diff --git a/benchmark/paddle/image/run_mkldnn.sh b/benchmark/paddle/image/run_mkldnn.sh index a4527e04968cf8c8c3c31d16f50bc3e28381f6d8..3cc779b48d082985f75ab1c053fbe262bc6d58aa 100755 --- a/benchmark/paddle/image/run_mkldnn.sh +++ b/benchmark/paddle/image/run_mkldnn.sh @@ -1,9 +1,7 @@ set -e function train() { - unset OMP_NUM_THREADS MKL_NUM_THREADS - export OMP_DYNAMIC="FALSE" - export KMP_AFFINITY="granularity=fine,compact,0,0" + unset OMP_NUM_THREADS MKL_NUM_THREADS OMP_DYNAMIC KMP_AFFINITY topology=$1 layer_num=$2 bs=$3 @@ -14,8 +12,6 @@ function train() { elif [ $4 == "False" ]; then thread=`nproc` # each trainer_count use only 1 core to avoid conflict - export OMP_NUM_THREADS=1 - export MKL_NUM_THREADS=1 log="logs/${topology}-${layer_num}-${thread}mklml-${bs}.log" else echo "Wrong input $3, use True or False." diff --git a/cmake/cross_compiling/ios.cmake b/cmake/cross_compiling/ios.cmake index 310450f7d009dc0cdae9c0079a96445af8ec8f95..d3f5bf6852b3b295f3b5806b0577a880b0ce6ba6 100644 --- a/cmake/cross_compiling/ios.cmake +++ b/cmake/cross_compiling/ios.cmake @@ -76,11 +76,9 @@ set(IOS_PLATFORM ${IOS_PLATFORM} CACHE STRING "Type of iOS Platform") # Set the architecture for iOS if(NOT DEFINED IOS_ARCH) if(IOS_PLATFORM STREQUAL "OS") - # FIXME(liuyiqun): support "armv7;armv7s;arm64" future - set(IOS_ARCH "arm64") + set(IOS_ARCH "armv7;armv7s;arm64") elseif(IOS_PLATFORM STREQUAL "SIMULATOR") - # FIXME(liuyiqun): support "i386;x86_64" future - set(IOS_ARCH "x86_64") + set(IOS_ARCH "i386;x86_64") endif() endif() set(CMAKE_OSX_ARCHITECTURES ${IOS_ARCH} CACHE string "Build architecture for iOS") @@ -248,7 +246,7 @@ set(IOS_COMPILER_FLAGS "${XCODE_IOS_PLATFORM_VERSION_FLAGS} ${XCODE_IOS_BITCODE_ # Hidden visibilty is required for cxx on iOS set(CMAKE_C_FLAGS "${IOS_COMPILER_FLAGS} ${CMAKE_C_FLAGS}" CACHE STRING "C flags") -set(CMAKE_CXX_FLAGS "${IOS_COMPILER_FLAGS} -fvisibility-inlines-hidden ${CMAKE_CXX_FLAGS}" CACHE STRING "CXX flags") +set(CMAKE_CXX_FLAGS "${IOS_COMPILER_FLAGS} -fvisibility=hidden -fvisibility-inlines-hidden ${CMAKE_CXX_FLAGS}" CACHE STRING "CXX flags") set(IOS_LINK_FLAGS "${XCODE_IOS_PLATFORM_VERSION_FLAGS} -Wl,-search_paths_first") diff --git a/cmake/external/openblas.cmake b/cmake/external/openblas.cmake index 324e29f931ecbb6beab2d363daa01a19b1a56b3e..4c4f59656dae68739f2f07f3febd510e727fe2dd 100644 --- a/cmake/external/openblas.cmake +++ b/cmake/external/openblas.cmake @@ -29,7 +29,7 @@ IF(NOT ${CBLAS_FOUND}) "${CBLAS_INSTALL_DIR}/lib/${CMAKE_STATIC_LIBRARY_PREFIX}openblas${CMAKE_STATIC_LIBRARY_SUFFIX}" CