corresponding backward operator's type by looking up the `OpInfoMap`. -2. Build two maps named `inputs` and `outputs` to temporary storage backward operator's inputs and outputs. Copy forward operator's `inputs_` and `outputs_` to map `inputs`, except these, are not necessary for gradient computing. +2. Build two maps named `inputs` and `outputs` to temporarily store backward operator's inputs and outputs. Copy forward operator's `inputs_` and `outputs_` to map `inputs`, except these, are not necessary for gradient computing. 3. Add forward inputs' gradient variables into map `output`, adding forward outputs' gradient variables into map `input`. @@ -56,11 +56,11 @@ The function `BuildGradOp` will sequentially execute following processes: ### Backward Network Building -A backward network is a series of backward operators. The main idea of building a backward network is creating backward operators in the inverted sequence and append them together one by one. There is some corner case need to process specially. +A backward network is a series of backward operators. The main idea of building a backward network is creating backward operators in the inverted sequence and appending them together one by one. There are some corner cases that need special processing. 1. Op - When the input forward network is an Op, return its gradient Operator Immediately. If all of its outputs are in no gradient set, then return a special `NOP`. + When the input forward network is an Op, return its gradient Operator immediately. If all of its outputs are in no gradient set, then return a special `NOP`. 2. NetOp @@ -68,33 +68,33 @@ A backward network is a series of backward operators. The main idea of building 3. RnnOp - RnnOp is a nested stepnet operator. Backward module need to recusively call `Backward` for every stepnet. + RnnOp is a nested stepnet operator. Backward module needs to recusively call `Backward` for every stepnet. 4. Sharing Variables - **sharing variables**. As illustrated in the pictures, two operator's share the same variable name of W@GRAD, which will overwrite their sharing input variable. + As illustrated in the figure 1 and figure 2, two operators share the same variable name **W@GRAD**, which will overwrite their shared input variable.


-​ pic 1. Sharing variables in operators. +​ Figure 1. Sharing variables in operators.

-​ Sharing variable between operators or same input variable used in multiple operators leads to a duplicate gradient variable. As demo show above, we need to rename gradient name recursively and add a generic add operator to replace the overwrite links. +​ Sharing variable between operators or same input variable used in multiple operators can lead to duplicate gradient variables. As illustrated in figure 2, we need to rename the gradient names recursively and add a generic add operator to prevent overwriting.


-​ pic 2. Replace sharing variable's gradient with `Add` operator. +​ Figure 2. Replace sharing variable's gradient with `Add` operator.

-​ Because our framework finds variables accord to their names, we need to rename the output links. We add a suffix of number to represent its position in clockwise. +​ Because the framework finds variables according to their names, we need to rename the output links. We add an integer suffix to represent its position in the clockwise direction. -5. Part of Gradient is Zero. +5. Part of the Gradient is Zero. - In the whole graph, there is some case of that one operator's gradient is not needed, but its input's gradient is a dependency link of other operator, we need to fill a same shape gradient matrix in the position. In our implement, we insert a special `fillZeroLike` operator. + In the whole graph, there is some case of that one operator's gradient is not needed, but its input's gradient is a dependency link of other operator, we need to fill a same shape gradient matrix in the position. In our implementation, we insert a special `fillZeroLike` operator. Follow these rules above, then collect the sub graph `OutputGradients`/`InputGradients` as the NetOp's and return it. diff --git a/paddle/framework/backward_test.cc b/paddle/framework/backward_test.cc index 6932f5b989a3e21ebc44ec4fec9f5223f2547d7a..3b7cbcd98927be829d185590147adf74cd3d10d1 100644 --- a/paddle/framework/backward_test.cc +++ b/paddle/framework/backward_test.cc @@ -15,30 +15,42 @@ #include "paddle/framework/backward.h" #include +#include "paddle/framework/block_desc.h" +#include "paddle/framework/op_desc.h" #include "paddle/framework/op_registry.h" #include "paddle/operators/net_op.h" namespace paddle { namespace framework { -using OperatorBase = framework::OperatorBase; -using OpProtoAndCheckerMaker = framework::OpProtoAndCheckerMaker; -using OpProto = framework::OpProto; -using OpAttrChecker = framework::OpAttrChecker; -using Scope = framework::Scope; using DeviceContext = platform::DeviceContext; class RowWiseAddOpMaker : public OpProtoAndCheckerMaker { public: RowWiseAddOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("X", "Input X of Add").NotInGradient(); - AddInput("b", "Bias of Add").NotInGradient(); - AddOutput("Out", "Out of Add").NotInGradient(); + AddInput("X", "Input X of Add"); + AddInput("b", "Bias of Add"); + AddOutput("Out", "Out of Add"); AddComment("Add Op"); } }; +class RowWiseAddGradMaker : public SingleGradOpDescMaker { + public: + using SingleGradOpDescMaker::SingleGradOpDescMaker; + + protected: + std::unique_ptr Apply() const override { + auto grad_op = new OpDescBind(); + grad_op->SetInput(GradVarName("Out"), OutputGrad("Out")); + grad_op->SetOutput(GradVarName("X"), InputGrad("X")); + grad_op->SetOutput(GradVarName("b"), InputGrad("b")); + grad_op->SetType("rowwise_add_grad"); + return std::unique_ptr(grad_op); + } +}; + class MulOpMaker : public OpProtoAndCheckerMaker { public: MulOpMaker(OpProto *proto, OpAttrChecker *op_checker) @@ -46,6 +58,8 @@ class MulOpMaker : public OpProtoAndCheckerMaker { AddInput("X", "A"); AddInput("Y", "B"); AddOutput("Out", "Out"); + AddAttr("x_num_col_dims", "").SetDefault(1).EqualGreaterThan(1); + AddAttr("y_num_col_dims", "").SetDefault(1).EqualGreaterThan(1); AddComment("Mul"); } }; @@ -133,42 +147,46 @@ class FillZeroOpMaker : public OpProtoAndCheckerMaker { } }; -class AddOpMaker : public OpProtoAndCheckerMaker { +class SumOpMaker : public framework::OpProtoAndCheckerMaker { + public: + SumOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "the input tensors of sum operator.").AsDuplicable(); + AddOutput("Out", "the output tensor of sum operator."); + AddComment(""); + } +}; + +class MultInOutOpMaker : public OpProtoAndCheckerMaker { public: - AddOpMaker(OpProto *proto, OpAttrChecker *op_checker) + MultInOutOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("X", "x").AsDuplicable(); - AddOutput("Out", "out"); + AddInput("X", "x"); + AddInput("H", "h"); + AddOutput("Y", "y"); + AddOutput("Z", "z"); AddComment(""); } }; + } // namespace framework } // namespace paddle namespace f = paddle::framework; namespace ops = paddle::operators; using EnforceNotMet = paddle::platform::EnforceNotMet; -REGISTER_OP(rowwise_add, f::NOP, f::RowWiseAddOpMaker, rowwise_add_grad, - f::NOP); +REGISTER_OPERATOR(rowwise_add, f::NOP, f::RowWiseAddOpMaker, + f::RowWiseAddGradMaker); +REGISTER_OPERATOR(rowwise_add_grad, f::NOP); REGISTER_OP(mul, f::NOP, f::MulOpMaker, mul_grad, f::NOP); REGISTER_OP(sigmoid, f::NOP, f::SigmoidOpMaker, sigmoid_grad, f::NOP); REGISTER_OP_WITHOUT_GRADIENT(nograd, f::NOP, f::NoGradOpMaker); REGISTER_OP_WITHOUT_GRADIENT(fill_zeros_like, f::NOP, f::FillZeroOpMaker); -REGISTER_OP(add, f::NOP, f::AddOpMaker, add_grad, f::NOP); +REGISTER_OP(sum, f::NOP, f::SumOpMaker, sum_grad, f::NOP); REGISTER_OP_WITHOUT_GRADIENT(fc, f::FcOp, f::FcOpMaker); REGISTER_OP(many_output_op, f::NOP, f::ManyOutputOpMaker, many_output_op_grad, f::NOP); - -TEST(Backward, simple_op_grad) { - auto fwd = f::OpRegistry::CreateOp( - "rowwise_add", {{"X", {"x"}}, {"b", {"b"}}}, {{"Out", {"out"}}}, {}); - ASSERT_NE(fwd, nullptr); - auto gop = f::OpRegistry::CreateGradOp(*fwd); - ASSERT_EQ(1UL, gop->Inputs().size()); - ASSERT_EQ("rowwise_add_grad", gop->Type()); - ASSERT_EQ(f::GradVarName("x"), gop->Output(f::GradVarName("X"))); - ASSERT_EQ(f::GradVarName("b"), gop->Output(f::GradVarName("b"))); -} +REGISTER_OP(mult_in_out, f::NOP, f::MultInOutOpMaker, mult_in_out_grad, f::NOP); TEST(Backward, simple_op_not_need_grad) { auto fwd = f::OpRegistry::CreateOp( @@ -283,18 +301,7 @@ TEST(Backward, net_shared_weight) { ASSERT_TRUE(bwd->IsNetOp()); auto bwd_net = static_cast(bwd.get()); ASSERT_EQ(3UL, bwd_net->ops_.size()); - ASSERT_EQ("add", bwd_net->ops_[2]->Type()); -} - -TEST(Backward, op_register_grad_not_for_network) { - auto fwd = - f::OpRegistry::CreateOp("fc", {{"X", {"x"}}, {"W", {"w"}}, {"b", {"b"}}}, - {{"mul_result", {"mul_out"}}, - {"add_result", {"add_out"}}, - {"Out", {"out1"}}}, - {{"temporary_index", std::vector{0, 1}}}); - - ASSERT_THROW(f::OpRegistry::CreateGradOp(*fwd), EnforceNotMet); + ASSERT_EQ("sum", bwd_net->ops_[2]->Type()); } TEST(Backward, op_all_input_are_not_need) { @@ -399,3 +406,315 @@ TEST(Backward, linear_net_intermediate_variable_has_no_grad) { EXPECT_EQ(bwd_net->ops_[2]->Inputs(all).size(), 0UL); EXPECT_EQ(bwd_net->ops_[2]->Outputs(all).size(), 0UL); } + +// =================================== // + +f::ProgramDesc *GetNewProgramDesc() { + auto *program_desc = new f::ProgramDesc(); + auto *root_block = program_desc->add_blocks(); + root_block->set_idx(0); + root_block->set_parent_idx(-1); + return program_desc; +} + +TEST(Backward, simple_single_op) { + f::ProgramDesc *program_desc = GetNewProgramDesc(); + f::ProgramDescBind &program = f::ProgramDescBind::Instance(program_desc); + f::BlockDescBind *block = program.Block(0); + f::OpDescBind *op = block->AppendOp(); + op->SetType("rowwise_add"); + op->SetInput("X", {"x"}); + op->SetInput("b", {"b"}); + op->SetOutput("Out", {"out"}); + + AppendBackward(program, {}); + + ASSERT_EQ(block->AllOps().size(), 2UL); + f::OpDescBind *grad_op = block->AllOps()[1]; + EXPECT_EQ(grad_op->Type(), "rowwise_add_grad"); + ASSERT_EQ(grad_op->InputNames().size(), 1UL); + ASSERT_EQ(grad_op->OutputNames().size(), 2UL); + EXPECT_EQ(grad_op->Input(f::GradVarName("Out")), + std::vector({f::GradVarName("out")})); + EXPECT_EQ(grad_op->Output(f::GradVarName("X")), + std::vector({f::GradVarName("x")})); + EXPECT_EQ(grad_op->Output(f::GradVarName("b")), + std::vector({f::GradVarName("b")})); +} + +TEST(Backward, default_attribute) { + f::ProgramDesc *program_desc = GetNewProgramDesc(); + f::ProgramDescBind &program = f::ProgramDescBind::Instance(program_desc); + f::BlockDescBind *block = program.Block(0); + f::OpDescBind *op = block->AppendOp(); + op->SetType("mul"); + op->SetInput("X", {"x"}); + op->SetInput("Y", {"y"}); + op->SetOutput("Out", {"out"}); + + AppendBackward(program, {}); + + ASSERT_EQ(block->AllOps().size(), 2UL); + EXPECT_EQ(boost::get(op->GetAttr("x_num_col_dims")), 1); + EXPECT_EQ(boost::get(op->GetAttr("y_num_col_dims")), 1); + + f::OpDescBind *grad_op = block->AllOps()[1]; + ASSERT_EQ(grad_op->Type(), "mul_grad"); + EXPECT_EQ(boost::get(grad_op->GetAttr("x_num_col_dims")), 1); + EXPECT_EQ(boost::get(grad_op->GetAttr("y_num_col_dims")), 1); +} + +TEST(Backward, simple_mult_op) { + f::ProgramDesc *program_desc = GetNewProgramDesc(); + f::ProgramDescBind &program = f::ProgramDescBind::Instance(program_desc); + f::BlockDescBind *block = program.Block(0); + f::OpDescBind *op1 = block->AppendOp(); + op1->SetType("rowwise_add"); + op1->SetInput("X", {"x1"}); + op1->SetInput("b", {"b1"}); + op1->SetOutput("Out", {"out1"}); + + f::OpDescBind *op2 = block->AppendOp(); + op2->SetType("mul"); + op2->SetInput("X", {"out1"}); + op2->SetInput("Y", {"y2"}); + op2->SetOutput("Out", {"out2"}); + + f::OpDescBind *op3 = block->AppendOp(); + op3->SetType("rowwise_add"); + op3->SetInput("X", {"out2"}); + op3->SetInput("b", {"b3"}); + op3->SetOutput("Out", {"out3"}); + + AppendBackward(program, {}); + + ASSERT_EQ(block->AllOps().size(), 6UL); + f::OpDescBind *grad_op1 = block->AllOps()[5]; + EXPECT_EQ(grad_op1->Type(), "rowwise_add_grad"); + ASSERT_EQ(grad_op1->InputNames().size(), 1UL); + ASSERT_EQ(grad_op1->OutputNames().size(), 2UL); + EXPECT_EQ(grad_op1->Input(f::GradVarName("Out")), + std::vector({f::GradVarName("out1")})); + EXPECT_EQ(grad_op1->Output(f::GradVarName("X")), + std::vector({f::GradVarName("x1")})); + EXPECT_EQ(grad_op1->Output(f::GradVarName("b")), + std::vector({f::GradVarName("b1")})); + + f::OpDescBind *grad_op2 = block->AllOps()[4]; + EXPECT_EQ(grad_op2->Type(), "mul_grad"); + ASSERT_EQ(grad_op2->InputNames().size(), 4UL); + ASSERT_EQ(grad_op2->OutputNames().size(), 2UL); + EXPECT_EQ(grad_op2->Input("X"), std::vector({"out1"})); + EXPECT_EQ(grad_op2->Input("Y"), std::vector({"y2"})); + EXPECT_EQ(grad_op2->Input("Out"), std::vector({"out2"})); + EXPECT_EQ(grad_op2->Input(f::GradVarName("Out")), + std::vector({f::GradVarName("out2")})); + EXPECT_EQ(grad_op2->Output(f::GradVarName("X")), + std::vector({f::GradVarName("out1")})); + EXPECT_EQ(grad_op2->Output(f::GradVarName("Y")), + std::vector({f::GradVarName("y2")})); + + f::OpDescBind *grad_op3 = block->AllOps()[3]; + EXPECT_EQ(grad_op3->Type(), "rowwise_add_grad"); + ASSERT_EQ(grad_op3->InputNames().size(), 1UL); + ASSERT_EQ(grad_op3->OutputNames().size(), 2UL); + EXPECT_EQ(grad_op3->Input(f::GradVarName("Out")), + std::vector({f::GradVarName("out3")})); + EXPECT_EQ(grad_op3->Output(f::GradVarName("X")), + std::vector({f::GradVarName("out2")})); + EXPECT_EQ(grad_op3->Output(f::GradVarName("b")), + std::vector({f::GradVarName("b3")})); +} + +TEST(Backward, intermedia_var_no_grad) { + f::ProgramDesc *program_desc = GetNewProgramDesc(); + f::ProgramDescBind &program = f::ProgramDescBind::Instance(program_desc); + f::BlockDescBind *block = program.Block(0); + f::OpDescBind *op1 = block->AppendOp(); + op1->SetType("rowwise_add"); + op1->SetInput("X", {"x1"}); + op1->SetInput("b", {"b1"}); + op1->SetOutput("Out", {"out1"}); + + f::OpDescBind *op2 = block->AppendOp(); + op2->SetType("mul"); + op2->SetInput("X", {"x2"}); + op2->SetInput("Y", {"y2"}); + op2->SetOutput("Out", {"out2"}); + + f::OpDescBind *op3 = block->AppendOp(); + op3->SetType("rowwise_add"); + op3->SetInput("X", {"out2"}); + op3->SetInput("b", {"b3"}); + op3->SetOutput("Out", {"out3"}); + + f::OpDescBind *op4 = block->AppendOp(); + op4->SetType("mul"); + op4->SetInput("X", {"out1"}); + op4->SetInput("Y", {"out3"}); + op4->SetOutput("Out", {"out4"}); + + AppendBackward(program, {"out3"}); + + ASSERT_EQ(block->AllOps().size(), 6UL); + f::OpDescBind *grad_op1 = block->AllOps()[5]; + EXPECT_EQ(grad_op1->Type(), "rowwise_add_grad"); + ASSERT_EQ(grad_op1->InputNames().size(), 1UL); + ASSERT_EQ(grad_op1->OutputNames().size(), 2UL); + EXPECT_EQ(grad_op1->Input(f::GradVarName("Out")), + std::vector({f::GradVarName("out1")})); + EXPECT_EQ(grad_op1->Output(f::GradVarName("X")), + std::vector({f::GradVarName("x1")})); + EXPECT_EQ(grad_op1->Output(f::GradVarName("b")), + std::vector({f::GradVarName("b1")})); + + f::OpDescBind *grad_op4 = block->AllOps()[4]; + EXPECT_EQ(grad_op4->Type(), "mul_grad"); + ASSERT_EQ(grad_op4->InputNames().size(), 4UL); + ASSERT_EQ(grad_op4->OutputNames().size(), 2UL); + EXPECT_EQ(grad_op4->Input("X"), std::vector({"out1"})); + EXPECT_EQ(grad_op4->Input("Y"), std::vector({"out3"})); + EXPECT_EQ(grad_op4->Input("Out"), std::vector({"out4"})); + EXPECT_EQ(grad_op4->Input(f::GradVarName("Out")), + std::vector({f::GradVarName("out4")})); + EXPECT_EQ(grad_op4->Output(f::GradVarName("X")), + std::vector({f::GradVarName("out1")})); + EXPECT_EQ(grad_op4->Output(f::GradVarName("Y")), + std::vector({f::kEmptyVarName})); +} + +TEST(Backward, var_no_grad) { + f::ProgramDesc *program_desc = GetNewProgramDesc(); + f::ProgramDescBind &program = f::ProgramDescBind::Instance(program_desc); + f::BlockDescBind *block = program.Block(0); + f::OpDescBind *op1 = block->AppendOp(); + op1->SetType("mult_in_out"); + op1->SetInput("X", {"x1"}); + op1->SetInput("H", {"h1"}); + op1->SetOutput("Y", {"y1"}); + op1->SetOutput("Z", {"z1"}); + + f::OpDescBind *op2 = block->AppendOp(); + op2->SetType("mult_in_out"); + op2->SetInput("X", {"y1"}); + op2->SetInput("H", {"z1"}); + op2->SetOutput("Y", {"y2"}); + op2->SetOutput("Z", {"z2"}); + + AppendBackward(program, {"z1"}); + + ASSERT_EQ(block->AllOps().size(), 5UL); + f::OpDescBind *grad_op2 = block->AllOps()[2]; + ASSERT_EQ(grad_op2->Type(), "mult_in_out_grad"); + ASSERT_EQ(grad_op2->InputNames().size(), 6UL); + ASSERT_EQ(grad_op2->OutputNames().size(), 2UL); + EXPECT_EQ(grad_op2->Input("X"), std::vector({"y1"})); + EXPECT_EQ(grad_op2->Input("H"), std::vector({"z1"})); + EXPECT_EQ(grad_op2->Input("Y"), std::vector({"y2"})); + EXPECT_EQ(grad_op2->Input("Z"), std::vector({"z2"})); + EXPECT_EQ(grad_op2->Input(f::GradVarName("Y")), + std::vector({f::GradVarName("y2")})); + EXPECT_EQ(grad_op2->Input(f::GradVarName("Z")), + std::vector({f::GradVarName("z2")})); + EXPECT_EQ(grad_op2->Output(f::GradVarName("X")), + std::vector({f::GradVarName("y1")})); + EXPECT_EQ(grad_op2->Output(f::GradVarName("H")), + std::vector({f::kEmptyVarName})); + + f::OpDescBind *fill_zero_op = block->AllOps()[3]; + ASSERT_EQ(fill_zero_op->Type(), "fill_zeros_like"); + ASSERT_EQ(fill_zero_op->InputNames().size(), 1UL); + ASSERT_EQ(fill_zero_op->OutputNames().size(), 1UL); + EXPECT_EQ(fill_zero_op->Input("X"), std::vector({"z1"})); + EXPECT_EQ(fill_zero_op->Output("Y"), + std::vector({std::string("z1") + f::kZeroVarSuffix})); + + f::OpDescBind *grad_op1 = block->AllOps()[4]; + ASSERT_EQ(grad_op1->Type(), "mult_in_out_grad"); + ASSERT_EQ(grad_op1->InputNames().size(), 6UL); + ASSERT_EQ(grad_op1->OutputNames().size(), 2UL); + EXPECT_EQ(grad_op1->Input("X"), std::vector({"x1"})); + EXPECT_EQ(grad_op1->Input("H"), std::vector({"h1"})); + EXPECT_EQ(grad_op1->Input("Y"), std::vector({"y1"})); + EXPECT_EQ(grad_op1->Input("Z"), std::vector({"z1"})); + EXPECT_EQ(grad_op1->Input(f::GradVarName("Y")), + std::vector({f::GradVarName("y1")})); + EXPECT_EQ(grad_op1->Input(f::GradVarName("Z")), + std::vector({std::string("z1") + f::kZeroVarSuffix})); + EXPECT_EQ(grad_op1->Output(f::GradVarName("X")), + std::vector({f::GradVarName("x1")})); + EXPECT_EQ(grad_op1->Output(f::GradVarName("H")), + std::vector({f::GradVarName("h1")})); +} + +TEST(Backward, shared_var) { + f::ProgramDesc *program_desc = GetNewProgramDesc(); + f::ProgramDescBind &program = f::ProgramDescBind::Instance(program_desc); + f::BlockDescBind *block = program.Block(0); + f::OpDescBind *op1 = block->AppendOp(); + op1->SetType("rowwise_add"); + op1->SetInput("X", {"x1"}); + op1->SetInput("b", {"b1"}); + op1->SetOutput("Out", {"out1"}); + + f::OpDescBind *op2 = block->AppendOp(); + op2->SetType("mul"); + op2->SetInput("X", {"out1"}); + op2->SetInput("Y", {"y2"}); + op2->SetOutput("Out", {"out2"}); + + f::OpDescBind *op3 = block->AppendOp(); + op3->SetType("rowwise_add"); + op3->SetInput("X", {"out1"}); + op3->SetInput("b", {"b3"}); + op3->SetOutput("Out", {"out3"}); + + AppendBackward(program, {}); + + ASSERT_EQ(block->AllOps().size(), 7UL); + f::OpDescBind *grad_op3 = block->AllOps()[3]; + ASSERT_EQ(grad_op3->Type(), "rowwise_add_grad"); + ASSERT_EQ(grad_op3->InputNames().size(), 1UL); + ASSERT_EQ(grad_op3->OutputNames().size(), 2UL); + EXPECT_EQ(grad_op3->Input(f::GradVarName("Out")), + std::vector({f::GradVarName("out3")})); + EXPECT_EQ(grad_op3->Output(f::GradVarName("X")), + std::vector({f::GradVarName("out1") + "@RENAME@0"})); + EXPECT_EQ(grad_op3->Output(f::GradVarName("b")), + std::vector({f::GradVarName("b3")})); + + f::OpDescBind *grad_op4 = block->AllOps()[4]; + ASSERT_EQ(grad_op4->Type(), "mul_grad"); + ASSERT_EQ(grad_op4->InputNames().size(), 4UL); + ASSERT_EQ(grad_op4->OutputNames().size(), 2UL); + EXPECT_EQ(grad_op4->Input("X"), std::vector({"out1"})); + EXPECT_EQ(grad_op4->Input("Y"), std::vector({"y2"})); + EXPECT_EQ(grad_op4->Input("Out"), std::vector({"out2"})); + EXPECT_EQ(grad_op4->Input(f::GradVarName("Out")), + std::vector({f::GradVarName("out2")})); + EXPECT_EQ(grad_op4->Output(f::GradVarName("X")), + std::vector({f::GradVarName("out1") + "@RENAME@1"})); + EXPECT_EQ(grad_op4->Output(f::GradVarName("Y")), + std::vector({f::GradVarName("y2")})); + + f::OpDescBind *sum_op = block->AllOps()[5]; + ASSERT_EQ(sum_op->Type(), "sum"); + ASSERT_EQ(sum_op->InputNames().size(), 1UL); + ASSERT_EQ(sum_op->OutputNames().size(), 1UL); + EXPECT_EQ(sum_op->Input("X"), + std::vector({f::GradVarName("out1") + "@RENAME@0", + f::GradVarName("out1") + "@RENAME@1"})); + EXPECT_EQ(sum_op->Output("Out"), + std::vector({f::GradVarName("out1")})); + + f::OpDescBind *grad_op1 = block->AllOps()[6]; + ASSERT_EQ(grad_op1->Type(), "rowwise_add_grad"); + ASSERT_EQ(grad_op1->InputNames().size(), 1UL); + ASSERT_EQ(grad_op1->OutputNames().size(), 2UL); + EXPECT_EQ(grad_op1->Input(f::GradVarName("Out")), + std::vector({f::GradVarName("out1")})); + EXPECT_EQ(grad_op1->Output(f::GradVarName("X")), + std::vector({f::GradVarName("x1")})); + EXPECT_EQ(grad_op1->Output(f::GradVarName("b")), + std::vector({f::GradVarName("b1")})); +} \ No newline at end of file diff --git a/paddle/framework/block_desc.cc b/paddle/framework/block_desc.cc new file mode 100644 index 0000000000000000000000000000000000000000..509aa235d3ee226adef15f08f5785866700499f1 --- /dev/null +++ b/paddle/framework/block_desc.cc @@ -0,0 +1,99 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/framework/block_desc.h" +#include "paddle/framework/program_desc.h" + +namespace paddle { +namespace framework { + +VarDescBind *BlockDescBind::NewVar(const std::string &name) { + need_update_ = true; + auto it = vars_.find(name); + PADDLE_ENFORCE(it == vars_.end(), "Duplicated variable %s", name); + auto var = new VarDescBind(name); + vars_[name].reset(var); + return var; +} + +VarDescBind *BlockDescBind::Var(const std::string &name) const { + auto it = vars_.find(name); + PADDLE_ENFORCE(it != vars_.end(), + "Can not find variable %s in current block.", name); + return it->second.get(); +} + +bool BlockDescBind::HasVar(const std::string &name) const { + return vars_.find(name) != vars_.end(); +} + +std::vector BlockDescBind::AllVars() const { + std::vector res; + for (const auto &p : vars_) { + res.push_back(p.second.get()); + } + return res; +} + +OpDescBind *BlockDescBind::AppendOp() { + need_update_ = true; + ops_.emplace_back(new OpDescBind()); + return ops_.back().get(); +} + +OpDescBind *BlockDescBind::PrependOp() { + need_update_ = true; + ops_.emplace_front(new OpDescBind()); + return ops_.front().get(); +} + +std::vector BlockDescBind::AllOps() const { + std::vector res; + for (const auto &op : ops_) { + res.push_back(op.get()); + } + return res; +} + +void BlockDescBind::Sync() { + if (need_update_) { + auto &op_field = *this->desc_->mutable_ops(); + op_field.Clear(); + op_field.Reserve(static_cast(ops_.size())); + for (auto &op_desc : ops_) { + op_field.AddAllocated(op_desc->Proto()); + } + auto &var_field = *this->desc_->mutable_vars(); + var_field.Clear(); + var_field.Reserve(static_cast(vars_.size())); + for (auto &var_desc : vars_) { + var_field.AddAllocated(var_desc.second->Proto()); + } + need_update_ = false; + } +} + +BlockDescBind *BlockDescBind::ParentBlock() const { + if (this->desc_->parent_idx() == -1) { + return nullptr; + } + return prog_->Block(static_cast(this->desc_->parent_idx())); +} + +void OpDescBind::SetBlockAttr(const std::string &name, BlockDescBind &block) { + BlockDesc *desc = block.RawPtr(); + this->attrs_[name] = desc; +} +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/block_desc.h b/paddle/framework/block_desc.h new file mode 100644 index 0000000000000000000000000000000000000000..3437e89923da8de79eeaa88d0466cf7eb0b5926d --- /dev/null +++ b/paddle/framework/block_desc.h @@ -0,0 +1,82 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include +#include +#include +#include "paddle/framework/op_desc.h" +#include "paddle/framework/var_desc.h" +#include "paddle/platform/macros.h" + +namespace paddle { +namespace framework { + +class ProgramDescBind; + +// Each Protobuf Message, we provide a XXXBind class. In that class, we optimize +// read/write speed. Only when we want the protobuf message, the local changes +// will be synchronized (by `Sync` method). + +class BlockDescBind { + public: + friend std::vector> MakeBlockBackward( + ProgramDescBind &program_desc, int block_idx, + std::unordered_set &no_grad_vars); + + friend void AppendBackward( + ProgramDescBind &program_desc, + const std::unordered_set &no_grad_vars); + + BlockDescBind(ProgramDescBind *prog, BlockDesc *desc) + : prog_(prog), desc_(desc), need_update_(false) {} + + int32_t ID() const { return desc_->idx(); } + + int32_t Parent() const { return desc_->parent_idx(); } + + VarDescBind *NewVar(const std::string &name_bytes); + + VarDescBind *Var(const std::string &name_bytes) const; + + bool HasVar(const std::string &var_name) const; + + std::vector AllVars() const; + + BlockDescBind *ParentBlock() const; + + OpDescBind *AppendOp(); + + OpDescBind *PrependOp(); + + std::vector AllOps() const; + + void Sync(); + + BlockDesc *RawPtr() { return desc_; } + + private: + ProgramDescBind *prog_; // not_own + BlockDesc *desc_; // not_own + bool need_update_; + + std::deque> ops_; + std::unordered_map> vars_; + + DISABLE_COPY_AND_ASSIGN(BlockDescBind); +}; +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/data_type.h b/paddle/framework/data_type.h new file mode 100644 index 0000000000000000000000000000000000000000..649899d42572c9a22adca5337dcd56b0bcf42e7c --- /dev/null +++ b/paddle/framework/data_type.h @@ -0,0 +1,35 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once +#include +#include "paddle/framework/framework.pb.h" + +namespace paddle { +namespace framework { + +inline DataType ToDataType(std::type_index type) { + if (typeid(float).hash_code() == type.hash_code()) { + return DataType::FP32; + } else if (typeid(double).hash_code() == type.hash_code()) { + return DataType::FP64; + } else if (typeid(int).hash_code() == type.hash_code()) { + return DataType::INT32; + } else { + PADDLE_THROW("Not supported"); + } +} + +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/details/op_registry.h b/paddle/framework/details/op_registry.h new file mode 100644 index 0000000000000000000000000000000000000000..daa474e8c5a223589018720da29a5c3363b5934d --- /dev/null +++ b/paddle/framework/details/op_registry.h @@ -0,0 +1,109 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once + +#include "paddle/framework/grad_op_desc_maker.h" +#include "paddle/framework/op_info.h" +#include "paddle/framework/op_proto_maker.h" +#include "paddle/framework/operator.h" + +namespace paddle { +namespace framework { +namespace details { + +enum OpInfoFillType { + kOperator = 0, + kOpProtoAndCheckerMaker = 1, + kGradOpDescMaker = 2 +}; + +template +struct OpInfoFillTypeID { + static constexpr OpInfoFillType ID() { + return std::is_base_of::value + ? kOperator + : (std::is_base_of::value + ? kOpProtoAndCheckerMaker + : (std::is_base_of::value + ? kGradOpDescMaker + : static_cast(-1))); + } +}; + +template ::ID()> +struct OpInfoFiller; + +template +class OperatorRegistrarRecursive; + +template +class OperatorRegistrarRecursive { + public: + using T = typename std::tuple_element>::type; + OperatorRegistrarRecursive(const char* op_type, OpInfo* info) { + OpInfoFiller fill; + fill(op_type, info); + constexpr auto size = sizeof...(ARGS); + OperatorRegistrarRecursive reg(op_type, + info); + (void)(reg); + } +}; + +template +class OperatorRegistrarRecursive { + public: + OperatorRegistrarRecursive(const char* op_type, OpInfo* info) {} +}; + +template +struct OpInfoFiller { + void operator()(const char* op_type, OpInfo* info) const { + info->creator_ = [](const std::string& type, const VariableNameMap& inputs, + const VariableNameMap& outputs, + const AttributeMap& attrs) { + return new T(type, inputs, outputs, attrs); + }; + } +}; + +template +struct OpInfoFiller { + void operator()(const char* op_type, OpInfo* info) const { + info->proto_ = new OpProto; + info->checker_ = new OpAttrChecker(); + auto maker = T(info->proto_, info->checker_); + maker.Validate(); + info->proto_->set_type(op_type); + PADDLE_ENFORCE( + info->proto_->IsInitialized(), + "Fail to initialize %s's OpProto, because %s is not initialized", + op_type, info->proto_->InitializationErrorString()); + } +}; + +template +struct OpInfoFiller { + void operator()(const char* op_type, OpInfo* info) const { + info->grad_op_maker_ = [](const OpDescBind& fwd_op) { + T maker(fwd_op); + return maker(); + }; + } +}; +} // namespace details + +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/framework.proto b/paddle/framework/framework.proto index 951c7afbc14e2d9119169c1351d38ff0b67bdc5b..b7a63f9ba10b77acff516d75cf1be0d4eeda40d4 100644 --- a/paddle/framework/framework.proto +++ b/paddle/framework/framework.proto @@ -13,6 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ syntax = "proto2"; +option optimize_for = LITE_RUNTIME; package paddle.framework; enum AttrType { @@ -66,7 +67,6 @@ message OpProto { optional bool duplicable = 3 [ default = false ]; optional bool intermediate = 4 [ default = false ]; - optional bool not_in_gradient = 5 [ default = false ]; } // AttrProto describes the C++ type Attribute. @@ -106,6 +106,7 @@ message LoDTensorDesc { message VarDesc { required string name = 1; optional LoDTensorDesc lod_tensor = 2; + optional bool persistable = 3 [ default = false ]; } message BlockDesc { @@ -115,4 +116,7 @@ message BlockDesc { repeated OpDesc ops = 4; } +// Please refer to +// https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/program.md +// for more details. message ProgramDesc { repeated BlockDesc blocks = 1; } diff --git a/paddle/framework/grad_op_builder.cc b/paddle/framework/grad_op_builder.cc deleted file mode 100644 index b02a599a800668b22e7fe39a10fa6dc132e305bd..0000000000000000000000000000000000000000 --- a/paddle/framework/grad_op_builder.cc +++ /dev/null @@ -1,58 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOpArgType::OUT WARRANTIES OR CONDITIONS OF ANY KOpArgType::IND, either -express or implied. See the License for the specific language governing -permissions and limitations under the License. */ - -#include "paddle/framework/grad_op_builder.h" -#include "paddle/framework/op_registry.h" - -namespace paddle { -namespace framework { -enum class OpArgType { IN, OUT }; - -static void TransOpArg(const OperatorBase* src_op, const OpArgType& src_type, - bool is_grad, VariableNameMap* vars) { - const auto& src_inout = - src_type == OpArgType::IN ? src_op->Inputs() : src_op->Outputs(); - auto& dst_inout = *vars; - auto& proto = OpInfoMap::Instance().Get(src_op->Type()).Proto(); - const auto& src_arg_list = - src_type == OpArgType::IN ? proto.inputs() : proto.outputs(); - for (const auto& arg : src_arg_list) { - if (arg.not_in_gradient() && !is_grad) continue; - const std::string src_name = arg.name(); - std::string dst_name = is_grad ? GradVarName(src_name) : src_name; - dst_inout[dst_name].reserve(src_inout.at(src_name).size()); - for (auto& var_name : src_inout.at(src_name)) { - std::string s = is_grad ? GradVarName(var_name) : var_name; - dst_inout[dst_name].emplace_back(s); - } - } -} - -OperatorBase* BuildGradOp(const OperatorBase* op) { - auto& info = OpInfoMap::Instance().Get(op->Type()); - PADDLE_ENFORCE(info.HasGradientOp()); - - VariableNameMap inputs; - VariableNameMap outputs; - TransOpArg(op, OpArgType::IN, false, &inputs); // I - TransOpArg(op, OpArgType::OUT, false, &inputs); // O - TransOpArg(op, OpArgType::OUT, true, &inputs); // OG - TransOpArg(op, OpArgType::IN, true, &outputs); // IG - - auto& grad_info = OpInfoMap::Instance().Get(info.grad_op_type_); - return grad_info.Creator()(info.grad_op_type_, inputs, outputs, op->Attrs()); -} - -} // namespace framework -} // namespace paddle diff --git a/paddle/framework/grad_op_builder_test.cc b/paddle/framework/grad_op_builder_test.cc deleted file mode 100644 index 9e3ca563c6765637f8471d142d32cec447f0b977..0000000000000000000000000000000000000000 --- a/paddle/framework/grad_op_builder_test.cc +++ /dev/null @@ -1,122 +0,0 @@ -#include "paddle/framework/grad_op_builder.h" -#include -#include "paddle/framework/op_registry.h" -#include "paddle/framework/operator.h" - -USE_OP(add); - -namespace paddle { -namespace framework { - -class MutiInOutOpMaker : public OpProtoAndCheckerMaker { - public: - MutiInOutOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("In1", "a single input"); - AddInput("In2_mult", "a multiple input").AsDuplicable(); - AddInput("In3", "another single input"); - AddOutput("Out1", "a single output"); - AddOutput("Out2_mult", "a multiple output").AsDuplicable(); - AddComment("test op with multiple inputs and outputs"); - } -}; - -class IOIgnoredOpMaker : public OpProtoAndCheckerMaker { - public: - IOIgnoredOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("In1", "a single input"); - AddInput("In2_mult", "a multiple input").AsDuplicable().NotInGradient(); - AddInput("In3_mult", "another multiple input").AsDuplicable(); - AddOutput("Out1_mult", "a multiple output").AsDuplicable(); - AddOutput("Out2", "a single output").NotInGradient(); - AddComment("op with inputs and outputs ignored in gradient calculating"); - } -}; - -} // namespace framework -} // namespace paddle - -namespace f = paddle::framework; - -TEST(GradOpBuilder, AddTwo) { - std::shared_ptr add_op(f::OpRegistry::CreateOp( - "add", {{"X", {"x"}}, {"Y", {"y"}}}, {{"Out", {"out"}}}, {})); - std::shared_ptr grad_add_op = - f::OpRegistry::CreateGradOp(*add_op); - EXPECT_EQ(grad_add_op->Inputs().size(), 4UL); - EXPECT_EQ(grad_add_op->Outputs().size(), 2UL); - EXPECT_EQ(grad_add_op->Input("X"), "x"); - EXPECT_EQ(grad_add_op->Input("Y"), "y"); - EXPECT_EQ(grad_add_op->Input("Out"), "out"); - EXPECT_EQ(grad_add_op->Input(f::GradVarName("Out")), f::GradVarName("out")); - EXPECT_EQ(grad_add_op->Output(f::GradVarName("X")), f::GradVarName("x")); - EXPECT_EQ(grad_add_op->Output(f::GradVarName("Y")), f::GradVarName("y")); -} - -REGISTER_OP(mult_io, f::NOP, f::MutiInOutOpMaker, mult_io_grad, f::NOP); -REGISTER_OP(io_ignored, f::NOP, f::IOIgnoredOpMaker, io_ignored_grad, f::NOP); - -TEST(GradOpBuilder, MutiInOut) { - std::shared_ptr test_op(f::OpRegistry::CreateOp( - "mult_io", {{"In1", {"in1"}}, - {"In2_mult", {"in2_1", "in2_2", "in2_3"}}, - {"In3", {"in3"}}}, - {{"Out1", {"out1"}}, {"Out2_mult", {"out2_1", "out2_2"}}}, {})); - std::shared_ptr grad_test_op = - f::OpRegistry::CreateGradOp(*test_op); - - ASSERT_EQ(grad_test_op->Inputs().size(), 3UL + 2UL + 2UL); - EXPECT_EQ(grad_test_op->Input("In1"), "in1"); - EXPECT_EQ(grad_test_op->Inputs("In2_mult"), - std::vector({"in2_1", "in2_2", "in2_3"})); - EXPECT_EQ(grad_test_op->Input("In3"), "in3"); - EXPECT_EQ(grad_test_op->Input("Out1"), "out1"); - EXPECT_EQ(grad_test_op->Inputs("Out2_mult"), - std::vector({"out2_1", "out2_2"})); - EXPECT_EQ(grad_test_op->Input(f::GradVarName("Out1")), - f::GradVarName("out1")); - EXPECT_EQ(grad_test_op->Inputs(f::GradVarName("Out2_mult")), - std::vector( - {f::GradVarName("out2_1"), f::GradVarName("out2_2")})); - - ASSERT_EQ(grad_test_op->Outputs().size(), 3UL); - EXPECT_EQ(grad_test_op->Output(f::GradVarName("In1")), f::GradVarName("in1")); - EXPECT_EQ(grad_test_op->Outputs(f::GradVarName("In2_mult")), - std::vector({f::GradVarName("in2_1"), - f::GradVarName("in2_2"), - f::GradVarName("in2_3")})); - EXPECT_EQ(grad_test_op->Output(f::GradVarName("In3")), f::GradVarName("in3")); -} - -TEST(GradOpBuilder, IOIgnoredInGradient) { - std::shared_ptr test_op(f::OpRegistry::CreateOp( - "io_ignored", {{"In1", {"in1"}}, - {"In2_mult", {"in2_1", "in2_2"}}, - {"In3_mult", {"in3_1", "in3_2"}}}, - {{"Out1_mult", {"out1_1", "out1_2"}}, {"Out2", {"out2"}}}, {})); - std::shared_ptr grad_test_op = - f::OpRegistry::CreateGradOp(*test_op); - - // 'In2' and 'Out2' are ignored in gradient calculating - ASSERT_EQ(grad_test_op->Inputs().size(), 2UL + 1UL + 2UL); - EXPECT_EQ(grad_test_op->Input("In1"), "in1"); - EXPECT_EQ(grad_test_op->Inputs("In3_mult"), - std::vector({"in3_1", "in3_2"})); - EXPECT_EQ(grad_test_op->Inputs("Out1_mult"), - std::vector({"out1_1", "out1_2"})); - EXPECT_EQ(grad_test_op->Inputs(f::GradVarName("Out1_mult")), - std::vector( - {f::GradVarName("out1_1"), f::GradVarName("out1_2")})); - EXPECT_EQ(grad_test_op->Input(f::GradVarName("Out2")), - f::GradVarName("out2")); - - ASSERT_EQ(grad_test_op->Outputs().size(), 3UL); - EXPECT_EQ(grad_test_op->Output(f::GradVarName("In1")), f::GradVarName("in1")); - EXPECT_EQ(grad_test_op->Outputs(f::GradVarName("In2_mult")), - std::vector( - {f::GradVarName("in2_1"), f::GradVarName("in2_2")})); - EXPECT_EQ(grad_test_op->Outputs(f::GradVarName("In3_mult")), - std::vector( - {f::GradVarName("in3_1"), f::GradVarName("in3_2")})); -} diff --git a/paddle/framework/grad_op_desc_maker.h b/paddle/framework/grad_op_desc_maker.h new file mode 100644 index 0000000000000000000000000000000000000000..e9ae6e22060850fe229998d3b651d08a5ca2033a --- /dev/null +++ b/paddle/framework/grad_op_desc_maker.h @@ -0,0 +1,124 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once +#include "paddle/framework/op_desc.h" +#include "paddle/framework/operator.h" + +namespace paddle { +namespace framework { + +class GradOpDescMakerBase { + public: + explicit GradOpDescMakerBase(const OpDescBind& fwd_op) : fwd_op_(fwd_op) {} + + virtual ~GradOpDescMakerBase() = default; + virtual std::vector> operator()() const = 0; + + protected: + static std::vector ToGradNames( + const std::vector& var_names) { + std::vector ret_val; + ret_val.reserve(var_names.size()); + std::transform(var_names.begin(), var_names.end(), + std::back_inserter(ret_val), GradVarName); + return ret_val; + } + + std::vector InputGrad(const std::string& name) const { + return ToGradNames(fwd_op_.Input(name)); + } + + std::vector OutputGrad(const std::string& name) const { + return ToGradNames(fwd_op_.Output(name)); + } + + std::vector InputNames() const { + return this->fwd_op_.InputNames(); + } + + std::vector OutputNames() const { + return this->fwd_op_.OutputNames(); + } + + std::vector Input(const std::string& name) const { + return fwd_op_.Input(name); + } + + std::vector Output(const std::string& name) const { + return fwd_op_.Output(name); + } + + const std::unordered_map& Attrs() const { + return fwd_op_.GetAttrMap(); + } + + const Attribute& GetAttr(const std::string& name) const { + auto& map = fwd_op_.GetAttrMap(); + auto it = map.find(name); + PADDLE_ENFORCE(it != map.end(), "Cannot find attribute %s", name); + return it->second; + } + + std::string ForwardOpType() const { return this->fwd_op_.Type(); } + + private: + const OpDescBind& fwd_op_; +}; + +class SingleGradOpDescMaker : public GradOpDescMakerBase { + public: + using GradOpDescMakerBase::GradOpDescMakerBase; + + std::vector> operator()() const { + std::vector> retv; + retv.emplace_back(this->Apply()); + return retv; + } + + protected: + virtual std::unique_ptr Apply() const = 0; +}; + +class DefaultGradOpDescMaker : public SingleGradOpDescMaker { + public: + using SingleGradOpDescMaker::SingleGradOpDescMaker; + + protected: + virtual std::unique_ptr Apply() const { + auto* grad = new OpDescBind(); + grad->SetType(this->GradOpType()); + + for (auto& input_param : this->InputNames()) { + grad->SetInput(input_param, this->Input(input_param)); + grad->SetOutput(GradVarName(input_param), this->InputGrad(input_param)); + } + + for (auto& output_param : this->OutputNames()) { + grad->SetInput(output_param, this->Output(output_param)); + grad->SetInput(GradVarName(output_param), this->OutputGrad(output_param)); + } + + grad->SetAttrMap(this->Attrs()); + + return std::unique_ptr(grad); + } + + virtual std::string GradOpType() const { + return this->ForwardOpType() + "_grad"; + } +}; + +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/lod_tensor.cc b/paddle/framework/lod_tensor.cc index 3c349637cdbe59b2cf9a1ea28e7715f4181f9293..5b7badf89c1714331bae9fc8cf94c8da2c66dbad 100644 --- a/paddle/framework/lod_tensor.cc +++ b/paddle/framework/lod_tensor.cc @@ -72,6 +72,22 @@ bool operator==(const LoD& a, const LoD& b) { return true; } +size_t LoDTensor::NumElements(size_t level, size_t idx) const { + PADDLE_ENFORCE_LT(level, NumLevels()); + PADDLE_ENFORCE_LT(idx, NumElements(level)); + // the last level of LoD, just return number of records in Tensor + if (level == NumLevels() - 1) { + return lod_[level][idx + 1] - lod_[level][idx]; + } + // high level of LoD, and there is another lower level, return number of + // lower-level elements + auto tmp = SliceInLevel(lod_, level, idx, idx + 1); + PADDLE_ENFORCE_GE(tmp.size(), 2); + // there is a 0 as a placeholder stored in LoD, so the number of elements + // equals lod.size() - 1 + return tmp[1].size() - 1; +} + void LoDTensor::ShrinkLevels(size_t level_begin, size_t level_end) { auto new_lod = framework::SliceLevels(lod_, level_begin, level_end); lod_ = new_lod; diff --git a/paddle/framework/lod_tensor.h b/paddle/framework/lod_tensor.h index 82f58464264c6871b51251e0feae3d5ca076cd2b..4db36ee76609ac6360fe2fc7b4a366e0284d1016 100644 --- a/paddle/framework/lod_tensor.h +++ b/paddle/framework/lod_tensor.h @@ -15,7 +15,7 @@ #pragma once #include -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA #include #include #include @@ -29,7 +29,7 @@ namespace paddle { namespace framework { -#ifdef PADDLE_ONLY_CPU +#ifndef PADDLE_WITH_CUDA template using Vector = std::vector; #else @@ -38,6 +38,18 @@ using Vector = thrust::host_vector< T, thrust::system::cuda::experimental::pinned_allocator>; #endif +/* + * 3-level LoD stores + * + * 0 10 20 + * 0 5 10 15 20 + * 0 2 5 7 10 12 15 20 + * + * - in a level, each element indicates offset in the underlying Tensor + * - the first element should be 0 and that indicates that this sequence start + * from 0 + * - each sequence's begin and end(no-inclusive) is level[id, id+1] + */ using LoD = std::vector>; LoD SliceLevels(const LoD& in, size_t level_begin, size_t level_end); @@ -65,11 +77,8 @@ class LoDTensor : public Tensor { * Get a element from LoD. */ size_t lod_element(size_t level, size_t elem) const { - PADDLE_ENFORCE(level < NumLevels(), "level [%d] out of range [%d]", level, - NumLevels()); - PADDLE_ENFORCE(elem < NumElements(level), - "element begin [%d] out of range [%d]", elem, - NumElements(level)); + PADDLE_ENFORCE_LT(level, NumLevels()); + PADDLE_ENFORCE_LT(elem, NumElements(level)); return (lod_)[level][elem]; } @@ -82,12 +91,23 @@ class LoDTensor : public Tensor { * Number of elements in a level. */ size_t NumElements(size_t level = 0) const { - PADDLE_ENFORCE(level < NumLevels(), "level [%d] out of range [%d]", level, - NumLevels()); + PADDLE_ENFORCE_LT(level, NumLevels()); // the last offset is the end of last element return (lod_)[level].size() - 1; } + /* + * Number of lower-level elements. + * For example, a 2-level lod-tensor + * + * 0-th level | | + * 1-th level || ||| + * + * NumElements(0, 0) get 2 + * NumElements(0, 1) get 3 + */ + size_t NumElements(size_t level, size_t idx) const; + /* * Shrink levels[level_begin:level_end] */ diff --git a/paddle/framework/lod_tensor.md b/paddle/framework/lod_tensor.md index 07bbdf9416c432052b3222757a61ac4bfd70fe14..d147f1c4257eec14664301edab8d1fe2f128d2b0 100644 --- a/paddle/framework/lod_tensor.md +++ b/paddle/framework/lod_tensor.md @@ -1,147 +1,175 @@ # Design Doc: LoD (Level-of-Detail) Tensor -PaddlePaddle's RNN doesn't require that all instances have the same length. To do so, we introduce an extension to Tensor, namely, LoD Tensor. +Like other deep learning systems, PaddlePaddle supports training models from sequence data. Also, like other systems, PaddlePaddle represent a mini-batch of sequences as a Tensor. What is different is that PaddlePaddle doesn't require all sequences in a mini-batch to be of the same length. Thus no need for padding zeros. -## Challenge of Variable-length Inputs +| | TensorFlow | PaddlePaddle | +|-----------------------|------------|--------------| +| RNN | Support | Support | +| recursive RNN | Support | Support | +| padding zeros | Must | No need | +| blob data type | Tensor | LoDTensor | -People usually represent a mini-batch by a Tensor. For example, a mini-batch of 10 images, each of size 32x32, is a 10x32x32 Tensor. So a transformation, T, of all images can be a matrix multiplication of the 10xOx32-dimensional tensor T and the 10x32x32 Tensor. +PaddlePaddle achieves this flexibility by passing through a new data type, *LoD Tensor*, which is a Tensor attached with segmentation index known as *LoD*, between operators. The LoD index doesn't only segment a tensor, but also recursively segments sub-sequences. This document presents the design of LoD and LoDTensor. -Another example is that each mini-batch contains 32 sentences, where each word is a D-dimensional one-hot vector. If all sentences have the same length L, we can represent this mini-batch by a 32xLxD tensor. However, in most cases, sentences have variable lengths, and we will need an index data structure to record these variable lengths. -## LoD as a Solution +## The Challenge: Variable-length Sequences -### Mini-Batch of variable-length sentences +Most deep learning systems represent a mini-batch as a Tensor. For example, a mini-batch of 10 images, each of size 32x32, is a 10x32x32 Tensor. Another example is that each mini-batch contains N sentences, where each word is a D-dimensional one-hot vector. Suppose that all sentences have the same length L, we can represent this mini-batch by a NxLxD tensor. -Let's imagine a mini-batch of 3 variable lengths sentences, containing 3, 1, and 2 words respectively. We can represent it by a (3+1+2)xD tensor plus some index information: +Both examples show that the elements of sequences are usually of the same size. In the first example, all images are 32x32, and in the second one, all words are D-dimensional vectors. It doesn't make sense to allow variable-sized images, as that would require transformations like convolution to handle variable-sized Tensors. + +The real challenge is that in most cases, sentences have variable lengths, and we will need an index data structure to segment the tensor into sequences. Also, sequences might consist of sub-sequences. + + +## A Solution: The LoD Index + +To understand our solution, it is best to look at some examples. + +### A Mini-Batch of Sentences + +Let's imagine a mini-batch of 3 variable lengths sentences composed of 3, 1, and 2 words, respectively. We can represent the mini-batch by a (3+1+2)xD tensor plus some index information: ``` - 3 3 1 2 ||| | || ``` -Each `|` represents a D-dimensional word vectors. The number 3 on top indicate 3 sentences, and numbers 3, 1, and 2 on the second level represent the number of words in each sentence. +where each `|` represents a D-dimensional word vector. The numbers, 3, 1, and 2, form a 1-level LoD. + +### Recursive Sequences + +Let check another example of a 2-level LoD Tensor. Consider a mini-batch of three articles with 3, 1, and 2 sentences, and each sentence consists of a variable number of words: + +``` +3 1 2 +3 2 4 1 2 3 +||| || |||| | || ||| +``` -### Mini-Batch of variable-length videos +### A Mini-Batch of Videos -This approach generalizes to the case where elements are not words, but higher dimensional objects, like images. Suppose that a mini-batch contains videos of the same frame size 640x480. If a mini-batch contains 3 videos of 3, 1, and 2 frames respectively. The underlying tensor is of size (3+1+2)x640x480. The index information illustrates as: +LoD tensors generalize to the case where elements are higher dimensional objects, like images. Suppose that a mini-batch contains videos of the same frame size 640x480. Here is a mini-batch of 3 videos with 3, 1, and 2 frames, respectively. ``` - 3 3 1 2 口口口 口 口口 ``` -where each `口` represents an image. +The underlying tensor is of size (3+1+2)x640x480, and each `口` represents a 640x480 image. -### Mini-Batch of fixed-size images +### A Mini-Batch of Images -Let's get back to a typical example, image classification, where each mini-batch has M fixed-sized images. The LoD Tensor representation is +In traditional cases like a mini-batch with N fixed-sized images, the LoD Tensor representation is as ``` - M 1 1 1 1 1 口口口口 ... 口 ``` -The many 1's on the second level seem duplicated. For this particular case of 2 levels and the second level always have length 1, we can ignore the LoD index. - -### Design and summarization +In this case, we don't lose any information by ignoring the many 1's in the index and simply considering this LoD Tensor as a usual Tensor: -In summary, as long as that the essential elements (words or images) have the same size, we can represent mini-batches by a LoD Tensor: +``` +口口口口 ... 口 +``` -- The underlying tensor has size LxD1xD2x..., where D1xD2... is the size of the essential elements, and -- The first dimension size L has an additonal property -- a LoD index as a nested vector: +### Model Parameters - ```c++ - typedef std::vector> LoD; - ``` +A model parameter is just a usual Tensor, which, just like the above example, is a **0-level LoD Tensor**. -- The LoD index is not necessary when there are only two levels and all elements of the second level have length 1. -## Slicing of LoD Tensor +## The LoD Tensor -Consider that we have a network with three levels of RNN: the top level one handles articles, the second level one handles sentences, and the basic level one handles words. This network requires that mini-batches represented by 3 level LoD Tensor, for example, +Let us revisit above example of the 2-level LoD Tensor ``` - 3 3 1 2 3 2 4 1 2 3 ||| || |||| | || ||| ``` -To allow each level of RNN to handle its input, we define **the slicing of a LoD Tensor is defined as getting the j-th sequence on level i, or the -slice** +It is indeed a tree, where leaves are elementary sequences identified by **branches**. + +For example, the third sentence in above example is identified by branch <0,2>, where 0 indicates the first article with length 3, and 2 indicates the third sentence in this article with length 4. + +### The LoD Index -For example, the <2,1>-slice of above slice is +We can save the LoD index in the above example ``` -2 -|| +3 1 2 +3 2 4 1 2 3 ``` -and the <1,2>-slice of above example is +in a not-full 2D matrix: +```c++ +typedef std::vector > LoD; ``` -2 -2 3 -|| ||| -``` -Let's go on slicing this slice. Its <1,1>-slice is +where + +- `LoD.size()` is the number of levels, or the maximum length of branches, +- `LoD[i][j]` is the length of the j-th segment at the i-th level. + +## The Offset Representation + +To quickly access elementary sequences, we adopt an offset representation -- instead of saving the lengths, we save the beginning and ending elements of sequences. + +In the above example, we accumulate the length of elementary sequences: ``` -1 -1 -| +3 2 4 1 2 3 ``` -### The Slicing Algorithm +into offsets -The algorithm, with over-simplified data structure, is defined as +``` +0 3 5 9 10 12 15 + = = = = = = + 3 2+3 4+5 1+9 2+10 3+12 +``` -```c++ -typedef std::vector> LoD; +so we know that the first sentence is from word 0 to word 3, and the second sentence from work 3 to word 5. -struct LoDTensor { - LoD lod_; - float* tensor_; -}; +Similarly, the lengths in the top level LoD -LoDTensor Slice(const LoDTensor& lodt, int level, int sequence); +``` +3 1 2 ``` -Let us revisit the example above +are transformed into offsets of elements/words as follows: ``` - 3 -3 1 2 -3 2 4 1 2 3 -||| || |||| | || ||| +0 9 10 15 + = = = + 3+2+4 1+9 2+3+10 ``` -Suppose that we want to retrieve the <1,2>-slice +so we can tell that the first article is from word 0 to word 9, and the second article is from word 9 to word 10. + +The complete offset representation is as follows: ``` -2 -2 3 -|| ||| +0 9 10 15 +0 3 5 9 10 12 15 + ||| || |||| | || ||| ``` -we will need to find out the starting position of this slice by summing over all leaf nodes in `LoD` to the left of the slice, i.e., 3 + 2 + 4 + 1 = 10. +## Slicing of LoD Tensors + +When we use the above 2-level LoD Tensor as the input to a nested-RNN, we need to retrieve certain sequences. Here we define the sequence identified by branch as the **-slice**. -To avoid the traversal of the LoD tree at slicing time, we can do it at the construction time -- instead of saving the lengths of the next level in the LoD tree, we can save the starting offset of the next level. For example, above LoD Tensor can be transformed into +For example, the <2>-slice of above example is ``` - 0 -0 9 10 -0 3 5 9 10 12 -||| || |||| | || ||| +10 15 +10 12 15 + || ||| ``` -We don't really need the 0 on top, so the LoD Tensor could be +and the <2,0>-slice of above slice is ``` -0 9 10 -0 3 5 9 10 12 -||| || |||| | || ||| +10 12 + || ``` diff --git a/paddle/framework/lod_tensor_test.cc b/paddle/framework/lod_tensor_test.cc index 486b839738ec077545163bc47e6a97ef188c3c2f..44f09f584fb752d7003baa804979f3bb5cd9d651 100644 --- a/paddle/framework/lod_tensor_test.cc +++ b/paddle/framework/lod_tensor_test.cc @@ -56,6 +56,12 @@ TEST_F(LoDTensorTester, NumElements) { ASSERT_EQ(lod_tensor_.NumElements(2), 8UL); } +TEST_F(LoDTensorTester, NumElements2) { + ASSERT_EQ(lod_tensor_.NumElements(0, 0), 2UL); + ASSERT_EQ(lod_tensor_.NumElements(0, 1), 2UL); + ASSERT_EQ(lod_tensor_.NumElements(1, 1), 2UL); +} + TEST_F(LoDTensorTester, ShrinkLevels) { // slice 1 level for (size_t level = 0; level < 3UL; ++level) { @@ -65,7 +71,7 @@ TEST_F(LoDTensorTester, ShrinkLevels) { ASSERT_EQ(new_lod_tensor.NumElements(0), lod_tensor_.NumElements(level)); ASSERT_EQ(new_lod_tensor.data(), lod_tensor_.data()); } - // slice 2 level + // shrink 2 level for (size_t level = 0; level < 2UL; ++level) { LoDTensor new_lod_tensor = lod_tensor_; new_lod_tensor.ShrinkLevels(level, level + 2); diff --git a/paddle/framework/lod_tensor_test.cu b/paddle/framework/lod_tensor_test.cu index 97e69cdb2e5e1e64031c899f5e04020665485ba8..647d07536dd070bc37137fc01f683ec07ba7d6f4 100644 --- a/paddle/framework/lod_tensor_test.cu +++ b/paddle/framework/lod_tensor_test.cu @@ -36,8 +36,8 @@ TEST(LoDTensor, LoDInGPU) { lod_tensor.mutable_data(place); lod_tensor.set_lod(src_lod); - CHECK_EQ(lod_tensor.lod_element(0, 2), 4); - CHECK_EQ(lod_tensor.lod_element(0, 4), 8); + CHECK_EQ(lod_tensor.lod_element(0, 2), 4UL); + CHECK_EQ(lod_tensor.lod_element(0, 4), 8UL); auto lod = lod_tensor.lod(); diff --git a/paddle/framework/op_desc.cc b/paddle/framework/op_desc.cc new file mode 100644 index 0000000000000000000000000000000000000000..e7538b4af3429e566a439d5a0db8496efcd94969 --- /dev/null +++ b/paddle/framework/op_desc.cc @@ -0,0 +1,225 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/framework/op_desc.h" +#include +#include +#include "paddle/framework/block_desc.h" +#include "paddle/framework/operator.h" + +namespace paddle { +namespace framework { + +OpDescBind::OpDescBind(const std::string &type, const VariableNameMap &inputs, + const VariableNameMap &outputs, + const AttributeMap &attrs) { + op_desc_.set_type(type); + inputs_ = inputs; + outputs_ = outputs; + attrs_ = attrs; + need_update_ = true; +} + +OpDesc *OpDescBind::Proto() { + Sync(); + return &op_desc_; +} + +const std::vector &OpDescBind::Input( + const std::string &name) const { + auto it = inputs_.find(name); + PADDLE_ENFORCE(it != inputs_.end(), "Input %s cannot be found in Op %s", name, + Type()); + return it->second; +} + +std::vector OpDescBind::InputArgumentNames() const { + std::vector retv; + for (auto &ipt : this->inputs_) { + retv.insert(retv.end(), ipt.second.begin(), ipt.second.end()); + } + return retv; +} + +void OpDescBind::SetInput(const std::string ¶m_name, + const std::vector &args) { + need_update_ = true; + inputs_[param_name] = args; +} + +const std::vector &OpDescBind::Output( + const std::string &name) const { + auto it = outputs_.find(name); + PADDLE_ENFORCE(it != outputs_.end(), "Output %s cannot be found in Op %s", + name, Type()); + return it->second; +} + +std::vector OpDescBind::OutputArgumentNames() const { + std::vector retv; + for (auto &ipt : this->outputs_) { + retv.insert(retv.end(), ipt.second.begin(), ipt.second.end()); + } + return retv; +} + +void OpDescBind::SetOutput(const std::string ¶m_name, + const std::vector &args) { + need_update_ = true; + this->outputs_[param_name] = args; +} + +AttrType OpDescBind::GetAttrType(const std::string &name) const { + auto it = attrs_.find(name); + PADDLE_ENFORCE(it != attrs_.end(), "Attribute %s is not found", name); + return static_cast(it->second.which() - 1); +} + +std::vector OpDescBind::AttrNames() const { + std::vector retv; + retv.reserve(attrs_.size()); + for (auto &attr : attrs_) { + retv.push_back(attr.first); + } + return retv; +} + +void OpDescBind::SetAttr(const std::string &name, const Attribute &v) { + this->attrs_[name] = v; + need_update_ = true; +} + +void OpDescBind::SetAttrMap( + const std::unordered_map &attr_map) { + attrs_ = attr_map; + need_update_ = true; +} + +Attribute OpDescBind::GetAttr(const std::string &name) const { + auto it = attrs_.find(name); + PADDLE_ENFORCE(it != attrs_.end(), "Attribute %s is not found", name); + return it->second; +} + +int OpDescBind::GetBlockAttr(const std::string &name) const { + auto it = attrs_.find(name); + PADDLE_ENFORCE(it != attrs_.end(), "Attribute %s is not found", name); + return boost::get(it->second)->idx(); +} + +const std::unordered_map &OpDescBind::GetAttrMap() + const { + return attrs_; +} + +void OpDescBind::Rename(const std::string &old_name, + const std::string &new_name) { + for (auto &input : inputs_) { + std::replace(input.second.begin(), input.second.end(), old_name, new_name); + } + for (auto &output : outputs_) { + std::replace(output.second.begin(), output.second.end(), old_name, + new_name); + } + need_update_ = true; +} + +struct SetAttrDescVisitor : public boost::static_visitor { + explicit SetAttrDescVisitor(OpDesc::Attr *attr) : attr_(attr) {} + mutable OpDesc::Attr *attr_; + void operator()(int v) const { attr_->set_i(v); } + void operator()(float v) const { attr_->set_f(v); } + void operator()(const std::string &v) const { attr_->set_s(v); } + void operator()(bool b) const { attr_->set_b(b); } + + void operator()(const std::vector &v) const { + VectorToRepeated(v, attr_->mutable_ints()); + } + void operator()(const std::vector &v) const { + VectorToRepeated(v, attr_->mutable_floats()); + } + void operator()(const std::vector &v) const { + VectorToRepeated(v, attr_->mutable_strings()); + } + void operator()(const std::vector &v) const { + VectorToRepeated(v, attr_->mutable_bools()); + } + void operator()(BlockDesc *desc) const { attr_->set_block_idx(desc->idx()); } + void operator()(boost::blank) const { PADDLE_THROW("Unexpected branch"); } +}; + +void OpDescBind::Sync() { + if (need_update_) { + this->op_desc_.mutable_inputs()->Clear(); + for (auto &ipt : inputs_) { + auto *input = op_desc_.add_inputs(); + input->set_parameter(ipt.first); + VectorToRepeated(ipt.second, input->mutable_arguments()); + } + + this->op_desc_.mutable_outputs()->Clear(); + for (auto &opt : outputs_) { + auto *output = op_desc_.add_outputs(); + output->set_parameter(opt.first); + VectorToRepeated(opt.second, output->mutable_arguments()); + } + + this->op_desc_.mutable_attrs()->Clear(); + for (auto &attr : attrs_) { + auto *attr_desc = op_desc_.add_attrs(); + attr_desc->set_name(attr.first); + attr_desc->set_type( + static_cast(attr.second.which() - 1)); + SetAttrDescVisitor visitor(attr_desc); + boost::apply_visitor(visitor, attr.second); + } + + need_update_ = false; + } +} + +using InferShapeFuncMap = + std::unordered_map>; + +static InferShapeFuncMap &InferShapeFuncs() { + static InferShapeFuncMap *g_map = nullptr; + if (g_map == nullptr) { + g_map = new InferShapeFuncMap(); + auto &info_map = OpInfoMap::Instance(); + // all registered kernels + for (auto &pair : OperatorWithKernel::AllOpKernels()) { + auto &info = info_map.Get(pair.first); + // use empty type here to avoid runtime checks. + auto op = + static_cast(info.Creator()("", {}, {}, {})); + g_map->insert( + {pair.first, [op](InferShapeContext *ctx) { op->InferShape(ctx); }}); + } + } + return *g_map; +} + +void OpDescBind::InferShape(const BlockDescBind &block) const { + auto &funcs = InferShapeFuncs(); + auto it = funcs.find(this->Type()); + if (it == funcs.end()) { + PADDLE_THROW("Operator %s has not been registered", this->Type()); + } + CompileTimeInferShapeContext ctx(*this, block); + it->second(&ctx); +} + +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/op_desc.h b/paddle/framework/op_desc.h new file mode 100644 index 0000000000000000000000000000000000000000..81c4225041157ac600d1db73ef2363ebcd4abfc0 --- /dev/null +++ b/paddle/framework/op_desc.h @@ -0,0 +1,126 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include +#include "paddle/framework/attribute.h" +#include "paddle/framework/type_defs.h" +#include "paddle/framework/var_desc.h" + +namespace paddle { +namespace framework { + +class BlockDescBind; + +class OpDescBind { + public: + OpDescBind() {} + + OpDescBind(const std::string &type, const VariableNameMap &inputs, + const VariableNameMap &outputs, const AttributeMap &attrs); + + OpDesc *Proto(); + + std::string Type() const { return op_desc_.type(); } + + void SetType(const std::string &type) { op_desc_.set_type(type); } + + const std::vector &Input(const std::string &name) const; + + std::vector InputArgumentNames() const; + + void SetInput(const std::string ¶m_name, + const std::vector &args); + + const std::vector &Output(const std::string &name) const; + + std::vector OutputArgumentNames() const; + + void SetOutput(const std::string ¶m_name, + const std::vector &args); + + bool HasAttr(const std::string &name) const { + return attrs_.find(name) != attrs_.end(); + } + + AttrType GetAttrType(const std::string &name) const; + + std::vector AttrNames() const; + + void SetAttr(const std::string &name, const Attribute &v); + + void SetBlockAttr(const std::string &name, BlockDescBind &block); + + Attribute GetAttr(const std::string &name) const; + + int GetBlockAttr(const std::string &name) const; + + void Rename(const std::string &old_name, const std::string &new_name); + + // Only be used in C++ + const AttributeMap &GetAttrMap() const; + + // Only be used in C++ + void SetAttrMap(const AttributeMap &attr_map); + + std::vector InputNames() const { return MapKeys(inputs_); } + std::vector OutputNames() const { return MapKeys(outputs_); } + + void SetInputMap(const VariableNameMap &input) { + this->inputs_ = input; + this->need_update_ = true; + } + + void SetOutputMap(const VariableNameMap &output) { + this->outputs_ = output; + this->need_update_ = true; + } + + void Sync(); + + const VariableNameMap &Inputs() const { return inputs_; } + + const VariableNameMap &Outputs() const { return outputs_; } + + AttributeMap *MutableAttrMap() { + this->need_update_ = true; + return &this->attrs_; + } + + void InferShape(const BlockDescBind &block) const; + + private: + template + static std::vector MapKeys(const MapType &map) { + std::vector ret_val; + ret_val.reserve(map.size()); + std::transform( + map.begin(), map.end(), std::back_inserter(ret_val), + [](const typename MapType::value_type &pair) { return pair.first; }); + return ret_val; + } + + OpDesc op_desc_; + VariableNameMap inputs_; + VariableNameMap outputs_; + AttributeMap attrs_; + + // need_update_ indicate there some local changes not be synchronized. If + // local changes should be synchronized, need_update_ should be set to true. + bool need_update_{false}; +}; +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/op_info.h b/paddle/framework/op_info.h index b98d8f23a14cf6fbe787953ad16b5c9ab99222ad..c504f69e30bb899c183bd4281d2eadb50fd3b376 100644 --- a/paddle/framework/op_info.h +++ b/paddle/framework/op_info.h @@ -19,21 +19,18 @@ #include #include "paddle/framework/attribute.h" +#include "paddle/framework/op_desc.h" +#include "paddle/framework/type_defs.h" +#include "paddle/platform/macros.h" namespace paddle { namespace framework { -class OperatorBase; -using VariableNameMap = std::map>; - -using OpCreator = std::function; struct OpInfo { OpCreator creator_; - std::string grad_op_type_; - OpProto* proto_; - OpAttrChecker* checker_; + GradOpMakerFN grad_op_maker_; + OpProto* proto_{nullptr}; + OpAttrChecker* checker_{nullptr}; bool HasOpProtoAndChecker() const { return proto_ != nullptr && checker_ != nullptr; @@ -46,30 +43,25 @@ struct OpInfo { return *proto_; } - const OpAttrChecker& Checker() const { - PADDLE_ENFORCE_NOT_NULL(checker_, - "Operator Checker has not been registered"); - return *checker_; - } - const OpCreator& Creator() const { PADDLE_ENFORCE_NOT_NULL(creator_, "Operator Creator has not been registered"); return creator_; } - bool HasGradientOp() const { return !grad_op_type_.empty(); } + const GradOpMakerFN& GradOpMaker() const { + PADDLE_ENFORCE_NOT_NULL(grad_op_maker_, + "Operator GradOpMaker has not been registered."); + return grad_op_maker_; + } + + const OpAttrChecker* Checker() const { return checker_; } }; class OpInfoMap { public: static OpInfoMap& Instance(); - OpInfoMap(const OpInfoMap& o) = delete; - OpInfoMap(OpInfoMap&& o) = delete; - OpInfoMap& operator=(const OpInfoMap& o) = delete; - OpInfoMap& operator=(OpInfoMap&& o) = delete; - bool Has(const std::string& op_type) const { return map_.find(op_type) != map_.end(); } @@ -105,6 +97,8 @@ class OpInfoMap { private: OpInfoMap() = default; std::unordered_map map_; + + DISABLE_COPY_AND_ASSIGN(OpInfoMap); }; } // namespace framework diff --git a/paddle/framework/op_proto_maker.h b/paddle/framework/op_proto_maker.h index 4d55a37db9f0a3deac7b3489c8bc288ea41f4799..a134befd90a1eaeff6f6ea62f11412df63cdc394 100644 --- a/paddle/framework/op_proto_maker.h +++ b/paddle/framework/op_proto_maker.h @@ -44,11 +44,6 @@ class OpProtoAndCheckerMaker { var_->set_intermediate(true); return *this; } - - VariableBuilder& NotInGradient() { - var_->set_not_in_gradient(true); - return *this; - } }; VariableBuilder AddInput(const std::string& name, const std::string& comment); diff --git a/paddle/framework/op_proto_maker_test.cc b/paddle/framework/op_proto_maker_test.cc index b01e30f75371ca4aa63dae86ddfb966b1d4c7830..988a14cf4de8fdf052ca7e8c41bff0c05ba2daaa 100644 --- a/paddle/framework/op_proto_maker_test.cc +++ b/paddle/framework/op_proto_maker_test.cc @@ -48,4 +48,4 @@ TEST(ProtoMaker, DuplicatedInOut) { paddle::framework::OpAttrChecker op_checker; auto proto_maker = TestInOutProtoMaker(&op_proto, &op_checker); ASSERT_THROW(proto_maker.Validate(), paddle::platform::EnforceNotMet); -} \ No newline at end of file +} diff --git a/paddle/framework/op_registry.cc b/paddle/framework/op_registry.cc index b0e85dd49f97da4a7f889fde0b5f060954947be8..b118edae17430c8a4dd5c96a2a0c675766e08166 100644 --- a/paddle/framework/op_registry.cc +++ b/paddle/framework/op_registry.cc @@ -23,7 +23,9 @@ std::unique_ptr OpRegistry::CreateOp( const std::string& type, const VariableNameMap& inputs, const VariableNameMap& outputs, AttributeMap attrs) { auto& info = OpInfoMap::Instance().Get(type); - info.Checker().Check(attrs); + if (info.Checker() != nullptr) { + info.Checker()->Check(attrs); + } auto op = info.Creator()(type, inputs, outputs, attrs); return std::unique_ptr(op); } @@ -52,9 +54,20 @@ std::unique_ptr OpRegistry::CreateOp(const OpDesc& op_desc) { return CreateOp(op_desc.type(), inputs, outputs, attrs); } -std::unique_ptr OpRegistry::CreateGradOp(const OperatorBase& op) { - PADDLE_ENFORCE(!op.IsNetOp(), "Use framework::Backward to get backward ops"); - return std::unique_ptr(BuildGradOp(&op)); +std::unique_ptr OpRegistry::CreateOp(const OpDescBind& op_desc) { + return CreateOp(op_desc.Type(), op_desc.Inputs(), op_desc.Outputs(), + op_desc.GetAttrMap()); +} + +std::vector> OpRegistry::CreateGradOpDescs( + OpDescBind* op_desc) { + auto& info = OpInfoMap::Instance().Get(op_desc->Type()); + + if (info.Checker() != nullptr) { + info.Checker()->Check(*op_desc->MutableAttrMap()); + } + + return info.grad_op_maker_(*op_desc); } } // namespace framework diff --git a/paddle/framework/op_registry.h b/paddle/framework/op_registry.h index 90077d0192421f3678a049a723972fcb1e8d67af..5ca3af52a6909eeee21f647d0e60c7a690f90190 100644 --- a/paddle/framework/op_registry.h +++ b/paddle/framework/op_registry.h @@ -21,49 +21,54 @@ limitations under the License. */ #include #include #include "paddle/framework/attribute.h" +#include "paddle/framework/details/op_registry.h" #include "paddle/framework/framework.pb.h" -#include "paddle/framework/grad_op_builder.h" -#include "paddle/framework/op_info.h" -#include "paddle/framework/op_proto_maker.h" +#include "paddle/framework/grad_op_desc_maker.h" +#include "paddle/framework/op_desc.h" #include "paddle/framework/operator.h" #include "paddle/framework/scope.h" namespace paddle { namespace framework { +class Registrar { + public: + // In our design, various kinds of classes, e.g., operators and kernels, + // have their corresponding registry and registrar. The action of + // registration is in the constructor of a global registrar variable, which, + // however, are not used in the code that calls package framework, and would + // be removed from the generated binary file by the linker. To avoid such + // removal, we add Touch to all registrar classes and make USE_OP macros to + // call this method. So, as long as the callee code calls USE_OP, the global + // registrar variable won't be removed by the linker. + void Touch() {} +}; + +template +struct OperatorRegistrar : public Registrar { + explicit OperatorRegistrar(const char* op_type) : op_type(op_type) { + PADDLE_ENFORCE(!OpInfoMap::Instance().Has(op_type), + "'%s' is registered more than once.", op_type); + static_assert(sizeof...(ARGS) != 0, + "OperatorRegistrar should be invoked at least by OpClass"); + details::OperatorRegistrarRecursive<0, false, ARGS...>(op_type, &info); + OpInfoMap::Instance().Insert(op_type, info); + } + + const char* op_type; + + OpInfo info; +}; class OpRegistry { public: template static void RegisterOp(const std::string& op_type, const std::string& grad_op_type) { - PADDLE_ENFORCE(!OpInfoMap::Instance().Has(op_type), - "'%s' is registered more than once.", op_type); - OpInfo op_info; - op_info.creator_ = []( - const std::string& type, const VariableNameMap& inputs, - const VariableNameMap& outputs, const AttributeMap& attrs) { - return new OpType(type, inputs, outputs, attrs); - }; - op_info.grad_op_type_ = grad_op_type; - if (std::type_index(typeid(ProtoMakerType)) != - std::type_index(typeid(NOPMaker))) { - op_info.proto_ = new OpProto; - op_info.checker_ = new OpAttrChecker; - auto maker = ProtoMakerType(op_info.proto_, op_info.checker_); - maker.Validate(); - op_info.proto_->set_type(op_type); - PADDLE_ENFORCE( - op_info.proto_->IsInitialized(), - "Fail to initialize %s's OpProto, because %s is not initialized", - op_type, op_info.proto_->InitializationErrorString()); - } else { - op_info.proto_ = nullptr; - op_info.checker_ = nullptr; - } - OpInfoMap::Instance().Insert(op_type, op_info); + OperatorRegistrar reg(op_type.c_str()); + reg.info.grad_op_type_ = grad_op_type; // register gradient op if (!grad_op_type.empty()) { - RegisterOp(grad_op_type, ""); + OperatorRegistrar grad_reg(grad_op_type.c_str()); } } @@ -74,20 +79,10 @@ class OpRegistry { static std::unique_ptr CreateOp(const OpDesc& op_desc); - static std::unique_ptr CreateGradOp(const OperatorBase& op); -}; + static std::vector> CreateGradOpDescs( + OpDescBind* op_desc); -class Registrar { - public: - // In our design, various kinds of classes, e.g., operators and kernels, - // have their corresponding registry and registrar. The action of - // registration is in the constructor of a global registrar variable, which, - // however, are not used in the code that calls package framework, and would - // be removed from the generated binary file by the linker. To avoid such - // removal, we add Touch to all registrar classes and make USE_OP macros to - // call this method. So, as long as the callee code calls USE_OP, the global - // registrar variable won't be removed by the linker. - void Touch() {} + static std::unique_ptr CreateOp(const OpDescBind& op_desc); }; template @@ -100,13 +95,39 @@ class OpRegistrar : public Registrar { } }; -template +template +struct OpKernelRegistrarFunctor; + +template +struct OpKernelRegistrarFunctor { + using KERNEL_TYPE = + typename std::tuple_element>::type; + + void operator()(const char* op_type) const { + using T = typename KERNEL_TYPE::ELEMENT_TYPE; + OperatorWithKernel::OpKernelKey key(ToDataType(std::type_index(typeid(T))), + PlaceType()); + OperatorWithKernel::AllOpKernels()[op_type][key].reset(new KERNEL_TYPE); + + constexpr auto size = std::tuple_size>::value; + OpKernelRegistrarFunctor + func; + func(op_type); + } +}; + +template +struct OpKernelRegistrarFunctor { + void operator()(const char* op_type) const {} +}; + +// User can register many kernel in one place. The data type could be different. +template class OpKernelRegistrar : public Registrar { public: explicit OpKernelRegistrar(const char* op_type) { - OperatorWithKernel::OpKernelKey key; - key.place_ = PlaceType(); - OperatorWithKernel::AllOpKernels()[op_type][key].reset(new KernelType); + OpKernelRegistrarFunctor func; + func(op_type); } }; @@ -119,33 +140,41 @@ class OpKernelRegistrar : public Registrar { __test_global_namespace_##uniq_name##__>::value, \ msg) +#define REGISTER_OPERATOR(op_type, op_class, ...) \ + STATIC_ASSERT_GLOBAL_NAMESPACE( \ + __reg_op__##op_type, \ + "REGISTER_OPERATOR must be called in global namespace"); \ + class _OpClass_##op_type##_ : public op_class { \ + public: \ + DEFINE_OP_CLONE_METHOD(_OpClass_##op_type##_); \ + DEFINE_OP_CONSTRUCTOR(_OpClass_##op_type##_, op_class); \ + }; \ + static ::paddle::framework::OperatorRegistrar<_OpClass_##op_type##_, \ + ##__VA_ARGS__> \ + __op_registrar_##op_type##__(#op_type); \ + int TouchOpRegistrar_##op_type() { \ + __op_registrar_##op_type##__.Touch(); \ + return 0; \ + } + /** * Macro to register Operator. */ -#define REGISTER_OP(op_type, op_class, op_maker_class, grad_op_type, \ - grad_op_class) \ - STATIC_ASSERT_GLOBAL_NAMESPACE( \ - __reg_op__##op_type, "REGISTER_OP must be called in global namespace"); \ - class _OpClass_##op_type##_ : public op_class { \ - public: \ - DEFINE_OP_CLONE_METHOD(_OpClass_##op_type##_); \ - DEFINE_OP_CONSTRUCTOR(_OpClass_##op_type##_, op_class); \ - }; \ - class _OpGradClass_##op_type##_ : public grad_op_class { \ - public: \ - DEFINE_OP_CLONE_METHOD(_OpGradClass_##op_type##_); \ - DEFINE_OP_CONSTRUCTOR(_OpGradClass_##op_type##_, grad_op_class); \ - }; \ - static ::paddle::framework::OpRegistrar< \ - _OpClass_##op_type##_, op_maker_class, _OpGradClass_##op_type##_> \ - __op_registrar_##op_type##__(#op_type, #grad_op_type); \ - int TouchOpRegistrar_##op_type() { \ - __op_registrar_##op_type##__.Touch(); \ - return 0; \ - } +#define REGISTER_OP(op_type, op_class, op_maker_class, grad_op_type, \ + grad_op_class) \ + REGISTER_OPERATOR(grad_op_type, grad_op_class); \ + class _GradOpDescMaker_##grad_op_type##_ \ + : public ::paddle::framework::DefaultGradOpDescMaker { \ + using ::paddle::framework::DefaultGradOpDescMaker::DefaultGradOpDescMaker; \ + \ + protected: \ + virtual std::string GradOpType() const { return #grad_op_type; } \ + }; \ + REGISTER_OPERATOR(op_type, op_class, _GradOpDescMaker_##grad_op_type##_, \ + op_maker_class); #define REGISTER_OP_WITHOUT_GRADIENT(op_type, op_class, op_maker_class) \ - REGISTER_OP(op_type, op_class, op_maker_class, , ::paddle::framework::NOP) + REGISTER_OPERATOR(op_type, op_class, op_maker_class) /** * Macro to register OperatorKernel. @@ -192,7 +221,7 @@ class OpKernelRegistrar : public Registrar { // TODO(fengjiayi): The following macros // seems ugly, do we have better method? -#ifdef PADDLE_ONLY_CPU +#ifndef PADDLE_WITH_CUDA #define USE_OP_KERNEL(op_type) USE_OP_DEVICE_KERNEL(op_type, CPU) #else #define USE_OP_KERNEL(op_type) \ diff --git a/paddle/framework/op_registry_test.cc b/paddle/framework/op_registry_test.cc index e00c6e8d904508ec9985537fc703c7c61a14e0de..b860fe6cac773d1e85adecc43f5dfec42b6c7661 100644 --- a/paddle/framework/op_registry_test.cc +++ b/paddle/framework/op_registry_test.cc @@ -10,7 +10,6 @@ class CosineOp : public OperatorBase { using OperatorBase::OperatorBase; void Run(const Scope& scope, const platform::DeviceContext& dev_ctx) const override {} - void InferShape(const Scope& scope) const override {} }; class CosineOpProtoAndCheckerMaker : public OpProtoAndCheckerMaker { @@ -29,7 +28,6 @@ class CosineOpProtoAndCheckerMaker : public OpProtoAndCheckerMaker { class MyTestOp : public OperatorBase { public: using OperatorBase::OperatorBase; - void InferShape(const Scope& scope) const override {} void Run(const Scope& scope, const platform::DeviceContext& dev_ctx) const override {} }; @@ -174,4 +172,15 @@ TEST(OpRegistry, CustomChecker) { op->Run(scope, dev_ctx); int test_attr = op->Attr("test_attr"); ASSERT_EQ(test_attr, 4); -} \ No newline at end of file +} + +class CosineOpComplete : public paddle::framework::CosineOp { + public: + DEFINE_OP_CONSTRUCTOR(CosineOpComplete, paddle::framework::CosineOp); + DEFINE_OP_CLONE_METHOD(CosineOpComplete); +}; + +TEST(OperatorRegistrar, Test) { + using namespace paddle::framework; + OperatorRegistrar reg("cos"); +} diff --git a/paddle/framework/operator.cc b/paddle/framework/operator.cc index fcbfc3e4377edd0ea55c8d4328c325fa18663001..2fca816f353635d3bff184323755961ee82fbb68 100644 --- a/paddle/framework/operator.cc +++ b/paddle/framework/operator.cc @@ -14,7 +14,7 @@ limitations under the License. */ #include "paddle/framework/operator.h" #include -#include "paddle/framework/op_registry.h" +#include namespace paddle { namespace framework { @@ -22,17 +22,35 @@ namespace framework { template <> Eigen::DefaultDevice& ExecutionContext::GetEigenDevice< platform::CPUPlace, Eigen::DefaultDevice>() const { - return *device_context_.get_eigen_device(); + return *device_context_.GetEigenDevice(); } -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA template <> Eigen::GpuDevice& ExecutionContext::GetEigenDevice() const { - return *device_context_.get_eigen_device(); + return *device_context_.GetEigenDevice(); } #endif +const Tensor* GetTensorFromVar(const Variable* var) { + if (var->IsType()) { + return &var->Get(); + } + PADDLE_ENFORCE(var->IsType(), + "The Input must be LoDTensor or Tensor."); + return &var->Get(); +} + +Tensor* GetTensorFromVar(Variable* var) { + if (var->IsType()) { + return var->GetMutable(); + } + PADDLE_ENFORCE(var->IsType(), + "The Input must be LoDTensor or Tensor."); + return var->GetMutable(); +} + std::string OperatorBase::Input(const std::string& name) const { auto& ins = Inputs(name); PADDLE_ENFORCE_LE(ins.size(), 1UL, @@ -187,13 +205,13 @@ void OperatorBase::GenerateTemporaryNames() { } template <> -const Tensor* InferShapeContext::Input(const std::string& name) const { +const Tensor* ExecutionContext::Input(const std::string& name) const { auto* var = InputVar(name); return var == nullptr ? nullptr : GetTensorFromVar(var); } template <> -const std::vector InferShapeContext::MultiInput( +const std::vector ExecutionContext::MultiInput( const std::string& name) const { auto names = op().Inputs(name); std::vector res; @@ -207,13 +225,13 @@ const std::vector InferShapeContext::MultiInput( } template <> -Tensor* InferShapeContext::Output(const std::string& name) const { +Tensor* ExecutionContext::Output(const std::string& name) const { auto var = OutputVar(name); return var == nullptr ? nullptr : var->GetMutable(); } template <> -std::vector InferShapeContext::MultiOutput( +std::vector ExecutionContext::MultiOutput( const std::string& name) const { auto names = op().Outputs(name); std::vector res; @@ -227,5 +245,12 @@ std::vector InferShapeContext::MultiOutput( return res; } +std::ostream& operator<<(std::ostream& os, + const OperatorWithKernel::OpKernelKey& kernel_key) { + os << "place[" << kernel_key.place_ << "]:data_type[" << kernel_key.data_type_ + << "]"; + return os; +} + } // namespace framework } // namespace paddle diff --git a/paddle/framework/operator.h b/paddle/framework/operator.h index 2d6d5510ef6dc83f1a016be6ff123f0b9bcaf230..1e9ace99876e00b9f733da3e66f37dc0dc2f8cad 100644 --- a/paddle/framework/operator.h +++ b/paddle/framework/operator.h @@ -15,15 +15,19 @@ limitations under the License. */ #pragma once #include +#include #include #include #include #include "op_info.h" #include "paddle/framework/attribute.h" +#include "paddle/framework/block_desc.h" +#include "paddle/framework/data_type.h" #include "paddle/framework/framework.pb.h" #include "paddle/framework/lod_tensor.h" #include "paddle/framework/scope.h" +#include "paddle/framework/shape_inference.h" #include "paddle/framework/tensor.h" #include "paddle/platform/device_context.h" #include "paddle/platform/place.h" @@ -53,9 +57,11 @@ inline std::string GradVarName(const std::string& var_name) { } class OperatorBase; -class InferShapeContext; class ExecutionContext; +extern const Tensor* GetTensorFromVar(const Variable* var); +extern Tensor* GetTensorFromVar(Variable* var); + /** * OperatorBase has the basic element that Net will call to do computation. * Only CreateOperator from OpRegistry will new Operator directly. User @@ -78,10 +84,6 @@ class OperatorBase { virtual std::string DebugString() const; - /// InferShape infer the size of Variables used by this Operator with - /// information inside scope - virtual void InferShape(const Scope& scope) const = 0; - /// Net will call this function to Run an op. virtual void Run(const Scope& scope, const platform::DeviceContext& dev_ctx) const = 0; @@ -159,7 +161,6 @@ class OperatorBase { class NOP : public OperatorBase { public: using OperatorBase::OperatorBase; - void InferShape(const Scope& scope) const override {} void Run(const Scope& scope, const platform::DeviceContext& dev_ctx) const override {} std::unique_ptr Clone() const override { @@ -167,10 +168,11 @@ class NOP : public OperatorBase { } }; -class InferShapeContext { +class ExecutionContext { public: - InferShapeContext(const OperatorBase& op, const Scope& scope) - : op_(op), scope_(scope) {} + ExecutionContext(const OperatorBase& op, const Scope& scope, + const platform::DeviceContext& device_context) + : op_(op), scope_(scope), device_context_(device_context) {} const OperatorBase& op() const { return op_; } @@ -262,15 +264,6 @@ class InferShapeContext { return res; } - const Tensor* GetTensorFromVar(const Variable* var) const { - if (var->IsType()) { - return &var->Get(); - } - PADDLE_ENFORCE(var->IsType(), - "The Input(%s) must be LoDTensor or Tensor."); - return &var->Get(); - } - void ShareLoD(const std::string& in, const std::string& out, size_t i = 0, size_t j = 0) const { PADDLE_ENFORCE_LT(i, InputSize(in)); @@ -285,62 +278,234 @@ class InferShapeContext { out_tensor->set_lod(in_tensor.lod()); } + template ::EigenDeviceType> + DeviceType& GetEigenDevice() const; + + platform::Place GetPlace() const { return device_context_.GetPlace(); } + + const platform::DeviceContext& device_context() const { + return device_context_; + } + private: const OperatorBase& op_; const Scope& scope_; + const platform::DeviceContext& device_context_; }; template <> -const Tensor* InferShapeContext::Input(const std::string& name) const; +const Tensor* ExecutionContext::Input(const std::string& name) const; template <> -const std::vector InferShapeContext::MultiInput( +const std::vector ExecutionContext::MultiInput( const std::string& name) const; template <> -Tensor* InferShapeContext::Output(const std::string& name) const; +Tensor* ExecutionContext::Output(const std::string& name) const; template <> -std::vector InferShapeContext::MultiOutput( +std::vector ExecutionContext::MultiOutput( const std::string& name) const; -template -struct EigenDeviceConverter; +class CompileTimeInferShapeContext : public InferShapeContext { + public: + CompileTimeInferShapeContext(const OpDescBind& op, const BlockDescBind& block) + : op_(op), block_(block) {} + + bool HasInput(const std::string& name) const override { + const std::vector& input_names = op_.Input(name); + auto length = input_names.size(); + PADDLE_ENFORCE_EQ(length, 1UL, + "Input(%s) should have only one value, " + "but it have %d now", + name, length); + return block_.HasVar(input_names[0]); + } -template <> -struct EigenDeviceConverter { - using EigenDeviceType = Eigen::DefaultDevice; -}; + bool HasOutput(const std::string& name) const override { + const std::vector& output_names = op_.Output(name); + auto length = output_names.size(); + PADDLE_ENFORCE_EQ(length, 1UL, + "Output(%s) should have only one value, " + "but it have %d now", + name, length); + return block_.HasVar(output_names[0]); + } -#ifndef PADDLE_ONLY_CPU -template <> -struct EigenDeviceConverter { - using EigenDeviceType = Eigen::GpuDevice; + bool HasInputs(const std::string& name) const override { + const std::vector& input_names = op_.Input(name); + PADDLE_ENFORCE(!input_names.empty(), "Inputs(%s) length is 0", name); + for (auto& input : input_names) { + if (!block_.HasVar(input)) return false; + } + return true; + } + + bool HasOutputs(const std::string& name) const override { + const std::vector& output_names = op_.Output(name); + PADDLE_ENFORCE(!output_names.empty(), "Inputs(%s) length is 0", name); + for (auto& output : output_names) { + if (!block_.HasVar(output)) return false; + } + return true; + } + + DDim GetInputDim(const std::string& name) const override { + std::vector ddims = GetInputsDim(name); + auto length = ddims.size(); + PADDLE_ENFORCE_EQ(length, 1UL, + "Input(%s) should have 1 value, " + "but it has %d now", + name, length); + return ddims[0]; + } + + void SetInputDim(const std::string& name, const DDim& dim) override { + SetInputsDim(name, {dim}); + } + + DDim GetOutputDim(const std::string& name) const override { + std::vector ddims = GetOutputsDim(name); + auto length = ddims.size(); + PADDLE_ENFORCE_EQ(length, 1UL, + "Output(%s) should have 1 value, " + "but it has %d now", + name, length); + return ddims[0]; + } + + void SetOutputDim(const std::string& name, const DDim& dim) override { + SetOutputsDim(name, {dim}); + } + + AttrReader Attrs() const override { return AttrReader(op_.GetAttrMap()); } + + const std::vector& Inputs( + const std::string& name) const override { + return op_.Input(name); + } + + const std::vector& Outputs( + const std::string& name) const override { + return op_.Output(name); + } + + private: + DDim GetDim(const std::string& name) const override { + return framework::make_ddim(block_.Var(name)->Shape()); + } + + void SetDim(const std::string& name, const DDim& dim) override { + block_.Var(name)->SetShape(framework::vectorize(dim)); + } + + const OpDescBind& op_; + const BlockDescBind& block_; }; -#endif -class ExecutionContext : public InferShapeContext { +class RuntimeInferShapeContext : public InferShapeContext { public: - ExecutionContext(const OperatorBase& op, const Scope& scope, - const platform::DeviceContext& device_context) - : InferShapeContext(op, scope), device_context_(device_context) {} + RuntimeInferShapeContext(const OperatorBase& op, const Scope& scope) + : op_(op), scope_(scope) {} - template ::EigenDeviceType> - DeviceType& GetEigenDevice() const; + bool HasInput(const std::string& name) const override { + auto ipt = op_.Input(name); + auto* var = ipt == kEmptyVarName ? nullptr : scope_.FindVar(ipt); + return var != nullptr; + } - platform::Place GetPlace() const { return device_context_.GetPlace(); } + bool HasOutput(const std::string& name) const override { + auto ipt = op_.Output(name); + auto* var = ipt == kEmptyVarName ? nullptr : scope_.FindVar(ipt); + return var != nullptr; + } - const platform::DeviceContext& device_context() const { - return device_context_; + bool HasInputs(const std::string& name) const override { + auto inputs = op_.Inputs(name); + if (inputs.empty()) { + return false; + } + for (auto& input : inputs) { + if (scope_.FindVar(input) == nullptr) { + return false; + } + } + return true; + } + + bool HasOutputs(const std::string& name) const override { + auto outputs = op_.Outputs(name); + if (outputs.empty()) { + return false; + } + for (auto& output : outputs) { + if (scope_.FindVar(output) == nullptr) { + return false; + } + } + return true; + } + + DDim GetInputDim(const std::string& name) const override { + return GetDim(op_.Input(name)); + } + + void SetInputDim(const std::string& name, const DDim& dim) override { + SetDim(op_.Input(name), dim); + } + + DDim GetOutputDim(const std::string& name) const override { + return GetDim(op_.Output(name)); + } + + void SetOutputDim(const std::string& name, const DDim& dim) override { + SetDim(op_.Output(name), dim); + } + + AttrReader Attrs() const override { return AttrReader(op_.Attrs()); } + + const std::vector& Inputs( + const std::string& name) const override { + return op_.Inputs(name); + } + + const std::vector& Outputs( + const std::string& name) const override { + return op_.Outputs(name); } private: - const platform::DeviceContext& device_context_; + template + Tensor* GetTensor(const std::string& name) const { + Tensor* t = nullptr; + auto* var = scope_.FindVar(name); + if (!var->IsType() && !var->IsType()) { + if (Allocate) { + t = var->GetMutable(); + } else { + PADDLE_THROW("Variable(%s) should be tensor", name); + } + } else { + t = GetTensorFromVar(scope_.FindVar(name)); + } + return t; + } + + DDim GetDim(const std::string& name) const override { + return GetTensor(name)->dims(); + } + + void SetDim(const std::string& name, const DDim& dim) override { + GetTensor(name)->Resize(dim); + } + + const OperatorBase& op_; + const Scope& scope_; }; -class OpKernel { +class OpKernelBase { public: /** * ExecutionContext is the only parameter of Kernel Run function. @@ -351,46 +516,77 @@ class OpKernel { virtual void Compute(const ExecutionContext& context) const = 0; - virtual ~OpKernel() {} + virtual ~OpKernelBase() = default; +}; + +template +class OpKernel : public OpKernelBase { + public: + using ELEMENT_TYPE = T; }; class OperatorWithKernel : public OperatorBase { public: struct OpKernelKey { platform::Place place_; + DataType data_type_; - OpKernelKey() = default; - explicit OpKernelKey(const platform::DeviceContext& dev_ctx) { - place_ = dev_ctx.GetPlace(); - } + OpKernelKey(DataType data_type, platform::Place place) + : place_(place), data_type_(data_type) {} + + OpKernelKey(DataType data_type, const platform::DeviceContext& dev_ctx) + : place_(dev_ctx.GetPlace()), data_type_(data_type) {} bool operator==(const OpKernelKey& o) const { - return platform::places_are_same_class(place_, o.place_); + return platform::places_are_same_class(place_, o.place_) && + data_type_ == o.data_type_; } }; struct OpKernelHash { - std::hash hash_; + std::hash hash_; size_t operator()(const OpKernelKey& key) const { - return hash_(platform::is_gpu_place(key.place_)); + int place = key.place_.which(); + int data_type = static_cast(key.data_type_); + int pre_hash = data_type << NUM_PLACE_TYPE_LIMIT_IN_BIT | + (place & ((1 << NUM_PLACE_TYPE_LIMIT_IN_BIT) - 1)); + return hash_(pre_hash); } }; using OpKernelMap = - std::unordered_map, OpKernelHash>; + std::unordered_map, + OpKernelHash>; OperatorWithKernel(const std::string& type, const VariableNameMap& inputs, const VariableNameMap& outputs, const AttributeMap& attrs) : OperatorBase(type, inputs, outputs, attrs) {} - void InferShape(const Scope& scope) const override { - InferShape(InferShapeContext(*this, scope)); - } - void Run(const Scope& scope, const platform::DeviceContext& dev_ctx) const final { - auto& opKernel = AllOpKernels().at(type_).at(OpKernelKey(dev_ctx)); - opKernel->Compute(ExecutionContext(*this, scope, dev_ctx)); + RuntimeInferShapeContext infer_shape_ctx(*this, scope); + this->InferShape(&infer_shape_ctx); + + ExecutionContext ctx(*this, scope, dev_ctx); + + // check if op[type] has kernel registered. + auto& all_op_kernels = AllOpKernels(); + auto kernels_iter = all_op_kernels.find(type_); + if (kernels_iter == all_op_kernels.end()) { + PADDLE_THROW("op[%s] has no kernel", type_); + } + + // check if op[type] have kernel for kernel_key + OpKernelMap& kernels = kernels_iter->second; + auto kernel_key = OpKernelKey(IndicateDataType(ctx), dev_ctx); + auto kernel_iter = kernels.find(kernel_key); + + if (kernel_iter == kernels.end()) { + PADDLE_THROW("op[%s] has no kernel with kernel_key[%s]", type_, + kernel_key); + } + + kernel_iter->second->Compute(ctx); } static std::unordered_map& @@ -400,14 +596,47 @@ class OperatorWithKernel : public OperatorBase { } bool SupportGPU() const override { - OperatorWithKernel::OpKernelKey key; - key.place_ = platform::GPUPlace(); - return OperatorWithKernel::AllOpKernels().at(type_).count(key) != 0; + auto& op_kernels = OperatorWithKernel::AllOpKernels().at(type_); + return std::any_of(op_kernels.begin(), op_kernels.end(), + [](OpKernelMap::const_reference kern_pair) { + return platform::is_gpu_place(kern_pair.first.place_); + }); } + virtual void InferShape(InferShapeContext* ctx) const = 0; + protected: - virtual void InferShape(const InferShapeContext& ctx) const = 0; + // indicate kernel DataType by input data. Defaultly all input data must be + // same. + virtual DataType IndicateDataType(const ExecutionContext& ctx) const { + auto& scope = ctx.scope(); + int data_type = -1; + for (auto& input : this->inputs_) { + for (auto& ipt_name : input.second) { + auto* var = scope.FindVar(ipt_name); + if (var != nullptr) { + const Tensor* t = nullptr; + if (var->IsType()) { + t = &var->Get(); + } else if (var->IsType()) { + t = &var->Get(); + } + if (t != nullptr) { + int tmp = static_cast(ToDataType(t->type())); + PADDLE_ENFORCE(tmp == data_type || data_type == -1, + "DataType of Paddle Op must be same."); + data_type = tmp; + } + } + } + } + PADDLE_ENFORCE(data_type != -1, "DataType should be indicated by input"); + return static_cast(data_type); + } }; +std::ostream& operator<<(std::ostream& os, + const OperatorWithKernel::OpKernelKey& kernel_key); + } // namespace framework } // namespace paddle diff --git a/paddle/framework/operator_test.cc b/paddle/framework/operator_test.cc index 0beab0fac5b94c78121261d2661a6f969289afc4..a02f4668bca2360995cc05206f7f97e027db0907 100644 --- a/paddle/framework/operator_test.cc +++ b/paddle/framework/operator_test.cc @@ -14,6 +14,7 @@ limitations under the License. */ #include "paddle/framework/operator.h" #include "gtest/gtest.h" +#include "paddle/framework/op_info.h" #include "paddle/framework/op_registry.h" namespace paddle { @@ -26,7 +27,6 @@ class OpWithoutKernelTest : public OperatorBase { OpWithoutKernelTest(const std::string& type, const VariableNameMap& inputs, const VariableNameMap& outputs, const AttributeMap& attrs) : OperatorBase(type, inputs, outputs, attrs), x(1) {} - void InferShape(const Scope& scope) const override {} void Run(const Scope& scope, const platform::DeviceContext& dev_ctx) const override { ++op_run_num; @@ -86,7 +86,6 @@ TEST(OperatorBase, all) { auto op = paddle::framework::OpRegistry::CreateOp(op_desc); scope.NewVar("OUT1"); ASSERT_EQ(paddle::framework::op_run_num, 0); - op->InferShape(scope); op->Run(scope, device_context); ASSERT_EQ(paddle::framework::op_run_num, 1); } @@ -114,11 +113,14 @@ class OpWithKernelTest : public OperatorWithKernel { using OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext& ctx) const override {} + void InferShape(framework::InferShapeContext* ctx) const override {} + DataType IndicateDataType(const ExecutionContext& ctx) const override { + return DataType::FP32; + } }; template -class CPUKernelTest : public OpKernel { +class CPUKernelTest : public OpKernel { public: void Compute(const ExecutionContext& ctx) const { std::cout << "this is cpu kernel" << std::endl; @@ -145,7 +147,7 @@ class OpKernelTestMultiInputsProtoAndCheckerMaker } }; -class CPUKernalMultiInputsTest : public OpKernel { +class CPUKernalMultiInputsTest : public OpKernel { public: void Compute(const ExecutionContext& ctx) const { auto xs = ctx.op().Inputs("xs"); @@ -254,7 +256,6 @@ class OperatorClone : public paddle::framework::OperatorBase { const paddle::framework::VariableNameMap& outputs, const paddle::framework::AttributeMap& attrs) : OperatorBase(type, inputs, outputs, attrs) {} - void InferShape(const paddle::framework::Scope& scope) const override {} void Run(const paddle::framework::Scope& scope, const paddle::platform::DeviceContext& dev_ctx) const override {} }; diff --git a/paddle/framework/program_desc.cc b/paddle/framework/program_desc.cc new file mode 100644 index 0000000000000000000000000000000000000000..e89f9a46d587b6378aa3be92306c5680093e1926 --- /dev/null +++ b/paddle/framework/program_desc.cc @@ -0,0 +1,60 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/framework/program_desc.h" +#include "paddle/framework/block_desc.h" + +namespace paddle { +namespace framework { + +using ProgDescMap = + std::unordered_map>; +static ProgDescMap *g_bind_map = nullptr; + +ProgramDescBind &ProgramDescBind::Instance(ProgramDesc *prog) { + if (g_bind_map == nullptr) { + g_bind_map = new ProgDescMap(); + } + auto &map = *g_bind_map; + auto &ptr = map[prog]; + + if (ptr == nullptr) { + ptr.reset(new ProgramDescBind(prog)); + } + return *ptr; +} + +BlockDescBind *ProgramDescBind::AppendBlock(const BlockDescBind &parent) { + auto *b = prog_->add_blocks(); + b->set_parent_idx(parent.ID()); + b->set_idx(prog_->blocks_size() - 1); + blocks_.emplace_back(new BlockDescBind(this, b)); + return blocks_.back().get(); +} + +ProgramDesc *ProgramDescBind::Proto() { + for (auto &block : blocks_) { + block->Sync(); + } + return prog_; +} + +ProgramDescBind::ProgramDescBind(ProgramDesc *prog) { + prog_ = prog; + for (auto &block : *prog->mutable_blocks()) { + blocks_.emplace_back(new BlockDescBind(this, &block)); + } +} +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/program_desc.h b/paddle/framework/program_desc.h new file mode 100644 index 0000000000000000000000000000000000000000..f29b1c54e7160ac477229f64e5471939131a2d8f --- /dev/null +++ b/paddle/framework/program_desc.h @@ -0,0 +1,50 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include +#include "paddle/framework/framework.pb.h" +#include "paddle/platform/macros.h" + +namespace paddle { +namespace framework { + +class BlockDescBind; + +class ProgramDescBind { + public: + static ProgramDescBind &Instance(ProgramDesc *prog); + + BlockDescBind *AppendBlock(const BlockDescBind &parent); + + BlockDescBind *Block(size_t idx) { return blocks_[idx].get(); } + + size_t Size() const { return blocks_.size(); } + + ProgramDesc *Proto(); + + private: + explicit ProgramDescBind(ProgramDesc *prog); + + // Not owned + ProgramDesc *prog_; + + std::vector> blocks_; + + DISABLE_COPY_AND_ASSIGN(ProgramDescBind); +}; +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/scope.h b/paddle/framework/scope.h index c93b03e48130afe9568089b6a7586c4185d1d5b4..7047f0d55e9844aec19892631fe4b5b387bf89ca 100644 --- a/paddle/framework/scope.h +++ b/paddle/framework/scope.h @@ -19,6 +19,7 @@ limitations under the License. */ #include #include "paddle/framework/variable.h" +#include "paddle/platform/macros.h" namespace paddle { namespace framework { @@ -38,11 +39,6 @@ class Scope { Scope() {} ~Scope(); - // Disable Copy, Assign, Move. - Scope(const Scope& other) = delete; - Scope& operator=(const Scope& other) = delete; - Scope(Scope&& other) = delete; - /// Create a sub-scope. Returns a reference other than a pointer so /// to prevent from manual deletion. /// Mark it to const because that new kid scope cannot change parent scope. @@ -73,6 +69,8 @@ class Scope { std::unordered_map vars_; mutable std::list kids_; Scope const* parent_{nullptr}; + + DISABLE_COPY_AND_ASSIGN(Scope); }; } // namespace framework diff --git a/paddle/framework/shape_inference.h b/paddle/framework/shape_inference.h new file mode 100644 index 0000000000000000000000000000000000000000..64aab16ae54d34fd614add348c7c420b4a8f771d --- /dev/null +++ b/paddle/framework/shape_inference.h @@ -0,0 +1,89 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "paddle/framework/ddim.h" + +namespace paddle { +namespace framework { + +// TODO(longfei): Once after both CompileTimeInferShapeContext and +// RuntimeInferShapeContext get merged, we can rename InferShapeContext into +// InferShapeContext so to replace the current InferShapeContext. +class InferShapeContext { + public: + virtual ~InferShapeContext() {} + virtual bool HasInput(const std::string &name) const = 0; + virtual bool HasOutput(const std::string &name) const = 0; + + virtual bool HasInputs(const std::string &name) const = 0; + virtual bool HasOutputs(const std::string &name) const = 0; + + virtual framework::DDim GetInputDim(const std::string &name) const = 0; + std::vector GetInputsDim(const std::string &name) const { + const std::vector &names = Inputs(name); + return GetDims(names); + } + virtual void SetInputDim(const std::string &name, + const framework::DDim &dim) = 0; + void SetInputsDim(const std::string &name, + const std::vector &dims) { + auto &names = Inputs(name); + SetDims(names, dims); + } + virtual framework::DDim GetOutputDim(const std::string &name) const = 0; + std::vector GetOutputsDim(const std::string &name) const { + const std::vector &names = Outputs(name); + return GetDims(names); + } + virtual void SetOutputDim(const std::string &name, const DDim &dim) = 0; + void SetOutputsDim(const std::string &name, + const std::vector &dims) { + auto &names = Outputs(name); + SetDims(names, dims); + } + virtual AttrReader Attrs() const = 0; + virtual const std::vector &Inputs( + const std::string &name) const = 0; + virtual const std::vector &Outputs( + const std::string &name) const = 0; + // TODO(qiao) implement this function + void ShareLoD(const std::string &in, const std::string &out, size_t i = 0, + size_t j = 0) const {} + + protected: + virtual framework::DDim GetDim(const std::string &name) const = 0; + virtual void SetDim(const std::string &name, const framework::DDim &dim) = 0; + std::vector GetDims( + const std::vector &names) const { + std::vector ret; + ret.reserve(names.size()); + std::transform( + names.begin(), names.end(), std::back_inserter(ret), + [this](const std::string &name) { return this->GetDim(name); }); + return ret; + } + void SetDims(const std::vector &names, + const std::vector &dims) { + size_t length = names.size(); + PADDLE_ENFORCE_EQ(length, dims.size()); + for (size_t i = 0; i < length; ++i) { + SetDim(names[i], dims[i]); + } + } +}; + +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/tensor.h b/paddle/framework/tensor.h index f040c09c089ec75c9773d752685be5e232e8f4b7..ba82127d9c028eb39b9dc1a7f34fcf546524142b 100644 --- a/paddle/framework/tensor.h +++ b/paddle/framework/tensor.h @@ -29,20 +29,10 @@ limitations under the License. */ namespace paddle { -namespace pybind { -namespace details { -template -struct CastToPyBufferImpl; -} -} // namespace pybind - namespace framework { class Tensor { public: - template - friend struct pybind::details::CastToPyBufferImpl; - template friend struct EigenTensor; @@ -105,6 +95,19 @@ class Tensor { template inline void CopyFrom(const Tensor& src, const platform::Place& dst_place); + /** + * @brief Copy the content of an external vector to a tensor. + * + * @param[in] src The external vector. + * @param[in] ctx The device context contains place where to store. + * + * * @note CopyFromVector assumes that the tensor has been resized + * before invoking. + */ + template + inline void CopyFromVector(const std::vector& src, + const platform::Place& dst_place); + /** * @brief Return the slice of the tensor. * @@ -119,6 +122,8 @@ class Tensor { return holder_->place(); } + std::type_index type() const { return holder_->type(); } + private: template inline void check_memory_size() const; diff --git a/paddle/framework/tensor_array.cc b/paddle/framework/tensor_array.cc new file mode 100644 index 0000000000000000000000000000000000000000..2728bce1c1af848285e80d8ee8b3b61ec046342e --- /dev/null +++ b/paddle/framework/tensor_array.cc @@ -0,0 +1,283 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + + + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/framework/tensor_array.h" + +#include +#include +#include + +namespace paddle { +namespace framework { + +namespace detail { + +/* + * Offer an iterator over the length-sorted lod-tensor's top level. The top + * level of a lod-tensor stores batch-size of sequences, each top-level sequence + * may contains several lower-level sequences, sort top-level lod by the numbers + * of lower-level sequences in descending order, so that during RNN's running, + * the batch-size will keep decreasing, the short sentences will end at the tail + * of each batch. + * + * Let's take a simple lod-tensor for example + * + * |(0) |(1) top-level has two instances + * ||| ||||| lower-level + * + * sort by lower-level's length + * + * |(1) |(0) + * ||||| ||| + * + * when RNN runs, it get 5 batches (equals the number of elements the longest + * sequence has) + * + * ||||| + * ||| + * + * the first three batches has two elements, the last two elements just has 1 + * element each. + */ +struct DynamicBatchUnpacker { + using value_type = float; + + DynamicBatchUnpacker(const LoDTensor& source, size_t level, + bool descend = true) + : source(&source), level(level) { + BuildLengthSortedMeta(descend); + } + + LoDTensor GetBatch(size_t index); + + std::vector meta; + + LoDTensor const* source; + size_t level; + + protected: + void BuildLengthSortedMeta(bool descend); +}; + +LoDTensor PackDynamicBatch(const std::vector& source, + const std::vector& meta, const LoD& lod, + size_t level); + +} // namespace detail + +const LoDTensor& TensorArray::Read(size_t index) const { + PADDLE_ENFORCE_LE(index, MAX_SIZE, "index[%d] too large", index); + if (index >= size()) { + values_.resize(index + 1); + } + return values_[index]; +} + +void TensorArray::Write(size_t index, const LoDTensor& value) { + PADDLE_ENFORCE_LE(index, MAX_SIZE, "index[%d] too large", index); + + if (index >= size()) { + values_.resize(index + 1); + } + + values_[index].Resize(value.dims()); + values_[index].mutable_data(platform::CPUPlace()); + values_[index].CopyFrom(value, platform::CPUPlace()); +} + +void TensorArray::WriteShared(size_t index, const LoDTensor& value) { + PADDLE_ENFORCE_LE(index, MAX_SIZE, "index[%d] too large", index); + if (index >= size()) { + values_.resize(index + 1); + } + + values_[index].ShareDataWith(value); +} + +LoDTensor TensorArray::Pack(size_t level, const std::vector& meta, + const LoD& lod) const { + return detail::PackDynamicBatch(values_, meta, lod, level); +} + +std::vector TensorArray::Unpack(const LoDTensor& source, int level, + bool length_desend) { + detail::DynamicBatchUnpacker unpacker(source, level, + length_desend /*descend*/); + + // find max length of all the sequences + size_t max_length = 0; + for (const auto& seq : unpacker.meta) { + max_length = std::max(max_length, seq.end - seq.begin); + } + + // write batches to values + for (size_t batch_id = 0; batch_id < max_length; batch_id++) { + Write(batch_id, unpacker.GetBatch(batch_id)); + } + + return unpacker.meta; +} + +LoDTensor TensorArray::Stack() const { + LoDTensor result; + if (size() == 0) return result; + + const auto& first_dims = values_.front().dims(); + // check all the values have the same shape + // TODO(superjom) check the same dtypes + for (size_t idx = 1; idx < size(); idx++) { + const auto& value_dims = values_[idx].dims(); + PADDLE_ENFORCE_EQ(first_dims, value_dims); + } + + // copy + auto result_dims = vectorize(first_dims); + result_dims.insert(result_dims.begin(), size()); + result.Resize(make_ddim(result_dims)); + result.mutable_data(platform::CPUPlace()); + + for (size_t idx = 0; idx < size(); idx++) { + result.Slice(idx, idx + 1) + .CopyFrom(Read(idx), platform::CPUPlace()); + } + return result; +} + +void TensorArray::Unstack(const LoDTensor& source) const { + Unstack(source, false /*data_shared*/); +} + +void TensorArray::UnstackShared(const LoDTensor& source) const { + Unstack(source, true /*data_shared*/); +} + +void TensorArray::Unstack(const LoDTensor& source, bool data_shared) const { + size_t first_dim = source.dims()[0]; + DDim value_dims = slice_ddim(source.dims(), 1, source.dims().size()); + PADDLE_ENFORCE_GT(first_dim, 0, + "source should have some data to be unstacked"); + + values_.resize(first_dim); + + for (size_t elem = 0; elem < first_dim; elem++) { + // create a new value + auto& value = values_[elem]; + if (data_shared) { + // share memory + value.ShareDataWith(source.Slice(elem, elem + 1)); + } else { + // copy + value.Resize(value_dims); + value.CopyFrom(source.Slice(elem, elem + 1), + platform::CPUPlace()); + } + } +} + +size_t TensorArray::size() const { return values_.size(); } + +namespace detail { + +void DynamicBatchUnpacker::BuildLengthSortedMeta(bool descend) { + PADDLE_ENFORCE(meta.empty(), "duplicate build meta"); + // collect meta for each sequence in some level + auto lod = SliceLevels(source->lod(), level, level + 1)[0]; + + for (size_t seq_id = 0; seq_id < lod.size() - 1; seq_id++) { + DySeqMeta seq_meta({lod[seq_id], lod[seq_id + 1], seq_id}); + meta.push_back(seq_meta); + } + + PADDLE_ENFORCE_GT(meta.size(), 0, "meta is empty"); + + // sort by length + sort(meta.begin(), meta.end(), + [descend](const DySeqMeta& a, const DySeqMeta& b) { + bool a_ge_b = (a.end - a.begin) > (b.end - b.begin); + return descend ? a_ge_b : !a_ge_b; + }); +} + +LoDTensor DynamicBatchUnpacker::GetBatch(size_t index) { + PADDLE_ENFORCE(!meta.empty(), "should build meta first"); + LoDTensor result; + + // collect indice need to copy to the batch + std::vector indice; + for (const auto& seq : meta) { + size_t id = seq.begin + index; + if (id >= seq.end) break; + indice.push_back(id); + } + PADDLE_ENFORCE(!indice.empty(), "invalid batch at %d", index); + + // copy the indice of records in LoDTensor + auto record_dims = slice_ddim(source->dims(), 1, source->dims().size()); + auto record_dims_vec = vectorize(record_dims); + record_dims_vec.insert(record_dims_vec.begin(), indice.size()); + result.Resize(make_ddim(record_dims_vec)); + result.mutable_data(platform::CPUPlace()); + + for (size_t i = 0; i < indice.size(); i++) { + auto index = indice[i]; + auto target = result.Slice(i, i + 1); + auto source_ = source->Slice(index, index + 1); + + target.CopyFrom(source_, platform::CPUPlace()); + } + + return result; +} + +// TODO(supejom) to cache lod if reasonable +LoDTensor PackDynamicBatch(const std::vector& source, + const std::vector& meta, const LoD& lod, + size_t level) { + PADDLE_ENFORCE(!source.empty()); + PADDLE_ENFORCE(!meta.empty()); + PADDLE_ENFORCE(!lod.empty()); + + LoDTensor result; + + // init result space + auto record_dims = slice_ddim(source[0].dims(), 1, source[0].dims().size()); + auto record_dims_vec = vectorize(record_dims); + auto height = lod[level].back(); + record_dims_vec.insert(record_dims_vec.begin(), height); + result.Resize(make_ddim(record_dims_vec)); + result.mutable_data(platform::CPUPlace()); + + for (size_t batch_id = 0; batch_id < source.size(); batch_id++) { + for (size_t seq_id = 0; seq_id < meta.size(); seq_id++) { + const auto& seq_meta = meta[seq_id]; + // source is source[batch_id][seq_id] + // target is result[index] + auto index = seq_meta.begin + batch_id; + if (index >= seq_meta.end) break; + auto source_ = source[batch_id].Slice(seq_id, seq_id + 1); + auto target = result.Slice(index, index + 1); + target.CopyFrom(source_, platform::CPUPlace()); + } + } + + result.set_lod(lod); + return result; +} + +} // namespace detail + +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/tensor_array.h b/paddle/framework/tensor_array.h new file mode 100644 index 0000000000000000000000000000000000000000..94a14c2df492b175cf6a643800937878e95c5f37 --- /dev/null +++ b/paddle/framework/tensor_array.h @@ -0,0 +1,113 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once +#include + +#include "paddle/framework/lod_tensor.h" + +namespace paddle { +namespace framework { + +/* + * DyBatchSeqPosition stores indices of the basic element in tensor. It is used + * after lod-tensor's re-assembling, its info can be used to recover the order + * in original lod-tensor. + */ +struct DySeqMeta { + DySeqMeta(size_t begin, size_t end, size_t ori_idx) + : begin(begin), end(end), ori_idx(ori_idx) {} + + size_t begin; + size_t end; // not included + size_t ori_idx; +}; + +/* + * TensorArray is a C-array-like array of tensors, it is meant to be used with + * dynamic iteration primitives such as while_loop. It is used to segment inputs + * and store states in all time steps. + * + * By providing some methods similar to a C++ array, the difinition of some + * state-based dynamic models such as RNN cound be more natural and highly + * flexible. + */ +class TensorArray { + public: + using value_type = float; + + // max number of values allowed to store. + const size_t MAX_SIZE{100000}; + + /* + * Read the value at location `index` in the `TensorArray`. + */ + const LoDTensor &Read(size_t index) const; + + /* + * Write value into the index of the TensorArray. + */ + void Write(size_t index, const LoDTensor &value); + + /* + * Write value into the index of the TensorArray, with memory shared. + */ + void WriteShared(size_t index, const LoDTensor &value); + + /* + * Recover the original LoD-arranged LoDTensor with the `values`, `level` and + * `indice_map`. + */ + LoDTensor Pack(size_t level, const std::vector &meta, + const LoD &lod) const; + + /* + * Split LoDTensor in some `level` and write the generated batches to + * `values`, if set `desend`, will sort by length in descending order else in + * ascending order. + */ + std::vector Unpack(const LoDTensor &source, int level, + bool length_desend); + + /* + * Pack the values into a tensor with rank one higher than each tensor in + * values. + */ + LoDTensor Stack() const; + + /* + * Unpacks the given division of a rank-`R` tensor into rank-`(R-1)` tensors. + */ + void Unstack(const LoDTensor &source) const; + + /* + * Unpacks the given division of a rank-`R` tensor into rank-`(R-1)` tensors, + * with memory of tensors shared. + */ + void UnstackShared(const LoDTensor &source) const; + + /* + * Return the number of values. + */ + size_t size() const; + + protected: + void Unstack(const LoDTensor &source, bool data_shared) const; + + private: + mutable std::vector values_; +}; // class TensorArray + +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/tensor_array_test.cc b/paddle/framework/tensor_array_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..d9f52509cdd1b79f6d53b5d4922f9e44279de08b --- /dev/null +++ b/paddle/framework/tensor_array_test.cc @@ -0,0 +1,130 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/framework/tensor_array.h" + +#include + +namespace paddle { +namespace framework { + +class TensorArrayTester : public ::testing::Test { + protected: + void SetUp() override { + LoDTensor source; + source.Resize(make_ddim({batch_size, dim})); + int* data = source.mutable_data(platform::CPUPlace()); + for (int i = 0; i < 16 * 32; i++) { + data[i] = i; + } + ta.Unstack(source); + } + + TensorArray ta; + const int batch_size = 16; + const int dim = 32; +}; + +TEST_F(TensorArrayTester, Read) { + for (int i = 0; i < batch_size; i++) { + const auto& tensor = ta.Read(i); + ASSERT_EQ(tensor.dims()[0], 1); + ASSERT_EQ(tensor.dims()[1], dim); + } +} + +TEST_F(TensorArrayTester, Write) { + LoDTensor source; + source.Resize(make_ddim({1, dim})); + for (int i = 0; i < dim; i++) { + *(source.mutable_data(platform::CPUPlace()) + i) = i; + } + + ta.Write(2, source); + + const auto& tensor = ta.Read(2); + for (int i = 0; i < dim; i++) { + EXPECT_EQ(*(tensor.data() + i), *(source.data() + i)); + } +} + +TEST_F(TensorArrayTester, WriteShared) { + LoDTensor source; + source.Resize(make_ddim({1, dim})); + for (int i = 0; i < dim; i++) { + *(source.mutable_data(platform::CPUPlace()) + i) = i; + } + + ta.WriteShared(2, source); + + const auto& tensor = ta.Read(2); + for (int i = 0; i < dim; i++) { + EXPECT_EQ(*(tensor.data() + i), *(source.data() + i)); + } + + EXPECT_EQ(source.data(), tensor.data()); +} + +class TensorArrayPackTester : public ::testing::Test { + protected: + virtual void SetUp() override { + lod.push_back(std::vector{0, 2, 9, 13}); + + source.set_lod(lod); + source.Resize(make_ddim({13, 128})); + source.mutable_data(platform::CPUPlace()); + + // content of each setence: 0 1 2 3 4 + const auto& level = lod.front(); + for (size_t i = 0; i < level.size() - 1; i++) { + size_t begin = level[i]; + size_t end = level[i + 1]; + for (size_t j = begin; j < end; j++) { + auto record = source.Slice(j, j + 1); + for (int dim = 0; dim < 128; dim++) { + record.mutable_data(platform::CPUPlace())[dim] = j - begin; + } + } + } + + // unpack + meta = ta.Unpack(source, 0, true); + } + + LoD lod; + TensorArray ta; + LoDTensor source; + std::vector meta; +}; + +TEST_F(TensorArrayPackTester, Unpack) { + ASSERT_EQ(ta.size(), 7UL); + + const auto& t0 = ta.Read(0); + const auto& t1 = ta.Read(1); + + ASSERT_EQ(t0.data()[0], int(0)); + ASSERT_EQ(t1.data()[0], int(1)); +} + +TEST_F(TensorArrayPackTester, Pack) { + LoDTensor packed = ta.Pack(0, meta, lod); +} + +TEST_F(TensorArrayTester, size) { + ASSERT_EQ(ta.size(), static_cast(batch_size)); +} + +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/tensor_impl.h b/paddle/framework/tensor_impl.h index a5405f9c31543b5733f9db923c2a6f8b968cfc2d..8ee9941982cdd8f78fdbace9dca085097b08eeb8 100644 --- a/paddle/framework/tensor_impl.h +++ b/paddle/framework/tensor_impl.h @@ -65,7 +65,7 @@ inline T* Tensor::mutable_data(platform::Place place) { holder_.reset(new PlaceholderImpl( boost::get(place), size)); } else if (platform::is_gpu_place(place)) { -#ifdef PADDLE_ONLY_CPU +#ifndef PADDLE_WITH_CUDA PADDLE_THROW("'GPUPlace' is not supported in CPU only device."); } #else @@ -103,7 +103,7 @@ inline void Tensor::CopyFrom(const Tensor& src, memory::Copy(boost::get(dst_place), dst_ptr, boost::get(src_place), src_ptr, size); } -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA else if (platform::is_gpu_place(src_place) && platform::is_cpu_place(dst_place)) { memory::Copy(boost::get(dst_place), dst_ptr, @@ -123,6 +123,29 @@ inline void Tensor::CopyFrom(const Tensor& src, #endif } +template +inline void Tensor::CopyFromVector(const std::vector& src, + const platform::Place& dst_place) { + auto src_ptr = static_cast(src.data()); + platform::CPUPlace src_place; + auto dst_ptr = static_cast(mutable_data(dst_place)); + auto size = src.size() * sizeof(T); + + if (platform::is_cpu_place(dst_place)) { + memory::Copy(boost::get(dst_place), dst_ptr, src_place, + src_ptr, size); + } +#ifdef PADDLE_WITH_CUDA + else if (platform::is_gpu_place(dst_place)) { + memory::Copy(boost::get(dst_place), dst_ptr, src_place, + src_ptr, size, 0); + } + PADDLE_ENFORCE(cudaStreamSynchronize(0), + "cudaStreamSynchronize failed in Tensor CopyFromVector"); + +#endif +} + template inline Tensor Tensor::Slice(const int& begin_idx, const int& end_idx) const { check_memory_size(); diff --git a/paddle/framework/tensor_test.cc b/paddle/framework/tensor_test.cc index e2ec738de35c90c6a06c9a46b062d4cce55f5eda..492eba69e1ea483eca1da782004231af61fc60be 100644 --- a/paddle/framework/tensor_test.cc +++ b/paddle/framework/tensor_test.cc @@ -74,7 +74,7 @@ TEST(Tensor, MutableData) { EXPECT_EQ(p1, p2); } -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA { Tensor src_tensor; float* p1 = nullptr; @@ -126,7 +126,7 @@ TEST(Tensor, ShareDataWith) { ASSERT_EQ(src_tensor.data(), dst_tensor.data()); } -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA { Tensor src_tensor; Tensor dst_tensor; @@ -163,7 +163,7 @@ TEST(Tensor, Slice) { EXPECT_EQ(src_data_address + 3 * 4 * 1 * sizeof(int), slice_data_address); } -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA { Tensor src_tensor; src_tensor.mutable_data(make_ddim({6, 9}), GPUPlace()); @@ -218,7 +218,7 @@ TEST(Tensor, CopyFrom) { EXPECT_EQ(dst_ptr[i], slice_ptr[i]); } } -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA { Tensor src_tensor; Tensor gpu_tensor; @@ -263,6 +263,93 @@ TEST(Tensor, CopyFrom) { #endif } +TEST(Tensor, CopyFromVector) { + using namespace paddle::framework; + using namespace paddle::platform; + { + std::vector src_vec = {1, 2, 3, 4, 5, 6, 7, 8, 9}; + Tensor cpu_tensor; + + // Copy to CPU Tensor + cpu_tensor.Resize(make_ddim({3, 3})); + auto cpu_place = new paddle::platform::CPUPlace(); + cpu_tensor.CopyFromVector(src_vec, *cpu_place); + + // Compare Tensors + const int* cpu_ptr = cpu_tensor.data(); + const int* src_ptr = src_vec.data(); + ASSERT_NE(src_ptr, cpu_ptr); + for (size_t i = 0; i < 9; ++i) { + EXPECT_EQ(src_ptr[i], cpu_ptr[i]); + } + + src_vec.erase(src_vec.begin(), src_vec.begin() + 5); + cpu_tensor.Resize(make_ddim({2, 2})); + cpu_tensor.CopyFromVector(src_vec, *cpu_place); + cpu_ptr = cpu_tensor.data(); + src_ptr = src_vec.data(); + ASSERT_NE(src_ptr, cpu_ptr); + for (size_t i = 0; i < 5; ++i) { + EXPECT_EQ(src_ptr[i], cpu_ptr[i]); + } + + delete cpu_place; + } + +#ifdef PADDLE_WITH_CUDA + { + std::vector src_vec = {1, 2, 3, 4, 5, 6, 7, 8, 9}; + Tensor cpu_tensor; + Tensor gpu_tensor; + Tensor dst_tensor; + + // Copy to CPU Tensor + cpu_tensor.Resize(make_ddim({3, 3})); + auto cpu_place = new paddle::platform::CPUPlace(); + cpu_tensor.CopyFromVector(src_vec, *cpu_place); + + // Copy to GPUTensor + gpu_tensor.Resize(make_ddim({3, 3})); + auto gpu_place = new paddle::platform::GPUPlace(); + gpu_tensor.CopyFromVector(src_vec, *gpu_place); + // Copy from GPU to CPU tensor for comparison + dst_tensor.CopyFrom(gpu_tensor, *cpu_place); + + // Compare Tensors + const int* src_ptr = src_vec.data(); + const int* cpu_ptr = cpu_tensor.data(); + const int* dst_ptr = dst_tensor.data(); + ASSERT_NE(src_ptr, cpu_ptr); + ASSERT_NE(src_ptr, dst_ptr); + for (size_t i = 0; i < 9; ++i) { + EXPECT_EQ(src_ptr[i], cpu_ptr[i]); + EXPECT_EQ(src_ptr[i], dst_ptr[i]); + } + + src_vec.erase(src_vec.begin(), src_vec.begin() + 5); + + cpu_tensor.Resize(make_ddim({2, 2})); + cpu_tensor.CopyFromVector(src_vec, *cpu_place); + gpu_tensor.Resize(make_ddim({2, 2})); + gpu_tensor.CopyFromVector(src_vec, *gpu_place); + dst_tensor.CopyFrom(gpu_tensor, *cpu_place); + + src_ptr = src_vec.data(); + cpu_ptr = cpu_tensor.data(); + dst_ptr = dst_tensor.data(); + ASSERT_NE(src_ptr, cpu_ptr); + ASSERT_NE(src_ptr, dst_ptr); + for (size_t i = 0; i < 5; ++i) { + EXPECT_EQ(src_ptr[i], cpu_ptr[i]); + EXPECT_EQ(src_ptr[i], dst_ptr[i]); + } + + delete cpu_place; + delete gpu_place; + } +#endif +} + TEST(Tensor, ReshapeToMatrix) { using namespace paddle::framework; using namespace paddle::platform; diff --git a/paddle/framework/type_defs.h b/paddle/framework/type_defs.h new file mode 100644 index 0000000000000000000000000000000000000000..6f65a942ba2a4073e6aa1047875ec5c3283c23a6 --- /dev/null +++ b/paddle/framework/type_defs.h @@ -0,0 +1,43 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once +#include +#include +#include +#include "paddle/platform/variant.h" + +namespace paddle { +namespace framework { +class OperatorBase; +class OpDescBind; +using VariableNameMap = std::map>; + +// The order should be as same as framework.proto +using Attribute = + boost::variant, + std::vector, std::vector, bool, + std::vector, BlockDesc*>; + +using AttributeMap = std::unordered_map; + +using OpCreator = std::function; + +using GradOpMakerFN = + std::function>(const OpDescBind&)>; + +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/var_desc.cc b/paddle/framework/var_desc.cc new file mode 100644 index 0000000000000000000000000000000000000000..13b9c5f3cdf98e6d22f4217fa1cf9a48910a78d8 --- /dev/null +++ b/paddle/framework/var_desc.cc @@ -0,0 +1,36 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/framework/var_desc.h" + +namespace paddle { +namespace framework { + +void VarDescBind::SetShape(const std::vector &dims) { + VectorToRepeated(dims, desc_.mutable_lod_tensor()->mutable_dims()); +} + +void VarDescBind::SetDataType(DataType data_type) { + desc_.mutable_lod_tensor()->set_data_type(data_type); +} + +std::vector VarDescBind::Shape() const { + return RepeatedToVector(desc_.lod_tensor().dims()); +} + +DataType VarDescBind::GetDataType() const { + return desc_.lod_tensor().data_type(); +} +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/var_desc.h b/paddle/framework/var_desc.h new file mode 100644 index 0000000000000000000000000000000000000000..4763bf09d004539ab24e4aad3bf429667f1fcc73 --- /dev/null +++ b/paddle/framework/var_desc.h @@ -0,0 +1,73 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include "paddle/framework/framework.pb.h" + +namespace paddle { +namespace framework { + +// convert between std::vector and protobuf repeated. +template +inline std::vector RepeatedToVector( + const google::protobuf::RepeatedField &repeated_field) { + std::vector ret; + ret.reserve(repeated_field.size()); + std::copy(repeated_field.begin(), repeated_field.end(), + std::back_inserter(ret)); + return ret; +} + +template +inline void VectorToRepeated(const std::vector &vec, + RepeatedField *repeated_field) { + repeated_field->Reserve(vec.size()); + for (const auto &elem : vec) { + *repeated_field->Add() = elem; + } +} + +// Specialize vector. +template +inline void VectorToRepeated(const std::vector &vec, + RepeatedField *repeated_field) { + repeated_field->Reserve(vec.size()); + for (auto elem : vec) { + *repeated_field->Add() = elem; + } +} + +class VarDescBind { + public: + explicit VarDescBind(const std::string &name) { desc_.set_name(name); } + + VarDesc *Proto() { return &desc_; } + + std::string Name() const { return desc_.name(); } + + void SetShape(const std::vector &dims); + + void SetDataType(DataType data_type); + + std::vector Shape() const; + + DataType GetDataType() const; + + private: + VarDesc desc_; +}; +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/variable.md b/paddle/framework/variable.md index f44d5ea46e7ce98dd443d684ad42308496bc4179..442ef6b718b227d79ca73031efcbb55817558252 100644 --- a/paddle/framework/variable.md +++ b/paddle/framework/variable.md @@ -7,7 +7,7 @@ Variable is also known as *blob* in MxNet and Caffe2. It is the input and outpu For the flexibility of a DL system, a variable should be able to contain any typed value -- a tensor in most cases, but could also be some integer IDs or a scope of other variables in the case of RNN. -To use the minimum amount of memory, we'd like that a variable to allocate memory when it has to, or, lazy memory allocation. Let's take the following example: +To use the minimum amount of memory, we would like that a variable allocates memory only when it has to, or, lazy memory allocation. Let's take the following example: ```cpp Variable vr, v1, v2; @@ -38,7 +38,7 @@ This syntax for lazy memory allocation when we call `Randomize` and `Mult`, thos To make memory allocation lazy, we cannot assume that we know the type held by a variable at definition time. In other words, `class Variable` cannot be a template `template class Variable`. -Because we don't know the type `T`, we cannot save a `T*` as `Variable's` data member. Instead, we save an interface object `Placeholder`, who can return the pointer to the saved object via `Placeholder::Ptr()` as `void*`. +Because we don't know the type `T`, we cannot save a `T*` as `Variable's` data member. Instead, we save an interface object `Placeholder`, which can return the pointer to the saved object via `Placeholder::Ptr()` as `void*`. But anyway, Variable needs to know `T` so could it `delete(ptr)` and so could `Variable::Get` checks the expected type and the saved object's type. @@ -49,4 +49,4 @@ Because `PlaceholderImpl` knows `T`, it can save and return `typeid(T)` for the ## Conclusion -The technique type hiding utilizes C++ class templates, interface and derivation, and C++ RTTI (typeid). This combination saves us from definition something like `caffe2::TypeMata`, which takes hundreds of lines of C++ code. +The technique type hiding utilizes C++ class templates, interface and derivation, and C++ RTTI (typeid). This combination saves us from defining something like `caffe2::TypeMeta`, which takes hundreds of lines of C++ code. diff --git a/paddle/function/BlockExpandOp.cpp b/paddle/function/BlockExpandOp.cpp index a89b6bba45843d81264819cad6ba053f28314f6b..bd0fe119ce46df9c333258c9c1ad7b5b2bdc544f 100644 --- a/paddle/function/BlockExpandOp.cpp +++ b/paddle/function/BlockExpandOp.cpp @@ -194,7 +194,7 @@ public: REGISTER_TYPED_FUNC(BlockExpand, CPU, BlockExpandForward); REGISTER_TYPED_FUNC(BlockExpandGrad, CPU, BlockExpandBackward); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA REGISTER_TYPED_FUNC(BlockExpand, GPU, BlockExpandForward); REGISTER_TYPED_FUNC(BlockExpandGrad, GPU, BlockExpandBackward); #endif diff --git a/paddle/function/ContextProjectionOp.cpp b/paddle/function/ContextProjectionOp.cpp index b87750b74247bd0eb822340bc5a85d41b86ecec2..23916c0f4b6319004ca0793bc9305b8a1dd0ae89 100644 --- a/paddle/function/ContextProjectionOp.cpp +++ b/paddle/function/ContextProjectionOp.cpp @@ -395,7 +395,7 @@ REGISTER_TYPED_FUNC(ContextProjectionForward, REGISTER_TYPED_FUNC(ContextProjectionBackward, CPU, ContextProjectionBackwardFunc); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA REGISTER_TYPED_FUNC(ContextProjectionForward, GPU, ContextProjectionForwardFunc); diff --git a/paddle/function/CosSimOp.cpp b/paddle/function/CosSimOp.cpp index 7ece7b2dfedaf460741c97b5a700eb632d85cabc..2e5c281f37d8ffb1062121b5dc5b4f790ab52089 100644 --- a/paddle/function/CosSimOp.cpp +++ b/paddle/function/CosSimOp.cpp @@ -233,7 +233,7 @@ private: REGISTER_TYPED_FUNC(CosSimForward, CPU, CosSimForwardFunc); REGISTER_TYPED_FUNC(CosSimBackward, CPU, CosSimBackwardFunc); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA REGISTER_TYPED_FUNC(CosSimForward, GPU, CosSimForwardFunc); REGISTER_TYPED_FUNC(CosSimBackward, GPU, CosSimBackwardFunc); #endif diff --git a/paddle/function/CropOp.cpp b/paddle/function/CropOp.cpp index f12ee43e3d72f9ac776eaff93914228850694dd2..46f98f12c1f150fdf3ed53a7a96e5cf0020e14a4 100644 --- a/paddle/function/CropOp.cpp +++ b/paddle/function/CropOp.cpp @@ -169,7 +169,7 @@ private: REGISTER_TYPED_FUNC(Crop, CPU, CropFunc); REGISTER_TYPED_FUNC(CropGrad, CPU, CropGradFunc); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA REGISTER_TYPED_FUNC(Crop, GPU, CropFunc); REGISTER_TYPED_FUNC(CropGrad, GPU, CropGradFunc); #endif diff --git a/paddle/function/CrossMapNormalOp.cpp b/paddle/function/CrossMapNormalOp.cpp index ef878bfbba961bdd3d5212e19fb83bb1e285e47f..9e88669d37bd50179dcc0464e8c1cd6e2fed74db 100644 --- a/paddle/function/CrossMapNormalOp.cpp +++ b/paddle/function/CrossMapNormalOp.cpp @@ -336,7 +336,7 @@ private: REGISTER_TYPED_FUNC(CrossMapNormal, CPU, CrossMapNormalFunc); REGISTER_TYPED_FUNC(CrossMapNormalGrad, CPU, CrossMapNormalGradFunc); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA REGISTER_TYPED_FUNC(CrossMapNormal, GPU, CrossMapNormalFunc); REGISTER_TYPED_FUNC(CrossMapNormalGrad, GPU, CrossMapNormalGradFunc); #endif diff --git a/paddle/function/DepthwiseConvOp.cpp b/paddle/function/DepthwiseConvOp.cpp index 2f3112fe657cd381891dc53c7179e7520911e8c9..9863e3ae1d5fcb1eece5267fd4f2a6b593b799df 100644 --- a/paddle/function/DepthwiseConvOp.cpp +++ b/paddle/function/DepthwiseConvOp.cpp @@ -292,7 +292,7 @@ REGISTER_TYPED_FUNC(DepthwiseConvGradInput, REGISTER_TYPED_FUNC(DepthwiseConvGradFilter, CPU, DepthwiseConvGradFilterFunction); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA REGISTER_TYPED_FUNC(DepthwiseConv, GPU, DepthwiseConvFunction); REGISTER_TYPED_FUNC(DepthwiseConvGradInput, GPU, diff --git a/paddle/function/DepthwiseConvOpTest.cpp b/paddle/function/DepthwiseConvOpTest.cpp index d8e8c889d5c23bf9b2b5fd0b0393395883188fd8..b1a90da7db2b647dd384e3772820294140e5ec9d 100644 --- a/paddle/function/DepthwiseConvOpTest.cpp +++ b/paddle/function/DepthwiseConvOpTest.cpp @@ -17,7 +17,7 @@ limitations under the License. */ namespace paddle { -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA TEST(DepthwiseConv, Forward) { DepthwiseConvolution( "GemmConv-CPU", "DepthwiseConv-GPU", forward); diff --git a/paddle/function/GemmConvOp.cpp b/paddle/function/GemmConvOp.cpp index f8cf4ebea8d724f0291b981647622b63e3d84495..bdb56ddac38b91d756fc6f31282f29c0489fd660 100644 --- a/paddle/function/GemmConvOp.cpp +++ b/paddle/function/GemmConvOp.cpp @@ -340,7 +340,7 @@ public: REGISTER_TYPED_FUNC(GemmConv, CPU, GemmConvFunction); REGISTER_TYPED_FUNC(GemmConvGradInput, CPU, GemmConvGradInputFunction); REGISTER_TYPED_FUNC(GemmConvGradFilter, CPU, GemmConvGradFilterFunction); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA REGISTER_TYPED_FUNC(GemmConv, GPU, GemmConvFunction); REGISTER_TYPED_FUNC(GemmConvGradInput, GPU, GemmConvGradInputFunction); REGISTER_TYPED_FUNC(GemmConvGradFilter, GPU, GemmConvGradFilterFunction); diff --git a/paddle/function/GemmConvOpTest.cpp b/paddle/function/GemmConvOpTest.cpp index 5283d79a5a53d979ae4e134f7e46b7ee106e9c44..b5b5e1f35b79e422b14f7495bc321533cc1d618a 100644 --- a/paddle/function/GemmConvOpTest.cpp +++ b/paddle/function/GemmConvOpTest.cpp @@ -24,7 +24,7 @@ TEST(GemmConv, NaiveConv) { "NaiveConv-CPU", "GemmConv-CPU", forward); } -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA TEST(GemmConv, Forward) { Convolution( "GemmConv-CPU", "GemmConv-GPU", forward); diff --git a/paddle/function/Im2ColTest.cpp b/paddle/function/Im2ColTest.cpp index acc88a553abe7ac58b629aba9b850df58cee7f81..a0a01a5fc7fc055dce6ddb3ee51c7ab18f8a4ca7 100644 --- a/paddle/function/Im2ColTest.cpp +++ b/paddle/function/Im2ColTest.cpp @@ -116,7 +116,7 @@ void TestIm2ColFunctor() { TEST(Im2ColFunctor, CPU) { TestIm2ColFunctor(); } -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA TEST(Im2ColFunctor, GPU) { TestIm2ColFunctor(); } diff --git a/paddle/function/MulOp.cpp b/paddle/function/MulOp.cpp index 25e41edad54bec0f76a3de4799fab14241407272..704a8c41325ef86067a3bd8ed6d772b77df147c5 100644 --- a/paddle/function/MulOp.cpp +++ b/paddle/function/MulOp.cpp @@ -341,7 +341,7 @@ private: }; REGISTER_TYPED_FUNC(MulOp, CPU, MulFunc); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA REGISTER_TYPED_FUNC(MulOp, GPU, MulFunc); #endif } // namespace paddle diff --git a/paddle/function/PadOp.cpp b/paddle/function/PadOp.cpp index adba7c92ece505eecc74edce6b393cf27fa10ccc..eed2f2e3089b6b6167ef7c5a7acb7ecaa08945e1 100644 --- a/paddle/function/PadOp.cpp +++ b/paddle/function/PadOp.cpp @@ -207,7 +207,7 @@ private: REGISTER_TYPED_FUNC(Pad, CPU, PadFunc); REGISTER_TYPED_FUNC(PadGrad, CPU, PadGradFunc); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA REGISTER_TYPED_FUNC(Pad, GPU, PadFunc); REGISTER_TYPED_FUNC(PadGrad, GPU, PadGradFunc); #endif diff --git a/paddle/function/RowConvOp.cpp b/paddle/function/RowConvOp.cpp index b6501e8f4db7fd33891cd80e07a6f36dd0b34532..7c802d66273c6f7aa56b2f460e3dff4401967517 100644 --- a/paddle/function/RowConvOp.cpp +++ b/paddle/function/RowConvOp.cpp @@ -217,7 +217,7 @@ public: REGISTER_TYPED_FUNC(RowConv, CPU, RowConvFunc); REGISTER_TYPED_FUNC(RowConvGrad, CPU, RowConvGradFunc); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA REGISTER_TYPED_FUNC(RowConv, GPU, RowConvFunc); REGISTER_TYPED_FUNC(RowConvGrad, GPU, RowConvGradFunc); #endif diff --git a/paddle/function/SwitchOp.cpp b/paddle/function/SwitchOp.cpp index 01e252a8dc0cd5fa1e964efa01d04cf282b3dfe7..597723a2dded6a6a116e05b7d4c942cd633e2c99 100644 --- a/paddle/function/SwitchOp.cpp +++ b/paddle/function/SwitchOp.cpp @@ -132,7 +132,7 @@ public: REGISTER_TYPED_FUNC(NCHW2NHWC, CPU, NCHW2NHWCFunc); REGISTER_TYPED_FUNC(NHWC2NCHW, CPU, NHWC2NCHWFunc); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA REGISTER_TYPED_FUNC(NCHW2NHWC, GPU, NCHW2NHWCFunc); REGISTER_TYPED_FUNC(NHWC2NCHW, GPU, NHWC2NCHWFunc); #endif diff --git a/paddle/function/neon/NeonDepthwiseConv.h b/paddle/function/neon/NeonDepthwiseConv.h index 33722d3cac61b62f5dce8f51105c1bf4e70c4a6c..98a86d278f39e70472793e6a1d38f7dae469fd62 100644 --- a/paddle/function/neon/NeonDepthwiseConv.h +++ b/paddle/function/neon/NeonDepthwiseConv.h @@ -18,7 +18,6 @@ limitations under the License. */ #include "neon_util.h" namespace paddle { - namespace neon { #if defined(__ARM_NEON__) || defined(__ARM_NEON) @@ -26,17 +25,20 @@ namespace neon { template struct DepthwiseConvKernel {}; -inline float32_t conv3x3(float32x4_t r0, - float32x4_t r1, - float32x4_t r2, +inline float32_t conv3x3(const float* r0, + const float* r1, + const float* r2, float32x4_t k0, float32x4_t k1, float32x4_t k2) { - float32x4_t tmp; - tmp = vmulq_f32(r0, k0); - tmp = vmlaq_f32(tmp, r1, k1); - tmp = vmlaq_f32(tmp, r2, k2); - return vaddvq_f32(tmp); + float32_t tmp[12]; + vst1q_f32(&(tmp[0]), k0); + vst1q_f32(&(tmp[4]), k1); + vst1q_f32(&(tmp[8]), k2); + float32_t sum0 = r0[0] * tmp[0] + r0[1] * tmp[1] + r0[2] * tmp[2]; + float32_t sum1 = r1[0] * tmp[4] + r1[1] * tmp[5] + r1[2] * tmp[6]; + float32_t sum2 = r2[0] * tmp[8] + r2[1] * tmp[9] + r2[2] * tmp[10]; + return sum0 + sum1 + sum2; } inline float32_t conv4x4(float32x4_t r0, @@ -136,10 +138,7 @@ struct DepthwiseConvKernel<3, 1> { } for (int r = 0; r < remain; r++) { - float32x4_t i0 = vld1q_f32(r0); - float32x4_t i1 = vld1q_f32(r1); - float32x4_t i2 = vld1q_f32(r2); - *outputData = conv3x3(i0, i1, i2, k[0], k[1], k[2]); + *outputData = conv3x3(r0, r1, r2, k[0], k[1], k[2]); r0++; r1++; r2++; @@ -243,10 +242,7 @@ struct DepthwiseConvKernel<3, 2> { } for (int r = 0; r < remain; r++) { - float32x4_t i0 = vld1q_f32(r0); - float32x4_t i1 = vld1q_f32(r1); - float32x4_t i2 = vld1q_f32(r2); - *outputData = conv3x3(i0, i1, i2, k[0], k[1], k[2]); + *outputData = conv3x3(r0, r1, r2, k[0], k[1], k[2]); r0 += 2; r1 += 2; r2 += 2; diff --git a/paddle/gserver/CMakeLists.txt b/paddle/gserver/CMakeLists.txt index 62cff9361ccba3ae3b9359ddb932f5b26146eb97..5f39167afc34affbea7858fa0794ef52b786a383 100644 --- a/paddle/gserver/CMakeLists.txt +++ b/paddle/gserver/CMakeLists.txt @@ -60,6 +60,36 @@ if(NOT WITH_PYTHON) dataproviders/PyDataProvider.h) endif() +if(MOBILE_INFERENCE) + # Remove evaluators + list(REMOVE_ITEM GSERVER_SOURCES + layers/ValidationLayer.cpp + evaluators/Evaluator.cpp + evaluators/DetectionMAPEvaluator.cpp + evaluators/CTCErrorEvaluator.cpp + evaluators/ChunkEvaluator.cpp) + + # Remove dataproviders + list(REMOVE_ITEM GSERVER_SOURCES + dataproviders/DataProvider.cpp + dataproviders/MultiDataProvider.cpp + dataproviders/ProtoDataProvider.cpp + dataproviders/PyDataProvider2.cpp + dataproviders/PyDataProvider.cpp) + + # Remove useless gradientmachines + list(REMOVE_ITEM GSERVER_SOURCES + gradientmachines/MultiNetwork.cpp + gradientmachines/RecurrentGradientMachine.cpp + gradientmachines/ParallelNeuralNetwork.cpp + gradientmachines/GradientMachineMode.cpp + gradientmachines/MultiGradientMachine.cpp) + + # Remove useless layers + list(REMOVE_ITEM GSERVER_SOURCES + layers/RecurrentLayerGroup.cpp) +endif() + if(WITH_GPU) cuda_add_library(paddle_gserver ${GSERVER_SOURCES}) else() diff --git a/paddle/gserver/activations/MKLDNNActivation.cpp b/paddle/gserver/activations/MKLDNNActivation.cpp index ac50937ef3e28c1ac5aae651f9cf266ad07abcc4..18c5638100065109fba1f0647a1c5f91256f7b9d 100644 --- a/paddle/gserver/activations/MKLDNNActivation.cpp +++ b/paddle/gserver/activations/MKLDNNActivation.cpp @@ -27,31 +27,53 @@ static ClassRegistrar gMKLDNNActivationRegistrar; #define MKLDNN_ACTIVATION_CLASS_NAME(ACT_TYPE) mkldnn_##ACT_TYPE##Activation /** - * @def DEFINE_MKLDNN_ELTWISE_ACTIVATION + * @def BEGIN_MKLDNN_ACTIVATION + */ +#define BEGIN_MKLDNN_ACTIVATION(ACT_TYPE, BASE_CLASS) \ + class MKLDNN_ACTIVATION_CLASS_NAME(ACT_TYPE) : public BASE_CLASS { +/** + * @def END_MKLDNN_ACTIVATION */ -#define DEFINE_MKLDNN_ELTWISE_ACTIVATION(ACT_TYPE, ALPHA, BWD_ALPHA) \ - class MKLDNN_ACTIVATION_CLASS_NAME(ACT_TYPE) \ - : public MKLDNNEltwiseActivation { \ - private: \ - static const std::string name; \ - static const float alpha; \ - static const float bwdAlpha; \ - \ - public: \ - const std::string& getName() const { return name; } \ - float getAlpha() const { return alpha; } \ - float getBwdAlpha() const { return bwdAlpha; } \ - }; \ - const std::string MKLDNN_ACTIVATION_CLASS_NAME(ACT_TYPE)::name = \ - "mkldnn_" #ACT_TYPE; \ - const float MKLDNN_ACTIVATION_CLASS_NAME(ACT_TYPE)::alpha = ALPHA; \ - const float MKLDNN_ACTIVATION_CLASS_NAME(ACT_TYPE)::bwdAlpha = BWD_ALPHA; \ - static InitFunction __reg_activation__mkldnn_##ACT_TYPE([] { \ - gMKLDNNActivationRegistrar \ - .registerClass( \ - "mkldnn_" #ACT_TYPE); \ +#define END_MKLDNN_ACTIVATION(ACT_TYPE) \ +private: \ + static const std::string name; \ + \ +public: \ + const std::string& getName() const { return name; } \ + } \ + ; \ + const std::string MKLDNN_ACTIVATION_CLASS_NAME(ACT_TYPE)::name = \ + "mkldnn_" #ACT_TYPE; \ + static InitFunction __reg_activation__mkldnn_##ACT_TYPE([] { \ + gMKLDNNActivationRegistrar \ + .registerClass( \ + "mkldnn_" #ACT_TYPE); \ }); +/** + * @def DEFINE_MKLDNN_ACTIVATION + */ +#define DEFINE_MKLDNN_ACTIVATION(ACT_TYPE, BASE_CLASS) \ + BEGIN_MKLDNN_ACTIVATION(ACT_TYPE, BASE_CLASS) \ + END_MKLDNN_ACTIVATION(ACT_TYPE) + +/** + * @def DEFINE_MKLDNN_ELTWISE_ACTIVATION + */ +#define DEFINE_MKLDNN_ELTWISE_ACTIVATION( \ + ACT_TYPE, BASE_CLASS, ALPHA, BWD_ALPHA) \ + BEGIN_MKLDNN_ACTIVATION(ACT_TYPE, BASE_CLASS) \ +private: \ + static const float alpha; \ + static const float bwdAlpha; \ + \ +public: \ + float getAlpha() const { return alpha; } \ + float getBwdAlpha() const { return bwdAlpha; } \ + END_MKLDNN_ACTIVATION(ACT_TYPE) \ + const float MKLDNN_ACTIVATION_CLASS_NAME(ACT_TYPE)::alpha = ALPHA; \ + const float MKLDNN_ACTIVATION_CLASS_NAME(ACT_TYPE)::bwdAlpha = BWD_ALPHA; + /** * @brief MKLDNN Relu Activation. * Actually mkldnn_relu is Leaky Relu. @@ -59,19 +81,129 @@ static ClassRegistrar gMKLDNNActivationRegistrar; * f(x) = negative_slope * x (x < 0) * @note the negative_slope should be -0.f in forward */ -DEFINE_MKLDNN_ELTWISE_ACTIVATION(relu, -0.f, 0.f) +DEFINE_MKLDNN_ELTWISE_ACTIVATION(relu, MKLDNNEltwiseActivation, -0.f, 0.f) /** * @brief MKLDNN Tanh Activation. */ -DEFINE_MKLDNN_ELTWISE_ACTIVATION(tanh, 0.f, 0.f) +DEFINE_MKLDNN_ELTWISE_ACTIVATION(tanh, MKLDNNEltwiseActivation, 0.f, 0.f) /** * @brief MKLDNN ELU(Exponential Linear Unit) Activation. * f(x) = x (x >= 0) * f(x) = negative_slope * (exp(x) - 1) (x < 0) */ -DEFINE_MKLDNN_ELTWISE_ACTIVATION(elu, 0.f, 0.f) +DEFINE_MKLDNN_ELTWISE_ACTIVATION(elu, MKLDNNEltwiseActivation, 0.f, 0.f) + +mkldnn::algorithm MKLDNNEltwiseActivation::getAlgo(std::string type) const { + const std::map algoMap = { + {"relu", algorithm::eltwise_relu}, + {"tanh", algorithm::eltwise_tanh}, + {"elu", algorithm::eltwise_elu}}; + type.erase(0, 7); // remove mkldnn_ + algorithm algo = (algorithm)0; + mapGet(type, algoMap, &algo); + return algo; +} + +void MKLDNNEltwiseActivation::resetFwd(Argument& act) { + if (cnt_ == act.value->getElementCnt()) { + return; + } + MKLDNNActivation::resetFwd(act); + // note: alpha represents the NegativeSlope when used in relu. + float alpha = getAlpha(); + float beta = getBeta(); + algorithm algo = getAlgo(this->getName()); + auto fwdDesc = eltwise_fwd::desc(mkldnn::prop_kind::forward_training, + algo, + val_->getMemoryDesc(), + alpha, + beta); + fwdPD_.reset(new eltwise_fwd::primitive_desc(fwdDesc, *engine_)); + // use inplace for forward but save input value before submit + inVal_ = val_; + copyInVal_ = nullptr; + if (act.grad && algo == algorithm::eltwise_tanh) { + // tanh need save src input for backward + inVal_ = MKLDNNMatrix::create(nullptr, val_->getPrimitiveDesc()); + copyInVal_ = std::make_shared(*val_, *inVal_); + CHECK(copyInVal_) << "should not be emptry"; + pipelineFwd_.push_back(*copyInVal_); + } + fwd_.reset(new eltwise_fwd(*fwdPD_, *val_, *val_)); + pipelineFwd_.push_back(*fwd_); + needResetBwd_ = true; +} + +void MKLDNNEltwiseActivation::resetBwd(Argument& act) { + if (!needResetBwd_) { + return; + } + VLOG(MKLDNN_BASE) << getName() << " reset mkldnn backward"; + needResetBwd_ = false; + algorithm algo = getAlgo(this->getName()); + float alpha = getBwdAlpha(); + float beta = getBeta(); + grad_ = MKLDNNMatrix::create(act.grad, val_->getPrimitiveDesc()); + auto eng = CPUEngine::Instance().getEngine(); + auto bwdDesc = eltwise_bwd::desc( + algo, grad_->getMemoryDesc(), val_->getMemoryDesc(), alpha, beta); + auto bwdPD = eltwise_bwd::primitive_desc(bwdDesc, eng, *fwdPD_); + CHECK(inVal_); + bwd_.reset(new eltwise_bwd(bwdPD, *inVal_, *grad_, *grad_)); + pipelineBwd_.clear(); + pipelineBwd_.push_back(*bwd_); +} + +/** + * @brief MKLDNN Softmax Activation + */ +DEFINE_MKLDNN_ACTIVATION(softmax, MKLDNNSoftmaxActivation) + +void MKLDNNSoftmaxActivation::resetFwd(Argument& act) { + if (cnt_ == act.value->getElementCnt()) { + return; + } + MKLDNNActivation::resetFwd(act); + int axis = 1; + auto fwdDesc = softmax_fwd::desc( + mkldnn::prop_kind::forward_scoring, val_->getMemoryDesc(), axis); + auto fwdPD = softmax_fwd::primitive_desc(fwdDesc, *engine_); + fwd_.reset(new softmax_fwd(fwdPD, *val_, *val_)); + pipelineFwd_.push_back(*fwd_); +} + +Error __must_check MKLDNNSoftmaxActivation::forward(Argument& act) { + resetFwd(act); + stream_->submit(pipelineFwd_); + real* v = act.value->getData(); + real threshold = exp(-64); +#pragma omp parallel for + for (size_t i = 0; i < act.value->getElementCnt(); ++i) { + v[i] = v[i] < threshold ? threshold : v[i]; + } + return Error(); +} + +Error __must_check MKLDNNSoftmaxActivation::backward(Argument& act) { + MatrixPtr outputV = act.value; + MatrixPtr outputG = act.grad; + Matrix::resizeOrCreate(sftMaxDot_, + outputG->getHeight(), + outputG->getWidth(), + /* trans */ false, + /* useGpu */ false); + Matrix::resizeOrCreate(sftMaxSum_, + outputG->getHeight(), + 1, + /* trans */ false, + /* useGpu */ false); + sftMaxDot_->dotMul(*outputG, *outputV); + sftMaxSum_->colMerge(*sftMaxDot_); + act.grad->softmaxDerivative(*act.value, *sftMaxSum_); + return Error(); +} ActivationFunction* MKLDNNActivation::create(const std::string& type) { return gMKLDNNActivationRegistrar.createByType(type); @@ -84,4 +216,34 @@ std::vector MKLDNNActivation::getAllRegisteredTypes() { return types; } +void MKLDNNActivation::resetFwd(Argument& act) { + VLOG(MKLDNN_BASE) << getName() << " reset mkldnn forward"; + cnt_ = act.value->getElementCnt(); + pipelineFwd_.clear(); + stream_.reset(new MKLDNNStream()); + engine_.reset(new mkldnn::engine(mkldnn::engine::cpu, 0)); + val_ = std::dynamic_pointer_cast(act.value); + if (val_ == nullptr) { + int bs = act.getBatchSize(); + int ih = act.getFrameHeight() > 0 ? act.getFrameHeight() : 1; + int iw = act.getFrameWidth() > 0 ? act.getFrameWidth() : 1; + int ic = cnt_ / bs / ih / iw; + CHECK_EQ(cnt_, (size_t)bs * ic * ih * iw); + val_ = MKLDNNMatrix::create( + act.value, {bs, ic, ih, iw}, mkldnn::memory::format::nchw, *engine_); + CHECK(val_); + val_->downSpatial(); + } +} + +Error __must_check MKLDNNActivation::forward(Argument& act) { + resetFwd(act); + stream_->submit(pipelineFwd_); + return Error(); +} +Error __must_check MKLDNNActivation::backward(Argument& act) { + resetBwd(act); + stream_->submit(pipelineBwd_); + return Error(); +} } // namespace paddle diff --git a/paddle/gserver/activations/MKLDNNActivation.h b/paddle/gserver/activations/MKLDNNActivation.h index 40dd8c618aa2b70d410130e12efc54520218afea..dd16421fd6e93b49c30b1d3b601f95980afec57b 100644 --- a/paddle/gserver/activations/MKLDNNActivation.h +++ b/paddle/gserver/activations/MKLDNNActivation.h @@ -36,6 +36,7 @@ protected: // mkldnn matrix, primitive, stream and pipeline MKLDNNMatrixPtr val_; MKLDNNMatrixPtr grad_; + std::shared_ptr engine_; std::shared_ptr stream_; std::shared_ptr fwd_; std::shared_ptr bwd_; @@ -48,8 +49,18 @@ public: static ActivationFunction* create(const std::string& type); static std::vector getAllRegisteredTypes(); virtual const std::string& getName() const = 0; - virtual Error __must_check forward(Argument& act) = 0; - virtual Error __must_check backward(Argument& act) = 0; + /** + * reset the forward primitives + */ + virtual void resetFwd(Argument& act); + /** + * reset the backward primitives, + * can not merge this functions into resetFwd as the grad data + * would be changing before backward. + */ + virtual void resetBwd(Argument& act) {} + virtual Error __must_check forward(Argument& act); + virtual Error __must_check backward(Argument& act); }; /** @@ -59,6 +70,7 @@ public: class MKLDNNEltwiseActivation : public MKLDNNActivation { typedef mkldnn::eltwise_forward eltwise_fwd; typedef mkldnn::eltwise_backward eltwise_bwd; + typedef mkldnn::algorithm algorithm; protected: // save the forward primitive desc, which can be used backward @@ -70,9 +82,7 @@ protected: public: MKLDNNEltwiseActivation() {} - ~MKLDNNEltwiseActivation() {} - virtual const std::string& getName() const = 0; // in common, the alpha of forward and backward should be equal. @@ -80,105 +90,30 @@ public: virtual float getAlpha() const = 0; virtual float getBwdAlpha() const = 0; virtual float getBeta() const { return 0.f; } - virtual mkldnn::algorithm getAlgo(const std::string& type) const { - if (type == "mkldnn_relu") { - return mkldnn::algorithm::eltwise_relu; - } else if (type == "mkldnn_tanh") { - return mkldnn::algorithm::eltwise_tanh; - } else if (type == "mkldnn_elu") { - return mkldnn::algorithm::eltwise_elu; - } else { - LOG(FATAL) << "Unkown eltwise activation type: " << type; - } - return (mkldnn::algorithm)0; - } - - /** - * reshape and reset the forward primitives - */ - void resetFwd(Argument& act) { - if (cnt_ == act.value->getElementCnt()) { - return; - } - VLOG(MKLDNN_BASE) << getName() << " reset mkldnn forward"; - cnt_ = act.value->getElementCnt(); - stream_.reset(new MKLDNNStream()); - auto eng = CPUEngine::Instance().getEngine(); - - // get algo setting - mkldnn::algorithm algo = getAlgo(this->getName()); - // note: alpha represents the NegativeSlope when used in relu. - float alpha = getAlpha(); - float beta = getBeta(); - - pipelineFwd_.clear(); - val_ = std::dynamic_pointer_cast(act.value); - if (val_ == nullptr) { - int bs = act.getBatchSize(); - int ih = act.getFrameHeight() > 0 ? act.getFrameHeight() : 1; - int iw = act.getFrameWidth() > 0 ? act.getFrameWidth() : 1; - int ic = cnt_ / bs / ih / iw; - CHECK_EQ(cnt_, (size_t)bs * ic * ih * iw); - val_ = MKLDNNMatrix::create( - act.value, {bs, ic, ih, iw}, mkldnn::memory::format::nchw, eng); - CHECK(val_); - } - auto fwdDesc = eltwise_fwd::desc(mkldnn::prop_kind::forward_training, - algo, - val_->getMemoryDesc(), - alpha, - beta); - fwdPD_.reset(new eltwise_fwd::primitive_desc(fwdDesc, eng)); - // use inplace for forward but save input value before submit - inVal_ = val_; - copyInVal_ = nullptr; - if (act.grad && algo == mkldnn::algorithm::eltwise_tanh) { - // tanh need save src input for backward - inVal_ = MKLDNNMatrix::create(nullptr, val_->getPrimitiveDesc()); - copyInVal_ = std::make_shared(*val_, *inVal_); - CHECK(copyInVal_) << "should not be emptry"; - pipelineFwd_.push_back(*copyInVal_); - } - fwd_.reset(new eltwise_fwd(*fwdPD_, *val_, *val_)); - pipelineFwd_.push_back(*fwd_); - needResetBwd_ = true; - } + virtual algorithm getAlgo(std::string type) const; + void resetFwd(Argument& act) override; + void resetBwd(Argument& act) override; +}; - /** - * reset the backward primitives, can not merge into resetFwd as the grad data - * would be changing before backward. - */ - void resetBwd(Argument& act) { - if (!needResetBwd_) { - return; - } - VLOG(MKLDNN_BASE) << getName() << " reset mkldnn backward"; - needResetBwd_ = false; - mkldnn::algorithm algo = getAlgo(this->getName()); - float alpha = getBwdAlpha(); - float beta = getBeta(); - grad_ = MKLDNNMatrix::create(act.grad, val_->getPrimitiveDesc()); - auto eng = CPUEngine::Instance().getEngine(); - auto bwdDesc = eltwise_bwd::desc( - algo, grad_->getMemoryDesc(), val_->getMemoryDesc(), alpha, beta); - auto bwdPD = eltwise_bwd::primitive_desc(bwdDesc, eng, *fwdPD_); - CHECK(inVal_); - bwd_.reset(new eltwise_bwd(bwdPD, *inVal_, *grad_, *grad_)); - pipelineBwd_.clear(); - pipelineBwd_.push_back(*bwd_); - } +/** + * @brief Base class of MKLDNN softmax Activation, + * only have mkldnn forward, use cpu implement for backward. + */ +class MKLDNNSoftmaxActivation : public MKLDNNActivation { + typedef mkldnn::softmax_forward softmax_fwd; - Error __must_check forward(Argument& act) { - resetFwd(act); - stream_->submit(pipelineFwd_); - return Error(); - } +private: + // for backward + MatrixPtr sftMaxSum_; + MatrixPtr sftMaxDot_; - Error __must_check backward(Argument& act) { - resetBwd(act); - stream_->submit(pipelineBwd_); - return Error(); - } +public: + MKLDNNSoftmaxActivation() {} + ~MKLDNNSoftmaxActivation() {} + virtual const std::string& getName() const = 0; + void resetFwd(Argument& act) override; + Error __must_check forward(Argument& act) override; + Error __must_check backward(Argument& act) override; }; } // namespace paddle diff --git a/paddle/gserver/gradientmachines/GradientMachine.cpp b/paddle/gserver/gradientmachines/GradientMachine.cpp index b44e4dc202f01956ed21c175aa897ced8e92546b..de5faf5e1e2b3e73bc07fe7f1635110f4efd7eec 100644 --- a/paddle/gserver/gradientmachines/GradientMachine.cpp +++ b/paddle/gserver/gradientmachines/GradientMachine.cpp @@ -17,12 +17,15 @@ limitations under the License. */ #include #include "paddle/utils/Logging.h" +#include "NeuralNetwork.h" +#include "hl_gpu.h" + +#ifndef PADDLE_MOBILE_INFERENCE #include "GradientMachineMode.h" #include "MultiGradientMachine.h" #include "MultiNetwork.h" -#include "NeuralNetwork.h" #include "ParallelNeuralNetwork.h" -#include "hl_gpu.h" +#endif namespace paddle { @@ -30,13 +33,16 @@ GradientMachine* GradientMachine::create( const ModelConfig& config, int mode, const std::vector& parameterTypes) { +#ifndef PADDLE_MOBILE_INFERENCE if (auto gm = IGradientMachineMode::tryCreateGradientMachine(mode, config)) { return gm; } if (FLAGS_trainer_count > 1) { return new MultiGradientMachine(config, FLAGS_use_gpu); } +#endif if (FLAGS_trainer_count == 1) { // single +#ifndef PADDLE_MOBILE_INFERENCE NeuralNetwork* nn; if (config.type() == "multi_nn") { /* multi submodel calculate, thread(s) will be initialized inside */ @@ -48,6 +54,9 @@ GradientMachine* GradientMachine::create( /* single thread calculate */ nn = NeuralNetwork::create(config); } +#else + NeuralNetwork* nn = NeuralNetwork::create(config); +#endif ParamInitCallback testParamInitCb = [](int paramId, Parameter* para) { para->enableType(PARAMETER_VALUE); }; diff --git a/paddle/gserver/gradientmachines/GradientMachine.h b/paddle/gserver/gradientmachines/GradientMachine.h index f9c82a2bef82b4e6bcbf0c73583505d2692f3926..ebfe0573cfdbfb2ef54a29b038e8b85356cc6c27 100644 --- a/paddle/gserver/gradientmachines/GradientMachine.h +++ b/paddle/gserver/gradientmachines/GradientMachine.h @@ -20,13 +20,16 @@ limitations under the License. */ #include "ModelConfig.pb.h" #include "TrainerConfig.pb.h" #include "paddle/gserver/dataproviders/DataProvider.h" -#include "paddle/gserver/evaluators/Evaluator.h" #include "paddle/gserver/layers/Layer.h" #include "paddle/math/Matrix.h" #include "paddle/parameter/Parameter.h" #include "paddle/parameter/ParameterUpdaterBase.h" #include "paddle/utils/Thread.h" +#ifndef PADDLE_MOBILE_INFERENCE +#include "paddle/gserver/evaluators/Evaluator.h" +#endif + namespace paddle { /** * @brief A gradient machine is capable of calculating some outputs given @@ -147,6 +150,7 @@ public: virtual void onPassEnd() = 0; +#ifndef PADDLE_MOBILE_INFERENCE /** * Create an evaluator which can be used for eval() */ @@ -156,6 +160,7 @@ public: * evaluate using the given evaluator */ virtual void eval(Evaluator* evaluator) const = 0; +#endif std::vector& getParameters() { return parameters_; } diff --git a/paddle/gserver/gradientmachines/NeuralNetwork.cpp b/paddle/gserver/gradientmachines/NeuralNetwork.cpp index 26cff3e67710b2f38d93572c5d58849aa94a5135..dcf0acb5a2cc7698625a2e254d4c12a32bc9631d 100644 --- a/paddle/gserver/gradientmachines/NeuralNetwork.cpp +++ b/paddle/gserver/gradientmachines/NeuralNetwork.cpp @@ -14,15 +14,17 @@ limitations under the License. */ #include "paddle/utils/Util.h" +#include "NeuralNetwork.h" +#include "hl_gpu.h" +#include "paddle/gserver/layers/AgentLayer.h" #include "paddle/utils/CustomStackTrace.h" #include "paddle/utils/Logging.h" +#include "paddle/utils/Stat.h" +#ifndef PADDLE_MOBILE_INFERENCE #include "MultiNetwork.h" -#include "NeuralNetwork.h" #include "RecurrentGradientMachine.h" -#include "hl_gpu.h" -#include "paddle/gserver/layers/AgentLayer.h" -#include "paddle/utils/Stat.h" +#endif namespace paddle { void parameterInitNN(int paramId, @@ -54,6 +56,7 @@ void parameterInitNN(int paramId, } NeuralNetwork* NeuralNetwork::create(const ModelConfig& config) { +#ifndef PADDLE_MOBILE_INFERENCE if (config.type() == "recurrent_nn") { return newNeuralNetwork("root"); } else if (config.type() == "multi_nn") { @@ -61,6 +64,9 @@ NeuralNetwork* NeuralNetwork::create(const ModelConfig& config) { } else { return newNeuralNetwork(); } +#else + return new NeuralNetwork(); +#endif } std::map NeuralNetwork::dllInitMap; @@ -304,6 +310,8 @@ void NeuralNetwork::onPassEnd() { } } +#ifndef PADDLE_MOBILE_INFERENCE + class CombinedEvaluator : public Evaluator { public: void addEvaluator(std::unique_ptr&& evaluator) { @@ -466,6 +474,8 @@ Evaluator* NeuralNetwork::makeEvaluator() const { void NeuralNetwork::eval(Evaluator* evaluator) const { evaluator->eval(*this); } +#endif + void NeuralNetwork::setOutputGrad(const std::vector& args) { CHECK_GE(outputLayers_.size(), args.size()); for (size_t i = 0; i < args.size(); ++i) { diff --git a/paddle/gserver/gradientmachines/NeuralNetwork.h b/paddle/gserver/gradientmachines/NeuralNetwork.h index 12810f642519b7965fc1b7d751290445e3350dd5..56a1ec78460731554c9b47cf3f517f7654dc314f 100644 --- a/paddle/gserver/gradientmachines/NeuralNetwork.h +++ b/paddle/gserver/gradientmachines/NeuralNetwork.h @@ -97,9 +97,12 @@ public: virtual void onPassEnd(); +#ifndef PADDLE_MOBILE_INFERENCE virtual Evaluator* makeEvaluator() const; virtual void eval(Evaluator* evaluator) const; +#endif + virtual void resetState(); virtual void setOutputGrad(const std::vector& args); diff --git a/paddle/gserver/layers/BatchNormBaseLayer.cpp b/paddle/gserver/layers/BatchNormBaseLayer.cpp index f7a80e23e1bd49549bec57b360587adc6b423794..bc7d1c83a48aefeb4bc6d3baa32b78aba712e58d 100644 --- a/paddle/gserver/layers/BatchNormBaseLayer.cpp +++ b/paddle/gserver/layers/BatchNormBaseLayer.cpp @@ -16,7 +16,7 @@ limitations under the License. */ #include "BatchNormalizationLayer.h" #include "Layer.h" #include "paddle/utils/Stat.h" -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA #include "CudnnBatchNormLayer.h" #endif diff --git a/paddle/gserver/layers/BatchNormalizationLayer.cpp b/paddle/gserver/layers/BatchNormalizationLayer.cpp index 412762d38475422be98ffeb87ffcfb028c3e035f..dacff25e5927daf9c991577a71be86b160228317 100644 --- a/paddle/gserver/layers/BatchNormalizationLayer.cpp +++ b/paddle/gserver/layers/BatchNormalizationLayer.cpp @@ -13,7 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/utils/Stat.h" -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA #include "hl_batch_transpose.h" #endif #include "BatchNormalizationLayer.h" @@ -90,7 +90,7 @@ void BatchNormalizationLayer::expandMat(const MatrixPtr& in, MatrixPtr& out) { size_t batchSize = in->getHeight(); CHECK_EQ(out->getHeight(), batchSize * imgPixels_); if (useGpu_) { -#ifdef PADDLE_ONLY_CPU +#ifndef PADDLE_WITH_CUDA LOG(FATAL) << "paddle is compiled only for cpu"; #else batchTranspose( @@ -127,7 +127,7 @@ void BatchNormalizationLayer::shrinkMat(const MatrixPtr& in, MatrixPtr& out) { } CHECK_EQ(in->getHeight(), static_cast(batchSize * imgPixels_)); if (useGpu_) { -#ifdef PADDLE_ONLY_CPU +#ifndef PADDLE_WITH_CUDA LOG(FATAL) << "paddle is compiled only for cpu"; #else batchTranspose( diff --git a/paddle/gserver/layers/Layer.cpp b/paddle/gserver/layers/Layer.cpp index e95f42c863b3733ca66055e1b3939e734cae8ad1..01f2aae6cf88d47296da804061b9b039cca593db 100644 --- a/paddle/gserver/layers/Layer.cpp +++ b/paddle/gserver/layers/Layer.cpp @@ -15,11 +15,14 @@ limitations under the License. */ #include "paddle/utils/Util.h" #include "CostLayer.h" -#include "ValidationLayer.h" #include "paddle/math/SparseMatrix.h" #include "paddle/utils/Error.h" #include "paddle/utils/Logging.h" +#ifndef PADDLE_MOBILE_INFERENCE +#include "ValidationLayer.h" +#endif + DEFINE_bool(log_error_clipping, false, "enable log error clipping or not"); namespace paddle { @@ -103,10 +106,12 @@ LayerPtr Layer::create(const LayerConfig& config) { return LayerPtr(new MultiClassCrossEntropy(config)); else if (type == "rank-cost") return LayerPtr(new RankingCost(config)); +#ifndef PADDLE_MOBILE_INFERENCE else if (type == "auc-validation") return LayerPtr(new AucValidation(config)); else if (type == "pnpair-validation") return LayerPtr(new PnpairValidation(config)); +#endif return LayerPtr(registrar_.createByType(config.type(), config)); } diff --git a/paddle/gserver/layers/MKLDNNConvLayer.cpp b/paddle/gserver/layers/MKLDNNConvLayer.cpp index 9a0abd291ae8fae43b0e95c7371f3ce35d1261ec..0d6742e909635c1097b4fe21bbb304f8a71af5cb 100644 --- a/paddle/gserver/layers/MKLDNNConvLayer.cpp +++ b/paddle/gserver/layers/MKLDNNConvLayer.cpp @@ -28,7 +28,7 @@ bool MKLDNNConvLayer::init(const LayerMap& layerMap, if (!MKLDNNLayer::init(layerMap, parameterMap)) { return false; } - CHECK_EQ(inputLayers_.size(), 1) << "Only support one input layer yet"; + CHECK_EQ(inputLayers_.size(), 1UL) << "Only support one input layer yet"; CHECK_EQ(inputLayers_.size(), parameters_.size()); CHECK(config_.shared_biases()) << "Only support shared biases yet"; diff --git a/paddle/gserver/layers/MKLDNNFcLayer.cpp b/paddle/gserver/layers/MKLDNNFcLayer.cpp index 8cbfbd0d2b9f2149f7c959aec5c4ae1de952f903..e829456d6afd7cc844f752d4571cd9f90c73997f 100644 --- a/paddle/gserver/layers/MKLDNNFcLayer.cpp +++ b/paddle/gserver/layers/MKLDNNFcLayer.cpp @@ -28,7 +28,7 @@ bool MKLDNNFcLayer::init(const LayerMap& layerMap, return false; } - CHECK_EQ(inputLayers_.size(), 1) << "Only support one input layer yet"; + CHECK_EQ(inputLayers_.size(), 1UL) << "Only support one input layer yet"; CHECK_EQ(inputLayers_.size(), parameters_.size()); CHECK(!parameters_[0]->isSparse()) << "Do not support sparse yet"; diff --git a/paddle/gserver/layers/PoolLayer.cpp b/paddle/gserver/layers/PoolLayer.cpp index 96d5c54accc047b685502a178de2d290f3158731..7b932d5a76e9c4fe7cbe5882bbc19eb3de4b503a 100644 --- a/paddle/gserver/layers/PoolLayer.cpp +++ b/paddle/gserver/layers/PoolLayer.cpp @@ -15,7 +15,7 @@ limitations under the License. */ #include "PoolLayer.h" #include "PoolProjectionLayer.h" #include "paddle/utils/Logging.h" -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA #include "CudnnPoolLayer.h" #endif namespace paddle { @@ -53,7 +53,7 @@ Layer* PoolLayer::create(const LayerConfig& config) { const std::string& pool = config.inputs(0).pool_conf().pool_type(); if (pool == "max-projection" || pool == "avg-projection") { return new PoolProjectionLayer(config); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA } else if (CudnnPoolLayer::typeCheck(pool)) { return new CudnnPoolLayer(config); #endif diff --git a/paddle/gserver/tests/CMakeLists.txt b/paddle/gserver/tests/CMakeLists.txt index de9b8e63dfc4291f8f42ca8c57cb5eb6baed8d8e..fcee19415c13e9731bd47eb53bbff9b52cf6450b 100644 --- a/paddle/gserver/tests/CMakeLists.txt +++ b/paddle/gserver/tests/CMakeLists.txt @@ -1,15 +1,17 @@ # gserver pacakge unittests +if(NOT MOBILE_INFERENCE) ################### test_ProtoDataProvider ############ -add_unittest_without_exec(test_ProtoDataProvider - test_ProtoDataProvider.cpp) - -# test_ProtoDataProvider will mkdir as same name, -# so if WORKING_DIRECTORY is default directory, then -# mkdir will get error. -add_test(NAME test_ProtoDataProvider - COMMAND ${CMAKE_CURRENT_BINARY_DIR}/test_ProtoDataProvider - WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle) + add_unittest_without_exec(test_ProtoDataProvider + test_ProtoDataProvider.cpp) + + # test_ProtoDataProvider will mkdir as same name, + # so if WORKING_DIRECTORY is default directory, then + # mkdir will get error. + add_test(NAME test_ProtoDataProvider + COMMAND ${CMAKE_CURRENT_BINARY_DIR}/test_ProtoDataProvider + WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle) +endif() ################# test_LayerGrad ####################### add_unittest_without_exec(test_LayerGrad @@ -98,9 +100,11 @@ add_unittest_without_exec(test_KmaxSeqScore add_test(NAME test_KmaxSeqScore COMMAND test_KmaxSeqScore) +if(NOT MOBILE_INFERENCE) ################## test_Evaluator ####################### -add_unittest(test_Evaluator - test_Evaluator.cpp) + add_unittest(test_Evaluator + test_Evaluator.cpp) +endif() ################ test_LinearChainCRF #################### add_simple_unittest(test_LinearChainCRF) @@ -131,27 +135,31 @@ if(NOT WITH_DOUBLE) WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle) endif() +if(NOT MOBILE_INFERENCE) ############### test_RecurrentGradientMachine ############### -# TODO(yuyang18): There is some bug in test_RecurrentGradientMachine -# I will fix it. -add_unittest_without_exec(test_RecurrentGradientMachine - test_RecurrentGradientMachine.cpp) -add_test(NAME test_RecurrentGradientMachine - COMMAND .set_python_path.sh -d - ${PADDLE_SOURCE_DIR}/python:${PADDLE_SOURCE_DIR}/paddle/gserver/tests - ${CMAKE_CURRENT_BINARY_DIR}/test_RecurrentGradientMachine - WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle) - -add_unittest_without_exec(test_NetworkCompare - test_NetworkCompare.cpp) -if(WITH_GPU) - add_test(NAME test_NetworkCompare - COMMAND .set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python ${CMAKE_CURRENT_BINARY_DIR}/test_NetworkCompare --use_gpu=true - WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle) -else() - add_test(NAME test_NetworkCompare - COMMAND .set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python ${CMAKE_CURRENT_BINARY_DIR}/test_NetworkCompare --use_gpu=false - WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle) + # TODO(yuyang18): There is some bug in test_RecurrentGradientMachine + # I will fix it. + add_unittest_without_exec(test_RecurrentGradientMachine + test_RecurrentGradientMachine.cpp) + add_test(NAME test_RecurrentGradientMachine + COMMAND .set_python_path.sh -d + ${PADDLE_SOURCE_DIR}/python:${PADDLE_SOURCE_DIR}/paddle/gserver/tests + ${CMAKE_CURRENT_BINARY_DIR}/test_RecurrentGradientMachine + WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle) +endif() + +if(NOT MOBILE_INFERENCE) + add_unittest_without_exec(test_NetworkCompare + test_NetworkCompare.cpp) + if(WITH_GPU) + add_test(NAME test_NetworkCompare + COMMAND .set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python ${CMAKE_CURRENT_BINARY_DIR}/test_NetworkCompare --use_gpu=true + WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle) + else() + add_test(NAME test_NetworkCompare + COMMAND .set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python ${CMAKE_CURRENT_BINARY_DIR}/test_NetworkCompare --use_gpu=false + WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle) + endif() endif() diff --git a/paddle/gserver/tests/LayerGradUtil.cpp b/paddle/gserver/tests/LayerGradUtil.cpp index a38880e14cdfcef05461dae567d198e5400c6bb1..cd957c7c0bca4c6089cc07e8f4226b8260190f07 100644 --- a/paddle/gserver/tests/LayerGradUtil.cpp +++ b/paddle/gserver/tests/LayerGradUtil.cpp @@ -674,7 +674,7 @@ void testLayerGradKernel(TestConfig testConf, bool useGpu, bool useWeight, float epsilon) { -#ifdef PADDLE_ONLY_CPU +#ifndef PADDLE_WITH_CUDA if (useGpu) return; #endif FLAGS_use_gpu = useGpu; diff --git a/paddle/gserver/tests/LayerGradUtil.h b/paddle/gserver/tests/LayerGradUtil.h index 88e831f78bd165f63806df6c081d84411be51502..e10a27eedfa3d207d77a9bf1c5bfb23480dcca69 100644 --- a/paddle/gserver/tests/LayerGradUtil.h +++ b/paddle/gserver/tests/LayerGradUtil.h @@ -15,7 +15,6 @@ limitations under the License. */ #pragma once #include "ModelConfig.pb.h" #include "paddle/gserver/layers/DataLayer.h" -#include "paddle/trainer/Trainer.h" #include "paddle/testing/TestUtil.h" using namespace std; // NOLINT diff --git a/paddle/gserver/tests/test_ActivationGrad.cpp b/paddle/gserver/tests/test_ActivationGrad.cpp index de93972a5880518dfbfb9f8582e17c594e54b9b8..f4c2a07c4426da36ff0b0570339a3a972dadec1f 100644 --- a/paddle/gserver/tests/test_ActivationGrad.cpp +++ b/paddle/gserver/tests/test_ActivationGrad.cpp @@ -17,7 +17,6 @@ limitations under the License. */ #include #include "ModelConfig.pb.h" #include "paddle/gserver/layers/DataLayer.h" -#include "paddle/trainer/Trainer.h" #include "LayerGradUtil.h" #include "paddle/testing/TestUtil.h" diff --git a/paddle/gserver/tests/test_BatchNorm.cpp b/paddle/gserver/tests/test_BatchNorm.cpp index 659eefa31bdb1f2433d03a59d5bf4782c71bdecf..41116f480957153eca33d211d09095903d6a00d9 100644 --- a/paddle/gserver/tests/test_BatchNorm.cpp +++ b/paddle/gserver/tests/test_BatchNorm.cpp @@ -17,7 +17,6 @@ limitations under the License. */ #include #include "ModelConfig.pb.h" #include "paddle/gserver/layers/DataLayer.h" -#include "paddle/trainer/Trainer.h" #include "paddle/utils/GlobalConstants.h" #include "LayerGradUtil.h" @@ -119,7 +118,7 @@ TEST(Layer, batchNorm) { CHECK_EQ(static_cast(convLayer->getOutputValue()->getWidth()), 576); } -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA void batchNormInference(int n, int c, int h, int w) { MatrixPtr input = std::make_shared(n, c * h * w); MatrixPtr cudnnOut = std::make_shared(n, c * h * w); diff --git a/paddle/gserver/tests/test_CRFLayerGrad.cpp b/paddle/gserver/tests/test_CRFLayerGrad.cpp index df14449291e9ec08f45718de07bbb101f6dbea58..f010066ebc6c33eff17715ba20b4e238583f1966 100644 --- a/paddle/gserver/tests/test_CRFLayerGrad.cpp +++ b/paddle/gserver/tests/test_CRFLayerGrad.cpp @@ -16,7 +16,6 @@ limitations under the License. */ #include "ModelConfig.pb.h" #include "paddle/gserver/layers/DataLayer.h" #include "paddle/gserver/layers/LinearChainCRF.h" -#include "paddle/trainer/Trainer.h" #include "LayerGradUtil.h" #include "paddle/testing/TestUtil.h" diff --git a/paddle/gserver/tests/test_ConvTrans.cpp b/paddle/gserver/tests/test_ConvTrans.cpp index 6035a866b4eee4c6a61fa93f3adbf5e1d2d549f7..5f2f9665478ad4bdfb00421ec57b3ecc1b41b417 100644 --- a/paddle/gserver/tests/test_ConvTrans.cpp +++ b/paddle/gserver/tests/test_ConvTrans.cpp @@ -18,7 +18,6 @@ limitations under the License. */ #include "ModelConfig.pb.h" #include "paddle/gserver/layers/DataLayer.h" #include "paddle/math/MathUtils.h" -#include "paddle/trainer/Trainer.h" #include "paddle/utils/GlobalConstants.h" #include "LayerGradUtil.h" diff --git a/paddle/gserver/tests/test_ConvUnify.cpp b/paddle/gserver/tests/test_ConvUnify.cpp index e7325e0cc3b7195b5fec77c878e3e087cfc643e0..8634355b5206f5cde0aa0717df50ade39e173ae7 100644 --- a/paddle/gserver/tests/test_ConvUnify.cpp +++ b/paddle/gserver/tests/test_ConvUnify.cpp @@ -18,7 +18,6 @@ limitations under the License. */ #include "ModelConfig.pb.h" #include "paddle/gserver/layers/DataLayer.h" #include "paddle/math/MathUtils.h" -#include "paddle/trainer/Trainer.h" #include "paddle/utils/GlobalConstants.h" #include "LayerGradUtil.h" @@ -117,7 +116,7 @@ MatrixPtr doOneConvTest(size_t imgSize, } TEST(Layer, convParaUnified) { -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA MatrixPtr input, resultCpu, resultGpu; /// TEST1 for conv /// diff --git a/paddle/gserver/tests/test_CrossEntropyOverBeamGrad.cpp b/paddle/gserver/tests/test_CrossEntropyOverBeamGrad.cpp index 538d18cdc3d262df0ddb031d9e6b38a3fea57606..477638426fe91f2c5b1f4d5011496385f07c2e90 100644 --- a/paddle/gserver/tests/test_CrossEntropyOverBeamGrad.cpp +++ b/paddle/gserver/tests/test_CrossEntropyOverBeamGrad.cpp @@ -18,7 +18,6 @@ limitations under the License. */ #include #include "ModelConfig.pb.h" #include "paddle/gserver/layers/DataLayer.h" -#include "paddle/trainer/Trainer.h" #include "LayerGradUtil.h" #include "paddle/testing/TestUtil.h" @@ -228,7 +227,7 @@ void genGroundTruth(vector& beamExpansions, curBeam.groundTruth[j] = *(start + n); curBeam.inBeam[j] = 1; } else { - CHECK_LE(curBeam.rowIdxInBeam[j] + 1, + CHECK_LE((size_t)curBeam.rowIdxInBeam[j] + 1, curBeam.subSeqStartPos.size() - 1); int start = curBeam.subSeqStartPos[curBeam.rowIdxInBeam[j]]; int end = curBeam.subSeqStartPos[curBeam.rowIdxInBeam[j] + 1]; diff --git a/paddle/gserver/tests/test_DetectionOutput.cpp b/paddle/gserver/tests/test_DetectionOutput.cpp index af43dc51fad35c834635b543b1a016f6d717de1e..dc39c97a87f8b346dc9cc09d6158b1b4069bcf2d 100644 --- a/paddle/gserver/tests/test_DetectionOutput.cpp +++ b/paddle/gserver/tests/test_DetectionOutput.cpp @@ -150,7 +150,7 @@ TEST(Layer, detectionOutputLayerFwd) { useGpu, result2); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA // GPU case 1. useGpu = true; inputLoc = Matrix::create(1, 16, false, useGpu); diff --git a/paddle/gserver/tests/test_Evaluator.cpp b/paddle/gserver/tests/test_Evaluator.cpp index 93996392d221d531f65caf465decbffdbc2d0384..62a131171fa5ae973cb3069151a582aaeac9ee0e 100644 --- a/paddle/gserver/tests/test_Evaluator.cpp +++ b/paddle/gserver/tests/test_Evaluator.cpp @@ -51,7 +51,7 @@ void testEvaluator(TestConfig testConf, string testEvaluatorName, size_t batchSize, bool useGpu) { -#ifdef PADDLE_ONLY_CPU +#ifndef PADDLE_WITH_CUDA if (useGpu) return; #endif FLAGS_use_gpu = useGpu; diff --git a/paddle/gserver/tests/test_KmaxSeqScore.cpp b/paddle/gserver/tests/test_KmaxSeqScore.cpp index 308abe6816428bc0f98ec32e892622fa4a23b1ae..ffe5cfb8dbb55d0b70a5699969abaa101f05f9ce 100644 --- a/paddle/gserver/tests/test_KmaxSeqScore.cpp +++ b/paddle/gserver/tests/test_KmaxSeqScore.cpp @@ -18,7 +18,6 @@ limitations under the License. */ #include #include "ModelConfig.pb.h" #include "paddle/gserver/layers/DataLayer.h" -#include "paddle/trainer/Trainer.h" #include "paddle/utils/GlobalConstants.h" #include "LayerGradUtil.h" @@ -97,7 +96,7 @@ TEST(Layer, kmaxSeqScoreLayer) { Matrix::create(subSeqStartPosition.back(), 1, false, false); std::vector mode = {false}; -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA mode.push_back(true); #endif diff --git a/paddle/gserver/tests/test_LayerGrad.cpp b/paddle/gserver/tests/test_LayerGrad.cpp index 090bde7b203652e3ffb1662b8f5b8937885d2608..1a46fb49153a0aa4228f58db481b950bc2d6de83 100644 --- a/paddle/gserver/tests/test_LayerGrad.cpp +++ b/paddle/gserver/tests/test_LayerGrad.cpp @@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA #include #endif #include @@ -21,7 +21,6 @@ limitations under the License. */ #include "ModelConfig.pb.h" #include "paddle/gserver/layers/DataLayer.h" #include "paddle/math/MathUtils.h" -#include "paddle/trainer/Trainer.h" #include "LayerGradUtil.h" #include "paddle/testing/TestUtil.h" @@ -258,7 +257,7 @@ void testProjectionConv(size_t groups, bool isDeconv) { true); } -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA TEST(Projection, conv) { /// test ConvProjection testProjectionConv(1, false); @@ -422,7 +421,7 @@ TEST(Layer, depthwiseConvLayer) { // 'depthwise_conv' is a sepecial case of 'exconv' whose // groups size equals to the input channels size. testDepthwiseConvLayer("exconv", /* useGpu= */ false); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA testDepthwiseConvLayer("exconv", /* useGpu= */ true); #endif } @@ -480,7 +479,7 @@ void testConvLayer(const string& type, bool trans, bool useGpu) { TEST(Layer, convLayer) { testConvLayer("exconv", /* trans= */ false, /* useGpu= */ false); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA testConvLayer("exconv", /* trans= */ false, /* useGpu= */ true); testConvLayer("cudnn_conv", /* trans= */ false, /* useGpu= */ true); #endif @@ -525,7 +524,7 @@ TEST(Layer, convTransLayer) { for (auto useGpu : {false, true}) { testConvTransLayer("exconvt", /* trans= */ false, /* useGpu= */ useGpu); } -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA testConvTransLayer("cudnn_convt", /* trans= */ false, /* useGpu= */ true); #endif } @@ -638,7 +637,7 @@ TEST(Layer, SelectiveFullyConnectedLayer) { /* trans= */ false, /* useGup= */ false, false); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA testLayerGrad(config, "selective_fc", 100, @@ -1210,7 +1209,7 @@ void testPoolLayer(const string& poolType, bool trans, bool useGpu) { testLayerGrad(config, "pool", 100, trans, useGpu); } -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA void testPoolLayer2(const string& poolType, bool trans, bool useGpu) { TestConfig config; config.inputDefs.push_back({INPUT_DATA, "layer_0", 3200, 0}); @@ -1236,7 +1235,7 @@ TEST(Layer, PoolLayer) { testPoolLayer("avg-projection", /* trans= */ false, /* useGpu= */ false); testPoolLayer("max-projection", /* trans= */ false, /* useGpu= */ false); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA testPoolLayer("avg-projection", /* trans= */ false, /* useGpu= */ true); testPoolLayer("max-projection", /* trans= */ false, /* useGpu= */ true); testPoolLayer("cudnn-max-pool", /* trans= */ false, /* useGpu= */ true); @@ -1309,7 +1308,7 @@ void testPool3DLayer(const string& poolType, bool trans, bool useGpu) { TEST(Layer, Pool3DLayer) { testPool3DLayer("avg", /* trans= */ false, /* useGpu= */ false); testPool3DLayer("max", /* trans= */ false, /* useGpu= */ false); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA testPool3DLayer("avg", /* trans= */ false, /* useGpu= */ true); testPool3DLayer("max", /* trans= */ false, /* useGpu= */ true); #endif @@ -1695,7 +1694,7 @@ void testBatchNormLayer(const string& type, bool trans, bool useGpu) { TEST(Layer, BatchNormalizationLayer) { testBatchNormLayer("batch_norm", false, false); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA testBatchNormLayer("batch_norm", false, true); if (hl_get_cudnn_lib_version() >= int(4000)) { testBatchNormLayer("cudnn_batch_norm", false, true); @@ -1744,7 +1743,7 @@ void testBatchNorm3DLayer(const string& type, bool trans, bool useGpu) { TEST(Layer, testBatchNorm3DLayer) { testBatchNorm3DLayer("batch_norm", false, false); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA testBatchNorm3DLayer("batch_norm", false, true); if (hl_get_cudnn_lib_version() >= int(4000)) { testBatchNorm3DLayer("cudnn_batch_norm", false, true); @@ -2262,7 +2261,7 @@ void test3DConvLayer(const string& type, bool trans, bool useGpu) { TEST(Layer, test3DConvLayer) { test3DConvLayer("conv3d", /* trans= */ false, /* useGpu= */ false); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA test3DConvLayer("conv3d", /* trans= */ false, /* useGpu= */ true); #endif } @@ -2339,7 +2338,7 @@ void test3DDeConvLayer(const string& type, bool trans, bool useGpu) { TEST(Layer, test3DDeConvLayer) { test3DDeConvLayer("deconv3d", /* trans= */ false, /* useGpu= */ false); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA test3DDeConvLayer("deconv3d", /* trans= */ false, /* useGpu= */ true); #endif } diff --git a/paddle/gserver/tests/test_MKLDNN.cpp b/paddle/gserver/tests/test_MKLDNN.cpp index 1bfbbde4246a10eaf86693a6a2f237f390966db3..a70b2f17f4f1130322f3c50d244f70fdcf34468b 100644 --- a/paddle/gserver/tests/test_MKLDNN.cpp +++ b/paddle/gserver/tests/test_MKLDNN.cpp @@ -215,15 +215,15 @@ struct testActDesc { static void getAddtoConfig(TestConfig& cfg, const testActDesc& pm) { cfg.biasSize = 0; cfg.layerConfig.set_type("addto"); - size_t layerSize = pm.ih * pm.ih * pm.iw; + size_t layerSize = pm.ic * pm.ih * pm.iw; cfg.layerConfig.set_size(layerSize); cfg.inputDefs.push_back({INPUT_DATA, "layer_0", layerSize, 0}); cfg.layerConfig.add_inputs(); } -void testActivation(std::string& actType, const testActDesc& pm) { - // TODO(TJ): mkldnn_softmax not implemented, paddle do not have elu activation - if (actType == "mkldnn_softmax" || actType == "mkldnn_elu") { +void testActivation(std::string actType, const testActDesc& pm) { + // TODO(TJ): remove me when paddle support elu activation + if (actType == "mkldnn_elu") { return; } const std::string compareTypes[] = {actType, actType.erase(0, 7)}; @@ -240,6 +240,7 @@ TEST(MKLDNNActivation, Activations) { for (auto type : types) { /* bs, c, h, w*/ testActivation(type, {16, 64, 32, 32}); + testActivation(type, {2, 8, 1, 1}); } } diff --git a/paddle/gserver/tests/test_NetworkCompare.cpp b/paddle/gserver/tests/test_NetworkCompare.cpp index d36f72360f8ebd2033fb3e8c0e1b30911abba362..2b92211936aad1a034369bda0830bed3438cf401 100644 --- a/paddle/gserver/tests/test_NetworkCompare.cpp +++ b/paddle/gserver/tests/test_NetworkCompare.cpp @@ -243,7 +243,7 @@ TEST(Compare, concat_slice) { compareNetwork(config_file_a, config_file_b); } -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA TEST(Compare, img_pool) { std::string config_file_a = "./gserver/tests/img_pool_a.conf"; std::string config_file_b = "./gserver/tests/img_pool_b.conf"; diff --git a/paddle/gserver/tests/test_PriorBox.cpp b/paddle/gserver/tests/test_PriorBox.cpp index ae0e3bc3d24c54eb84c7b5f5053e629607ef4310..8dc5568784295b5a2e7d4decd178d612432a1a18 100644 --- a/paddle/gserver/tests/test_PriorBox.cpp +++ b/paddle/gserver/tests/test_PriorBox.cpp @@ -151,7 +151,7 @@ TEST(Layer, priorBoxLayerFwd) { useGpu, result); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA // reset the input parameters variance[1] = 0.1; variance[3] = 0.2; diff --git a/paddle/gserver/tests/test_ProtoDataProvider.cpp b/paddle/gserver/tests/test_ProtoDataProvider.cpp index e11bf402c27898b8fdbd3fceeb8aeff8906352db..af6472619d1840e82787974d265d601b4a406c09 100644 --- a/paddle/gserver/tests/test_ProtoDataProvider.cpp +++ b/paddle/gserver/tests/test_ProtoDataProvider.cpp @@ -485,7 +485,7 @@ TEST(ProtoDataProvider, test) { // Currently in async mode, useGpu is not supported continue; } -#ifdef PADDLE_ONLY_CPU +#ifndef PADDLE_WITH_CUDA if (useGpu) { continue; } @@ -525,7 +525,7 @@ TEST(ProtoDataProvider, constant_slots) { for (int numConstantSlots : {1, 2}) { for (int useGpu : numTwoArray) { for (int dataCompression : numTwoArray) { -#ifdef PADDLE_ONLY_CPU +#ifndef PADDLE_WITH_CUDA if (useGpu) { continue; } @@ -708,7 +708,7 @@ TEST(ProtoSequenceDataProvider, test) { // Currently in async mode, useGpu is not supported continue; } -#ifdef PADDLE_ONLY_CPU +#ifndef PADDLE_WITH_CUDA if (useGpu) { continue; } diff --git a/paddle/gserver/tests/test_PyDataProvider.cpp b/paddle/gserver/tests/test_PyDataProvider.cpp index db883543c306c1938eb9da188ce20ed768018efb..fe54799259d86064c4fcaec0e53707247981a1b4 100644 --- a/paddle/gserver/tests/test_PyDataProvider.cpp +++ b/paddle/gserver/tests/test_PyDataProvider.cpp @@ -37,7 +37,7 @@ TEST(PyDataProvider, py_fill_slots) { config.clear_files(); std::string dataFile = "gserver/tests/pyDataProvider/pyDataProviderList"; config.set_files(dataFile); -#ifdef PADDLE_ONLY_CPU +#ifndef PADDLE_WITH_CUDA bool useGpu = false; #else bool useGpu = true; @@ -71,7 +71,7 @@ TEST(PyDataProvider, py_fill_nest_slots) { std::string dataFile = "gserver/tests/pyDataProvider/pyDataProviderList"; config.set_files(dataFile); EXPECT_EQ(config.IsInitialized(), true); -#ifdef PADDLE_ONLY_CPU +#ifndef PADDLE_WITH_CUDA bool useGpu = false; #else bool useGpu = true; diff --git a/paddle/gserver/tests/test_SelectiveFCLayer.cpp b/paddle/gserver/tests/test_SelectiveFCLayer.cpp index ab23d00a2cb6077147f5b89664a8e2437b4cd63b..d164e382c4a804aef2417135b64cf709474d12f1 100644 --- a/paddle/gserver/tests/test_SelectiveFCLayer.cpp +++ b/paddle/gserver/tests/test_SelectiveFCLayer.cpp @@ -24,7 +24,6 @@ limitations under the License. */ #include "paddle/gserver/layers/Layer.h" #include "paddle/gserver/layers/SelectiveFullyConnectedLayer.h" #include "paddle/math/CpuSparseMatrix.h" -#include "paddle/trainer/Trainer.h" using namespace paddle; // NOLINT using namespace std; // NOLINT @@ -321,7 +320,7 @@ TEST(Layer, SelectiveFcLayer_train_dense_mul) { "filelist=gserver/tests/SelectiveFcTest/dense_mul_list"; for (auto useGpu : {false, true}) { -#ifdef PADDLE_ONLY_CPU +#ifndef PADDLE_WITH_CUDA if (useGpu) { break; } @@ -388,7 +387,7 @@ void testSelectiveFcLayerTrainSparseMul(const LayerConfig& config, outMatSelfc->getWidth(), outMatSelfc->getElementCnt())); cpuOutMatSelfc->copyFrom(*outMatSelfc, HPPL_STREAM_DEFAULT); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA if (useGpu) { hl_stream_synchronize(HPPL_STREAM_DEFAULT); } @@ -418,7 +417,7 @@ void testSelectiveFcLayerTrainSparseMul(const LayerConfig& config, MatrixPtr cpuOutMatFc( new CpuMatrix(outMatFc->getHeight(), outMatFc->getWidth())); cpuOutMatFc->copyFrom(*outMatFc, HPPL_STREAM_DEFAULT); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA if (useGpu) { hl_stream_synchronize(HPPL_STREAM_DEFAULT); } @@ -443,7 +442,7 @@ TEST(Layer, SelectiveFcLayer_train_sparse_mul) { selLayerConfig.set_size(fcLayerWidth); testSelectiveFcLayerTrainSparseMul(selLayerConfig, false); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA testSelectiveFcLayerTrainSparseMul(selLayerConfig, true); #endif } diff --git a/paddle/gserver/tests/test_SeqSliceLayerGrad.cpp b/paddle/gserver/tests/test_SeqSliceLayerGrad.cpp index e1d4ae16176433b898ba88dd60550e44b4fe37be..3dbffc563462973bdc1da529d486b2a2d5a677d3 100644 --- a/paddle/gserver/tests/test_SeqSliceLayerGrad.cpp +++ b/paddle/gserver/tests/test_SeqSliceLayerGrad.cpp @@ -15,7 +15,6 @@ limitations under the License. */ #include #include "ModelConfig.pb.h" #include "paddle/gserver/layers/DataLayer.h" -#include "paddle/trainer/Trainer.h" #include "LayerGradUtil.h" #include "paddle/testing/TestUtil.h" @@ -195,7 +194,7 @@ TEST(Layer, SeqSliceLayer) { vector> ends; std::vector mode = {false}; -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA mode.push_back(true); #endif genSeqInfo(seqStartPos, subSeqStartPos); diff --git a/paddle/gserver/tests/test_WarpCTCLayer.cpp b/paddle/gserver/tests/test_WarpCTCLayer.cpp index 55427e2f12fd7b77c6eea1f65b3229e6fd29d71d..da829460061d38f363317e33daeb65cfa705bb8e 100644 --- a/paddle/gserver/tests/test_WarpCTCLayer.cpp +++ b/paddle/gserver/tests/test_WarpCTCLayer.cpp @@ -199,7 +199,7 @@ TEST(Layer, WarpCTCLayer) { for (auto batchSize : {1, 10, 32}) { for (auto normByTimes : {false, true}) { for (auto useGpu : {false, true}) { -#ifdef PADDLE_ONLY_CPU +#ifndef PADDLE_WITH_CUDA if (useGpu) continue; #endif LOG(INFO) << "layerSize=" << layerSize << " batchSize=" << batchSize diff --git a/paddle/math/Matrix.cpp b/paddle/math/Matrix.cpp index 0023b4d0f5da500f380ecb836b7c54e050b13d67..c3e34d5309d9ca8a32d7b0a8043e668cdb5be54b 100644 --- a/paddle/math/Matrix.cpp +++ b/paddle/math/Matrix.cpp @@ -670,7 +670,7 @@ void GpuMatrix::leftMul(Matrix& a, real scaleAB, real scaleT) { } void GpuMatrix::selectRows(Matrix& table, IVector& ids) { -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA CHECK(dynamic_cast(&table)); CHECK(table.useGpu()); CHECK(ids.useGpu()); @@ -694,7 +694,7 @@ void GpuMatrix::selectRows(Matrix& table, IVector& ids) { } void GpuMatrix::addToRows(Matrix& table, IVector& ids) { -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA CHECK(dynamic_cast(&table)); CHECK(table.useGpu()); CHECK(ids.useGpu()); @@ -741,7 +741,7 @@ void GpuMatrix::rowMax(Matrix& max) { } void GpuMatrix::rowMax(IVector& maxIds, Matrix& maxVal) { -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA CHECK(maxIds.useGpu() && maxVal.useGpu()) << "Matrix type are not equal"; size_t numSamples = getHeight(); size_t beam = maxVal.getWidth(); diff --git a/paddle/math/RowBuffer.h b/paddle/math/RowBuffer.h index dbb829c4e24a659e4a97c0a3ba4c5c78b68815d3..9ef5b89680b00981188d78cb312dc75e2c0a79ee 100644 --- a/paddle/math/RowBuffer.h +++ b/paddle/math/RowBuffer.h @@ -99,7 +99,11 @@ public: /** * @brief clear local buffer. It only affect auto-growth buffer. */ - inline void clear() { rowStore_.clear(); } + inline void clear() { + // swap an empty vector to it to free the memory. + std::vector> empty; + rowStore_.swap(empty); + } /** * @brief get current number of rows. diff --git a/paddle/math/SparseMatrix.cpp b/paddle/math/SparseMatrix.cpp index 6370c77386688a334fa0de8b4e2b272882e9e2b0..284b68d590ba655395c0186d8ea86d6855c6fc50 100644 --- a/paddle/math/SparseMatrix.cpp +++ b/paddle/math/SparseMatrix.cpp @@ -836,7 +836,7 @@ void GpuSparseMatrix::zeroMem() { } void GpuSparseMatrix::rowMax(IVector& maxIds, Matrix& maxVal) { -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA CHECK(maxIds.useGpu() && maxVal.useGpu()) << "Matrix type are not equal"; size_t numSamples = getHeight(); size_t beam = maxVal.getWidth(); diff --git a/paddle/math/Vector.cpp b/paddle/math/Vector.cpp index eb87ee9bb7936d27c0c32a1a4b35ff49871c0a10..ff72672e3ab77212b309fcfea835839a916fa632 100644 --- a/paddle/math/Vector.cpp +++ b/paddle/math/Vector.cpp @@ -172,7 +172,7 @@ void GpuVectorT::isEqualTo(const VectorT& b, const T& value) { template void GpuVectorT::selectFrom(const VectorT& src, const VectorT& ids) { -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA hl_vector_select_from(this->getData(), this->getSize(), src.getData(), @@ -850,7 +850,7 @@ CpuGpuVectorT::CpuGpuVectorT(CpuGpuVectorT& src, size_t size) : sync_(nullptr) { CHECK_LE(offset + size, static_cast(src.getSize())); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA SyncedFlag* flag = src.getSync(); if (*flag == DATA_AT_CPU) { src.copyToGpu(); // will set synchronous data between CPU and GPU @@ -861,7 +861,7 @@ CpuGpuVectorT::CpuGpuVectorT(CpuGpuVectorT& src, auto cMemHandle = (src.getVector(false))->getMemoryHandle(); cpuVectorT_ = std::make_shared>( size, std::dynamic_pointer_cast(cMemHandle), offset); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA auto gMemHandle = (src.getVector(true))->getMemoryHandle(); gpuVectorT_ = std::make_shared>( size, std::dynamic_pointer_cast(gMemHandle), offset); diff --git a/paddle/math/tests/test_Allocator.cpp b/paddle/math/tests/test_Allocator.cpp index 1ca70ea84c867b83013625eaee141f5b75fad4ae..1fecf659e5080c7d25f5f76b92b15f75eaab6ce3 100644 --- a/paddle/math/tests/test_Allocator.cpp +++ b/paddle/math/tests/test_Allocator.cpp @@ -68,7 +68,7 @@ void testPoolAllocator() { TEST(Allocator, Pool) { testPoolAllocator(); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA testPoolAllocator(); #endif } @@ -92,7 +92,7 @@ TEST(MemoryHandle, Cpu) { EXPECT_EQ(ptr1, ptr2); } -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA TEST(MemoryHandle, Gpu) { int numGpu = hl_get_device_count(); diff --git a/paddle/math/tests/test_BaseMatrix.cpp b/paddle/math/tests/test_BaseMatrix.cpp index 22ce39701fca7b650fc03794cb0701e0987d2dae..1766257860b0b13e9f0ce898438e7c2d644f545e 100644 --- a/paddle/math/tests/test_BaseMatrix.cpp +++ b/paddle/math/tests/test_BaseMatrix.cpp @@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA /** * This test file use autotest::AutoCompare and cmpWithoutArg to compares the * implementation of CPU and GPU member function in diff --git a/paddle/math/tests/test_CpuGpuVector.cpp b/paddle/math/tests/test_CpuGpuVector.cpp index 58bc43a38ba9465a832fcd0652e6309c403577e3..c72f89c8244b1209e490b09387c2ee6352426ce1 100644 --- a/paddle/math/tests/test_CpuGpuVector.cpp +++ b/paddle/math/tests/test_CpuGpuVector.cpp @@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA #include #include "paddle/math/Vector.h" diff --git a/paddle/math/tests/test_ExecViaCpu.cpp b/paddle/math/tests/test_ExecViaCpu.cpp index 04c856453d2ec4ad764e37ae430e3e30ac0dea0b..25e0ba11ded96dd78aedc3c297507d0555d80d74 100644 --- a/paddle/math/tests/test_ExecViaCpu.cpp +++ b/paddle/math/tests/test_ExecViaCpu.cpp @@ -94,7 +94,7 @@ void testWrapper(F&& f) { } } -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA TEST(ExecViaCpu, test1) { testWrapper(f); testWrapper(&f); diff --git a/paddle/math/tests/test_GpuProfiler.cpp b/paddle/math/tests/test_GpuProfiler.cpp index e6b5dba446b5a0022ade76b188895c4e0e2a22b4..d9f146f0d1f63480ddee784071b43ff85da0b15c 100644 --- a/paddle/math/tests/test_GpuProfiler.cpp +++ b/paddle/math/tests/test_GpuProfiler.cpp @@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA #include #include "paddle/math/Matrix.h" @@ -162,4 +162,4 @@ int main(int argc, char** argv) { return RUN_ALL_TESTS(); } -#endif /* PADDLE_ONLY_CPU */ +#endif diff --git a/paddle/math/tests/test_Matrix.cpp b/paddle/math/tests/test_Matrix.cpp index 1c21da5b76e95603258a5006d0c57b00126e65b9..2f99fa3581e14b91acc0b294856619f4ae2b3483 100644 --- a/paddle/math/tests/test_Matrix.cpp +++ b/paddle/math/tests/test_Matrix.cpp @@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA /** * This test file use autotest::AutoCompare and cmpWithArg to compares the * implementation of CPU and GPU member function in Matrix.cpp. diff --git a/paddle/math/tests/test_SparseMatrix.cpp b/paddle/math/tests/test_SparseMatrix.cpp index c0572dfdbf738a4dfad04811b3a3e1b65487ff6d..8abbe8d82e02b7d1738fe7e6d0c8d494166e7892 100644 --- a/paddle/math/tests/test_SparseMatrix.cpp +++ b/paddle/math/tests/test_SparseMatrix.cpp @@ -47,7 +47,7 @@ struct MatrixPara { SparseFormat format; }; -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA void test_sparse_matrix_mul(MatrixPara paraA, MatrixPara paraB, MatrixPara paraC) { @@ -452,7 +452,7 @@ TEST(Matrix, SparseMatrixCSRFormatTrimFrom) { matB->trimFrom(*mat); checkSMatrixEqual2(matA, matB); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA GpuSparseMatrixPtr matC = std::make_shared( height, trimedWidth, height, FLOAT_VALUE, SPARSE_CSR, true); matC->trimFrom(*mat); @@ -546,7 +546,7 @@ TEST(Matrix, SparseMatrixCSCFormatTrimFrom) { matB->trimFrom(*mat); checkSMatrixEqual2(matA, matB); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA GpuSparseMatrixPtr matC = std::make_shared( height, trimedWidth, height, FLOAT_VALUE, SPARSE_CSC, true); matC->trimFrom(*mat); diff --git a/paddle/math/tests/test_Tensor.cu b/paddle/math/tests/test_Tensor.cu index 31b693afa8bd50f77a8efb67769e6215dd755bd3..d03698dee25fdd6dd49f2a3fdb5c605333440f49 100644 --- a/paddle/math/tests/test_Tensor.cu +++ b/paddle/math/tests/test_Tensor.cu @@ -270,7 +270,7 @@ TEST(Unary, BaseOp) { TestUnaryVectorT testCpuIVector( testUnaryBaseOpInt); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_GPU TestUnaryMatrix testGpuMatrix(testUnaryBaseOp); TestUnaryVectorT testGpuVector(testUnaryBaseOp); TestUnaryVectorT testGpuIVector( @@ -317,7 +317,7 @@ void testUnayrMathOp(Tensor& A1, Tensor& A2) { TEST(Unary, MathOp) { TestUnaryMatrix testCpu(testUnayrMathOp); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_GPU TestUnaryMatrix testGpu(testUnayrMathOp); #endif } @@ -374,7 +374,7 @@ void testUnayrCompareOp(Tensor& A1, Tensor& A2) { TEST(Unary, CompareOp) { TestUnaryMatrix testCpu(testUnayrCompareOp); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_GPU TestUnaryMatrix testGpu(testUnayrCompareOp); #endif } @@ -536,7 +536,7 @@ void testBinaryBaseOp(Tensor& A1, Tensor& A2, Tensor& B) { TEST(Binary, BaseOp) { TestBinaryMatrix testCpu(testBinaryBaseOp); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_GPU TestBinaryMatrix testGpu(testBinaryBaseOp); #endif } @@ -710,7 +710,7 @@ void testBinaryMathOp(Tensor& A1, Tensor& A2, Tensor& B) { TEST(Binary, MathOp) { TestBinaryMatrix testCpu(testBinaryMathOp); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_GPU TestBinaryMatrix testGpu(testBinaryMathOp); #endif } @@ -810,7 +810,7 @@ void testBinaryCompareOp(Tensor& A1, Tensor& A2, Tensor& B) { TEST(Binary, CompareOp) { TestBinaryMatrix testCpu(testBinaryCompareOp); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_GPU TestBinaryMatrix testGpu(testBinaryCompareOp); #endif } @@ -955,7 +955,7 @@ void testTernaryBaseOp(Tensor& A1, Tensor& A2, Tensor& B, Tensor& C) { TEST(Ternary, BaseOp) { TestTernaryMatrix testCpu(testTernaryBaseOp); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_GPU TestTernaryMatrix testGpu(testTernaryBaseOp); #endif } @@ -1058,7 +1058,7 @@ void testTernaryCompareOp(Tensor& A1, Tensor& A2, Tensor& B, Tensor& C) { TEST(Ternary, CompareOp) { TestTernaryMatrix testCpu(testTernaryCompareOp); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_GPU TestTernaryMatrix testGpu(testTernaryCompareOp); #endif } @@ -1086,7 +1086,7 @@ void testQuaternaryAdd( TEST(Quaternary, BaseOp) { TestQuaternaryMatrix testCpu(testQuaternaryAdd); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_GPU TestQuaternaryMatrix testGpu(testQuaternaryAdd); #endif } @@ -1156,7 +1156,7 @@ void testQuaternaryCompareOp( TEST(Quaternary, CompareOp) { TestQuaternaryMatrix testCpu(testQuaternaryCompareOp); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_GPU TestQuaternaryMatrix testGpu(testQuaternaryCompareOp); #endif } diff --git a/paddle/math/tests/test_TrainingAlgorithm.cpp b/paddle/math/tests/test_TrainingAlgorithm.cpp index 4a88844b43ef40af988d2b391d2bef4568dea9b7..5ae0aa036f6bfc1e5bd4e955277c4efff8c739ce 100644 --- a/paddle/math/tests/test_TrainingAlgorithm.cpp +++ b/paddle/math/tests/test_TrainingAlgorithm.cpp @@ -91,7 +91,7 @@ int VectorCheckErr(const VectorPtr& vector1, const VectorPtr& vector2) { typedef std::function testMatrixFunc; void testCase(testMatrixFunc matrixFunc) { -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA for (auto useGpu : {false, true}) { #else for (auto useGpu : {false}) { diff --git a/paddle/math/tests/test_batchTranspose.cpp b/paddle/math/tests/test_batchTranspose.cpp index 4eb9837909ffaaf0f483ab65ece7a0b29fd49319..b70a61976402fd0a7cfee8382fd926fcf28486d5 100644 --- a/paddle/math/tests/test_batchTranspose.cpp +++ b/paddle/math/tests/test_batchTranspose.cpp @@ -17,7 +17,7 @@ limitations under the License. */ using namespace paddle; // NOLINT -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA TEST(MatrixBatchTransTest, test_batch_matrix_transpose) { const int nx = 100; const int ny = 50; diff --git a/paddle/math/tests/test_lazyAssign.cu b/paddle/math/tests/test_lazyAssign.cu index 92afab4ff7f5ff4acc219c5ac783733340c5726a..04f23cff55db45c39049538545430bc5996cce5d 100644 --- a/paddle/math/tests/test_lazyAssign.cu +++ b/paddle/math/tests/test_lazyAssign.cu @@ -72,7 +72,7 @@ void testLazyAssign(int height, int width) { TEST(lazyAssign, CPU) { testMatrixCase(testLazyAssign); } -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_GPU TEST(lazyAssign, GPU) { testMatrixCase(testLazyAssign); } #endif @@ -142,6 +142,6 @@ void testSgdUpdate(int height, int width) { TEST(sgdUpdate, CPU) { testMatrixCase(testSgdUpdate); } -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_GPU TEST(sgdUpdate, GPU) { testMatrixCase(testSgdUpdate); } #endif diff --git a/paddle/math/tests/test_matrixCompare.cpp b/paddle/math/tests/test_matrixCompare.cpp index 061fb22e3fd744d9d9895fd1008089e4a6ce6a0f..7e5a1db44a5302e3b4e5d2768755824666e880ba 100644 --- a/paddle/math/tests/test_matrixCompare.cpp +++ b/paddle/math/tests/test_matrixCompare.cpp @@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA /// This unittest checks GpuMatrix/CpuMatrix get same result, so disable when /// only cpu version. diff --git a/paddle/math/tests/test_perturbation.cpp b/paddle/math/tests/test_perturbation.cpp index 60ebae015381a3901c14d0cd4c1225e54ac5726f..c7c07c817a08d78ddcbf8218e8c4a9d22f4990bc 100644 --- a/paddle/math/tests/test_perturbation.cpp +++ b/paddle/math/tests/test_perturbation.cpp @@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA #include #include diff --git a/paddle/math/tests/test_sparseMatrixCompare.cpp b/paddle/math/tests/test_sparseMatrixCompare.cpp index a9185a4b24b13ca0287b0f67375c4599e8b9ac78..2b2a391b9d04a9f7fa4986a6b6dd5cd8e5385f1f 100644 --- a/paddle/math/tests/test_sparseMatrixCompare.cpp +++ b/paddle/math/tests/test_sparseMatrixCompare.cpp @@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA /// This unittest checks GpuSparseMatrix/CpuSparseMatrix get same result, // so disable when /// only cpu version. diff --git a/paddle/memory/.clang-format b/paddle/memory/.clang-format deleted file mode 100644 index 29282dc87e2c499988c17d90d47d44cd5cf7f115..0000000000000000000000000000000000000000 --- a/paddle/memory/.clang-format +++ /dev/null @@ -1,5 +0,0 @@ ---- -Language: Cpp -BasedOnStyle: Google -Standard: Cpp11 -... diff --git a/paddle/memory/.clang-format b/paddle/memory/.clang-format new file mode 120000 index 0000000000000000000000000000000000000000..7d28cb3924707d39dafe20f4664fb17b5538996c --- /dev/null +++ b/paddle/memory/.clang-format @@ -0,0 +1 @@ +../framework/.clang-format \ No newline at end of file diff --git a/paddle/memory/detail/buddy_allocator.cc b/paddle/memory/detail/buddy_allocator.cc index bb44970109c05d239e6b92d90b2079b752fa0104..e212f7737a4093125857126cabb5b1a7b3e055b1 100644 --- a/paddle/memory/detail/buddy_allocator.cc +++ b/paddle/memory/detail/buddy_allocator.cc @@ -175,14 +175,14 @@ void* BuddyAllocator::SystemAlloc(size_t size) { } BuddyAllocator::PoolSet::iterator BuddyAllocator::RefillPool() { -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA if (system_allocator_->UseGpu()) { if ((total_used_ + total_free_) == 0) { // Compute the maximum allocation size for the first allocation. max_chunk_size_ = platform::GpuMaxChunkSize(); } } -#endif // PADDLE_ONLY_CPU +#endif // Allocate a new maximum sized block size_t index = 0; diff --git a/paddle/memory/detail/system_allocator.cc b/paddle/memory/detail/system_allocator.cc index a270bd59581520859d43cddd2fc0cfa72080f46d..33166d9ce23a4a345fc00a65adf63281b13643c3 100644 --- a/paddle/memory/detail/system_allocator.cc +++ b/paddle/memory/detail/system_allocator.cc @@ -62,7 +62,7 @@ void CPUAllocator::Free(void* p, size_t size, size_t index) { bool CPUAllocator::UseGpu() const { return false; } -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA void* GPUAllocator::Alloc(size_t& index, size_t size) { // CUDA documentation doesn't explain if cudaMalloc returns nullptr @@ -134,7 +134,7 @@ void GPUAllocator::Free(void* p, size_t size, size_t index) { bool GPUAllocator::UseGpu() const { return true; } -#endif // PADDLE_ONLY_CPU +#endif } // namespace detail } // namespace memory diff --git a/paddle/memory/detail/system_allocator.h b/paddle/memory/detail/system_allocator.h index 82ba322e057575c460b1d51d719c9b0fa459273e..552cab4f96ff21a6f3c66209eb62150e92996826 100644 --- a/paddle/memory/detail/system_allocator.h +++ b/paddle/memory/detail/system_allocator.h @@ -40,7 +40,7 @@ class CPUAllocator : public SystemAllocator { virtual bool UseGpu() const; }; -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA class GPUAllocator : public SystemAllocator { public: virtual void* Alloc(size_t& index, size_t size); @@ -51,7 +51,7 @@ class GPUAllocator : public SystemAllocator { size_t gpu_alloc_size_ = 0; size_t fallback_alloc_size_ = 0; }; -#endif // PADDLE_ONLY_CPU +#endif } // namespace detail } // namespace memory diff --git a/paddle/memory/detail/system_allocator_test.cc b/paddle/memory/detail/system_allocator_test.cc index ba44e06ddb68e92e4086a8006b868557b0c89b50..6a8558937bf0c924e5f48605ff066e2789fd59b6 100644 --- a/paddle/memory/detail/system_allocator_test.cc +++ b/paddle/memory/detail/system_allocator_test.cc @@ -56,10 +56,10 @@ TEST(CPUAllocator, LockMem) { TestAllocator(a, 0); } -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA TEST(GPUAllocator, Alloc) { paddle::memory::detail::GPUAllocator a; TestAllocator(a, 2048); TestAllocator(a, 0); } -#endif // PADDLE_ONLY_CPU +#endif diff --git a/paddle/memory/memcpy.cc b/paddle/memory/memcpy.cc index 19ec9ba9b26f5919796181a19a048b7edb508bdd..1df88a6da9fb0c50d0d7ecd083c0533d8a886a67 100644 --- a/paddle/memory/memcpy.cc +++ b/paddle/memory/memcpy.cc @@ -26,7 +26,7 @@ void Copy(platform::CPUPlace, void* dst, std::memcpy(dst, src, num); } -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA template <> void Copy(platform::CPUPlace dst_place, void* dst, @@ -80,7 +80,16 @@ void Copy(platform::GPUPlace dst_place, platform::GpuMemcpySync(dst, src, num, cudaMemcpyHostToDevice); } -#endif // PADDLE_ONLY_CPU +template <> +void Copy(platform::GPUPlace dst_place, + void* dst, + platform::GPUPlace src_place, + const void* src, size_t num) { + platform::SetDeviceId(dst_place.device); + platform::GpuMemcpySync(dst, src, num, cudaMemcpyDeviceToDevice); +} + +#endif } // namespace memory } // namespace paddle diff --git a/paddle/memory/memcpy.h b/paddle/memory/memcpy.h index 2b9c0eada6e8406fc81baec7f331a8dd5b8b0ec1..9b36182c2b619317da31310141823442d8fd3f94 100644 --- a/paddle/memory/memcpy.h +++ b/paddle/memory/memcpy.h @@ -33,7 +33,7 @@ namespace memory { template void Copy(DstPlace, void* dst, SrcPlace, const void* src, size_t num); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA /** * \brief Copy memory from one place to another place. @@ -53,7 +53,7 @@ template void Copy(DstPlace, void* dst, SrcPlace, const void* src, size_t num, cudaStream_t stream); -#endif // PADDLE_ONLY_CPU +#endif } // namespace memory } // namespace paddle diff --git a/paddle/memory/memory.cc b/paddle/memory/memory.cc index 29bc26f9d3bca0e30896657431f9a9bb1dac0d1d..5087c02385f7f37d78d134b739f3f22522977fb8 100644 --- a/paddle/memory/memory.cc +++ b/paddle/memory/memory.cc @@ -62,7 +62,7 @@ size_t Used(platform::CPUPlace place) { return GetCPUBuddyAllocator()->Used(); } -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA BuddyAllocator* GetGPUBuddyAllocator(int gpu_id) { using BuddyAllocVec = std::vector; @@ -77,7 +77,7 @@ BuddyAllocator* GetGPUBuddyAllocator(int gpu_id) { // GPU buddy allocator initialization std::call_once(gpu_allocator_flag, [&]() { - int gpu_num = platform::GetDeviceCount(); + int gpu_num = platform::GetCUDADeviceCount(); allocators.reserve(gpu_num); for (int gpu = 0; gpu < gpu_num; gpu++) { platform::SetDeviceId(gpu); @@ -111,7 +111,7 @@ size_t Used(platform::GPUPlace place) { return GetGPUBuddyAllocator(place.device)->Used(); } -#endif // PADDLE_ONLY_CPU +#endif } // namespace memory } // namespace paddle diff --git a/paddle/memory/memory_test.cc b/paddle/memory/memory_test.cc index 53cc63a098d0802479e3a371717adb7596c249ed..2444931e26774ae80b916fbb7bd46ff93025d9ed 100644 --- a/paddle/memory/memory_test.cc +++ b/paddle/memory/memory_test.cc @@ -80,7 +80,7 @@ TEST(BuddyAllocator, CPUMultAlloc) { } } -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA size_t align(size_t size, paddle::platform::GPUPlace place) { size += sizeof(paddle::memory::detail::Metadata); @@ -135,4 +135,4 @@ TEST(BuddyAllocator, GPUMultAlloc) { } } -#endif // PADDLE_ONLY_CPU +#endif diff --git a/paddle/operators/.clang-format b/paddle/operators/.clang-format deleted file mode 100644 index 47b8a85206ab457e2b3cb90a68b7a82a0753d327..0000000000000000000000000000000000000000 --- a/paddle/operators/.clang-format +++ /dev/null @@ -1,5 +0,0 @@ ---- -Language: Cpp -BasedOnStyle: Google -Standard: Cpp11 -... diff --git a/paddle/operators/.clang-format b/paddle/operators/.clang-format new file mode 120000 index 0000000000000000000000000000000000000000..7d28cb3924707d39dafe20f4664fb17b5538996c --- /dev/null +++ b/paddle/operators/.clang-format @@ -0,0 +1 @@ +../framework/.clang-format \ No newline at end of file diff --git a/paddle/operators/CMakeLists.txt b/paddle/operators/CMakeLists.txt index f8b0bce6815ff17a60ef64b0eec34a7cc9d16e72..d132c1813e6871add95c5017d563da35a7912fe4 100644 --- a/paddle/operators/CMakeLists.txt +++ b/paddle/operators/CMakeLists.txt @@ -55,12 +55,33 @@ function(op_library TARGET) set(pybind_flag 1) endif() + # pool_op contains several operators + if ("${TARGET}" STREQUAL "pool_op") + set(pybind_flag 1) + # It's enough to just adding one operator to pybind + file(APPEND ${pybind_file} "USE_OP(pool2d);\n") + endif() + + # pool_with_index_op contains several operators + if ("${TARGET}" STREQUAL "pool_with_index_op") + set(pybind_flag 1) + # It's enough to just adding one operator to pybind + file(APPEND ${pybind_file} "USE_OP(max_pool2d_with_index);\n") + endif() + # activation_op contains several operators if ("${TARGET}" STREQUAL "activation_op") set(pybind_flag 1) # It's enough to just adding one operator to pybind file(APPEND ${pybind_file} "USE_OP(sigmoid);\n") endif() + + # reduce_op contains several operators + if ("${TARGET}" STREQUAL "reduce_op") + set(pybind_flag 1) + # It's enough to just adding one operator to pybind + file(APPEND ${pybind_file} "USE_OP(reduce_sum);\n") + endif() # pybind USE_NO_KERNEL_OP file(READ ${TARGET}.cc TARGET_CONTENT) @@ -88,10 +109,18 @@ add_subdirectory(math) set(DEPS_OPS recurrent_op - cond_op) + cond_op + cross_entropy_op + softmax_with_cross_entropy_op + sum_op) + + op_library(recurrent_op SRCS recurrent_op.cc rnn/recurrent_op_utils.cc DEPS framework_proto tensor net_op) op_library(cond_op SRCS cond_op.cc DEPS framework_proto tensor operator net_op) +op_library(cross_entropy_op DEPS cross_entropy) +op_library(softmax_with_cross_entropy_op DEPS cross_entropy softmax) +op_library(sum_op DEPS net_op) list(REMOVE_ITEM GENERAL_OPS ${DEPS_OPS}) foreach(src ${GENERAL_OPS}) diff --git a/paddle/operators/accuracy_op.cc b/paddle/operators/accuracy_op.cc index 70e4f9da1221ab300e2b507a3da2f7c5da93f2e4..c5fb113e0f4455ec852fd48542ea2917398df5f3 100644 --- a/paddle/operators/accuracy_op.cc +++ b/paddle/operators/accuracy_op.cc @@ -22,25 +22,23 @@ class AccuracyOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL( - ctx.InputVar("Inference"), - "Input(Inference) of AccuracyOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Label"), - "Input(Label) of AccuracyOp should not be null."); - PADDLE_ENFORCE_NOT_NULL( - ctx.OutputVar("Accuracy"), - "Output(Accuracy) of AccuracyOp should not be null."); + void InferShape(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Inference"), + "Input(Inference) of AccuracyOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Label"), + "Input(Label) of AccuracyOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Accuracy"), + "Output(Accuracy) of AccuracyOp should not be null."); - auto *inference = ctx.Input("Inference"); - auto *label = ctx.Input("Label"); + auto inference_dim = ctx->GetInputDim("Inference"); + auto label_dim = ctx->GetInputDim("Label"); - PADDLE_ENFORCE_EQ(label->dims().size(), 1, "label must be a vector"); - PADDLE_ENFORCE_EQ(inference->dims()[0], label->dims()[0], + PADDLE_ENFORCE_EQ(label_dim.size(), 1, "label must be a vector"); + PADDLE_ENFORCE_EQ(inference_dim[0], label_dim[0], "inference size must be the same as label size"); - ctx.Output("Accuracy")->Resize({1}); - ctx.ShareLoD("Inference", /*->*/ "Accuracy"); + ctx->SetOutputDim("Accuracy", {1}); + ctx->ShareLoD("Inference", /*->*/ "Accuracy"); } }; diff --git a/paddle/operators/accuracy_op.cu b/paddle/operators/accuracy_op.cu index 0a6a0fd15c73330902552f7a9aa6339de24c1a18..0ca9ef941d4cb15619caea2b6baed197e4b15e5a 100644 --- a/paddle/operators/accuracy_op.cu +++ b/paddle/operators/accuracy_op.cu @@ -47,7 +47,7 @@ __global__ void AccuracyCudaKernel(const int N, const int D, const int* Xdata, } template -class AccuracyOpCUDAKernel : public framework::OpKernel { +class AccuracyOpCUDAKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), @@ -69,8 +69,12 @@ class AccuracyOpCUDAKernel : public framework::OpKernel { return; } - AccuracyCudaKernel<<<1, PADDLE_CUDA_NUM_THREADS>>>( - num_samples, infer_width, inference_data, label_data, accuracy_data); + AccuracyCudaKernel<<< + 1, PADDLE_CUDA_NUM_THREADS, 0, + reinterpret_cast( + ctx.device_context()) + .stream()>>>(num_samples, infer_width, inference_data, label_data, + accuracy_data); } }; diff --git a/paddle/operators/accuracy_op.h b/paddle/operators/accuracy_op.h index fe704efe1c979f4fc6a5a37184e51b416f5e517f..12c6b9aac8819caedbc02017cee81b37322bb72a 100644 --- a/paddle/operators/accuracy_op.h +++ b/paddle/operators/accuracy_op.h @@ -35,7 +35,7 @@ template ; template -class AccuracyKernel : public framework::OpKernel { +class AccuracyKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { auto* inference = ctx.Input("Inference"); diff --git a/paddle/operators/activation_op.cc b/paddle/operators/activation_op.cc index 06654702bc42cc7cf4917b00693334b1d36ce371..92db62907924d8e9e3e6acde88f3d66b7f69ec0a 100644 --- a/paddle/operators/activation_op.cc +++ b/paddle/operators/activation_op.cc @@ -22,10 +22,9 @@ class ActivationOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - ctx.Output("Y")->Resize( - ctx.Input("X")->dims()); - ctx.ShareLoD("X", /*->*/ "Y"); + void InferShape(framework::InferShapeContext *ctx) const override { + ctx->SetOutputDim("Y", ctx->GetInputDim("X")); + ctx->ShareLoD("X", /*->*/ "Y"); } }; @@ -34,9 +33,8 @@ class ActivationOpGrad : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - ctx.Output(framework::GradVarName("X")) - ->Resize(ctx.Input("Y")->dims()); + void InferShape(framework::InferShapeContext *ctx) const override { + ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("Y")); } }; @@ -71,6 +69,22 @@ class ReluOpMaker : public framework::OpProtoAndCheckerMaker { } }; +template +class LeakyReluOpMaker : public framework::OpProtoAndCheckerMaker { + public: + LeakyReluOpMaker(framework::OpProto *proto, + framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "Input of LeakyRelu operator"); + AddOutput("Y", "Output of LeakyRelu operator"); + AddComment( + "LeakyRelu activation operator, " + "leaky_relu = max(x, alpha * x)"); + AddAttr("alpha", "The small negative slope") + .SetDefault(static_cast(0.02f)); + } +}; + class TanhOpMaker : public framework::OpProtoAndCheckerMaker { public: TanhOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) @@ -83,6 +97,17 @@ class TanhOpMaker : public framework::OpProtoAndCheckerMaker { } }; +class TanhShrinkOpMaker : public framework::OpProtoAndCheckerMaker { + public: + TanhShrinkOpMaker(framework::OpProto *proto, + framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "Input of TanhShrink operator"); + AddOutput("Y", "Output of TanhShrink operator"); + AddComment("TanhShrink activation operator, tanhshrink(x) = x - tanh(x)"); + } +}; + class SqrtOpMaker : public framework::OpProtoAndCheckerMaker { public: SqrtOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) @@ -134,6 +159,17 @@ class SquareOpMaker : public framework::OpProtoAndCheckerMaker { } }; +class SoftsignOpMaker : public framework::OpProtoAndCheckerMaker { + public: + SoftsignOpMaker(framework::OpProto *proto, + framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "Input of Softsign operator"); + AddOutput("Y", "Output of Softsign operator"); + AddComment("Softsign activation operator, softsign(x) = x / (1 + |x|)"); + } +}; + template class BReluOpMaker : public framework::OpProtoAndCheckerMaker { public: @@ -165,6 +201,40 @@ class SoftReluOpMaker : public framework::OpProtoAndCheckerMaker { } }; +template +class ELUOpMaker : public framework::OpProtoAndCheckerMaker { + public: + ELUOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", + "(Tensor) The input of ELU operator, it shouldn't be empty. Input " + "is flattened and treated as a 1D array."); + AddOutput("Y", + "(Tensor) The output of ELU operator. It has the same shape as " + "the input."); + AddAttr( + "alpha", "(float, default 1.0) Alpha value in the elu formulation.") + .SetDefault(static_cast(1.)); + AddComment(R"DOC( + ELU activation operator. It applies this element-wise computation on + the input: f(x) = max(0, x) + min(0, alpha * (exp(x) - 1)). + Check .. _Link: https://arxiv.org/abs/1511.07289 for more details.)DOC"); + } +}; + +template +class Relu6OpMaker : public framework::OpProtoAndCheckerMaker { + public: + Relu6OpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "Input of Relu6 operator"); + AddOutput("Y", "Output of Relu6 operator"); + AddComment("Relu6 activation operator, relu6 = min(max(0, x), 6)"); + AddAttr("threshold", "The threshold value of Relu6") + .SetDefault(static_cast(6)); + } +}; + template class PowOpMaker : public framework::OpProtoAndCheckerMaker { public: @@ -197,111 +267,67 @@ class STanhOpMaker : public framework::OpProtoAndCheckerMaker { } // namespace paddle namespace ops = paddle::operators; + REGISTER_OP(sigmoid, ops::ActivationOp, ops::SigmoidOpMaker, sigmoid_grad, ops::ActivationOpGrad); -REGISTER_OP_CPU_KERNEL(sigmoid, - ops::ActivationKernel>); -REGISTER_OP_CPU_KERNEL( - sigmoid_grad, ops::ActivationGradKernel>); REGISTER_OP(exp, ops::ActivationOp, ops::ExpOpMaker, exp_grad, ops::ActivationOpGrad); -REGISTER_OP_CPU_KERNEL( - exp, - ops::ActivationKernel); -REGISTER_OP_CPU_KERNEL(exp_grad, - ops::ActivationGradKernel); REGISTER_OP(relu, ops::ActivationOp, ops::ReluOpMaker, relu_grad, ops::ActivationOpGrad); -REGISTER_OP_CPU_KERNEL(relu, - ops::ActivationKernel>); -REGISTER_OP_CPU_KERNEL( - relu_grad, ops::ActivationGradKernel>); REGISTER_OP(tanh, ops::ActivationOp, ops::TanhOpMaker, tanh_grad, ops::ActivationOpGrad); -REGISTER_OP_CPU_KERNEL( - tanh, - ops::ActivationKernel); -REGISTER_OP_CPU_KERNEL( - tanh_grad, ops::ActivationGradKernel>); + +REGISTER_OP(tanh_shrink, ops::ActivationOp, ops::TanhShrinkOpMaker, + tanh_shrink_grad, ops::ActivationOpGrad); REGISTER_OP(sqrt, ops::ActivationOp, ops::SqrtOpMaker, sqrt_grad, ops::ActivationOpGrad); -REGISTER_OP_CPU_KERNEL( - sqrt, - ops::ActivationKernel); -REGISTER_OP_CPU_KERNEL( - sqrt_grad, ops::ActivationGradKernel>); REGISTER_OP(abs, ops::ActivationOp, ops::AbsOpMaker, abs_grad, ops::ActivationOpGrad); -REGISTER_OP_CPU_KERNEL( - abs, - ops::ActivationKernel); -REGISTER_OP_CPU_KERNEL(abs_grad, - ops::ActivationGradKernel); REGISTER_OP(reciprocal, ops::ActivationOp, ops::ReciprocalOpMaker, reciprocal_grad, ops::ActivationOpGrad); -REGISTER_OP_CPU_KERNEL(reciprocal, - ops::ActivationKernel>); -REGISTER_OP_CPU_KERNEL( - reciprocal_grad, - ops::ActivationGradKernel>); REGISTER_OP(log, ops::ActivationOp, ops::LogOpMaker, log_grad, ops::ActivationOpGrad); -REGISTER_OP_CPU_KERNEL( - log, - ops::ActivationKernel); -REGISTER_OP_CPU_KERNEL( - log_grad, ops::ActivationGradKernel>); REGISTER_OP(square, ops::ActivationOp, ops::SquareOpMaker, square_grad, ops::ActivationOpGrad); -REGISTER_OP_CPU_KERNEL(square, - ops::ActivationKernel); -REGISTER_OP_CPU_KERNEL( - square_grad, ops::ActivationGradKernel>); + +REGISTER_OP(softsign, ops::ActivationOp, ops::SoftsignOpMaker, softsign_grad, + ops::ActivationOpGrad); REGISTER_OP(brelu, ops::ActivationOp, ops::BReluOpMaker, brelu_grad, ops::ActivationOpGrad); -REGISTER_OP_CPU_KERNEL(brelu, - ops::BReluKernel); -REGISTER_OP_CPU_KERNEL(brelu_grad, - ops::BReluGradKernel); + +REGISTER_OP(leaky_relu, ops::ActivationOp, ops::LeakyReluOpMaker, + leaky_relu_grad, ops::ActivationOpGrad); REGISTER_OP(soft_relu, ops::ActivationOp, ops::SoftReluOpMaker, soft_relu_grad, ops::ActivationOpGrad); -REGISTER_OP_CPU_KERNEL(soft_relu, - ops::SoftReluKernel); -REGISTER_OP_CPU_KERNEL( - soft_relu_grad, ops::SoftReluGradKernel); + +REGISTER_OP(elu, ops::ActivationOp, ops::ELUOpMaker, elu_grad, + ops::ActivationOpGrad); + +REGISTER_OP(relu6, ops::ActivationOp, ops::Relu6OpMaker, relu6_grad, + ops::ActivationOpGrad); REGISTER_OP(pow, ops::ActivationOp, ops::PowOpMaker, pow_grad, ops::ActivationOpGrad); -REGISTER_OP_CPU_KERNEL(pow, ops::PowKernel); -REGISTER_OP_CPU_KERNEL(pow_grad, - ops::PowGradKernel); REGISTER_OP(stanh, ops::ActivationOp, ops::STanhOpMaker, stanh_grad, ops::ActivationOpGrad); -REGISTER_OP_CPU_KERNEL(stanh, - ops::STanhKernel); -REGISTER_OP_CPU_KERNEL(stanh_grad, - ops::STanhGradKernel); + +#define REGISTER_ACTIVATION_CPU_KERNEL(act_type, functor, grad_functor) \ + REGISTER_OP_CPU_KERNEL( \ + act_type, \ + ops::ActivationKernel>); \ + REGISTER_OP_CPU_KERNEL(act_type##_grad, \ + ops::ActivationGradKernel>); + +FOR_EACH_KERNEL_FUNCTOR(REGISTER_ACTIVATION_CPU_KERNEL); diff --git a/paddle/operators/activation_op.cu b/paddle/operators/activation_op.cu index feed1302b292a546f88fa35457c86aa2cfdaa307..7b7644519d4e9cadcc4ca62ccb599262feffa660 100644 --- a/paddle/operators/activation_op.cu +++ b/paddle/operators/activation_op.cu @@ -17,84 +17,12 @@ namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL(sigmoid, - ops::ActivationKernel>); -REGISTER_OP_GPU_KERNEL( - sigmoid_grad, ops::ActivationGradKernel>); - -REGISTER_OP_GPU_KERNEL( - exp, - ops::ActivationKernel); -REGISTER_OP_GPU_KERNEL(exp_grad, - ops::ActivationGradKernel); -REGISTER_OP_GPU_KERNEL(relu, - ops::ActivationKernel>); -REGISTER_OP_GPU_KERNEL( - relu_grad, ops::ActivationGradKernel>); - -REGISTER_OP_GPU_KERNEL( - tanh, - ops::ActivationKernel); -REGISTER_OP_GPU_KERNEL( - tanh_grad, ops::ActivationGradKernel>); - -REGISTER_OP_GPU_KERNEL( - sqrt, - ops::ActivationKernel); -REGISTER_OP_GPU_KERNEL( - sqrt_grad, ops::ActivationGradKernel>); - -REGISTER_OP_GPU_KERNEL( - abs, - ops::ActivationKernel); -REGISTER_OP_GPU_KERNEL(abs_grad, - ops::ActivationGradKernel); - -REGISTER_OP_GPU_KERNEL(reciprocal, - ops::ActivationKernel>); -REGISTER_OP_GPU_KERNEL( - reciprocal_grad, - ops::ActivationGradKernel>); - -REGISTER_OP_GPU_KERNEL( - log, - ops::ActivationKernel); -REGISTER_OP_GPU_KERNEL( - log_grad, ops::ActivationGradKernel>); - -REGISTER_OP_GPU_KERNEL(square, - ops::ActivationKernel); -REGISTER_OP_GPU_KERNEL( - square_grad, ops::ActivationGradKernel>); - -REGISTER_OP_GPU_KERNEL(brelu, - ops::BReluKernel); -REGISTER_OP_GPU_KERNEL(brelu_grad, - ops::BReluGradKernel); - -REGISTER_OP_GPU_KERNEL(soft_relu, - ops::SoftReluKernel); -REGISTER_OP_GPU_KERNEL( - soft_relu_grad, ops::SoftReluGradKernel); - -REGISTER_OP_GPU_KERNEL(pow, ops::PowKernel); -REGISTER_OP_GPU_KERNEL(pow_grad, - ops::PowGradKernel); - -REGISTER_OP_GPU_KERNEL(stanh, - ops::STanhKernel); -REGISTER_OP_GPU_KERNEL(stanh_grad, - ops::STanhGradKernel); +#define REGISTER_ACTIVATION_GPU_KERNEL(act_type, functor, grad_functor) \ + REGISTER_OP_GPU_KERNEL( \ + act_type, \ + ops::ActivationKernel>); \ + REGISTER_OP_GPU_KERNEL(act_type##_grad, \ + ops::ActivationGradKernel>); + +FOR_EACH_KERNEL_FUNCTOR(REGISTER_ACTIVATION_GPU_KERNEL); diff --git a/paddle/operators/activation_op.h b/paddle/operators/activation_op.h index 15f8afb4ba45cc989fe7576b82b8bf853b1df7de..123f0c4dbca6537c9bd167ca74a06987db6e1893 100644 --- a/paddle/operators/activation_op.h +++ b/paddle/operators/activation_op.h @@ -19,9 +19,12 @@ namespace paddle { namespace operators { -template -class ActivationKernel : public framework::OpKernel { +template +class ActivationKernel + : public framework::OpKernel { public: + using T = typename Functor::ELEMENT_TYPE; + void Compute(const framework::ExecutionContext& context) const override { auto* X = context.Input("X"); auto* Y = context.Output("Y"); @@ -31,13 +34,20 @@ class ActivationKernel : public framework::OpKernel { auto y = framework::EigenVector::Flatten(*Y); auto place = context.GetEigenDevice(); Functor functor; + + auto attrs = functor.GetAttrs(); + for (auto& attr : attrs) { + *attr.second = context.Attr(attr.first); + } functor(place, x, y); } }; -template -class ActivationGradKernel : public framework::OpKernel { +template +class ActivationGradKernel + : public framework::OpKernel { public: + using T = typename Functor::ELEMENT_TYPE; void Compute(const framework::ExecutionContext& context) const override { auto* X = context.Input("X"); auto* Y = context.Input("Y"); @@ -51,303 +61,430 @@ class ActivationGradKernel : public framework::OpKernel { auto dx = framework::EigenVector::Flatten(*dX); auto place = context.GetEigenDevice(); Functor functor; + auto attrs = functor.GetAttrs(); + for (auto& attr : attrs) { + *attr.second = context.Attr(attr.first); + } functor(place, x, y, dy, dx); } }; +template +struct BaseActivationFunctor { + using ELEMENT_TYPE = T; + + using AttrPair = std::vector>; + + AttrPair GetAttrs() { return AttrPair(); } +}; + // sigmoid(x) = 1 / (1 + exp(-x)) template -struct SigmoidFunctor { +struct SigmoidFunctor : public BaseActivationFunctor { template - void operator()(Device d, X x, Y y) { + void operator()(Device d, X x, Y y) const { y.device(d) = static_cast(1) / (static_cast(1) + (-x).exp()); } }; template -struct SigmoidGradFunctor { +struct SigmoidGradFunctor : public BaseActivationFunctor { template - void operator()(Device d, X x, Y y, dY dy, dX dx) { + void operator()(Device d, X x, Y y, dY dy, dX dx) const { dx.device(d) = dy * y * (static_cast(1) - y); } }; // exp(x) = e^x -struct ExpFunctor { +template +struct ExpFunctor : public BaseActivationFunctor { template - void operator()(Device d, X x, Y y) { + void operator()(Device d, X x, Y y) const { y.device(d) = x.exp(); } }; -struct ExpGradFunctor { +template +struct ExpGradFunctor : public BaseActivationFunctor { template - void operator()(Device d, X x, Y y, dY dy, dX dx) { + void operator()(Device d, X x, Y y, dY dy, dX dx) const { dx.device(d) = dy * y; } }; // relu(x) = max(x, 0) template -struct ReluFunctor { +struct ReluFunctor : public BaseActivationFunctor { template - void operator()(Device d, X x, Y y) { + void operator()(Device d, X x, Y y) const { y.device(d) = x.cwiseMax(static_cast(0)); } }; template -struct ReluGradFunctor { +struct ReluGradFunctor : public BaseActivationFunctor { template - void operator()(Device d, X x, Y y, dY dy, dX dx) { + void operator()(Device d, X x, Y y, dY dy, dX dx) const { dx.device(d) = dy * (x > static_cast(0)).template cast(); } }; // tanh(x) = (exp(x) - exp(-x)) / (exp(x) + exp(-x)) -struct TanhFunctor { +template +struct TanhFunctor : public BaseActivationFunctor { template - void operator()(Device d, X x, Y y) { + void operator()(Device d, X x, Y y) const { y.device(d) = x.tanh(); } }; template -struct TanhGradFunctor { +struct TanhGradFunctor : public BaseActivationFunctor { template - void operator()(Device d, X x, Y y, dY dy, dX dx) { + void operator()(Device d, X x, Y y, dY dy, dX dx) const { dx.device(d) = dy * (static_cast(1) - y * y); } }; +// tanhshrink(x) = x - tanh(x) +// where tanh(x) = (exp(x) - exp(-x)) / (exp(x) + exp(-x)) +template +struct TanhShrinkFunctor : public BaseActivationFunctor { + template + void operator()(Device d, X x, Y y) const { + y.device(d) = x - x.tanh(); + } +}; + +template +struct TanhShrinkGradFunctor : public BaseActivationFunctor { + template + void operator()(Device d, X x, Y y, dY dy, dX dx) const { + dx.device(d) = dy * (x.tanh() * x.tanh()); + } +}; + // sqrt(x) = x^(1/2) -struct SqrtFunctor { +template +struct SqrtFunctor : public BaseActivationFunctor { template - void operator()(Device d, X x, Y y) { + void operator()(Device d, X x, Y y) const { y.device(d) = x.sqrt(); } }; template -struct SqrtGradFunctor { +struct SqrtGradFunctor : public BaseActivationFunctor { template - void operator()(Device d, X x, Y y, dY dy, dX dx) { + void operator()(Device d, X x, Y y, dY dy, dX dx) const { const Y y_conj = Eigen::numext::conj(y); dx.device(d) = static_cast(0.5) * dy / y_conj; } }; // abs(x) = |x| -struct AbsFunctor { +template +struct AbsFunctor : public BaseActivationFunctor { template - void operator()(Device d, X x, Y y) { + void operator()(Device d, X x, Y y) const { y.device(d) = x.abs(); } }; -struct AbsGradFunctor { +template +struct AbsGradFunctor : public BaseActivationFunctor { template - void operator()(Device d, X x, Y y, dY dy, dX dx) { + void operator()(Device d, X x, Y y, dY dy, dX dx) const { dx.device(d) = dy * x.sign(); } }; // reciprocal(x) = 1 / x template -struct ReciprocalFunctor { +struct ReciprocalFunctor : public BaseActivationFunctor { template - void operator()(Device d, X x, Y y) { + void operator()(Device d, X x, Y y) const { y.device(d) = static_cast(1) / x; } }; template -struct ReciprocalGradFunctor { +struct ReciprocalGradFunctor : public BaseActivationFunctor { template - void operator()(Device d, X x, Y y, dY dy, dX dx) { + void operator()(Device d, X x, Y y, dY dy, dX dx) const { dx.device(d) = dy * static_cast(-1) * y * y; } }; // log(x) = natural logarithm of x -struct LogFunctor { +template +struct LogFunctor : public BaseActivationFunctor { template - void operator()(Device d, X x, Y y) { + void operator()(Device d, X x, Y y) const { y.device(d) = x.log(); } }; template -struct LogGradFunctor { +struct LogGradFunctor : public BaseActivationFunctor { template - void operator()(Device d, X x, Y y, dY dy, dX dx) { + void operator()(Device d, X x, Y y, dY dy, dX dx) const { dx.device(d) = dy * (static_cast(1) / x); } }; // square(x) = x^2 -struct SquareFunctor { +template +struct SquareFunctor : public BaseActivationFunctor { template - void operator()(Device d, X x, Y y) { + void operator()(Device d, X x, Y y) const { y.device(d) = x.square(); } }; template -struct SquareGradFunctor { +struct SquareGradFunctor : public BaseActivationFunctor { template - void operator()(Device d, X x, Y y, dY dy, dX dx) { + void operator()(Device d, X x, Y y, dY dy, dX dx) const { dx.device(d) = dy * static_cast(2) * x; } }; -template -class BReluKernel : public framework::OpKernel { - public: - void Compute(const framework::ExecutionContext& context) const override { - auto* X = context.Input("X"); - auto* Y = context.Output("Y"); - auto t_min = static_cast(context.Attr("t_min")); - auto t_max = static_cast(context.Attr("t_max")); - Y->mutable_data(context.GetPlace()); +template +struct BReluFunctor : public BaseActivationFunctor { + float t_min; + float t_max; + + // NOTE: Explicit hides the `BaseActivationFunctor::GetAttrs` + // not polymorphism for speed. + typename BaseActivationFunctor::AttrPair GetAttrs() { + return {{"t_min", &t_min}, {"t_max", &t_max}}; + } - auto x = framework::EigenVector::Flatten(*X); - auto y = framework::EigenVector::Flatten(*Y); - auto place = context.GetEigenDevice(); - y.device(place) = x.cwiseMax(t_min).cwiseMin(t_max); + template + void operator()(Device d, X x, Y y) const { + y.device(d) = x.cwiseMax(t_min).cwiseMin(t_max); } }; -template -class BReluGradKernel : public framework::OpKernel { - public: - void Compute(const framework::ExecutionContext& context) const override { - auto* X = context.Input("X"); - auto* dY = context.Input(framework::GradVarName("Y")); - auto* dX = context.Output(framework::GradVarName("X")); - auto t_min = static_cast(context.Attr("t_min")); - auto t_max = static_cast(context.Attr("t_max")); - dX->mutable_data(context.GetPlace()); +template +struct BReluGradFunctor : public BaseActivationFunctor { + float t_min; + float t_max; + typename BaseActivationFunctor::AttrPair GetAttrs() { + return {{"t_min", &t_min}, {"t_max", &t_max}}; + } + template + void operator()(Device d, X x, Y y, dY dy, dX dx) const { + dx.device(d) = dy * ((x > t_min) * (x < t_max)).template cast(); + } +}; - auto dy = framework::EigenVector::Flatten(*dY); - auto x = framework::EigenVector::Flatten(*X); - auto dx = framework::EigenVector::Flatten(*dX); - auto place = context.GetEigenDevice(); +// relu6(x) = min(max(0, x), 6) +template +struct Relu6Functor : public BaseActivationFunctor { + float threshold; - dx.device(place) = dy * ((x > t_min) * (x < t_max)).template cast(); + // NOTE: Explicit hides the `BaseActivationFunctor::GetAttrs` + // not polymorphism for speed. + typename BaseActivationFunctor::AttrPair GetAttrs() { + return {{"threshold", &threshold}}; + } + + template + void operator()(Device d, X x, Y y) const { + y.device(d) = x.cwiseMax(static_cast(0)).cwiseMin(threshold); } }; -template -class SoftReluKernel : public framework::OpKernel { - public: - void Compute(const framework::ExecutionContext& context) const override { - auto* X = context.Input("X"); - auto* Y = context.Output("Y"); - auto threshold = static_cast(context.Attr("threshold")); - Y->mutable_data(context.GetPlace()); +template +struct Relu6GradFunctor : public BaseActivationFunctor { + float threshold; + typename BaseActivationFunctor::AttrPair GetAttrs() { + return {{"threshold", &threshold}}; + } + template + void operator()(Device d, X x, Y y, dY dy, dX dx) const { + dx.device(d) = + dy * ((x > static_cast(0)) * (x < threshold)).template cast(); + } +}; - auto x = framework::EigenVector::Flatten(*X); - auto y = framework::EigenVector::Flatten(*Y); - auto place = context.GetEigenDevice(); - auto temp = x.cwiseMax(-threshold).cwiseMin(threshold).eval(); - y.device(place) = (static_cast(1) + temp.exp()).log(); +// softsign(x) = x / (1 + |x|) +template +struct SoftsignFunctor : public BaseActivationFunctor { + template + void operator()(Device d, X x, Y y) { + y.device(d) = x / (static_cast(1) + x.abs()); } }; -template -class SoftReluGradKernel : public framework::OpKernel { - public: - void Compute(const framework::ExecutionContext& context) const override { - auto* X = context.Input("X"); - auto* Y = context.Input("Y"); - auto* dY = context.Input(framework::GradVarName("Y")); - auto* dX = context.Output(framework::GradVarName("X")); - auto threshold = static_cast(context.Attr("threshold")); - dX->mutable_data(context.GetPlace()); +// d(softsign(x))/dx = 1 / (1 + |x|)^2 +// Taken from https://en.wikipedia.org/wiki/Activation_function +template +struct SoftsignGradFunctor : public BaseActivationFunctor { + template + void operator()(Device d, X x, Y y, dY dy, dX dx) { + dx.device(d) = + dy * (static_cast(1) / (static_cast(1) + x.abs()).square()); + } +}; - auto x = framework::EigenVector::Flatten(*X); - auto y = framework::EigenVector::Flatten(*Y); - auto dy = framework::EigenVector::Flatten(*dY); - auto dx = framework::EigenVector::Flatten(*dX); - auto place = context.GetEigenDevice(); +template +struct SoftReluFunctor : public BaseActivationFunctor { + float threshold; + typename BaseActivationFunctor::AttrPair GetAttrs() { + return {{"threshold", &threshold}}; + } + + template + void operator()(Device d, X x, Y y) const { + auto temp = x.cwiseMax(-threshold).cwiseMin(threshold); + y.device(d) = (static_cast(1) + temp.exp()).log(); + } +}; + +template +struct SoftReluGradFunctor : public BaseActivationFunctor { + float threshold; + typename BaseActivationFunctor::AttrPair GetAttrs() { + return {{"threshold", &threshold}}; + } + template + void operator()(Device d, X x, Y y, dY dy, dX dx) const { auto temp = ((x > -threshold) * (x < threshold)).template cast().eval(); - dx.device(place) = dy * (static_cast(1) - (-y).exp()) * temp; + dx.device(d) = dy * (static_cast(1) - (-y).exp()) * temp; } }; -template -class PowKernel : public framework::OpKernel { - public: - void Compute(const framework::ExecutionContext& context) const override { - auto* X = context.Input("X"); - auto* Y = context.Output("Y"); - auto factor = static_cast(context.Attr("factor")); - Y->mutable_data(context.GetPlace()); +template +struct LeakyReluFunctor : public BaseActivationFunctor { + float alpha; + typename BaseActivationFunctor::AttrPair GetAttrs() { + return {{"alpha", &alpha}}; + } - auto x = framework::EigenVector::Flatten(*X); - auto y = framework::EigenVector::Flatten(*Y); - auto place = context.GetEigenDevice(); - y.device(place) = x.pow(factor); + template + void operator()(Device d, X x, Y y) const { + y.device(d) = x.cwiseMax(alpha * x); } }; -template -class PowGradKernel : public framework::OpKernel { - public: - void Compute(const framework::ExecutionContext& context) const override { - auto* X = context.Input("X"); - auto* dY = context.Input(framework::GradVarName("Y")); - auto* dX = context.Output(framework::GradVarName("X")); - auto factor = static_cast(context.Attr("factor")); - dX->mutable_data(context.GetPlace()); +template +struct LeakyReluGradFunctor : public BaseActivationFunctor { + float alpha; + typename BaseActivationFunctor::AttrPair GetAttrs() { + return {{"alpha", &alpha}}; + } + template + void operator()(Device d, X x, Y y, dY dy, dX dx) const { + auto temp1 = alpha * (x < static_cast(0)).template cast().eval(); + auto temp2 = (x >= static_cast(0)).template cast().eval(); + dx.device(d) = dy * (temp1 + temp2).template cast(); + } +}; - auto dy = framework::EigenVector::Flatten(*dY); - auto x = framework::EigenVector::Flatten(*X); - auto dx = framework::EigenVector::Flatten(*dX); - auto place = context.GetEigenDevice(); +template +struct ELUFunctor : public BaseActivationFunctor { + float alpha; + typename BaseActivationFunctor::AttrPair GetAttrs() { + return {{"alpha", &alpha}}; + } - dx.device(place) = dy * factor * x.pow(factor - static_cast(1)); + template + void operator()(Device d, X x, Y y) const { + y.device(d) = + x.cwiseMax(static_cast(0)) + + (alpha * (x.exp() - static_cast(1))).cwiseMin(static_cast(0)); } }; -template -class STanhKernel : public framework::OpKernel { - public: - void Compute(const framework::ExecutionContext& context) const override { - auto* X = context.Input("X"); - auto* Y = context.Output("Y"); - auto scale_a = static_cast(context.Attr("scale_a")); - auto scale_b = static_cast(context.Attr("scale_b")); - Y->mutable_data(context.GetPlace()); +template +struct ELUGradFunctor : public BaseActivationFunctor { + float alpha; + typename BaseActivationFunctor::AttrPair GetAttrs() { + return {{"alpha", &alpha}}; + } + template + void operator()(Device d, X x, Y y, dY dy, dX dx) const { + dx.device(d) = + dy * (x > static_cast(0)).template cast() + + dy * (y + alpha) * (x < static_cast(0)).template cast(); + } +}; - auto x = framework::EigenVector::Flatten(*X); - auto y = framework::EigenVector::Flatten(*Y); - auto place = context.GetEigenDevice(); - y.device(place) = scale_b * (scale_a * x).tanh(); +template +struct PowFunctor : public BaseActivationFunctor { + float factor; + typename BaseActivationFunctor::AttrPair GetAttrs() { + return {{"factor", &factor}}; + } + template + void operator()(Device d, X x, Y y) const { + y.device(d) = x.pow(factor); } }; -template -class STanhGradKernel : public framework::OpKernel { - public: - void Compute(const framework::ExecutionContext& context) const override { - auto* X = context.Input("X"); - auto* dY = context.Input(framework::GradVarName("Y")); - auto* dX = context.Output(framework::GradVarName("X")); - auto scale_a = static_cast(context.Attr("scale_a")); - auto scale_b = static_cast(context.Attr("scale_b")); - dX->mutable_data(context.GetPlace()); +template +struct PowGradFunctor : public BaseActivationFunctor { + float factor; + typename BaseActivationFunctor::AttrPair GetAttrs() { + return {{"factor", &factor}}; + } + template + void operator()(Device d, X x, Y y, dY dy, dX dx) const { + dx.device(d) = dy * factor * x.pow(factor - static_cast(1)); + } +}; - auto dy = framework::EigenVector::Flatten(*dY); - auto x = framework::EigenVector::Flatten(*X); - auto dx = framework::EigenVector::Flatten(*dX); - auto place = context.GetEigenDevice(); +template +struct STanhFunctor : public BaseActivationFunctor { + float scale_a; + float scale_b; + typename BaseActivationFunctor::AttrPair GetAttrs() { + return {{"scale_a", &scale_a}, {"scale_b", &scale_b}}; + } + + template + void operator()(Device d, X x, Y y) const { + y.device(d) = scale_b * (scale_a * x).tanh(); + } +}; +template +struct STanhGradFunctor : public BaseActivationFunctor { + float scale_a; + float scale_b; + typename BaseActivationFunctor::AttrPair GetAttrs() { + return {{"scale_a", &scale_a}, {"scale_b", &scale_b}}; + } + + template + void operator()(Device d, X x, Y y, dY dy, dX dx) const { auto temp = (scale_a * x).tanh() * (scale_a * x).tanh(); - dx.device(place) = dy * scale_a * scale_b * (static_cast(1) - temp); + dx.device(d) = dy * scale_a * scale_b * (static_cast(1) - temp); } }; } // namespace operators } // namespace paddle + +#define FOR_EACH_KERNEL_FUNCTOR(__macro) \ + __macro(sigmoid, SigmoidFunctor, SigmoidGradFunctor); \ + __macro(exp, ExpFunctor, ExpGradFunctor); \ + __macro(relu, ReluFunctor, ReluGradFunctor); \ + __macro(tanh, TanhFunctor, TanhGradFunctor); \ + __macro(sqrt, SqrtFunctor, SqrtGradFunctor); \ + __macro(abs, AbsFunctor, AbsGradFunctor); \ + __macro(reciprocal, ReciprocalFunctor, ReciprocalGradFunctor); \ + __macro(log, LogFunctor, LogGradFunctor); \ + __macro(square, SquareFunctor, SquareGradFunctor); \ + __macro(brelu, BReluFunctor, BReluGradFunctor); \ + __macro(soft_relu, SoftReluFunctor, SoftReluGradFunctor); \ + __macro(pow, PowFunctor, PowGradFunctor); \ + __macro(stanh, STanhFunctor, STanhGradFunctor); \ + __macro(softsign, SoftsignFunctor, SoftsignGradFunctor); \ + __macro(leaky_relu, LeakyReluFunctor, LeakyReluGradFunctor); \ + __macro(relu6, Relu6Functor, Relu6GradFunctor); \ + __macro(tanh_shrink, TanhShrinkFunctor, TanhShrinkGradFunctor); \ + __macro(elu, ELUFunctor, ELUGradFunctor) diff --git a/paddle/operators/adadelta_op.cc b/paddle/operators/adadelta_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..cf1bca1658f3fb4cd14465791a0f94865871a120 --- /dev/null +++ b/paddle/operators/adadelta_op.cc @@ -0,0 +1,115 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/operators/adadelta_op.h" + +namespace paddle { +namespace operators { + +class AdadeltaOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Param"), + "Input(Param) of AdadeltaOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Grad"), + "Input(Grad) of AdadeltaOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("AvgSquaredGrad"), + "Input(AvgSquaredGrad) of AdadeltaOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("AvgSquaredUpdate"), + "Input(AvgSquaredUpdate) of AdadeltaOp should not be null."); + + PADDLE_ENFORCE(ctx->HasOutput("ParamOut"), + "Output(ParamOut) of AdadeltaOp should not be null."); + PADDLE_ENFORCE( + ctx->HasOutput("AvgSquaredGradOut"), + "Output(AvgSquaredGradOut) of AdadeltaOp should not be null."); + PADDLE_ENFORCE( + ctx->HasOutput("AvgSquaredUpdateOut"), + "Output(AvgSquaredUpdateOut) of AdadeltaOp should not be null."); + + auto param_dim = ctx->GetInputDim("Param"); + PADDLE_ENFORCE_EQ( + param_dim, ctx->GetInputDim("Grad"), + "param and grad input of AdadeltaOp should have same dimension"); + PADDLE_ENFORCE_EQ(param_dim, ctx->GetInputDim("AvgSquaredGrad"), + "Param and AvgSquaredGrad input of AdadeltaOp " + "should have same dimension"); + PADDLE_ENFORCE_EQ(param_dim, ctx->GetInputDim("AvgSquaredUpdate"), + "Param and AvgSquaredUpdate input of AdadeltaOp " + "should have same dimension"); + + ctx->SetOutputDim("ParamOut", param_dim); + ctx->SetOutputDim("AvgSquaredGradOut", param_dim); + ctx->SetOutputDim("AvgSquaredUpdateOut", param_dim); + } +}; + +class AdadeltaOpMaker : public framework::OpProtoAndCheckerMaker { + public: + AdadeltaOpMaker(framework::OpProto *proto, + framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("Param", "(Tensor) Input parameter"); + AddInput("Grad", "(Tensor) Input gradient"); + AddInput("AvgSquaredGrad", + "(Tensor) Input expectation of squared gradient"); + AddInput("AvgSquaredUpdate", + "(Tensor) Input expectation of squared parameter updates"); + + AddOutput("ParamOut", "(Tensor) Output parameter"); + AddOutput("AvgSquaredGradOut", + "(Tensor) Output expectation of squared gradient"); + AddOutput("AvgSquaredUpdateOut", + "(Tensor) Output expectation of squared parameter updates"); + + AddAttr("rho", + "(float, default 0.95) Exponential decay rate " + "for squared gradients.") + .SetDefault(0.95f); + AddAttr("epsilon", + "(float, default 1.0e-6) Constant for " + "numerical stability") + .SetDefault(1.0e-6f); + AddComment(R"DOC( +Adadelta Updates Operator. + +This implements the Adadelta optimizer[1]. Adadelta is a per-dimension +adaptive learning rate method for gradient descent. + +Adadelta updates: + +avg_squared_grad_out = rho * avg_squared_grad + (1 - rho) * grad * grad +param_update = - sqrt((avg_squared_update + epsilon) / + (avg_squared_grad_out + epsilon)) * grad +avg_squared_update_out = rho * avg_squared_update + (1 - rho) * param_update**2 +param_out = param + param_update + +References: + [1] ADADELTA: An Adaptive Learning Rate Method + https://arxiv.org/abs/1212.5701 + +)DOC"); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP_WITHOUT_GRADIENT(adadelta, ops::AdadeltaOp, ops::AdadeltaOpMaker); +REGISTER_OP_CPU_KERNEL( + adadelta, ops::AdadeltaOpKernel); diff --git a/paddle/operators/adadelta_op.cu b/paddle/operators/adadelta_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..3af1c8c8e9861138a33b3156818f704c3b20363f --- /dev/null +++ b/paddle/operators/adadelta_op.cu @@ -0,0 +1,20 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#define EIGEN_USE_GPU +#include "paddle/operators/adadelta_op.h" + +namespace ops = paddle::operators; +REGISTER_OP_GPU_KERNEL( + adadelta, ops::AdadeltaOpKernel); diff --git a/paddle/operators/adadelta_op.h b/paddle/operators/adadelta_op.h new file mode 100644 index 0000000000000000000000000000000000000000..d29e15c43583bd447fbacb548a326f303f7d1463 --- /dev/null +++ b/paddle/operators/adadelta_op.h @@ -0,0 +1,69 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" + +namespace paddle { +namespace operators { + +template +class AdadeltaOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto param_out_tensor = ctx.Output("ParamOut"); + auto avg_squared_grad_out_tensor = + ctx.Output("AvgSquaredGradOut"); + auto avg_squared_update_out_tensor = + ctx.Output("AvgSquaredUpdateOut"); + + param_out_tensor->mutable_data(ctx.GetPlace()); + avg_squared_grad_out_tensor->mutable_data(ctx.GetPlace()); + avg_squared_update_out_tensor->mutable_data(ctx.GetPlace()); + + float rho = ctx.Attr("rho"); + float epsilon = ctx.Attr("epsilon"); + + auto param = framework::EigenVector::Flatten( + *ctx.Input("Param")); + auto grad = framework::EigenVector::Flatten( + *ctx.Input("Grad")); + // Squared gradient accumulator + auto avg_squared_grad = framework::EigenVector::Flatten( + *ctx.Input("AvgSquaredGrad")); + // Squared updates accumulator + auto avg_squared_update = framework::EigenVector::Flatten( + *ctx.Input("AvgSquaredUpdate")); + auto param_out = framework::EigenVector::Flatten(*param_out_tensor); + auto avg_squared_grad_out = + framework::EigenVector::Flatten(*avg_squared_grad_out_tensor); + auto avg_squared_update_out = + framework::EigenVector::Flatten(*avg_squared_update_out_tensor); + auto place = ctx.GetEigenDevice(); + + avg_squared_grad_out.device(place) = + rho * avg_squared_grad + (1 - rho) * grad.square(); + auto update = + -((avg_squared_update + epsilon) / (avg_squared_grad_out + epsilon)) + .sqrt() * + grad; + avg_squared_update_out.device(place) = + rho * avg_squared_update + (1 - rho) * update.square(); + param_out.device(place) = param + update; + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/adagrad_op.cc b/paddle/operators/adagrad_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..a17747efb79d76a6beffb27bfc03ee1b62c5618d --- /dev/null +++ b/paddle/operators/adagrad_op.cc @@ -0,0 +1,93 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/operators/adagrad_op.h" + +namespace paddle { +namespace operators { + +class AdagradOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Param"), + "Input(Param) of AdagradOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Grad"), + "Input(Grad) of AdagradOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Moment"), + "Input(Moment) of AdagradOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("LearningRate"), + "Input(LearningRate) of AdagradOp should not be null."); + + PADDLE_ENFORCE(ctx->HasOutput("ParamOut"), + "Output(ParamOut) of AdagradOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("MomentOut"), + "Output(MomentOut) of AdagradOp should not be null."); + + auto lr_dims = ctx->GetInputDim("LearningRate"); + PADDLE_ENFORCE_EQ(framework::product(lr_dims), 1, + "LearningRate should have one element"); + auto param_dims = ctx->GetInputDim("Param"); + PADDLE_ENFORCE_EQ( + param_dims, ctx->GetInputDim("Grad"), + "Param and Grad input of AdagradOp should have the same dimension."); + PADDLE_ENFORCE_EQ( + param_dims, ctx->GetInputDim("Moment"), + "Param and Moment input of AdagradOp should have the same dimension."); + + ctx->SetOutputDim("ParamOut", param_dims); + ctx->SetOutputDim("MomentOut", param_dims); + } +}; + +class AdagradOpMaker : public framework::OpProtoAndCheckerMaker { + public: + AdagradOpMaker(framework::OpProto *proto, + framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("Param", "(Tensor) Input parameter"); + AddInput("Grad", "(Tensor) Input gradient"); + AddInput("Moment", "(Tensor) Second moment"); + AddInput("LearningRate", "(Tensor) Learning rate"); + + AddOutput("ParamOut", "(Tensor) Output parameter"); + AddOutput("MomentOut", "(Tensor) Output second moment"); + + AddAttr("epsilon", + "(float, default 1.0e-6) " + "Constant for numerical stability") + .SetDefault(1.0e-6f); + AddComment(R"DOC( + +Adaptive Gradient Algorithm (Adagrad). + +moment_out = moment + grad * grad +param_out = param - learning_rate * grad / (sqrt(moment_out) + epsilon) + +The original paper(http://www.jmlr.org/papers/volume12/duchi11a/duchi11a.pdf) +does not have the epsilon attribute. It is added here for numerical stability +by avoiding division by zero. + +)DOC"); + } +}; +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP_WITHOUT_GRADIENT(adagrad, ops::AdagradOp, ops::AdagradOpMaker); +REGISTER_OP_CPU_KERNEL(adagrad, + ops::AdagradOpKernel); diff --git a/paddle/operators/adagrad_op.cu b/paddle/operators/adagrad_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..a5b7951121360f78612f9008a522235104708112 --- /dev/null +++ b/paddle/operators/adagrad_op.cu @@ -0,0 +1,20 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#define EIGEN_USE_GPU +#include "paddle/operators/adagrad_op.h" + +namespace ops = paddle::operators; +REGISTER_OP_GPU_KERNEL(adagrad, + ops::AdagradOpKernel); diff --git a/paddle/operators/adagrad_op.h b/paddle/operators/adagrad_op.h new file mode 100644 index 0000000000000000000000000000000000000000..c5d8f751d3527f89b96d4274328ba0bb5f6efa44 --- /dev/null +++ b/paddle/operators/adagrad_op.h @@ -0,0 +1,55 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" + +namespace paddle { +namespace operators { + +template +class AdagradOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto param_out_tensor = ctx.Output("ParamOut"); + auto moment_out_tensor = ctx.Output("MomentOut"); + + param_out_tensor->mutable_data(ctx.GetPlace()); + moment_out_tensor->mutable_data(ctx.GetPlace()); + + float epsilon = ctx.Attr("epsilon"); + + auto param = framework::EigenVector::Flatten( + *ctx.Input("Param")); + auto grad = framework::EigenVector::Flatten( + *ctx.Input("Grad")); + auto moment = framework::EigenVector::Flatten( + *ctx.Input("Moment")); + auto lr = framework::EigenVector::Flatten( + *ctx.Input("LearningRate")); + + auto param_out = framework::EigenVector::Flatten(*param_out_tensor); + auto moment_out = framework::EigenVector::Flatten(*moment_out_tensor); + auto place = ctx.GetEigenDevice(); + + moment_out.device(place) = moment + grad * grad; + Eigen::DSizes m_dsize(moment_out_tensor->numel()); + param_out.device(place) = + param - lr.broadcast(m_dsize) * grad / (moment_out.sqrt() + epsilon); + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/adamax_op.cc b/paddle/operators/adamax_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..5cf727742c770f8f0d59cd49345c5c055722a56c --- /dev/null +++ b/paddle/operators/adamax_op.cc @@ -0,0 +1,139 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/operators/adamax_op.h" + +namespace paddle { +namespace operators { + +class AdamaxOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Param"), + "Input(Param) of AdamaxOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Grad"), + "Input(Grad) of AdamaxOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Moment"), + "Input(Moment) of AdamaxOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("InfNorm"), + "Input(InfNorm) of AdamaxOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("LearningRate"), + "Input(LearningRate) of AdamaxOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Beta1Pow"), + "Input(Beta1Pow) of AdamaxOp should not be null."); + + PADDLE_ENFORCE(ctx->HasOutput("ParamOut"), + "Output(ParamOut) of AdamaxOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("MomentOut"), + "Output(MomentOut) of AdamaxOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("InfNormOut"), + "Output(InfNormOut) of AdamaxOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Beta1PowOut"), + "Output(Beta1PowOut) of AdamaxOp should not be null."); + + auto lr_dims = ctx->GetInputDim("LearningRate"); + PADDLE_ENFORCE_EQ(framework::product(lr_dims), 1, + "Learning rate should have 1 dimension"); + auto beta1_pow_dims = ctx->GetInputDim("Beta1Pow"); + PADDLE_ENFORCE_EQ(framework::product(beta1_pow_dims), 1, + "Beta1 power accumulator should have 1 dimension"); + auto param_dims = ctx->GetInputDim("Param"); + PADDLE_ENFORCE_EQ( + param_dims, ctx->GetInputDim("Grad"), + "Param and Grad input of AdamaxOp should have same dimension"); + PADDLE_ENFORCE_EQ( + param_dims, ctx->GetInputDim("Moment"), + "Param and Moment input of AdamaxOp should have same dimension"); + PADDLE_ENFORCE_EQ( + param_dims, ctx->GetInputDim("InfNorm"), + "Param and InfNorm input of AdamaxOp should have same dimension"); + + ctx->SetOutputDim("ParamOut", param_dims); + ctx->SetOutputDim("MomentOut", param_dims); + ctx->SetOutputDim("InfNormOut", param_dims); + ctx->SetOutputDim("Beta1PowOut", beta1_pow_dims); + } +}; + +class AdamaxOpMaker : public framework::OpProtoAndCheckerMaker { + public: + AdamaxOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("Param", "(Tensor) Input parameter"); + AddInput("Grad", "(Tensor) Input gradient"); + AddInput("LearningRate", "(Tensor) Learning rate"); + AddInput("Moment", "(Tensor) First moment"); + AddInput("InfNorm", + "(Tensor) " + "Input exponentially weighted infinity norm"); + AddInput("Beta1Pow", "(Tensor) Input beta1 power accumulator"); + + AddOutput("ParamOut", "(Tensor) Output parameter"); + AddOutput("MomentOut", "(Tensor) Output first moment"); + AddOutput("InfNormOut", + "(Tensor) " + "Output exponentially weighted infinity norm"); + AddOutput("Beta1PowOut", "(Tensor) Output beta1 power accumulator"); + + AddAttr("beta1", + "(float, default 0.9) " + "Exponential decay rate for the " + "1st moment estimates.") + .SetDefault(0.9f); + AddAttr("beta2", + "(float, default 0.999) " + "exponential decay rate for the weighted " + "infinity norm estimates.") + .SetDefault(0.999f); + AddAttr("epsilon", + "(float, default 1.0e-8) " + "Constant for numerical stability") + .SetDefault(1.0e-8f); + AddComment(R"DOC( +Adamax Updates Operator. + +This implements the Adamax optimizer from Section 7 of the Adam +paper[1]. Adamax is a variant of the +Adam algorithm based on the infinity norm. + +Adamax updates: + +moment_out = beta1 * moment + (1 - beta1) * grad +inf_norm_out = max(beta2 * inf_norm + epsilon, abs(grad)) +beta1_pow_out = beta1_pow * beta1 +learning_rate_t = learning_rate/(1 - beta1_pow_out) +param_out = param - learning_rate_t * moment_out/inf_norm_out + +The original paper does not have an epsilon attribute. +However, it is added here for numerical stability +by preventing divide by 0. + +References: + [1] Adam: A Method for Stochastic Optimization + (https://arxiv.org/abs/1412.6980) + +)DOC"); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP_WITHOUT_GRADIENT(adamax, ops::AdamaxOp, ops::AdamaxOpMaker); +REGISTER_OP_CPU_KERNEL(adamax, + ops::AdamaxOpKernel); diff --git a/paddle/operators/add_op.cu b/paddle/operators/adamax_op.cu similarity index 80% rename from paddle/operators/add_op.cu rename to paddle/operators/adamax_op.cu index d9c6d20a6c320b59e57ed25da3dd8b093833f8c7..fee3b6fc6b656917d79b84f48da8e63be7683890 100644 --- a/paddle/operators/add_op.cu +++ b/paddle/operators/adamax_op.cu @@ -12,7 +12,9 @@ See the License for the specific language governing permissions and limitations under the License. */ -#include "paddle/operators/add_op.h" +#define EIGEN_USE_GPU +#include "paddle/operators/adamax_op.h" namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL(add, ops::AddKernel); +REGISTER_OP_GPU_KERNEL(adamax, + ops::AdamaxOpKernel); diff --git a/paddle/operators/adamax_op.h b/paddle/operators/adamax_op.h new file mode 100644 index 0000000000000000000000000000000000000000..9677b1bb786002aadfaeb571b2ba2e6aa2481ca5 --- /dev/null +++ b/paddle/operators/adamax_op.h @@ -0,0 +1,72 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" + +namespace paddle { +namespace operators { + +template +class AdamaxOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto param_out_tensor = ctx.Output("ParamOut"); + auto moment_out_tensor = ctx.Output("MomentOut"); + auto inf_norm_out_tensor = ctx.Output("InfNormOut"); + auto beta1_pow_out_tensor = ctx.Output("Beta1PowOut"); + + param_out_tensor->mutable_data(ctx.GetPlace()); + moment_out_tensor->mutable_data(ctx.GetPlace()); + inf_norm_out_tensor->mutable_data(ctx.GetPlace()); + beta1_pow_out_tensor->mutable_data(ctx.GetPlace()); + + float beta1 = ctx.Attr("beta1"); + float beta2 = ctx.Attr("beta2"); + float epsilon = ctx.Attr("epsilon"); + + auto param = framework::EigenVector::Flatten( + *ctx.Input("Param")); + auto grad = framework::EigenVector::Flatten( + *ctx.Input("Grad")); + auto moment = framework::EigenVector::Flatten( + *ctx.Input("Moment")); + auto inf_norm = framework::EigenVector::Flatten( + *ctx.Input("InfNorm")); + auto lr = framework::EigenVector::Flatten( + *ctx.Input("LearningRate")); + auto beta1_pow = framework::EigenVector::Flatten( + *ctx.Input("Beta1Pow")); + auto param_out = framework::EigenVector::Flatten(*param_out_tensor); + auto moment_out = framework::EigenVector::Flatten(*moment_out_tensor); + auto inf_norm_out = + framework::EigenVector::Flatten(*inf_norm_out_tensor); + auto beta1_pow_out = + framework::EigenVector::Flatten(*beta1_pow_out_tensor); + auto place = ctx.GetEigenDevice(); + + moment_out.device(place) = beta1 * moment + (1 - beta1) * grad; + inf_norm_out.device(place) = + grad.abs().cwiseMax((beta2 * inf_norm) + epsilon); + beta1_pow_out.device(place) = beta1_pow * beta1; + auto lr_t = lr / (1 - beta1_pow_out); + Eigen::DSizes m_dsize(moment_out_tensor->numel()); + param_out.device(place) = + param - lr_t.broadcast(m_dsize) * (moment_out / inf_norm_out); + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/add_op.cc b/paddle/operators/add_op.cc deleted file mode 100644 index ed11d096974341022637676537793645f46738f0..0000000000000000000000000000000000000000 --- a/paddle/operators/add_op.cc +++ /dev/null @@ -1,70 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "paddle/operators/add_op.h" - -namespace paddle { -namespace operators { - -class AddOp : public framework::OperatorWithKernel { - public: - using framework::OperatorWithKernel::OperatorWithKernel; - - protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), - "Input(X) of AddOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), - "Input(Y) of AddOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"), - "Output(Out) of AddOp should not be null."); - - PADDLE_ENFORCE_EQ(ctx.Input("X")->dims(), - ctx.Input("Y")->dims(), - "Two input of Add Op's dimension must be same."); - ctx.Output("Out")->Resize( - ctx.Input("X")->dims()); - } -}; - -class AddOpMaker : public framework::OpProtoAndCheckerMaker { - public: - AddOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("X", "The first input of add op"); - AddInput("Y", "The second input of add op"); - AddOutput("Out", "The output of add op"); - AddComment(R"DOC( -Two Element Add Operator. - -The equation is: Out = X + Y -)DOC"); - } -}; - -class AddOpGrad : public framework::OperatorWithKernel { - public: - using framework::OperatorWithKernel::OperatorWithKernel; - - protected: - void InferShape(const framework::InferShapeContext &ctx) const override {} -}; - -} // namespace operators -} // namespace paddle - -namespace ops = paddle::operators; -REGISTER_OP(add, ops::AddOp, ops::AddOpMaker, add_grad, ops::AddOpGrad); - -REGISTER_OP_CPU_KERNEL(add, ops::AddKernel); diff --git a/paddle/operators/clip_op.cc b/paddle/operators/clip_op.cc index e5a54bc4b226fd24337050fdd84b2de9c49f7949..3e9b0d82ba06a918d52124e791715277640fd4ff 100644 --- a/paddle/operators/clip_op.cc +++ b/paddle/operators/clip_op.cc @@ -22,28 +22,28 @@ class ClipOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), - "Input(X) of ClipOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"), - "Output(Out) of ClipOp should not be null."); - auto x_dims = ctx.Input("X")->dims(); + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of ClipOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of ClipOp should not be null."); + auto x_dims = ctx->GetInputDim("X"); auto max = Attr("max"); auto min = Attr("min"); PADDLE_ENFORCE_LT(min, max, "max should be greater than min."); - ctx.Output("Out")->Resize(x_dims); - ctx.ShareLoD("X", /*->*/ "Out"); + ctx->SetOutputDim("Out", x_dims); + ctx->ShareLoD("X", /*->*/ "Out"); } }; template class ClipOpMaker : public framework::OpProtoAndCheckerMaker { public: - ClipOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + ClipOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "(Tensor)The input of clip op." - "The input should be a k-D tensor(k > 0 and k < 7)"); + "The number of dimensions must be between [1, 9]."); AddOutput("Out", "(Tensor)The output of clip op with shape as input(X)"); AddAttr( "min", "(float)Minimum value, under which element is replaced by min."); @@ -61,14 +61,13 @@ class ClipOpGrad : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) should not be null"); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")), - "Input(Out@GRAD) should not be null"); - auto x_dims = ctx.Input("X")->dims(); - auto *x_grad = ctx.Output(framework::GradVarName("X")); - if (x_grad != nullptr) { - x_grad->Resize(x_dims); + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null"); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), + "Input(Out@GRAD) should not be null"); + auto x_dims = ctx->GetInputDim("X"); + if (ctx->HasOutput(framework::GradVarName("X"))) { + ctx->SetOutputDim(framework::GradVarName("X"), x_dims); } } }; diff --git a/paddle/operators/clip_op.h b/paddle/operators/clip_op.h index ce1d4e1f460414e6e4acee4fa3207f309c55d86b..ac702e9935201ba5263a80ebeb1ab22fa0bd1340 100644 --- a/paddle/operators/clip_op.h +++ b/paddle/operators/clip_op.h @@ -56,7 +56,7 @@ class ClipGradFunctor { }; template -class ClipKernel : public framework::OpKernel { +class ClipKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto max = context.Attr("max"); @@ -73,7 +73,7 @@ class ClipKernel : public framework::OpKernel { }; template -class ClipGradKernel : public framework::OpKernel { +class ClipGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto max = context.Attr("max"); diff --git a/paddle/operators/concat_op.cc b/paddle/operators/concat_op.cc index 07f847079e834716904dcc038d2097efd268bd3e..235c4449ace79e87d209615199d6628f269961a9 100644 --- a/paddle/operators/concat_op.cc +++ b/paddle/operators/concat_op.cc @@ -24,31 +24,32 @@ class ConcatOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"), - "Output(Out) of ConcatOp should not be null."); + void InferShape(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE_GE(ctx->Inputs("X").size(), 1UL, + "Inputs(X) of ConcatOp should be empty.") + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of ConcatOp should not be null."); - auto ins = ctx.MultiInput("X"); - auto *out = ctx.Output("Out"); - size_t axis = static_cast(ctx.Attr("axis")); - size_t n = ins.size(); + auto ins = ctx->GetInputsDim("X"); + size_t axis = static_cast(ctx->Attrs().Get("axis")); + const size_t n = ins.size(); PADDLE_ENFORCE_GT(n, 1, "Input tensors count should > 1."); - auto out_dims = ins[0]->dims(); + auto out_dims = ins[0]; size_t in_zero_dims_size = out_dims.size(); for (size_t i = 1; i < n; i++) { for (size_t j = 0; j < in_zero_dims_size; j++) { if (j == axis) { - out_dims[axis] += ins[i]->dims()[j]; + out_dims[axis] += ins[i][j]; continue; } - PADDLE_ENFORCE_EQ(out_dims[j], ins[i]->dims()[j], + PADDLE_ENFORCE_EQ(out_dims[j], ins[i][j], "Input tensors should have the same " "elements except the specify axis.") } } - out->Resize(out_dims); + ctx->SetOutputDim("Out", out_dims); } }; @@ -73,10 +74,27 @@ class ConcatOpMaker : public framework::OpProtoAndCheckerMaker { } }; +class ConcatOpGrad : public framework::OperatorWithKernel { + public: + ConcatOpGrad(const std::string &type, + const framework::VariableNameMap &inputs, + const framework::VariableNameMap &outputs, + const framework::AttributeMap &attrs) + : OperatorWithKernel(type, inputs, outputs, attrs) {} + + protected: + void InferShape(framework::InferShapeContext *ctx) const override { + ctx->SetOutputsDim(framework::GradVarName("X"), ctx->GetInputsDim("X")); + } +}; + } // namespace operators } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP_WITHOUT_GRADIENT(concat, ops::ConcatOp, ops::ConcatOpMaker) +REGISTER_OP(concat, ops::ConcatOp, ops::ConcatOpMaker, concat_grad, + ops::ConcatOpGrad) REGISTER_OP_CPU_KERNEL(concat, ops::ConcatKernel) +REGISTER_OP_CPU_KERNEL(concat_grad, + ops::ConcatGradKernel) diff --git a/paddle/operators/concat_op.cu b/paddle/operators/concat_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..ede832ddcd486729db56bba016683b33875f8837 --- /dev/null +++ b/paddle/operators/concat_op.cu @@ -0,0 +1,20 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/operators/concat_op.h" +namespace ops = paddle::operators; +REGISTER_OP_GPU_KERNEL(concat, + ops::ConcatKernel); +REGISTER_OP_GPU_KERNEL( + concat_grad, ops::ConcatGradKernel); diff --git a/paddle/operators/concat_op.h b/paddle/operators/concat_op.h index f977054fdf8aa0164db726b94a21c57f770dd674..c113f19fb5cf806709bff845ee0f1078b34014bb 100644 --- a/paddle/operators/concat_op.h +++ b/paddle/operators/concat_op.h @@ -16,46 +16,51 @@ limitations under the License. */ #include #include "paddle/framework/op_registry.h" +#include "paddle/operators/strided_memcpy.h" namespace paddle { namespace operators { template -class ConcatKernel : public framework::OpKernel { +class ConcatKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { auto ins = ctx.MultiInput("X"); auto* out = ctx.Output("Out"); int64_t axis = static_cast(ctx.Attr("axis")); - size_t n = ins.size(); - size_t output_axis_dim = 0; - size_t before = 1, after = 1; - for (size_t i = 0; i < n; i++) { - output_axis_dim += ins[i]->dims()[axis]; - } - auto& input_zero = ins[0]; - for (int64_t i = 0; i < input_zero->dims().size(); i++) { - if (i == axis) { - continue; - } - if (i < axis) { - before *= input_zero->dims()[i]; - } else { - after *= input_zero->dims()[i]; - } - } + const size_t n = ins.size(); size_t output_offset = 0; + out->mutable_data(ctx.GetPlace()); + auto out_stride = framework::stride(out->dims()); for (size_t i = 0; i < n; i++) { auto& in = ins[i]; auto axis_dim = in->dims()[axis]; - for (size_t j = 0; j < before; j++) { - size_t len = axis_dim * after * sizeof(T); - const T* src = in->data() + axis_dim * after * j; - T* out_data = out->mutable_data(platform::CPUPlace()); - T* dest = out_data + output_offset + output_axis_dim * after * j; - memcpy(dest, src, len); - } - output_offset += axis_dim * after; + auto in_stride = framework::stride(in->dims()); + StridedMemcpy(ctx.device_context(), in->data(), in_stride, + in->dims(), out_stride, out->data() + output_offset); + output_offset += axis_dim * in_stride[axis]; + } + } +}; + +template +class ConcatGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const { + auto* in = ctx.Input(framework::GradVarName("Out")); + auto outs = ctx.MultiOutput(framework::GradVarName("X")); + int64_t axis = static_cast(ctx.Attr("axis")); + const size_t n = outs.size(); + size_t input_offset = 0; + auto in_stride = framework::stride(in->dims()); + for (size_t i = 0; i < n; i++) { + auto& out = outs[i]; + out->mutable_data(ctx.GetPlace()); + size_t axis_dim = out->dims()[axis]; + auto out_stride = framework::stride(out->dims()); + StridedMemcpy(ctx.device_context(), in->data() + input_offset, + in_stride, out->dims(), out_stride, out->data()); + input_offset += axis_dim * in_stride[axis]; } } }; diff --git a/paddle/operators/cond_op.cc b/paddle/operators/cond_op.cc index 8262a7a5c8c13c86c5f6c123a14fa89696358c57..2737104a205cbc1e18ce4a3a45592a416d38a874 100644 --- a/paddle/operators/cond_op.cc +++ b/paddle/operators/cond_op.cc @@ -14,12 +14,7 @@ limitations under the License. */ #include "paddle/operators/cond_op.h" -#include -#include - -#include "paddle/framework/op_registry.h" #include "paddle/operators/gather.h" -#include "paddle/operators/net_op.h" #include "paddle/operators/scatter.h" namespace paddle { @@ -31,175 +26,183 @@ using Tensor = framework::Tensor; using LoDTensor = framework::LoDTensor; using DDim = framework::DDim; -void CondOp::CreateScope(const Scope& scope) const { +framework::Scope& CondOp::AddSubScope(const Scope& scope) const { auto sub_scopes_var = scope.FindVar("SubScopes"); PADDLE_ENFORCE_NOT_NULL(sub_scopes_var, "Output(SubScopes) of CondOp should not be null."); auto sub_scopes = sub_scopes_var->GetMutable>(); auto& sub_scope = scope.NewScope(); sub_scopes->push_back(&sub_scope); + return sub_scope; } -void CondOp::CreateIndexTensor(const Scope& scope) const { +std::vector& CondOp::GetSubScopes( + const framework::Scope& scope) const { + auto sub_scopes_var = scope.FindVar("SubScopes"); + PADDLE_ENFORCE_NOT_NULL(sub_scopes_var, + "Output(SubScopes) of CondOp should not be null."); + return *sub_scopes_var->GetMutable>(); +} + +LoDTensor& CondOp::AddIndexTensor(const Scope& scope) const { auto index_tensors_var = scope.FindVar("IndexTensors"); PADDLE_ENFORCE_NOT_NULL(index_tensors_var, "Output(IndexTensors) of CondOp should not be null."); auto& index_tensors = *index_tensors_var->GetMutable>(); index_tensors.push_back(LoDTensor()); + return index_tensors.back(); } -void CondOp::InferShape(const Scope& scope) const { - auto sub_scopes_var = scope.FindVar("SubScopes"); - PADDLE_ENFORCE_NOT_NULL(sub_scopes_var, - "Output(SubScopes) of CondOp should not be null."); - auto& sub_scopes = *sub_scopes_var->GetMutable>(); - - for (int i = 0; i < 2; ++i) { - // Create two sub scopes for true and false branches - // sub_scopes[0] for the true branch and sub_scopes[1] for the false - // branch - CreateScope(scope); - - // Create two tensors for true and false indices - // index_tensors[0] for the true branch and index_tensors[1] for the false - // branch - CreateIndexTensor(scope); - - PADDLE_ENFORCE(!Inputs("Xs").empty(), - "Inputs(Xs) of CondOp can't be empty."); - for (auto& input : Inputs("Xs")) { - // Create a new tensor in sub-scope for input-type tensor - Variable* v = sub_scopes[i]->NewVar(input); - LoDTensor* sub_input = v->GetMutable(); - sub_input->Resize(scope.FindVar(input)->GetMutable()->dims()); - } - - for (auto& output : (*sub_net_op_[i]).Outputs()) { - for (auto& var_name : output.second) { - sub_scopes[i]->NewVar(var_name); - } - } - - // each net calls InferShape - sub_net_op_[i]->InferShape(*sub_scopes[i]); - } - - for (auto& output : Outputs("Outs")) { - LoDTensor* tensor_t_out = - sub_scopes[0]->FindVar(output)->GetMutable(); - PADDLE_ENFORCE_NOT_NULL(tensor_t_out, "True output should not be NULL"); - LoDTensor* tensor_f_out = - sub_scopes[1]->FindVar(output)->GetMutable(); - PADDLE_ENFORCE_NOT_NULL(tensor_f_out, "False output should not be NULL"); - - auto* tensor_out_var = scope.FindVar(output); - PADDLE_ENFORCE_NOT_NULL(tensor_out_var, "Output not found"); - LoDTensor* tensor_out = tensor_out_var->GetMutable(); - PADDLE_ENFORCE_NOT_NULL(tensor_t_out, - "True output tensor should not be NULL"); - - // check output size should be same - PADDLE_ENFORCE_EQ(tensor_t_out->dims(), tensor_f_out->dims(), - "Outputs not of the same shape"); - tensor_out->Resize(tensor_t_out->dims()); - // tensor_out->mutable_data(tensor_out->dims(), - // platform::CPUPlace()); - tensor_out->mutable_data(platform::CPUPlace()); - } -} - -void CondOp::Run(const Scope& scope, - const platform::DeviceContext& dev_ctx) const { - auto* sub_scopes_var = scope.FindVar("SubScopes"); - PADDLE_ENFORCE_NOT_NULL(sub_scopes_var, - "Output(SubScopes) of CondOp should not be null."); - auto sub_scopes = sub_scopes_var->Get>(); +std::vector& CondOp::GetIndexTensors( + const framework::Scope& scope) const { auto* index_tensors_var = scope.FindVar("IndexTensors"); PADDLE_ENFORCE_NOT_NULL(index_tensors_var, "Output(IndexTensors) of CondOp should not be null."); - auto index_tensors = index_tensors_var->Get>(); + return *index_tensors_var->GetMutable>(); +} - std::string cond_name = Input("Cond"); - Variable* cond_var = scope.FindVar(cond_name); +void CondOp::PrepareDataForSubnet( + const framework::Scope& scope, + const platform::DeviceContext& dev_ctx) const { + PADDLE_ENFORCE(!Inputs("Xs").empty(), "Inputs(Xs) of CondOp can't be empty."); + + for (int i = 0; i < BRANCH_NUM; ++i) { + // Create two sub scopes for true and false branches + // sub_scopes[0] for the true branch + // sub_scopes[1] for the false branch + AddSubScope(scope); + // Create two tensors for true and false indices: + // index_tensors[0] for the true branch + // index_tensors[1] for the false branch + AddIndexTensor(scope); + } + + Variable* cond_var = scope.FindVar(Input("Cond")); PADDLE_ENFORCE_NOT_NULL(cond_var, "Input(Cond) of CondOp should not be null."); const LoDTensor* cond = cond_var->GetMutable(); - // Step 1: get the true/false index at runtime - // index_[0]: vector, contains all index for cond[i] == true - // index_[1]: vector, contains all index for cond[i] == false - for (int i = 0; i < 2; ++i) index_[i].clear(); + // get the true/false index at runtime according to cond tensor + // index_vectors[0]: vector, contains all index for cond[i] == true + // index_vectors[1]: vector, contains all index for cond[i] == false + std::vector> index_vectors; + index_vectors.resize(BRANCH_NUM); const int* cond_data = cond->data(); for (int i = 0; i < cond->dims()[0]; ++i) { if (cond_data[i]) - index_[0].push_back(i); + index_vectors[TRUE_BRANCH].push_back(i); else - index_[1].push_back(i); + index_vectors[FALSE_BRANCH].push_back(i); } - // put index_[0] and index_[1] into two tensors: - // index_tensor_[0] and index_tensor_[1] - DDim dim = paddle::framework::make_ddim({0}); - for (int i = 0; i < 2; ++i) { - dim[0] = index_[i].size(); - int* tmp_ptr = + // put index_vectors[0] and index_vectors[1] into two tensors: + // index_tensors[0] and index_tensors[1] + std::vector& index_tensors = GetIndexTensors(scope); + std::vector& sub_scopes = GetSubScopes(scope); + + for (int i = 0; i < BRANCH_NUM; ++i) { + DDim dim = {static_cast(index_vectors[i].size())}; + int* index_tensor_data_ptr = index_tensors[i].mutable_data(dim, platform::CPUPlace()); - index_tensors[i].Resize(dim); - memcpy(tmp_ptr, index_[i].data(), dim[0] * sizeof(int)); + memcpy(index_tensor_data_ptr, index_vectors[i].data(), + dim[0] * sizeof(int)); } - // Step 2: collect data by calling gather - for (int i = 0; i < 2; ++i) { - // i= 0/i for True and False branches respectively - for (auto& input : Inputs("Xs")) { - // find Tensor - Variable* v = scope.FindVar(input); - PADDLE_ENFORCE_NOT_NULL(v); - LoDTensor* tensor_parent = v->GetMutable(); + // create input in subscopes according to index_vectors + for (auto& input : Inputs("Xs")) { + Variable* var_parent = scope.FindVar(input); + PADDLE_ENFORCE_NOT_NULL(var_parent); + const auto* tensor_parent = &var_parent->Get(); - v = sub_scopes[i]->FindVar(input); - PADDLE_ENFORCE_NOT_NULL(v); - LoDTensor* tensor_child = v->GetMutable(); + for (int i = 0; i < BRANCH_NUM; ++i) { + Variable* var_child = sub_scopes[i]->FindVar(input); + PADDLE_ENFORCE_NOT_NULL(var_child); + auto* tensor_child = var_child->GetMutable(); // Resize child - DDim dim = tensor_child->dims(); - dim[0] = index_[i].size(); - tensor_child->Resize(dim); + DDim dim = tensor_parent->dims(); + dim[0] = index_tensors[i].dims()[0]; tensor_child->mutable_data(dim, platform::CPUPlace()); - Gather(dev_ctx.GetPlace(), tensor_parent, &index_tensors[i], - tensor_child); + CPUGather(dev_ctx, *tensor_parent, index_tensors[i], tensor_child); } } - // Step 3: run - for (int i = 0; i < 2; ++i) { - sub_net_op_[i]->Run(*sub_scopes[i], dev_ctx); + // create output_tensors in subscope for sub_net + for (int i = 0; i < BRANCH_NUM; ++i) { + for (auto& output : (*sub_net_op_[i]).Outputs()) { + for (auto& var_name : output.second) { + sub_scopes[i]->NewVar(var_name); + } + } } +} - // Step 4: merge output results +void CondOp::MergeDataFromSubnet(const framework::Scope& scope, + const platform::DeviceContext& dev_ctx) const { + std::vector& sub_scopes = GetSubScopes(scope); + const std::vector& index_tensors = + GetIndexTensors(scope); + + // Infer the output dim, out_dim[0] = true_dim[0] + false_dim[0] PADDLE_ENFORCE(!Outputs("Outs").empty(), "Outputs(Outs) of CondOp can't be empty."); - for (int i = 0; i < 2; ++i) { - // i= 0/i for True and False branches respectively - for (auto& output : Outputs("Outs")) { - // find Tensor - Variable* v = scope.FindVar(output); - PADDLE_ENFORCE_NOT_NULL(v); - LoDTensor* tensor_parent = v->GetMutable(); - - v = sub_scopes[i]->FindVar(output); - PADDLE_ENFORCE_NOT_NULL(v); - LoDTensor* tensor_child = v->GetMutable(); - - ScatterUpdate(dev_ctx.GetPlace(), tensor_child, &index_tensors[i], + for (auto& output : Outputs("Outs")) { + const LoDTensor* tensor_t_out = + &sub_scopes[TRUE_BRANCH]->FindVar(output)->Get(); + PADDLE_ENFORCE_NOT_NULL(tensor_t_out, "True output should not be NULL"); + const LoDTensor* tensor_f_out = + &sub_scopes[FALSE_BRANCH]->FindVar(output)->Get(); + PADDLE_ENFORCE_NOT_NULL(tensor_f_out, "False output should not be NULL"); + + auto* var_out = scope.FindVar(output); + PADDLE_ENFORCE_NOT_NULL(var_out, "Output not found"); + LoDTensor* tensor_out = var_out->GetMutable(); + PADDLE_ENFORCE_NOT_NULL(tensor_t_out, + "True output tensor should not be NULL"); + + DDim true_dim = tensor_t_out->dims(); + DDim false_dim = tensor_f_out->dims(); + true_dim[0] = 0; + false_dim[0] = 0; + PADDLE_ENFORCE_EQ(true_dim, false_dim, + "Outputs not of the same shape except the first dim"); + + DDim out_dim = tensor_t_out->dims(); + out_dim[0] = tensor_t_out->dims()[0] + tensor_f_out->dims()[0]; + tensor_out->Resize(out_dim); + tensor_out->mutable_data(platform::CPUPlace()); + } + + // merge output results: + // output_tensor = true_output_tensor + false_output_tensor + for (auto& output : Outputs("Outs")) { + Variable* var_parent = scope.FindVar(output); + PADDLE_ENFORCE_NOT_NULL(var_parent); + auto* tensor_parent = var_parent->GetMutable(); + + for (int i = 0; i < BRANCH_NUM; ++i) { + Variable* var_child = sub_scopes[i]->FindVar(output); + PADDLE_ENFORCE_NOT_NULL(var_child); + auto* tensor_child = &var_child->Get(); + ScatterAssign(dev_ctx, *tensor_child, index_tensors[i], tensor_parent); } } } +void CondOp::Run(const Scope& scope, + const platform::DeviceContext& dev_ctx) const { + PrepareDataForSubnet(scope, dev_ctx); + std::vector& sub_scopes = GetSubScopes(scope); + for (int i = 0; i < BRANCH_NUM; ++i) { + sub_net_op_[i]->Run(*sub_scopes[i], dev_ctx); + } + MergeDataFromSubnet(scope, dev_ctx); +} + class CondOpProtoAndCheckerMaker : public framework::OpProtoAndCheckerMaker { public: CondOpProtoAndCheckerMaker(framework::OpProto* proto, @@ -215,7 +218,7 @@ class CondOpProtoAndCheckerMaker : public framework::OpProtoAndCheckerMaker { AddComment(R"DOC( Sample dependent Cond Operator: Given Cond[i] as a 1/0 vector to indicate true/false -The equation is: +The equation is: Out[i] = subnet_t[i], if Cond[i] == true Out[i] = subnet_t[i], if Cond[i] == false )DOC"); diff --git a/paddle/operators/cond_op.h b/paddle/operators/cond_op.h index b09e32331e66c53555c88c06d7b1456276050eaa..93121fb31be287794249b5a62386d5a8dd268a0c 100644 --- a/paddle/operators/cond_op.h +++ b/paddle/operators/cond_op.h @@ -40,8 +40,7 @@ class CondOp : public framework::OperatorBase { const framework::VariableNameMap& outputs, const framework::AttributeMap& attrs) : OperatorBase(type, inputs, outputs, attrs) { - index_.resize(2); - sub_net_op_.resize(2); + sub_net_op_.resize(BRANCH_NUM); } CondOp(const CondOp& o) @@ -51,40 +50,44 @@ class CondOp : public framework::OperatorBase { PADDLE_THROW("Not implemented"); } - void CreateScope(const framework::Scope& scope) const; + framework::Scope& AddSubScope(const framework::Scope& scope) const; + std::vector& GetSubScopes( + const framework::Scope& scope) const; - void CreateIndexTensor(const framework::Scope& scope) const; + framework::LoDTensor& AddIndexTensor(const framework::Scope& scope) const; + std::vector& GetIndexTensors( + const framework::Scope& scope) const; - /* - * InferShape must be called before Run. - */ - void InferShape(const framework::Scope& scope) const override; + void PrepareDataForSubnet(const framework::Scope& scope, + const platform::DeviceContext& dev_ctx) const; + void MergeDataFromSubnet(const framework::Scope& scope, + const platform::DeviceContext& dev_ctx) const; /* * Set True Block */ void set_truenet(std::unique_ptr&& net) { - sub_net_op_[0] = std::move(net); + sub_net_op_[TRUE_BRANCH] = std::move(net); } /* * Set False Block */ void set_falsenet(std::unique_ptr&& net) { - sub_net_op_[1] = std::move(net); + sub_net_op_[FALSE_BRANCH] = std::move(net); } void Run(const framework::Scope& scope, const platform::DeviceContext& dev_ctx) const override; private: + const int TRUE_BRANCH = 0; + const int FALSE_BRANCH = 1; + const int BRANCH_NUM = 2; + // sub_net_op_[0]: subnet_t // sub_net_op_[1]: subnet_f std::vector> sub_net_op_; - - // index_[0]: True_index; - // index_[1]: False_index; - mutable std::vector> index_; }; } // namespace operators diff --git a/paddle/operators/conv2d_op.cc b/paddle/operators/conv2d_op.cc index c3281db0964de6d7dd6be629fbcc55cabb9fef9d..6325d4248f10ea8a12ae5398d9fe0e579db3f7ae 100644 --- a/paddle/operators/conv2d_op.cc +++ b/paddle/operators/conv2d_op.cc @@ -27,27 +27,25 @@ class Conv2DOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Input"), - "Input(Input) of Conv2DOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Filter"), - "Input(Filter) of Conv2DOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Output"), - "Output(Output) of Conv2DOp should not be null."); - - auto in = ctx.Input("Input"); - auto filter = ctx.Input("Filter"); - auto out = ctx.Output("Output"); - std::vector strides = Attr>("strides"); - std::vector paddings = Attr>("paddings"); - int groups = Attr("groups"); - int input_channels = in->dims()[1]; - int output_channels = filter->dims()[0]; - - PADDLE_ENFORCE_EQ(in->dims().size(), 4, "Conv2DOp input should be 4-D."); - PADDLE_ENFORCE_EQ(filter->dims().size(), 4, - "Conv2DOp filter should be 4-D."); - PADDLE_ENFORCE_EQ(input_channels, filter->dims()[1] * groups, + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Input"), + "Input(Input) of Conv2DOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Filter"), + "Input(Filter) of Conv2DOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Output"), + "Output(Output) of Conv2DOp should not be null."); + + auto in_dims = ctx->GetInputDim("Input"); + auto filter_dims = ctx->GetInputDim("Filter"); + std::vector strides = ctx->Attrs().Get>("strides"); + std::vector paddings = ctx->Attrs().Get>("paddings"); + int groups = ctx->Attrs().Get("groups"); + int input_channels = in_dims[1]; + int output_channels = filter_dims[0]; + + PADDLE_ENFORCE_EQ(in_dims.size(), 4, "Conv2DOp input should be 4-D."); + PADDLE_ENFORCE_EQ(filter_dims.size(), 4, "Conv2DOp filter should be 4-D."); + PADDLE_ENFORCE_EQ(input_channels, filter_dims[1] * groups, "The number of input channels should be equal to filter " "channels * groups."); PADDLE_ENFORCE_EQ( @@ -55,17 +53,17 @@ class Conv2DOp : public framework::OperatorWithKernel { "The number of output channels should be divided by groups."); auto output_height = - outputSize(in->dims()[2], filter->dims()[2], paddings[0], strides[0]); + outputSize(in_dims[2], filter_dims[2], paddings[0], strides[0]); auto output_width = - outputSize(in->dims()[3], filter->dims()[3], paddings[1], strides[1]); - out->Resize( - {in->dims()[0], filter->dims()[0], output_height, output_width}); + outputSize(in_dims[3], filter_dims[3], paddings[1], strides[1]); + ctx->SetOutputDim( + "Output", {in_dims[0], filter_dims[0], output_height, output_width}); } }; class Conv2DOpMaker : public framework::OpProtoAndCheckerMaker { public: - Conv2DOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + Conv2DOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput( "Input", @@ -108,14 +106,15 @@ class Conv2DOpGrad : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - auto in = ctx.Input("Input"); - auto filter = ctx.Input("Filter"); - auto d_in = ctx.Output(framework::GradVarName("Input")); - auto d_filter = - ctx.Output(framework::GradVarName("Filter")); - if (d_in) d_in->Resize(in->dims()); - if (d_filter) d_filter->Resize(filter->dims()); + void InferShape(framework::InferShapeContext* ctx) const override { + auto in_dims = ctx->GetInputDim("Input"); + auto filter_dims = ctx->GetInputDim("Filter"); + if (ctx->HasOutput(framework::GradVarName("Input"))) { + ctx->SetOutputDim(framework::GradVarName("Input"), in_dims); + } + if (ctx->HasOutput(framework::GradVarName("Filter"))) { + ctx->SetOutputDim(framework::GradVarName("Filter"), filter_dims); + } } }; diff --git a/paddle/operators/conv_shift_op.cc b/paddle/operators/conv_shift_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..e1e321ed5fce6ce4e4089cc5c5e488a2cbad6c82 --- /dev/null +++ b/paddle/operators/conv_shift_op.cc @@ -0,0 +1,206 @@ +/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/operators/conv_shift_op.h" +#include "paddle/framework/eigen.h" + +namespace paddle { +namespace operators { + +using framework::Tensor; +template +using EigenMatrix = framework::EigenMatrix; + +class ConvShiftOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should be not null."); + PADDLE_ENFORCE(ctx->HasInput("Y"), "Input(Y) should be not null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) should be not null."); + + auto x_dims = ctx->GetInputDim("X"); + auto y_dims = ctx->GetInputDim("Y"); + PADDLE_ENFORCE_EQ(x_dims.size(), 2, "Input(X)'s rank should be 2."); + PADDLE_ENFORCE_EQ(y_dims.size(), 2, "Input(Y)'s rank should be 2."); + PADDLE_ENFORCE_EQ(x_dims[0], y_dims[0], + "The 1st dimension of Input(X) and Input(Y) should " + "be equal."); + PADDLE_ENFORCE_EQ(y_dims[1] % 2, 1, + "The 2nd dimension of Input(Y) should be odd."); + PADDLE_ENFORCE_LE(y_dims[1], x_dims[1], + "The 2nd dimension of Input(Y) should be less than or " + "equal to the 2nd dimension of Input(X)."); + ctx->SetOutputDim("Out", x_dims); + ctx->ShareLoD("X", /*->*/ "Out"); + } +}; + +class ConvShiftGradOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should be not null."); + PADDLE_ENFORCE(ctx->HasInput("Y"), "Input(Y) should be not null."); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), + "Input(Out@GRAD) should be not null."); + + auto x_grad_name = framework::GradVarName("X"); + if (ctx->HasOutput(x_grad_name)) { + auto x_dims = ctx->GetInputDim("X"); + ctx->SetOutputDim(x_grad_name, x_dims); + } + + auto y_grad_name = framework::GradVarName("Y"); + if (ctx->HasOutput(y_grad_name)) { + auto y_dims = ctx->GetInputDim("Y"); + ctx->SetOutputDim(y_grad_name, y_dims); + } + } +}; + +class ConvShiftOpMaker : public framework::OpProtoAndCheckerMaker { + public: + ConvShiftOpMaker(framework::OpProto *proto, + framework::OpAttrChecker *op_checker) + : framework::OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", + "(Tensor, default Tensor), a 2-D tensor with shape B x M, " + "where B is the batch size and M is the data dimension."); + AddInput("Y", + "(Tensor, default Tensor), a 2-D tensor with shape B x N, " + "where B is the batch size and N is the data dimension. N must " + "be odd."); + AddOutput("Out", + "(Tensor, default Tensor), a 2-D tensor with shape B x M, " + "i.e., the same shape as X."); + AddComment(R"DOC( +ConvShift Operator. + +A layer for circular convolution of two vectors, +as used in the Neural Turing Machine: https://arxiv.org/abs/1410.5401 + +The equation is: + + \f[ + Out[i] = \sum_{j=-(N-1)/2}^{(N-1)/2} X_{i+j} * Y_{j} + \f] + +where X's index is computed modulo M, and b's index is computed modulo N. + +Both of the input `X` and `Y` can carry LoD (Level of Details) information. +However, the output only shares the LoD information with input `X`. +)DOC"); + } +}; + +template +class ConvShiftKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &context) const override { + auto *X = context.Input("X"); + auto *Y = context.Input("Y"); + auto *Out = context.Output("Out"); + Out->mutable_data(context.GetPlace()); + + auto x = EigenMatrix::From(*X); + auto y = EigenMatrix::From(*Y); + auto out = EigenMatrix::From(*Out); + out.setZero(); + + size_t batch_size = X->dims()[0]; + size_t x_width = X->dims()[1]; + size_t y_width = Y->dims()[1]; + size_t y_half_width = (y_width - 1) / 2; + + for (size_t k = 0; k < batch_size; ++k) { + for (size_t i = 0; i < x_width; ++i) { + for (size_t j = 0; j < y_width; ++j) { + int index = (i + j - y_half_width + x_width) % x_width; + out(k, i) += x(k, index) * y(k, j); + } + } + } + } +}; + +template +class ConvShiftGradKernel + : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &context) const override { + auto *X = context.Input("X"); + auto *Y = context.Input("Y"); + auto *dOut = context.Input(framework::GradVarName("Out")); + auto *dX = context.Output(framework::GradVarName("X")); + auto *dY = context.Output(framework::GradVarName("Y")); + + auto x = EigenMatrix::From(*X); + auto y = EigenMatrix::From(*Y); + auto dout = EigenMatrix::From(*dOut); + + auto x_dims = X->dims(); + auto y_dims = Y->dims(); + size_t batch_size = x_dims[0]; + size_t x_width = x_dims[1]; + size_t y_width = y_dims[1]; + size_t y_half_width = (y_width - 1) / 2; + + // The below trades code duplication for efficiency (keeping the if + // statement outside of the loop). + if (dX) { + dX->mutable_data(context.GetPlace()); + auto dx = EigenMatrix::From(*dX); + dx.setZero(); + for (size_t k = 0; k < batch_size; ++k) { + for (size_t i = 0; i < x_width; ++i) { + for (size_t j = 0; j < y_width; ++j) { + int index = (i + j - y_half_width + x_width) % x_width; + dx(k, index) += dout(k, i) * y(k, j); + } + } + } + } + + if (dY) { + dY->mutable_data(context.GetPlace()); + auto dy = EigenMatrix::From(*dY); + dy.setZero(); + for (size_t k = 0; k < batch_size; ++k) { + for (size_t i = 0; i < x_width; ++i) { + for (size_t j = 0; j < y_width; ++j) { + int index = (i + j - y_half_width + x_width) % x_width; + dy(k, j) += x(k, index) * dout(k, i); + } + } + } + } + } +}; +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP(conv_shift, ops::ConvShiftOp, ops::ConvShiftOpMaker, + conv_shift_grad, ops::ConvShiftGradOp); +REGISTER_OP_CPU_KERNEL(conv_shift, + ops::ConvShiftKernel); +REGISTER_OP_CPU_KERNEL( + conv_shift_grad, + ops::ConvShiftGradKernel); diff --git a/paddle/operators/conv_shift_op.cu b/paddle/operators/conv_shift_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..145e966fe9caa68f7485bb258fa78fd34bfd4c04 --- /dev/null +++ b/paddle/operators/conv_shift_op.cu @@ -0,0 +1,194 @@ +/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/operators/conv_shift_op.h" +#include "paddle/platform/cuda_helper.h" + +namespace paddle { +namespace operators { + +using framework::Tensor; + +namespace { + +inline int div_up(int x, int y) { return (x + y - 1) / y; } + +// Some notes on the design: +// +// Each thread is responsible for computing a single output out[k, i]. +// Thread blocks are based on tiles of x with height 1 in the batch dimension. +// +// This design is based on the typical use case where the filter +// y is fairly small. For large y, it would probably be more efficient +// to also tile across y. +template +__global__ void conv_shift_forward(const T *x, const T *y, T *out, int x_width, + int y_width, int y_half_width, + int batch_size) { + extern __shared__ T mem[]; + + int tx = threadIdx.x; + int i = blockIdx.x * blockDim.x + tx; // global x index + int k = blockIdx.y; // batch index + + // Check if we are in a boundary block with fewer x's to process than + // blockDim.x. + int num_x = + (blockIdx.x == gridDim.x - 1) ? (x_width % blockDim.x) : blockDim.x; + + T *sx = mem; + T *sx_pad = &mem[num_x]; + T *sy = &mem[blockDim.x + y_width]; + + // Collaboratively load y[k, :] and length-y padding of x into shared memory. + int pad_start = blockIdx.x * blockDim.x + num_x + x_width - y_half_width; + for (int j = tx; j < y_width; j += blockDim.x) { + sy[j] = y[k * y_width + j]; + sx_pad[j] = x[k * x_width + (pad_start + j) % x_width]; + } + + // Load a cyclically shifted slice of x into shared memory. + if (tx < num_x) { + int load_i = (i - y_half_width + x_width) % x_width; + sx[tx] = x[k * x_width + load_i]; + } else { + return; + } + __syncthreads(); + + // Compute dot product of sx[tx:tx + y_width] and sy. + T sum = 0; + for (int j = 0; j < y_width; ++j) { + sum += sx[tx + j] * sy[j]; + } + + // Save to out[k, i]. + out[k * x_width + i] = sum; +} + +// Compute x gradient - initial naive implementation with atomic add. +template +__global__ void conv_shift_dx(const T *dout, const T *y, T *dx, int x_width, + int y_width, int y_half_width, int batch_size) { + int i = blockIdx.x * blockDim.x + threadIdx.x; // x index + int j = blockIdx.y; // y index + int k = blockIdx.z; // batch index + + if (i < x_width) { + int index = (i + j - y_half_width + x_width) % x_width; + atomicAdd(&dx[k * x_width + index], + dout[k * x_width + i] * y[k * y_width + j]); + } +} + +// Compute y gradient - initial naive implementation with atomic add. +template +__global__ void conv_shift_dy(const T *x, const T *dout, T *dy, int x_width, + int y_width, int y_half_width, int batch_size) { + int i = blockIdx.x * blockDim.x + threadIdx.x; // x index + int j = blockIdx.y; // y index + int k = blockIdx.z; // batch index + + if (i < x_width) { + int index = (i + j - y_half_width + x_width) % x_width; + atomicAdd(&dy[k * y_width + j], + x[k * x_width + index] * dout[k * x_width + i]); + } +} +} // namespace + +template +class ConvShiftKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &context) const override { + const Tensor *X = context.Input("X"); + const Tensor *Y = context.Input("Y"); + Tensor *Out = context.Output("Out"); + const T *x_data = X->data(); + const T *y_data = Y->data(); + T *out_data = Out->mutable_data(context.GetPlace()); + + int batch_size = X->dims()[0]; + int x_width = X->dims()[1]; + int y_width = Y->dims()[1]; + int y_half_width = (y_width - 1) / 2; + + const int x_per_block = 256; + int num_x_blocks = div_up(x_width, x_per_block); + int mem_per_block = (x_per_block + 2 * y_width) * sizeof(T); + + dim3 grid_dim(num_x_blocks, batch_size); + + auto stream = reinterpret_cast( + context.device_context()) + .stream(); + + conv_shift_forward<<>>( + x_data, y_data, out_data, x_width, y_width, y_half_width, batch_size); + } +}; + +template +class ConvShiftGradKernel + : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &context) const override { + const Tensor *X = context.Input("X"); + const Tensor *Y = context.Input("Y"); + const Tensor *dOut = context.Input(framework::GradVarName("Out")); + const T *x_data = X->data(); + const T *y_data = Y->data(); + const T *dout_data = dOut->data(); + + Tensor *dX = context.Output(framework::GradVarName("X")); + Tensor *dY = context.Output(framework::GradVarName("Y")); + + int batch_size = X->dims()[0]; + int x_width = X->dims()[1]; + int y_width = Y->dims()[1]; + int y_half_width = (y_width - 1) / 2; + + auto stream = reinterpret_cast( + context.device_context()) + .stream(); + + const int x_per_block = 256; + int num_x_blocks = div_up(x_width, x_per_block); + dim3 grid_dim(num_x_blocks, y_width, batch_size); + + if (dX) { + T *dx_data = dX->mutable_data(context.GetPlace()); + cudaMemsetAsync(dx_data, 0, dX->numel() * sizeof(T), stream); + conv_shift_dx<<>>( + dout_data, y_data, dx_data, x_width, y_width, y_half_width, + batch_size); + } + if (dY) { + T *dy_data = dY->mutable_data(context.GetPlace()); + cudaMemsetAsync(dy_data, 0, dY->numel() * sizeof(T), stream); + conv_shift_dy<<>>( + x_data, dout_data, dy_data, x_width, y_width, y_half_width, + batch_size); + } + } +}; +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP_GPU_KERNEL(conv_shift, + ops::ConvShiftKernel); +REGISTER_OP_GPU_KERNEL( + conv_shift_grad, + ops::ConvShiftGradKernel); diff --git a/paddle/operators/conv_shift_op.h b/paddle/operators/conv_shift_op.h new file mode 100644 index 0000000000000000000000000000000000000000..5a160b0f1696c70868fc48d219b38cde2018e8a3 --- /dev/null +++ b/paddle/operators/conv_shift_op.h @@ -0,0 +1,33 @@ +/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once +#include "paddle/framework/op_registry.h" + +namespace paddle { +namespace operators { + +template +class ConvShiftKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &context) const override; +}; + +template +class ConvShiftGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &context) const override; +}; +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/cos_sim_op.cc b/paddle/operators/cos_sim_op.cc index b56ee2047b811e212b4bf74bf7fbba753a6bcb11..2b4c4b9c45d22e9d0b124145b55cfaeeaf1583d5 100644 --- a/paddle/operators/cos_sim_op.cc +++ b/paddle/operators/cos_sim_op.cc @@ -24,22 +24,22 @@ class CosSimOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { + void InferShape(framework::InferShapeContext* ctx) const override { // notnull check - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), - "Input(X) of CosSimOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), - "Input(Y) of CosSimOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"), - "Output(Out) of CosSimOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("XNorm"), - "Output(XNorm) of CosSimOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("YNorm"), - "Output(YNorm) of CosSimOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of CosSimOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Y"), + "Input(Y) of CosSimOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of CosSimOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("XNorm"), + "Output(XNorm) of CosSimOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("YNorm"), + "Output(YNorm) of CosSimOp should not be null."); // shape check - auto x_dims = ctx.Input("X")->dims(); - auto y_dims = ctx.Input("Y")->dims(); + auto x_dims = ctx->GetInputDim("X"); + auto y_dims = ctx->GetInputDim("Y"); PADDLE_ENFORCE_EQ(x_dims.size(), y_dims.size(), "Ranks of Input(X) and Input(Y) must be equal."); @@ -54,16 +54,16 @@ class CosSimOp : public framework::OperatorWithKernel { " just 1 (which will be broadcasted to match Input(X))."); // resize tensor - ctx.Output("Out")->Resize({x_dims[0], 1}); - ctx.Output("XNorm")->Resize({x_dims[0], 1}); - ctx.Output("YNorm")->Resize({y_dims[0], 1}); - ctx.ShareLoD("X", /*->*/ "Out"); + ctx->SetOutputDim("Out", {x_dims[0], 1}); + ctx->SetOutputDim("XNorm", {x_dims[0], 1}); + ctx->SetOutputDim("YNorm", {y_dims[0], 1}); + ctx->ShareLoD("X", /*->*/ "Out"); } }; class CosSimOpMaker : public framework::OpProtoAndCheckerMaker { public: - CosSimOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + CosSimOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "The 1st input of cos_sim op."); AddInput("Y", "The 2nd input of cos_sim op."); @@ -98,27 +98,23 @@ class CosSimOpGrad : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { + void InferShape(framework::InferShapeContext* ctx) const override { // notnull check - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) must not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), "Input(Y) must not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("XNorm"), - "Input(XNorm) must not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("YNorm"), - "Input(YNorm) must not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Out"), - "Input(Out) must not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")), - "Input(Out@GRAD) must not be null."); + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) must not be null."); + PADDLE_ENFORCE(ctx->HasInput("Y"), "Input(Y) must not be null."); + PADDLE_ENFORCE(ctx->HasInput("XNorm"), "Input(XNorm) must not be null."); + PADDLE_ENFORCE(ctx->HasInput("YNorm"), "Input(YNorm) must not be null."); + PADDLE_ENFORCE(ctx->HasInput("Out"), "Input(Out) must not be null."); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), + "Input(Out@GRAD) must not be null."); // shape check - auto x_dims = ctx.Input("X")->dims(); - auto y_dims = ctx.Input("Y")->dims(); - auto xnorm_dims = ctx.Input("XNorm")->dims(); - auto ynorm_dims = ctx.Input("YNorm")->dims(); - auto out_dims = ctx.Input("Out")->dims(); - auto out_grad_dims = - ctx.Input(framework::GradVarName("Out"))->dims(); + auto x_dims = ctx->GetInputDim("X"); + auto y_dims = ctx->GetInputDim("Y"); + auto xnorm_dims = ctx->GetInputDim("XNorm"); + auto ynorm_dims = ctx->GetInputDim("YNorm"); + auto out_dims = ctx->GetInputDim("Out"); + auto out_grad_dims = ctx->GetInputDim(framework::GradVarName("Out")); PADDLE_ENFORCE_GE(x_dims.size(), y_dims.size(), "Ranks of Input(X) and Input(Y) must be equal."); @@ -143,10 +139,14 @@ class CosSimOpGrad : public framework::OperatorWithKernel { "Shape of Input(Out@Grad) must be [X.Dim(0), 1]."); // resize tensor - auto *x_grad = ctx.Output(framework::GradVarName("X")); - auto *y_grad = ctx.Output(framework::GradVarName("Y")); - if (x_grad) x_grad->Resize(x_dims); - if (y_grad) y_grad->Resize(y_dims); + auto x_grad_name = framework::GradVarName("X"); + auto y_grad_name = framework::GradVarName("Y"); + if (ctx->HasOutput(x_grad_name)) { + ctx->SetOutputDim(x_grad_name, x_dims); + } + if (ctx->HasOutput(y_grad_name)) { + ctx->SetOutputDim(y_grad_name, y_dims); + } } }; diff --git a/paddle/operators/cos_sim_op.h b/paddle/operators/cos_sim_op.h index bcf6f758cae561a2e22f5be6c7a242647ef1c144..68c56f531f941e1b8f66ac7ba6bf318881642c4f 100644 --- a/paddle/operators/cos_sim_op.h +++ b/paddle/operators/cos_sim_op.h @@ -28,7 +28,7 @@ template ; template -class CosSimKernel : public framework::OpKernel { +class CosSimKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { // get Tensor @@ -67,7 +67,7 @@ class CosSimKernel : public framework::OpKernel { }; template -class CosSimGradKernel : public framework::OpKernel { +class CosSimGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { // get Tensor diff --git a/paddle/operators/crop_op.cc b/paddle/operators/crop_op.cc index 52a1123348b10e39bcfa1ba062c893e5f20ed862..a1424993cc3076f2755164393c5429d79fc1ee78 100644 --- a/paddle/operators/crop_op.cc +++ b/paddle/operators/crop_op.cc @@ -25,16 +25,14 @@ class CropOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), - "Input(X) of CropOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"), - "Output(Out) of CropOp should not be null."); - auto x_dim = ctx.Input("X")->dims(); - auto *y = ctx.Input("Y"); - auto *out = ctx.Output("Out"); - if (y == nullptr) { - auto shape = Attr>("shape"); + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of CropOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of CropOp should not be null."); + auto x_dim = ctx->GetInputDim("X"); + if (!ctx->HasInput("Y")) { + auto shape = ctx->Attrs().Get>("shape"); PADDLE_ENFORCE_EQ( int64_t(shape.size()), x_dim.size(), "Shape size should be equal to dimention size of input tensor."); @@ -42,19 +40,20 @@ class CropOp : public framework::OperatorWithKernel { for (size_t i = 0; i < shape.size(); ++i) { tensor_shape[i] = static_cast(shape[i]); } - out->Resize(framework::make_ddim(tensor_shape)); + ctx->SetOutputDim("Out", framework::make_ddim(tensor_shape)); } else { - PADDLE_ENFORCE_EQ(framework::arity(x_dim), framework::arity(y->dims()), + auto y_dim = ctx->GetInputDim("Y"); + PADDLE_ENFORCE_EQ(framework::arity(x_dim), framework::arity(y_dim), "Tensor rank of both CropOp's " "inputs must be same."); - out->Resize(y->dims()); + ctx->SetOutputDim("Out", y_dim); } } }; class CropOpMaker : public framework::OpProtoAndCheckerMaker { public: - CropOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + CropOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "The input of pad op. " @@ -78,12 +77,12 @@ class CropOpMaker : public framework::OpProtoAndCheckerMaker { Crop Operator. Crop input into output, as specified by offsets and shape. -There are two ways to set shape: +There are two ways to set shape: 1. referenc input: crop input X as shape as reference input. - The dimension of reference input should + The dimension of reference input should be as same as input X. 2. shape list: crop input X by shape described by a list. - The size of shape list should be as same as + The size of shape list should be as same as dimension size of input X. The input should be a k-D tensor(k > 0 and k < 7). As an example: @@ -94,15 +93,15 @@ Given: [0, 3, 4, 0, 0] [0, 0, 0, 0, 0]] -and +and offsets = [0, 1] and - + shape = [2, 2] -then we get +then we get Out = [[1, 2], [3, 4]] @@ -116,14 +115,14 @@ class CropOpGrad : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) should not be null"); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")), - "Input(Out@GRAD) should not be null"); - auto x_dims = ctx.Input("X")->dims(); - auto *x_grad = ctx.Output(framework::GradVarName("X")); - if (x_grad != nullptr) { - x_grad->Resize(x_dims); + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null"); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), + "Input(Out@GRAD) should not be null"); + auto x_dims = ctx->GetInputDim("X"); + auto x_grad_name = framework::GradVarName("X"); + if (ctx->HasOutput(x_grad_name)) { + ctx->SetOutputDim(x_grad_name, x_dims); } } }; diff --git a/paddle/operators/crop_op.h b/paddle/operators/crop_op.h index ac3aeaf41e206c1deb74c7022c36f02c4777a84b..2e72583d68d0acf0e2f5044637dba55de3b57209 100644 --- a/paddle/operators/crop_op.h +++ b/paddle/operators/crop_op.h @@ -27,7 +27,7 @@ using EigenTensor = framework::EigenTensor; using framework::Tensor; template -class CropKernel : public framework::OpKernel { +class CropKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto* x = context.Input("X"); @@ -69,7 +69,7 @@ void CropGradFunction(const framework::ExecutionContext& context) { } template -class CropGradKernel : public framework::OpKernel { +class CropGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { size_t rank = diff --git a/paddle/operators/cross_entropy_op.cc b/paddle/operators/cross_entropy_op.cc index b11dc1472d153dd188a0b3553d6950774216a3fd..708e80e96a0f007be1594d8b4781d1296d6b398d 100644 --- a/paddle/operators/cross_entropy_op.cc +++ b/paddle/operators/cross_entropy_op.cc @@ -22,32 +22,36 @@ class CrossEntropyOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) must not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Label"), - "Input(Label) must not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Y"), "Output(Y) must not be null."); - - auto x = ctx.Input("X"); - auto label = ctx.Input("Label"); - PADDLE_ENFORCE_EQ(x->dims().size(), 2, "Input(X)'s rank must be 2."); - PADDLE_ENFORCE_EQ(label->dims().size(), 2, - "Input(Label)'s rank must be 2."); - PADDLE_ENFORCE_EQ(x->dims()[0], label->dims()[0], - "The 1st dimension of Input(X) and Input(Label) must " + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should be not null."); + PADDLE_ENFORCE(ctx->HasInput("Label"), "Input(Label) should be not null."); + PADDLE_ENFORCE(ctx->HasOutput("Y"), "Output(Y) should be not null."); + + auto x_dims = ctx->GetInputDim("X"); + auto label_dims = ctx->GetInputDim("Label"); + PADDLE_ENFORCE_EQ(x_dims.size(), 2, "Input(X)'s rank should be 2."); + PADDLE_ENFORCE_EQ(label_dims.size(), 2, "Input(Label)'s rank should be 2."); + PADDLE_ENFORCE_EQ(x_dims[0], label_dims[0], + "The 1st dimension of Input(X) and Input(Label) should " "be equal."); - if (ctx.Attr("soft_label")) { - PADDLE_ENFORCE_EQ(x->dims()[1], label->dims()[1], - "If Attr(soft_label) == true, The 2nd dimension of " - "Input(X) and Input(Label) must be equal."); + if (ctx->Attrs().Get("softLabel")) { + PADDLE_ENFORCE_EQ(x_dims[1], label_dims[1], + "If Attr(softLabel) == true, the 2nd dimension of " + "Input(X) and Input(Label) should be equal."); } else { - PADDLE_ENFORCE_EQ(label->dims()[1], 1, - "If Attr(soft_label) == false, The 2nd dimension of " - "Input(Label) must be 1."); + PADDLE_ENFORCE_EQ(label_dims[1], 1, + "If Attr(softLabel) == false, the 2nd dimension of " + "Input(Label) should be 1."); } - ctx.Output("Y")->Resize({x->dims()[0], 1}); - ctx.ShareLoD("X", /*->*/ "Y"); + ctx->SetOutputDim("Y", {x_dims[0], 1}); + ctx->ShareLoD("X", /*->*/ "Y"); + } + + // CrossEntropy's data type just determined by "X" + framework::DataType IndicateDataType( + const framework::ExecutionContext& ctx) const override { + return framework::ToDataType(ctx.Input("X")->type()); } }; @@ -56,66 +60,85 @@ class CrossEntropyGradientOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) must not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Label"), - "Input(Label) must not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Y")), - "Input(Y@GRAD) must not be null."); - - auto x = ctx.Input("X"); - auto label = ctx.Input("Label"); - auto dy = ctx.Input(framework::GradVarName("Y")); - PADDLE_ENFORCE_EQ(x->dims().size(), 2, "Input(X)'s rank must be 2."); - PADDLE_ENFORCE_EQ(dy->dims().size(), 2, "Input(Y@Grad)'s rank must be 2."); - PADDLE_ENFORCE_EQ(label->dims().size(), 2, - "Input(Label)'s rank must be 2."); - PADDLE_ENFORCE_EQ(x->dims()[0], label->dims()[0], - "The 1st dimension of Input(X) and Input(Label) must " + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should be not null."); + PADDLE_ENFORCE(ctx->HasInput("Label"), "Input(Label) should be not null."); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Y")), + "Input(Y@GRAD) shoudl be not null."); + PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("X")), + "Output(X@GRAD) should be not null."); + + auto x_dims = ctx->GetInputDim("X"); + auto label_dims = ctx->GetInputDim("Label"); + auto dy_dims = ctx->GetInputDim(framework::GradVarName("Y")); + PADDLE_ENFORCE_EQ(x_dims.size(), 2, "Input(X)'s rank should be 2."); + PADDLE_ENFORCE_EQ(dy_dims.size(), 2, "Input(Y@Grad)'s rank should be 2."); + PADDLE_ENFORCE_EQ(label_dims.size(), 2, "Input(Label)'s rank should be 2."); + PADDLE_ENFORCE_EQ(x_dims[0], label_dims[0], + "The 1st dimension of Input(X) and Input(Label) should " "be equal."); - PADDLE_ENFORCE_EQ(x->dims()[0], dy->dims()[0], - "The 1st dimension of Input(X) and Input(Y@Grad) must " + PADDLE_ENFORCE_EQ(x_dims[0], dy_dims[0], + "The 1st dimension of Input(X) and Input(Y@Grad) should " "be equal."); - PADDLE_ENFORCE_EQ(dy->dims()[1], 1, - "The 2nd dimension of Input(Y@Grad) must be 1."); - if (ctx.Attr("soft_label")) { - PADDLE_ENFORCE_EQ(x->dims()[1], label->dims()[1], - "If Attr(soft_label) == true, The 2nd dimension of " - "Input(X) and Input(Label) must be equal."); + PADDLE_ENFORCE_EQ(dy_dims[1], 1, + "The 2nd dimension of Input(Y@Grad) should be 1."); + if (ctx->Attrs().Get("softLabel")) { + PADDLE_ENFORCE_EQ(x_dims[1], label_dims[1], + "When Attr(softLabel) == true, the 2nd dimension of " + "Input(X) and Input(Label) should be equal."); } else { - PADDLE_ENFORCE_EQ(label->dims()[1], 1, - "If Attr(soft_label) == false, The 2nd dimension of " - "Input(Label) must be 1."); + PADDLE_ENFORCE_EQ(label_dims[1], 1, + "When Attr(softLabel) == false, the 2nd dimension of " + "Input(Label) should be 1."); } + ctx->SetOutputDim(framework::GradVarName("X"), x_dims); + } - auto dx = ctx.Output(framework::GradVarName("X")); - dx->Resize(x->dims()); + // CrossEntropy's data type just determined by "X" + framework::DataType IndicateDataType( + const framework::ExecutionContext& ctx) const override { + return framework::ToDataType(ctx.Input("X")->type()); } }; class CrossEntropyOpMaker : public framework::OpProtoAndCheckerMaker { public: - CrossEntropyOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + CrossEntropyOpMaker(framework::OpProto* proto, + framework::OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("X", "The first input of CrossEntropyOp"); - AddInput("Label", "The second input of CrossEntropyOp"); - AddOutput("Y", "The output of CrossEntropyOp"); - AddAttr("soft_label", "Is soft label. Default zero.") + AddInput("X", + "(Tensor, default Tensor), a 2-D tensor with shape N x D, " + "where N is the batch size and D is the number of classes. " + "This input is a probability computed by the previous operator, " + "which is almost always the result of a softmax operator."); + AddInput( + "Label", + "(Tensor, default Tensor), the ground truth which is " + "a 2-D tensor. " + "When softLabel is set to false, `Label` is a Tensor with shape " + "[N x 1]. " + "When softLabel is set to true, `Label` is a Tensor " + "with shape [N x K]."); + AddOutput("Y", + "(Tensor, default Tensor), a 2-D tensor " + "with shape [N x 1]. The cross entropy loss."); + AddAttr( + "softLabel", + "(bool, default false), a flag to indicate whether to interpretate " + "the given labels as soft labels.") .SetDefault(false); - AddComment(R"DOC( CrossEntropy Operator. It supports both standard cross-entropy and soft-label cross-entropy loss computation. 1) One-hot cross-entropy: - soft_label = False, Label[i, 0] indicates the class index for sample i: + softLabel = false, Label[i, 0] indicates the class index for sample i: Y[i] = -log(X[i, Label[i]]) 2) Soft-label cross-entropy: - soft_label = True, Label[i, j] indicates the soft label of class j + softLabel = true, Label[i, j] indicates the soft label of class j for sample i: Y[i] = \sum_j{-Label[i, j] * log(X[i, j])} diff --git a/paddle/operators/cross_entropy_op.cu b/paddle/operators/cross_entropy_op.cu index 1d6361a81472a49729958120c52060b1dff803f2..5e2024e0ea9040b758e1cec4dbaa4b329bbb727e 100644 --- a/paddle/operators/cross_entropy_op.cu +++ b/paddle/operators/cross_entropy_op.cu @@ -12,47 +12,12 @@ See the License for the specific language governing permissions and limitations under the License. */ -#include "paddle/framework/op_registry.h" #include "paddle/operators/cross_entropy_op.h" -#include "paddle/platform/assert.h" -#include "paddle/platform/hostdevice.h" namespace paddle { namespace operators { -template -__global__ void CrossEntropyKernel(T* Y, const T* X, const int* label, - const int N, const int D) { - // TOOD(qingqing) define CUDA_1D_KERNEL_LOOP macro in a common file. - // CUDA_1D_KERNEL_LOOP(i, N) { - for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < N; - i += blockDim.x * gridDim.x) { - PADDLE_ASSERT(label[i] >= 0 && label[i] < D); - Y[i] = -tolerable_value(log(X[i * D + label[i]])); - } -} - -template -__global__ void SoftCrossEntropyKernel(T* Y, const T* X, const T* label, - const int N, const int D) { - for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < N; - i += blockDim.x * gridDim.x) { - T sum = static_cast(0); - for (int j = 0; j < D; j++) { - sum += label[i * D + j] * tolerable_value(log(X[i * D + j])); - } - Y[i] = -sum; - } -} - -// TODO(qingqing): make zero setting an common function. -template -__global__ void zero(T* X, const int N) { - for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < N; - i += blockDim.x * gridDim.x) { - X[i] = 0.0; - } -} +namespace { template __global__ void CrossEntropyGradientKernel(T* dX, const T* dY, const T* X, @@ -71,80 +36,69 @@ template __global__ void SoftCrossEntropyGradientKernel(T* dX, const T* dY, const T* X, const T* label, const int N, const int D) { - // TOOD(qingqing): optimize for this kernel - for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < N; - i += blockDim.x * gridDim.x) { - for (int j = 0; j < D; ++j) { - int idx = i * D + j; - dX[idx] = -label[idx] * dY[i] / X[idx]; - } + int ids = blockIdx.x * blockDim.x + threadIdx.x; + if (ids < N * D) { + int row_ids = ids / D; + dX[ids] = -label[ids] * dY[row_ids] / X[ids]; } } +} // namespace template -class CrossEntropyOpCUDAKernel : public framework::OpKernel { +class CrossEntropyOpCUDAKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), - "It must use GPUPlace."); - - auto x = ctx.Input("X"); - auto y = ctx.Output("Y"); - auto label = ctx.Input("Label"); - - auto* x_data = x->data(); + "This kernel only runs on GPU device."); + const Tensor* x = ctx.Input("X"); + const Tensor* label = ctx.Input("Label"); + Tensor* y = ctx.Output("Y"); y->mutable_data(ctx.GetPlace()); - auto* y_data = y->data(); - int n = x->dims()[0]; - int d = x->dims()[1]; - int block = 512; - int grid = (n + block - 1) / block; - // TODO(qingqing) launch kernel on specified stream - // base on ExecutionContext. - if (ctx.Attr("soft_label")) { - auto* label_data = ctx.Input("Label")->data(); - SoftCrossEntropyKernel<<>>(y_data, x_data, label_data, n, - d); - } else { - auto* label_data = ctx.Input("Label")->data(); - CrossEntropyKernel<<>>(y_data, x_data, label_data, n, d); - } + math::CrossEntropyFunctor()( + ctx.device_context(), y, x, label, ctx.Attr("softLabel")); } }; template -class CrossEntropyGradientOpCUDAKernel : public framework::OpKernel { +class CrossEntropyGradientOpCUDAKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), - "It must use GPUPlace."); + "This kernel only runs on GPU device."); + + const Tensor* x = ctx.Input("X"); + const Tensor* label = ctx.Input("Label"); + Tensor* dx = ctx.Output(framework::GradVarName("X")); + dx->mutable_data(ctx.GetPlace()); - auto x = ctx.Input("X"); - auto dx = ctx.Output(framework::GradVarName("X")); - auto dy = ctx.Input(framework::GradVarName("Y")); - auto label = ctx.Input("Label"); + const T* dy_data = + ctx.Input(framework::GradVarName("Y"))->data(); + T* dx_data = dx->mutable_data(ctx.GetPlace()); + const T* x_data = x->data(); - auto* dx_data = dx->mutable_data(ctx.GetPlace()); - auto* dy_data = dy->data(); - auto* x_data = x->data(); + int batch_size = x->dims()[0]; + int class_num = x->dims()[1]; - int n = x->dims()[0]; - int d = x->dims()[1]; int block = 512; - int grid = (n * d + block - 1) / block; - zero<<>>(dx_data, n * d); - grid = (n + block - 1) / block; - // TODO(qingqing): launch kernel on specified stream - // base on ExecutionContext. - if (ctx.Attr("soft_label")) { + int grid = (batch_size * class_num + block - 1) / block; + + if (ctx.Attr("softLabel")) { auto* label_data = label->data(); - SoftCrossEntropyGradientKernel<<>>( - dx_data, dy_data, x_data, label_data, n, d); + SoftCrossEntropyGradientKernel<<< + grid, block, 0, reinterpret_cast( + ctx.device_context()) + .stream()>>>(dx_data, dy_data, x_data, label_data, + batch_size, class_num); } else { + math::SetConstant(ctx.device_context(), dx, 0); auto* label_data = label->data(); - CrossEntropyGradientKernel<<>>(dx_data, dy_data, x_data, - label_data, n, d); + grid = (batch_size + block - 1) / block; + CrossEntropyGradientKernel<<< + grid, block, 0, reinterpret_cast( + ctx.device_context()) + .stream()>>>(dx_data, dy_data, x_data, label_data, + batch_size, class_num); } } }; diff --git a/paddle/operators/cross_entropy_op.h b/paddle/operators/cross_entropy_op.h index 69caba5ff31f60df2c24cef0e6331f058f6ba8d6..d2d321aa7ed8e32cc19d5a171beea34d36195b10 100644 --- a/paddle/operators/cross_entropy_op.h +++ b/paddle/operators/cross_entropy_op.h @@ -13,97 +13,65 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once +#include "paddle/framework/eigen.h" #include "paddle/framework/op_registry.h" -#include "paddle/platform/hostdevice.h" +#include "paddle/operators/math/cross_entropy.h" +#include "paddle/operators/math/math_function.h" namespace paddle { namespace operators { using Tensor = framework::Tensor; +template +using EigenMatrix = framework::EigenMatrix; template -HOSTDEVICE T tolerable_value(const T x) { - PADDLE_ASSERT(std::is_floating_point::value); - const T kApproInf = 1e20; - if (x == INFINITY) { - return kApproInf; - } - if (x == -INFINITY) { - return -kApproInf; - } - return x; -} - -template -class CrossEntropyOpKernel : public framework::OpKernel { +class CrossEntropyOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()), - "It must use CPUPlace."); - - auto x = ctx.Input("X"); - auto y = ctx.Output("Y"); - - auto* x_data = x->data(); + "This kernel only runs on CPU."); + const Tensor* x = ctx.Input("X"); + const Tensor* labels = ctx.Input("Label"); + Tensor* y = ctx.Output("Y"); y->mutable_data(ctx.GetPlace()); - auto* y_data = y->data(); - int batch_size = x->dims()[0]; - int class_num = x->dims()[1]; - - if (ctx.Attr("soft_label")) { - auto* label_data = ctx.Input("Label")->data(); - int index = 0; - for (int i = 0; i < batch_size; ++i) { - T sum = static_cast(0); - for (int j = 0; j < class_num; ++j) { - sum += label_data[index] * tolerable_value(std::log(x_data[index])); - y_data[i] = -sum; - index++; - } - } - } else { - auto* label_data = ctx.Input("Label")->data(); - for (int i = 0; i < batch_size; ++i) { - int index = i * class_num + label_data[i]; - y_data[i] = -tolerable_value(std::log(x_data[index])); - } - } + math::CrossEntropyFunctor()( + ctx.device_context(), y, x, labels, ctx.Attr("softLabel")); } }; template -class CrossEntropyGradientOpKernel : public framework::OpKernel { +class CrossEntropyGradientOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()), - "It must use CPUPlace."); - - auto x = ctx.Input("X"); - auto dx = ctx.Output(framework::GradVarName("X")); - auto dy = ctx.Input(framework::GradVarName("Y")); - auto label = ctx.Input("Label"); - - auto* dx_data = dx->mutable_data(ctx.GetPlace()); - auto* dy_data = dy->data(); - auto* x_data = x->data(); + "This kernel only runs on CPU."); + const Tensor* x = ctx.Input("X"); + const Tensor* dy = ctx.Input(framework::GradVarName("Y")); + const Tensor* label = ctx.Input("Label"); + Tensor* dx = ctx.Output(framework::GradVarName("X")); + T* dx_data = dx->mutable_data(ctx.GetPlace()); - int batch_size = x->dims()[0]; int class_num = x->dims()[1]; - - // TODO(qingqing): make zero setting an common function. - if (ctx.Attr("soft_label")) { - auto* label_data = ctx.Input("Label")->data(); - int index = 0; - for (int i = 0; i < batch_size; ++i) { - for (int j = 0; j < class_num; ++j) { - dx_data[index] = -label_data[index] * dy_data[i] / x_data[index]; - index++; - } - } + if (ctx.Attr("softLabel")) { + auto x_mat = EigenMatrix::From(*x); + auto dy_mat = EigenMatrix::From(*dy); + auto lbl_mat = EigenMatrix::From(*label); + auto dx_mat = EigenMatrix::From(*dx); + + dx_mat.device(ctx.GetEigenDevice()) = + -(lbl_mat * dy_mat.broadcast(Eigen::DSizes(1, class_num)) / + x_mat); } else { - auto* label_data = label->data(); - memset(dx_data, 0, sizeof(T) * batch_size * class_num); + int batch_size = x->dims()[0]; + const T* dy_data = dy->data(); + const T* x_data = x->data(); + const int* label_data = label->data(); + + math::SetConstant(ctx.device_context(), dx, 0); + for (int i = 0; i < batch_size; ++i) { PADDLE_ASSERT(label_data[i] >= 0 || label_data[i] < class_num); int index = i * class_num + label_data[i]; diff --git a/paddle/operators/detail/strided_memcpy.h b/paddle/operators/detail/strided_memcpy.h index b165224b37fb091c094a823179256c3dd40a37c9..068c82f399316a1587d7322d8dab75823656800e 100644 --- a/paddle/operators/detail/strided_memcpy.h +++ b/paddle/operators/detail/strided_memcpy.h @@ -34,7 +34,7 @@ struct StridedMemcpyFunctor { auto& cpu_place = boost::get(place); memory::Copy(cpu_place, dst, cpu_place, src, sizeof(T) * dst_dim.head); } else { -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA auto& gpu_place = boost::get(place); auto& cuda_ctx = reinterpret_cast(dev_ctx); diff --git a/paddle/operators/dropout_op.cc b/paddle/operators/dropout_op.cc index 2130eda6a42c893d8ec251a7022a0bfa44433bb7..708ccfa0bfc163ea75fcd9dbcb5ef6140b4b5f36 100644 --- a/paddle/operators/dropout_op.cc +++ b/paddle/operators/dropout_op.cc @@ -24,25 +24,25 @@ class DropoutOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) must not be null."); - PADDLE_ENFORCE_GE(ctx.Attr("dropout_prob"), 0); - PADDLE_ENFORCE_LE(ctx.Attr("dropout_prob"), 1); - - auto dims = ctx.Input("X")->dims(); - ctx.Output("Out")->Resize(dims); - if (ctx.Attr("is_training")) { - ctx.Output("Mask")->Resize(dims); + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) must not be null."); + PADDLE_ENFORCE_GE(ctx->Attrs().Get("dropout_prob"), 0); + PADDLE_ENFORCE_LE(ctx->Attrs().Get("dropout_prob"), 1); + + auto x_dims = ctx->GetInputDim("X"); + ctx->SetOutputDim("Out", x_dims); + if (ctx->Attrs().Get("is_training") == 1) { + ctx->SetOutputDim("Mask", x_dims); } - ctx.ShareLoD("X", /*->*/ "Out"); + ctx->ShareLoD("X", /*->*/ "Out"); } }; template class DropoutOpMaker : public framework::OpProtoAndCheckerMaker { public: - DropoutOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + DropoutOpMaker(framework::OpProto* proto, + framework::OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddAttr("dropout_prob", "Probability of setting units to zero.") .SetDefault(.5f); @@ -70,27 +70,26 @@ class DropoutOpGrad : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE(ctx.Attr("is_training"), - "GradOp is only callable when is_training is true"); - - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) must not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Mask"), "Mask must not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")), - "Input(Out@GRAD) must not be null."); - - PADDLE_ENFORCE_GE(ctx.Attr("dropout_prob"), 0); - PADDLE_ENFORCE_LE(ctx.Attr("dropout_prob"), 1); - auto x_dims = ctx.Input("X")->dims(); - auto out_dims = ctx.Input(framework::GradVarName("Out"))->dims(); + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE_EQ(ctx->Attrs().Get("is_training"), 1, + "GradOp is only callable when is_training is true"); + + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) must not be null."); + PADDLE_ENFORCE(ctx->HasInput("Mask"), "Mask must not be null."); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), + "Input(Out@GRAD) must not be null."); + + PADDLE_ENFORCE_GE(ctx->Attrs().Get("dropout_prob"), 0); + PADDLE_ENFORCE_LE(ctx->Attrs().Get("dropout_prob"), 1); + auto x_dims = ctx->GetInputDim("X"); + auto out_dims = ctx->GetInputDim(framework::GradVarName("Out")); PADDLE_ENFORCE_EQ(x_dims, out_dims, "Dimensions of Input(X) and Out@Grad must be the same."); - auto mask_dims = ctx.Input("Mask")->dims(); + auto mask_dims = ctx->GetInputDim("Mask"); PADDLE_ENFORCE_EQ(x_dims, mask_dims, "Dimensions of Input(X) and Mask must be the same."); - auto *x_grad = ctx.Output(framework::GradVarName("X")); - x_grad->Resize(x_dims); + ctx->SetOutputDim(framework::GradVarName("X"), x_dims); } }; diff --git a/paddle/operators/dropout_op.cu b/paddle/operators/dropout_op.cu index a04e4a22cc09d4e8106a528e490ccf8e90681c08..30c769000f2b98c69eaa78a4c139630dd0956386 100644 --- a/paddle/operators/dropout_op.cu +++ b/paddle/operators/dropout_op.cu @@ -47,7 +47,7 @@ struct MaskGenerator { // Use std::random and thrust::random(thrust is a std library in CUDA) to // implement uniform random. template -class GPUDropoutKernel : public framework::OpKernel { +class GPUDropoutKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto* x = context.Input("X"); diff --git a/paddle/operators/dropout_op.h b/paddle/operators/dropout_op.h index d57f64afcb3558aeea6aed23fae06866e9af874a..745525fe81dadb22cbb64d66203f5a75608d3718 100644 --- a/paddle/operators/dropout_op.h +++ b/paddle/operators/dropout_op.h @@ -26,7 +26,7 @@ template ; template -class CPUDropoutKernel : public framework::OpKernel { +class CPUDropoutKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto* x = context.Input("X"); @@ -62,7 +62,7 @@ class CPUDropoutKernel : public framework::OpKernel { }; template -class DropoutGradKernel : public framework::OpKernel { +class DropoutGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { PADDLE_ENFORCE(context.Attr("is_training"), diff --git a/paddle/operators/elementwise_add_op.cc b/paddle/operators/elementwise_add_op.cc index 5f7b654d69f081dfa85b0d61960eb52b7982faa1..d9bc80c869c023caebf0b45ed24f2def3f0b1dd8 100644 --- a/paddle/operators/elementwise_add_op.cc +++ b/paddle/operators/elementwise_add_op.cc @@ -13,6 +13,7 @@ limitations under the License. */ #include "paddle/operators/elementwise_add_op.h" +#include "paddle/operators/elementwise_op.h" namespace paddle { namespace operators { diff --git a/paddle/operators/elementwise_add_op.h b/paddle/operators/elementwise_add_op.h index 9e9f1ffba6fb23f5394713c67aa4363b85717f50..f04fe3ec6069ab1bf227be6a3a5c10ee908e4824 100644 --- a/paddle/operators/elementwise_add_op.h +++ b/paddle/operators/elementwise_add_op.h @@ -14,13 +14,13 @@ #pragma once -#include "paddle/operators/elementwise_op.h" +#include "paddle/operators/elementwise_op_function.h" namespace paddle { namespace operators { template -class ElementwiseAddKernel : public framework::OpKernel { +class ElementwiseAddKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { ElementwiseCompute(ctx); @@ -101,7 +101,7 @@ struct ElementwiseAddBroadCast2GradFunctor { }; template -class ElementwiseAddGradKernel : public framework::OpKernel { +class ElementwiseAddGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { ElementwiseGradCompute, diff --git a/paddle/operators/elementwise_div_op.cc b/paddle/operators/elementwise_div_op.cc index c6898150d310d0c4fdefae5a58a5792a72f9889e..3f56344d0007b5f14fd9b5b9b44a9b29d3c42f2a 100644 --- a/paddle/operators/elementwise_div_op.cc +++ b/paddle/operators/elementwise_div_op.cc @@ -13,6 +13,7 @@ limitations under the License. */ #include "paddle/operators/elementwise_div_op.h" +#include "paddle/operators/elementwise_op.h" namespace paddle { namespace operators { diff --git a/paddle/operators/elementwise_div_op.h b/paddle/operators/elementwise_div_op.h index 9bd7c8ea548c46ec9b4c5a085e4e70d5dd162f3a..8946ff3d25c2aff3dc3aa69368f0083371cd2fef 100644 --- a/paddle/operators/elementwise_div_op.h +++ b/paddle/operators/elementwise_div_op.h @@ -14,13 +14,13 @@ #pragma once -#include "paddle/operators/elementwise_op.h" +#include "paddle/operators/elementwise_op_function.h" namespace paddle { namespace operators { template -class ElementwiseDivKernel : public framework::OpKernel { +class ElementwiseDivKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { ElementwiseCompute(ctx); @@ -103,7 +103,7 @@ struct ElementwiseDivBroadCast2GradFunctor { }; template -class ElementwiseDivGradKernel : public framework::OpKernel { +class ElementwiseDivGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { ElementwiseGradCompute, diff --git a/paddle/operators/elementwise_mul_op.cc b/paddle/operators/elementwise_mul_op.cc index f2544b54d6bc543a50d8de03d482333b485bc076..da7765aa6a7a81c9e0b4f462022cad54c16aec47 100644 --- a/paddle/operators/elementwise_mul_op.cc +++ b/paddle/operators/elementwise_mul_op.cc @@ -13,6 +13,7 @@ limitations under the License. */ #include "paddle/operators/elementwise_mul_op.h" +#include "paddle/operators/elementwise_op.h" namespace paddle { namespace operators { @@ -35,7 +36,9 @@ REGISTER_OP(elementwise_mul, ops::ElementwiseOp, ops::ElementwiseMulOpMaker, elementwise_mul_grad, ops::ElementwiseOpGrad); REGISTER_OP_CPU_KERNEL( elementwise_mul, - ops::ElementwiseMulKernel); + ops::ElementwiseMulKernel, + ops::ElementwiseMulKernel); REGISTER_OP_CPU_KERNEL( elementwise_mul_grad, - ops::ElementwiseMulGradKernel); + ops::ElementwiseMulGradKernel, + ops::ElementwiseMulGradKernel); diff --git a/paddle/operators/elementwise_mul_op.cu b/paddle/operators/elementwise_mul_op.cu index da08a75596c4d3b89dc8892bd4405464fec96389..056f081d3e6ac349978ff00689700c035bed8e39 100644 --- a/paddle/operators/elementwise_mul_op.cu +++ b/paddle/operators/elementwise_mul_op.cu @@ -19,7 +19,9 @@ namespace ops = paddle::operators; REGISTER_OP_GPU_KERNEL( elementwise_mul, - ops::ElementwiseMulKernel); + ops::ElementwiseMulKernel, + ops::ElementwiseMulKernel); REGISTER_OP_GPU_KERNEL( elementwise_mul_grad, - ops::ElementwiseMulGradKernel); + ops::ElementwiseMulGradKernel, + ops::ElementwiseMulGradKernel); diff --git a/paddle/operators/elementwise_mul_op.h b/paddle/operators/elementwise_mul_op.h index 1eaf2e3efc97a32739efcaf37066817ee173fadc..4469b07eaa08a3b011a88e58f1d645dd30b10ced 100644 --- a/paddle/operators/elementwise_mul_op.h +++ b/paddle/operators/elementwise_mul_op.h @@ -13,13 +13,13 @@ limitations under the License. */ #pragma once -#include "paddle/operators/elementwise_op.h" +#include "paddle/operators/elementwise_op_function.h" namespace paddle { namespace operators { template -class ElementwiseMulKernel : public framework::OpKernel { +class ElementwiseMulKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { ElementwiseCompute(ctx); @@ -102,7 +102,7 @@ struct ElementwiseMulBroadCast2GradFunctor { }; template -class ElementwiseMulGradKernel : public framework::OpKernel { +class ElementwiseMulGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { ElementwiseGradCompute, diff --git a/paddle/operators/elementwise_op.h b/paddle/operators/elementwise_op.h index f224722c1bec6716e68de9da2509250f7d4b37ae..66f1910a4738f737c7581e17729c201a7e3eaeae 100644 --- a/paddle/operators/elementwise_op.h +++ b/paddle/operators/elementwise_op.h @@ -1,222 +1,44 @@ /* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - Licensed under the Apache License, Version 2.0 (the "License"); - you may not use this file except in compliance with the License. - You may obtain a copy of the License at +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at - http://www.apache.org/licenses/LICENSE-2.0 + http://www.apache.org/licenses/LICENSE-2.0 - Unless required by applicable law or agreed to in writing, software - distributed under the License is distributed on an "AS IS" BASIS, - WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - See the License for the specific language governing permissions and - limitations under the License. */ +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ #pragma once -#include -#include "paddle/framework/eigen.h" #include "paddle/framework/op_registry.h" -#include "paddle/operators/math/math_function.h" +#include "paddle/framework/operator.h" namespace paddle { namespace operators { -/* - * Out = X ⊙ Y - * If Y's shape does not match X' shape, they will be reshaped. - * For example: - * 1. shape(X) = (2, 3, 4, 5), shape(Y) = (3, 4), with axis=1 - * pre=2, n=3*4, post=5 - * x.shape(2, 12, 5) * y.shape(1,12,1).broadcast(2,12,5) - * 2. shape(X) = (2, 3, 4, 5), shape(Y) = (4,5) - * pre=2*3, n=4*5, post=1 - * x.shape(2, 3, 20) * y.shape(1,1,20).broadcast(2,3,20) - */ -inline void get_mid_dims(const framework::DDim& x_dims, - const framework::DDim& y_dims, const int axis, - int& pre, int& n, int& post) { - pre = 1; - n = 1; - post = 1; - for (int i = 0; i < axis; ++i) { - pre *= x_dims[i]; - } - - for (int i = 0; i < y_dims.size(); ++i) { - PADDLE_ENFORCE_EQ(x_dims[i + axis], y_dims[i], - "Broadcast dimension mismatch."); - n *= y_dims[i]; - } - - for (int i = axis + y_dims.size(); i < x_dims.size(); ++i) { - post *= x_dims[i]; - } -} - -#define EIGEN_FUNCTOR(name, eigen_op) \ - struct Eigen##name##Functor { \ - template \ - inline void Run(const framework::Tensor* x, const framework::Tensor* y, \ - framework::Tensor* z, \ - const framework::ExecutionContext& ctx) { \ - auto x_e = framework::EigenVector::Flatten(*x); \ - auto y_e = framework::EigenVector::Flatten(*y); \ - auto z_e = framework::EigenVector::Flatten(*z); \ - z_e.device(ctx.GetEigenDevice()) = eigen_op(x_e, y_e); \ - } \ - template \ - inline void RunBroadCast(const framework::Tensor* x, \ - const framework::Tensor* y, framework::Tensor* z, \ - const framework::ExecutionContext& ctx, int pre, \ - int n) { \ - auto x_e = framework::EigenVector::Flatten(*x); \ - auto y_e = framework::EigenVector::Flatten(*y); \ - auto z_e = framework::EigenVector::Flatten(*z); \ - auto y_bcast = y_e.reshape(Eigen::DSizes(1, n)) \ - .broadcast(Eigen::DSizes(pre, 1)) \ - .reshape(Eigen::DSizes(x_e.size())); \ - z_e.device(ctx.GetEigenDevice()) = eigen_op(x_e, y_bcast); \ - } \ - template \ - inline void RunBroadCast2(const framework::Tensor* x, \ - const framework::Tensor* y, \ - framework::Tensor* z, \ - const framework::ExecutionContext& ctx, int pre, \ - int n, int post) { \ - auto x_e = framework::EigenVector::Flatten(*x); \ - auto y_e = framework::EigenVector::Flatten(*y); \ - auto z_e = framework::EigenVector::Flatten(*z); \ - auto y_bcast = y_e.reshape(Eigen::DSizes(1, n, 1)) \ - .broadcast(Eigen::DSizes(pre, 1, post)) \ - .reshape(Eigen::DSizes(x_e.size())); \ - z_e.device(ctx.GetEigenDevice()) = eigen_op(x_e, y_bcast); \ - } \ - } - -template -void ElementwiseCompute(const framework::ExecutionContext& ctx) { - using Tensor = framework::Tensor; - - auto* x = ctx.Input("X"); - auto* y = ctx.Input("Y"); - auto* z = ctx.Output("Out"); - z->mutable_data(ctx.GetPlace()); - - auto x_dims = x->dims(); - auto y_dims = y->dims(); - PADDLE_ENFORCE_GE(x_dims.size(), y_dims.size(), - "Rank of first input must >= rank of second input.") - - if (x_dims == y_dims || product(y_dims) == 1) { - functor f; - f.template Run(x, y, z, ctx); - return; - } - - int axis = ctx.Attr("axis"); - axis = (axis == -1 ? x_dims.size() - y_dims.size() : axis); - PADDLE_ENFORCE(axis >= 0 && axis < x_dims.size(), - "Axis should be in range [0, x_dims)"); - - int pre, n, post; - get_mid_dims(x_dims, y_dims, axis, pre, n, post); - if (post == 1) { - functor f; - f.template RunBroadCast(x, y, z, ctx, pre, n); - return; - } else { - functor f; - f.template RunBroadCast2(x, y, z, ctx, pre, n, post); - return; - } -} - -#define EIGEN_ADD(x, y) ((x) + (y)) -EIGEN_FUNCTOR(Add, EIGEN_ADD); - -#define EIGEN_SUB(x, y) ((x) - (y)) -EIGEN_FUNCTOR(Sub, EIGEN_SUB); - -#define EIGEN_MUL(x, y) ((x) * (y)) -EIGEN_FUNCTOR(Mul, EIGEN_MUL); - -#define EIGEN_DIV(x, y) ((x) / (y)) -EIGEN_FUNCTOR(Div, EIGEN_DIV); - -template -void ElementwiseGradCompute(const framework::ExecutionContext& ctx) { - using Tensor = framework::Tensor; - - auto* x = ctx.Input("X"); - auto* y = ctx.Input("Y"); - auto* out = ctx.Input("Out"); - auto* dout = ctx.Input(framework::GradVarName("Out")); - - auto place = ctx.GetEigenDevice(); - - auto x_dims = x->dims(); - auto y_dims = y->dims(); - - auto* dx = ctx.Output(framework::GradVarName("X")); - auto* dy = ctx.Output(framework::GradVarName("Y")); - if (dx) { - dx->mutable_data(ctx.GetPlace()); - } - if (dy) { - dy->mutable_data(ctx.GetPlace()); - } - - if (x_dims == y_dims) { - functor f; - f(place, x, y, out, dx, dy, dout); - return; - } - - if (product(y_dims) == 1) { - functor1 f; - f(place, x, y, out, dx, dy, dout); - return; - } - - int axis = ctx.Attr("axis"); - axis = (axis == -1 ? x_dims.size() - y_dims.size() : axis); - - int pre, n, post; - get_mid_dims(x_dims, y_dims, axis, pre, n, post); - - if (post == 1) { - broadcastfunctor f; - f(place, x, y, out, dx, dy, dout, pre, n); - return; - } else { - broadcast2functor f; - f(place, x, y, out, dx, dy, dout, pre, n, post); - return; - } -} - class ElementwiseOp : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; protected: using Tensor = framework::Tensor; - void InferShape(const framework::InferShapeContext& ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), - "Input(X) of elementwise op should not be null"); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), - "Input(Y) of elementwise op should not be null"); - PADDLE_ENFORCE_NOT_NULL( - ctx.OutputVar("Out"), - "Output(Out) of elementwise op should not be null."); - - auto x_dim = ctx.Input("X")->dims(); - auto y_dim = ctx.Input("Y")->dims(); + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of elementwise op should not be null"); + PADDLE_ENFORCE(ctx->HasInput("Y"), + "Input(Y) of elementwise op should not be null"); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of elementwise op should not be null."); + + auto x_dim = ctx->GetInputDim("X"); + auto y_dim = ctx->GetInputDim("Y"); PADDLE_ENFORCE_GE(x_dim.size(), y_dim.size(), "Rank of first input must >= rank of second input.") - ctx.Output("Out")->Resize(x_dim); - ctx.ShareLoD("X", /*->*/ "Out"); + ctx->SetOutputDim("Out", x_dim); + ctx->ShareLoD("X", /*->*/ "Out"); } }; @@ -234,7 +56,7 @@ must be small or equal to X's dimensions. )DOC"); AddAttr("axis", R"DOC( -When the shape(Y) does not equal the shape(X),Y will be broadcasted +When the shape(Y) does not equal the shape(X),Y will be broadcasted to match the shape of X and axis should be dimension index Y in X )DOC") .SetDefault(-1) @@ -244,7 +66,7 @@ to match the shape of X and axis should be dimension index Y in X comment_ = R"DOC( Limited elementwise {name} operator.The equation is: Out = {equation}. 1. The shape of Y should be same with X or -2. Y's shape is a subset of X. +2. Y's shape is a subset of X. Y will be broadcasted to match the shape of X and axis should be dimension index Y in X. example: @@ -284,27 +106,26 @@ class ElementwiseOpGrad : public framework::OperatorWithKernel { using Tensor = framework::Tensor; protected: - void InferShape(const framework::InferShapeContext& ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) should not be null"); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), "Input(Y) should not be null"); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")), - "Input(Out@GRAD) should not be null"); + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null"); + PADDLE_ENFORCE(ctx->HasInput("Y"), "Input(Y) should not be null"); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), + "Input(Out@GRAD) should not be null"); - auto x_dims = ctx.Input("X")->dims(); - auto y_dims = ctx.Input("Y")->dims(); - auto out_dims = ctx.Input(framework::GradVarName("Out"))->dims(); - auto* x_grad = ctx.Output(framework::GradVarName("X")); - auto* y_grad = ctx.Output(framework::GradVarName("Y")); + auto x_dims = ctx->GetInputDim("X"); + auto y_dims = ctx->GetInputDim("Y"); + auto out_dims = ctx->GetInputDim(framework::GradVarName("Out")); PADDLE_ENFORCE_GE(x_dims.size(), y_dims.size(), "Rank of first input must >= rank of second input.") - if (x_grad) { - x_grad->Resize(x_dims); + auto x_grad_name = framework::GradVarName("X"); + auto y_grad_name = framework::GradVarName("Y"); + if (ctx->HasOutput(x_grad_name)) { + ctx->SetOutputDim(x_grad_name, x_dims); } - - if (y_grad) { - y_grad->Resize(y_dims); + if (ctx->HasOutput(y_grad_name)) { + ctx->SetOutputDim(y_grad_name, y_dims); } } }; diff --git a/paddle/operators/elementwise_op_function.h b/paddle/operators/elementwise_op_function.h new file mode 100644 index 0000000000000000000000000000000000000000..3eb97f60b59848d23bcd15ea1e3d2f21b721f6a4 --- /dev/null +++ b/paddle/operators/elementwise_op_function.h @@ -0,0 +1,200 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" +#include "paddle/framework/operator.h" + +#include "paddle/operators/math/math_function.h" + +namespace paddle { +namespace operators { + +/* + * Out = X ⊙ Y + * If Y's shape does not match X' shape, they will be reshaped. + * For example: + * 1. shape(X) = (2, 3, 4, 5), shape(Y) = (3, 4), with axis=1 + * pre=2, n=3*4, post=5 + * x.shape(2, 12, 5) * y.shape(1,12,1).broadcast(2,12,5) + * 2. shape(X) = (2, 3, 4, 5), shape(Y) = (4,5) + * pre=2*3, n=4*5, post=1 + * x.shape(2, 3, 20) * y.shape(1,1,20).broadcast(2,3,20) + */ +inline void get_mid_dims(const framework::DDim& x_dims, + const framework::DDim& y_dims, const int axis, + int& pre, int& n, int& post) { + pre = 1; + n = 1; + post = 1; + for (int i = 0; i < axis; ++i) { + pre *= x_dims[i]; + } + + for (int i = 0; i < y_dims.size(); ++i) { + PADDLE_ENFORCE_EQ(x_dims[i + axis], y_dims[i], + "Broadcast dimension mismatch."); + n *= y_dims[i]; + } + + for (int i = axis + y_dims.size(); i < x_dims.size(); ++i) { + post *= x_dims[i]; + } +} + +#define EIGEN_FUNCTOR(name, eigen_op) \ + struct Eigen##name##Functor { \ + template \ + inline void Run(const framework::Tensor* x, const framework::Tensor* y, \ + framework::Tensor* z, \ + const framework::ExecutionContext& ctx) { \ + auto x_e = framework::EigenVector::Flatten(*x); \ + auto y_e = framework::EigenVector::Flatten(*y); \ + auto z_e = framework::EigenVector::Flatten(*z); \ + z_e.device(ctx.GetEigenDevice()) = eigen_op(x_e, y_e); \ + } \ + template \ + inline void RunBroadCast(const framework::Tensor* x, \ + const framework::Tensor* y, framework::Tensor* z, \ + const framework::ExecutionContext& ctx, int pre, \ + int n) { \ + auto x_e = framework::EigenVector::Flatten(*x); \ + auto y_e = framework::EigenVector::Flatten(*y); \ + auto z_e = framework::EigenVector::Flatten(*z); \ + auto y_bcast = y_e.reshape(Eigen::DSizes(1, n)) \ + .broadcast(Eigen::DSizes(pre, 1)) \ + .reshape(Eigen::DSizes(x_e.size())); \ + z_e.device(ctx.GetEigenDevice()) = eigen_op(x_e, y_bcast); \ + } \ + template \ + inline void RunBroadCast2(const framework::Tensor* x, \ + const framework::Tensor* y, \ + framework::Tensor* z, \ + const framework::ExecutionContext& ctx, int pre, \ + int n, int post) { \ + auto x_e = framework::EigenVector::Flatten(*x); \ + auto y_e = framework::EigenVector::Flatten(*y); \ + auto z_e = framework::EigenVector::Flatten(*z); \ + auto y_bcast = y_e.reshape(Eigen::DSizes(1, n, 1)) \ + .broadcast(Eigen::DSizes(pre, 1, post)) \ + .reshape(Eigen::DSizes(x_e.size())); \ + z_e.device(ctx.GetEigenDevice()) = eigen_op(x_e, y_bcast); \ + } \ + } + +template +void ElementwiseCompute(const framework::ExecutionContext& ctx) { + using Tensor = framework::Tensor; + + auto* x = ctx.Input("X"); + auto* y = ctx.Input("Y"); + auto* z = ctx.Output("Out"); + z->mutable_data(ctx.GetPlace()); + + auto x_dims = x->dims(); + auto y_dims = y->dims(); + PADDLE_ENFORCE_GE(x_dims.size(), y_dims.size(), + "Rank of first input must >= rank of second input.") + + if (x_dims == y_dims || product(y_dims) == 1) { + functor f; + f.template Run(x, y, z, ctx); + return; + } + + int axis = ctx.Attr("axis"); + axis = (axis == -1 ? x_dims.size() - y_dims.size() : axis); + PADDLE_ENFORCE(axis >= 0 && axis < x_dims.size(), + "Axis should be in range [0, x_dims)"); + + int pre, n, post; + get_mid_dims(x_dims, y_dims, axis, pre, n, post); + if (post == 1) { + functor f; + f.template RunBroadCast(x, y, z, ctx, pre, n); + return; + } else { + functor f; + f.template RunBroadCast2(x, y, z, ctx, pre, n, post); + return; + } +} + +#define EIGEN_ADD(x, y) ((x) + (y)) +EIGEN_FUNCTOR(Add, EIGEN_ADD); + +#define EIGEN_SUB(x, y) ((x) - (y)) +EIGEN_FUNCTOR(Sub, EIGEN_SUB); + +#define EIGEN_MUL(x, y) ((x) * (y)) +EIGEN_FUNCTOR(Mul, EIGEN_MUL); + +#define EIGEN_DIV(x, y) ((x) / (y)) +EIGEN_FUNCTOR(Div, EIGEN_DIV); + +template +void ElementwiseGradCompute(const framework::ExecutionContext& ctx) { + using Tensor = framework::Tensor; + + auto* x = ctx.Input("X"); + auto* y = ctx.Input("Y"); + auto* out = ctx.Input("Out"); + auto* dout = ctx.Input(framework::GradVarName("Out")); + + auto place = ctx.GetEigenDevice(); + + auto x_dims = x->dims(); + auto y_dims = y->dims(); + + auto* dx = ctx.Output(framework::GradVarName("X")); + auto* dy = ctx.Output(framework::GradVarName("Y")); + if (dx) { + dx->mutable_data(ctx.GetPlace()); + } + if (dy) { + dy->mutable_data(ctx.GetPlace()); + } + + if (x_dims == y_dims) { + functor f; + f(place, x, y, out, dx, dy, dout); + return; + } + + if (product(y_dims) == 1) { + functor1 f; + f(place, x, y, out, dx, dy, dout); + return; + } + + int axis = ctx.Attr("axis"); + axis = (axis == -1 ? x_dims.size() - y_dims.size() : axis); + + int pre, n, post; + get_mid_dims(x_dims, y_dims, axis, pre, n, post); + + if (post == 1) { + broadcastfunctor f; + f(place, x, y, out, dx, dy, dout, pre, n); + return; + } else { + broadcast2functor f; + f(place, x, y, out, dx, dy, dout, pre, n, post); + return; + } +} +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/elementwise_sub_op.cc b/paddle/operators/elementwise_sub_op.cc index 31c37ff7ab5595c29f973929387d3945b6f3aaf8..3e4f98fdb35b148931a67d511fe41958eb523f99 100644 --- a/paddle/operators/elementwise_sub_op.cc +++ b/paddle/operators/elementwise_sub_op.cc @@ -13,6 +13,7 @@ limitations under the License. */ #include "paddle/operators/elementwise_sub_op.h" +#include "paddle/operators/elementwise_op.h" namespace paddle { namespace operators { diff --git a/paddle/operators/elementwise_sub_op.h b/paddle/operators/elementwise_sub_op.h index f6bc66cd0e1594a8bc7070e2f182401b92d1c88e..3f40c1c5bcea5e8473765b039de4ee2a16054f0c 100644 --- a/paddle/operators/elementwise_sub_op.h +++ b/paddle/operators/elementwise_sub_op.h @@ -13,13 +13,13 @@ limitations under the License. */ #pragma once -#include "paddle/operators/elementwise_op.h" +#include "paddle/operators/elementwise_op_function.h" namespace paddle { namespace operators { template -class ElementwiseSubKernel : public framework::OpKernel { +class ElementwiseSubKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { ElementwiseCompute(ctx); @@ -102,7 +102,7 @@ struct ElementwiseSubBroadCast2GradFunctor { }; template -class ElementwiseSubGradKernel : public framework::OpKernel { +class ElementwiseSubGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { ElementwiseGradCompute, diff --git a/paddle/operators/fc_op.cc b/paddle/operators/fc_op.cc index 5ac0e8cc45f007d42f1b6d7f86333f5cbedb3ea8..7c422c81fc479fa2e317bdee1b66017096381d27 100644 --- a/paddle/operators/fc_op.cc +++ b/paddle/operators/fc_op.cc @@ -100,7 +100,7 @@ class FCOp : public NetOp { add_out = Output("AddOut"); AppendOp(framework::OpRegistry::CreateOp( - "rowwise_add", {{"X", {sum_out}}, {"b", {Input("B")}}}, + "elementwise_add", {{"X", {sum_out}}, {"Y", {Input("B")}}}, {{"Out", {add_out}}}, {})); } else { if (Output("AddOut") != framework::kEmptyVarName) { diff --git a/paddle/operators/fill_constant_op.cc b/paddle/operators/fill_constant_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..65d03d5fa4426ef4229c1155753c67e59ce98857 --- /dev/null +++ b/paddle/operators/fill_constant_op.cc @@ -0,0 +1,68 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/operators/fill_constant_op.h" + +namespace paddle { +namespace operators { + +class FillConstantOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of FillConstantOp should not be null."); + auto &shape = ctx->Attrs().Get>("shape"); + std::vector shape_int64(shape.size(), 0); + std::transform(shape.begin(), shape.end(), shape_int64.begin(), + [](int a) { return static_cast(a); }); + auto dims = framework::make_ddim(shape_int64); + ctx->SetOutputDim("Out", dims); + } + + framework::DataType IndicateDataType( + const framework::ExecutionContext &ctx) const override { + return static_cast(ctx.Attr("dataType")); + } +}; + +class FillConstantOpMaker : public framework::OpProtoAndCheckerMaker { + public: + FillConstantOpMaker(framework::OpProto *proto, + framework::OpAttrChecker *op_checker) + : framework::OpProtoAndCheckerMaker(proto, op_checker) { + AddAttr("dataType", + "(int, default 5 (FP32)) " + "Output data type") + .SetDefault(framework::DataType::FP32); + AddAttr>("shape", "(vector) The shape of the output"); + AddAttr("value", "(float, default 0) The value to be filled") + .SetDefault(0.0f); + AddOutput("Out", + "(Tensor) Tensor of specified shape will be filled " + "with the specified value"); + AddComment(R"DOC(Fill up a variable with specified constant value.)DOC"); + } +}; +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP_WITHOUT_GRADIENT(fill_constant, ops::FillConstantOp, + ops::FillConstantOpMaker); +REGISTER_OP_CPU_KERNEL( + fill_constant, + ops::FillConstantOpKernel); diff --git a/paddle/operators/rowwise_add_op.cu b/paddle/operators/fill_constant_op.cu similarity index 75% rename from paddle/operators/rowwise_add_op.cu rename to paddle/operators/fill_constant_op.cu index 4a57f64c890ce99d6060faec6a4a01b107403344..eef8fcbd7f65a9891126e039c4d46a106a6daa60 100644 --- a/paddle/operators/rowwise_add_op.cu +++ b/paddle/operators/fill_constant_op.cu @@ -13,11 +13,10 @@ limitations under the License. */ #define EIGEN_USE_GPU -#include "paddle/operators/rowwise_add_op.h" +#include "paddle/framework/op_registry.h" +#include "paddle/operators/fill_constant_op.h" namespace ops = paddle::operators; REGISTER_OP_GPU_KERNEL( - rowwise_add, ops::RowwiseAddKernel); -REGISTER_OP_GPU_KERNEL( - rowwise_add_grad, - ops::RowwiseAddGradKernel); + fill_constant, + ops::FillConstantOpKernel); diff --git a/paddle/operators/add_op.h b/paddle/operators/fill_constant_op.h similarity index 52% rename from paddle/operators/add_op.h rename to paddle/operators/fill_constant_op.h index a7307b6818aa3d10ff215d06281e2b53196fd101..53b8b548eca6dfe035c326d95f91d3e279f63318 100644 --- a/paddle/operators/add_op.h +++ b/paddle/operators/fill_constant_op.h @@ -19,28 +19,17 @@ limitations under the License. */ namespace paddle { namespace operators { -using Tensor = framework::Tensor; -template -using EigenVector = framework::EigenVector; - template -class AddKernel : public framework::OpKernel { +class FillConstantOpKernel : public framework::OpKernel { public: - void Compute(const framework::ExecutionContext& context) const override { - auto* input0 = context.Input("X"); - auto* input1 = context.Input("Y"); - auto* output = context.Output("Out"); - - output->mutable_data(context.GetPlace()); - - auto X = EigenVector::Flatten(*input0); - auto Y = EigenVector::Flatten(*input1); - auto Z = EigenVector::Flatten(*output); - - auto place = context.GetEigenDevice(); - - Z.device(place) = X + Y; + void Compute(const framework::ExecutionContext& ctx) const override { + auto* out = ctx.Output("Out"); + out->mutable_data(ctx.GetPlace()); + auto value = ctx.Attr("value"); + + auto out_eigen = framework::EigenVector::Flatten(*out); + auto place = ctx.GetEigenDevice(); + out_eigen.device(place) = out_eigen.constant(static_cast(value)); } }; diff --git a/paddle/operators/fill_zeros_like_op.cc b/paddle/operators/fill_zeros_like_op.cc index 761a527a5574edc779340ec595dfe1bc1964438a..4c70b9a36bb8f3a2184a8f2ac738378e9d3271e2 100644 --- a/paddle/operators/fill_zeros_like_op.cc +++ b/paddle/operators/fill_zeros_like_op.cc @@ -22,15 +22,13 @@ class FillZerosLikeOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), - "Input(X) of FillZerosLikeOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Y"), - "Output(Y) of FillZerosLikeOp should not be null."); - - ctx.Output("Y")->Resize( - ctx.Input("X")->dims()); - ctx.ShareLoD("X", /*->*/ "Y"); + void InferShape(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of FillZerosLikeOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Y"), + "Output(Y) of FillZerosLikeOp should not be null."); + ctx->SetOutputDim("Y", ctx->GetInputDim("X")); + ctx->ShareLoD("X", /*->*/ "Y"); } }; diff --git a/paddle/operators/fill_zeros_like_op.h b/paddle/operators/fill_zeros_like_op.h index 4474581784531faee1741f0b143743e31cc3788f..cdf56a723b117fe7b08ef2749aa2c2978c923d44 100644 --- a/paddle/operators/fill_zeros_like_op.h +++ b/paddle/operators/fill_zeros_like_op.h @@ -20,7 +20,7 @@ namespace paddle { namespace operators { template -class FillZerosLikeKernel : public framework::OpKernel { +class FillZerosLikeKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto* output = context.Output("Y"); diff --git a/paddle/operators/gather.cu.h b/paddle/operators/gather.cu.h new file mode 100644 index 0000000000000000000000000000000000000000..8d04ecd284226c7b4c6cdd5531915fee2d94ce61 --- /dev/null +++ b/paddle/operators/gather.cu.h @@ -0,0 +1,79 @@ +/* Copyright (c) 2016 PaddlePaddle Authors All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once +#include "paddle/framework/tensor.h" +#include "paddle/platform/place.h" + +namespace paddle { +namespace operators { + +using framework::Tensor; +using platform::Place; + +#define CUDA_1D_KERNEL_LOOP(i, n) \ + for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < (n); \ + i += blockDim.x * gridDim.x) + +template +__global__ void GatherCUDAKernel(const T* params, const int* indices, T* output, + size_t index_size, size_t slice_size) { + CUDA_1D_KERNEL_LOOP(i, index_size * slice_size) { + int indices_i = i / slice_size; + int slice_i = i - indices_i * slice_size; // offset inside the slice + int gather_i = indices[indices_i]; + int params_i = gather_i * slice_size + slice_i; + *(output + i) = *(params + params_i); + } +} + +/** + * A thin wrapper on gpu tensor + * Return a new tensor from source tensor, gathered according to index + * input[src]: type-T source Tensor + * input[index]: type-int index Tensor (1-D) + * return: output tensor + */ +template +void GPUGather(const platform::DeviceContext& ctx, const Tensor& src, + const Tensor& index, Tensor* output) { + // PADDLE_ENFORCE(platform::is_gpu_place(place)); + // check index of shape 1-D + PADDLE_ENFORCE(index.dims().size() == 1); + int index_size = index.dims()[0]; + + auto src_dims = src.dims(); + framework::DDim output_dims(src_dims); + output_dims[0] = index_size; + + // slice size + int slice_size = 1; + for (int i = 1; i < src_dims.size(); ++i) slice_size *= src_dims[i]; + + const T* p_src = src.data(); + const int* p_index = index.data(); + T* p_output = output->data(); + + int block = 512; + int n = slice_size * index_size; + int grid = (n + block - 1) / block; + + GatherCUDAKernel<<< + grid, block, 0, + reinterpret_cast(ctx).stream()>>>( + p_src, p_index, p_output, index_size, slice_size); +} + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/gather.h b/paddle/operators/gather.h index 92fb51ec17709bc6f8abb2f516a9240fb5dc3a77..052db49cb3c2594eca8b9a5e3716689480089703 100644 --- a/paddle/operators/gather.h +++ b/paddle/operators/gather.h @@ -24,49 +24,40 @@ limitations under the License. */ namespace paddle { namespace operators { -// Implementation of CPU copy -template -void CPUGather(const T* src, const int* indices, const int slice_size, - const int index_size, T* output) { - const size_t slice_bytes = slice_size * sizeof(T); - - for (int i = 0; i < index_size; ++i) { - int index_ = indices[i]; - memcpy(output + i * slice_size, src + index_ * slice_size, slice_bytes); - } -} - -// Implementation of GPU copy: -template -void GPUGather(const T* src, const int* index, const int slice_size, - const int index_size, T* output); +using framework::Tensor; /** + * A thin wrapper for gathering on cpu tensor * Return a new tensor from source tensor, gathered according to index * input[src]: type-T source Tensor * input[index]: type-int index Tensor (1-D) * return: output tensor */ template -void Gather(const platform::Place& place, const paddle::framework::Tensor* src, - const paddle::framework::Tensor* index, - paddle::framework::Tensor* output) { +void CPUGather(const platform::DeviceContext& ctx, const Tensor& src, + const Tensor& index, Tensor* output) { + PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace())); // check index of shape 1-D - PADDLE_ENFORCE(index->dims().size() == 1); - int index_size = index->dims()[0]; + PADDLE_ENFORCE(index.dims().size() == 1); + int index_size = index.dims()[0]; - auto src_dims = src->dims(); + auto src_dims = src.dims(); framework::DDim output_dims(src_dims); output_dims[0] = index_size; + const T* p_src = src.data(); + const int* p_index = index.data(); + T* p_output = output->data(); + // slice size int slice_size = 1; for (int i = 1; i < src_dims.size(); ++i) slice_size *= src_dims[i]; - // Gathering - if (platform::is_cpu_place(place)) { - CPUGather(src->data(), index->data(), slice_size, index_size, - output->data()); + const size_t slice_bytes = slice_size * sizeof(T); + + for (int i = 0; i < index_size; ++i) { + int index_ = p_index[i]; + memcpy(p_output + i * slice_size, p_src + index_ * slice_size, slice_bytes); } } diff --git a/paddle/operators/gather_op.cc b/paddle/operators/gather_op.cc index fecd1ce2147a1e6f2f7928266be74ed7b647c5b9..fb99c6c01670fcbacd012573b0679ac755638c57 100644 --- a/paddle/operators/gather_op.cc +++ b/paddle/operators/gather_op.cc @@ -23,19 +23,26 @@ class GatherOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), - "Input(X) of GatherOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Index"), - "Input(Index) of GatherOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"), - "Output(Out) of GatherOp should not be null."); + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of GatherOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Index"), + "Input(Index) of GatherOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of GatherOp should not be null."); - int batch_size = ctx.Input("Index")->dims()[0]; + auto index_dims = ctx->GetInputDim("Index"); + PADDLE_ENFORCE(index_dims.size() == 1); + int batch_size = ctx->GetInputDim("Index")[0]; PADDLE_ENFORCE_GE(batch_size, 0, "Batch size must be >0"); - framework::DDim output_dims(ctx.Input("X")->dims()); + framework::DDim output_dims(ctx->GetInputDim("X")); output_dims[0] = batch_size; - ctx.Output("Out")->Resize(output_dims); + ctx->SetOutputDim("Out", output_dims); + } + + framework::DataType IndicateDataType( + const framework::ExecutionContext& ctx) const override { + return framework::ToDataType(ctx.Input("X")->type()); } }; @@ -44,23 +51,25 @@ class GatherGradOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - auto X_grad = ctx.Output(framework::GradVarName("X")); - auto X = ctx.Input("X"); + void InferShape(framework::InferShapeContext* ctx) const override { + ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X")); + } - X_grad->Resize(X->dims()); + framework::DataType IndicateDataType( + const framework::ExecutionContext& ctx) const override { + return framework::ToDataType(ctx.Input("X")->type()); } }; class GatherOpMaker : public framework::OpProtoAndCheckerMaker { public: - GatherOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + GatherOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "The source input of gather op"); AddInput("Index", "The index input of gather op"); AddOutput("Out", "The output of add op"); AddComment(R"DOC( -Gather Operator by selecting from the first axis, +Gather Operator by selecting from the first axis, Out = X[Index] )DOC"); @@ -72,8 +81,5 @@ Out = X[Index] namespace ops = paddle::operators; REGISTER_OP(gather, ops::GatherOp, ops::GatherOpMaker, gather_grad, ops::GatherGradOp); -REGISTER_OP_CPU_KERNEL(gather, - ops::GatherOpKernel); -REGISTER_OP_CPU_KERNEL( - gather_grad, - ops::GatherGradientOpKernel); +REGISTER_OP_CPU_KERNEL(gather, ops::GatherOpKernel); +REGISTER_OP_CPU_KERNEL(gather_grad, ops::GatherGradientOpKernel); diff --git a/paddle/operators/gather_op.cu b/paddle/operators/gather_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..92219d6a433e6db0bb9886ed8670cbafaa843ff8 --- /dev/null +++ b/paddle/operators/gather_op.cu @@ -0,0 +1,64 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "gather.cu.h" +#include "paddle/framework/eigen.h" +#include "paddle/operators/gather_op.h" +#include "scatter.cu.h" + +namespace paddle { +namespace operators { + +template +class GatherOpCUDAKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &ctx) const override { + PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), + "This kernel only runs on GPU device."); + auto *x = ctx.Input("X"); + auto *index = ctx.Input("Index"); + auto *output = ctx.Output("Out"); + + output->mutable_data(ctx.GetPlace()); + + GPUGather(ctx.device_context(), *x, *index, output); + } +}; + +template +class GatherGradOpCUDAKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &ctx) const override { + PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), + "This kernel only runs on GPU device."); + auto *Index = ctx.Input("Index"); + auto *dX = ctx.Output(framework::GradVarName("X")); + auto *dO = ctx.Input(framework::GradVarName("Out")); + auto *x = ctx.Input("X"); + + dX->mutable_data(ctx.GetPlace()); + auto dxt = framework::EigenVector::Flatten(*dX); + auto place = ctx.GetEigenDevice(); + dxt.device(place) = dxt.constant(static_cast(0)); + + GPUScatterAssign(ctx.device_context(), *dO, *Index, dX); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP_GPU_KERNEL(gather, ops::GatherOpCUDAKernel); +REGISTER_OP_GPU_KERNEL(gather_grad, ops::GatherGradOpCUDAKernel); diff --git a/paddle/operators/gather_op.h b/paddle/operators/gather_op.h index 381854f301870beadb72d9e9b4eb17ff199960fb..8276ed0d3d8b676aafab45fae70942e78b72b8e6 100644 --- a/paddle/operators/gather_op.h +++ b/paddle/operators/gather_op.h @@ -23,29 +23,40 @@ namespace operators { using Tensor = framework::Tensor; -template -class GatherOpKernel : public framework::OpKernel { +template +class GatherOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext &ctx) const override { - auto *X = ctx.Input("X"); - auto *Index = ctx.Input("Index"); - auto *Y = ctx.Output("Out"); + PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()), + "This kernel only runs on CPU."); + + auto *x = ctx.Input("X"); + auto *index = ctx.Input("Index"); + auto *output = ctx.Output("Out"); + + output->mutable_data(ctx.GetPlace()); - Y->mutable_data(ctx.GetPlace()); - Gather(ctx.GetPlace(), X, Index, Y); + CPUGather(ctx.device_context(), *x, *index, output); } }; -template -class GatherGradientOpKernel : public framework::OpKernel { +template +class GatherGradientOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext &ctx) const override { + PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()), + "This kernel only runs on CPU."); + auto *Index = ctx.Input("Index"); auto *dX = ctx.Output(framework::GradVarName("X")); auto *dO = ctx.Input(framework::GradVarName("Out")); dX->mutable_data(ctx.GetPlace()); - ScatterUpdate(ctx.GetPlace(), dO, Index, dX); + auto dxt = framework::EigenVector::Flatten(*dX); + auto place = ctx.GetEigenDevice(); + dxt.device(place) = dxt.constant(static_cast(0)); + + ScatterAssign(ctx.device_context(), *dO, *Index, dX); } }; diff --git a/paddle/operators/gather_test.cc b/paddle/operators/gather_test.cc index 0ae1e99452973feb6d085dd6ef51e2afca988f59..cbd86b87961ee24aa889e208de5ac38e03a33135 100644 --- a/paddle/operators/gather_test.cc +++ b/paddle/operators/gather_test.cc @@ -41,7 +41,9 @@ TEST(Gather, GatherData) { int* p_output = output->mutable_data(make_ddim({2, 4}), CPUPlace()); - Gather(CPUPlace(), src, index, output); + auto* cpu_place = new paddle::platform::CPUPlace(); + paddle::platform::CPUDeviceContext ctx(*cpu_place); + CPUGather(ctx, *src, *index, output); for (int i = 0; i < 4; ++i) EXPECT_EQ(p_output[i], i + 4); for (int i = 4; i < 8; ++i) EXPECT_EQ(p_output[i], i - 4); diff --git a/paddle/operators/gaussian_random_op.cc b/paddle/operators/gaussian_random_op.cc index 5b7cbb5cc7bcb7e43b15363d37d7b8f2cbf0fbdc..ca7fb385057dda4ee6f7dea7edc21d9ebbf6b48e 100644 --- a/paddle/operators/gaussian_random_op.cc +++ b/paddle/operators/gaussian_random_op.cc @@ -16,7 +16,7 @@ namespace paddle { namespace operators { template -class CPUGaussianRandomKernel : public framework::OpKernel { +class CPUGaussianRandomKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { float mean = context.Attr("mean"); @@ -43,13 +43,10 @@ class GaussianRandomOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext& ctx) const override { - PADDLE_ENFORCE_NOT_NULL( - ctx.OutputVar("Out"), - "Output(Out) of GaussianRandomOp should not be null."); - - auto* tensor = ctx.Output("Out"); - auto dims = Attr>("dims"); + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of GaussianRandomOp should not be null."); + auto dims = ctx->Attrs().Get>("dims"); std::vector temp; temp.reserve(dims.size()); for (auto dim : dims) { @@ -57,7 +54,12 @@ class GaussianRandomOp : public framework::OperatorWithKernel { } PADDLE_ENFORCE(dims.size() > 0UL, "dims can be one int or array. dims must be set."); - tensor->Resize(framework::make_ddim(temp)); + ctx->SetOutputDim("Out", framework::make_ddim(temp)); + } + + framework::DataType IndicateDataType( + const framework::ExecutionContext& ctx) const override { + return static_cast(Attr("data_type")); } }; @@ -79,6 +81,8 @@ Use to initialize tensor with gaussian random generator. "Random seed of generator." "0 means use system wide seed") .SetDefault(0); + AddAttr("data_type", "output data type") + .SetDefault(framework::DataType::FP32); } }; diff --git a/paddle/operators/gaussian_random_op.cu b/paddle/operators/gaussian_random_op.cu index 2d63b3049988cfc3135a87a57dad56b970df3eab..315560bf1ba8a66b9a3b7d79510d202885e845d6 100644 --- a/paddle/operators/gaussian_random_op.cu +++ b/paddle/operators/gaussian_random_op.cu @@ -37,7 +37,7 @@ struct GaussianGenerator { }; template -class GPUGaussianRandomKernel : public framework::OpKernel { +class GPUGaussianRandomKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto* tensor = context.Output("Out"); diff --git a/paddle/operators/gemm_conv2d_op.h b/paddle/operators/gemm_conv2d_op.h index 5c9e81732aa72211c2021382cf9a907880c53c17..323e3f7c3bd506c6b63bf4d1152384649f5da575 100644 --- a/paddle/operators/gemm_conv2d_op.h +++ b/paddle/operators/gemm_conv2d_op.h @@ -25,7 +25,7 @@ namespace operators { using Tensor = framework::Tensor; template -class GemmConv2DKernel : public framework::OpKernel { +class GemmConv2DKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { const Tensor* input = context.Input("Input"); @@ -98,7 +98,7 @@ class GemmConv2DKernel : public framework::OpKernel { }; template -class GemmConvGrad2DKernel : public framework::OpKernel { +class GemmConvGrad2DKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { const Tensor* input = context.Input("Input"); diff --git a/paddle/operators/interp_op.cc b/paddle/operators/interp_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..d02b01c3f3a1b30ec27253140203b076a98ce0c2 --- /dev/null +++ b/paddle/operators/interp_op.cc @@ -0,0 +1,113 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/framework/op_registry.h" +#include "paddle/operators/net_op.h" + +namespace paddle { +namespace operators { + +class InterpOp : public NetOp { + public: + InterpOp(const std::string &type, const framework::VariableNameMap &inputs, + const framework::VariableNameMap &outputs, + const framework::AttributeMap &attrs) + : NetOp(type, inputs, outputs, attrs) { + PADDLE_ENFORCE_NE(Input("X"), framework::kEmptyVarName, + "Input(X) of InterpOp should not be null."); + PADDLE_ENFORCE_NE(Input("Y"), framework::kEmptyVarName, + "Input(Y) of InterpOp should not be null."); + PADDLE_ENFORCE_NE(Input("W"), framework::kEmptyVarName, + "Input(W) of InterpOp should not be null."); + PADDLE_ENFORCE_NE(Output("SubOut"), framework::kEmptyVarName, + "Output(SubOut) of InterpOp should not be null."); + PADDLE_ENFORCE_NE(Output("MulOut"), framework::kEmptyVarName, + "Output(MulOut) of InterpOp should not be null."); + PADDLE_ENFORCE_NE(Output("Out"), framework::kEmptyVarName, + "Output(Out) of InterpOp should not be null."); + + // SubOut = X - Y + auto x = Input("X"); + auto y = Input("Y"); + auto sub_out = Output("SubOut"); + AppendOp(framework::OpRegistry::CreateOp( + "elementwise_sub", {{"X", {x}}, {"Y", {y}}}, {{"Out", {sub_out}}}, {})); + + // MulOut = SubOut * W = (X - Y) * W + auto w = Input("W"); + auto mul_out = Output("MulOut"); + AppendOp(framework::OpRegistry::CreateOp( + "elementwise_mul", {{"X", {sub_out}}, {"Y", {w}}}, {{"Out", {mul_out}}}, + {{"axis", 0}})); + + // Out = MulOut + Y = (X - Y) * W + Y = X * W + Y * (1 - W) + AppendOp(framework::OpRegistry::CreateOp("elementwise_add", + {{"X", {mul_out}}, {"Y", {y}}}, + {{"Out", {Output("Out")}}}, {})); + + CompleteAddOp(false); + } +}; + +class InterpOpMaker : public framework::OpProtoAndCheckerMaker { + public: + InterpOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", + "(Tensor), 2-D Matrix of shape [batch_size, data_dim]" + "containing data samples, the first input of interp_op"); + AddInput("Y", + "(Tensor), 2-D Matrix of shape `[batch_size, data_dim]`" + "containing data samples, the second input of interp_op"); + AddInput("W", + "(Tensor), 1-D Vector of shape [batch_size]," + "the interpolated values in the half-open interval [0.0, 1.0)"); + AddOutput("SubOut", + "(Tensor), the intermediate subtraction outputs, saving X - Y.") + .AsIntermediate(); + AddOutput("MulOut", + "(Tensor), the intermediate multiplication outputs," + "saving the elementwise multiplication of (X - Y) and W.") + .AsIntermediate(); + AddOutput("Out", + "(Tensor), the output of interp_op, same shape with X," + "returns the first-dimensional piecewise linear interpolant " + "between X and Y"); + AddComment(R"DOC( + Linear Interpolation with two inputs, used in NEURAL TURING MACHINE. + + Equation: + Out.row[i] = X.row[i] * W[i] + Y.row[i] * (1 - W[i]) + = (X.row[i] - Y.row[i]) * W[i] + Y.row[i] + + Example: + X = [[1,2],[3,4]], + Y = [[2,1],[4,3]], + W = [0.3, 0.4] + + Then, Out = [[1.7,1.3],[3.6,3.4]] + + where 1.7 = 1*0.3+2*(1-0.3), + 1.3 = 2*0.3+1*(1-0.3), + 3.6 = 3*0.4+4*(1-0.4), + 3.4 = 4*0.4+3*(1-0.4) +)DOC"); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP_WITHOUT_GRADIENT(interp, ops::InterpOp, ops::InterpOpMaker); diff --git a/paddle/operators/lookup_table_op.cc b/paddle/operators/lookup_table_op.cc index 04ac24662e9cfec6a49cd213cb76bdebc7b730c8..3f8d4ab85756ec007004e397a55ea60e2fa704ae 100644 --- a/paddle/operators/lookup_table_op.cc +++ b/paddle/operators/lookup_table_op.cc @@ -22,27 +22,31 @@ class LookupTableOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("W"), - "Input(W) of LookupTableOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Ids"), - "Input(Ids) of LookupTableOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"), - "Output(Out) of LookupTableOp should not be null."); - - auto table_t = ctx.Input("W"); - auto ids_t = ctx.Input("Ids"); - auto output_t = ctx.Output("Out"); - - output_t->Resize({ids_t->dims()[0], table_t->dims()[1]}); - ctx.ShareLoD("Ids", /*->*/ "Out"); + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("W"), + "Input(W) of LookupTableOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Ids"), + "Input(Ids) of LookupTableOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of LookupTableOp should not be null."); + + auto table_dims = ctx->GetInputDim("W"); + auto ids_dims = ctx->GetInputDim("Ids"); + + ctx->SetOutputDim("Out", {ids_dims[0], table_dims[1]}); + ctx->ShareLoD("Ids", /*->*/ "Out"); + } + + framework::DataType IndicateDataType( + const framework::ExecutionContext& ctx) const override { + return framework::ToDataType(ctx.Input("W")->type()); } }; class LookupTableOpMaker : public framework::OpProtoAndCheckerMaker { public: - LookupTableOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + LookupTableOpMaker(framework::OpProto* proto, + framework::OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("W", "An input represents embedding tensors," @@ -66,11 +70,14 @@ class LookupTableOpGrad : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &context) const override { - auto table = context.Input("W"); - auto d_table = - context.Output(framework::GradVarName("W")); - d_table->Resize(table->dims()); + void InferShape(framework::InferShapeContext* ctx) const override { + auto table_dims = ctx->GetInputDim("W"); + ctx->SetOutputDim(framework::GradVarName("W"), table_dims); + } + + framework::DataType IndicateDataType( + const framework::ExecutionContext& ctx) const override { + return framework::ToDataType(ctx.Input("W")->type()); } }; diff --git a/paddle/operators/lookup_table_op.cu b/paddle/operators/lookup_table_op.cu index 708344046760691aa2da562eb1ee3d8b130c5f18..c3808fa9a8de031fcae3ac0417e8c4330b2f5aad 100644 --- a/paddle/operators/lookup_table_op.cu +++ b/paddle/operators/lookup_table_op.cu @@ -61,7 +61,7 @@ __global__ void LookupTableGrad(T* table, const T* output, const int32_t* ids, } template -class LookupTableCUDAKernel : public framework::OpKernel { +class LookupTableCUDAKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto table_t = context.Input("W"); @@ -77,12 +77,15 @@ class LookupTableCUDAKernel : public framework::OpKernel { dim3 threads(128, 8); dim3 grids(8, 1); - LookupTable<<>>(output, table, ids, N, K, D); + LookupTable<<< + grids, threads, 0, reinterpret_cast( + context.device_context()) + .stream()>>>(output, table, ids, N, K, D); } }; template -class LookupTableGradCUDAKernel : public framework::OpKernel { +class LookupTableGradCUDAKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto ids_t = context.Input("Ids"); @@ -102,8 +105,10 @@ class LookupTableGradCUDAKernel : public framework::OpKernel { dim3 threads(128, 8); dim3 grids(8, 1); - LookupTableGrad<<>>(d_table, d_output, ids, N, - K, D); + LookupTableGrad<<< + grids, threads, 0, reinterpret_cast( + context.device_context()) + .stream()>>>(d_table, d_output, ids, N, K, D); } }; diff --git a/paddle/operators/lookup_table_op.h b/paddle/operators/lookup_table_op.h index a1298906dd4b4209644fe06584f70169519de01c..dfead2fc5b25b9be26bb19cd74a3a94daf62cca6 100644 --- a/paddle/operators/lookup_table_op.h +++ b/paddle/operators/lookup_table_op.h @@ -23,7 +23,7 @@ namespace operators { using Tensor = framework::Tensor; template -class LookupTableKernel : public framework::OpKernel { +class LookupTableKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto table_t = context.Input("W"); // float tensor @@ -44,7 +44,7 @@ class LookupTableKernel : public framework::OpKernel { }; template -class LookupTableGradKernel : public framework::OpKernel { +class LookupTableGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto ids_t = context.Input("Ids"); diff --git a/paddle/operators/lstm_unit_op.cc b/paddle/operators/lstm_unit_op.cc index 3600f199770c4b8c9a6561b4c270a91bc8b20c0b..13a45ec246988558b0c090ab14e627b5b0e44532 100644 --- a/paddle/operators/lstm_unit_op.cc +++ b/paddle/operators/lstm_unit_op.cc @@ -22,37 +22,35 @@ class LstmUnitOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), - "Input(X) of LSTM should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("C_prev"), - "Input(C_prev) of LSTM should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("C"), - "Output(C) of LSTM should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("H"), - "Output(H) of LSTM should not be null."); - - auto *x = ctx.Input("X"); - auto *c_prev = ctx.Input("C_prev"); - - PADDLE_ENFORCE_EQ(x->dims().size(), 2, "Input(X)'s rank must be 2."); - PADDLE_ENFORCE(x->dims()[0] == c_prev->dims()[0], + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) of LSTM should not be null."); + PADDLE_ENFORCE(ctx->HasInput("C_prev"), + "Input(C_prev) of LSTM should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("C"), + "Output(C) of LSTM should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("H"), + "Output(H) of LSTM should not be null."); + + auto x_dims = ctx->GetInputDim("X"); + auto c_prev_dims = ctx->GetInputDim("C_prev"); + + PADDLE_ENFORCE_EQ(x_dims.size(), 2, "Input(X)'s rank must be 2."); + PADDLE_ENFORCE(x_dims[0] == c_prev_dims[0], "Batch size of inputs and states must be equal"); - PADDLE_ENFORCE(x->dims()[1] == c_prev->dims()[1] * 4, + PADDLE_ENFORCE(x_dims[1] == c_prev_dims[1] * 4, "Dimension of FC should equal to prev state * 4"); - int b_size = c_prev->dims()[0]; // batch size - int s_dim = c_prev->dims()[1]; // state dim - ctx.Output("C")->Resize({b_size, s_dim}); - ctx.Output("H")->Resize({b_size, s_dim}); + int b_size = c_prev_dims[0]; // batch size + int s_dim = c_prev_dims[1]; // state dim + ctx->SetOutputDim("C", {b_size, s_dim}); + ctx->SetOutputDim("H", {b_size, s_dim}); } }; -template class LstmUnitOpMaker : public framework::OpProtoAndCheckerMaker { public: - LstmUnitOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + LstmUnitOpMaker(framework::OpProto* proto, + framework::OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "FC input before the non-linear activation."); AddInput( @@ -63,13 +61,13 @@ class LstmUnitOpMaker : public framework::OpProtoAndCheckerMaker { AddComment(R"DOC(Lstm-Unit Operator -Equation: +Equation: i, f, o, j = split(X) C = C_prev * sigm(f + forget_bias) + sigm(i) * tanh(j) H = C * sigm(o) - + )DOC"); - AddAttr("forget_bias", "The forget bias of Lstm Unit.") + AddAttr("forget_bias", "The forget bias of Lstm Unit.") .SetDefault(0.0); } }; @@ -79,15 +77,14 @@ class LstmUnitGradOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("C")), - "Input(C@GRAD) should not be null"); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("H")), - "Input(H@GRAD) should not be null"); - ctx.Output(framework::GradVarName("X")) - ->Resize(ctx.Input("X")->dims()); - ctx.Output(framework::GradVarName("C_prev")) - ->Resize(ctx.Input("C_prev")->dims()); + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("C")), + "Input(C@GRAD) should not be null"); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("H")), + "Input(H@GRAD) should not be null"); + ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X")); + ctx->SetOutputDim(framework::GradVarName("C_prev"), + ctx->GetInputDim("C_prev")); } }; @@ -95,9 +92,11 @@ class LstmUnitGradOp : public framework::OperatorWithKernel { } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP(lstm_unit, ops::LstmUnitOp, ops::LstmUnitOpMaker, - lstm_unit_grad, ops::LstmUnitGradOp); +REGISTER_OP(lstm_unit, ops::LstmUnitOp, ops::LstmUnitOpMaker, lstm_unit_grad, + ops::LstmUnitGradOp); REGISTER_OP_CPU_KERNEL(lstm_unit, - ops::LstmUnitKernel); + ops::LstmUnitKernel, + ops::LstmUnitKernel); REGISTER_OP_CPU_KERNEL( - lstm_unit_grad, ops::LstmUnitGradKernel); + lstm_unit_grad, ops::LstmUnitGradKernel, + ops::LstmUnitGradKernel); diff --git a/paddle/operators/lstm_unit_op.cu b/paddle/operators/lstm_unit_op.cu index 6e5e4978994c281416a65af5f8ffdec688768d63..49ea550b6f49a13bf31d14321d7a9eb13a834d4b 100644 --- a/paddle/operators/lstm_unit_op.cu +++ b/paddle/operators/lstm_unit_op.cu @@ -89,8 +89,8 @@ __global__ void LSTMUnitGradientKernel(const int nthreads, const int dim, } } -template -class LstmUnitOpCUDAKernel : public framework::OpKernel { +template +class LstmUnitOpCUDAKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), @@ -101,7 +101,7 @@ class LstmUnitOpCUDAKernel : public framework::OpKernel { auto* c_tensor = ctx.Output("C"); auto* h_tensor = ctx.Output("H"); - auto forget_bias = static_cast(ctx.Attr("forget_bias")); + auto forget_bias = static_cast(ctx.Attr("forget_bias")); int b_size = c_tensor->dims()[0]; int D = c_tensor->dims()[1]; @@ -120,8 +120,8 @@ class LstmUnitOpCUDAKernel : public framework::OpKernel { } }; -template -class LstmUnitGradOpCUDAKernel : public framework::OpKernel { +template +class LstmUnitGradOpCUDAKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), @@ -153,7 +153,7 @@ class LstmUnitGradOpCUDAKernel : public framework::OpKernel { int N = c_tensor->dims()[0]; int D = c_tensor->dims()[1]; - auto forget_bias = static_cast(ctx.Attr("forget_bias")); + auto forget_bias = static_cast(ctx.Attr("forget_bias")); int block = 512; int n = N * D; @@ -169,5 +169,7 @@ class LstmUnitGradOpCUDAKernel : public framework::OpKernel { } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL(lstm_unit, ops::LstmUnitOpCUDAKernel); -REGISTER_OP_GPU_KERNEL(lstm_unit_grad, ops::LstmUnitGradOpCUDAKernel); +REGISTER_OP_GPU_KERNEL(lstm_unit, ops::LstmUnitOpCUDAKernel, + ops::LstmUnitOpCUDAKernel); +REGISTER_OP_GPU_KERNEL(lstm_unit_grad, ops::LstmUnitGradOpCUDAKernel, + ops::LstmUnitGradOpCUDAKernel); diff --git a/paddle/operators/lstm_unit_op.h b/paddle/operators/lstm_unit_op.h index 683034fe15df8cabfdff5e856adb5c0467055064..a0ff498c1d3ed2aaa10f5473ef91de168c250649 100644 --- a/paddle/operators/lstm_unit_op.h +++ b/paddle/operators/lstm_unit_op.h @@ -32,8 +32,8 @@ inline T tanh(T x) { return 2. * sigmoid(2. * x) - 1.; } -template -class LstmUnitKernel : public framework::OpKernel { +template +class LstmUnitKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()), @@ -44,7 +44,7 @@ class LstmUnitKernel : public framework::OpKernel { auto* c_tensor = ctx.Output("C"); auto* h_tensor = ctx.Output("H"); - auto forget_bias = static_cast(ctx.Attr("forget_bias")); + auto forget_bias = static_cast(ctx.Attr("forget_bias")); int b_size = c_tensor->dims()[0]; int D = c_tensor->dims()[1]; @@ -75,8 +75,8 @@ class LstmUnitKernel : public framework::OpKernel { } }; -template -class LstmUnitGradKernel : public framework::OpKernel { +template +class LstmUnitGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()), @@ -108,7 +108,7 @@ class LstmUnitGradKernel : public framework::OpKernel { int N = c_tensor->dims()[0]; int D = c_tensor->dims()[1]; - auto forget_bias = static_cast(ctx.Attr("forget_bias")); + auto forget_bias = static_cast(ctx.Attr("forget_bias")); for (int n = 0; n < N; ++n) { for (int d = 0; d < D; ++d) { diff --git a/paddle/operators/math/CMakeLists.txt b/paddle/operators/math/CMakeLists.txt index f8333f34f7b4c7b0f9a0f14a7a33f9d98e1d331c..a0ceb029e3abee2fe591325ffa3100168c3aa8e3 100644 --- a/paddle/operators/math/CMakeLists.txt +++ b/paddle/operators/math/CMakeLists.txt @@ -1,10 +1,13 @@ - if(WITH_GPU) - nv_library(math_function SRCS math_function.cc math_function.cu im2col.cc - im2col.cu DEPS cblas device_context) + nv_library(math_function SRCS math_function.cc math_function.cu im2col.cc im2col.cu pooling.cc pooling.cu DEPS cblas device_context operator) + nv_test(math_function_test SRCS math_function_test.cc DEPS math_function tensor) + nv_library(softmax SRCS softmax.cc softmax.cu DEPS operator) + nv_library(cross_entropy SRCS cross_entropy.cc cross_entropy.cu DEPS operator) else() - cc_library(math_function SRCS math_function.cc im2col.cc DEPS cblas device_context) + cc_library(math_function SRCS math_function.cc im2col.cc pooling.cc DEPS cblas device_context operator) + cc_test(math_function_test SRCS math_function_test.cc DEPS math_function tensor) + cc_library(softmax SRCS softmax.cc DEPS operator) + cc_library(cross_entropy SRCS cross_entropy.cc DEPS operator) endif() -nv_test(math_function_test SRCS math_function_test.cc DEPS math_function tensor) cc_test(im2col_test SRCS im2col_test.cc DEPS math_function tensor) diff --git a/paddle/operators/math/cross_entropy.cc b/paddle/operators/math/cross_entropy.cc new file mode 100644 index 0000000000000000000000000000000000000000..150a65f2751aaeac17f9403404d2efd990a0c72b --- /dev/null +++ b/paddle/operators/math/cross_entropy.cc @@ -0,0 +1,59 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/operators/math/cross_entropy.h" + +namespace paddle { +namespace operators { +namespace math { + +using Tensor = framework::Tensor; +template +using EigenMatrix = framework::EigenMatrix; + +template +class CrossEntropyFunctor { + public: + void operator()(const platform::DeviceContext& ctx, framework::Tensor* out, + const framework::Tensor* prob, + const framework::Tensor* labels, const bool softLabel) { + const int batch_size = prob->dims()[0]; + if (softLabel) { + auto in = EigenMatrix::From(*prob); + auto lbl = EigenMatrix::From(*labels); + auto loss = EigenMatrix::From(*out); + + loss.device(*ctx.GetEigenDevice()) = + -((lbl * in.log().unaryExpr(math::TolerableValue())) + .sum(Eigen::DSizes(1)) + .reshape(Eigen::DSizes(batch_size, 1))); + } else { + const int class_num = prob->dims()[1]; + const T* prob_data = prob->data(); + T* loss_data = out->data(); + + const int* label_data = labels->data(); + for (int i = 0; i < batch_size; ++i) { + int index = i * class_num + label_data[i]; + loss_data[i] = -math::TolerableValue()(std::log(prob_data[index])); + } + } + } +}; + +template class CrossEntropyFunctor; +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/math/cross_entropy.cu b/paddle/operators/math/cross_entropy.cu new file mode 100644 index 0000000000000000000000000000000000000000..367190e6b0682ec62550e869e2f04c3a2b2cbec3 --- /dev/null +++ b/paddle/operators/math/cross_entropy.cu @@ -0,0 +1,109 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/operators/math/cross_entropy.h" + +namespace paddle { +namespace operators { +namespace math { + +namespace { +template +__global__ void CrossEntropyKernel(T* Y, const T* X, const int* label, + const int N, const int D) { + // TOOD(qingqing) define CUDA_1D_KERNEL_LOOP macro in a common file. + // CUDA_1D_KERNEL_LOOP(i, N) { + for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < N; + i += blockDim.x * gridDim.x) { + PADDLE_ASSERT(label[i] >= 0 && label[i] < D); + Y[i] = -math::TolerableValue()(log(X[i * D + label[i]])); + } +} + +template +__device__ __forceinline__ T sum_single_warp(T val) { + val += __shfl_down(val, 16); + val += __shfl_down(val, 8); + val += __shfl_down(val, 4); + val += __shfl_down(val, 2); + val += __shfl_down(val, 1); + return val; +} + +template +__global__ void SoftCrossEntropyKernel(T* Y, const T* X, const T* label, + const int class_num) { + int tid = threadIdx.x; + extern __shared__ T d_sum[]; + d_sum[tid] = 0; + + int cur_idx = tid; + int next_idx = blockIdx.x * class_num + tid; + while (cur_idx < class_num) { + d_sum[tid] += + math::TolerableValue()(std::log(X[next_idx])) * label[next_idx]; + next_idx += blockDim.x; + cur_idx += blockDim.x; + } + __syncthreads(); + + for (unsigned int stride = blockDim.x >> 1; stride >= 32; stride >>= 1) { + if (tid < stride) d_sum[tid] += d_sum[tid + stride]; + __syncthreads(); + } + + T val = d_sum[tid]; + val = sum_single_warp(val); + if (tid == 0) Y[blockIdx.x] = -val; +} +} // namespace + +using Tensor = framework::Tensor; + +template +class CrossEntropyFunctor { + public: + void operator()(const platform::DeviceContext& ctx, framework::Tensor* out, + const framework::Tensor* prob, + const framework::Tensor* labels, bool softLabel) { + const T* prob_data = prob->data(); + T* loss_data = out->mutable_data(ctx.GetPlace()); + + int batch_size = prob->dims()[0]; + int class_num = prob->dims()[1]; + + if (softLabel) { + const T* label_data = labels->data(); + int block = class_num > 512 ? 512 : pow(2, int(std::log2(class_num))); + + SoftCrossEntropyKernel<<< + batch_size, block, block * sizeof(T), + reinterpret_cast(ctx).stream()>>>( + loss_data, prob_data, label_data, class_num); + } else { + const int* label_data = labels->data(); + int block = 512; + int grid = (batch_size + block - 1) / block; + CrossEntropyKernel<<< + grid, block, 0, + reinterpret_cast(ctx).stream()>>>( + loss_data, prob_data, label_data, batch_size, class_num); + } + } +}; + +template class CrossEntropyFunctor; +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/math/cross_entropy.h b/paddle/operators/math/cross_entropy.h new file mode 100644 index 0000000000000000000000000000000000000000..0ab6827ffa8f8b90b432a801607a97206e010cf4 --- /dev/null +++ b/paddle/operators/math/cross_entropy.h @@ -0,0 +1,46 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once +#include "paddle/framework/eigen.h" +#include "paddle/framework/operator.h" +#include "paddle/framework/tensor.h" +#include "paddle/platform/hostdevice.h" + +namespace paddle { +namespace operators { +namespace math { + +template +struct TolerableValue { + HOSTDEVICE T operator()(const T& x) const { + PADDLE_ASSERT(std::is_floating_point::value); + const T kApproInf = 1e20; + + if (x == INFINITY) return kApproInf; + if (x == -INFINITY) return -kApproInf; + return x; + } +}; + +template +class CrossEntropyFunctor { + public: + void operator()(const platform::DeviceContext& context, + framework::Tensor* out, const framework::Tensor* prob, + const framework::Tensor* labels, const bool softLabel); +}; +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/math/im2col_test.cc b/paddle/operators/math/im2col_test.cc index f0b8c885918afe7f80edc465c6d9be7c11ac066f..40bdbfe73351a609a4ab9fdc27ac5ff6710df2a2 100644 --- a/paddle/operators/math/im2col_test.cc +++ b/paddle/operators/math/im2col_test.cc @@ -71,7 +71,7 @@ void testIm2col() { context = new paddle::platform::CPUDeviceContext(paddle::platform::CPUPlace()); } else { -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA context = new paddle::platform::CUDADeviceContext(paddle::platform::GPUPlace()); #else @@ -116,7 +116,7 @@ void testIm2col() { TEST(math, im2col) { testIm2col(); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA testIm2col(); #endif } diff --git a/paddle/operators/math/math_function.h b/paddle/operators/math/math_function.h index 43306fca73387b7b212f556a2b187df113a1b327..473eff4d198ca9b17b6af8eebd6dfe39d49d138d 100644 --- a/paddle/operators/math/math_function.h +++ b/paddle/operators/math/math_function.h @@ -52,6 +52,7 @@ int LAPACKE_dgetri(int matrix_layout, int n, double* a, int lda, #include +#include "paddle/framework/eigen.h" #include "paddle/framework/tensor.h" #include "paddle/platform/device_context.h" #include "paddle/platform/enforce.h" @@ -84,6 +85,13 @@ void matmul(const platform::DeviceContext& context, const framework::Tensor& matrix_b, bool trans_b, T alpha, framework::Tensor* matrix_out, T beta); +template +void SetConstant(const platform::DeviceContext& context, + framework::Tensor* tensor, T num) { + auto t = framework::EigenVector::Flatten(*tensor); + t.device(*context.GetEigenDevice()) = t.constant(static_cast(num)); +} + } // namespace math } // namespace operators } // namespace paddle diff --git a/paddle/operators/math/math_function_test.cc b/paddle/operators/math/math_function_test.cc index f272f7e5135e7092618b8c94ee55faf1cfd8e8a5..9945ba101d719848aa0c06fa65629d59f167c083 100644 --- a/paddle/operators/math/math_function_test.cc +++ b/paddle/operators/math/math_function_test.cc @@ -1,7 +1,7 @@ #include "paddle/operators/math/math_function.h" #include "gtest/gtest.h" -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA TEST(math_function, notrans_mul_trans) { paddle::framework::Tensor input1; paddle::framework::Tensor input1_gpu; @@ -243,3 +243,24 @@ TEST(math_function, gemm_trans_clbas) { EXPECT_EQ(input3_ptr[6], 86); EXPECT_EQ(input3_ptr[7], 99); } + +TEST(math_function, zero) { + paddle::framework::Tensor tensor; + auto* cpu_place = new paddle::platform::CPUPlace(); + float* t = tensor.mutable_data({2, 2}, *cpu_place); + paddle::platform::CPUDeviceContext context(*cpu_place); + paddle::operators::math::SetConstant( + context, &tensor, 0); + EXPECT_EQ(t[0], 0); + EXPECT_EQ(t[1], 0); + EXPECT_EQ(t[2], 0); + EXPECT_EQ(t[3], 0); + + paddle::operators::math::SetConstant( + context, &tensor, 1); + + EXPECT_EQ(t[0], 1); + EXPECT_EQ(t[1], 1); + EXPECT_EQ(t[2], 1); + EXPECT_EQ(t[3], 1); +} diff --git a/paddle/operators/math/pooling.cc b/paddle/operators/math/pooling.cc new file mode 100644 index 0000000000000000000000000000000000000000..50cfb88bb5700dda3785e63e0ccc6457cc928da0 --- /dev/null +++ b/paddle/operators/math/pooling.cc @@ -0,0 +1,740 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/operators/math/pooling.h" + +namespace paddle { +namespace operators { +namespace math { + +/* + * All tensors are in NCHW format. + * Ksize, strides, paddings are two elements. These two elements represent + * height and width, respectively. + */ +template +class Pool2dFunctor { + public: + void operator()(const platform::DeviceContext& context, + const framework::Tensor& input, framework::Tensor& output, + std::vector& ksize, std::vector& strides, + std::vector& paddings, PoolProcess pool_process) { + const int batch_size = input.dims()[0]; + const int input_height = input.dims()[2]; + const int input_width = input.dims()[3]; + const int output_channels = output.dims()[1]; + const int output_height = output.dims()[2]; + const int output_width = output.dims()[3]; + const int ksize_height = ksize[0]; + const int ksize_width = ksize[1]; + const int stride_height = strides[0]; + const int stride_width = strides[1]; + const int padding_height = paddings[0]; + const int padding_width = paddings[1]; + + const int input_stride = input_height * input_width; + const int output_stride = output_height * output_width; + + const T* input_data = input.data(); + T* output_data = output.mutable_data(context.GetPlace()); + + for (int i = 0; i < batch_size; i++) { + for (int c = 0; c < output_channels; ++c) { + for (int ph = 0; ph < output_height; ++ph) { + int hstart = ph * stride_height - padding_height; + int hend = std::min(hstart + ksize_height, input_height); + hstart = std::max(hstart, 0); + for (int pw = 0; pw < output_width; ++pw) { + int wstart = pw * stride_width - padding_width; + int wend = std::min(wstart + ksize_width, input_width); + wstart = std::max(wstart, 0); + + T ele = pool_process.initial(); + for (int h = hstart; h < hend; ++h) { + for (int w = wstart; w < wend; ++w) { + pool_process.compute(ele, input_data[h * input_width + w]); + } + } + int pool_size = (hend - hstart) * (wend - wstart); + pool_process.finalize(ele, (static_cast(pool_size))); + output_data[ph * output_width + pw] = ele; + } + } + input_data += input_stride; + output_data += output_stride; + } + } + } +}; + +/* +* All tensors are in NCHW format. +* Ksize, strides, paddings are two elements. These two elements represent height +* and width, respectively. +*/ +template +class Pool2dGradFunctor { + public: + void operator()(const platform::DeviceContext& context, + const framework::Tensor& input, framework::Tensor& input_grad, + const framework::Tensor& output, + const framework::Tensor& output_grad, std::vector& ksize, + std::vector& strides, std::vector& paddings, + PoolProcess pool_grad_process) { + const int batch_size = input.dims()[0]; + const int input_height = input.dims()[2]; + const int input_width = input.dims()[3]; + const int output_channels = output.dims()[1]; + const int output_height = output.dims()[2]; + const int output_width = output.dims()[3]; + const int ksize_height = ksize[0]; + const int ksize_width = ksize[1]; + const int stride_height = strides[0]; + const int stride_width = strides[1]; + const int padding_height = paddings[0]; + const int padding_width = paddings[1]; + const int input_stride = input_height * input_width; + const int output_stride = output_height * output_width; + + const T* input_data = input.data(); + const T* output_data = output.data(); + const T* output_grad_data = output_grad.data(); + T* input_grad_data = input_grad.mutable_data(context.GetPlace()); + + for (int i = 0; i < batch_size; i++) { + for (int c = 0; c < output_channels; ++c) { + for (int ph = 0; ph < output_height; ++ph) { + int hstart = ph * stride_height - padding_height; + int hend = std::min(hstart + ksize_height, input_height); + hstart = std::max(hstart, 0); + for (int pw = 0; pw < output_width; ++pw) { + int wstart = pw * stride_width - padding_width; + int wend = std::min(wstart + ksize_width, input_width); + wstart = std::max(wstart, 0); + int pool_size = (hend - hstart) * (wend - wstart); + float scale = 1.0 / pool_size; + for (int h = hstart; h < hend; ++h) { + for (int w = wstart; w < wend; ++w) { + pool_grad_process.compute( + input_data[h * input_width + w], + output_data[ph * output_width + pw], + output_grad_data[ph * output_width + pw], + input_grad_data[h * input_width + w], + static_cast(scale)); + } + } + } + } + input_data += input_stride; + output_data += output_stride; + input_grad_data += input_stride; + output_grad_data += output_stride; + } + } + } +}; + +/* + * All tensors are in NCHW format. + * Ksize, strides, paddings are two elements. These two elements represent + * height and width, respectively. + */ +template +class MaxPool2dGradFunctor { + public: + void operator()(const platform::DeviceContext& context, + const framework::Tensor& input, framework::Tensor& input_grad, + const framework::Tensor& output, + const framework::Tensor& output_grad, std::vector& ksize, + std::vector& strides, std::vector& paddings) { + const int batch_size = input.dims()[0]; + const int input_height = input.dims()[2]; + const int input_width = input.dims()[3]; + const int output_channels = output.dims()[1]; + const int output_height = output.dims()[2]; + const int output_width = output.dims()[3]; + const int ksize_height = ksize[0]; + const int ksize_width = ksize[1]; + const int stride_height = strides[0]; + const int stride_width = strides[1]; + const int padding_height = paddings[0]; + const int padding_width = paddings[1]; + const int input_stride = input_height * input_width; + const int output_stride = output_height * output_width; + + const T* input_data = input.data(); + const T* output_data = output.data(); + const T* output_grad_data = output_grad.data(); + T* input_grad_data = input_grad.mutable_data(context.GetPlace()); + + for (int i = 0; i < batch_size; i++) { + for (int c = 0; c < output_channels; ++c) { + for (int ph = 0; ph < output_height; ++ph) { + int hstart = ph * stride_height - padding_height; + int hend = std::min(hstart + ksize_height, input_height); + hstart = std::max(hstart, 0); + for (int pw = 0; pw < output_width; ++pw) { + int wstart = pw * stride_width - padding_width; + int wend = std::min(wstart + ksize_width, input_width); + wstart = std::max(wstart, 0); + + bool stop = false; + for (int h = hstart; h < hend && !stop; ++h) { + for (int w = wstart; w < wend && !stop; ++w) { + int input_idx = h * input_width + w; + int output_idx = ph * output_width + pw; + if (input_data[input_idx] == output_data[output_idx]) { + input_grad_data[input_idx] += output_grad_data[output_idx]; + stop = true; + } + } + } + } + } + input_data += input_stride; + output_data += output_stride; + input_grad_data += input_stride; + output_grad_data += output_stride; + } + } + } +}; + +template class MaxPool2dGradFunctor; +template class MaxPool2dGradFunctor; + +template class Pool2dFunctor, float>; +template class Pool2dFunctor, float>; +template class Pool2dGradFunctor< + platform::CPUPlace, paddle::operators::math::MaxPoolGrad, float>; +template class Pool2dGradFunctor< + platform::CPUPlace, paddle::operators::math::AvgPoolGrad, float>; +template class Pool2dFunctor, double>; +template class Pool2dFunctor, double>; +template class Pool2dGradFunctor< + platform::CPUPlace, paddle::operators::math::MaxPoolGrad, double>; +template class Pool2dGradFunctor< + platform::CPUPlace, paddle::operators::math::AvgPoolGrad, double>; + +/* + * All tensors are in NCDHW format. + * Ksize, strides, paddings are three elements. These three elements represent + * depth, height and width, respectively. + */ +template +class Pool3dFunctor { + public: + void operator()(const platform::DeviceContext& context, + const framework::Tensor& input, framework::Tensor& output, + std::vector& ksize, std::vector& strides, + std::vector& paddings, PoolProcess pool_process) { + const int batch_size = input.dims()[0]; + const int input_depth = input.dims()[2]; + const int input_height = input.dims()[3]; + const int input_width = input.dims()[4]; + const int output_channels = output.dims()[1]; + const int output_depth = output.dims()[2]; + const int output_height = output.dims()[3]; + const int output_width = output.dims()[4]; + const int ksize_depth = ksize[0]; + const int ksize_height = ksize[1]; + const int ksize_width = ksize[2]; + const int stride_depth = strides[0]; + const int stride_height = strides[1]; + const int stride_width = strides[2]; + const int padding_depth = paddings[0]; + const int padding_height = paddings[1]; + const int padding_width = paddings[2]; + + const int input_stride = input_depth * input_height * input_width; + const int output_stride = output_depth * output_height * output_width; + + const T* input_data = input.data(); + T* output_data = output.mutable_data(context.GetPlace()); + + for (int i = 0; i < batch_size; i++) { + for (int c = 0; c < output_channels; ++c) { + for (int pd = 0; pd < output_depth; ++pd) { + int dstart = pd * stride_depth - padding_depth; + int dend = std::min(dstart + ksize_depth, input_depth); + dstart = std::max(dstart, 0); + for (int ph = 0; ph < output_height; ++ph) { + int hstart = ph * stride_height - padding_height; + int hend = std::min(hstart + ksize_height, input_height); + hstart = std::max(hstart, 0); + for (int pw = 0; pw < output_width; ++pw) { + int wstart = pw * stride_width - padding_width; + int wend = std::min(wstart + ksize_width, input_width); + wstart = std::max(wstart, 0); + int output_idx = (pd * output_height + ph) * output_width + pw; + T ele = pool_process.initial(); + for (int d = dstart; d < dend; ++d) { + for (int h = hstart; h < hend; ++h) { + for (int w = wstart; w < wend; ++w) { + pool_process.compute( + ele, + input_data[(d * input_height + h) * input_width + w]); + } + } + } + int pool_size = + (dend - dstart) * (hend - hstart) * (wend - wstart); + pool_process.finalize(ele, static_cast(pool_size)); + output_data[output_idx] = ele; + } + } + } + input_data += input_stride; + output_data += output_stride; + } + } + } +}; + +/* + * All tensors are in NCDHW format. + * Ksize, strides, paddings are three elements. These three elements represent + * depth, height and width, respectively. + */ +template +class Pool3dGradFunctor { + public: + void operator()(const platform::DeviceContext& context, + const framework::Tensor& input, framework::Tensor& input_grad, + const framework::Tensor& output, + const framework::Tensor& output_grad, std::vector& ksize, + std::vector& strides, std::vector& paddings, + PoolProcess pool_grad_process) { + const int batch_size = input.dims()[0]; + const int input_depth = input.dims()[2]; + const int input_height = input.dims()[3]; + const int input_width = input.dims()[4]; + const int output_channels = output.dims()[1]; + const int output_depth = output.dims()[2]; + const int output_height = output.dims()[3]; + const int output_width = output.dims()[4]; + const int ksize_depth = ksize[0]; + const int ksize_height = ksize[1]; + const int ksize_width = ksize[2]; + const int stride_depth = strides[0]; + const int stride_height = strides[1]; + const int stride_width = strides[2]; + const int padding_depth = paddings[0]; + const int padding_height = paddings[1]; + const int padding_width = paddings[2]; + const int input_stride = input_depth * input_height * input_width; + const int output_stride = output_depth * output_height * output_width; + + const T* input_data = input.data(); + const T* output_data = output.data(); + const T* output_grad_data = output_grad.data(); + T* input_grad_data = input_grad.mutable_data(context.GetPlace()); + + for (int i = 0; i < batch_size; i++) { + for (int c = 0; c < output_channels; ++c) { + for (int pd = 0; pd < output_depth; ++pd) { + int dstart = pd * stride_depth - padding_depth; + int dend = std::min(dstart + ksize_depth, input_depth); + dstart = std::max(dstart, 0); + for (int ph = 0; ph < output_height; ++ph) { + int hstart = ph * stride_height - padding_height; + int hend = std::min(hstart + ksize_height, input_height); + hstart = std::max(hstart, 0); + + for (int pw = 0; pw < output_width; ++pw) { + int wstart = pw * stride_width - padding_width; + int wend = std::min(wstart + ksize_width, input_width); + wstart = std::max(wstart, 0); + + int pool_size = + (dend - dstart) * (hend - hstart) * (wend - wstart); + float scale = 1.0 / pool_size; + for (int d = dstart; d < dend; ++d) { + for (int h = hstart; h < hend; ++h) { + for (int w = wstart; w < wend; ++w) { + int input_idx = (d * input_height + h) * input_width + w; + int output_idx = + (pd * output_height + ph) * output_width + pw; + pool_grad_process.compute( + input_data[input_idx], output_data[output_idx], + output_grad_data[output_idx], + input_grad_data[input_idx], static_cast(scale)); + } + } + } + } + } + } + input_data += input_stride; + output_data += output_stride; + input_grad_data += input_stride; + output_grad_data += output_stride; + } + } + } +}; + +/* + * All tensors are in NCDHW format. + * Ksize, strides, paddings are three elements. These three elements represent + * depth, height and width, respectively. + */ +template +class MaxPool3dGradFunctor { + public: + void operator()(const platform::DeviceContext& context, + const framework::Tensor& input, framework::Tensor& input_grad, + const framework::Tensor& output, + const framework::Tensor& output_grad, std::vector& ksize, + std::vector& strides, std::vector& paddings) { + const int batch_size = input.dims()[0]; + const int input_depth = input.dims()[2]; + const int input_height = input.dims()[3]; + const int input_width = input.dims()[4]; + const int output_channels = output.dims()[1]; + const int output_depth = output.dims()[2]; + const int output_height = output.dims()[3]; + const int output_width = output.dims()[4]; + const int ksize_depth = ksize[0]; + const int ksize_height = ksize[1]; + const int ksize_width = ksize[2]; + const int stride_depth = strides[0]; + const int stride_height = strides[1]; + const int stride_width = strides[2]; + const int padding_depth = paddings[0]; + const int padding_height = paddings[1]; + const int padding_width = paddings[2]; + const int input_stride = input_depth * input_height * input_width; + const int output_stride = output_depth * output_height * output_width; + + const T* input_data = input.data(); + const T* output_data = output.data(); + const T* output_grad_data = output_grad.data(); + T* input_grad_data = input_grad.mutable_data(context.GetPlace()); + + for (int i = 0; i < batch_size; i++) { + for (int c = 0; c < output_channels; ++c) { + for (int pd = 0; pd < output_depth; ++pd) { + int dstart = pd * stride_depth - padding_depth; + int dend = std::min(dstart + ksize_depth, input_depth); + dstart = std::max(dstart, 0); + for (int ph = 0; ph < output_height; ++ph) { + int hstart = ph * stride_height - padding_height; + int hend = std::min(hstart + ksize_height, input_height); + hstart = std::max(hstart, 0); + for (int pw = 0; pw < output_width; ++pw) { + int wstart = pw * stride_width - padding_width; + int wend = std::min(wstart + ksize_width, input_width); + wstart = std::max(wstart, 0); + bool stop = false; + for (int d = dstart; d < dend && !stop; ++d) { + for (int h = hstart; h < hend && !stop; ++h) { + for (int w = wstart; w < wend && !stop; ++w) { + int input_idx = (d * input_height + h) * input_width + w; + int output_idx = + (pd * output_height + ph) * output_width + pw; + + if (input_data[input_idx] == output_data[output_idx]) { + input_grad_data[input_idx] += + output_grad_data[output_idx]; + stop = true; + } + } + } + } + } + } + } + input_data += input_stride; + output_data += output_stride; + input_grad_data += input_stride; + output_grad_data += output_stride; + } + } + } +}; + +template class MaxPool3dGradFunctor; +template class MaxPool3dGradFunctor; + +template class Pool3dFunctor, float>; +template class Pool3dFunctor, float>; +template class Pool3dGradFunctor< + platform::CPUPlace, paddle::operators::math::MaxPoolGrad, float>; +template class Pool3dGradFunctor< + platform::CPUPlace, paddle::operators::math::AvgPoolGrad, float>; +template class Pool3dFunctor, double>; +template class Pool3dFunctor, double>; +template class Pool3dGradFunctor< + platform::CPUPlace, paddle::operators::math::MaxPoolGrad, double>; +template class Pool3dGradFunctor< + platform::CPUPlace, paddle::operators::math::AvgPoolGrad, double>; + +/* + * All tensors are in NCHW format. + * Ksize, strides, paddings are two elements. These two elements represent + * height and width, respectively. + */ +template +class MaxPool2dWithIndexFunctor { + public: + void operator()(const platform::DeviceContext& context, + const framework::Tensor& input, framework::Tensor& output, + framework::Tensor& mask, std::vector& ksize, + std::vector& strides, std::vector& paddings) { + const int batch_size = input.dims()[0]; + const int input_height = input.dims()[2]; + const int input_width = input.dims()[3]; + const int output_channels = output.dims()[1]; + const int output_height = output.dims()[2]; + const int output_width = output.dims()[3]; + const int ksize_height = ksize[0]; + const int ksize_width = ksize[1]; + const int stride_height = strides[0]; + const int stride_width = strides[1]; + const int padding_height = paddings[0]; + const int padding_width = paddings[1]; + const int input_stride = input_height * input_width; + const int output_stride = output_height * output_width; + + const T* input_data = input.data(); + T* output_data = output.mutable_data(context.GetPlace()); + T* mask_data = mask.mutable_data(context.GetPlace()); + + for (int i = 0; i < batch_size; i++) { + for (int c = 0; c < output_channels; ++c) { + for (int ph = 0; ph < output_height; ++ph) { + int hstart = ph * stride_height - padding_height; + int hend = std::min(hstart + ksize_height, input_height); + hstart = std::max(hstart, 0); + for (int pw = 0; pw < output_width; ++pw) { + int wstart = pw * stride_width - padding_width; + int wend = std::min(wstart + ksize_width, input_width); + wstart = std::max(wstart, 0); + + T ele = static_cast(-FLT_MAX); + int index = -1; + for (int h = hstart; h < hend; ++h) { + for (int w = wstart; w < wend; ++w) { + if (ele < input_data[h * input_width + w]) { + ele = input_data[h * input_width + w]; + index = h * input_width + w; + } + } + } + output_data[ph * output_width + pw] = ele; + mask_data[ph * output_width + pw] = index; + } + } + // offset + input_data += input_stride; + output_data += output_stride; + mask_data += output_stride; + } + } + } +}; + +/* + * All tensors are in NCHW format. + * Ksize, strides, paddings are two elements. These two elements represent + * height and width, respectively. + */ +template +class MaxPool2dWithIndexGradFunctor { + public: + void operator()(const platform::DeviceContext& context, + framework::Tensor& input_grad, + const framework::Tensor& output_grad, + const framework::Tensor& mask, std::vector& ksize, + std::vector& strides, std::vector& paddings) { + const int batch_size = input_grad.dims()[0]; + const int input_height = input_grad.dims()[2]; + const int input_width = input_grad.dims()[3]; + const int output_channels = output_grad.dims()[1]; + const int output_height = output_grad.dims()[2]; + const int output_width = output_grad.dims()[3]; + const int input_stride = input_height * input_width; + const int output_stride = output_height * output_width; + + const T* mask_data = mask.data(); + const T* output_grad_data = output_grad.data(); + T* input_grad_data = input_grad.mutable_data(context.GetPlace()); + + for (int n = 0; n < batch_size; ++n) { + for (int c = 0; c < output_channels; ++c) { + for (int ph = 0; ph < output_height; ++ph) { + for (int pw = 0; pw < output_width; ++pw) { + const int output_idx = ph * output_width + pw; + const int input_idx = static_cast(mask_data[output_idx]); + input_grad_data[input_idx] += output_grad_data[output_idx]; + } + } + // offset + input_grad_data += input_stride; + output_grad_data += output_stride; + mask_data += output_stride; + } + } + } +}; + +template class MaxPool2dWithIndexFunctor; +template class MaxPool2dWithIndexGradFunctor; +template class MaxPool2dWithIndexFunctor; +template class MaxPool2dWithIndexGradFunctor; + +/* + * All tensors are in NCDHW format. + * Ksize, strides, paddings are three elements. These three elements represent + * depth, height and width, respectively. + */ +template +class MaxPool3dWithIndexFunctor { + public: + void operator()(const platform::DeviceContext& context, + const framework::Tensor& input, framework::Tensor& output, + framework::Tensor& mask, std::vector& ksize, + std::vector& strides, std::vector& paddings) { + const int batch_size = input.dims()[0]; + const int input_depth = input.dims()[2]; + const int input_height = input.dims()[3]; + const int input_width = input.dims()[4]; + const int output_channels = output.dims()[1]; + const int output_depth = output.dims()[2]; + const int output_height = output.dims()[3]; + const int output_width = output.dims()[4]; + const int ksize_depth = ksize[0]; + const int ksize_height = ksize[1]; + const int ksize_width = ksize[2]; + const int stride_depth = strides[0]; + const int stride_height = strides[1]; + const int stride_width = strides[2]; + const int padding_depth = paddings[0]; + const int padding_height = paddings[1]; + const int padding_width = paddings[2]; + const int input_stride = input_depth * input_height * input_width; + const int output_stride = output_depth * output_height * output_width; + + const T* input_data = input.data(); + T* output_data = output.mutable_data(context.GetPlace()); + T* mask_data = mask.mutable_data(context.GetPlace()); + + for (int i = 0; i < batch_size; i++) { + for (int c = 0; c < output_channels; ++c) { + for (int pd = 0; pd < output_depth; ++pd) { + int dstart = pd * stride_depth - padding_depth; + int dend = std::min(dstart + ksize_depth, input_depth); + dstart = std::max(dstart, 0); + for (int ph = 0; ph < output_height; ++ph) { + int hstart = ph * stride_height - padding_height; + int hend = std::min(hstart + ksize_height, input_height); + hstart = std::max(hstart, 0); + for (int pw = 0; pw < output_width; ++pw) { + int wstart = pw * stride_width - padding_width; + int wend = std::min(wstart + ksize_width, input_width); + wstart = std::max(wstart, 0); + + int output_idx = (pd * output_height + ph) * output_width + pw; + T ele = static_cast(-FLT_MAX); + int index = -1; + for (int d = dstart; d < dend; ++d) { + for (int h = hstart; h < hend; ++h) { + for (int w = wstart; w < wend; ++w) { + int input_idx = (d * input_height + h) * input_width + w; + if (ele < input_data[input_idx]) { + index = input_idx; + ele = input_data[input_idx]; + } + } + } + } + output_data[output_idx] = ele; + mask_data[output_idx] = index; + } + } + } + // offset + input_data += input_stride; + output_data += output_stride; + mask_data += output_stride; + } + } + } +}; + +/* + * All tensors are in NCDHW format. + * Ksize, strides, paddings are three elements. These three elements represent + * depth, height and width, respectively. + */ +template +class MaxPool3dWithIndexGradFunctor { + public: + void operator()(const platform::DeviceContext& context, + framework::Tensor& input_grad, + const framework::Tensor& output_grad, + const framework::Tensor& mask, std::vector& ksize, + std::vector& strides, std::vector& paddings) { + const int batch_size = input_grad.dims()[0]; + const int input_depth = input_grad.dims()[2]; + const int input_height = input_grad.dims()[3]; + const int input_width = input_grad.dims()[4]; + const int output_channels = output_grad.dims()[1]; + const int output_depth = output_grad.dims()[2]; + const int output_height = output_grad.dims()[3]; + const int output_width = output_grad.dims()[4]; + const int input_stride = input_depth * input_height * input_width; + const int output_stride = output_depth * output_height * output_width; + + const T* mask_data = mask.data(); + const T* output_grad_data = output_grad.data(); + T* input_grad_data = input_grad.mutable_data(context.GetPlace()); + + for (int n = 0; n < batch_size; ++n) { + for (int c = 0; c < output_channels; ++c) { + for (int pd = 0; pd < output_depth; ++pd) { + for (int ph = 0; ph < output_height; ++ph) { + for (int pw = 0; pw < output_width; ++pw) { + const int output_idx = + (pd * output_height + ph) * output_width + pw; + const int input_idx = static_cast(mask_data[output_idx]); + input_grad_data[input_idx] += output_grad_data[output_idx]; + } + } + } + // offset + input_grad_data += input_stride; + output_grad_data += output_stride; + mask_data += output_stride; + } + } + } +}; + +template class MaxPool3dWithIndexFunctor; +template class MaxPool3dWithIndexGradFunctor; +template class MaxPool3dWithIndexFunctor; +template class MaxPool3dWithIndexGradFunctor; +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/math/pooling.cu b/paddle/operators/math/pooling.cu new file mode 100644 index 0000000000000000000000000000000000000000..736327f4b7b9e9df9ce8f7f60b0437fc1d2d373a --- /dev/null +++ b/paddle/operators/math/pooling.cu @@ -0,0 +1,1059 @@ +/* Copyright (c) 2016 paddlepaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/operators/math/pooling.h" +#include "paddle/platform/cuda_helper.h" + +namespace paddle { +namespace operators { +namespace math { + +template +__global__ void KernelPool2D(const int nthreads, const T* input_data, + T* output_data, const int channels, + const int input_height, const int input_width, + const int output_height, const int output_width, + const int ksize_height, const int ksize_width, + const int stride_height, const int stride_width, + const int padding_height, const int padding_width, + PoolProcess pool_process) { + for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < nthreads; + index += blockDim.x * gridDim.x) { + int pw = index % output_width; + int ph = (index / output_width) % output_height; + int c = (index / output_width / output_height) % channels; + int batch_idx = index / output_width / output_height / channels; + + int hstart = ph * stride_height - padding_height; + int hend = min(hstart + ksize_height, input_height); + hstart = max(hstart, 0); + + int wstart = pw * stride_width - padding_width; + int wend = min(wstart + ksize_width, input_width); + wstart = max(wstart, 0); + + input_data += (batch_idx * channels + c) * input_height * input_width; + T ele = pool_process.initial(); + for (int h = hstart; h < hend; ++h) { + for (int w = wstart; w < wend; ++w) { + pool_process.compute(ele, input_data[h * input_width + w]); + } + } + int pool_size = (hend - hstart) * (wend - wstart); + pool_process.finalize(ele, (static_cast(pool_size))); + output_data[index] = ele; + } +} + +template +__global__ void KernelPool2DGrad( + const int nthreads, const T* input_data, const T* output_data, + const T* output_grad, T* input_grad, const int channels, + const int input_height, const int input_width, const int output_height, + const int output_width, const int ksize_height, const int ksize_width, + const int stride_height, const int stride_width, const int padding_height, + const int padding_width, PoolProcess pool_process) { + for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < nthreads; + index += blockDim.x * gridDim.x) { + int offsetW = index % input_width + padding_width; + int offsetH = (index / input_width) % input_height + padding_height; + int offsetC = (index / input_width / input_height) % channels; + int batch_idx = index / input_width / input_height / channels; + + int phstart = (offsetH < ksize_height) + ? 0 + : (offsetH - ksize_height) / stride_height + 1; + int pwstart = (offsetW < ksize_width) + ? 0 + : (offsetW - ksize_width) / stride_width + 1; + int phend = min(offsetH / stride_height + 1, output_height); + int pwend = min(offsetW / stride_width + 1, output_width); + T gradient = 0; + T input = input_data[index]; + int output_idx = + (batch_idx * channels + offsetC) * output_height * output_width; + output_data += output_idx; + output_grad += output_idx; + for (int ph = phstart; ph < phend; ++ph) { + for (int pw = pwstart; pw < pwend; ++pw) { + int hstart = ph * stride_height - padding_height; + int wstart = pw * stride_width - padding_width; + int hend = min(hstart + ksize_height, input_height); + int wend = min(wstart + ksize_width, input_width); + hstart = max(hstart, 0); + wstart = max(wstart, 0); + int pool_size = (hend - hstart) * (wend - wstart); + int output_sub_idx = ph * output_width + pw; + pool_process.compute(input, output_data[output_sub_idx], + output_grad[output_sub_idx], gradient, + static_cast(1.0 / pool_size)); + } + } + input_grad[index] = gradient; + } +} + +template +__global__ void KernelMaxPool2DGrad( + const int nthreads, const T* input_data, const T* output_data, + const T* output_grad, T* input_grad, const int channels, + const int input_height, const int input_width, const int output_height, + const int output_width, const int ksize_height, const int ksize_width, + const int stride_height, const int stride_width, const int padding_height, + const int padding_width) { + for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < nthreads; + index += blockDim.x * gridDim.x) { + int pw = index % output_width; + int ph = (index / output_width) % output_height; + int c = (index / output_width / output_height) % channels; + int batch_idx = index / output_width / output_height / channels; + + int hstart = ph * stride_height - padding_height; + int hend = min(hstart + ksize_height, input_height); + hstart = max(hstart, 0); + + int wstart = pw * stride_width - padding_width; + int wend = min(wstart + ksize_width, input_width); + wstart = max(wstart, 0); + + input_data += (batch_idx * channels + c) * input_height * input_width; + input_grad += (batch_idx * channels + c) * input_height * input_width; + + T ele = output_data[index]; + int maxIndex = -1; + bool stop = false; + for (int h = hstart; h < hend && !stop; ++h) { + for (int w = wstart; w < wend && !stop; ++w) { + if (ele == input_data[h * input_width + w]) { + maxIndex = h * input_width + w; + stop = true; + } + } + } + + if (maxIndex != -1) { + // atomic add + platform::CudaAtomicAdd(input_grad + maxIndex, output_grad[index]); + } + } +} + +/* + * All tensors are in NCHW format. + * Ksize, strides, paddings are two elements. These two elements represent + * height and width, respectively. + */ +template +class Pool2dFunctor { + public: + void operator()(const platform::DeviceContext& context, + const framework::Tensor& input, framework::Tensor& output, + std::vector& ksize, std::vector& strides, + std::vector& paddings, PoolProcess pool_process) { + const int batch_size = input.dims()[0]; + const int input_channels = input.dims()[1]; + const int input_height = input.dims()[2]; + const int input_width = input.dims()[3]; + const int output_channels = output.dims()[1]; + const int output_height = output.dims()[2]; + const int output_width = output.dims()[3]; + const int ksize_height = ksize[0]; + const int ksize_width = ksize[1]; + const int stride_height = strides[0]; + const int stride_width = strides[1]; + const int padding_height = paddings[0]; + const int padding_width = paddings[1]; + + const T* input_data = input.data(); + T* output_data = output.mutable_data(context.GetPlace()); + + int nthreads = batch_size * output_channels * output_height * output_width; + int blocks = (nthreads + 1024 - 1) / 1024; + dim3 threads(1024, 1); + dim3 grid(blocks, 1); + + KernelPool2D< + PoolProcess, + T><<(context) + .stream()>>>(nthreads, input_data, output_data, input_channels, + input_height, input_width, output_height, + output_width, ksize_height, ksize_width, + stride_height, stride_width, padding_height, + padding_width, pool_process); + } +}; + +/* + * All tensors are in NCHW format. + * Ksize, strides, paddings are two elements. These two elements represent + * height and width, respectively. + */ +template +class Pool2dGradFunctor { + public: + void operator()(const platform::DeviceContext& context, + const framework::Tensor& input, framework::Tensor& input_grad, + const framework::Tensor& output, + const framework::Tensor& output_grad, std::vector& ksize, + std::vector& strides, std::vector& paddings, + PoolProcess pool_process) { + const int batch_size = input.dims()[0]; + const int input_channels = input.dims()[1]; + const int input_height = input.dims()[2]; + const int input_width = input.dims()[3]; + const int output_height = output.dims()[2]; + const int output_width = output.dims()[3]; + const int ksize_height = ksize[0]; + const int ksize_width = ksize[1]; + const int stride_height = strides[0]; + const int stride_width = strides[1]; + const int padding_height = paddings[0]; + const int padding_width = paddings[1]; + + const T* input_data = input.data(); + const T* output_data = output.data(); + const T* output_grad_data = output_grad.data(); + T* input_grad_data = input_grad.mutable_data(context.GetPlace()); + + int nthreads = batch_size * input_channels * input_height * input_width; + int blocks = (nthreads + 1024 - 1) / 1024; + dim3 threads(1024, 1); + dim3 grid(blocks, 1); + + KernelPool2DGrad< + PoolProcess, + T><<(context) + .stream()>>>( + nthreads, input_data, output_data, output_grad_data, input_grad_data, + input_channels, input_height, input_width, output_height, output_width, + ksize_height, ksize_width, stride_height, stride_width, padding_height, + padding_width, pool_process); + } +}; + +/* + * All tensors are in NCHW format. + * Ksize, strides, paddings are two elements. These two elements represent + * height and width, respectively. + */ +template +class MaxPool2dGradFunctor { + public: + void operator()(const platform::DeviceContext& context, + const framework::Tensor& input, framework::Tensor& input_grad, + const framework::Tensor& output, + const framework::Tensor& output_grad, std::vector& ksize, + std::vector& strides, std::vector& paddings) { + const int batch_size = input.dims()[0]; + const int input_channels = input.dims()[1]; + const int input_height = input.dims()[2]; + const int input_width = input.dims()[3]; + const int output_channels = output.dims()[1]; + const int output_height = output.dims()[2]; + const int output_width = output.dims()[3]; + const int ksize_height = ksize[0]; + const int ksize_width = ksize[1]; + const int stride_height = strides[0]; + const int stride_width = strides[1]; + const int padding_height = paddings[0]; + const int padding_width = paddings[1]; + + const T* input_data = input.data(); + const T* output_data = output.data(); + const T* output_grad_data = output_grad.data(); + T* input_grad_data = input_grad.mutable_data(context.GetPlace()); + + int nthreads = batch_size * output_channels * output_height * output_width; + int blocks = (nthreads + 1024 - 1) / 1024; + dim3 threads(1024, 1); + dim3 grid(blocks, 1); + + KernelMaxPool2DGrad< + T><<(context) + .stream()>>>( + nthreads, input_data, output_data, output_grad_data, input_grad_data, + input_channels, input_height, input_width, output_height, output_width, + ksize_height, ksize_width, stride_height, stride_width, padding_height, + padding_width); + } +}; + +template class MaxPool2dGradFunctor; +template class MaxPool2dGradFunctor; + +template class Pool2dFunctor, float>; +template class Pool2dFunctor, float>; +template class Pool2dGradFunctor< + platform::GPUPlace, paddle::operators::math::MaxPoolGrad, float>; +template class Pool2dGradFunctor< + platform::GPUPlace, paddle::operators::math::AvgPoolGrad, float>; +template class Pool2dFunctor, double>; +template class Pool2dFunctor, double>; +template class Pool2dGradFunctor< + platform::GPUPlace, paddle::operators::math::MaxPoolGrad, double>; +template class Pool2dGradFunctor< + platform::GPUPlace, paddle::operators::math::AvgPoolGrad, double>; + +template +__global__ void KernelPool3D( + const int nthreads, const T* input_data, T* output_data, const int channels, + const int input_depth, const int input_height, const int input_width, + const int output_depth, const int output_height, const int output_width, + const int ksize_depth, const int ksize_height, const int ksize_width, + const int stride_depth, const int stride_height, const int stride_width, + const int padding_depth, const int padding_height, const int padding_width, + PoolProcess pool_process) { + for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < nthreads; + index += blockDim.x * gridDim.x) { + int pw = index % output_width; + int ph = (index / output_width) % output_height; + int pd = (index / output_width / output_height) % output_depth; + int c = (index / output_width / output_height / output_depth) % channels; + int batch_idx = + index / output_width / output_height / output_depth / channels; + int dstart = pd * stride_depth - padding_depth; + int hstart = ph * stride_height - padding_height; + int wstart = pw * stride_width - padding_width; + int dend = min(dstart + ksize_depth, input_depth); + int hend = min(hstart + ksize_height, input_height); + int wend = min(wstart + ksize_width, input_width); + dstart = max(dstart, 0); + hstart = max(hstart, 0); + wstart = max(wstart, 0); + T ele = pool_process.initial(); + input_data += + (batch_idx * channels + c) * input_depth * input_height * input_width; + for (int d = dstart; d < dend; ++d) { + for (int h = hstart; h < hend; ++h) { + for (int w = wstart; w < wend; ++w) { + pool_process.compute( + ele, input_data[(d * input_height + h) * input_width + w]); + } + } + } + int pool_size = (dend - dstart) * (hend - hstart) * (wend - wstart); + pool_process.finalize(ele, static_cast(pool_size)); + output_data[index] = ele; + } +} + +template +__global__ void KernelPool3DGrad( + const int nthreads, const T* input_data, const T* output_data, + const T* output_grad, T* input_grad, const int channels, + const int input_depth, const int input_height, const int input_width, + const int output_depth, const int output_height, const int output_width, + const int ksize_depth, const int ksize_height, const int ksize_width, + const int stride_depth, const int stride_height, const int stride_width, + const int padding_depth, const int padding_height, const int padding_width, + PoolProcess pool_process) { + for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < nthreads; + index += blockDim.x * gridDim.x) { + int offsetW = index % input_width + padding_width; + int offsetH = (index / input_width) % input_height + padding_height; + int offsetD = + (index / input_width / input_height) % input_depth + padding_depth; + int offsetC = (index / input_width / input_height / input_depth) % channels; + int batch_idx = index / input_width / input_height / input_depth / channels; + + int pdstart = (offsetD < ksize_depth) + ? 0 + : (offsetD - ksize_depth) / stride_depth + 1; + int phstart = (offsetH < ksize_height) + ? 0 + : (offsetH - ksize_height) / stride_height + 1; + int pwstart = (offsetW < ksize_width) + ? 0 + : (offsetW - ksize_width) / stride_width + 1; + int pdend = min((offsetD) / stride_depth + 1, output_depth); + int phend = min((offsetH) / stride_height + 1, output_height); + int pwend = min((offsetW) / stride_width + 1, output_width); + + T gradient = 0; + T input = input_data[index]; + int output_idx = (batch_idx * channels + offsetC) * output_depth * + output_height * output_width; + output_data += output_idx; + output_grad += output_idx; + + for (int pd = pdstart; pd < pdend; ++pd) { + for (int ph = phstart; ph < phend; ++ph) { + for (int pw = pwstart; pw < pwend; ++pw) { + // figure out the pooling size + int dstart = pd * stride_depth - padding_depth; + int hstart = ph * stride_height - padding_height; + int wstart = pw * stride_width - padding_width; + int dend = min(dstart + ksize_depth, input_depth); + int hend = min(hstart + ksize_height, input_height); + int wend = min(wstart + ksize_width, input_width); + dstart = max(dstart, 0); + hstart = max(hstart, 0); + wstart = max(wstart, 0); + int pool_size = (dend - dstart) * (hend - hstart) * (wend - wstart); + int output_sub_idx = (pd * output_height + ph) * output_width + pw; + pool_process.compute(input, output_data[output_sub_idx], + output_grad[output_sub_idx], gradient, + static_cast(1.0 / pool_size)); + } + } + } + input_grad[index] = gradient; + } +} + +template +__global__ void KernelMaxPool3DGrad( + const int nthreads, const T* input_data, const T* output_data, + const T* output_grad, T* input_grad, const int channels, + const int input_depth, const int input_height, const int input_width, + const int output_depth, const int output_height, const int output_width, + const int ksize_depth, const int ksize_height, const int ksize_width, + const int stride_depth, const int stride_height, const int stride_width, + const int padding_depth, const int padding_height, + const int padding_width) { + for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < nthreads; + index += blockDim.x * gridDim.x) { + int pw = index % output_width; + int ph = (index / output_width) % output_height; + int pd = (index / output_width / output_height) % output_depth; + int c = (index / output_width / output_height / output_depth) % channels; + int batch_idx = + index / output_width / output_height / output_depth / channels; + int dstart = pd * stride_depth - padding_depth; + int hstart = ph * stride_height - padding_height; + int wstart = pw * stride_width - padding_width; + int dend = min(dstart + ksize_depth, input_depth); + int hend = min(hstart + ksize_height, input_height); + int wend = min(wstart + ksize_width, input_width); + dstart = max(dstart, 0); + hstart = max(hstart, 0); + wstart = max(wstart, 0); + T ele = output_data[index]; + bool stop = false; + int maxIdx = -1; + input_data += + (batch_idx * channels + c) * input_depth * input_height * input_width; + input_grad += + (batch_idx * channels + c) * input_depth * input_height * input_width; + + for (int d = dstart; d < dend && !stop; ++d) { + for (int h = hstart; h < hend && !stop; ++h) { + for (int w = wstart; w < wend && !stop; ++w) { + if (ele == input_data[(d * input_height + h) * input_width + w]) { + stop = true; + maxIdx = (d * input_height + h) * input_width + w; + } + } + } + } + if (maxIdx != -1) { + // atomic add + platform::CudaAtomicAdd(input_grad + maxIdx, output_grad[index]); + } + } +} + +/* + * All tensors are in NCDHW format. + * Ksize, strides, paddings are three elements. These three elements represent + * depth, height and width, respectively. + */ +template +class Pool3dFunctor { + public: + void operator()(const platform::DeviceContext& context, + const framework::Tensor& input, framework::Tensor& output, + std::vector& ksize, std::vector& strides, + std::vector& paddings, PoolProcess pool_process) { + const int batch_size = input.dims()[0]; + const int input_channels = input.dims()[1]; + const int input_depth = input.dims()[2]; + const int input_height = input.dims()[3]; + const int input_width = input.dims()[4]; + const int output_channels = output.dims()[1]; + const int output_depth = output.dims()[2]; + const int output_height = output.dims()[3]; + const int output_width = output.dims()[4]; + const int ksize_depth = ksize[0]; + const int ksize_height = ksize[1]; + const int ksize_width = ksize[2]; + const int stride_depth = strides[0]; + const int stride_height = strides[1]; + const int stride_width = strides[2]; + const int padding_depth = paddings[0]; + const int padding_height = paddings[1]; + const int padding_width = paddings[2]; + + const T* input_data = input.data(); + T* output_data = output.mutable_data(context.GetPlace()); + + int nthreads = batch_size * output_channels * output_depth * output_height * + output_width; + int blocks = (nthreads + 1024 - 1) / 1024; + dim3 threads(1024, 1); + dim3 grid(blocks, 1); + + KernelPool3D< + PoolProcess, + T><<(context) + .stream()>>>( + nthreads, input_data, output_data, input_channels, input_depth, + input_height, input_width, output_depth, output_height, output_width, + ksize_depth, ksize_height, ksize_width, stride_depth, stride_height, + stride_width, padding_depth, padding_height, padding_width, + pool_process); + } +}; + +/* + * All tensors are in NCDHW format. + * Ksize, strides, paddings are three elements. These three elements represent + * depth, height and width, respectively. + */ +template +class Pool3dGradFunctor { + public: + void operator()(const platform::DeviceContext& context, + const framework::Tensor& input, framework::Tensor& input_grad, + const framework::Tensor& output, + const framework::Tensor& output_grad, std::vector& ksize, + std::vector& strides, std::vector& paddings, + PoolProcess pool_process) { + const int batch_size = input.dims()[0]; + const int input_channels = input.dims()[1]; + const int input_depth = input.dims()[2]; + const int input_height = input.dims()[3]; + const int input_width = input.dims()[4]; + const int output_channels = output.dims()[1]; + const int output_depth = output.dims()[2]; + const int output_height = output.dims()[3]; + const int output_width = output.dims()[4]; + const int ksize_depth = ksize[0]; + const int ksize_height = ksize[1]; + const int ksize_width = ksize[2]; + const int stride_depth = strides[0]; + const int stride_height = strides[1]; + const int stride_width = strides[2]; + const int padding_depth = paddings[0]; + const int padding_height = paddings[1]; + const int padding_width = paddings[2]; + + const T* input_data = input.data(); + const T* output_data = output.data(); + const T* output_grad_data = output_grad.data(); + T* input_grad_data = input_grad.mutable_data(context.GetPlace()); + + int nthreads = + batch_size * input_channels * input_depth * input_height * input_width; + int blocks = (nthreads + 1024 - 1) / 1024; + dim3 threads(1024, 1); + dim3 grid(blocks, 1); + + KernelPool3DGrad< + PoolProcess, + T><<(context) + .stream()>>>( + nthreads, input_data, output_data, output_grad_data, input_grad_data, + input_channels, input_depth, input_height, input_width, output_depth, + output_height, output_width, ksize_depth, ksize_height, ksize_width, + stride_depth, stride_height, stride_width, padding_depth, + padding_height, padding_width, pool_process); + } +}; + +/* + * All tensors are in NCDHW format. + * Ksize, strides, paddings are three elements. These three elements represent + * depth, height and width, respectively. + */ +template +class MaxPool3dGradFunctor { + public: + void operator()(const platform::DeviceContext& context, + const framework::Tensor& input, framework::Tensor& input_grad, + const framework::Tensor& output, + const framework::Tensor& output_grad, std::vector& ksize, + std::vector& strides, std::vector& paddings) { + const int batch_size = input.dims()[0]; + const int input_channels = input.dims()[1]; + const int input_depth = input.dims()[2]; + const int input_height = input.dims()[3]; + const int input_width = input.dims()[4]; + const int output_channels = output.dims()[1]; + const int output_depth = output.dims()[2]; + const int output_height = output.dims()[3]; + const int output_width = output.dims()[4]; + const int ksize_depth = ksize[0]; + const int ksize_height = ksize[1]; + const int ksize_width = ksize[2]; + const int stride_depth = strides[0]; + const int stride_height = strides[1]; + const int stride_width = strides[2]; + const int padding_depth = paddings[0]; + const int padding_height = paddings[1]; + const int padding_width = paddings[2]; + + const T* input_data = input.data(); + const T* output_data = output.data(); + const T* output_grad_data = output_grad.data(); + T* input_grad_data = input_grad.mutable_data(context.GetPlace()); + + int nthreads = batch_size * output_channels * output_depth * output_height * + output_width; + int blocks = (nthreads + 1024 - 1) / 1024; + dim3 threads(1024, 1); + dim3 grid(blocks, 1); + + KernelMaxPool3DGrad< + T><<(context) + .stream()>>>( + nthreads, input_data, output_data, output_grad_data, input_grad_data, + input_channels, input_depth, input_height, input_width, output_depth, + output_height, output_width, ksize_depth, ksize_height, ksize_width, + stride_depth, stride_height, stride_width, padding_depth, + padding_height, padding_width); + } +}; + +template class MaxPool3dGradFunctor; +template class MaxPool3dGradFunctor; + +template class Pool3dFunctor, float>; +template class Pool3dFunctor, float>; +template class Pool3dGradFunctor< + platform::GPUPlace, paddle::operators::math::MaxPoolGrad, float>; +template class Pool3dGradFunctor< + platform::GPUPlace, paddle::operators::math::AvgPoolGrad, float>; +template class Pool3dFunctor, double>; +template class Pool3dFunctor, double>; +template class Pool3dGradFunctor< + platform::GPUPlace, paddle::operators::math::MaxPoolGrad, double>; +template class Pool3dGradFunctor< + platform::GPUPlace, paddle::operators::math::AvgPoolGrad, double>; + +template +__global__ void KernelMaxPool2dWithIdx( + const int nthreads, const T* input_data, T* output_data, T* mask_data, + const int channels, const int input_height, const int input_width, + const int output_height, const int output_width, const int ksize_height, + const int ksize_width, const int stride_height, const int stride_width, + const int padding_height, const int padding_width) { + for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < nthreads; + index += blockDim.x * gridDim.x) { + int pw = index % output_width; + int ph = (index / output_width) % output_height; + int c = (index / output_width / output_height) % channels; + int batch_idx = index / output_width / output_height / channels; + + int hstart = ph * stride_height - padding_height; + int hend = min(hstart + ksize_height, input_height); + hstart = max(hstart, 0); + + int wstart = pw * stride_width - padding_width; + int wend = min(wstart + ksize_width, input_width); + wstart = max(wstart, 0); + + input_data += (batch_idx * channels + c) * input_height * input_width; + T ele = -FLT_MAX; + int max_index = -1; + for (int h = hstart; h < hend; ++h) { + for (int w = wstart; w < wend; ++w) { + int input_index = h * input_width + w; + if (ele < input_data[input_index]) { + max_index = input_index; + ele = input_data[input_index]; + } + } + } + output_data[index] = ele; + mask_data[index] = max_index; + } +} + +template +__global__ void KernelMaxPool2DWithIdxGrad( + const int nthreads, T* input_grad, const T* output_grad, const T* mask_data, + const int channels, const int input_height, const int input_width, + const int output_height, const int output_width, const int ksize_height, + const int ksize_width, const int stride_height, const int stride_width, + const int padding_height, const int padding_width) { + for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < nthreads; + index += blockDim.x * gridDim.x) { + int w_offset = index % input_width; + int h_offset = (index / input_width) % input_height; + int c_offset = (index / input_width / input_height) % channels; + int batch_idx = index / input_width / input_height / channels; + + int ph_start = + (h_offset + padding_height < ksize_height) + ? 0 + : (h_offset + padding_height - ksize_height) / stride_height + 1; + int pw_start = + (w_offset + padding_width < ksize_width) + ? 0 + : (w_offset + padding_width - ksize_width) / stride_width + 1; + int ph_end = + min((h_offset + padding_height) / stride_height + 1, output_height); + int pw_end = + min((w_offset + padding_width) / stride_width + 1, output_width); + + T gradient = 0; + int input_current_featuremap_idx = h_offset * input_width + w_offset; + int output_idx = + (batch_idx * channels + c_offset) * output_height * output_width; + + mask_data += output_idx; + output_grad += output_idx; + for (int ph = ph_start; ph < ph_end; ++ph) { + for (int pw = pw_start; pw < pw_end; ++pw) { + if (mask_data[ph * output_width + pw] == input_current_featuremap_idx) + gradient += output_grad[ph * output_width + pw]; + } + } + input_grad[index] = gradient; + } +} + +/* + * All tensors are in NCHW format. + * Ksize, strides, paddings are two elements. These two elements represent + * height and width, respectively. + */ +template +class MaxPool2dWithIndexFunctor { + public: + void operator()(const platform::DeviceContext& context, + const framework::Tensor& input, framework::Tensor& output, + framework::Tensor& mask, std::vector& ksize, + std::vector& strides, std::vector& paddings) { + const int batch_size = input.dims()[0]; + const int input_channels = input.dims()[1]; + const int input_height = input.dims()[2]; + const int input_width = input.dims()[3]; + const int output_channels = output.dims()[1]; + const int output_height = output.dims()[2]; + const int output_width = output.dims()[3]; + const int ksize_height = ksize[0]; + const int ksize_width = ksize[1]; + const int stride_height = strides[0]; + const int stride_width = strides[1]; + const int padding_height = paddings[0]; + const int padding_width = paddings[1]; + + const T* input_data = input.data(); + T* output_data = output.mutable_data(context.GetPlace()); + T* mask_data = mask.mutable_data(context.GetPlace()); + + int nthreads = batch_size * output_channels * output_height * output_width; + int blocks = (nthreads + 1024 - 1) / 1024; + dim3 threads(1024, 1); + dim3 grid(blocks, 1); + + KernelMaxPool2dWithIdx< + T><<(context) + .stream()>>>(nthreads, input_data, output_data, mask_data, + input_channels, input_height, input_width, + output_height, output_width, ksize_height, + ksize_width, stride_height, stride_width, + padding_height, padding_width); + } +}; + +/* + * All tensors are in NCHW format. + * Ksize, strides, paddings are two elements. These two elements represent + * height and width, respectively. + */ +template +class MaxPool2dWithIndexGradFunctor { + public: + void operator()(const platform::DeviceContext& context, + framework::Tensor& input_grad, + const framework::Tensor& output_grad, + const framework::Tensor& mask, std::vector& ksize, + std::vector& strides, std::vector& paddings) { + const int batch_size = input_grad.dims()[0]; + const int input_channels = input_grad.dims()[1]; + const int input_height = input_grad.dims()[2]; + const int input_width = input_grad.dims()[3]; + const int output_height = output_grad.dims()[2]; + const int output_width = output_grad.dims()[3]; + const int ksize_height = ksize[0]; + const int ksize_width = ksize[1]; + const int stride_height = strides[0]; + const int stride_width = strides[1]; + const int padding_height = paddings[0]; + const int padding_width = paddings[1]; + + const T* mask_data = mask.data(); + const T* output_grad_data = output_grad.data(); + T* input_grad_data = input_grad.mutable_data(context.GetPlace()); + + int nthreads = batch_size * input_channels * input_height * input_width; + int blocks = (nthreads + 1024 - 1) / 1024; + dim3 threads(1024, 1); + dim3 grid(blocks, 1); + + KernelMaxPool2DWithIdxGrad< + T><<(context) + .stream()>>>(nthreads, input_grad_data, output_grad_data, + mask_data, input_channels, input_height, + input_width, output_height, output_width, + ksize_height, ksize_width, stride_height, + stride_width, padding_height, padding_width); + } +}; + +template class MaxPool2dWithIndexFunctor; +template class MaxPool2dWithIndexGradFunctor; +template class MaxPool2dWithIndexFunctor; +template class MaxPool2dWithIndexGradFunctor; + +template +__global__ void KernelMaxPool3DWithIdx( + const int nthreads, const T* input_data, T* output_data, T* mask_data, + const int channels, const int input_depth, const int input_height, + const int input_width, const int output_depth, const int output_height, + const int output_width, const int ksize_depth, const int ksize_height, + const int ksize_width, const int stride_depth, const int stride_height, + const int stride_width, const int padding_depth, const int padding_height, + const int padding_width) { + for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < nthreads; + index += blockDim.x * gridDim.x) { + int pw = index % output_width; + int ph = (index / output_width) % output_height; + int pd = (index / output_width / output_height) % output_depth; + int c = (index / output_width / output_height / output_depth) % channels; + int batch_idx = + index / output_width / output_height / output_depth / channels; + + int dstart = pd * stride_depth - padding_depth; + int hstart = ph * stride_height - padding_height; + int wstart = pw * stride_width - padding_width; + int dend = min(dstart + ksize_depth, input_depth); + int hend = min(hstart + ksize_height, input_height); + int wend = min(wstart + ksize_width, input_width); + dstart = max(dstart, 0); + hstart = max(hstart, 0); + wstart = max(wstart, 0); + + T ele = -FLT_MAX; + int max_index = -1; + input_data += + (batch_idx * channels + c) * input_depth * input_height * input_width; + + for (int d = dstart; d < dend; ++d) { + for (int h = hstart; h < hend; ++h) { + for (int w = wstart; w < wend; ++w) { + if (ele < input_data[(d * input_height + h) * input_width + w]) { + max_index = (d * input_height + h) * input_width + w; + ele = input_data[max_index]; + } + } + } + } + output_data[index] = ele; + mask_data[index] = max_index; + } +} + +template +__global__ void KernelMaxPool3DWithIdxGrad( + const int nthreads, T* input_grad, const T* output_grad, const T* mask, + const int channels, const int input_depth, const int input_height, + const int input_width, const int output_depth, const int output_height, + const int output_width, const int ksize_depth, const int ksize_height, + const int ksize_width, const int stride_depth, const int stride_height, + const int stride_width, const int padding_depth, const int padding_height, + const int padding_width) { + for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < nthreads; + index += blockDim.x * gridDim.x) { + int w_offset = index % input_width; + int h_offset = (index / input_width) % input_height; + int d_offset = (index / input_width / input_height) % input_depth; + int c_offset = + (index / input_width / input_height / input_depth) % channels; + int batch_idx = index / input_width / input_height / input_depth / channels; + + int pd_start = + (d_offset + padding_depth < ksize_depth) + ? 0 + : (d_offset + padding_depth - ksize_depth) / stride_depth + 1; + int ph_start = + (h_offset + padding_height < ksize_height) + ? 0 + : (h_offset + padding_height - ksize_height) / stride_height + 1; + int pw_start = + (w_offset + padding_width < ksize_width) + ? 0 + : (w_offset + padding_width - ksize_width) / stride_width + 1; + int pd_end = + min((d_offset + padding_depth) / stride_depth + 1, output_depth); + int ph_end = + min((h_offset + padding_height) / stride_height + 1, output_height); + int pw_end = + min((w_offset + padding_width) / stride_width + 1, output_width); + + T gradient = 0; + int input_current_feature_map_idx = + (d_offset * input_height + h_offset) * input_width + w_offset; + int output_idx = (batch_idx * channels + c_offset) * output_depth * + output_height * output_width; + mask += output_idx; + output_grad += output_idx; + + for (int pd = pd_start; pd < pd_end; ++pd) { + for (int ph = ph_start; ph < ph_end; ++ph) { + for (int pw = pw_start; pw < pw_end; ++pw) { + if (mask[(pd * output_height + ph) * output_width + pw] == + input_current_feature_map_idx) + gradient += + output_grad[(pd * output_height + ph) * output_width + pw]; + } + } + } + input_grad[index] = gradient; + } +} + +/* + * All tensors are in NCDHW format. + * Ksize, strides, paddings are three elements. These three elements represent + * depth, height and width, respectively. + */ +template +class MaxPool3dWithIndexFunctor { + public: + void operator()(const platform::DeviceContext& context, + const framework::Tensor& input, framework::Tensor& output, + framework::Tensor& mask, std::vector& ksize, + std::vector& strides, std::vector& paddings) { + const int batch_size = input.dims()[0]; + const int input_channels = input.dims()[1]; + const int input_depth = input.dims()[2]; + const int input_height = input.dims()[3]; + const int input_width = input.dims()[4]; + const int output_channels = output.dims()[1]; + const int output_depth = output.dims()[2]; + const int output_height = output.dims()[3]; + const int output_width = output.dims()[4]; + const int ksize_depth = ksize[0]; + const int ksize_height = ksize[1]; + const int ksize_width = ksize[2]; + const int stride_depth = strides[0]; + const int stride_height = strides[1]; + const int stride_width = strides[2]; + const int padding_depth = paddings[0]; + const int padding_height = paddings[1]; + const int padding_width = paddings[2]; + + const T* input_data = input.data(); + T* output_data = output.mutable_data(context.GetPlace()); + T* mask_data = mask.mutable_data(context.GetPlace()); + + int nthreads = batch_size * output_channels * output_depth * output_height * + output_width; + int blocks = (nthreads + 1024 - 1) / 1024; + dim3 threads(1024, 1); + dim3 grid(blocks, 1); + + KernelMaxPool3DWithIdx< + T><<(context) + .stream()>>>( + nthreads, input_data, output_data, mask_data, input_channels, + input_depth, input_height, input_width, output_depth, output_height, + output_width, ksize_depth, ksize_height, ksize_width, stride_depth, + stride_height, stride_width, padding_depth, padding_height, + padding_width); + } +}; + +/* + * All tensors are in NCDHW format. + * Ksize, strides, paddings are three elements. These three elements represent + * depth, height and width, respectively. + */ +template +class MaxPool3dWithIndexGradFunctor { + public: + void operator()(const platform::DeviceContext& context, + framework::Tensor& input_grad, + const framework::Tensor& output_grad, + const framework::Tensor& mask, std::vector& ksize, + std::vector& strides, std::vector& paddings) { + const int batch_size = input_grad.dims()[0]; + const int input_channels = input_grad.dims()[1]; + const int input_depth = input_grad.dims()[2]; + const int input_height = input_grad.dims()[3]; + const int input_width = input_grad.dims()[4]; + const int output_depth = output_grad.dims()[2]; + const int output_height = output_grad.dims()[3]; + const int output_width = output_grad.dims()[4]; + const int ksize_depth = ksize[0]; + const int ksize_height = ksize[1]; + const int ksize_width = ksize[2]; + const int stride_depth = strides[0]; + const int stride_height = strides[1]; + const int stride_width = strides[2]; + const int padding_depth = paddings[0]; + const int padding_height = paddings[1]; + const int padding_width = paddings[2]; + + const T* output_grad_data = output_grad.data(); + const T* mask_data = mask.data(); + T* input_grad_data = input_grad.mutable_data(context.GetPlace()); + + int nthreads = + batch_size * input_channels * input_depth * input_height * input_width; + int blocks = (nthreads + 1024 - 1) / 1024; + dim3 threads(1024, 1); + dim3 grid(blocks, 1); + + KernelMaxPool3DWithIdxGrad< + T><<(context) + .stream()>>>( + nthreads, input_grad_data, output_grad_data, mask_data, input_channels, + input_depth, input_height, input_width, output_depth, output_height, + output_width, ksize_depth, ksize_height, ksize_width, stride_depth, + stride_height, stride_width, padding_depth, padding_height, + padding_width); + } +}; + +template class MaxPool3dWithIndexFunctor; +template class MaxPool3dWithIndexGradFunctor; +template class MaxPool3dWithIndexFunctor; +template class MaxPool3dWithIndexGradFunctor; + +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/math/pooling.h b/paddle/operators/math/pooling.h new file mode 100644 index 0000000000000000000000000000000000000000..c50c57b5c52cdc5c12425cb119b80502aef5451e --- /dev/null +++ b/paddle/operators/math/pooling.h @@ -0,0 +1,194 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include "paddle/framework/eigen.h" +#include "paddle/framework/tensor.h" +#include "paddle/platform/device_context.h" +#include "paddle/platform/hostdevice.h" + +namespace paddle { +namespace operators { +namespace math { + +#define FLT_MAX \ + __FLT_MAX__ // It might need to be placed in another file, but I'm still + // wondering where to put it. + +/* + * \brief Extracting simple operations from pooling. + * Both MaxPool and AvgPool need "initial", "compute" and "finalize" + * operation. + * MaxPool initializes temp variable to the negative maximum to find the + * maximum value in the pooling field. + * AvgPool initializes temp variable to the zero to accumulate all values + * in pool pooling, and finally takes the average. + * MaxPoolGrad and AvgPoolGrad are gradient operations respectively. + */ +template +class MaxPool { + public: + DEVICE inline T initial() { return static_cast(-FLT_MAX); } + DEVICE inline void compute(T& y, const T& x) { y = y > x ? y : x; } + DEVICE inline void finalize(T& y, const T& pool_field) {} +}; + +template +class AvgPool { + public: + DEVICE inline T initial() { return static_cast(0); } + DEVICE inline void compute(T& y, const T& x) { y += x; } + DEVICE inline void finalize(T& y, const T& pool_field) { y /= pool_field; } +}; + +template +class MaxPoolGrad { + public: + DEVICE inline void compute(const T& x, const T& y, const T& dy, T& dx, + T scale) { + dx += dy * (x == y); + } +}; + +template +class AvgPoolGrad { + public: + DEVICE inline void compute(const T& x, const T& y, const T& dy, T& dx, + T scale) { + dx += (scale * dy); + } +}; + +/* + * \brief Getting pooling results, and calculating gradient. + * + * In pool2d, all tensors are in NCHW format. Where N is batch size, C is the + * number of channels, H and W is the height and width of feature. + * In pool3d, all tensors are in NCDHW format. Where N is batch size, C is the + * number of channels, D, H and W is the depth, height and width of feature. + * + * In max pooling, it is possible that the pooling region has multiple maximum + * elements. In this case, we should compute the gradient of the first maximum + * element. + * This is different from average pooling. So we rewrite the max_pool_grad: + * MaxPool2dGradFunctor, MaxPool3dGradFunctor. + */ +template +class Pool2dFunctor { + public: + void operator()(const platform::DeviceContext& context, + const framework::Tensor& input, framework::Tensor& output, + std::vector& ksize, std::vector& strides, + std::vector& paddings, PoolProcess pool_compute); +}; + +template +class Pool2dGradFunctor { + public: + void operator()(const platform::DeviceContext& context, + const framework::Tensor& input, framework::Tensor& input_grad, + const framework::Tensor& output, + const framework::Tensor& output_grad, std::vector& ksize, + std::vector& strides, std::vector& paddings, + PoolProcess pool_compute); +}; + +template +class MaxPool2dGradFunctor { + public: + void operator()(const platform::DeviceContext& context, + const framework::Tensor& input, framework::Tensor& input_grad, + const framework::Tensor& output, + const framework::Tensor& output_grad, std::vector& ksize, + std::vector& strides, std::vector& paddings); +}; + +template +class Pool3dFunctor { + public: + void operator()(const platform::DeviceContext& context, + const framework::Tensor& input, framework::Tensor& output, + std::vector& ksize, std::vector& strides, + std::vector& paddings, PoolProcess pool_compute); +}; + +template +class Pool3dGradFunctor { + public: + void operator()(const platform::DeviceContext& context, + const framework::Tensor& input, framework::Tensor& input_grad, + const framework::Tensor& output, + const framework::Tensor& output_grad, std::vector& ksize, + std::vector& strides, std::vector& paddings, + PoolProcess pool_compute); +}; + +template +class MaxPool3dGradFunctor { + public: + void operator()(const platform::DeviceContext& context, + const framework::Tensor& input, framework::Tensor& input_grad, + const framework::Tensor& output, + const framework::Tensor& output_grad, std::vector& ksize, + std::vector& strides, std::vector& paddings); +}; + +/* + * \brief Getting max pooling results and corresponding max index, and + * calculating gradient. + * In up-sampling-pooling, it is necessary to know max element index. + * In pool2d, all tensors are in NCHW format. In pool3d, all tensors are in + * NCDHW format. + */ +template +class MaxPool2dWithIndexFunctor { + public: + void operator()(const platform::DeviceContext& context, + const framework::Tensor& input, framework::Tensor& output, + framework::Tensor& mask, std::vector& ksize, + std::vector& strides, std::vector& paddings); +}; + +template +class MaxPool2dWithIndexGradFunctor { + public: + void operator()(const platform::DeviceContext& context, + framework::Tensor& input_grad, + const framework::Tensor& output_grad, + const framework::Tensor& mask, std::vector& ksize, + std::vector& strides, std::vector& paddings); +}; + +template +class MaxPool3dWithIndexFunctor { + public: + void operator()(const platform::DeviceContext& context, + const framework::Tensor& input, framework::Tensor& output, + framework::Tensor& mask, std::vector& ksize, + std::vector& strides, std::vector& paddings); +}; + +template +class MaxPool3dWithIndexGradFunctor { + public: + void operator()(const platform::DeviceContext& context, + framework::Tensor& input_grad, + const framework::Tensor& output_grad, + const framework::Tensor& mask, std::vector& ksize, + std::vector& strides, std::vector& paddings); +}; + +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/framework/grad_op_builder.h b/paddle/operators/math/softmax.cc similarity index 71% rename from paddle/framework/grad_op_builder.h rename to paddle/operators/math/softmax.cc index 998f8ebbb5f2f4fb8b7e938b5916afd0f8a7930d..0ba8197ab8b64649c8adcf67771ba01eca7f1d10 100644 --- a/paddle/framework/grad_op_builder.h +++ b/paddle/operators/math/softmax.cc @@ -12,14 +12,15 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#pragma once - -#include "paddle/framework/operator.h" +#include "paddle/operators/math/softmax.h" namespace paddle { -namespace framework { +namespace operators { +namespace math { -OperatorBase* BuildGradOp(const OperatorBase* op); +template class SoftmaxFunctor; +template class SoftmaxGradFunctor; -} // namespace framework +} // namespace math +} // namespace operators } // namespace paddle diff --git a/paddle/operators/math/softmax.cu b/paddle/operators/math/softmax.cu new file mode 100644 index 0000000000000000000000000000000000000000..99f988d51e4b16c3f3bfd9c76b411bb53619603e --- /dev/null +++ b/paddle/operators/math/softmax.cu @@ -0,0 +1,28 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#define EIGEN_USE_GPU + +#include "paddle/operators/math/softmax.h" + +namespace paddle { +namespace operators { +namespace math { + +template class SoftmaxFunctor; +template class SoftmaxGradFunctor; + +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/math/softmax.h b/paddle/operators/math/softmax.h new file mode 100644 index 0000000000000000000000000000000000000000..b7f627eee7f8fe68a83595a3390a55d438c97afb --- /dev/null +++ b/paddle/operators/math/softmax.h @@ -0,0 +1,104 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include "paddle/framework/eigen.h" +#include "paddle/framework/operator.h" +#include "paddle/framework/tensor.h" + +namespace paddle { +namespace operators { +namespace math { + +template +using EigenMatrix = framework::EigenMatrix; + +template +struct ValueClip { + HOSTDEVICE T operator()(const T& x) const { + const T kThreshold = -64.; + return x < kThreshold ? kThreshold : x; + } +}; + +template +class SoftmaxFunctor { + public: + void operator()(const platform::DeviceContext& context, + const framework::Tensor* X, framework::Tensor* Y) { + auto logits = EigenMatrix::From(*X); + auto softmax = EigenMatrix::From(*Y); + + const int kBatchDim = 0; + const int kClassDim = 1; + + const int batch_size = logits.dimension(kBatchDim); + const int num_classes = logits.dimension(kClassDim); + + Eigen::DSizes along_class(kClassDim); + Eigen::DSizes batch_by_one(batch_size, 1); + Eigen::DSizes one_by_class(1, num_classes); + + auto shifted_logits = (logits - + logits.maximum(along_class) + .eval() + .reshape(batch_by_one) + .broadcast(one_by_class)) + .unaryExpr(ValueClip()); + + softmax.device(*context.GetEigenDevice()) = shifted_logits.exp(); + softmax.device(*context.GetEigenDevice()) = + (softmax * + softmax.sum(along_class) + .inverse() + .eval() + .reshape(batch_by_one) + .broadcast(one_by_class)); + } +}; + +template +class SoftmaxGradFunctor { + public: + void operator()(const platform::DeviceContext& context, + const framework::Tensor* y, const framework::Tensor* y_grad, + framework::Tensor* x_grad) { + auto softmax = EigenMatrix::From(*y); + auto softmax_grad = EigenMatrix::From(*y_grad); + auto logits_grad = EigenMatrix::From(*x_grad); + + const int kBatchDim = 0; + const int kClassDim = 1; + + const int batch_size = softmax.dimension(kBatchDim); + const int num_classes = softmax.dimension(kClassDim); + + Eigen::DSizes along_class(kClassDim); + Eigen::DSizes batch_by_one(batch_size, 1); + Eigen::DSizes one_by_class(1, num_classes); + + auto dot = (softmax * softmax_grad) + .sum(along_class) + .eval() + .reshape(batch_by_one) + .broadcast(one_by_class); + logits_grad.device(*context.GetEigenDevice()) = + (softmax_grad - dot) * softmax; + } +}; + +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/mean_op.cc b/paddle/operators/mean_op.cc index b04384bda81b93f5db0be3206eee10ad5e854540..441543049faf96c8fbc14e84448a348c3287c1e2 100644 --- a/paddle/operators/mean_op.cc +++ b/paddle/operators/mean_op.cc @@ -22,21 +22,21 @@ class MeanOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), - "Input(X) of MeanOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"), - "Output(Out) of MeanOp should not be null."); - ctx.Output("Out")->Resize({1}); + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of MeanOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of MeanOp should not be null."); + ctx->SetOutputDim("Out", {1}); } }; class MeanOpMaker : public framework::OpProtoAndCheckerMaker { public: - MeanOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + MeanOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "The input of mean op"); - AddOutput("Out", "The output of mean op").NotInGradient(); + AddOutput("Out", "The output of mean op"); AddComment(R"DOC( Mean Operator )DOC"); } @@ -47,9 +47,23 @@ class MeanGradOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - ctx.Output(framework::GradVarName("X")) - ->Resize(ctx.Input("X")->dims()); + void InferShape(framework::InferShapeContext* ctx) const override { + ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X")); + } +}; + +class MeanGradMaker : public framework::SingleGradOpDescMaker { + public: + using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; + + protected: + std::unique_ptr Apply() const override { + auto* grad_op = new framework::OpDescBind(); + grad_op->SetType("mean_grad"); + grad_op->SetInput("X", Input("X")); + grad_op->SetInput(framework::GradVarName("Out"), OutputGrad("Out")); + grad_op->SetOutput(framework::GradVarName("X"), InputGrad("X")); + return std::unique_ptr(grad_op); } }; @@ -57,7 +71,8 @@ class MeanGradOp : public framework::OperatorWithKernel { } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP(mean, ops::MeanOp, ops::MeanOpMaker, mean_grad, ops::MeanGradOp); +REGISTER_OPERATOR(mean, ops::MeanOp, ops::MeanOpMaker, ops::MeanGradMaker); +REGISTER_OPERATOR(mean_grad, ops::MeanGradOp); REGISTER_OP_CPU_KERNEL(mean, ops::MeanKernel); REGISTER_OP_CPU_KERNEL(mean_grad, diff --git a/paddle/operators/mean_op.h b/paddle/operators/mean_op.h index ce31e178d8e375dc59be80a6c05133201308da70..c99286a5b928f1edcd845b01b21b95654c25db07 100644 --- a/paddle/operators/mean_op.h +++ b/paddle/operators/mean_op.h @@ -28,7 +28,7 @@ template ; template -class MeanKernel : public framework::OpKernel { +class MeanKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto* input = context.Input("X"); @@ -45,7 +45,7 @@ class MeanKernel : public framework::OpKernel { }; template -class MeanGradKernel : public framework::OpKernel { +class MeanGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto OG = context.Input(framework::GradVarName("Out")); diff --git a/paddle/operators/minus_op.cc b/paddle/operators/minus_op.cc index 29cb85489bd05f6c1e7143d962eac0af26e75825..d7fd2f901b766c3105a5ea8847fb8a3fe456a945 100644 --- a/paddle/operators/minus_op.cc +++ b/paddle/operators/minus_op.cc @@ -26,22 +26,22 @@ class MinusOp : public framework::OperatorWithKernel { : OperatorWithKernel(type, inputs, outputs, attrs) {} protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), - "Input(X) of MinusOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), - "Input(Y) of MinusOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"), - "Output(Out) of MinusOp should not be null."); + void InferShape(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of MinusOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Y"), + "Input(Y) of MinusOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of MinusOp should not be null."); - auto *left_tensor = ctx.Input("X"); - auto *right_tensor = ctx.Input("Y"); + auto x_dims = ctx->GetInputDim("X"); + auto y_dims = ctx->GetInputDim("Y"); PADDLE_ENFORCE_EQ( - left_tensor->numel(), right_tensor->numel(), + x_dims, y_dims, "Minus operator must take two tensor with same num of elements"); - ctx.Output("Out")->Resize(left_tensor->dims()); - ctx.ShareLoD("X", /*->*/ "Out"); + ctx->SetOutputDim("Out", x_dims); + ctx->ShareLoD("X", /*->*/ "Out"); } }; @@ -49,9 +49,9 @@ class MinusOpMaker : public framework::OpProtoAndCheckerMaker { public: MinusOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("X", "The left tensor of minus operator.").NotInGradient(); - AddInput("Y", "The right tensor of minus operator.").NotInGradient(); - AddOutput("Out", "The output tensor of minus operator.").NotInGradient(); + AddInput("X", "The left tensor of minus operator."); + AddInput("Y", "The right tensor of minus operator."); + AddOutput("Out", "The output tensor of minus operator."); AddComment(R"DOC(Minus Operator @@ -64,26 +64,35 @@ or not. But the output only shares the LoD with input `X`. )DOC"); } }; -template -class MinusGradOp : public NetOp { + +class MinusGradMaker : public framework::GradOpDescMakerBase { public: - MinusGradOp(const std::string &type, const framework::VariableNameMap &inputs, - const framework::VariableNameMap &outputs, - const framework::AttributeMap &attrs) - : NetOp(type, inputs, outputs, attrs) { - auto out_grad = Input(framework::GradVarName("Out")); - auto x_grad = Output(framework::GradVarName("X")); - auto y_grad = Output(framework::GradVarName("Y")); - - // x_grad = out_grad - AppendOp(framework::OpRegistry::CreateOp("identity", {{"X", {out_grad}}}, - {{"Y", {x_grad}}}, {})); - - framework::AttributeMap scale_attr; - scale_attr["scale"] = static_cast(-1); - AppendOp(framework::OpRegistry::CreateOp("scale", {{"X", {out_grad}}}, - {{"Out", {y_grad}}}, scale_attr)); - CompleteAddOp(false); + using framework::GradOpDescMakerBase::GradOpDescMakerBase; + + std::vector> operator()() + const override { + std::vector> ops; + auto x_g = InputGrad("X"); + if (!x_g.empty()) { + auto *x_g_op = new framework::OpDescBind(); + x_g_op->SetType("scale"); + x_g_op->SetInput("X", OutputGrad("Out")); + x_g_op->SetOutput("Out", x_g); + x_g_op->SetAttr("scale", 1.0f); + ops.emplace_back(x_g_op); + } + + auto y_g = InputGrad("Y"); + if (!y_g.empty()) { + auto *y_g_op = new framework::OpDescBind(); + y_g_op->SetType("scale"); + y_g_op->SetInput("X", OutputGrad("Out")); + y_g_op->SetOutput("Out", y_g); + y_g_op->SetAttr("scale", -1.0f); + ops.emplace_back(y_g_op); + } + + return ops; } }; @@ -91,7 +100,6 @@ class MinusGradOp : public NetOp { } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP(minus, ops::MinusOp, ops::MinusOpMaker, minus_grad, - ops::MinusGradOp); +REGISTER_OPERATOR(minus, ops::MinusOp, ops::MinusOpMaker, ops::MinusGradMaker); REGISTER_OP_CPU_KERNEL(minus, ops::MinusKernel); diff --git a/paddle/operators/minus_op.h b/paddle/operators/minus_op.h index 6310a4fd5141516cff4fc7acbe1d17913a1b5506..bd9a2790aa2b208c2d3dfc792031283eb6c42397 100644 --- a/paddle/operators/minus_op.h +++ b/paddle/operators/minus_op.h @@ -20,7 +20,7 @@ namespace paddle { namespace operators { template -class MinusKernel : public framework::OpKernel { +class MinusKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto* left_tensor = context.Input("X"); diff --git a/paddle/operators/modified_huber_loss_op.cc b/paddle/operators/modified_huber_loss_op.cc index 8606c0d1e1bf7a52299528d30af0367d9f93edd2..6522327fdcf646a60df76909d049341b5a9021c9 100644 --- a/paddle/operators/modified_huber_loss_op.cc +++ b/paddle/operators/modified_huber_loss_op.cc @@ -22,20 +22,19 @@ class ModifiedHuberLossOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext& context) const override { - PADDLE_ENFORCE_NOT_NULL(context.InputVar("X"), "X must be initialized."); - PADDLE_ENFORCE_NOT_NULL(context.InputVar("Y"), "Y must be initialized."); + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "X must be initialized."); + PADDLE_ENFORCE(ctx->HasInput("Y"), "Y must be initialized."); - auto* x = context.Input("X"); - auto* y = context.Input("Y"); + auto x_dims = ctx->GetInputDim("X"); + auto y_dims = ctx->GetInputDim("Y"); - PADDLE_ENFORCE_EQ(x->dims(), y->dims(), - "The shape of X and Y must be the same."); - PADDLE_ENFORCE_EQ(x->dims().size(), 2, "The tensor rank of X must be 2."); - PADDLE_ENFORCE_EQ(x->dims()[1], 1, "The 2nd dimension of X must be 1."); + PADDLE_ENFORCE_EQ(x_dims, y_dims, "The shape of X and Y must be the same."); + PADDLE_ENFORCE_EQ(x_dims.size(), 2, "The tensor rank of X must be 2."); + PADDLE_ENFORCE_EQ(x_dims[1], 1, "The 2nd dimension of X must be 1."); - context.Output("IntermediateVal")->Resize(x->dims()); - context.Output("Out")->Resize({x->dims()[0], 1}); + ctx->SetOutputDim("IntermediateVal", x_dims); + ctx->SetOutputDim("Out", {x_dims[0], 1}); } }; @@ -75,27 +74,28 @@ class ModifiedHuberLossGradOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext& context) const override { - auto* x = context.Input("X"); - auto* y = context.Input("Y"); - auto* intermediate_val = context.Input("IntermediateVal"); - auto* out_grad = context.Input(framework::GradVarName("Out")); - auto* x_grad = - context.Output(framework::GradVarName("X")); - - PADDLE_ENFORCE_NOT_NULL(x, "X must be initialized."); - PADDLE_ENFORCE_NOT_NULL(y, "Y must be initialized."); - PADDLE_ENFORCE_NOT_NULL(intermediate_val, - "Intermediate value must not be null."); - PADDLE_ENFORCE_NOT_NULL(out_grad, "Input(Out@Grad) must not be null."); + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "X must be initialized."); + PADDLE_ENFORCE(ctx->HasInput("Y"), "Y must be initialized."); + PADDLE_ENFORCE(ctx->HasInput("IntermediateVal"), + "Intermediate value must not be null."); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), + "Input(Out@Grad) must not be null."); + + auto x_dims = ctx->GetInputDim("X"); + auto y_dims = ctx->GetInputDim("Y"); + auto intermediate_dims = ctx->GetInputDim("IntermediateVal"); + auto out_grad_dims = ctx->GetInputDim(framework::GradVarName("Out")); PADDLE_ENFORCE_EQ( - intermediate_val->dims(), x->dims(), + intermediate_dims, x_dims, "The shape of X and intermediate value must be the same."); - PADDLE_ENFORCE_EQ(out_grad->dims(), x->dims(), + PADDLE_ENFORCE_EQ(out_grad_dims, x_dims, "The shape of Input(Out@Grad) and X must be the same."); - if (x_grad) x_grad->Resize(x->dims()); + if (ctx->HasOutput(framework::GradVarName("X"))) { + ctx->SetOutputDim(framework::GradVarName("X"), x_dims); + } } }; diff --git a/paddle/operators/modified_huber_loss_op.cu b/paddle/operators/modified_huber_loss_op.cu index bce760f95e72cfec05b07591e0fa1250168b112f..8854e166cd99ce914d7f9f9bcead3234b0649506 100644 --- a/paddle/operators/modified_huber_loss_op.cu +++ b/paddle/operators/modified_huber_loss_op.cu @@ -39,7 +39,7 @@ struct ModifiedHuberLossBackward { }; template -class ModifiedHuberLossGradGPUKernel : public framework::OpKernel { +class ModifiedHuberLossGradGPUKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto* in0 = context.Input("Y"); diff --git a/paddle/operators/modified_huber_loss_op.h b/paddle/operators/modified_huber_loss_op.h index cb51007749e3c59572d4852959f4119ac377decc..aba75efad9c19e3e113b4f09bc1fbd4732f4e187 100644 --- a/paddle/operators/modified_huber_loss_op.h +++ b/paddle/operators/modified_huber_loss_op.h @@ -47,7 +47,7 @@ struct ModifiedHuberLossForward { }; template -class ModifiedHuberLossKernel : public framework::OpKernel { +class ModifiedHuberLossKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto* in0 = context.Input("X"); @@ -73,7 +73,7 @@ class ModifiedHuberLossKernel : public framework::OpKernel { // CPU backward kernel template -class ModifiedHuberLossGradCPUKernel : public framework::OpKernel { +class ModifiedHuberLossGradCPUKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto* in0 = context.Input("Y"); diff --git a/paddle/operators/mul_op.cc b/paddle/operators/mul_op.cc index 7047718a3f1bf7e9598952efa1d9bcb20d5cf5b4..ec0683d8875a945746b676a1f859e614af557441 100644 --- a/paddle/operators/mul_op.cc +++ b/paddle/operators/mul_op.cc @@ -1,16 +1,16 @@ /* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - Licensed under the Apache License, Version 2.0 (the "License"); - you may not use this file except in compliance with the License. - You may obtain a copy of the License at +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at - http://www.apache.org/licenses/LICENSE-2.0 + http://www.apache.org/licenses/LICENSE-2.0 - Unless required by applicable law or agreed to in writing, software - distributed under the License is distributed on an "AS IS" BASIS, - WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - See the License for the specific language governing permissions and - limitations under the License. */ +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ #include "paddle/operators/mul_op.h" @@ -24,27 +24,25 @@ class MulOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), - "Input(X) of MulOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), - "Input(Y) of MulOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"), - "Output(Out) of MulOp should not be null."); - - auto x_dims = ctx.Input("X")->dims(); - auto y_dims = ctx.Input("Y")->dims(); - int x_num_col_dims = Attr("x_num_col_dims"); - int y_num_col_dims = Attr("y_num_col_dims"); - - PADDLE_ENFORCE(x_dims.size() > x_num_col_dims, - "The rank of input tensor X(%s) should be larger than " - "`mul_op`'s `x_num_col_dims`.", - ctx.op().Input("X")); - PADDLE_ENFORCE(y_dims.size() > y_num_col_dims, - "The rank of input tensor Y(%s) should be larger than " - "`mul_op`'s `y_num_col_dims`.", - ctx.op().Input("Y")); + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) of MulOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Y"), "Input(Y) of MulOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of MulOp should not be null."); + + auto x_dims = ctx->GetInputDim("X"); + auto y_dims = ctx->GetInputDim("Y"); + int x_num_col_dims = ctx->Attrs().Get("x_num_col_dims"); + int y_num_col_dims = ctx->Attrs().Get("y_num_col_dims"); + + PADDLE_ENFORCE_GT( + x_dims.size(), x_num_col_dims, + "The input tensor X's rank of MulOp should be larger than " + "x_num_col_dims."); + PADDLE_ENFORCE_GT( + y_dims.size(), y_num_col_dims, + "The input tensor Y's rank of MulOp should be larger than " + "y_num_col_dims."); auto x_mat_dims = framework::flatten_to_2d(x_dims, x_num_col_dims); auto y_mat_dims = framework::flatten_to_2d(y_dims, y_num_col_dims); @@ -52,24 +50,23 @@ class MulOp : public framework::OperatorWithKernel { PADDLE_ENFORCE_EQ( x_mat_dims[1], y_mat_dims[0], "First matrix's width must be equal with second matrix's height."); - ctx.Output("Out")->Resize( - {x_mat_dims[0], y_mat_dims[1]}); - ctx.ShareLoD("X", /*->*/ "Out"); + ctx->SetOutputDim("Out", {x_mat_dims[0], y_mat_dims[1]}); + ctx->ShareLoD("X", /*->*/ "Out"); } }; class MulOpMaker : public framework::OpProtoAndCheckerMaker { public: - MulOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + MulOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "The first input of mul op"); AddInput("Y", "The second input of mul op"); AddOutput("Out", "The output of mul op"); AddAttr( "x_num_col_dims", - R"DOC(mul_op can take tensors with more than two dimensions as input `X`, - in that case, tensors will be reshaped to a matrix. The matrix's first - dimension(column length) will be the product of tensor's last + R"DOC(mul_op can take tensors with more than two dimensions as input `X`, + in that case, tensors will be reshaped to a matrix. The matrix's first + dimension(column length) will be the product of tensor's last `num_col_dims` dimensions, and the matrix's second dimension(row length) will be the product of tensor's first `rank - num_col_dims` dimensions. )DOC") @@ -100,16 +97,14 @@ class MulOpGrad : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) should not be null"); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), "Input(Y) should not be null"); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")), - "Input(Out@GRAD) should not be null"); - auto x_dims = ctx.Input("X")->dims(); - auto y_dims = ctx.Input("Y")->dims(); - auto out_dims = ctx.Input(framework::GradVarName("Out"))->dims(); - auto *x_grad = ctx.Output(framework::GradVarName("X")); - auto *y_grad = ctx.Output(framework::GradVarName("Y")); + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null"); + PADDLE_ENFORCE(ctx->HasInput("Y"), "Input(Y) should not be null"); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), + "Input(Out@GRAD) should not be null"); + auto x_dims = ctx->GetInputDim("X"); + auto y_dims = ctx->GetInputDim("Y"); + auto out_dims = ctx->GetInputDim(framework::GradVarName("Out")); auto x_mat_dims = framework::flatten_to_2d(x_dims, Attr("x_num_col_dims")); @@ -125,8 +120,15 @@ class MulOpGrad : public framework::OperatorWithKernel { "The second dimension of Out@GRAD must equal to the second " "dimension of the second operand."); - if (x_grad) x_grad->Resize(x_dims); - if (y_grad) y_grad->Resize(y_dims); + auto x_grad_name = framework::GradVarName("X"); + auto y_grad_name = framework::GradVarName("Y"); + + if (ctx->HasOutput(x_grad_name)) { + ctx->SetOutputDim(x_grad_name, x_dims); + } + if (ctx->HasOutput(y_grad_name)) { + ctx->SetOutputDim(y_grad_name, y_dims); + } } }; diff --git a/paddle/operators/mul_op.h b/paddle/operators/mul_op.h index ac7136a76933d1f3ead86518c65d589747227631..684b1ea0c0c8ddabc9809cc05ed985e0cc250955 100644 --- a/paddle/operators/mul_op.h +++ b/paddle/operators/mul_op.h @@ -28,7 +28,7 @@ template ; template -class MulKernel : public framework::OpKernel { +class MulKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { const Tensor* x = context.Input("X"); @@ -52,7 +52,7 @@ class MulKernel : public framework::OpKernel { }; template -class MulGradKernel : public framework::OpKernel { +class MulGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { int x_num_col_dims = ctx.template Attr("x_num_col_dims"); diff --git a/paddle/operators/multiplex_op.cc b/paddle/operators/multiplex_op.cc index 6e77b86b5698a263b850a973cd1b8644a0aa2201..a86685b6dde4761cf74f9521bd9609b0864b9bdf 100644 --- a/paddle/operators/multiplex_op.cc +++ b/paddle/operators/multiplex_op.cc @@ -18,61 +18,69 @@ namespace paddle { namespace operators { using Tensor = framework::Tensor; -using LoDTensor = framework::LoDTensor; class MultiplexOp : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE(!ctx.MultiInputVar("X").empty(), - "Input(X) should not be null"); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"), - "Output(Out) shouldn't be null."); - auto ins = ctx.MultiInput("X"); - auto *out = ctx.Output("Out"); - auto num_ins = ins.size(); - PADDLE_ENFORCE(num_ins > 2, - "multiplex operator should have more than 2 inputs."); - PADDLE_ENFORCE_EQ(ins[0]->dims().size(), 1, - "The first input must be a index vector."); - auto in_dim = ins[1]->dims(); - - for (size_t i = 2; i < num_ins; i++) { - auto dim = ins[i]->dims(); - PADDLE_ENFORCE( - in_dim == dim, - "All the input tensors except the first one must have the same size"); + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Ids"), "Input(Ids) shouldn't be null."); + PADDLE_ENFORCE(!ctx->Inputs("X").empty(), + "MultiInput(X) shouldn't be empty."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) shouldn't be null."); + auto ids_dim = ctx->GetInputDim("Ids"); + PADDLE_ENFORCE( + ids_dim.size() == 2 && ids_dim[1] == 1, + "The index tensor must be a vector with size batchSize x 1."); + + auto ins_dims = ctx->GetInputsDim("X"); + auto num_ins = ins_dims.size(); + PADDLE_ENFORCE(num_ins > 1, + "multiplex operator should have more than " + "one candidate input tensors."); + + auto in_dim = ins_dims[0]; + PADDLE_ENFORCE(in_dim.size() >= 2, + "The rank of candidate tensors must be not less than 2."); + for (size_t i = 1; i < num_ins; i++) { + auto dim = ins_dims[i]; + PADDLE_ENFORCE(in_dim == dim, + "All the candidate tensors must have the same size."); } - out->Resize(in_dim); + ctx->SetOutputDim("Out", in_dim); + } + + framework::DataType IndicateDataType( + const framework::ExecutionContext& ctx) const override { + return framework::ToDataType(ctx.MultiInput("X")[0]->type()); } }; class MultiplexOpMaker : public framework::OpProtoAndCheckerMaker { public: - MultiplexOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + MultiplexOpMaker(framework::OpProto* proto, + framework::OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("X", "The input tensors of multiplex operator.").AsDuplicable(); + AddInput("Ids", "The index tensor of multiplex operator."); + AddInput("X", "The candidate tensors of multiplex operator.") + .AsDuplicable(); AddOutput("Out", "The output tensor of multiplex operator."); AddComment(R"DOC(Multiplex operator -Multiplex multiple tensors according to the index provided by the first -input tensor. +Multiplex multiple tensors according to the index provided by the index tensor. -ins[0]: the index tensor. -ins[1:N]: the candidate output tensors. +Ids: the index tensor. +X[0 : N - 1]: the candidate tensors for output (N >= 2). For each index i from 0 to batchSize - 1, the output is the i-th row of the -the (index[i] + 1)-th tensor. +the (Ids[i])-th tensor. For i-th row of the output tensor: -y[i][j] = x_{k}[i][j], j = 0,1, ... , (x_{1}.width - 1) +y[i] = x_{k}[i] where y is the output tensor. `x_{k}` is the k-th input tensor -and `k = x{0}[i] + 1`. - +and `k = Ids[i]`. )DOC"); } }; @@ -82,21 +90,24 @@ class MultiplexGradOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE(!ctx.MultiInputVar("X").empty(), - "Input(X) should not be null"); - PADDLE_ENFORCE(!ctx.MultiOutputVar(framework::GradVarName("X")).empty(), - "Output(X@Grad) should not be null"); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")), - "Input(Out@GRAD) shouldn't be null."); - auto d_ins = ctx.MultiOutput(framework::GradVarName("X")); - auto ins = ctx.MultiInput("X"); - // don't compute gradient for index (ins[0]) - for (size_t i = 1; i < ins.size(); i++) { - if (d_ins[i]) { - d_ins[i]->Resize(ins[i]->dims()); - } + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(!ctx->Inputs("X").empty(), "Input(X) should not be null."); + PADDLE_ENFORCE(!ctx->Outputs(framework::GradVarName("X")).empty(), + "Output(X@Grad) should not be null."); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), + "Input(Out@GRAD) should not be null."); + std::vector d_ins; + auto ins = ctx->GetInputsDim("X"); + // No need to compute gradient for Input(Ids) + for (size_t i = 0; i < ins.size(); i++) { + d_ins.push_back(ins[i]); } + ctx->SetOutputsDim(framework::GradVarName("X"), d_ins); + } + + framework::DataType IndicateDataType( + const framework::ExecutionContext& ctx) const override { + return framework::ToDataType(ctx.MultiInput("X")[0]->type()); } }; diff --git a/paddle/operators/multiplex_op.cu b/paddle/operators/multiplex_op.cu index 4736f15bd594178168e3bcf799142d0fc18bff13..72b1f96eafde37976b4b067b534112b17e02b807 100644 --- a/paddle/operators/multiplex_op.cu +++ b/paddle/operators/multiplex_op.cu @@ -18,28 +18,31 @@ namespace paddle { namespace operators { +using Tensor = framework::Tensor; + template -class MultiplexGPUKernel : public framework::OpKernel { +class MultiplexGPUKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const { - auto ins = ctx.MultiInput("X"); - auto* out = ctx.Output("Out"); - + auto ins = ctx.MultiInput("X"); + auto* ids = ctx.Input("Ids"); + auto* out = ctx.Output("Out"); out->mutable_data(ctx.GetPlace()); - auto rows = ins[1]->dims()[0]; - auto cols = ins[1]->dims()[1]; + auto rows = ins[0]->dims()[0]; + auto cols = ins[0]->numel() / rows; // copy index to cpu - framework::Tensor index_t_cpu; - index_t_cpu.CopyFrom(*(ins[0]), platform::CPUPlace()); - auto* index = index_t_cpu.data(); + Tensor index_t_cpu; + index_t_cpu.CopyFrom(*ids, platform::CPUPlace()); + auto* index = index_t_cpu.data(); auto stream = reinterpret_cast( ctx.device_context()) .stream(); Place place = boost::get(ctx.GetPlace()); for (auto i = 0; i < rows; i++) { - int k = (int)index[i] + 1; - PADDLE_ENFORCE_LT(k, ins.size(), + int32_t k = index[i]; + PADDLE_ENFORCE_GE(k, 0, "index must be nonnegative."); + PADDLE_ENFORCE_LT((size_t)k, ins.size(), "index exceeds the number of candidate tensors."); memory::Copy(place, out->data() + i * cols, place, ins[k]->data() + i * cols, cols * sizeof(T), stream); @@ -48,14 +51,14 @@ class MultiplexGPUKernel : public framework::OpKernel { }; template -class MultiplexGradGPUKernel : public framework::OpKernel { +class MultiplexGradGPUKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const { - auto* d_out = ctx.Input(framework::GradVarName("Out")); - auto ins = ctx.MultiInput("X"); - auto d_ins = - ctx.MultiOutput(framework::GradVarName("X")); - for (size_t i = 1; i < d_ins.size(); i++) { + auto* d_out = ctx.Input(framework::GradVarName("Out")); + auto ins = ctx.MultiInput("X"); + auto* ids = ctx.Input("Ids"); + auto d_ins = ctx.MultiOutput(framework::GradVarName("X")); + for (size_t i = 0; i < d_ins.size(); i++) { if (d_ins[i]) { d_ins[i]->mutable_data(ctx.GetPlace()); auto t = framework::EigenVector::Flatten(*d_ins[i]); @@ -63,19 +66,19 @@ class MultiplexGradGPUKernel : public framework::OpKernel { } } - auto rows = ins[1]->dims()[0]; - auto cols = ins[1]->dims()[1]; + auto rows = ins[0]->dims()[0]; + auto cols = ins[0]->numel() / rows; // copy index to cpu - framework::Tensor index_t_cpu; - index_t_cpu.CopyFrom(*(ins[0]), platform::CPUPlace()); - auto* index = index_t_cpu.data(); + Tensor index_t_cpu; + index_t_cpu.CopyFrom(*ids, platform::CPUPlace()); + auto* index = index_t_cpu.data(); auto stream = reinterpret_cast( ctx.device_context()) .stream(); Place place = boost::get(ctx.GetPlace()); for (auto i = 0; i < rows; i++) { - int k = (int)index[i] + 1; + size_t k = static_cast(index[i]); if (d_ins[k]) { memory::Copy(place, d_ins[k]->data() + i * cols, place, d_out->data() + i * cols, cols * sizeof(T), stream); diff --git a/paddle/operators/multiplex_op.h b/paddle/operators/multiplex_op.h index 98466426bd90bc30a22ecf74e6739e2d4ad1d21d..ab3cafaa324a29d6f249cf1f73db92e1364eebc8 100644 --- a/paddle/operators/multiplex_op.h +++ b/paddle/operators/multiplex_op.h @@ -23,20 +23,22 @@ namespace paddle { namespace operators { template -class MultiplexCPUKernel : public framework::OpKernel { +class MultiplexCPUKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const { auto ins = ctx.MultiInput("X"); - auto* out = ctx.Output("Out"); + auto ids = ctx.Input("Ids"); + auto* out = ctx.Output("Out"); out->mutable_data(ctx.GetPlace()); - auto rows = ins[1]->dims()[0]; - auto cols = ins[1]->dims()[1]; - auto* index = ins[0]->data(); + auto rows = ins[0]->dims()[0]; + auto cols = ins[0]->numel() / rows; + auto index = ids->data(); Place place = boost::get(ctx.GetPlace()); for (auto i = 0; i < rows; i++) { - int k = (int)index[i] + 1; + int32_t k = index[i]; + PADDLE_ENFORCE_GE(k, 0, "index must be nonnegative."); PADDLE_ENFORCE_LT(static_cast(k), ins.size(), "index exceeds the number of candidate tensors."); memory::Copy(place, out->data() + i * cols, place, @@ -46,14 +48,15 @@ class MultiplexCPUKernel : public framework::OpKernel { }; template -class MultiplexGradCPUKernel : public framework::OpKernel { +class MultiplexGradCPUKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const { auto* d_out = ctx.Input(framework::GradVarName("Out")); + auto* ids = ctx.Input("Ids"); auto ins = ctx.MultiInput("X"); auto d_ins = ctx.MultiOutput(framework::GradVarName("X")); - for (size_t i = 1; i < d_ins.size(); i++) { + for (size_t i = 0; i < d_ins.size(); i++) { if (d_ins[i]) { d_ins[i]->mutable_data(ctx.GetPlace()); auto t = framework::EigenVector::Flatten(*d_ins[i]); @@ -61,12 +64,12 @@ class MultiplexGradCPUKernel : public framework::OpKernel { } } - auto rows = ins[1]->dims()[0]; - auto cols = ins[1]->dims()[1]; - auto* index = ins[0]->data(); + auto rows = ins[0]->dims()[0]; + auto cols = ins[0]->numel() / rows; + auto* index = ids->data(); Place place = boost::get(ctx.GetPlace()); for (auto i = 0; i < rows; i++) { - int k = (int)index[i] + 1; + size_t k = static_cast(index[i]); if (d_ins[k]) { memory::Copy(place, d_ins[k]->data() + i * cols, place, d_out->data() + i * cols, cols * sizeof(T)); diff --git a/paddle/operators/net_op.h b/paddle/operators/net_op.h index fcd8134b2c19cae6a4d006a4cd6fe32d2d627c34..ebeb262d9621fa35c870b6407992f6b6d2bf7c70 100644 --- a/paddle/operators/net_op.h +++ b/paddle/operators/net_op.h @@ -14,6 +14,7 @@ limitations under the License. */ #pragma once +#include #include "paddle/framework/framework.pb.h" #include "paddle/framework/op_registry.h" @@ -53,16 +54,6 @@ class NetOp : public framework::OperatorBase { this->CompleteAddOp(); } - /** - * Infer all the operators' input and output variables' shapes, will be called - * before every mini-batch - */ - void InferShape(const framework::Scope& scope) const override { - for (auto& op : ops_) { - op->InferShape(scope); - } - } - /** * @brief Run the network. * diff --git a/paddle/operators/net_op_test.cc b/paddle/operators/net_op_test.cc index f2e98ee7a1e14ee739abba01e97608845ce557f4..63bebd5b44719868a38ddf2b023955d1ab05245c 100644 --- a/paddle/operators/net_op_test.cc +++ b/paddle/operators/net_op_test.cc @@ -7,14 +7,12 @@ namespace operators { using Scope = framework::Scope; using DeviceContext = platform::DeviceContext; -static int infer_shape_cnt = 0; static int run_cnt = 0; class TestOp : public framework::OperatorBase { public: using framework::OperatorBase::OperatorBase; DEFINE_OP_CLONE_METHOD(TestOp); - void InferShape(const Scope& scope) const override { ++infer_shape_cnt; } void Run(const Scope& scope, const platform::DeviceContext& dev_ctx) const override { ++run_cnt; diff --git a/paddle/operators/pad_op.cc b/paddle/operators/pad_op.cc index 375d8a35acc0716259071c31bc332fdf5aabce1c..2f26ada85e477a6e5cd24c881f6369548331e8d0 100644 --- a/paddle/operators/pad_op.cc +++ b/paddle/operators/pad_op.cc @@ -24,14 +24,13 @@ class PadOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), - "Input(X) of PadOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"), - "Output(Out) of PadOp should not be null."); - - auto x_dim = ctx.Input("X")->dims(); - auto paddings = Attr>("paddings"); + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) of PadOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of PadOp should not be null."); + + auto x_dim = ctx->GetInputDim("X"); + auto paddings = ctx->Attrs().Get>("paddings"); PADDLE_ENFORCE_EQ(x_dim.size() * 2, int64_t(paddings.size()), "Size of paddings should be equal to 2 * dimension size " "of input tensor."); @@ -39,27 +38,25 @@ class PadOp : public framework::OperatorWithKernel { for (int i = 0; i < x_dim.size(); ++i) { out_dims[i] = x_dim[i] + paddings[i * 2] + paddings[i * 2 + 1]; } - ctx.Output("Out")->Resize( - framework::make_ddim(out_dims)); + ctx->SetOutputDim("Out", framework::make_ddim(out_dims)); if (out_dims[0] == x_dim[0]) { // Only pass LoD when the first dimension is equal between // output and input. - ctx.ShareLoD("X", /*->*/ "Out"); + ctx->ShareLoD("X", /*->*/ "Out"); } } }; class PadOpMaker : public framework::OpProtoAndCheckerMaker { public: - PadOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + PadOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "The input of pad op. " "The input should be a k-D tensor(k > 0 and k < 7)"); AddOutput("Out", "The output of pad op." - "A tensor with the same shape as X.") - .NotInGradient(); + "A tensor with the same shape as X."); AddComment(R"DOC( Pad input into output, as specified by paddings and pad_value. The input should be a k-D tensor(k > 0 and k < 7). As an example: @@ -68,15 +65,15 @@ Given: X = [[1, 2], [3, 4]] -and +and paddings = [0, 1, 1, 2] and - -pad_value = 0 -then we get +pad_value = 0 + +then we get Out = [[0, 1, 2, 0, 0] [0, 3, 4, 0, 0] @@ -101,23 +98,41 @@ class PadOpGrad : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) should not be null"); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")), - "Input(Out@GRAD) should not be null"); - auto x_dims = ctx.Input("X")->dims(); - auto *x_g = ctx.Output(framework::GradVarName("X")); - if (x_g != nullptr) { - x_g->Resize(x_dims); + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null"); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), + "Input(Out@GRAD) should not be null"); + auto x_dims = ctx->GetInputDim("X"); + auto x_grad_name = framework::GradVarName("X"); + if (ctx->HasOutput(x_grad_name)) { + ctx->SetOutputDim(x_grad_name, x_dims); } } }; +class PadOpGradMaker : public framework::SingleGradOpDescMaker { + public: + using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; + + protected: + std::unique_ptr Apply() const override { + auto* bind = new framework::OpDescBind(); + bind->SetInput("X", Input("X")); + bind->SetInput(framework::GradVarName("Out"), OutputGrad("Out")); + bind->SetOutput(framework::GradVarName("X"), InputGrad("X")); + bind->SetAttrMap(Attrs()); + bind->SetType("pad_grad"); + return std::unique_ptr(bind); + } +}; + } // namespace operators } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP(pad, ops::PadOp, ops::PadOpMaker, pad_grad, ops::PadOpGrad); + +REGISTER_OPERATOR(pad, ops::PadOp, ops::PadOpMaker, ops::PadOpGradMaker); +REGISTER_OPERATOR(pad_grad, ops::PadOpGrad); REGISTER_OP_CPU_KERNEL(pad, ops::PadKernel); REGISTER_OP_CPU_KERNEL(pad_grad, ops::PadGradKernel); diff --git a/paddle/operators/pad_op.h b/paddle/operators/pad_op.h index 2cc3b945ae5b2e2e93d8531c7f99e4c215d1d806..9534dbf54529e3b9ae2b6640d51fe291e9521927 100644 --- a/paddle/operators/pad_op.h +++ b/paddle/operators/pad_op.h @@ -47,7 +47,7 @@ void PadFunction(const framework::ExecutionContext& context) { } template -class PadKernel : public framework::OpKernel { +class PadKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { int rank = context.Input("X")->dims().size(); @@ -97,7 +97,7 @@ void PadGradFunction(const framework::ExecutionContext& context) { } template -class PadGradKernel : public framework::OpKernel { +class PadGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { size_t rank = diff --git a/paddle/operators/pool_op.cc b/paddle/operators/pool_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..ba3b5ed2075ceca284b49ecddb90ba5950b820c3 --- /dev/null +++ b/paddle/operators/pool_op.cc @@ -0,0 +1,195 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/operators/pool_op.h" + +namespace paddle { +namespace operators { + +int OutputSizePool(int input_size, int filter_size, int padding, int stride) { + int output_size = (input_size - filter_size + 2 * padding) / stride + 1; + return output_size; +} + +class PoolOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "X(Input) of Pooling should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Out(Output) of Pooling should not be null."); + + auto in_x_dims = ctx->GetInputDim("X"); + + std::string pooling_type = ctx->Attrs().Get("poolingType"); + std::vector ksize = ctx->Attrs().Get>("ksize"); + std::vector strides = ctx->Attrs().Get>("strides"); + std::vector paddings = ctx->Attrs().Get>("paddings"); + + PADDLE_ENFORCE(pooling_type == "max" || pooling_type == "avg", + "pooling_type should be 'max' or 'avg'"); + PADDLE_ENFORCE(in_x_dims.size() == 4 || in_x_dims.size() == 5, + "Pooling intput should be 4-D or 5-D"); + + if (ctx->Attrs().Get("globalPooling")) { + ksize.resize(static_cast(in_x_dims.size()) - 2); + for (size_t i = 0; i < ksize.size(); ++i) + ksize[i] = static_cast(in_x_dims[i + 2]); + } + + PADDLE_ENFORCE(in_x_dims.size() - ksize.size() == 2U, + "Input size and Pooling size should be consistent."); + PADDLE_ENFORCE(ksize.size() == 2 || ksize.size() == 3, + "Pooling size should be 2 elements. or 3 elements."); + PADDLE_ENFORCE_EQ(ksize.size(), strides.size(), + "strides size and pooling size should be the same."); + PADDLE_ENFORCE_EQ(ksize.size(), paddings.size(), + "paddings size and pooling size should be the same."); + + std::vector output_shape({in_x_dims[0], in_x_dims[1]}); + for (size_t i = 0; i < ksize.size(); ++i) { + output_shape.push_back( + OutputSizePool(in_x_dims[i + 2], ksize[i], paddings[i], strides[i])); + } + ctx->SetOutputDim("Out", framework::make_ddim(output_shape)); + } +}; + +class PoolOpGrad : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "X(Input) of Pooling should not be null."); + PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("X")), + "Input@Grad of Pooling should not be null."); + ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X")); + } +}; + +class Pool2dOpMaker : public framework::OpProtoAndCheckerMaker { + public: + Pool2dOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput( + "X", + "The input tensor of pooling operator. " + "The format of input tensor is NCHW. Where N is batch size, C is the " + "number of channels, H and W is the height and width of feature."); + AddOutput("Out", + "The output tensor of pooling operator." + "The format of output tensor is also NCHW."); + + AddAttr("poolingType", + "PoolingType of pooling operator." + "Str constant equal to 'max' or 'avg'.") + .InEnum({"max", "avg"}); + AddAttr>( + "ksize", + "Pooling size(depth, height, width) of pooling operator." + "If globalPooling = true, ksize is ignored and need not be " + "specified."); // TODO(Add checker) + AddAttr( + "globalPooling", + "Whether to use the globalPooling." + "Bool constant equal to false or true." + "Default false." + "If globalPooling = true, ksize is ignored and need not be specified.") + .SetDefault(false); + AddAttr>("strides", + "Strides(height, width) of pooling operator." + "Default {1,1}") + .SetDefault({1, 1}); // TODO(Add checker) + AddAttr>("paddings", + "Paddings(height, width) of pooling operator." + "Default {0,0}.") + .SetDefault({0, 0}); // TODO(Add checker) + AddComment(R"DOC( +The pooling2d operation calculates the output based on +the input, poolingType and ksize, strides, paddings parameters. +)DOC"); + } +}; + +class Pool3dOpMaker : public framework::OpProtoAndCheckerMaker { + public: + Pool3dOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", + "The input tensor of pooling operator. " + "The format of input tensor is NCDHW. Where N is batch size, C is " + "the " + "number of channels, D, H and W is the depth, height and width of " + "feature."); + AddOutput("Out", + "The output tensor of pooling operator." + "The format of output tensor is also NCDHW."); + + AddAttr("poolingType", + "PoolingType of pooling operator." + "str constant equal to 'max' or 'avg'.") + .InEnum({"max", "avg"}); + AddAttr>( + "ksize", + "Pooling size(depth, height, width) of pooling operator." + "If globalPooling = true, ksize is ignored and need not be " + "specified."); // TODO(Add checker) + AddAttr( + "globalPooling", + "Whether to use the globalPooling." + "Bool constant equal to false or true." + "Default false." + "If globalPooling = true, ksize is ignored and need not be specified.") + .SetDefault(false); + AddAttr>( + "strides", + "Strides(depth, height, width) of pooling operator." + "Default {1,1,1}.") + .SetDefault({1, 1, 1}); // TODO(Add checker) + AddAttr>( + "paddings", + "Paddings(depth, height, width) of pooling operator." + "Default {0,0,0}.") + .SetDefault({0, 0, 0}); // TODO(Add checker) + AddComment(R"DOC( +The pooling3d operation calculates the output based on +the input, poolingType and ksize, strides, paddings parameters. +)DOC"); + } +}; +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; + +REGISTER_OP(pool2d, ops::PoolOp, ops::Pool2dOpMaker, pool2d_grad, + ops::PoolOpGrad); + +REGISTER_OP_CPU_KERNEL(pool2d, + ops::PoolKernel); +REGISTER_OP_CPU_KERNEL(pool2d_grad, + ops::PoolGradKernel) + +REGISTER_OP(pool3d, ops::PoolOp, ops::Pool3dOpMaker, pool3d_grad, + ops::PoolOpGrad); + +REGISTER_OP_CPU_KERNEL(pool3d, + ops::PoolKernel); +REGISTER_OP_CPU_KERNEL(pool3d_grad, + ops::PoolGradKernel); diff --git a/paddle/operators/pool_op.cu b/paddle/operators/pool_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..0e3b80868f7b9d1697d619889160856d65ad59a3 --- /dev/null +++ b/paddle/operators/pool_op.cu @@ -0,0 +1,27 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/operators/pool_op.h" + +namespace ops = paddle::operators; + +REGISTER_OP_GPU_KERNEL(pool2d, + ops::PoolKernel); +REGISTER_OP_GPU_KERNEL(pool2d_grad, + ops::PoolGradKernel); + +REGISTER_OP_GPU_KERNEL(pool3d, + ops::PoolKernel); +REGISTER_OP_GPU_KERNEL(pool3d_grad, + ops::PoolGradKernel); diff --git a/paddle/operators/pool_op.h b/paddle/operators/pool_op.h new file mode 100644 index 0000000000000000000000000000000000000000..c2bc358def42959f2cc8f61cb00436fae1b7514b --- /dev/null +++ b/paddle/operators/pool_op.h @@ -0,0 +1,147 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" +#include "paddle/operators/math/math_function.h" +#include "paddle/operators/math/pooling.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; + +template +class PoolKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + const Tensor* in_x = context.Input("X"); + Tensor* out = context.Output("Out"); + + std::string pooling_type = context.Attr("poolingType"); + std::vector ksize = context.Attr>("ksize"); + std::vector strides = context.Attr>("strides"); + std::vector paddings = context.Attr>("paddings"); + if (context.Attr("globalPooling")) { + for (size_t i = 0; i < ksize.size(); ++i) { + ksize[i] = static_cast(in_x->dims()[i + 2]); + } + } + + switch (ksize.size()) { + case 2: { + if (pooling_type == "max") { + paddle::operators::math::Pool2dFunctor< + Place, paddle::operators::math::MaxPool, T> + pool2d_forward; + paddle::operators::math::MaxPool pool_process; + pool2d_forward(context.device_context(), *in_x, *out, ksize, strides, + paddings, pool_process); + + } else if (pooling_type == "avg") { + paddle::operators::math::Pool2dFunctor< + Place, paddle::operators::math::AvgPool, T> + pool2d_forward; + paddle::operators::math::AvgPool pool_process; + pool2d_forward(context.device_context(), *in_x, *out, ksize, strides, + paddings, pool_process); + } + } break; + case 3: { + if (pooling_type == "max") { + paddle::operators::math::Pool3dFunctor< + Place, paddle::operators::math::MaxPool, T> + pool3d_forward; + paddle::operators::math::MaxPool pool_process; + pool3d_forward(context.device_context(), *in_x, *out, ksize, strides, + paddings, pool_process); + } else if (pooling_type == "avg") { + paddle::operators::math::Pool3dFunctor< + Place, paddle::operators::math::AvgPool, T> + pool3d_forward; + paddle::operators::math::AvgPool pool_process; + pool3d_forward(context.device_context(), *in_x, *out, ksize, strides, + paddings, pool_process); + } + } break; + } + } +}; + +template +class PoolGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + const Tensor* in_x = context.Input("X"); + const Tensor* out = context.Input("Out"); + const Tensor* out_grad = + context.Input(framework::GradVarName("Out")); + Tensor* in_x_grad = context.Output(framework::GradVarName("X")); + + std::string pooling_type = context.Attr("poolingType"); + std::vector ksize = context.Attr>("ksize"); + std::vector strides = context.Attr>("strides"); + std::vector paddings = context.Attr>("paddings"); + + if (context.Attr("globalPooling")) { + for (size_t i = 0; i < ksize.size(); ++i) + ksize[i] = static_cast(in_x->dims()[i + 2]); + } + + if (in_x_grad) { + in_x_grad->mutable_data(context.GetPlace()); + auto temp = framework::EigenVector::Flatten(*in_x_grad); + temp.device(context.GetEigenDevice()) = + temp.constant(static_cast(0)); + + switch (ksize.size()) { + case 2: { + if (pooling_type == "max") { + paddle::operators::math::MaxPool2dGradFunctor + pool2d_backward; + pool2d_backward(context.device_context(), *in_x, *in_x_grad, *out, + *out_grad, ksize, strides, paddings); + } else if (pooling_type == "avg") { + paddle::operators::math::Pool2dGradFunctor< + Place, paddle::operators::math::AvgPoolGrad, T> + pool2d_backward; + paddle::operators::math::AvgPoolGrad pool_process; + pool2d_backward(context.device_context(), *in_x, *in_x_grad, *out, + *out_grad, ksize, strides, paddings, pool_process); + } + } break; + case 3: { + if (pooling_type == "max") { + paddle::operators::math::MaxPool3dGradFunctor + pool3d_backward; + pool3d_backward(context.device_context(), *in_x, *in_x_grad, *out, + *out_grad, ksize, strides, paddings); + } else if (pooling_type == "avg") { + paddle::operators::math::Pool3dGradFunctor< + Place, paddle::operators::math::AvgPoolGrad, T> + pool3d_backward; + paddle::operators::math::AvgPoolGrad pool_process; + pool3d_backward(context.device_context(), *in_x, *in_x_grad, *out, + *out_grad, ksize, strides, paddings, pool_process); + } + } break; + } + } + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/pool_with_index_op.cc b/paddle/operators/pool_with_index_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..7b6afcfd1f7e30624cb6859228892677cba58856 --- /dev/null +++ b/paddle/operators/pool_with_index_op.cc @@ -0,0 +1,228 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/operators/pool_with_index_op.h" + +namespace paddle { +namespace operators { + +inline int OutputSizeMaxPool(int input_size, int filter_size, int padding, + int stride) { + int output_size = (input_size - filter_size + 2 * padding) / stride + 1; + return output_size; +} + +class MaxPoolWithIndexOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "X(Input) of Pooling should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Out(Output) of Pooling should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Mask"), + "Mask(Output) of Pooling should not be null."); + + auto in_x_dims = ctx->GetInputDim("X"); + + std::vector ksize = ctx->Attrs().Get>("ksize"); + std::vector strides = ctx->Attrs().Get>("strides"); + std::vector paddings = ctx->Attrs().Get>("paddings"); + + PADDLE_ENFORCE(in_x_dims.size() == 4 || in_x_dims.size() == 5, + "Pooling intput should be 4-D or 5-D"); + + if (ctx->Attrs().Get("globalPooling")) { + ksize.resize(static_cast(in_x_dims.size()) - 2); + for (size_t i = 0; i < ksize.size(); ++i) + ksize[i] = static_cast(in_x_dims[i + 2]); + } + + PADDLE_ENFORCE(in_x_dims.size() - ksize.size() == 2U, + "Intput size and pooling size should be consistent."); + PADDLE_ENFORCE_EQ(ksize.size(), strides.size(), + "Strides size and pooling size should be the same."); + PADDLE_ENFORCE_EQ(ksize.size(), paddings.size(), + "Paddings size and pooling size should be the same."); + + std::vector output_shape({in_x_dims[0], in_x_dims[1]}); + for (size_t i = 0; i < ksize.size(); ++i) { + output_shape.push_back(OutputSizeMaxPool(in_x_dims[i + 2], ksize[i], + paddings[i], strides[i])); + } + ctx->SetOutputDim("Out", framework::make_ddim(output_shape)); + ctx->SetOutputDim("Mask", framework::make_ddim(output_shape)); + } +}; + +class MaxPoolWithIndexOpGrad : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) must not be null."); + PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("X")), + "Input(X@GRAD) should not be null."); + ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X")); + } +}; + +class MaxPool2dWithIndexOpMaker : public framework::OpProtoAndCheckerMaker { + public: + MaxPool2dWithIndexOpMaker(framework::OpProto *proto, + framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput( + "X", + "The input tensor of pooling operator. " + "The format of input tensor is NCHW. Where N is batch size, C is the " + "number of channels, H and W is the height and width of image."); + AddOutput("Out", + "The output tensor of pooling operator." + "The format of output tensor is also NCHW." + "Where N is batch size, C is " + "the number of channels, H and W is the height and " + "width of image."); + AddOutput("Mask", + "The Mask tensor of pooling operator." + "The format of output tensor is also NCHW." + "Where N is batch size, C is the number of channels, H and W " + "is the height and width of image." + "The value in it is the index in current feature map"); + + AddAttr>( + "ksize", + "The pooling size(height, width) of pooling operator." + "If globalPooling = true, ksize is ignored and need not be " + "specified."); // TODO(Chengduo): Add checker. (Currently, + // TypedAttrChecker don't support vector type.) + AddAttr( + "globalPooling", + "Whether to use the globalPooling." + "Bool constant equal to false or true." + "Default false." + "If globalPooling = true, ksize is ignored and need not be specified.") + .SetDefault(false); + AddAttr>("strides", + "Strides(height, width) of pooling operator." + "Default {1,1}.") + .SetDefault({1, 1}); // TODO(Chengduo): Add checker. (Currently, + // TypedAttrChecker don't support vector type.) + AddAttr>("paddings", + "Paddings(height, width) of pooling operator." + "Default {0,0}.") + .SetDefault({0, 0}); // TODO(Chengduo): Add checker. (Currently, + // TypedAttrChecker don't support vector type.) + + AddComment(R"DOC( +The maxPooling2d with index operation calculates the output and the mask +based on the input and ksize, strides, paddings parameters. Input(X) and +output(Out, Mask) are in NCHW format. Where N is batch size, C is the +number of channels, H and W is the height and width of feature. +Parameters(ksize, strides, paddings) are two elements. +These two elements represent height and width, respectively. +)DOC"); + } +}; + +class MaxPool3dWithIndexOpMaker : public framework::OpProtoAndCheckerMaker { + public: + MaxPool3dWithIndexOpMaker(framework::OpProto *proto, + framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput( + "X", + "The input tensor of pooling operator. " + "The format of input tensor is NCDHW. Where N is batch size, C is " + "the number of channels, D, H and W is the depth, height and width of " + "image."); + AddOutput("Out", + "The output tensor of pooling operator." + "The format of output tensor is also NCDHW." + "Where N is batch size, C is " + "the number of channels, D, H and W is the depth, height and " + "width of image."); + AddOutput("Mask", + "The Mask tensor of pooling operator." + "The format of output tensor is also NCDHW." + "Where N is batch size, C is the number of channels, D, H and W " + "is the depth, height and width of image." + "The value in it is the index in current feature map"); + + AddAttr>( + "ksize", + "The pooling size(depth, height, width) of pooling operator." + "If globalPooling = true, ksize is ignored and need not be " + "specified."); // TODO(Chengduo): Add checker. (Currently, + // TypedAttrChecker don't support vector type.) + AddAttr( + "globalPooling", + "Whether to use the globalPooling." + "Bool constant equal to false or true." + "Default false." + "If globalPooling = true, ksize is ignored and need not be specified.") + .SetDefault(false); + AddAttr>( + "strides", + "Strides(depth, height, width) of pooling operator." + "Default {1,1,1}.") + .SetDefault({1, 1, 1}); // TODO(Chengduo): Add checker. (Currently, + // TypedAttrChecker don't support vector type.) + AddAttr>( + "paddings", + "Paddings(depth, height, width) of pooling operator." + "Default {0,0,0}.") + .SetDefault({0, 0, 0}); // TODO(Chengduo): Add checker. (Currently, + // TypedAttrChecker don't support vector type.) + + AddComment(R"DOC( +The maxpooling3d with index operation calculates the output and the mask +based on the input and ksize, strides, paddings parameters. +Input(X) and output(Out, Mask) are in NCDHW format. Where N is batch +size, C is the number of channels, D, H and W is the depth, height and +width of feature. Parameters(ksize, strides, paddings) are three elements. +These three elements represent depth, height and width, respectively. +)DOC"); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; + +REGISTER_OP(max_pool2d_with_index, ops::MaxPoolWithIndexOp, + ops::MaxPool2dWithIndexOpMaker, max_pool2d_with_index_grad, + ops::MaxPoolWithIndexOpGrad); + +REGISTER_OP_CPU_KERNEL( + max_pool2d_with_index, + ops::MaxPoolWithIndexKernel); +REGISTER_OP_CPU_KERNEL( + max_pool2d_with_index_grad, + ops::MaxPoolWithIndexGradKernel) + +REGISTER_OP(max_pool3d_with_index, ops::MaxPoolWithIndexOp, + ops::MaxPool3dWithIndexOpMaker, max_pool3d_with_index_grad, + ops::MaxPoolWithIndexOpGrad); + +REGISTER_OP_CPU_KERNEL( + max_pool3d_with_index, + ops::MaxPoolWithIndexKernel); +REGISTER_OP_CPU_KERNEL( + max_pool3d_with_index_grad, + ops::MaxPoolWithIndexGradKernel) diff --git a/paddle/operators/pool_with_index_op.cu b/paddle/operators/pool_with_index_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..287657d4b1c57f354ef050885f71261092bdc062 --- /dev/null +++ b/paddle/operators/pool_with_index_op.cu @@ -0,0 +1,31 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/operators/pool_with_index_op.h" + +namespace ops = paddle::operators; + +REGISTER_OP_GPU_KERNEL( + max_pool2d_with_index, + ops::MaxPoolWithIndexKernel); +REGISTER_OP_GPU_KERNEL( + max_pool2d_with_index_grad, + ops::MaxPoolWithIndexGradKernel) + +REGISTER_OP_GPU_KERNEL( + max_pool3d_with_index, + ops::MaxPoolWithIndexKernel); +REGISTER_OP_GPU_KERNEL( + max_pool3d_with_index_grad, + ops::MaxPoolWithIndexGradKernel) diff --git a/paddle/operators/pool_with_index_op.h b/paddle/operators/pool_with_index_op.h new file mode 100644 index 0000000000000000000000000000000000000000..01b961ca8295f723bea7335e43ec5ab100dfc65c --- /dev/null +++ b/paddle/operators/pool_with_index_op.h @@ -0,0 +1,103 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" +#include "paddle/operators/math/math_function.h" +#include "paddle/operators/math/pooling.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; + +template +class MaxPoolWithIndexKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + const Tensor* in_x = context.Input("X"); + Tensor* out = context.Output("Out"); + Tensor* mask = context.Output("Mask"); + + std::vector ksize = context.Attr>("ksize"); + std::vector strides = context.Attr>("strides"); + std::vector paddings = context.Attr>("paddings"); + if (context.Attr("globalPooling")) { + for (size_t i = 0; i < ksize.size(); ++i) { + ksize[i] = static_cast(in_x->dims()[i + 2]); + } + } + + switch (ksize.size()) { + case 2: { + paddle::operators::math::MaxPool2dWithIndexFunctor + pool2d_forward; + pool2d_forward(context.device_context(), *in_x, *out, *mask, ksize, + strides, paddings); + } break; + case 3: { + paddle::operators::math::MaxPool3dWithIndexFunctor + pool3d_forward; + pool3d_forward(context.device_context(), *in_x, *out, *mask, ksize, + strides, paddings); + } break; + } + } +}; + +template +class MaxPoolWithIndexGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + const Tensor* mask = context.Input("Mask"); + const Tensor* out_grad = + context.Input(framework::GradVarName("Out")); + Tensor* in_x_grad = context.Output(framework::GradVarName("X")); + + std::vector ksize = context.Attr>("ksize"); + std::vector strides = context.Attr>("strides"); + std::vector paddings = context.Attr>("paddings"); + if (context.Attr("globalPooling")) { + for (size_t i = 0; i < ksize.size(); ++i) { + ksize[i] = static_cast(in_x_grad->dims()[i + 2]); + } + } + + if (in_x_grad) { + in_x_grad->mutable_data(context.GetPlace()); + auto temp = framework::EigenVector::Flatten(*in_x_grad); + temp.device(context.GetEigenDevice()) = + temp.constant(static_cast(0)); + + switch (ksize.size()) { + case 2: { + paddle::operators::math::MaxPool2dWithIndexGradFunctor + pool2d_backward; + pool2d_backward(context.device_context(), *in_x_grad, *out_grad, + *mask, ksize, strides, paddings); + } break; + case 3: { + paddle::operators::math::MaxPool3dWithIndexGradFunctor + pool3d_backward; + pool3d_backward(context.device_context(), *in_x_grad, *out_grad, + *mask, ksize, strides, paddings); + } break; + } + } + } +}; +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/prelu_op.cc b/paddle/operators/prelu_op.cc index 912196c190b5ddbd4e3482a5314e949186b94368..166fe2682422c61a264c88b8683752ee323e940e 100644 --- a/paddle/operators/prelu_op.cc +++ b/paddle/operators/prelu_op.cc @@ -26,19 +26,14 @@ class PReluOp : public framework::OperatorWithKernel { : OperatorWithKernel(type, inputs, outputs, attrs) {} protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) should not be null"); - auto *in = ctx.Input("X"); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Alpha"), - "Input(Alpha) should not be null"); - auto *alpha = ctx.Input("Alpha"); - PADDLE_ENFORCE(alpha->numel() == 1, "Size of weight Alpha must be one."); - - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"), - "Output(Out) should not be null"); - auto *out = ctx.Output("Out"); - out->Resize(in->dims()); - ctx.ShareLoD("X", /*->*/ "Out"); + void InferShape(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null"); + PADDLE_ENFORCE(ctx->HasInput("Alpha"), "Input(Alpha) should not be null"); + PADDLE_ENFORCE(product(ctx->GetInputDim("Alpha")) == 1, + "Size of weight Alpha must be one."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) should not be null"); + ctx->SetOutputDim("Out", ctx->GetInputDim("X")); + ctx->ShareLoD("X", /*->*/ "Out"); } }; @@ -68,19 +63,13 @@ class PReluGradOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) must not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")), - "Input(Out@GRAD) should not be null"); - auto *dx = ctx.Output(framework::GradVarName("X")); - auto *x = ctx.Input("X"); - - auto *dalpha = - ctx.Output(framework::GradVarName("Alpha")); - auto *alpha = ctx.Input("Alpha"); - - dx->Resize(x->dims()); - dalpha->Resize(alpha->dims()); + void InferShape(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) must not be null."); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), + "Input(Out@GRAD) should not be null"); + ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X")); + ctx->SetOutputDim(framework::GradVarName("Alpha"), + ctx->GetInputDim("Alpha")); } }; diff --git a/paddle/operators/prelu_op.h b/paddle/operators/prelu_op.h index 6b78ed295cbac060d816fb3dd27a4b80145cb1ce..5ad31c2203ae6c9bf6f48bb9ecf9a714597e7da8 100644 --- a/paddle/operators/prelu_op.h +++ b/paddle/operators/prelu_op.h @@ -40,7 +40,7 @@ class PReluFunctor { }; template -class PReluKernel : public framework::OpKernel { +class PReluKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto* x = context.Input("X"); @@ -77,7 +77,7 @@ class PReluGradFunctor { }; template -class PReluGradKernel : public framework::OpKernel { +class PReluGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto* dx = context.Output(framework::GradVarName("X")); diff --git a/paddle/operators/rank_loss_op.cc b/paddle/operators/rank_loss_op.cc index 39af08c8751c3b95cf5fdef7395186a0176a20a2..e0abbc4db1723e297c313d50fc1f9cbb77acb9f3 100644 --- a/paddle/operators/rank_loss_op.cc +++ b/paddle/operators/rank_loss_op.cc @@ -25,22 +25,21 @@ class RankLossOp : public framework::OperatorWithKernel { : OperatorWithKernel(type, inputs, outputs, attrs) {} protected: - void InferShape(const framework::InferShapeContext &ctx) const override { + void InferShape(framework::InferShapeContext *ctx) const override { // input check - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Label"), - "Input(Label) shouldn't be null"); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Left"), - "Input(Left) shouldn't be null"); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Right"), - "Input(Right) shouldn't be null"); - auto label_dims = ctx.Input("Label")->dims(); - auto left_dims = ctx.Input("Left")->dims(); - auto right_dims = ctx.Input("Right")->dims(); + PADDLE_ENFORCE(ctx->HasInput("Label"), "Input(Label) shouldn't be null"); + PADDLE_ENFORCE(ctx->HasInput("Left"), "Input(Left) shouldn't be null"); + PADDLE_ENFORCE(ctx->HasInput("Right"), "Input(Right) shouldn't be null"); + + auto label_dims = ctx->GetInputDim("Label"); + auto left_dims = ctx->GetInputDim("Left"); + auto right_dims = ctx->GetInputDim("Right"); + PADDLE_ENFORCE((label_dims == left_dims) && (left_dims == right_dims), "All inputs must have the same size"); PADDLE_ENFORCE((label_dims.size() == 2) && (label_dims[1] == 1), "All inputs must be row vector with size batch_size x 1."); - ctx.Output("Out")->Resize(label_dims); + ctx->SetOutputDim("Out", label_dims); } }; @@ -91,25 +90,22 @@ class RankLossGradOp : public framework::OperatorWithKernel { : OperatorWithKernel(type, inputs, outputs, attrs) {} protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Label"), - "Input(Label) shouldn't be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Left"), - "Input(Left) shouldn't be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Right"), - "Input(Right) shouldn't be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")), - "Input(Out@GRAD) shouldn't be null."); - auto dims = ctx.Input("Left")->dims(); - auto *left_grad = - ctx.Output(framework::GradVarName("Left")); - auto *right_grad = - ctx.Output(framework::GradVarName("Right")); - if (left_grad) { - left_grad->Resize(dims); + void InferShape(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Label"), "Input(Label) shouldn't be null."); + PADDLE_ENFORCE(ctx->HasInput("Left"), "Input(Left) shouldn't be null."); + PADDLE_ENFORCE(ctx->HasInput("Right"), "Input(Right) shouldn't be null."); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), + "Input(Out@GRAD) shouldn't be null."); + auto dims = ctx->GetInputDim("Left"); + auto left_grad_name = framework::GradVarName("Left"); + auto right_grad_name = framework::GradVarName("Right"); + + if (ctx->HasOutput(left_grad_name)) { + ctx->SetOutputDim(left_grad_name, dims); } - if (right_grad) { - right_grad->Resize(dims); + + if (ctx->HasOutput(right_grad_name)) { + ctx->SetOutputDim(right_grad_name, dims); } } }; diff --git a/paddle/operators/rank_loss_op.h b/paddle/operators/rank_loss_op.h index 7df195ff47ecfd79388385eed4bd37b8c9b45979..f184d6efcb496a1d7f38540712b6c431f816482e 100644 --- a/paddle/operators/rank_loss_op.h +++ b/paddle/operators/rank_loss_op.h @@ -21,7 +21,7 @@ namespace paddle { namespace operators { template -class RankLossKernel : public framework::OpKernel { +class RankLossKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const { auto* out_t = ctx.Output("Out"); @@ -42,7 +42,7 @@ class RankLossKernel : public framework::OpKernel { }; template -class RankLossGradKernel : public framework::OpKernel { +class RankLossGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const { auto* d_left_t = diff --git a/paddle/operators/recurrent_op.cc b/paddle/operators/recurrent_op.cc index e7deaf9940699b938e4f36358c2c7f3ba15e918b..04c4c24951f5db572486ded5edfc26948a821682 100644 --- a/paddle/operators/recurrent_op.cc +++ b/paddle/operators/recurrent_op.cc @@ -28,61 +28,41 @@ using Variable = framework::Variable; using Tensor = framework::Tensor; using LoDTensor = framework::LoDTensor; -void RecurrentAlgorithm::InferShape(const Scope& scope) const { +void RecurrentAlgorithm::Run(const Scope& scope, + const platform::DeviceContext& dev_ctx) const { auto* input0 = scope.FindVar(arg_->inlinks[0]); PADDLE_ENFORCE_NOT_NULL(input0); - seq_len_ = input0->GetMutable()->dims()[0]; - PADDLE_ENFORCE_GT(seq_len_, 0); - - CreateScopes(scope); - auto step_scopes = GetStepScopes(scope); - rnn::SegmentInputs(step_scopes, arg_->inlinks, seq_len_, - true /*infer_shape_mode*/); - InitMemories(step_scopes[0], true /*infer_shape_mode*/); - - for (size_t i = 0; i < seq_len_; i++) { - if (i > 0) { - rnn::LinkMemories(step_scopes, arg_->memories, i, -1, - true /*infer_shape_mode*/); - } - (*stepnet_)->InferShape(*step_scopes[i]); - } - rnn::ConcatOutputs(step_scopes, arg_->outlinks, seq_len_, - true /*infer_shape_mode*/); -} + size_t seq_len = input0->GetMutable()->dims()[0]; + PADDLE_ENFORCE_GT(seq_len, 0); -void RecurrentAlgorithm::Run(const Scope& scope, - const platform::DeviceContext& dev_ctx) const { - auto step_scopes = GetStepScopes(scope); - rnn::SegmentInputs(step_scopes, arg_->inlinks, seq_len_, - false /*infer_shape_mode*/); - InitMemories(step_scopes[0], false /*infer_shape_mode*/); + CreateScopes(scope, seq_len); + auto& step_scopes = GetStepScopes(scope); + rnn::SegmentInputs(step_scopes, arg_->inlinks, seq_len); + InitMemories(step_scopes[0]); - for (size_t step_id = 0; step_id < seq_len_; step_id++) { - // create output alias variables + for (size_t step_id = 0; step_id < seq_len; step_id++) { if (step_id > 0) { - rnn::LinkMemories(step_scopes, arg_->memories, step_id, -1, - false /*infer_shape_mode*/); + rnn::LinkMemories(step_scopes, arg_->memories, step_id, -1); } (*stepnet_)->Run(*step_scopes[step_id], dev_ctx); } - rnn::ConcatOutputs(step_scopes, arg_->outlinks, seq_len_, - false /*infer_shape_mode*/); + rnn::ConcatOutputs(step_scopes, arg_->outlinks, seq_len); } -void RecurrentAlgorithm::CreateScopes(const Scope& scope) const { +void RecurrentAlgorithm::CreateScopes(const Scope& scope, + size_t seq_len) const { // TODO(superjom) Only two scopes are needed for inference, this case will be // supported later. - auto step_scopes_var = scope.FindVar(arg_->step_scopes); + auto* step_scopes_var = scope.FindVar(arg_->step_scopes); PADDLE_ENFORCE(step_scopes_var != nullptr, ""); - auto step_scopes = step_scopes_var->GetMutable>(); + auto* step_scopes = step_scopes_var->GetMutable>(); // Now all variables in scope must be created outside of op. PADDLE_ENFORCE_NOT_NULL(stepnet_); PADDLE_ENFORCE(!(*stepnet_)->Outputs().empty(), "stepnet_ op has no outputs"); - if (seq_len_ > step_scopes->size()) { - for (size_t i = step_scopes->size(); i < seq_len_; ++i) { + if (seq_len > step_scopes->size()) { + for (size_t i = step_scopes->size(); i < seq_len; ++i) { auto& step_scope = scope.NewScope(); // create step net's temp inputs @@ -105,8 +85,7 @@ void RecurrentAlgorithm::CreateScopes(const Scope& scope) const { } } -void RecurrentAlgorithm::InitMemories(Scope* step_scope, - bool infer_shape_mode) const { +void RecurrentAlgorithm::InitMemories(Scope* step_scope) const { for (auto& attr : arg_->memories) { auto* pre_mem = step_scope->NewVar(attr.pre_var)->GetMutable(); PADDLE_ENFORCE(step_scope->FindVar(attr.boot_var) != nullptr, @@ -114,12 +93,9 @@ void RecurrentAlgorithm::InitMemories(Scope* step_scope, attr.boot_var); auto* boot_mem = step_scope->FindVar(attr.boot_var)->GetMutable(); - if (infer_shape_mode) { - pre_mem->Resize(boot_mem->dims()); - PADDLE_ENFORCE_EQ(pre_mem->dims().size(), 2); - } else { - pre_mem->ShareDataWith(*boot_mem); - } + pre_mem->Resize(boot_mem->dims()); + PADDLE_ENFORCE_EQ(pre_mem->dims().size(), 2); + pre_mem->ShareDataWith(*boot_mem); } } @@ -169,23 +145,23 @@ class RecurrentAlgorithmProtoAndCheckerMaker void RecurrentGradientAlgorithm::Run( const Scope& scope, const platform::DeviceContext& dev_ctx) const { - auto step_scopes = GetStepScopes(scope); - rnn::SegmentInputs(step_scopes, arg_->inlinks, seq_len_, - false /*infer_shape_mode*/); - for (int step_id = seq_len_ - 1; step_id >= 0; --step_id) { - if (static_cast(step_id) != seq_len_ - 1) { - rnn::LinkMemories(step_scopes, arg_->memories, step_id, 1, - false /*infer_shape_mode*/); + auto* input0 = scope.FindVar(arg_->inlinks[0]); + PADDLE_ENFORCE_NOT_NULL(input0); + size_t seq_len = input0->GetMutable()->dims()[0]; + auto& step_scopes = GetStepScopes(scope); + rnn::SegmentInputs(step_scopes, arg_->inlinks, seq_len); + for (int step_id = seq_len - 1; step_id >= 0; --step_id) { + if (step_id != seq_len - 1) { + rnn::LinkMemories(step_scopes, arg_->memories, step_id, 1); } (*stepnet_)->Run(*step_scopes[step_id], dev_ctx); } - LinkBootMemoryGradients(step_scopes[0], false); - rnn::ConcatOutputs(step_scopes, arg_->outlinks, seq_len_, - false /*infer_shape_mode*/); + rnn::ConcatOutputs(step_scopes, arg_->outlinks, seq_len); + LinkBootMemoryGradients(step_scopes[0]); } void RecurrentGradientAlgorithm::LinkBootMemoryGradients( - Scope* step_scope, bool infer_shape_mode) const { + Scope* step_scope) const { for (auto& attr : arg_->memories) { PADDLE_ENFORCE(step_scope->FindVar(attr.var) != nullptr, "memory variable [%s] does not exists", attr.var); @@ -194,30 +170,9 @@ void RecurrentGradientAlgorithm::LinkBootMemoryGradients( auto* mem_grad = step_scope->NewVar(attr.var)->GetMutable(); auto* boot_mem_grad = step_scope->NewVar(attr.boot_var)->GetMutable(); - if (infer_shape_mode) { - boot_mem_grad->Resize(mem_grad->dims()); - } else { - boot_mem_grad->ShareDataWith(*mem_grad); - } - } -} - -void RecurrentGradientAlgorithm::InferShape(const Scope& scope) const { - seq_len_ = - scope.FindVar(arg_->inlinks[0])->GetMutable()->dims()[0]; - auto step_scopes = GetStepScopes(scope); - rnn::SegmentInputs(step_scopes, arg_->inlinks, seq_len_, - true /*infer_shape_mode*/); - for (int step_id = seq_len_ - 1; step_id >= 0; --step_id) { - if (static_cast(step_id) != seq_len_ - 1) { - rnn::LinkMemories(step_scopes, arg_->memories, step_id, 1, - true /*infer_shape_mode*/); - } - (*stepnet_)->InferShape(*step_scopes[step_id]); + boot_mem_grad->Resize(mem_grad->dims()); + boot_mem_grad->ShareDataWith(*mem_grad); } - rnn::ConcatOutputs(step_scopes, arg_->outlinks, seq_len_, - true /*infer_shape_mode*/); - LinkBootMemoryGradients(step_scopes[0], true /*infer_shape_mode*/); } RecurrentGradientOp::RecurrentGradientOp( diff --git a/paddle/operators/recurrent_op.h b/paddle/operators/recurrent_op.h index ad4df9e55b91dbe89c34762945cd9edefde86e08..253d7e3284360ceaddce9ef5f8f9a3ea4793d740 100644 --- a/paddle/operators/recurrent_op.h +++ b/paddle/operators/recurrent_op.h @@ -41,11 +41,6 @@ class RecurrentAlgorithm { stepnet_ = stepnet; } - /** - * InferShape must be called before Run. - */ - void InferShape(const framework::Scope& scope) const; - protected: /* * The step scopes will be stored in the father scope as a variable. @@ -53,7 +48,7 @@ class RecurrentAlgorithm { * NOTE the scopes are reused in both the forward and backward, so just * create once and expand its size if more steps need. */ - void CreateScopes(const framework::Scope& scope) const; + void CreateScopes(const framework::Scope& scope, size_t seq_len) const; const std::vector& GetStepScopes( const framework::Scope& scope) const { @@ -61,12 +56,11 @@ class RecurrentAlgorithm { ->GetMutable>(); } - void InitMemories(framework::Scope* step_scopes, bool infer_shape_mode) const; + void InitMemories(framework::Scope* step_scopes) const; private: std::unique_ptr* stepnet_; rnn::Argument* arg_; - mutable size_t seq_len_; }; class RecurrentGradientAlgorithm { @@ -91,13 +85,7 @@ class RecurrentGradientAlgorithm { void Run(const framework::Scope& scope, const platform::DeviceContext& dev_ctx) const; - void LinkBootMemoryGradients(framework::Scope* step_scopes, - bool infer_shape_mode) const; - - /** - * InferShape must be called before Run. - */ - void InferShape(const framework::Scope& scope) const; + void LinkBootMemoryGradients(framework::Scope* step_scopes) const; protected: inline const std::vector& GetStepScopes( @@ -108,7 +96,6 @@ class RecurrentGradientAlgorithm { private: rnn::Argument* arg_; - mutable size_t seq_len_; std::unique_ptr* stepnet_; }; @@ -124,12 +111,6 @@ class RecurrentOp : public framework::OperatorBase { // TODO(yuyang18): Implement copy ctor well. PADDLE_THROW("Not implemented"); } - /** - * InferShape must be called before Run. - */ - void InferShape(const framework::Scope& scope) const override { - alg_.InferShape(scope); - } void Run(const framework::Scope& scope, const platform::DeviceContext& dev_ctx) const override { @@ -139,6 +120,7 @@ class RecurrentOp : public framework::OperatorBase { void set_stepnet(std::unique_ptr net) { stepnet_ = std::move(net); } + const OperatorBase& stepnet() const { return *stepnet_; } static const rnn::ArgumentName kArgName; @@ -163,13 +145,6 @@ class RecurrentGradientOp : public framework::OperatorBase { PADDLE_THROW("Not Implemented"); } - /** - * InferShape must be called before Run. - */ - void InferShape(const framework::Scope& scope) const override { - alg_.InferShape(scope); - } - void Run(const framework::Scope& scope, const platform::DeviceContext& dev_ctx) const override { alg_.Run(scope, dev_ctx); diff --git a/paddle/operators/reduce_op.cc b/paddle/operators/reduce_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..005f88b57cb47b2c00f9a81da8afa49439665a22 --- /dev/null +++ b/paddle/operators/reduce_op.cc @@ -0,0 +1,189 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/operators/reduce_op.h" + +namespace paddle { +namespace operators { + +using framework::Tensor; + +class ReduceOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of ReduceOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of ReduceOp should not be null."); + auto x_dims = ctx->GetInputDim("X"); + auto x_rank = x_dims.size(); + PADDLE_ENFORCE_LE(x_rank, 6, "Tensors with rank at most 6 are supported."); + int dim = ctx->Attrs().Get("dim"); + if (dim < 0) dim = x_rank + dim; + PADDLE_ENFORCE_LT( + dim, x_rank, + "The dim should be in the range [-rank(input), rank(input))."); + bool keep_dim = ctx->Attrs().Get("keep_dim"); + auto dims_vector = vectorize(x_dims); + if (keep_dim || x_rank == 1) { + dims_vector[dim] = 1; + } else { + dims_vector.erase(dims_vector.begin() + dim); + } + auto out_dims = framework::make_ddim(dims_vector); + ctx->SetOutputDim("Out", out_dims); + if (dim != 0) { + // Only pass LoD when not reducing on the first dim. + ctx->ShareLoD("X", /*->*/ "Out"); + } + } +}; + +class ReduceGradOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null."); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), + "Input(Out@GRAD) should not be null."); + auto x_dims = ctx->GetInputDim("X"); + auto x_rank = x_dims.size(); + PADDLE_ENFORCE_LE(x_rank, 6, "Tensors with rank at most 6 are supported."); + int dim = ctx->Attrs().Get("dim"); + if (dim < 0) dim = x_rank + dim; + PADDLE_ENFORCE_LT( + dim, x_rank, + "The dim should be in the range [-rank(input), rank(input))."); + auto x_grad_name = framework::GradVarName("X"); + if (ctx->HasOutput(x_grad_name)) { + ctx->SetOutputDim(x_grad_name, x_dims); + } + } +}; + +class ReduceOpMaker : public framework::OpProtoAndCheckerMaker { + public: + ReduceOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput( + "X", + "(Tensor) The input tensor. Tensors with rank at most 6 are supported"); + AddOutput("Out", "(Tensor) The result tensor."); + AddAttr( + "dim", + "(int, default 1) The dimension to reduce. " + "Must be in the range [-rank(input), rank(input)). " + "If `dim < 0`, the dim to reduce is `rank + dim`. " + "Noting that reducing on the first dim will make the LoD info lost.") + .SetDefault(0); + AddAttr("keep_dim", + "(bool, default false) " + "If true, retain the reduced dimension with length 1.") + .SetDefault(false); + comment_ = R"DOC( +{ReduceOP} operator computes the {reduce} of input tensor along the given dimension. +The result tensor has 1 fewer dimension than the input unless `keep_dim` is true. +)DOC"; + AddComment(comment_); + } + + protected: + std::string comment_; + + void Replace(std::string &src, std::string from, std::string to) { + std::size_t len_from = std::strlen(from.c_str()); + std::size_t len_to = std::strlen(to.c_str()); + for (std::size_t pos = src.find(from); pos != std::string::npos; + pos = src.find(from, pos + len_to)) { + src.replace(pos, len_from, to); + } + } + + void SetComment(std::string name, std::string op) { + Replace(comment_, "{ReduceOP}", name); + Replace(comment_, "{reduce}", op); + } +}; + +class ReduceSumOpMaker : public ReduceOpMaker { + public: + ReduceSumOpMaker(framework::OpProto *proto, + framework::OpAttrChecker *op_checker) + : ReduceOpMaker(proto, op_checker) { + SetComment("ReduceSum", "sum"); + AddComment(comment_); + } +}; + +class ReduceMeanOpMaker : public ReduceOpMaker { + public: + ReduceMeanOpMaker(framework::OpProto *proto, + framework::OpAttrChecker *op_checker) + : ReduceOpMaker(proto, op_checker) { + SetComment("ReduceMean", "mean"); + AddComment(comment_); + } +}; + +class ReduceMaxOpMaker : public ReduceOpMaker { + public: + ReduceMaxOpMaker(framework::OpProto *proto, + framework::OpAttrChecker *op_checker) + : ReduceOpMaker(proto, op_checker) { + SetComment("ReduceMax", "max"); + AddComment(comment_); + } +}; + +class ReduceMinOpMaker : public ReduceOpMaker { + public: + ReduceMinOpMaker(framework::OpProto *proto, + framework::OpAttrChecker *op_checker) + : ReduceOpMaker(proto, op_checker) { + SetComment("ReduceMin", "min"); + AddComment(comment_); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; + +REGISTER_OP(reduce_sum, ops::ReduceOp, ops::ReduceSumOpMaker, reduce_sum_grad, + ops::ReduceGradOp); + +REGISTER_OP(reduce_mean, ops::ReduceOp, ops::ReduceMeanOpMaker, + reduce_mean_grad, ops::ReduceGradOp); + +REGISTER_OP(reduce_max, ops::ReduceOp, ops::ReduceMaxOpMaker, reduce_max_grad, + ops::ReduceGradOp); + +REGISTER_OP(reduce_min, ops::ReduceOp, ops::ReduceMinOpMaker, reduce_min_grad, + ops::ReduceGradOp); + +#define REGISTER_REDUCE_CPU_KERNEL(reduce_type, functor, grad_functor) \ + REGISTER_OP_CPU_KERNEL( \ + reduce_type, \ + ops::ReduceKernel); \ + REGISTER_OP_CPU_KERNEL(reduce_type##_grad, \ + ops::ReduceGradKernel); + +FOR_EACH_KERNEL_FUNCTOR(REGISTER_REDUCE_CPU_KERNEL); diff --git a/paddle/operators/reduce_op.cu b/paddle/operators/reduce_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..d306e1a24096d737438d71d4d4abc35328d160cb --- /dev/null +++ b/paddle/operators/reduce_op.cu @@ -0,0 +1,28 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#define EIGEN_USE_GPU +#include "paddle/operators/reduce_op.h" + +namespace ops = paddle::operators; + +#define REGISTER_REDUCE_GPU_KERNEL(reduce_type, functor, grad_functor) \ + REGISTER_OP_GPU_KERNEL( \ + reduce_type, \ + ops::ReduceKernel); \ + REGISTER_OP_GPU_KERNEL(reduce_type##_grad, \ + ops::ReduceGradKernel); + +FOR_EACH_KERNEL_FUNCTOR(REGISTER_REDUCE_GPU_KERNEL); diff --git a/paddle/operators/reduce_op.h b/paddle/operators/reduce_op.h new file mode 100644 index 0000000000000000000000000000000000000000..45043c440bc8017e97f8be00d08f1cb60d201e20 --- /dev/null +++ b/paddle/operators/reduce_op.h @@ -0,0 +1,206 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once + +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; +using DDim = framework::DDim; +template +using EigenTensor = framework::EigenTensor; + +struct SumFunctor { + template + void operator()(const Place& place, X& x, Y& y, const Dim& dim) { + y.device(place) = x.sum(dim); + } +}; + +struct SumGradFunctor { + template + void operator()(const Place& place, X& x, Y& y, DX& dx, DY& dy, + const Dim& dim, int size) { + dx.device(place) = dy.broadcast(dim); + } +}; + +struct MeanFunctor { + template + void operator()(const Place& place, X& x, Y& y, const Dim& dim) { + y.device(place) = x.mean(dim); + } +}; + +struct MeanGradFunctor { + template + void operator()(const Place& place, X& x, Y& y, DX& dx, DY& dy, + const Dim& dim, int size) { + dx.device(place) = dy.broadcast(dim) / dx.constant(size); + } +}; + +struct MaxFunctor { + template + void operator()(const Place& place, X& x, Y& y, const Dim& dim) { + y.device(place) = x.maximum(dim); + } +}; + +struct MinFunctor { + template + void operator()(const Place& place, X& x, Y& y, const Dim& dim) { + y.device(place) = x.minimum(dim); + } +}; + +struct MaxOrMinGradFunctor { + template + void operator()(const Place& place, X& x, Y& y, DX& dx, DY& dy, + const Dim& dim, int size) { + auto equals = x == y.broadcast(dim); + auto ones = dx.constant(1); + auto zeros = dx.constant(0); + // If there are multiple minimum or maximum elements, the subgradient of + // each is the set [0, 1], and we pass gradient to all of them here. + dx.device(place) = dy.broadcast(dim) * equals.select(ones, zeros); + } +}; + +template +class ReduceKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + int rank = context.Input("X")->dims().size(); + switch (rank) { + case 1: + ReduceCompute<1>(context); + break; + case 2: + ReduceCompute<2>(context); + break; + case 3: + ReduceCompute<3>(context); + break; + case 4: + ReduceCompute<4>(context); + break; + case 5: + ReduceCompute<5>(context); + break; + case 6: + ReduceCompute<6>(context); + break; + } + } + + private: + template + void ReduceCompute(const framework::ExecutionContext& context) const { + auto* input = context.Input("X"); + auto* output = context.Output("Out"); + output->mutable_data(context.GetPlace()); + + auto x = EigenTensor::From(*input); + auto x_rank = static_cast(x.dimensions().size()); + int dim = static_cast(context.Attr("dim")); + if (dim < 0) dim = x_rank + dim; + auto reduce_dim = Eigen::array({{dim}}); + // construct the squeezed output tensor + bool keep_dim = context.Attr("keep_dim"); + DDim dims = output->dims(); + auto dims_vector = vectorize(dims); + if (keep_dim && x_rank > 1) { + dims_vector.erase(dims_vector.begin() + dim); + dims = framework::make_ddim(dims_vector); + } + auto out = EigenTensor < T, D == 1 ? 1 : (D - 1) > ::From(*output, dims); + auto& place = context.GetEigenDevice(); + Functor functor; + functor(place, x, out, reduce_dim); + } +}; + +template +class ReduceGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + int rank = context.Input("X")->dims().size(); + switch (rank) { + case 1: + ReduceGradCompute<1>(context); + break; + case 2: + ReduceGradCompute<2>(context); + break; + case 3: + ReduceGradCompute<3>(context); + break; + case 4: + ReduceGradCompute<4>(context); + break; + case 5: + ReduceGradCompute<5>(context); + break; + case 6: + ReduceGradCompute<6>(context); + break; + } + } + + private: + template + void ReduceGradCompute(const framework::ExecutionContext& context) const { + auto* input0 = context.Input("X"); + auto* input1 = context.Input("Out"); + auto* input2 = context.Input(framework::GradVarName("Out")); + auto* output = context.Output(framework::GradVarName("X")); + + output->mutable_data(context.GetPlace()); + auto x = EigenTensor::From(*input0); + auto x_grad = EigenTensor::From(*output); + auto x_rank = static_cast(x.dimensions().size()); + int dim = static_cast(context.Attr("dim")); + if (dim < 0) dim = x_rank + dim; + DDim dims = input0->dims(); + dims[dim] = 1; + auto x_reduce = EigenTensor::From(*input1, dims); + auto x_reduce_grad = EigenTensor::From(*input2, dims); + + Eigen::array braodcast_dim; + for (size_t i = 0; i < D; ++i) braodcast_dim[i] = 1; + braodcast_dim[dim] = input0->dims()[dim]; + auto& place = context.GetEigenDevice(); + Functor functor; + functor(place, x, x_reduce, x_grad, x_reduce_grad, braodcast_dim, + braodcast_dim[dim]); + } +}; + +} // namespace operators +} // namespace paddle + +#define FOR_EACH_KERNEL_FUNCTOR(__macro) \ + __macro(reduce_sum, SumFunctor, SumGradFunctor); \ + __macro(reduce_mean, MeanFunctor, MeanGradFunctor); \ + __macro(reduce_max, MaxFunctor, MaxOrMinGradFunctor); \ + __macro(reduce_min, MinFunctor, MaxOrMinGradFunctor); diff --git a/paddle/operators/reshape_op.cc b/paddle/operators/reshape_op.cc index ddb93007e21e4d1ae4be3650019c8bc6a680252d..3cd54930a0919b7c6aad624b442eb15167b4c354 100644 --- a/paddle/operators/reshape_op.cc +++ b/paddle/operators/reshape_op.cc @@ -26,14 +26,14 @@ class ReshapeOp : public framework::OperatorWithKernel { : OperatorWithKernel(type, inputs, outputs, attrs) {} protected: - void InferShape(const framework::InferShapeContext &ctx) const override { + void InferShape(framework::InferShapeContext *ctx) const override { // input check - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), - "Input(X) of ReshapeOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"), - "Output(Out) of ReshapeOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of ReshapeOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of ReshapeOp should not be null."); - auto shape = ctx.Attr>("shape"); + auto shape = ctx->Attrs().Get>("shape"); PADDLE_ENFORCE(shape.size() > 0, "Attr(shape) shouldn't be empty."); for (auto dim : shape) { PADDLE_ENFORCE(dim > 0, "Each dimension of shape must be positive."); @@ -41,8 +41,8 @@ class ReshapeOp : public framework::OperatorWithKernel { // capacity check int64_t capacity = std::accumulate(shape.begin(), shape.end(), 1, std::multiplies()); - auto *in = ctx.Input("X"); - int64_t in_size = framework::product(in->dims()); + auto x_dims = ctx->GetInputDim("X"); + int64_t in_size = framework::product(x_dims); PADDLE_ENFORCE_EQ(capacity, in_size, "The size of Input(X) mismatches with Attr(shape)."); // resize output @@ -50,11 +50,11 @@ class ReshapeOp : public framework::OperatorWithKernel { std::transform(shape.begin(), shape.end(), shape_int64.begin(), [](int a) { return static_cast(a); }); auto out_dims = framework::make_ddim(shape_int64); - ctx.Output("Out")->Resize(out_dims); - if (shape[0] == in->dims()[0]) { + ctx->SetOutputDim("Out", out_dims); + if (shape[0] == x_dims[0]) { // Only pass LoD when the first dimension is equal between // output and input. - ctx.ShareLoD("X", /*->*/ "Out"); + ctx->ShareLoD("X", /*->*/ "Out"); } } }; @@ -76,7 +76,7 @@ Given a 2-D tensor X with 2 rows and 2 columns [[1, 2], [3, 4]] -with target shape = [1, 4], the reshape operator will transform +with target shape = [1, 4], the reshape operator will transform the tensor X into a 1-D tensor: [1, 2, 3, 4] @@ -94,13 +94,11 @@ class ReshapeGradOp : public framework::OperatorWithKernel { : OperatorWithKernel(type, inputs, outputs, attrs) {} protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) shouldn't be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")), - "Input(Out@GRAD) shouldn't be null."); - auto dims = ctx.Input("X")->dims(); - auto *d_in = ctx.Output(framework::GradVarName("X")); - d_in->Resize(dims); + void InferShape(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) shouldn't be null."); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), + "Input(Out@GRAD) shouldn't be null."); + ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X")); } }; diff --git a/paddle/operators/reshape_op.h b/paddle/operators/reshape_op.h index 873acf30782d390cdca5e7e864c76e1f743f9a7c..628dfe4c0fadcfeec188d8ae5049a994e3281bc1 100644 --- a/paddle/operators/reshape_op.h +++ b/paddle/operators/reshape_op.h @@ -21,7 +21,7 @@ namespace paddle { namespace operators { template -class ReshapeKernel : public framework::OpKernel { +class ReshapeKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const { auto* out = ctx.Output("Out"); @@ -39,7 +39,7 @@ class ReshapeKernel : public framework::OpKernel { }; template -class ReshapeGradKernel : public framework::OpKernel { +class ReshapeGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const { auto* d_out = ctx.Input(framework::GradVarName("Out")); diff --git a/paddle/operators/rmsprop_op.cc b/paddle/operators/rmsprop_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..ada6f2bc3cd5de0424139144d8c0696158c319d9 --- /dev/null +++ b/paddle/operators/rmsprop_op.cc @@ -0,0 +1,120 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/operators/rmsprop_op.h" + +namespace paddle { +namespace operators { + +class RmspropOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Param"), + "Input(Param) of RmspropOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("MeanSquare"), + "Input(MeanSquare) of RmspropOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("LearningRate"), + "Input(LearningRate) of RmspropOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Grad"), + "Input(Grad) of RmspropOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Moment"), + "Input(Moment) of RmspropOp should not be null."); + + PADDLE_ENFORCE(ctx->HasOutput("ParamOut"), + "Output(param_out) of RmspropOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("MomentOut"), + "Output(Momentum_out) of RmspropOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("MeanSquareOut"), + "Output(MeanSquareOut) of RmspropOp should not be null."); + + auto param_dim = ctx->GetInputDim("Param"); + PADDLE_ENFORCE_EQ( + param_dim, ctx->GetInputDim("Grad"), + "Param and grad input of RmspropOp should have the same dimension."); + PADDLE_ENFORCE_EQ(param_dim, ctx->GetInputDim("Moment"), + "Param and Momentum input of RmspropOp " + "should have the same dimension."); + PADDLE_ENFORCE_EQ(param_dim, ctx->GetInputDim("MeanSquare"), + "Param and Momentum input of RmspropOp " + "should have the same dimension."); + + auto lr_dim = ctx->GetInputDim("LearningRate"); + PADDLE_ENFORCE_EQ(framework::product(lr_dim), 1, + "Learning Rate should be a scalar."); + + ctx->SetOutputDim("ParamOut", param_dim); + ctx->SetOutputDim("MomentOut", param_dim); + ctx->SetOutputDim("MeanSquareOut", param_dim); + } +}; + +class RmspropOpMaker : public framework::OpProtoAndCheckerMaker { + public: + RmspropOpMaker(framework::OpProto *proto, + framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("Param", + "(Tensor, default Tensor) " + "Input parameter value that has to be updated"); + AddInput("MeanSquare", + "(Tensor, default Tensor)" + " The mean square value that gets updated"); + AddInput("LearningRate", + "(Tensor, default Tensor) " + "The learning rate should be a tensor of size 1"); + AddInput("Grad", + "(Tensor, default Tensor) " + "Input gradient of the parameter"); + AddInput("Moment", + "(Tensor, default Tensor) The moment that gets updated"); + + AddOutput("ParamOut", "(Tensor) Output updated parameter value"); + AddOutput("MomentOut", "(Tensor) Output updated moment"); + AddOutput("MeanSquareOut", "(Tensor) Output Mean squared updated value"); + + AddAttr("epsilon", + "(float, default 1e-10) Constant " + "for numerical stability.") + .SetDefault(1.0e-10f); + AddAttr("decay", + "(float, default 0.9) " + "Discounting factor for coming gradient.") + .SetDefault(0.9f); + AddAttr("momentum", "(float, default 0.0) Constant value") + .SetDefault(0.0f); + AddComment(R"DOC( + +RMSprop + +MeanSquareOut = decay * MeanSquare + (1 - decay) * Grad * Grad +MomentOut = momentum * Moment + + LearningRate * Grad / sqrt(MeanSquareOut + epsilon) +ParamOut = Param - MomentOut + +The original slides that proposed RMSprop: Slide 29 of +http://www.cs.toronto.edu/~tijmen/csc321/slides/lecture_slides_lec6.pdf) + +)DOC"); + } +}; +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP_WITHOUT_GRADIENT(rmsprop, ops::RmspropOp, ops::RmspropOpMaker); +REGISTER_OP_CPU_KERNEL(rmsprop, + ops::RmspropOpKernel); diff --git a/paddle/operators/rmsprop_op.cu b/paddle/operators/rmsprop_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..52634a54816bcd5ad0ba82a56f1df95110112265 --- /dev/null +++ b/paddle/operators/rmsprop_op.cu @@ -0,0 +1,20 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#define EIGEN_USE_GPU +#include "paddle/operators/rmsprop_op.h" + +namespace ops = paddle::operators; +REGISTER_OP_GPU_KERNEL(rmsprop, + ops::RmspropOpKernel); diff --git a/paddle/operators/rmsprop_op.h b/paddle/operators/rmsprop_op.h new file mode 100644 index 0000000000000000000000000000000000000000..7bf2129010f994966d79ef11d5cec30159b47068 --- /dev/null +++ b/paddle/operators/rmsprop_op.h @@ -0,0 +1,67 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; +template +using EigenVector = framework::EigenVector; + +template +class RmspropOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* param_out = ctx.Output("ParamOut"); + auto* moment_out = ctx.Output("MomentOut"); + auto* mean_square_out = ctx.Output("MeanSquareOut"); + + auto grad = ctx.Input("Grad"); + + param_out->mutable_data(ctx.GetPlace()); + moment_out->mutable_data(ctx.GetPlace()); + mean_square_out->mutable_data(ctx.GetPlace()); + + float epsilon = ctx.Attr("epsilon"); + float rho = ctx.Attr("decay"); + float momentum = ctx.Attr("momentum"); + + auto p = EigenVector::Flatten(*ctx.Input("Param")); + auto ms = EigenVector::Flatten(*ctx.Input("MeanSquare")); + auto lr = EigenVector::Flatten(*ctx.Input("LearningRate")); + auto g = EigenVector::Flatten(*grad); + auto mom = EigenVector::Flatten(*ctx.Input("Moment")); + + auto p_out = EigenVector::Flatten(*param_out); + auto mom_out = EigenVector::Flatten(*moment_out); + auto ms_out = EigenVector::Flatten(*mean_square_out); + auto place = ctx.GetEigenDevice(); + + Eigen::DSizes grad_dsize(grad->numel()); + + ms_out.device(place) = rho * ms + (1 - rho) * g * g; + mom_out.device(place) = + momentum * mom + + lr.broadcast(grad_dsize) * g / (ms_out + epsilon).sqrt(); + p_out.device(place) = p - mom_out; + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/rnn/recurrent_op_utils.cc b/paddle/operators/rnn/recurrent_op_utils.cc index a767009d2366e20d2ebd35f562b8df7d408f2d4e..ef317a71f12c6de974bd8715bb08122b761fae37 100644 --- a/paddle/operators/rnn/recurrent_op_utils.cc +++ b/paddle/operators/rnn/recurrent_op_utils.cc @@ -25,7 +25,7 @@ using LoDTensor = framework::LoDTensor; void SegmentInputs(const std::vector& step_scopes, const std::vector& inlinks, - const size_t seq_len, bool infer_shape_mode) { + const size_t seq_len) { PADDLE_ENFORCE(!inlinks.empty(), "no in links are provided."); for (size_t i = 0; i < inlinks.size(); ++i) { // global inputs @@ -41,11 +41,9 @@ void SegmentInputs(const std::vector& step_scopes, for (size_t j = 0; j < seq_len; j++) { Tensor* step_input = step_scopes[j]->NewVar(inlinks[i])->GetMutable(); - if (!infer_shape_mode) { - // The input of operators of each step is Tensor here. - // Maybe need to modify Slice function. - *step_input = input->Slice(j, j + 1); - } + // The input of operators of each step is Tensor here. + // Maybe need to modify Slice function. + *step_input = input->Slice(j, j + 1); step_input->Resize(step_dims); } } @@ -53,39 +51,35 @@ void SegmentInputs(const std::vector& step_scopes, void ConcatOutputs(const std::vector& step_scopes, const std::vector& outlinks, - const size_t seq_len, bool infer_shape_mode) { + const size_t seq_len) { for (size_t i = 0; i < outlinks.size(); i++) { - auto output_var = step_scopes[0]->parent().FindVar(outlinks[i]); + auto* output_var = step_scopes[0]->parent().FindVar(outlinks[i]); PADDLE_ENFORCE_NOT_NULL(output_var, "output link [%s] is not in scope.", outlinks[i]); LoDTensor* output = output_var->GetMutable(); - if (infer_shape_mode) { - auto step_scope_var = step_scopes[0]->FindVar(outlinks[i]); - PADDLE_ENFORCE_NOT_NULL(step_scope_var, "%s not in scope", outlinks[i]); - f::DDim step_dims = - step_scope_var->template GetMutable()->dims(); - std::vector dims_vec = vectorize(step_dims); - dims_vec.insert(dims_vec.begin(), seq_len); - output->Resize(f::make_ddim(dims_vec)); - } else { - output->mutable_data(platform::CPUPlace()); - for (size_t j = 0; j < seq_len; j++) { - LoDTensor* step_output = - step_scopes[j]->FindVar(outlinks[i])->GetMutable(); - // TODO(luotao02) data type and platform::DeviceContext() should set - // correctly - (output->Slice(j, j + 1)) - .CopyFrom(*step_output, platform::CPUPlace()); - } + auto* step_scope_var = step_scopes[0]->FindVar(outlinks[i]); + PADDLE_ENFORCE_NOT_NULL(step_scope_var, "%s not in scope", outlinks[i]); + f::DDim step_dims = + step_scope_var->template GetMutable()->dims(); + std::vector dims_vec = vectorize(step_dims); + dims_vec.insert(dims_vec.begin(), seq_len); + output->Resize(f::make_ddim(dims_vec)); + output->mutable_data(platform::CPUPlace()); + for (size_t j = 0; j < seq_len; j++) { + LoDTensor* step_output = + step_scopes[j]->FindVar(outlinks[i])->GetMutable(); + // TODO(luotao02) data type and platform::DeviceContext() should set + // correctly + (output->Slice(j, j + 1)) + .CopyFrom(*step_output, platform::CPUPlace()); } } } void LinkMemories(const std::vector& scopes, const std::vector& memories, - const size_t step_id, const int offset, - bool infer_shape_mode) { + const size_t step_id, const int offset) { PADDLE_ENFORCE_LT(step_id, scopes.size(), "step [%d] is out of range of step scopes' size [%d]", step_id, scopes.size()); @@ -95,16 +89,13 @@ void LinkMemories(const std::vector& scopes, step_id + offset, scopes.size(), "offset [%d] is out of range, it must be less than (%d - %d)", offset, scopes.size(), step_id); - auto scope = scopes[step_id]; - auto linked_scope = scopes[step_id + offset]; + auto* scope = scopes[step_id]; + auto* linked_scope = scopes[step_id + offset]; for (auto& attr : memories) { - auto mem = scope->FindVar(attr.pre_var)->GetMutable(); - auto linked_mem = linked_scope->FindVar(attr.var)->GetMutable(); - if (infer_shape_mode) { - mem->Resize(linked_mem->dims()); - } else { - mem->ShareDataWith(*linked_mem); - } + auto* mem = scope->FindVar(attr.pre_var)->GetMutable(); + auto* linked_mem = linked_scope->FindVar(attr.var)->GetMutable(); + mem->Resize(linked_mem->dims()); + mem->ShareDataWith(*linked_mem); } } @@ -115,11 +106,11 @@ void InitArgument(const ArgumentName& name, Argument* arg, arg->inlinks = op.Inputs(name.inlinks); arg->outlinks = op.Outputs(name.outlinks); - auto boot_memories = + auto& boot_memories = is_grad ? op.Outputs(name.boot_memories) : op.Inputs(name.boot_memories); // attributes - auto memories = op.Attr>(name.memories); - auto pre_memories = op.Attr>(name.pre_memories); + auto& memories = op.Attr>(name.memories); + auto& pre_memories = op.Attr>(name.pre_memories); PADDLE_ENFORCE(memories.size() == boot_memories.size(), "the size of memories, boot_memories don't match:%d,%d", diff --git a/paddle/operators/rnn/recurrent_op_utils.h b/paddle/operators/rnn/recurrent_op_utils.h index 9c777f1e9067a3e2ceb9d23f7bf7d3c73343c91f..fd17b9b88915cf458ff2836b5c5d8f84cd9b65b5 100644 --- a/paddle/operators/rnn/recurrent_op_utils.h +++ b/paddle/operators/rnn/recurrent_op_utils.h @@ -64,18 +64,18 @@ struct ArgumentName { */ void SegmentInputs(const std::vector& step_scopes, const std::vector& inlinks, - const size_t seq_len, bool infer_shape_mode); + const size_t seq_len); /** * Process outputs of step nets and merge to variables. */ void ConcatOutputs(const std::vector& step_scopes, const std::vector& outlinks, - const size_t seq_len, bool infer_shape_mode); + const size_t seq_len); void LinkMemories(const std::vector& step_scopes, const std::vector& memories, const size_t step_id, - const int offset, bool infer_shape_mode); + const int offset); void InitArgument(const ArgumentName& name, Argument* arg, const framework::OperatorBase& op, bool is_grad = false); diff --git a/paddle/operators/rowwise_add_op.cc b/paddle/operators/rowwise_add_op.cc deleted file mode 100644 index fc3ad721f210213491617452141dfa8834b067c0..0000000000000000000000000000000000000000 --- a/paddle/operators/rowwise_add_op.cc +++ /dev/null @@ -1,104 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - - Licensed under the Apache License, Version 2.0 (the "License"); - you may not use this file except in compliance with the License. - You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - - Unless required by applicable law or agreed to in writing, software - distributed under the License is distributed on an "AS IS" BASIS, - WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - See the License for the specific language governing permissions and - limitations under the License. */ - -#include "paddle/operators/rowwise_add_op.h" - -namespace paddle { -namespace operators { - -using framework::Tensor; - -class RowwiseAddOp : public framework::OperatorWithKernel { - public: - using framework::OperatorWithKernel::OperatorWithKernel; - - protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), - "Input(X) of RowwiseAddOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("b"), - "Input(b) of RowwiseAddOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"), - "Output(Out) of RowwiseAddOp should not be null."); - - auto x_dims = ctx.Input("X")->dims(); - auto b_dims = ctx.Input("b")->dims(); - PADDLE_ENFORCE_GT( - x_dims.size(), b_dims.size(), - "The rank of input `X` must be larger than the one of input `b`."); - - int num_col_dims = x_dims.size() - b_dims.size(); - - PADDLE_ENFORCE_EQ( - framework::slice_ddim(x_dims, num_col_dims, x_dims.size()), b_dims, - "The width of two operands must be same"); - PADDLE_ENFORCE_EQ(ctx.OutputSize("Out"), 1, "The output size must be 1"); - ctx.Output("Out")->Resize(x_dims); - ctx.ShareLoD("X", /*->*/ "Out"); - } -}; - -class RowwiseAddOpMaker : public framework::OpProtoAndCheckerMaker { - public: - RowwiseAddOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("X", "The left input of row-wise add op, must be matrix"); - AddInput("b", "The right input of row-wise add op, must be vector"); - AddOutput("Out", "The output of row-wise add op"); - AddComment(R"DOC(Row-wise Add operator - -for i in xrange(X.shape[0]): - Out = X[i] + b -)DOC"); - } -}; -class RowwiseAddGradOp : public framework::OperatorWithKernel { - public: - using framework::OperatorWithKernel::OperatorWithKernel; - - protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "X should not be null"); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("b"), "b should not be null"); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")), - "Input(Out@GRAD) should not be null"); - auto x_dims = ctx.Input("X")->dims(); - auto b_dims = ctx.Input("b")->dims(); - PADDLE_ENFORCE_GT( - x_dims.size(), b_dims.size(), - "The rank of input `X` must be larger than the one of input `b`."); - - int num_col_dims = x_dims.size() - b_dims.size(); - PADDLE_ENFORCE_EQ( - framework::slice_ddim(x_dims, num_col_dims, x_dims.size()), b_dims, - "The width of two operands must be same"); - auto *dx = ctx.Output(framework::GradVarName("X")); - auto *db = ctx.Output(framework::GradVarName("b")); - if (dx) dx->Resize(x_dims); - if (db) db->Resize(b_dims); - } -}; - -} // namespace operators -} // namespace paddle - -namespace ops = paddle::operators; -REGISTER_OP(rowwise_add, ops::RowwiseAddOp, ops::RowwiseAddOpMaker, - rowwise_add_grad, ops::RowwiseAddGradOp); -REGISTER_OP_CPU_KERNEL( - rowwise_add, ops::RowwiseAddKernel); -REGISTER_OP_CPU_KERNEL( - rowwise_add_grad, - ops::RowwiseAddGradKernel); diff --git a/paddle/operators/rowwise_add_op.h b/paddle/operators/rowwise_add_op.h deleted file mode 100644 index 35774b940926f77167b8f19597027e74d3477e5b..0000000000000000000000000000000000000000 --- a/paddle/operators/rowwise_add_op.h +++ /dev/null @@ -1,80 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once -#include "paddle/framework/eigen.h" -#include "paddle/framework/op_registry.h" - -namespace paddle { -namespace operators { - -using Tensor = framework::Tensor; -template -using EigenVector = framework::EigenVector; -template -using EigenMatrix = framework::EigenMatrix; - -template -class RowwiseAddKernel : public framework::OpKernel { - public: - void Compute(const framework::ExecutionContext& context) const override { - auto out = context.Output("Out"); - out->mutable_data(context.GetPlace()); - int num_col_dims = context.Input("X")->dims().size() - - context.Input("b")->dims().size(); - auto input = - EigenMatrix::Reshape(*context.Input("X"), num_col_dims); - auto bias = EigenVector::Flatten(*context.Input("b")); - auto output = EigenMatrix::Reshape(*out, num_col_dims); - - const int bias_size = bias.dimension(0); - const int rest_size = input.size() / bias_size; - Eigen::DSizes one_d(input.size()); - Eigen::DSizes bcast(rest_size); - output.reshape(one_d).device(context.GetEigenDevice()) = - input.reshape(one_d) + bias.broadcast(bcast).reshape(one_d); - } -}; - -template -class RowwiseAddGradKernel : public framework::OpKernel { - public: - void Compute(const framework::ExecutionContext& context) const override { - auto* dout = context.Input(framework::GradVarName("Out")); - auto* dx = context.Output(framework::GradVarName("X")); - auto* db = context.Output(framework::GradVarName("b")); - int num_col_dims = context.Input("X")->dims().size() - - context.Input("b")->dims().size(); - - auto out_grad = EigenMatrix::Reshape(*dout, num_col_dims); - auto place = context.GetEigenDevice(); - - if (dx) { - dx->mutable_data(context.GetPlace()); - EigenMatrix::Reshape(*dx, num_col_dims).device(place) = out_grad; - } - - if (db) { - db->mutable_data(context.GetPlace()); - // https://eigen.tuxfamily.org/dox/unsupported/TensorBase_8h_source.html - // colwise add - Eigen::array dims{{0}}; /* dimension to reduce */ - EigenVector::Flatten(*db).device(place) = out_grad.sum(dims); - } - } -}; -} // namespace operators -} // namespace paddle diff --git a/paddle/operators/scale_op.cc b/paddle/operators/scale_op.cc index 1ae77a9722ef1a5548a6c4100c32fdddcee8c5cd..ac297da6b72250eaa5eb6ea2177fbcc7a6c1ec56 100644 --- a/paddle/operators/scale_op.cc +++ b/paddle/operators/scale_op.cc @@ -26,16 +26,13 @@ class ScaleOp : public framework::OperatorWithKernel { : OperatorWithKernel(type, inputs, outputs, attrs) {} protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), - "Input(X) of ScaleOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"), - "Output(Out) of ScaleOp should not be null."); - - auto *in = ctx.Input("X"); - auto *out = ctx.Output("Out"); - out->Resize(in->dims()); - ctx.ShareLoD("X", /*->*/ "Out"); + void InferShape(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of ScaleOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of ScaleOp should not be null."); + ctx->SetOutputDim("Out", ctx->GetInputDim("X")); + ctx->ShareLoD("X", /*->*/ "Out"); } }; @@ -44,8 +41,8 @@ class ScaleOpMaker : public framework::OpProtoAndCheckerMaker { public: ScaleOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("X", "The input tensor of scale operator.").NotInGradient(); - AddOutput("Out", "The output tensor of scale operator.").NotInGradient(); + AddInput("X", "The input tensor of scale operator."); + AddOutput("Out", "The output tensor of scale operator."); AddComment(R"DOC(Scale operator The equation is: Out = scale*X @@ -55,21 +52,18 @@ The equation is: Out = scale*X } }; -// The operator to calculate gradients of a scale operator is just the scale -// operator itself. -// Grad(Out=scale(X)) => Grad(X) = scale(Grad(Out)) -template -class ScaleGradOp : public NetOp { +class ScaleGradMaker : public framework::SingleGradOpDescMaker { public: - ScaleGradOp(const std::string &type, const framework::VariableNameMap &inputs, - const framework::VariableNameMap &outputs, - const framework::AttributeMap &attrs) - : NetOp(type, inputs, outputs, attrs) { - AppendOp(framework::OpRegistry::CreateOp( - "scale", {{"X", {Input(framework::GradVarName("Out"))}}}, - {{"Out", {Output(framework::GradVarName("X"))}}}, - {{"scale", Attr("scale")}})); - CompleteAddOp(false); + using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; + + protected: + std::unique_ptr Apply() const override { + auto *grad_op = new framework::OpDescBind(); + grad_op->SetType("scale"); + grad_op->SetInput("X", OutputGrad("Out")); + grad_op->SetOutput("Out", InputGrad("X")); + grad_op->SetAttr("scale", GetAttr("scale")); + return std::unique_ptr(grad_op); } }; @@ -78,7 +72,7 @@ class ScaleGradOp : public NetOp { namespace ops = paddle::operators; -REGISTER_OP(scale, ops::ScaleOp, ops::ScaleOpMaker, scale_grad, - ops::ScaleGradOp); +REGISTER_OPERATOR(scale, ops::ScaleOp, ops::ScaleOpMaker, + ops::ScaleGradMaker); REGISTER_OP_CPU_KERNEL(scale, ops::ScaleKernel); diff --git a/paddle/operators/scale_op.h b/paddle/operators/scale_op.h index 02fbdc52bbf89c9f2acc5eeaa1197e4ccbca9d31..dc6bc768997f4fdd049bb63bdc11252ab52fcda9 100644 --- a/paddle/operators/scale_op.h +++ b/paddle/operators/scale_op.h @@ -20,7 +20,7 @@ namespace paddle { namespace operators { template -class ScaleKernel : public framework::OpKernel { +class ScaleKernel : public framework::OpKernel { public: virtual void Compute(const framework::ExecutionContext& context) const { auto* tensor = context.Output("Out"); diff --git a/paddle/operators/scatter.cu.h b/paddle/operators/scatter.cu.h new file mode 100644 index 0000000000000000000000000000000000000000..d95436be4f25b9df4aaef57ddb249ecf944f0666 --- /dev/null +++ b/paddle/operators/scatter.cu.h @@ -0,0 +1,80 @@ +/* Copyright (c) 2016 PaddlePaddle Authors All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once +#include "paddle/framework/tensor.h" +#include "paddle/platform/place.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; + +#define CUDA_1D_KERNEL_LOOP(i, n) \ + for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < (n); \ + i += blockDim.x * gridDim.x) + +template +__global__ void ScatterCUDAKernel(const T* params, const int* indices, + T* output, size_t index_size, + size_t slice_size) { + CUDA_1D_KERNEL_LOOP(i, index_size * slice_size) { + int indices_i = i / slice_size; + int slice_i = i - indices_i * slice_size; // offset inside the slice + int scatter_i = indices[indices_i]; + int out_i = scatter_i * slice_size + slice_i; + *(output + out_i) = *(params + i); + } +} + +/** + * A thin wrapper on gpu tensor + * Return a new updated tensor from source tensor, scatter-assigned according to + * index + * input[src]: type-T source Tensor + * input[index]: type-int index Tensor (1-D) + * return: output tensor + */ +template +void GPUScatterAssign(const platform::DeviceContext& ctx, const Tensor& src, + const Tensor& index, Tensor* output) { + // PADDLE_ENFORCE(platform::is_gpu_place(place)); + // check index of shape 1-D + PADDLE_ENFORCE(index.dims().size() == 1); + int index_size = index.dims()[0]; + + auto src_dims = src.dims(); + framework::DDim output_dims(src_dims); + output_dims[0] = index_size; + + // slice size + int slice_size = 1; + for (int i = 1; i < src_dims.size(); ++i) slice_size *= src_dims[i]; + + const T* p_src = src.data(); + const int* p_index = index.data(); + T* p_output = output->data(); + + int block = 512; + int n = slice_size * index_size; + int grid = (n + block - 1) / block; + + ScatterCUDAKernel<<< + grid, block, 0, + reinterpret_cast(ctx).stream()>>>( + p_src, p_index, p_output, index_size, slice_size); +} + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/scatter.h b/paddle/operators/scatter.h index 6b542675c291607b35f180123cf42fee6a783a85..c1fb844ebd2ff7ca7dbdb8e8ac3c1fff4c0c6607 100644 --- a/paddle/operators/scatter.h +++ b/paddle/operators/scatter.h @@ -24,67 +24,42 @@ namespace paddle { namespace operators { using Tensor = framework::Tensor; -template -using EigenVector = framework::EigenVector; - -// Implementation of CPU copy -template -void CPUScatterUpdate(const paddle::framework::Tensor* src, const int* index, - const size_t index_size, - paddle::framework::Tensor* output) { - paddle::framework::DDim output_dims = output->dims(); - - for (size_t i = 0; i < index_size; ++i) { - int index_ = index[i]; - - paddle::framework::Tensor src_ = *src; - paddle::framework::Tensor output_ = *output; - if (index_size > 1) src_ = src->Slice(i, i + 1); - if (output_dims[0] > 1) output_ = output->Slice(index_, index_ + 1); - - auto X = EigenVector::Flatten(src_); - auto Y = EigenVector::Flatten(output_); - - Y = X + Y; - } -} - -// Implementation of GPU scatter: -template -void GPUScatterUpdate(const T* src, const int* index, const int slice_size, - const int index_size, T* output); /** * Return a updated tensor from source tensor, scattered according to index: - * dst[i] += src[index[i]] + * dst[i] = src[index[i]] * input[src]: type-T source Tensor * input[index]: type-int index Tensor (1-D) * return: output tensor */ template -void ScatterUpdate(const platform::Place& place, - const paddle::framework::Tensor* src, - const paddle::framework::Tensor* index, - paddle::framework::Tensor* output) { +void ScatterAssign(const platform::DeviceContext& ctx, const Tensor& src, + const Tensor& index, Tensor* output) { + PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace())); // check index of shape 1-D - PADDLE_ENFORCE(index->dims().size() == 1); - int index_size = index->dims()[0]; + PADDLE_ENFORCE(index.dims().size() == 1); + int index_size = index.dims()[0]; - auto src_dims = src->dims(); + auto src_dims = src.dims(); auto dst_dims = output->dims(); + const T* p_src = src.data(); + const int* p_index = index.data(); + T* p_output = output->data(); + // check src shape and dst shape should match for (int i = 1; i < src_dims.size(); i++) PADDLE_ENFORCE(src_dims[i] == dst_dims[i]); // slice size size_t slice_size = 1; - for (int i = 0; i < src_dims.size(); ++i) slice_size *= src_dims[i]; + for (int i = 1; i < src_dims.size(); ++i) slice_size *= src_dims[i]; + + const size_t slice_bytes = slice_size * sizeof(T); - if (platform::is_cpu_place(place)) { - CPUScatterUpdate(src, index->data(), index_size, output); - } else { + for (int i = 0; i < index_size; ++i) { + int index_ = p_index[i]; + memcpy(p_output + index_ * slice_size, p_src + i * slice_size, slice_bytes); } } diff --git a/paddle/operators/scatter_op.cc b/paddle/operators/scatter_op.cc index 3f02081a060281dec533c02b346f0667da28b8c3..fbea01a8db007a5b8cfe17f545f1cce6ecff3afb 100644 --- a/paddle/operators/scatter_op.cc +++ b/paddle/operators/scatter_op.cc @@ -23,29 +23,35 @@ class ScatterOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Ref"), - "Input(Ref) of ScatterOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Index"), - "Input(Index) of ScatterOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Updates"), - "Input(Updates) of ScatterOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"), - "Output(Out) of ScatterOp should not be null."); - - PADDLE_ENFORCE_EQ(ctx.Input("Index")->dims().size(), 1, + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Ref"), + "Input(Ref) of ScatterOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Index"), + "Input(Index) of ScatterOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Updates"), + "Input(Updates) of ScatterOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of ScatterOp should not be null."); + + auto updates_dims = ctx->GetInputDim("Updates"); + auto ref_dims = ctx->GetInputDim("Ref"); + PADDLE_ENFORCE_EQ(ctx->GetInputDim("Index").size(), 1, "Update Index should be 1-D."); - PADDLE_ENFORCE_EQ(ctx.Input("Ref")->dims().size(), - ctx.Input("Updates")->dims().size(), + PADDLE_ENFORCE_EQ(ref_dims.size(), updates_dims.size(), "Reference and Updates should have the same shape size"); - PADDLE_ENFORCE_EQ(ctx.Input("Updates")->dims()[0], - ctx.Input("Index")->dims()[0], + PADDLE_ENFORCE_EQ(ctx->GetInputDim("Updates")[0], + ctx->GetInputDim("Index")[0], "Updates and Index should have same batch-size."); - framework::DDim data_dim(ctx.Input("Updates")->dims()); - for (int i = 1; i < data_dim.size(); ++i) - PADDLE_ENFORCE_EQ(data_dim[i], ctx.Input("Updates")->dims()[i]); - ctx.Output("Out")->Resize( - ctx.Input("Ref")->dims()); + framework::DDim data_dim(updates_dims); + for (int i = 1; i < data_dim.size(); ++i) { + PADDLE_ENFORCE_EQ(data_dim[i], updates_dims[i]); + } + ctx->SetOutputDim("Out", ref_dims); + } + + framework::DataType IndicateDataType( + const framework::ExecutionContext& ctx) const override { + return framework::ToDataType(ctx.Input("Ref")->type()); } }; @@ -54,22 +60,22 @@ class ScatterGradOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - auto *dUpdates = - ctx.Output(framework::GradVarName("Updates")); - auto *Updates = ctx.Input("Updates"); - auto *dRef = ctx.Output(framework::GradVarName("Ref")); - auto *Ref = ctx.Input("Ref"); - - dRef->Resize(Ref->dims()); - dUpdates->Resize(Updates->dims()); + void InferShape(framework::InferShapeContext* ctx) const override { + ctx->SetOutputDim(framework::GradVarName("Updates"), + ctx->GetInputDim("Updates")); + ctx->SetOutputDim(framework::GradVarName("Ref"), ctx->GetInputDim("Ref")); + } + + framework::DataType IndicateDataType( + const framework::ExecutionContext& ctx) const override { + return framework::ToDataType(ctx.Input("Ref")->type()); } }; class ScatterOpMaker : public framework::OpProtoAndCheckerMaker { public: - ScatterOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + ScatterOpMaker(framework::OpProto* proto, + framework::OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("Ref", "The source input of scatter op"); AddInput("Index", @@ -77,21 +83,19 @@ class ScatterOpMaker : public framework::OpProtoAndCheckerMaker { AddInput("Updates", "The updated value of updates op"); AddOutput("Out", "The output of add op"); AddComment(R"DOC( -Scatter Operator by selecting from the first axis, +Scatter Operator by selecting from the first axis, Out = Ref Out[Index] = Ref[Index] + Updates )DOC"); } }; + } // namespace operators } // namespace paddle namespace ops = paddle::operators; REGISTER_OP(scatter, ops::ScatterOp, ops::ScatterOpMaker, scatter_grad, ops::ScatterGradOp); -REGISTER_OP_CPU_KERNEL(scatter, - ops::ScatterOpKernel); -REGISTER_OP_CPU_KERNEL( - scatter_grad, - ops::ScatterGradientOpKernel); +REGISTER_OP_CPU_KERNEL(scatter, ops::ScatterOpKernel); +REGISTER_OP_CPU_KERNEL(scatter_grad, ops::ScatterGradientOpKernel); diff --git a/paddle/operators/scatter_op.cu b/paddle/operators/scatter_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..06f4d759447b6dcd28b50576dfc246fc466d9336 --- /dev/null +++ b/paddle/operators/scatter_op.cu @@ -0,0 +1,63 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "gather.cu.h" +#include "paddle/operators/gather_op.h" +#include "scatter.cu.h" + +namespace paddle { +namespace operators { + +template +class ScatterOpCUDAKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &ctx) const override { + PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), + "This kernel only runs on GPU device."); + auto *Ref = ctx.Input("Ref"); + auto *Index = ctx.Input("Index"); + auto *Updates = ctx.Input("Updates"); + auto *Out = ctx.Output("Out"); + + Out->ShareDataWith(*Ref); + + GPUScatterAssign(ctx.device_context(), *Updates, *Index, Out); + } +}; + +template +class ScatterGradOpCUDAKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &ctx) const override { + PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), + "This kernel only runs on GPU device."); + auto *dRef = ctx.Output(framework::GradVarName("Ref")); + auto *dUpdates = ctx.Output(framework::GradVarName("Updates")); + auto *Index = ctx.Input("Index"); + auto *dOut = ctx.Input(framework::GradVarName("Out")); + + // In place gradient: dRef = dO + dRef->ShareDataWith(*dOut); + dUpdates->mutable_data(ctx.GetPlace()); + // Gradient by Gather: dUpdates = dO[Index] + GPUGather(ctx.device_context(), *dOut, *Index, dUpdates); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP_GPU_KERNEL(scatter, ops::ScatterOpCUDAKernel); +REGISTER_OP_GPU_KERNEL(scatter_grad, ops::ScatterGradOpCUDAKernel); diff --git a/paddle/operators/scatter_op.h b/paddle/operators/scatter_op.h index e9595638a86a4a4536ddad4e6f20fd80a54b1608..6101219006414e4865f676e3ca5d2a88949ad17a 100644 --- a/paddle/operators/scatter_op.h +++ b/paddle/operators/scatter_op.h @@ -23,10 +23,12 @@ namespace operators { using Tensor = framework::Tensor; -template -class ScatterOpKernel : public framework::OpKernel { +template +class ScatterOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext &ctx) const override { + PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()), + "This kernel only runs on CPU."); auto *Ref = ctx.Input("Ref"); auto *Index = ctx.Input("Index"); auto *Updates = ctx.Input("Updates"); @@ -35,14 +37,16 @@ class ScatterOpKernel : public framework::OpKernel { // In place output: Out = Ref, Out[Index] += Updates Out->ShareDataWith(*Ref); // Apply ScatterUpdate: Out[index] += Updates[:] - ScatterUpdate(ctx.GetPlace(), Updates, Index, Out); + ScatterAssign(ctx.device_context(), *Updates, *Index, Out); } }; -template -class ScatterGradientOpKernel : public framework::OpKernel { +template +class ScatterGradientOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext &ctx) const override { + PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()), + "This kernel only runs on CPU."); auto *dRef = ctx.Output(framework::GradVarName("Ref")); auto *dUpdates = ctx.Output(framework::GradVarName("Updates")); auto *Index = ctx.Input("Index"); @@ -52,7 +56,7 @@ class ScatterGradientOpKernel : public framework::OpKernel { dRef->ShareDataWith(*dOut); dUpdates->mutable_data(ctx.GetPlace()); // Gradient by Gather: dUpdates += dO[Index] - Gather(ctx.GetPlace(), dOut, Index, dUpdates); + CPUGather(ctx.device_context(), *dOut, *Index, dUpdates); } }; diff --git a/paddle/operators/scatter_test.cc b/paddle/operators/scatter_test.cc index 26fdaff1460a297fa638181641991f732533fe52..00dbdacbfef7af826790472acc6caa285c259e0e 100644 --- a/paddle/operators/scatter_test.cc +++ b/paddle/operators/scatter_test.cc @@ -40,7 +40,9 @@ TEST(scatter, ScatterUpdate) { float* p_output = output->mutable_data(make_ddim({4, 4}), CPUPlace()); - ScatterUpdate(CPUPlace(), src, index, output); + auto* cpu_place = new paddle::platform::CPUPlace(); + paddle::platform::CPUDeviceContext ctx(*cpu_place); + ScatterAssign(ctx, *src, *index, output); for (size_t i = 0; i < 4; ++i) EXPECT_EQ(p_output[i], float(0)); for (size_t i = 0; i < 4; ++i) EXPECT_EQ(output->data()[i], float(0)); diff --git a/paddle/operators/sequence_pool_op.cc b/paddle/operators/sequence_pool_op.cc index 73f9cb879a2ef690909428b3b672b12717a6a02c..06c00d31eac9ff09084db951ce2929850f9d14c2 100644 --- a/paddle/operators/sequence_pool_op.cc +++ b/paddle/operators/sequence_pool_op.cc @@ -22,23 +22,12 @@ class SequencePoolOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext& ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), - "Input(X) of SequencePoolOp should not be null."); - PADDLE_ENFORCE_NOT_NULL( - ctx.OutputVar("Out"), - "Output(Out) of SequencePoolOp should not be null."); - - auto* x = ctx.Input("X"); - auto dims = x->dims(); - auto lod = x->lod(); - PADDLE_ENFORCE_EQ(lod.size(), 1UL, "Only support one level sequence now."); - PADDLE_ENFORCE_GE( - dims[0], - /*batch size = */ static_cast(lod[0].size() - 1), - "The first dimension of Input(X) must be large than batch size."); - dims[0] = lod[0].size() - 1; - ctx.Output("Out")->Resize({dims}); + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of SequencePoolOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of SequencePoolOp should not be null."); + ctx->SetOutputDim("Out", ctx->GetInputDim("X")); } }; @@ -61,17 +50,17 @@ class SequencePoolOpMaker : public framework::OpProtoAndCheckerMaker { SequencePoolOp pools features of all time-steps of each instance. For a mini-batch of 3 variable lengths sentences, containing 2, 3, and 2 time-steps: - + Assume X is a [7,M,N] float LoDTensor, and X->lod()[0] = [0, 2, 5, 7]. - Besides, for the sake of simplicity, we assume M=1 and N=1, + Besides, for the sake of simplicity, we assume M=1 and N=1, and the value of X = [[1, 3], [2, 4, 6], [5, 1]]. Thus, Out is a [3,1,1] float LoDTensor, but Out->lod() is nullptr. - And for different strategy, the value of Out is as follows: + And for different strategy, the value of Out is as follows: - AVERAGE: [2, 4, 3], where 2=(1+3)/2, 4=(2+4+6)/3, 3=(5+1)/2 - SUM: [4, 12, 6], where 4=1+3, 12=2+4+6, 6=5+1 - - SQRT: [2.82, 6.93, 4.24], where 2.82=(1+3)/sqrt(2), + - SQRT: [2.82, 6.93, 4.24], where 2.82=(1+3)/sqrt(2), 6.93=(2+4+6)/sqrt(3), 4.24=(5+1)/sqrt(2) - MAX: [3, 6, 5], where 3=max(1,3), 6=max(2,4,6), 5=max(5,1) - LAST: [3, 6, 1], where 3=last(1,3), 6=last(2,4,6), 1=last(5,1) @@ -85,22 +74,18 @@ class SequencePoolGradOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext& ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")), - "Gradient of Out should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), - "The input X should not be null."); - auto og_dims = - ctx.Input(framework::GradVarName("Out"))->dims(); - auto x_dims = ctx.Input("X")->dims(); + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), + "Gradient of Out should not be null."); + PADDLE_ENFORCE(ctx->HasInput("X"), "The input X should not be null."); + auto og_dims = ctx->GetInputDim(framework::GradVarName("Out")); + auto x_dims = ctx->GetInputDim("X"); PADDLE_ENFORCE_EQ(og_dims.size(), x_dims.size(), "The rank of output grad must equal to Input(X)."); for (int64_t i = 1; i < og_dims.size(); ++i) { PADDLE_ENFORCE_EQ(og_dims[i], x_dims[i], "The dimension mismatch."); } - auto* x_grad = - ctx.Output(framework::GradVarName("X")); - x_grad->Resize(x_dims); + ctx->SetOutputDim(framework::GradVarName("X"), x_dims); } }; diff --git a/paddle/operators/sequence_pool_op.h b/paddle/operators/sequence_pool_op.h index 231614b4c1cb0eb1901b1720e933aed5cbb25f77..752d714125578b2d1f926765b183495ec5cc203e 100644 --- a/paddle/operators/sequence_pool_op.h +++ b/paddle/operators/sequence_pool_op.h @@ -38,7 +38,7 @@ enum SeqPoolType { }; template -class SequencePoolKernel : public framework::OpKernel { +class SequencePoolKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto* in = context.Input("X"); @@ -46,16 +46,27 @@ class SequencePoolKernel : public framework::OpKernel { int strategy = context.Attr("strategy"); auto dims = in->dims(); - auto lod = in->lod()[0]; + auto lod = in->lod(); int64_t w = in->numel() / dims[0]; + // InferShape by lod + PADDLE_ENFORCE_EQ(lod.size(), 1UL, "Only support one level sequence now."); + PADDLE_ENFORCE_GE( + dims[0], + /*batch size = */ static_cast(lod[0].size() - 1), + "The first dimension of Input(X) must be large than batch size."); + dims[0] = lod[0].size() - 1; + out->Resize({dims}); + + auto lod_level_0 = lod[0]; + out->mutable_data(context.GetPlace()); auto place = context.GetEigenDevice(); - for (int i = 0; i < static_cast(lod.size()) - 1; ++i) { - Tensor in_t = - in->Slice(static_cast(lod[i]), static_cast(lod[i + 1])); + for (int i = 0; i < static_cast(lod_level_0.size()) - 1; ++i) { + Tensor in_t = in->Slice(static_cast(lod_level_0[i]), + static_cast(lod_level_0[i + 1])); Tensor out_t = out->Slice(i, i + 1); - int64_t h = static_cast(lod[i + 1] - lod[i]); + int64_t h = static_cast(lod_level_0[i + 1] - lod_level_0[i]); auto in_e = EigenMatrix::From(in_t, framework::make_ddim({h, w})); auto out_e = EigenVector::Flatten(out_t); @@ -74,7 +85,7 @@ class SequencePoolKernel : public framework::OpKernel { }; template -class SequencePoolGradKernel : public framework::OpKernel { +class SequencePoolGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto* in = context.Input("X"); diff --git a/paddle/operators/sequence_softmax_op.cc b/paddle/operators/sequence_softmax_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..ea217ba4597c24ec323c6509f097874d977fcf81 --- /dev/null +++ b/paddle/operators/sequence_softmax_op.cc @@ -0,0 +1,103 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/operators/sequence_softmax_op.h" + +namespace paddle { +namespace operators { + +class SequenceSoftmaxOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of SequenceSoftmaxOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of SequenceSoftmaxOp should not be null."); + ctx->SetOutputDim("Out", ctx->GetInputDim("X")); + ctx->ShareLoD("X", /*->*/ "Out"); + } +}; + +class SequenceSoftmaxOpMaker : public framework::OpProtoAndCheckerMaker { + public: + SequenceSoftmaxOpMaker(framework::OpProto* proto, + framework::OpAttrChecker* op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", + "(LoDTensor) 1-D or 2-D input LoDTensor with the 2-nd dimension " + "of length 1."); + AddOutput("Out", + "(LoDTensor) 1-D or 2-D output LoDTensor with the 2-nd dimension " + "of length 1."); + AddComment(R"DOC( +SequenceSoftmaxOp computes softmax activation among all time-steps for each +sequence. The dimension of each time-step should be 1. Thus, the shape of +input Tensor can be either [N, 1] or [N], where N is the sum of all sequences' +lengths. + +Equation: + for i-th sequence in a mini-batch: + Out(X[lod[i]:lod[i+1]], :) = + exp(X[lod[i]:lod[i+1], :]) / sum(exp(X[lod[i]:lod[i+1], :])) + +For example, for a mini-batch of 3 sequences with variable-length, +each containing 2, 3, 2 time-steps, the lod of which is [0, 2, 5, 7], +then softmax will be computed among X[0:2, :], X[2:5, :], X[5:7, :] +and N turns out to be 7. +)DOC"); + } +}; + +class SequenceSoftmaxGradOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Out"), + "Input(Out) of SequenceSoftmaxGradOp should not be null."); + PADDLE_ENFORCE( + ctx->HasInput(framework::GradVarName("Out")), + "Input(Out@GRAD) of SequenceSoftmaxGradOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of SequenceSoftmaxOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("X")), + "Output(X@GRAD) of SequenceSoftmaxOp should not be null."); + + PADDLE_ENFORCE_EQ( + ctx->GetInputDim("Out"), + ctx->GetInputDim(framework::GradVarName("Out")), + "Input(Out) and Input(Out@GRAD) of SequenceSoftmaxGradOp should be of " + "the same shape."); + + ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X")); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP(sequence_softmax, ops::SequenceSoftmaxOp, + ops::SequenceSoftmaxOpMaker, sequence_softmax_grad, + ops::SequenceSoftmaxGradOp); +REGISTER_OP_CPU_KERNEL( + sequence_softmax, + ops::SequenceSoftmaxKernel); +REGISTER_OP_CPU_KERNEL( + sequence_softmax_grad, + ops::SequenceSoftmaxGradKernel); diff --git a/paddle/operators/sequence_softmax_op.cu b/paddle/operators/sequence_softmax_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..f2a1e3d5e31ef21b95a51b287bdd1d4aa9221e89 --- /dev/null +++ b/paddle/operators/sequence_softmax_op.cu @@ -0,0 +1,25 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#define EIGEN_USE_GPU + +#include "paddle/operators/sequence_softmax_op.h" + +namespace ops = paddle::operators; +REGISTER_OP_GPU_KERNEL( + sequence_softmax, + ops::SequenceSoftmaxKernel) +REGISTER_OP_GPU_KERNEL( + sequence_softmax_grad, + ops::SequenceSoftmaxGradKernel); diff --git a/paddle/operators/sequence_softmax_op.h b/paddle/operators/sequence_softmax_op.h new file mode 100644 index 0000000000000000000000000000000000000000..96d87c404d217280d74bd088e7a23f539ef6e7ce --- /dev/null +++ b/paddle/operators/sequence_softmax_op.h @@ -0,0 +1,94 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" +#include "paddle/operators/math/softmax.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; +using LoDTensor = framework::LoDTensor; + +template +class SequenceSoftmaxKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* x = ctx.Input("X"); + auto* out = ctx.Output("Out"); + + auto lod = x->lod(); + auto dims = x->dims(); + + const size_t level = lod.size() - 1; + PADDLE_ENFORCE_EQ(dims[0], static_cast(lod[level].back()), + "The first dimension of Input(X) should be equal to the " + "sum of all sequences' lengths."); + PADDLE_ENFORCE_EQ(dims[0], x->numel(), + "The width of each timestep in Input(X) of " + "SequenceSoftmaxOp should be 1."); + + out->mutable_data(ctx.GetPlace()); + for (int i = 0; i < static_cast(lod[level].size()) - 1; ++i) { + int start_pos = static_cast(lod[level][i]); + int end_pos = static_cast(lod[level][i + 1]); + Tensor x_i = x->Slice(start_pos, end_pos); + Tensor out_i = out->Slice(start_pos, end_pos); + + // Reshape from (end_pos - start_pos) x 1UL to 1UL x (end_pos - start_pos) + framework::DDim dims_i = framework::make_ddim({1UL, end_pos - start_pos}); + x_i.Resize(dims_i); + out_i.Resize(dims_i); + math::SoftmaxFunctor()(ctx.device_context(), &x_i, &out_i); + } + } +}; + +template +class SequenceSoftmaxGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* out = ctx.Input("Out"); + auto* out_grad = ctx.Input(framework::GradVarName("Out")); + auto* x = ctx.Input("X"); + auto* x_grad = ctx.Output(framework::GradVarName("X")); + + auto lod = x->lod(); + const size_t level = lod.size() - 1; + + x_grad->mutable_data(ctx.GetPlace()); + for (int i = 0; i < static_cast(lod[level].size()) - 1; ++i) { + int start_pos = static_cast(lod[level][i]); + int end_pos = static_cast(lod[level][i + 1]); + + Tensor out_i = out->Slice(start_pos, end_pos); + Tensor out_grad_i = out_grad->Slice(start_pos, end_pos); + Tensor x_grad_i = x_grad->Slice(start_pos, end_pos); + + // Reshape from (end_pos - start_pos) x 1UL to 1UL x (end_pos - start_pos) + framework::DDim dims_i = framework::make_ddim({1UL, end_pos - start_pos}); + out_i.Resize(dims_i); + out_grad_i.Resize(dims_i); + x_grad_i.Resize(dims_i); + math::SoftmaxGradFunctor()(ctx.device_context(), &out_i, + &out_grad_i, &x_grad_i); + } + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/sgd_op.cc b/paddle/operators/sgd_op.cc index b063e2427217f20eb89f7cd1af0354ad0e400feb..2a6a162a02507fa42c5ecbba59384f9c32aba7a9 100644 --- a/paddle/operators/sgd_op.cc +++ b/paddle/operators/sgd_op.cc @@ -22,19 +22,23 @@ class SGDOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("param"), - "Input(param) of SGDOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("grad"), - "Input(grad) of SGDOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("param_out"), - "Output(param_out) of SGDOp should not be null."); + void InferShape(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Param"), + "Input(Param) of SGDOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Grad"), + "Input(Grad) of SGDOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("LearningRate"), + "Input(LearningRate) of SGDOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("ParamOut"), + "Output(ParamOut) of SGDOp should not be null."); - PADDLE_ENFORCE_EQ(ctx.Input("param")->dims(), - ctx.Input("grad")->dims(), + auto lr_dims = ctx->GetInputDim("LearningRate"); + PADDLE_ENFORCE_EQ(framework::product(lr_dims), 1, + "Learning rate should have 1 element"); + auto param_dim = ctx->GetInputDim("Param"); + PADDLE_ENFORCE_EQ(param_dim, ctx->GetInputDim("Grad"), "Two input of SGD Op's dimension must be same."); - ctx.Output("param_out") - ->Resize(ctx.Input("param")->dims()); + ctx->SetOutputDim("ParamOut", param_dim); } }; @@ -42,10 +46,10 @@ class SGDOpMaker : public framework::OpProtoAndCheckerMaker { public: SGDOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("param", "input parameter"); - AddInput("grad", "input gradient"); - AddOutput("param_out", "output parameter"); - AddAttr("learning_rate", "learning rate of sgd"); + AddInput("Param", "Input parameter"); + AddInput("LearningRate", "Learning rate of SGD"); + AddInput("Grad", "Input gradient"); + AddOutput("ParamOut", "output parameter"); AddComment(R"DOC( Simplest sgd algorithm. diff --git a/paddle/operators/sgd_op.h b/paddle/operators/sgd_op.h index f8888f9c362e1c39af42236bb3a23be37aa3ae15..26f4012f258771794c736dbfad4af174b017f410 100644 --- a/paddle/operators/sgd_op.h +++ b/paddle/operators/sgd_op.h @@ -19,28 +19,25 @@ limitations under the License. */ namespace paddle { namespace operators { -using Tensor = framework::Tensor; -template -using EigenVector = framework::EigenVector; - template -class SGDOpKernel : public framework::OpKernel { +class SGDOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { - auto param = ctx.Input("param"); - auto grad = ctx.Input("grad"); - auto param_out = ctx.Output("param_out"); - float lr = ctx.Attr("learning_rate"); + auto param = ctx.Input("Param"); + auto grad = ctx.Input("Grad"); + auto param_out = ctx.Output("ParamOut"); + auto learning_rate = ctx.Input("LearningRate"); param_out->mutable_data(ctx.GetPlace()); - auto p = EigenVector::Flatten(*param); - auto g = EigenVector::Flatten(*grad); - auto o = EigenVector::Flatten(*param_out); + auto p = framework::EigenVector::Flatten(*param); + auto g = framework::EigenVector::Flatten(*grad); + auto o = framework::EigenVector::Flatten(*param_out); + auto lr = framework::EigenVector::Flatten(*learning_rate); auto place = ctx.GetEigenDevice(); - o.device(place) = p - lr * g; + Eigen::DSizes grad_dsize(grad->numel()); + o.device(place) = p - lr.broadcast(grad_dsize) * g; } }; diff --git a/paddle/operators/sigmoid_cross_entropy_with_logits_op.cc b/paddle/operators/sigmoid_cross_entropy_with_logits_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..b6653e1cc754bd626f04c28c42ce59f83c2fc87f --- /dev/null +++ b/paddle/operators/sigmoid_cross_entropy_with_logits_op.cc @@ -0,0 +1,150 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/operators/sigmoid_cross_entropy_with_logits_op.h" + +namespace paddle { +namespace operators { + +using framework::Tensor; + +class SigmoidCrossEntropyWithLogitsOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should be not null."); + PADDLE_ENFORCE(ctx->HasInput("Labels"), + "Input(Labels) should be not null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) should be not null."); + + auto x_dims = ctx->GetInputDim("X"); + auto labels_dims = ctx->GetInputDim("Labels"); + PADDLE_ENFORCE_EQ(x_dims.size(), 2, "Input(X)'s rank should be 2."); + PADDLE_ENFORCE_EQ(labels_dims.size(), 2, + "Input(Labels)'s rank should be 2."); + PADDLE_ENFORCE_EQ(x_dims[0], labels_dims[0], + "The 1st dimension of Input(X) and Input(Labels) should " + "be equal."); + PADDLE_ENFORCE_EQ(x_dims[1], labels_dims[1], + "The 2nd dimension of Input(X) and Input(Labels) should " + "be equal."); + + ctx->SetOutputDim("Out", x_dims); + ctx->ShareLoD("X", /*->*/ "Out"); + } +}; + +class SigmoidCrossEntropyWithLogitsGradOp + : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should be not null."); + PADDLE_ENFORCE(ctx->HasInput("Labels"), + "Input(Labels) should be not null."); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), + "Input(Out@GRAD) shoudl be not null."); + PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("X")), + "Output(X@GRAD) should be not null."); + + auto x_dims = ctx->GetInputDim("X"); + auto labels_dims = ctx->GetInputDim("Labels"); + auto dout_dims = ctx->GetInputDim(framework::GradVarName("Out")); + PADDLE_ENFORCE_EQ(x_dims.size(), 2, "Input(X)'s rank should be 2."); + PADDLE_ENFORCE_EQ(labels_dims.size(), 2, + "Input(Labels)'s rank should be 2."); + PADDLE_ENFORCE_EQ(dout_dims.size(), 2, + "Input(Out@Grad)'s rank should be 2."); + PADDLE_ENFORCE_EQ(x_dims[0], labels_dims[0], + "The 1st dimension of Input(X) and Input(Labels) should " + "be equal."); + PADDLE_ENFORCE_EQ(x_dims[1], labels_dims[1], + "The 2nd dimension of Input(X) and Input(Labels) should " + "be equal."); + PADDLE_ENFORCE_EQ(x_dims[0], dout_dims[0], + "The 1st dimension of Input(X) and Input(Out@Grad) " + "should be equal."); + PADDLE_ENFORCE_EQ(x_dims[1], dout_dims[1], + "The 2nd dimension of Input(X) and Input(Out@Grad) " + "should be equal."); + + ctx->SetOutputDim(framework::GradVarName("X"), x_dims); + } +}; + +class SigmoidCrossEntropyWithLogitsOpMaker + : public framework::OpProtoAndCheckerMaker { + public: + SigmoidCrossEntropyWithLogitsOpMaker(framework::OpProto* proto, + framework::OpAttrChecker* op_checker) + : framework::OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", + "(Tensor, default Tensor), a 2-D tensor with shape N x D, " + "where N is the batch size and D is the number of classes. " + "This input is a tensor of logits computed by the previous " + " operator. Logits are unscaled log probabilities given as " + "log(p/(1-p))."); + AddInput("Labels", + "(Tensor, default Tensor), a 2-D tensor of the same type " + "and shape as X. This input is a tensor of probabalistic labels " + "for each logit"); + AddOutput("Out", + "(Tensor, default Tensor), a 2-D tensor with shape N x D " + " of elementwise logistic losses."); + AddComment(R"DOC( +SigmoidCrossEntropyWithLogits Operator. + +This measures the elementwise probability error in discrete classification tasks +in which each class is independent. This can be thought of as predicting labels +for a data-point that are not mutually exclusive. For example, a news article +can be about politics, technology or sports at the same time or none of these. + +The logistic loss is given as follows: + + loss = -Labels * log(sigmoid(X)) - (1 - Labels) * log(1 - sigmoid(X)) + +We know that sigmoid(X) = (1 / (1 + exp(-X))). By substituting this we get + + loss = X - X * Labels + log(1 + exp(-X)) + +For stability and to prevent overflow of exp(-X) when X < 0, +we can reformulate the loss as follows: + + loss = max(X, 0) - X * Labels + log(1 + exp(-abs(X))) + +Both the input `X` and `Labels` can carry the LoD (Level of Details) information. +However the output only shares the LoD with input `X`. +)DOC"); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP(sigmoid_cross_entropy_with_logits, + ops::SigmoidCrossEntropyWithLogitsOp, + ops::SigmoidCrossEntropyWithLogitsOpMaker, + sigmoid_cross_entropy_with_logits_grad, + ops::SigmoidCrossEntropyWithLogitsGradOp); +REGISTER_OP_CPU_KERNEL(sigmoid_cross_entropy_with_logits, + ops::SigmoidCrossEntropyWithLogitsKernel< + paddle::platform::CPUPlace, float>); +REGISTER_OP_CPU_KERNEL(sigmoid_cross_entropy_with_logits_grad, + ops::SigmoidCrossEntropyWithLogitsGradKernel< + paddle::platform::CPUPlace, float>); diff --git a/paddle/operators/sigmoid_cross_entropy_with_logits_op.cu b/paddle/operators/sigmoid_cross_entropy_with_logits_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..32a39956a14a206373b7b4c141dad19577d171f0 --- /dev/null +++ b/paddle/operators/sigmoid_cross_entropy_with_logits_op.cu @@ -0,0 +1,24 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#define EIGEN_USE_GPU +#include "paddle/operators/sigmoid_cross_entropy_with_logits_op.h" + +namespace ops = paddle::operators; +REGISTER_OP_GPU_KERNEL(sigmoid_cross_entropy_with_logits, + ops::SigmoidCrossEntropyWithLogitsKernel< + paddle::platform::GPUPlace, float>); +REGISTER_OP_GPU_KERNEL(sigmoid_cross_entropy_with_logits_grad, + ops::SigmoidCrossEntropyWithLogitsGradKernel< + paddle::platform::GPUPlace, float>); diff --git a/paddle/operators/sigmoid_cross_entropy_with_logits_op.h b/paddle/operators/sigmoid_cross_entropy_with_logits_op.h new file mode 100644 index 0000000000000000000000000000000000000000..41c619f181c878f08959a8ca461c60af5ffdff2a --- /dev/null +++ b/paddle/operators/sigmoid_cross_entropy_with_logits_op.h @@ -0,0 +1,75 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" + +namespace paddle { +namespace operators { + +// Out = max(X, 0) - X * Labels + log(1 + exp(-abs(X))) +template +class SigmoidCrossEntropyWithLogitsKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &context) const override { + const framework::Tensor *X = context.Input("X"); + const framework::Tensor *Labels = + context.Input("Labels"); + framework::Tensor *Out = context.Output("Out"); + Out->mutable_data(context.GetPlace()); + + auto x = framework::EigenVector::Flatten(*X); + auto labels = framework::EigenVector::Flatten(*Labels); + auto out = framework::EigenVector::Flatten(*Out); + auto place = context.GetEigenDevice(); + + // term1 = max(x, 0) + auto term1 = x.cwiseMax(static_cast(0)); + // term2 = x * labels + auto term2 = x * labels; + // term3 = log(1 + exp(-abs(x))) + auto term3 = (static_cast(1) + (-(x.abs())).exp()).log(); + + out.device(place) = term1 - term2 + term3; + } +}; + +// dX = sigmoid(X) - labels +template +class SigmoidCrossEntropyWithLogitsGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &context) const override { + const framework::Tensor *X = context.Input("X"); + const framework::Tensor *Labels = + context.Input("Labels"); + const framework::Tensor *dOut = + context.Input(framework::GradVarName("Out")); + framework::Tensor *dX = + context.Output(framework::GradVarName("X")); + dX->mutable_data(context.GetPlace()); + + auto x = framework::EigenVector::Flatten(*X); + auto labels = framework::EigenVector::Flatten(*Labels); + auto dout = framework::EigenVector::Flatten(*dOut); + auto dx = framework::EigenVector::Flatten(*dX); + auto place = context.GetEigenDevice(); + + auto sigmoid_x = static_cast(1) / (static_cast(1) + (-x).exp()); + dx.device(place) = dout * (sigmoid_x - labels); + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/smooth_l1_loss_op.cc b/paddle/operators/smooth_l1_loss_op.cc index ae6d1c80b300690b070024d6266a1b99bf2ef04f..91391dc945be3d3d94966339bf9232eb21dcc38f 100644 --- a/paddle/operators/smooth_l1_loss_op.cc +++ b/paddle/operators/smooth_l1_loss_op.cc @@ -22,33 +22,28 @@ class SmoothL1LossOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext& ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "X must be initialized."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), "Y must be initialized."); - - auto* x = ctx.Input("X"); - auto* y = ctx.Input("Y"); - PADDLE_ENFORCE_EQ(x->dims(), y->dims(), - "The shape of X and Y must be the same."); - PADDLE_ENFORCE_GE(x->dims().size(), 2, + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "X must be initialized."); + PADDLE_ENFORCE(ctx->HasInput("Y"), "Y must be initialized."); + + auto x_dims = ctx->GetInputDim("X"); + auto y_dims = ctx->GetInputDim("Y"); + PADDLE_ENFORCE_EQ(x_dims, y_dims, "The shape of X and Y must be the same."); + PADDLE_ENFORCE_GE(x_dims.size(), 2, "The tensor rank of X must be at least 2."); - auto* inside_weight = ctx.Input("InsideWeight"); - if (inside_weight) { - auto* outside_weight = ctx.Input("OutsideWeight"); - PADDLE_ENFORCE_NOT_NULL(outside_weight, - "If weights are provided, must specify both " - "inside and outside weights."); - PADDLE_ENFORCE_EQ(inside_weight->dims(), x->dims(), + if (ctx->HasInput("InsideWeight")) { + PADDLE_ENFORCE(ctx->HasInput("OutsideWeight"), + "If weights are provided, must specify both " + "inside and outside weights."); + PADDLE_ENFORCE_EQ(ctx->GetInputDim("InsideWeight"), x_dims, "The shape of InsideWeight must be same as X."); - PADDLE_ENFORCE_EQ(outside_weight->dims(), x->dims(), + PADDLE_ENFORCE_EQ(ctx->GetInputDim("OutsideWeight"), x_dims, "The shape of OutsideWeight must be same as X."); } - auto* diff = ctx.Output("Diff"); - auto* out = ctx.Output("Out"); - diff->Resize(x->dims()); + ctx->SetOutputDim("Diff", x_dims); // loss is a two-rank tensor - out->Resize({x->dims()[0], 1}); + ctx->SetOutputDim("Out", {x_dims[0], 1}); } }; @@ -99,12 +94,9 @@ class SmoothL1LossGradOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext& ctx) const override { - auto in_dims = ctx.Input("X")->dims(); - auto out_dims = - ctx.Input(framework::GradVarName("Out"))->dims(); - auto* x_grad = ctx.Output(framework::GradVarName("X")); - auto* y_grad = ctx.Output(framework::GradVarName("Y")); + void InferShape(framework::InferShapeContext* ctx) const override { + auto in_dims = ctx->GetInputDim("X"); + auto out_dims = ctx->GetInputDim(framework::GradVarName("Out")); PADDLE_ENFORCE_GE(out_dims.size(), 2, "The tensor rank of Input(Out@Grad) should be 2."); @@ -114,8 +106,14 @@ class SmoothL1LossGradOp : public framework::OperatorWithKernel { PADDLE_ENFORCE_EQ(out_dims[1], 1, "The 2nd dimension of Input(Out@Grad) must be 1."); - if (x_grad) x_grad->Resize(in_dims); - if (y_grad) y_grad->Resize(in_dims); + auto x_grad_name = framework::GradVarName("X"); + auto y_grad_name = framework::GradVarName("Y"); + if (ctx->HasOutput(x_grad_name)) { + ctx->SetOutputDim(x_grad_name, in_dims); + } + if (ctx->HasOutput(y_grad_name)) { + ctx->SetOutputDim(y_grad_name, in_dims); + } } }; diff --git a/paddle/operators/smooth_l1_loss_op.h b/paddle/operators/smooth_l1_loss_op.h index 0604fb5e1c2f17c702208520a1d23bd5c3c65b5d..39d0070b6c8909b8f433de48038240e851d9d6cf 100644 --- a/paddle/operators/smooth_l1_loss_op.h +++ b/paddle/operators/smooth_l1_loss_op.h @@ -45,7 +45,7 @@ struct SmoothL1LossForward { }; template -class SmoothL1LossKernel : public framework::OpKernel { +class SmoothL1LossKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto* in0 = context.Input("X"); @@ -115,7 +115,7 @@ struct SmoothL1LossBackward { }; template -class SmoothL1LossGradKernel : public framework::OpKernel { +class SmoothL1LossGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto* in0 = context.Input("InsideWeight"); diff --git a/paddle/operators/softmax_op.cc b/paddle/operators/softmax_op.cc index e15cfe485016552971924a40a172e74a90629dce..4c131ed44de1c1567193c7cabff6ddf2dc63d9be 100644 --- a/paddle/operators/softmax_op.cc +++ b/paddle/operators/softmax_op.cc @@ -22,22 +22,23 @@ class SoftmaxOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), - "Input(X) of SoftmaxOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Y"), - "Output(Y) of SoftmaxOp should not be null."); - - PADDLE_ENFORCE(ctx.Input("X")->dims().size() == 2UL, + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of SoftmaxOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Y"), + "Output(Y) of SoftmaxOp should not be null."); + + auto x_dims = ctx->GetInputDim("X"); + PADDLE_ENFORCE(x_dims.size() == 2UL, "The input of softmax op must be a matrix."); - ctx.Output("Y")->Resize(ctx.Input("X")->dims()); + ctx->SetOutputDim("Y", x_dims); } }; class SoftmaxOpMaker : public framework::OpProtoAndCheckerMaker { public: - SoftmaxOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + SoftmaxOpMaker(framework::OpProto* proto, + framework::OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "The input tensor of softmax. " @@ -68,16 +69,15 @@ class SoftmaxOpGrad : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), "Input(Y) should be not null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Y")), - "Input(Y@GRAD) should be not null."); - PADDLE_ENFORCE_EQ(ctx.Input("Y")->dims(), - ctx.Input(framework::GradVarName("Y"))->dims(), + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Y"), "Input(Y) should be not null."); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Y")), + "Input(Y@GRAD) should be not null."); + PADDLE_ENFORCE_EQ(ctx->GetInputDim("Y"), + ctx->GetInputDim(framework::GradVarName("Y")), "Input(Y) and its gradients should have a same shape."); - ctx.Output(framework::GradVarName("X")) - ->Resize(ctx.Input("X")->dims()); + ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X")); } }; diff --git a/paddle/operators/softmax_op.h b/paddle/operators/softmax_op.h index 8a3a5ab927c0e2937936fcc973f000d4d95c3dbc..2c08853f4f615bfe95f51aa20776ddddcdaa8f61 100644 --- a/paddle/operators/softmax_op.h +++ b/paddle/operators/softmax_op.h @@ -15,6 +15,7 @@ limitations under the License. */ #pragma once #include "paddle/framework/eigen.h" #include "paddle/framework/op_registry.h" +#include "paddle/operators/math/softmax.h" namespace paddle { namespace operators { @@ -25,73 +26,31 @@ template ; template -class SoftmaxKernel : public framework::OpKernel { +class SoftmaxKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { - auto X = context.Input("X"); - auto Y = context.Output("Y"); - Y->mutable_data(context.GetPlace()); - - auto logits = EigenMatrix::From(*X); - auto softmax = EigenMatrix::From(*Y); - - const int kBatchDim = 0; - const int kClassDim = 1; - - const int batch_size = logits.dimension(kBatchDim); - const int num_classes = logits.dimension(kClassDim); - - Eigen::DSizes along_class(kClassDim); - Eigen::DSizes batch_by_one(batch_size, 1); - Eigen::DSizes one_by_class(1, num_classes); + auto* X = context.Input("X"); + auto* Y = context.Output("Y"); - auto shifted_logits = (logits - - logits.maximum(along_class) - .eval() - .reshape(batch_by_one) - .broadcast(one_by_class)); - - softmax.device(context.GetEigenDevice()) = shifted_logits.exp(); + // allocate memory on device. + Y->mutable_data(context.GetPlace()); - softmax.device(context.GetEigenDevice()) = - (softmax * - softmax.sum(along_class) - .inverse() - .eval() - .reshape(batch_by_one) - .broadcast(one_by_class)); + math::SoftmaxFunctor()(context.device_context(), X, Y); } }; template -class SoftmaxGradKernel : public framework::OpKernel { +class SoftmaxGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { - std::shared_ptr scale_ = std::make_shared(); + auto* Y = context.Input("Y"); + auto* dY = context.Input(framework::GradVarName("Y")); + auto* dX = context.Output(framework::GradVarName("X")); - auto Y = context.Input("Y"); - auto dY = context.Input(framework::GradVarName("Y")); - auto dX = context.Output(framework::GradVarName("X")); + // allocate memory on device. dX->mutable_data(context.GetPlace()); - const int batch_size = Y->dims()[0]; - const int class_num = Y->dims()[1]; - - Eigen::DSizes along_class(1); - Eigen::DSizes batch_by_one(batch_size, 1); - Eigen::DSizes one_by_class(1, class_num); - - auto Y_eigen = EigenMatrix::From(*Y); - auto dY_eigen = EigenMatrix::From(*dY); - auto dX_eigen = EigenMatrix::From(*dX); - auto place = context.GetEigenDevice(); - - auto dot = (Y_eigen * dY_eigen) - .sum(along_class) - .eval() - .reshape(batch_by_one) - .broadcast(one_by_class); - dX_eigen.device(place) = (dY_eigen - dot) * Y_eigen; + math::SoftmaxGradFunctor()(context.device_context(), Y, dY, dX); } }; diff --git a/paddle/operators/softmax_with_cross_entropy_op.cc b/paddle/operators/softmax_with_cross_entropy_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..5431a1657c300d9c60100616ca7ea2e1196824ec --- /dev/null +++ b/paddle/operators/softmax_with_cross_entropy_op.cc @@ -0,0 +1,197 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/operators/softmax_with_cross_entropy_op.h" +#include +#include + +namespace paddle { +namespace operators { + +class SoftmaxWithCrossEntropyOpMaker + : public framework::OpProtoAndCheckerMaker { + public: + SoftmaxWithCrossEntropyOpMaker(framework::OpProto* proto, + framework::OpAttrChecker* op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("Logits", + "(Tensor, default: Tensor), The unscaled log probabilities " + "which is a 2-D tensor with shape [N x K]. N is the batch_size, " + "and K is the class number."); + AddInput("Label", + "(Tensor, default: Tensor), The ground truth which is a 2-D " + "tensor. " + "If softLable is set to 0, Label is a Tensor with shape [N x " + "1]. " + "If softLable is set to 1, Label is a Tensor " + "with shape [N x K]."); + AddOutput( + "Softmax", + "(Tensor, default: Tensor), A 2-D tensor with shape [N x K]. " + "The outputs value of softmax activation by given the input batch, " + "which will be used in backward calculation.") + .AsIntermediate(); + AddOutput("Loss", + "(Tensor, default: Tensor), A 2-D tensor. The cross " + "entropy loss with shape [N x 1]."); + AddAttr( + "softLabel", + "(bool, default: false), A flag to indicate whether to interpretate " + "the given labels as soft labels.") + .SetDefault(false); + AddComment(R"DOC( +Cross entropy loss with softmax are used as the output layer extensively. This +operator computes the softmax normalized values for each row of the input +tensor, after which cross-entropy loss is then computed. This provides a more +numerically stable gradient. + +Because this operators performs a softmax on logits internally, it expects +unscaled logits. Please do not call this op with the output of softmax operator, +which will produce incorrect results. + +This operators expects mutually exclusive hard labels, each sample in a batch +is in exactly one class with probabilities 1. Each sample in the batch with one +and only one label. + +Equation: + +1) hard label (one-hot label) + +Loss_j = -\text{Logit}_{Label_j} + \log\left(\sum_{i=0}^{K}\exp(\text{Logit}_i)\right), j = 1, ..., K + +2) soft label (a distribution over all classes) + +Loss_j = -\sum_{i=0}^{K}\text{Label}_i\left(\text{Logit}_i-\log\left(\sum_{i=0}^{K}\exp(\text{Logit}_i)\right)\right), j = 1,...,K + +)DOC"); + } +}; + +class SoftmaxWithCrossEntropyOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Logits"), + "Input(Logits) should be not null."); + PADDLE_ENFORCE(ctx->HasInput("Label"), "Input(Label) should be not null."); + + PADDLE_ENFORCE(ctx->HasOutput("Softmax"), + "Output(Softmax) should be not null."); + PADDLE_ENFORCE(ctx->HasOutput("Loss"), "Output(Loss) should be not null."); + + auto logits_dims = ctx->GetInputDim("Logits"); + auto labels_dims = ctx->GetInputDim("Label"); + PADDLE_ENFORCE_EQ( + logits_dims.size(), 2UL, + "The input of softmax_with_cross_entropy should be a 2-D tensor."); + PADDLE_ENFORCE_EQ(labels_dims.size(), 2UL, + "The labels should be a 2-D tensor."); + + if (ctx->Attrs().Get("softLabel")) { + PADDLE_ENFORCE_EQ(logits_dims[1], labels_dims[1], + "If Attr(softLabel) == true, the 2nd dimension of " + "Input(X) and Input(Label) should be equal."); + } else { + PADDLE_ENFORCE_EQ(labels_dims[1], 1UL, + "If Attr(softLabel) == false, the 2nd dimension of " + "Input(Label) should be 1."); + } + + ctx->SetOutputDim("Softmax", logits_dims); + ctx->SetOutputDim("Loss", {logits_dims[0], 1}); + + ctx->ShareLoD("Logits", /*->*/ "Softmax"); + ctx->ShareLoD("Logits", /*->*/ "Loss"); + } + + framework::DataType IndicateDataType( + const framework::ExecutionContext& ctx) const override { + return framework::ToDataType(ctx.Input("Logits")->type()); + } +}; + +class SoftmaxWithCrossEntropyOpGrad : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Loss")), + "Input(Loss@Grad) should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Softmax"), + "Input(Softmax) should be not null."); + PADDLE_ENFORCE(ctx->HasInput("Label"), "Input(Label) should be not null."); + PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("Logits")), + "Output(Logits@Grad) should be not null."); + + auto softmax_dims = ctx->GetInputDim("Softmax"); + auto labels_dims = ctx->GetInputDim("Label"); + PADDLE_ENFORCE_EQ(labels_dims.size(), 2UL, + "The labels should be a 2-D tensor."); + + if (ctx->Attrs().Get("softLabel")) { + PADDLE_ENFORCE_EQ(softmax_dims[1], labels_dims[1], + "When Attr(softLabel) == true, the 2nd dimension of " + "Input(X) and Input(Label) should be equal."); + } else { + PADDLE_ENFORCE_EQ(labels_dims[1], 1UL, + "When Attr(softLabel) == false, the 2nd dimension of " + "Input(Label) should be 1."); + } + + ctx->SetOutputDim(framework::GradVarName("Logits"), + ctx->GetInputDim("Softmax")); + } + + framework::DataType IndicateDataType( + const framework::ExecutionContext& ctx) const override { + return framework::ToDataType( + ctx.Input(framework::GradVarName("Loss"))->type()); + } +}; + +class SoftmaxGradMaker : public framework::SingleGradOpDescMaker { + public: + using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; + + protected: + std::unique_ptr Apply() const override { + auto* grad_op = new framework::OpDescBind(); + grad_op->SetType("softmax_with_cross_entropy_grad"); + grad_op->SetInput("Label", Input("Label")); + grad_op->SetInput("Softmax", Output("Softmax")); + grad_op->SetInput("Loss", Output("Loss")); + grad_op->SetInput(framework::GradVarName("Softmax"), OutputGrad("Softmax")); + grad_op->SetInput(framework::GradVarName("Loss"), OutputGrad("Loss")); + grad_op->SetOutput(framework::GradVarName("Logits"), InputGrad("Logits")); + grad_op->SetAttrMap(Attrs()); + return std::unique_ptr(grad_op); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; + +REGISTER_OPERATOR(softmax_with_cross_entropy, ops::SoftmaxWithCrossEntropyOp, + ops::SoftmaxWithCrossEntropyOpMaker, ops::SoftmaxGradMaker); +REGISTER_OPERATOR(softmax_with_cross_entropy_grad, + ops::SoftmaxWithCrossEntropyOpGrad); +REGISTER_OP_CPU_KERNEL(softmax_with_cross_entropy, + ops::SoftmaxWithCrossEntropyKernel); +REGISTER_OP_CPU_KERNEL(softmax_with_cross_entropy_grad, + ops::SoftmaxWithCrossEntropyGradKernel); diff --git a/paddle/operators/softmax_with_cross_entropy_op.cu b/paddle/operators/softmax_with_cross_entropy_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..2bc53ecf871eb1800a920ba85e8eac31d7037efe --- /dev/null +++ b/paddle/operators/softmax_with_cross_entropy_op.cu @@ -0,0 +1,121 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#define EIGEN_USE_GPU + +#include "paddle/operators/softmax_with_cross_entropy_op.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; + +namespace { +template +__global__ void CrossEntropyGrad(T* out_grad, const T* in_grad, + const int* labels, const int batch_size, + const int class_num) { + int tid = blockIdx.x * blockDim.x + threadIdx.x; + int sample_idx = tid / class_num; + + if (tid < batch_size * class_num) out_grad[tid] *= in_grad[sample_idx]; + __syncthreads(); + + if (tid < batch_size) { + PADDLE_ASSERT(labels[sample_idx] >= 0 && labels[sample_idx] < class_num); + out_grad[tid * class_num + labels[tid]] -= 1.; + } +} + +template +__global__ void SoftCrossEntropyGradientKernel(T* logit_grad, + const T* loss_grad, + const T* labels, + const int batch_size, + const int class_num) { + int ids = blockIdx.x * blockDim.x + threadIdx.x; + if (ids < batch_size * class_num) { + int row_ids = ids / class_num; + logit_grad[ids] = logit_grad[ids] * loss_grad[row_ids] - labels[ids]; + } +} +} // namespace + +template +class SoftmaxWithCrossEntropyCUDAKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + PADDLE_ENFORCE(platform::is_gpu_place(context.GetPlace()), + "This kernel only runs on GPU device."); + const Tensor* logits = context.Input("Logits"); + const Tensor* labels = context.Input("Label"); + Tensor* softmax = context.Output("Softmax"); + + Tensor* loss = context.Output("Loss"); + softmax->mutable_data(context.GetPlace()); + loss->mutable_data(context.GetPlace()); + + math::SoftmaxFunctor()(context.device_context(), + logits, softmax); + math::CrossEntropyFunctor()( + context.device_context(), loss, softmax, labels, + context.Attr("softLabel")); + } +}; + +template +class SoftmaxWithCrossEntropyGradCUDAKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + PADDLE_ENFORCE(platform::is_gpu_place(context.GetPlace()), + "This kernel only runs on GPU device."); + const Tensor* labels = context.Input("Label"); + const T* loss_grad_data = + context.Input(framework::GradVarName("Loss"))->data(); + Tensor* logit_grad = + context.Output(framework::GradVarName("Logits")); + logit_grad->ShareDataWith(*context.Input("Softmax")); + T* logit_grad_data = logit_grad->data(); + + const int batch_size = logit_grad->dims()[0]; + const int class_num = logit_grad->dims()[1]; + int block = 512; + int grid = (batch_size * class_num + block - 1) / block; + + if (context.Attr("softLabel")) { + const T* label_data = labels->data(); + SoftCrossEntropyGradientKernel<<< + grid, block, 0, reinterpret_cast( + context.device_context()) + .stream()>>>(logit_grad_data, loss_grad_data, + label_data, batch_size, class_num); + } else { + const int* label_data = labels->data(); + CrossEntropyGrad<<< + grid, block, 0, reinterpret_cast( + context.device_context()) + .stream()>>>(logit_grad_data, loss_grad_data, + label_data, batch_size, class_num); + } + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP_GPU_KERNEL(softmax_with_cross_entropy, + ops::SoftmaxWithCrossEntropyCUDAKernel); +REGISTER_OP_GPU_KERNEL(softmax_with_cross_entropy_grad, + ops::SoftmaxWithCrossEntropyGradCUDAKernel); diff --git a/paddle/operators/softmax_with_cross_entropy_op.h b/paddle/operators/softmax_with_cross_entropy_op.h new file mode 100644 index 0000000000000000000000000000000000000000..cffd422f1827b646a8abcd881fdcb5455e6a663a --- /dev/null +++ b/paddle/operators/softmax_with_cross_entropy_op.h @@ -0,0 +1,88 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" +#include "paddle/operators/math/cross_entropy.h" +#include "paddle/operators/math/softmax.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; +template +using EigenMatrix = framework::EigenMatrix; + +template +class SoftmaxWithCrossEntropyKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + PADDLE_ENFORCE(platform::is_cpu_place(context.GetPlace()), + "This kernel only runs on CPU."); + const Tensor* logits = context.Input("Logits"); + const Tensor* labels = context.Input("Label"); + Tensor* softmax = context.Output("Softmax"); + Tensor* loss = context.Output("Loss"); + + softmax->mutable_data(context.GetPlace()); + loss->mutable_data(context.GetPlace()); + + math::SoftmaxFunctor()(context.device_context(), + logits, softmax); + math::CrossEntropyFunctor()( + context.device_context(), loss, softmax, labels, + context.Attr("softLabel")); + } +}; + +template +class SoftmaxWithCrossEntropyGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + const Tensor* out_grad = + context.Input(framework::GradVarName("Loss")); + const Tensor* labels = context.Input("Label"); + Tensor* logit_grad = + context.Output(framework::GradVarName("Logits")); + logit_grad->ShareDataWith(*context.Input("Softmax")); + + const int class_num = logit_grad->dims()[1]; + if (context.Attr("softLabel")) { + auto out_grad_mat = EigenMatrix::From(*out_grad); + auto logit_grad_mat = EigenMatrix::From(*logit_grad); + auto lbl_mat = EigenMatrix::From(*labels); + + logit_grad_mat.device(context.GetEigenDevice()) = + logit_grad_mat * + out_grad_mat.broadcast(Eigen::DSizes(1, class_num)) - + lbl_mat; + } else { + const int batch_size = logit_grad->dims()[0]; + const int* label_data = labels->data(); + const T* out_grad_data = out_grad->data(); + T* logit_grad_data = logit_grad->data(); + + for (int i = 0; i < batch_size; ++i) { + int index = i * class_num + label_data[i]; + logit_grad_data[index] = + (out_grad_data[i] * logit_grad_data[index] - 1.); + } + } + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/split_op.cc b/paddle/operators/split_op.cc index a9d35b4fb79ae83379552ae2c2b4d694bd8f86dd..d5dd4df2a2a6f424421ae7f117c26acff08d37d0 100644 --- a/paddle/operators/split_op.cc +++ b/paddle/operators/split_op.cc @@ -24,40 +24,43 @@ class SplitOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - // infershape - auto *in = ctx.Input("X"); - auto outs = ctx.MultiOutput("Out"); - size_t axis = static_cast(ctx.Attr("axis")); - size_t num = static_cast(ctx.Attr("num")); - std::vector sections = - static_cast>(ctx.Attr>("sections")); - const size_t n = outs.size(); + void InferShape(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of SplitOp should not be null."); + PADDLE_ENFORCE_GE(ctx->Outputs("Out").size(), 1UL, + "Outputs(Out) of SplitOp should not be empty."); + auto in_dims = ctx->GetInputDim("X"); + auto outs_names = ctx->Outputs("Out"); + size_t axis = static_cast(ctx->Attrs().Get("axis")); + size_t num = static_cast(ctx->Attrs().Get("num")); + std::vector sections = static_cast>( + ctx->Attrs().Get>("sections")); + const size_t outs_number = outs_names.size(); + std::vector outs_dims; + outs_dims.reserve(outs_number); if (num > 0) { - int64_t in_axis_dim = in->dims()[axis]; + int64_t in_axis_dim = in_dims[axis]; PADDLE_ENFORCE_EQ(in_axis_dim % num, 0, "tensor split does not result" " in an equal division"); size_t out_axis_dim = in_axis_dim / num; - for (size_t i = 0; i < n; ++i) { - auto dim = in->dims(); + for (size_t i = 0; i < outs_number; ++i) { + auto dim = in_dims; dim[axis] = out_axis_dim; - outs[i]->Resize(dim); + outs_dims.push_back(dim); } } else if (sections.size() > 0) { - PADDLE_ENFORCE_EQ(sections.size(), n, + PADDLE_ENFORCE_EQ(sections.size(), outs_number, "tensor split sections size" "should be equal to output size."); - for (size_t i = 0; i < n; ++i) { - auto dim = in->dims(); + for (size_t i = 0; i < outs_number; ++i) { + auto dim = in_dims; dim[axis] = sections[i]; - outs[i]->Resize(dim); + outs_dims.push_back(dim); } - } else { - PADDLE_ENFORCE_NOT_NULL(nullptr, "split operator should", - " specify indices or sections."); } + ctx->SetOutputsDim("Out", outs_dims); } }; @@ -115,4 +118,4 @@ USE_CPU_ONLY_OP(concat); REGISTER_OP(split, ops::SplitOp, ops::SplitOpMaker, split_grad, ops::SplitOpGrad); REGISTER_OP_CPU_KERNEL(split, - ops::SplitKernel); + ops::SplitOpKernel); diff --git a/paddle/operators/split_op.cu b/paddle/operators/split_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..93d1fc3c44cbc146c945c51af1abe6494572d1ae --- /dev/null +++ b/paddle/operators/split_op.cu @@ -0,0 +1,18 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/operators/split_op.h" +namespace ops = paddle::operators; +REGISTER_OP_GPU_KERNEL(split, + ops::SplitOpKernel); diff --git a/paddle/operators/split_op.h b/paddle/operators/split_op.h index 860690ee895075fda9ddef08776a2102642efff9..fa26e5f677b18c84b45dd583004d02cab4c1d375 100644 --- a/paddle/operators/split_op.h +++ b/paddle/operators/split_op.h @@ -16,44 +16,29 @@ limitations under the License. */ #include #include "paddle/framework/op_registry.h" +#include "paddle/operators/strided_memcpy.h" namespace paddle { namespace operators { template -class SplitKernel : public framework::OpKernel { +class SplitOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { auto* in = ctx.Input("X"); auto outs = ctx.MultiOutput("Out"); + auto in_stride = framework::stride(in->dims()); int64_t axis = static_cast(ctx.Attr("axis")); - size_t before = 1, after = 1; const size_t n = outs.size(); - size_t input_axis_dim = in->dims()[axis]; - - for (int64_t i = 0; i < in->dims().size(); ++i) { - if (i == axis) { - continue; - } - if (i < axis) { - before *= in->dims()[i]; - } else { - after *= in->dims()[i]; - } - } size_t input_offset = 0; for (size_t i = 0; i < n; i++) { auto& out = outs[i]; + out->mutable_data(ctx.GetPlace()); size_t axis_dim = out->dims()[axis]; - for (size_t j = 0; j < before; j++) { - size_t len = axis_dim * after * sizeof(T); - T* dest = - out->mutable_data(platform::CPUPlace()) + axis_dim * after * j; - const T* src = - in->data() + input_offset + input_axis_dim * after * j; - memcpy(dest, src, len); - } - input_offset += axis_dim * after; + auto out_stride = framework::stride(out->dims()); + StridedMemcpy(ctx.device_context(), in->data() + input_offset, + in_stride, out->dims(), out_stride, out->data()); + input_offset += axis_dim * in_stride[axis]; } } }; diff --git a/paddle/operators/squared_l2_distance_op.cc b/paddle/operators/squared_l2_distance_op.cc index 33a564b05b1b490c6d23b7d17cef45b7740dfa39..cce4e527c36bc0f4735165ff88ebba85b00eb898 100644 --- a/paddle/operators/squared_l2_distance_op.cc +++ b/paddle/operators/squared_l2_distance_op.cc @@ -22,24 +22,19 @@ class SquaredL2DistanceOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext& ctx) const override { - PADDLE_ENFORCE_NOT_NULL( - ctx.InputVar("X"), - "Input(X) of SquaredL2DistanceOp should not be null."); - PADDLE_ENFORCE_NOT_NULL( - ctx.InputVar("Y"), - "Input(Y) of SquaredL2DistanceOp should not be null."); - PADDLE_ENFORCE_NOT_NULL( - ctx.OutputVar("sub_result"), + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of SquaredL2DistanceOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Y"), + "Input(Y) of SquaredL2DistanceOp should not be null."); + PADDLE_ENFORCE( + ctx->HasOutput("sub_result"), "Output(sub_result) of SquaredL2DistanceOp should not be null."); - PADDLE_ENFORCE_NOT_NULL( - ctx.OutputVar("Out"), - "Output(Out) of SquaredL2DistanceOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of SquaredL2DistanceOp should not be null."); - auto* x = ctx.Input("X"); - auto x_dims = x->dims(); - auto* y = ctx.Input("Y"); - auto y_dims = y->dims(); + auto x_dims = ctx->GetInputDim("X"); + auto y_dims = ctx->GetInputDim("Y"); PADDLE_ENFORCE_EQ(framework::arity(x_dims), framework::arity(y_dims), "Tensor rank of both SquaredL2DistanceOp's " @@ -47,17 +42,16 @@ class SquaredL2DistanceOp : public framework::OperatorWithKernel { int rank = framework::arity(x_dims); PADDLE_ENFORCE_GE(rank, 2, "Tensor rank should be at least equal to 2."); - PADDLE_ENFORCE_EQ(x->numel() / x_dims[0], y->numel() / y_dims[0], + PADDLE_ENFORCE_EQ(product(x_dims) / x_dims[0], product(y_dims) / y_dims[0], "Product of dimensions expcet the first dimension of " "input and target must be equal."); PADDLE_ENFORCE(y_dims[0] == 1 || y_dims[0] == x_dims[0], "First dimension of target must be equal to input " "or to 1."); - ctx.Output("sub_result") - ->Resize({x_dims[0], x->numel() / x_dims[0]}); - ctx.Output("Out")->Resize({x_dims[0], 1}); - ctx.ShareLoD("X", /*->*/ "Out"); + ctx->SetOutputDim("sub_result", {x_dims[0], product(x_dims) / x_dims[0]}); + ctx->SetOutputDim("Out", {x_dims[0], 1}); + ctx->ShareLoD("X", /*->*/ "Out"); } }; @@ -92,22 +86,22 @@ class SquaredL2DistanceGradOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext& ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")), - "Gradient of Out should not be null"); - auto out_dims = ctx.Input(framework::GradVarName("Out"))->dims(); - auto x_dims = ctx.Input("X")->dims(); - auto y_dims = ctx.Input("Y")->dims(); + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), + "Gradient of Out should not be null"); + auto out_dims = ctx->GetInputDim(framework::GradVarName("Out")); + auto x_dims = ctx->GetInputDim("X"); + auto y_dims = ctx->GetInputDim("Y"); PADDLE_ENFORCE_EQ(out_dims[0], x_dims[0], "First dimension of output gradient and " "input value must be equal."); PADDLE_ENFORCE_EQ(out_dims[1], 1, "Second dimension of output gradient " "must be 1."); - auto* x_grad = ctx.Output(framework::GradVarName("X")); - auto* y_grad = ctx.Output(framework::GradVarName("Y")); - if (x_grad) x_grad->Resize(x_dims); - if (y_grad) y_grad->Resize(y_dims); + auto x_grad_name = framework::GradVarName("X"); + auto y_grad_name = framework::GradVarName("Y"); + if (ctx->HasOutput(x_grad_name)) ctx->SetOutputDim(x_grad_name, x_dims); + if (ctx->HasOutput(y_grad_name)) ctx->SetOutputDim(y_grad_name, y_dims); } }; diff --git a/paddle/operators/squared_l2_distance_op.h b/paddle/operators/squared_l2_distance_op.h index 097ac04fc09a10b3b624f491a847e281e41a802c..259ef4029646914f83a112b9c6d7fdf8401483f6 100644 --- a/paddle/operators/squared_l2_distance_op.h +++ b/paddle/operators/squared_l2_distance_op.h @@ -28,7 +28,7 @@ template ; template -class SquaredL2DistanceKernel : public framework::OpKernel { +class SquaredL2DistanceKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto* in0 = context.Input("X"); @@ -68,7 +68,7 @@ class SquaredL2DistanceKernel : public framework::OpKernel { }; template -class SquaredL2DistanceGradKernel : public framework::OpKernel { +class SquaredL2DistanceGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto* in0 = context.Input("sub_result"); diff --git a/paddle/operators/strided_memcpy_test.cc b/paddle/operators/strided_memcpy_test.cc index 05882a88738cfc9cc23480efe0afe504008377ca..68f064eaee5851333ddf9767b7138da83a28503d 100644 --- a/paddle/operators/strided_memcpy_test.cc +++ b/paddle/operators/strided_memcpy_test.cc @@ -72,7 +72,7 @@ TEST(StridedMemcpy, CPUConcat) { } } -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA TEST(StridedMemcpy, GPUCrop) { // clang-format off int src[] = { @@ -157,4 +157,4 @@ TEST(StridedMemcpy, GPUConcat) { #endif } // namespace operators -} // namespace paddle \ No newline at end of file +} // namespace paddle diff --git a/paddle/operators/sum_op.cc b/paddle/operators/sum_op.cc index 437fc262f359525045a4d772ee2c204ef571caa7..ffb0cb92111bfb8490d35e4f5cfc9e405b0e3250 100644 --- a/paddle/operators/sum_op.cc +++ b/paddle/operators/sum_op.cc @@ -11,6 +11,7 @@ limitations under the License. */ #include "paddle/operators/sum_op.h" #include +#include "paddle/operators/net_op.h" namespace paddle { namespace operators { @@ -21,31 +22,28 @@ class SumOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE(!ctx.MultiInputVar("X").empty(), - "Input(X) of SumOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"), - "Output(Out) of SumOp should not be null."); - - auto ins = ctx.MultiInput("X"); - auto *out = ctx.Output("Out"); - int N = ins.size(); - - auto in_dim = ins[0]->dims(); + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInputs("X"), "Inputs(X) should not be null"); + auto x_dims = ctx->GetInputsDim("X"); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of SumOp should not be null."); + size_t N = x_dims.size(); PADDLE_ENFORCE_GT(N, 1, "Input tensors count should > 1."); - for (int i = 1; i < N; i++) { - auto dim = ins[i]->dims(); + + auto in_dim = x_dims[0]; + for (size_t i = 1; i < N; i++) { + auto dim = x_dims[i]; PADDLE_ENFORCE(in_dim == dim, "Input tensors must have same shape"); } - out->Resize(in_dim); - ctx.ShareLoD("X", /*->*/ "Out"); + ctx->SetOutputDim("Out", in_dim); + ctx->ShareLoD("X", /*->*/ "Out"); } }; class SumOpMaker : public framework::OpProtoAndCheckerMaker { public: - SumOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + SumOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "the input tensors of sum operator.").AsDuplicable(); AddOutput("Out", "the output tensor of sum operator."); @@ -58,18 +56,26 @@ or not. But the output only shares the LoD with the first input. } }; -class SumGradOp : public framework::OperatorWithKernel { +class SumGradMaker : public framework::GradOpDescMakerBase { public: - using framework::OperatorWithKernel::OperatorWithKernel; + using framework::GradOpDescMakerBase::GradOpDescMakerBase; - protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - auto outputs = - ctx.MultiOutput(framework::GradVarName("X")); - auto dims = ctx.Input(framework::GradVarName("Out"))->dims(); - for (auto output : outputs) { - output->Resize(dims); - } + std::vector> operator()() + const override { + auto x_grads = InputGrad("X"); + std::vector> grad_ops; + grad_ops.reserve(x_grads.size()); + auto og = OutputGrad("Out"); + std::transform(x_grads.begin(), x_grads.end(), std::back_inserter(grad_ops), + [&og](const std::string& x_grad) { + auto* grad_op = new framework::OpDescBind(); + grad_op->SetType("scale"); + grad_op->SetInput("X", og); + grad_op->SetOutput("Out", {x_grad}); + grad_op->SetAttr("scale", 1.0f); + return std::unique_ptr(grad_op); + }); + return grad_ops; } }; @@ -77,7 +83,6 @@ class SumGradOp : public framework::OperatorWithKernel { } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP(sum, ops::SumOp, ops::SumOpMaker, sum_grad, ops::SumGradOp); + +REGISTER_OPERATOR(sum, ops::SumOp, ops::SumOpMaker, ops::SumGradMaker); REGISTER_OP_CPU_KERNEL(sum, ops::SumKernel); -REGISTER_OP_CPU_KERNEL(sum_grad, - ops::SumGradKernel); diff --git a/paddle/operators/sum_op.cu b/paddle/operators/sum_op.cu index a465cf3659ba7c51338abadfc62962fb6755a39d..b1896d3cd87f47bd2573287ee37b1b72ae9ec6e8 100644 --- a/paddle/operators/sum_op.cu +++ b/paddle/operators/sum_op.cu @@ -14,5 +14,3 @@ limitations under the License. */ namespace ops = paddle::operators; REGISTER_OP_GPU_KERNEL(sum, ops::SumKernel); -REGISTER_OP_GPU_KERNEL(sum_grad, - ops::SumGradKernel); diff --git a/paddle/operators/sum_op.h b/paddle/operators/sum_op.h index 0b1e9ebaa38d455fb5e3ce8c1a39cbbcdad9a940..91e5da8b40d452db8715990cdbe2731b3aea44b9 100644 --- a/paddle/operators/sum_op.h +++ b/paddle/operators/sum_op.h @@ -22,7 +22,7 @@ template ; template -class SumKernel : public framework::OpKernel { +class SumKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto ins = context.MultiInput("X"); @@ -42,24 +42,5 @@ class SumKernel : public framework::OpKernel { } }; -template -class SumGradKernel : public framework::OpKernel { - public: - void Compute(const framework::ExecutionContext& context) const override { - auto* input = context.Input(framework::GradVarName("Out")); - auto outs = context.MultiOutput(framework::GradVarName("X")); - for (auto out : outs) { - out->mutable_data(context.GetPlace()); - } - - auto place = context.GetEigenDevice(); - auto in = EigenVector::Flatten(*input); - for (auto out : outs) { - auto result = EigenVector::Flatten(*out); - result.device(place) = in; - } - } -}; - } // namespace operators } // namespace paddle diff --git a/paddle/operators/top_k_op.cc b/paddle/operators/top_k_op.cc index a6e43964e9825cd1ced9e7c1bc8d691422248fee..c95481991229dd40e162094ccbf15de8ab627c8b 100644 --- a/paddle/operators/top_k_op.cc +++ b/paddle/operators/top_k_op.cc @@ -22,26 +22,26 @@ class TopkOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), - "Input(X) of TopkOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"), - "Output(Out) of TopkOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Indices"), - "Output(Indices) of TopkOp should not be null."); + void InferShape(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of TopkOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of TopkOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Indices"), + "Output(Indices) of TopkOp should not be null."); - auto *input = ctx.Input("X"); - const int k = static_cast(ctx.Attr("k")); + auto input_dims = ctx->GetInputDim("X"); + const int k = static_cast(ctx->Attrs().Get("k")); PADDLE_ENFORCE_GE(k, 1, "k must >= 1"); - PADDLE_ENFORCE_GE(input->dims().size(), 1, "input must have >= 1d shape"); - PADDLE_ENFORCE_GE(input->dims()[input->dims().size() - 1], k, + PADDLE_ENFORCE_GE(input_dims.size(), 1, "input must have >= 1d shape"); + PADDLE_ENFORCE_GE(input_dims[input_dims.size() - 1], k, "input must have >= k columns"); - framework::DDim dims = input->dims(); + framework::DDim dims = input_dims; dims[dims.size() - 1] = k; - ctx.Output("Out")->Resize(dims); - ctx.Output("Indices")->Resize(dims); + ctx->SetOutputDim("Out", dims); + ctx->SetOutputDim("Indices", dims); } }; diff --git a/paddle/operators/top_k_op.cu b/paddle/operators/top_k_op.cu index afe4d149c53819c45e20353bc9d16393f3f61e0f..7be6932f1e301d06e0e232367a38bfa673ff45be 100644 --- a/paddle/operators/top_k_op.cu +++ b/paddle/operators/top_k_op.cu @@ -279,7 +279,7 @@ __global__ void KeMatrixTopK(T* output, int output_stride, int* indices, } template -class TopkOpCUDAKernel : public framework::OpKernel { +class TopkOpCUDAKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), @@ -301,14 +301,16 @@ class TopkOpCUDAKernel : public framework::OpKernel { // NOTE: pass lds and dim same to input width. // NOTE: old matrix implementation of stride is different to eigen. - // TODO(typhoonzero): launch kernel on specified stream. // TODO(typhoonzero): refine this kernel. dim3 threads(256, 1); dim3 grid(input_height, 1); - KeMatrixTopK<<>>( - output_data, output->dims()[1], indices_data, input_data, input_width, - input_width, int(k)); + KeMatrixTopK<<< + grid, threads, 0, reinterpret_cast( + ctx.device_context()) + .stream()>>>(output_data, output->dims()[1], + indices_data, input_data, + input_width, input_width, int(k)); } }; diff --git a/paddle/operators/top_k_op.h b/paddle/operators/top_k_op.h index ef66acc1d569282a42be64b7a5e90f3fbdb20690..4b248faa120bcfd20e70d288cce2d485d3e6371e 100644 --- a/paddle/operators/top_k_op.h +++ b/paddle/operators/top_k_op.h @@ -28,7 +28,7 @@ template ; template -class TopkKernel : public framework::OpKernel { +class TopkKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { // Get the top k elements of each row of input tensor diff --git a/paddle/operators/transpose_op.cc b/paddle/operators/transpose_op.cc index 017a05326e9b397185d7c3530891884b11784783..1101bbe3efe3d313400663ade561063da4e34dfd 100644 --- a/paddle/operators/transpose_op.cc +++ b/paddle/operators/transpose_op.cc @@ -24,12 +24,11 @@ class TransposeOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) should not be null"); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"), - "Output(Out) should not be null"); - auto x_dims = ctx.Input("X")->dims(); - std::vector axis = ctx.Attr>("axis"); + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null"); + PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) should not be null"); + auto x_dims = ctx->GetInputDim("X"); + std::vector axis = ctx->Attrs().Get>("axis"); size_t x_rank = x_dims.size(); size_t axis_size = axis.size(); @@ -51,14 +50,14 @@ class TransposeOp : public framework::OperatorWithKernel { for (size_t i = 0; i < axis_size; i++) { out_dims[i] = x_dims[axis[i]]; } - ctx.Output("Out")->Resize(out_dims); + ctx->SetOutputDim("Out", out_dims); } }; class TransposeOpMaker : public framework::OpProtoAndCheckerMaker { public: - TransposeOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + TransposeOpMaker(framework::OpProto* proto, + framework::OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput( "X", @@ -79,7 +78,7 @@ For example: [3, 4, 5]]) >> axis = [1, 0] >> output = input.transpose(axis) - >> output + >> output array([[0, 3], [1, 4], [2, 5]]) @@ -94,14 +93,15 @@ class TransposeOpGrad : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) should not be null"); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")), - "Input(Out@GRAD) should not be null"); - auto x_dims = ctx.Input("X")->dims(); - auto *x_grad = ctx.Output(framework::GradVarName("X")); - - if (x_grad) x_grad->Resize(x_dims); + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null"); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), + "Input(Out@GRAD) should not be null"); + auto x_dims = ctx->GetInputDim("X"); + ctx->SetOutputDim(framework::GradVarName("X"), x_dims); + if (ctx->HasOutput(framework::GradVarName("X"))) { + ctx->SetOutputDim(framework::GradVarName("X"), x_dims); + } } }; diff --git a/paddle/operators/transpose_op.h b/paddle/operators/transpose_op.h index ea299dce72ad340b0a65ee50582dc156b5ad7abb..aaa3f47ab5545accd4d1108e0ad6f5a3062186d0 100644 --- a/paddle/operators/transpose_op.h +++ b/paddle/operators/transpose_op.h @@ -38,7 +38,7 @@ void EigenTranspose(const framework::ExecutionContext& context, } template -class TransposeKernel : public framework::OpKernel { +class TransposeKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto* x = context.Input("X"); @@ -73,7 +73,7 @@ class TransposeKernel : public framework::OpKernel { }; template -class TransposeGradKernel : public framework::OpKernel { +class TransposeGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto* out_grad = diff --git a/paddle/operators/uniform_random_op.cc b/paddle/operators/uniform_random_op.cc index 17ea48361bc597ccfeb80884d51900e6567aa057..e330877fc4283b796dcb5c5d745881884ae491ae 100644 --- a/paddle/operators/uniform_random_op.cc +++ b/paddle/operators/uniform_random_op.cc @@ -21,20 +21,20 @@ namespace operators { // Use std::random and thrust::random(thrust is a std library in CUDA) to // implement uniform random. template -class CPUUniformRandomKernel : public framework::OpKernel { +class CPUUniformRandomKernel : public framework::OpKernel { public: - void Compute(const framework::ExecutionContext& context) const override { - auto* tensor = context.Output("Out"); - T* data = tensor->mutable_data(context.GetPlace()); - unsigned int seed = static_cast(context.Attr("seed")); + void Compute(const framework::ExecutionContext& ctx) const override { + auto* tensor = ctx.Output("Out"); + T* data = tensor->mutable_data(ctx.GetPlace()); + unsigned int seed = static_cast(ctx.Attr("seed")); std::minstd_rand engine; if (seed == 0) { seed = std::random_device()(); } engine.seed(seed); std::uniform_real_distribution dist( - static_cast(context.Attr("min")), - static_cast(context.Attr("max"))); + static_cast(ctx.Attr("min")), + static_cast(ctx.Attr("max"))); int64_t size = tensor->numel(); for (int64_t i = 0; i < size; ++i) { data[i] = dist(engine); @@ -47,21 +47,25 @@ class UniformRandomOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext& ctx) const override { - PADDLE_ENFORCE_NOT_NULL( - ctx.OutputVar("Out"), - "Output(Out) of UniformRandomOp should not be null."); + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of UniformRandomOp should not be null."); - PADDLE_ENFORCE(Attr("min") < Attr("max"), - "uniform_random's min must less then max"); - auto* tensor = ctx.Output("Out"); + PADDLE_ENFORCE( + ctx->Attrs().Get("min") < ctx->Attrs().Get("max"), + "uniform_random's min must less then max"); auto dims = Attr>("dims"); std::vector temp; temp.reserve(dims.size()); for (auto dim : dims) { temp.push_back(static_cast(dim)); } - tensor->Resize(framework::make_ddim(temp)); + ctx->SetOutputDim("Out", framework::make_ddim(temp)); + } + + framework::DataType IndicateDataType( + const framework::ExecutionContext& ctx) const override { + return static_cast(Attr("data_type")); } }; @@ -81,6 +85,8 @@ Used to initialize tensor with uniform random generator. "Random seed of uniform random. " "0 means generate a seed by system") .SetDefault(0); + AddAttr("data_type", "output tensor data type") + .SetDefault(framework::DataType::FP32); } }; } // namespace operators diff --git a/paddle/operators/uniform_random_op.cu b/paddle/operators/uniform_random_op.cu index 6614b53b3f990d10c82633f3c1f079acea0cd827..5612ce9eb1c644d6271b4a9bb949f685848e05c0 100644 --- a/paddle/operators/uniform_random_op.cu +++ b/paddle/operators/uniform_random_op.cu @@ -40,7 +40,7 @@ struct UniformGenerator { // Use std::random and thrust::random(thrust is a std library in CUDA) to // implement uniform random. template -class GPUUniformRandomKernel : public framework::OpKernel { +class GPUUniformRandomKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto* tensor = context.Output("Out"); diff --git a/paddle/parameter/FirstOrderOptimizer.h b/paddle/parameter/FirstOrderOptimizer.h index caa78acd98ea4b35fc69643689cfce23026275e0..895e8d6a63d1fad0ee7a6f5647402435d418b2f1 100644 --- a/paddle/parameter/FirstOrderOptimizer.h +++ b/paddle/parameter/FirstOrderOptimizer.h @@ -15,6 +15,7 @@ limitations under the License. */ #pragma once #include "ParameterOptimizer.h" +#include "ParameterUpdateFunctions.h" #include "Regularizer.h" namespace paddle { @@ -37,6 +38,15 @@ public: real torch_learningRate = optConfig_.learning_method() == "torch_momentum" ? 1.0 - paraConfig.momentum() : 1.0; +#ifdef PADDLE_USE_MKLDNN + sgdUpdate(learningRate_ * paraConfig.learning_rate() * + (firstTime_ ? 1.0 : torch_learningRate), + paraConfig.momentum(), + applyDecay_ ? paraConfig.decay_rate() : 0, + vecs[PARAMETER_VALUE].get(), + vecs[PARAMETER_GRADIENT].get(), + vecs[PARAMETER_MOMENTUM].get()); +#else vecs[PARAMETER_VALUE]->sgdUpdate( *vecs[PARAMETER_GRADIENT], *vecs[PARAMETER_MOMENTUM], @@ -44,6 +54,7 @@ public: (firstTime_ ? 1.0 : torch_learningRate), paraConfig.momentum(), applyDecay_ ? paraConfig.decay_rate() : 0); +#endif } virtual void finishBatch() { firstTime_ = false; } }; diff --git a/paddle/parameter/ParameterUpdateFunctions.cpp b/paddle/parameter/ParameterUpdateFunctions.cpp index c8af7105c78dcbf9f625a348b7f38efcf278469e..8b3be062b654a52e667626199be8c8bb4a2a96d7 100644 --- a/paddle/parameter/ParameterUpdateFunctions.cpp +++ b/paddle/parameter/ParameterUpdateFunctions.cpp @@ -30,6 +30,9 @@ void sgdUpdateCpu(real learningRate, const real* grad, real* momentumVec) { decayRate *= learningRate; +#ifdef PADDLE_USE_MKLDNN +#pragma omp parallel for +#endif for (size_t i = 0; i < size; ++i) { momentumVec[i] = momentum * momentumVec[i] - learningRate * grad[i] - decayRate * value[i]; diff --git a/paddle/platform/device_context.cc b/paddle/platform/device_context.cc index 93b472b41c8a4c3a2bfada9d4fbf0e9e1b0cc736..36450e926891342f37424447703781a33c1190ae 100644 --- a/paddle/platform/device_context.cc +++ b/paddle/platform/device_context.cc @@ -16,8 +16,8 @@ namespace paddle { namespace platform { template <> -Eigen::DefaultDevice* DeviceContext::get_eigen_device() - const { +Eigen::DefaultDevice* DeviceContext::GetEigenDevice< + platform::CPUPlace, Eigen::DefaultDevice>() const { return reinterpret_cast(this)->eigen_device(); } @@ -35,7 +35,13 @@ Eigen::DefaultDevice* CPUDeviceContext::eigen_device() const { Place CPUDeviceContext::GetPlace() const { return CPUPlace(); } -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA + +template <> +Eigen::GpuDevice* +DeviceContext::GetEigenDevice() const { + return reinterpret_cast(this)->eigen_device(); +} class EigenCudaStreamDevice : public Eigen::StreamInterface { public: @@ -90,11 +96,6 @@ class EigenCudaStreamDevice : public Eigen::StreamInterface { mutable unsigned int* semaphore_; }; -template <> -Eigen::GpuDevice* DeviceContext::get_eigen_device() const { - return reinterpret_cast(this)->eigen_device(); -} - CUDADeviceContext::CUDADeviceContext(GPUPlace place) : place_(place) { SetDeviceId(place_.device); PADDLE_ENFORCE(cudaStreamCreate(&stream_)); @@ -135,7 +136,7 @@ cudnnHandle_t CUDADeviceContext::cudnn_handle() const { return cudnn_handle_; } cudaStream_t CUDADeviceContext::stream() const { return stream_; } -#endif // PADDLE_ONLY_CPU +#endif } // namespace platform } // namespace paddle diff --git a/paddle/platform/device_context.h b/paddle/platform/device_context.h index f6a39a8e26c301296aac0af7f4e8b2c6c97ece24..ef5f19214d9ccb23b9c946bee28cb764122bd7cd 100644 --- a/paddle/platform/device_context.h +++ b/paddle/platform/device_context.h @@ -14,7 +14,7 @@ limitations under the License. */ #include "paddle/platform/enforce.h" #include "paddle/platform/place.h" -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA #include "paddle/platform/dynload/cublas.h" #include "paddle/platform/dynload/cudnn.h" #include "paddle/platform/gpu_info.h" @@ -27,13 +27,23 @@ limitations under the License. */ namespace paddle { namespace platform { +template +struct EigenDeviceConverter; + +template <> +struct EigenDeviceConverter { + using EigenDeviceType = Eigen::DefaultDevice; +}; + class DeviceContext { public: virtual ~DeviceContext() {} virtual Place GetPlace() const = 0; - template - DeviceType* get_eigen_device() const; + template ::EigenDeviceType> + DeviceType* GetEigenDevice() const; virtual void Wait() const {} }; @@ -51,7 +61,12 @@ class CPUDeviceContext : public DeviceContext { std::unique_ptr eigen_device_; }; -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA +template <> +struct EigenDeviceConverter { + using EigenDeviceType = Eigen::GpuDevice; +}; + class EigenCudaStreamDevice; class CUDADeviceContext : public DeviceContext { diff --git a/paddle/platform/device_context_test.cc b/paddle/platform/device_context_test.cc index 5883a55272f0f24c94d48bc43c62ddb7bef15465..8bf5174c4a5579f6f5602dd38e5a87ed3ef444a7 100644 --- a/paddle/platform/device_context_test.cc +++ b/paddle/platform/device_context_test.cc @@ -20,11 +20,11 @@ TEST(Device, Init) { using paddle::platform::CUDADeviceContext; using paddle::platform::GPUPlace; - int count = paddle::platform::GetDeviceCount(); + int count = paddle::platform::GetCUDADeviceCount(); for (int i = 0; i < count; i++) { DeviceContext* device_context = new CUDADeviceContext(GPUPlace(i)); Eigen::GpuDevice* gpu_device = - device_context->template get_eigen_device(); + device_context->template GetEigenDevice(); ASSERT_NE(nullptr, gpu_device); delete device_context; } @@ -34,7 +34,7 @@ TEST(Device, CUDADeviceContext) { using paddle::platform::CUDADeviceContext; using paddle::platform::GPUPlace; - int count = paddle::platform::GetDeviceCount(); + int count = paddle::platform::GetCUDADeviceCount(); for (int i = 0; i < count; i++) { CUDADeviceContext* device_context = new CUDADeviceContext(GPUPlace(i)); Eigen::GpuDevice* gpu_device = device_context->eigen_device(); diff --git a/paddle/platform/enforce.h b/paddle/platform/enforce.h index df5f71ed760952ed042d7ffa40a4319a73fb93bf..cd906c3fa9375cd6edaed0377a596771e25043d4 100644 --- a/paddle/platform/enforce.h +++ b/paddle/platform/enforce.h @@ -29,7 +29,7 @@ limitations under the License. */ #include // for __cxa_demangle #endif -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA #include "paddle/platform/dynload/cublas.h" #include "paddle/platform/dynload/cudnn.h" @@ -41,7 +41,7 @@ limitations under the License. */ #include #include -#endif // PADDLE_ONLY_CPU +#endif namespace paddle { namespace platform { @@ -107,13 +107,13 @@ struct EnforceNotMet : public std::exception { template inline typename std::enable_if::type throw_on_error( - int stat, const Args&... args) { + bool stat, const Args&... args) { if (UNLIKELY(!(stat))) { throw std::runtime_error(string::Sprintf(args...)); } } -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA template inline typename std::enable_if::type throw_on_error( @@ -185,7 +185,7 @@ inline void throw_on_error(T e) { std::make_exception_ptr( \ std::runtime_error(paddle::string::Sprintf(__VA_ARGS__))), \ __FILE__, __LINE__); \ - } while (0) + } while (false) #define PADDLE_ENFORCE(...) \ do { \ @@ -195,7 +195,7 @@ inline void throw_on_error(T e) { throw ::paddle::platform::EnforceNotMet(std::current_exception(), \ __FILE__, __LINE__); \ } \ - } while (0) + } while (false) /* * Some enforce helpers here, usage: diff --git a/paddle/platform/enforce_test.cc b/paddle/platform/enforce_test.cc index 80bdee3d9dfbe38ef707a6ba60cdb7f7b99714de..8206a055eabf4abf584962e921610d5029e2f571 100644 --- a/paddle/platform/enforce_test.cc +++ b/paddle/platform/enforce_test.cc @@ -213,4 +213,4 @@ TEST(ENFORCE_USER_DEFINED_CLASS, EQ) { TEST(ENFORCE_USER_DEFINED_CLASS, NE) { Dims a{{1, 2, 3, 4}}, b{{5, 6, 7, 8}}; ASSERT_THROW(PADDLE_ENFORCE_EQ(a, b), paddle::platform::EnforceNotMet); -} \ No newline at end of file +} diff --git a/paddle/platform/gpu_info.cc b/paddle/platform/gpu_info.cc index be381a4e26cf0eb41f5b3de88bd03ad8901683cc..70ad611d5dd61937e6bf7d980e34b5c9023977b2 100644 --- a/paddle/platform/gpu_info.cc +++ b/paddle/platform/gpu_info.cc @@ -26,11 +26,11 @@ DEFINE_double(fraction_of_gpu_memory_to_use, 0.95, namespace paddle { namespace platform { -int GetDeviceCount() { +int GetCUDADeviceCount() { int count; PADDLE_ENFORCE( cudaGetDeviceCount(&count), - "cudaGetDeviceCount failed in paddle::platform::GetDeviceCount"); + "cudaGetDeviceCount failed in paddle::platform::GetCUDADeviceCount"); return count; } diff --git a/paddle/platform/gpu_info.h b/paddle/platform/gpu_info.h index f0c825bd9b0bc41396b8fdb95f0b4337cbe3db02..37665b97d764fbcfe0964127d230b1d28d90b687 100644 --- a/paddle/platform/gpu_info.h +++ b/paddle/platform/gpu_info.h @@ -14,7 +14,7 @@ limitations under the License. */ #pragma once -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA #include #include @@ -28,7 +28,7 @@ const std::string kEnvFractionGpuMemoryToUse = "PADDLE_FRACTION_GPU_MEMORY_TO_USE"; //! Get the total number of GPU devices in system. -int GetDeviceCount(); +int GetCUDADeviceCount(); //! Get the current GPU device id in system. int GetCurrentDeviceId(); @@ -63,4 +63,4 @@ void GpuMemcpyPeer(void *dst, int dst_device, const void *src, int src_device, } // namespace platform } // namespace paddle -#endif // PADDLE_ONLY_CPU +#endif diff --git a/paddle/platform/hostdevice.h b/paddle/platform/hostdevice.h index e7de86b7b2f75d206e730ec409bbee5d0a08942e..eb2df291cceef553d6422e6166e1fef2c63e2a47 100644 --- a/paddle/platform/hostdevice.h +++ b/paddle/platform/hostdevice.h @@ -2,8 +2,10 @@ #ifdef __CUDACC__ #define HOSTDEVICE __host__ __device__ +#define DEVICE __device__ #define HOST __host__ #else #define HOSTDEVICE +#define DEVICE #define HOST #endif diff --git a/paddle/platform/macros.h b/paddle/platform/macros.h index 4a04a38c0c6f905639004dea2f4416ecc57c8620..feae7bdd77e3a0d02f33fb33991648408f542d0e 100644 --- a/paddle/platform/macros.h +++ b/paddle/platform/macros.h @@ -16,8 +16,10 @@ limitations under the License. */ // Disable the copy and assignment operator for a class. #ifndef DISABLE_COPY_AND_ASSIGN -#define DISABLE_COPY_AND_ASSIGN(classname) \ - private: \ - classname(const classname&) = delete; \ - classname& operator=(const classname&) = delete +#define DISABLE_COPY_AND_ASSIGN(classname) \ + private: \ + classname(const classname&) = delete; \ + classname(const classname&&) = delete; \ + classname& operator=(const classname&) = delete; \ + classname& operator=(const classname&&) = delete #endif diff --git a/paddle/platform/place.cc b/paddle/platform/place.cc index b31515e1f028acac885a506ff1c20479407a05e3..856e54df89c1c18ade040957188a2fbda0901473 100644 --- a/paddle/platform/place.cc +++ b/paddle/platform/place.cc @@ -47,7 +47,7 @@ bool is_cpu_place(const Place &p) { } bool places_are_same_class(const Place &p1, const Place &p2) { - return is_gpu_place(p1) == is_gpu_place(p2); + return p1.which() == p2.which(); } std::ostream &operator<<(std::ostream &os, const Place &p) { diff --git a/paddle/platform/place.h b/paddle/platform/place.h index 1117476bb37f1b0f3876c55e610803d5ee2558ce..0efc6932349a5b3ad295d195a16737a642e18943 100644 --- a/paddle/platform/place.h +++ b/paddle/platform/place.h @@ -15,6 +15,7 @@ limitations under the License. */ #pragma once #include + #include "paddle/platform/variant.h" namespace paddle { @@ -46,8 +47,18 @@ struct IsGPUPlace : public boost::static_visitor { bool operator()(const GPUPlace &gpu) const { return true; } }; +// Define the max number of Place in bit length. i.e., the max number of places +// should be less equal than 2^(NUM_PLACE_TYPE_LIMIT_IN_BIT) +#define NUM_PLACE_TYPE_LIMIT_IN_BIT 4 + typedef boost::variant Place; +// static check number of place types is less equal than +// 2^(NUM_PLACE_TYPE_LIMIT_IN_BIT) +BOOST_MPL_ASSERT((boost::mpl::less_equal< + Place::types::size, + boost::mpl::long_<1 << NUM_PLACE_TYPE_LIMIT_IN_BIT>>)); + void set_place(const Place &); const Place &get_place(); diff --git a/paddle/platform/variant.h b/paddle/platform/variant.h index c2257af1b5dd1a1e284979bf17e1a947072baa85..619897ca19eb2e6f4dbfd9160edf8c4bc58c89a9 100644 --- a/paddle/platform/variant.h +++ b/paddle/platform/variant.h @@ -16,7 +16,7 @@ #include -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA // Because boost's variadic templates has bug on nvcc, boost will disable // variadic template support when GPU enabled on nvcc. @@ -29,4 +29,6 @@ #endif #endif +#include +#include #include diff --git a/paddle/pserver/test/SocketTest.cpp b/paddle/pserver/test/SocketTest.cpp index 6f6c9e596cfb7a2547d5b6c5de69381eb9c29132..b43461d61bab21747e85090bbf7af21a87a670c6 100644 --- a/paddle/pserver/test/SocketTest.cpp +++ b/paddle/pserver/test/SocketTest.cpp @@ -215,7 +215,7 @@ int main(int argc, char** argv) { uint64_t dataSize = FLAGS_dim * sizeof(real); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA GpuVector gpuParam(FLAGS_dim); GpuVector gpuGrad(FLAGS_dim); #else diff --git a/paddle/pserver/test/test_ProtoServer.cpp b/paddle/pserver/test/test_ProtoServer.cpp index 04236fda2fb62b928b5c06ff38acfd3eb7217b08..ad8ffed9c1c8e4bdef27689ab21950db6b5cf0a2 100644 --- a/paddle/pserver/test/test_ProtoServer.cpp +++ b/paddle/pserver/test/test_ProtoServer.cpp @@ -99,7 +99,7 @@ TEST(ProtoServer, regular) { } TEST(ProtoServer, extended) { -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA ProtoClient* client; if (FLAGS_rdma_tcp == "rdma") client = new ProtoClient(FLAGS_server_addr, FLAGS_port, F_RDMA); diff --git a/paddle/pybind/.clang-format b/paddle/pybind/.clang-format new file mode 120000 index 0000000000000000000000000000000000000000..7d28cb3924707d39dafe20f4664fb17b5538996c --- /dev/null +++ b/paddle/pybind/.clang-format @@ -0,0 +1 @@ +../framework/.clang-format \ No newline at end of file diff --git a/paddle/pybind/CMakeLists.txt b/paddle/pybind/CMakeLists.txt index 4f05406c7f74113d8fb10aa6914166e553858338..97364f2db9523c0629616692631d8372657a2128 100644 --- a/paddle/pybind/CMakeLists.txt +++ b/paddle/pybind/CMakeLists.txt @@ -1,6 +1,6 @@ if(WITH_PYTHON) cc_library(paddle_pybind SHARED - SRCS pybind.cc - DEPS pybind python backward + SRCS pybind.cc exception.cc protobuf.cc + DEPS pybind python backward proto_desc tensor_array ${GLOB_OP_LIB}) endif(WITH_PYTHON) diff --git a/paddle/pybind/exception.cc b/paddle/pybind/exception.cc new file mode 100644 index 0000000000000000000000000000000000000000..ff79b12ee4b28c53ee04f4c170b5bca9ca28d14a --- /dev/null +++ b/paddle/pybind/exception.cc @@ -0,0 +1,34 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/pybind/exception.h" + +namespace paddle { +namespace pybind { + +void BindException(pybind11::module& m) { + static pybind11::exception exc(m, "EnforceNotMet"); + pybind11::register_exception_translator([](std::exception_ptr p) { + try { + if (p) std::rethrow_exception(p); + } catch (const platform::EnforceNotMet& e) { + exc(e.what()); + } + }); + + m.def("__unittest_throw_exception__", [] { PADDLE_THROW("test exception"); }); +} + +} // namespace pybind +} // namespace paddle diff --git a/paddle/pybind/exception.h b/paddle/pybind/exception.h new file mode 100644 index 0000000000000000000000000000000000000000..70beac146046f74e23f747bab130483901a7d443 --- /dev/null +++ b/paddle/pybind/exception.h @@ -0,0 +1,24 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once +#include +#include "paddle/platform/enforce.h" +#include "pybind11/pybind11.h" +namespace paddle { +namespace pybind { + +extern void BindException(pybind11::module& m); +} // namespace pybind +} // namespace paddle diff --git a/paddle/pybind/protobuf.cc b/paddle/pybind/protobuf.cc new file mode 100644 index 0000000000000000000000000000000000000000..6333cc332e9a4da8ecfe7e92cd2bca622b56fc18 --- /dev/null +++ b/paddle/pybind/protobuf.cc @@ -0,0 +1,204 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/pybind/protobuf.h" +#include +#include +#include "paddle/framework/block_desc.h" +#include "paddle/framework/op_desc.h" +#include "paddle/framework/program_desc.h" +#include "paddle/framework/var_desc.h" + +// Cast boost::variant for PyBind. +// Copy from +// https://github.com/pybind/pybind11/issues/576#issuecomment-269563199 +namespace pybind11 { +namespace detail { + +// Can be replaced by a generic lambda in C++14 +struct variant_caster_visitor : public boost::static_visitor { + return_value_policy policy; + handle parent; + + variant_caster_visitor(return_value_policy policy, handle parent) + : policy(policy), parent(parent) {} + + template + handle operator()(T const &src) const { + return make_caster::cast(src, policy, parent); + } +}; + +template +struct variant_caster; + +template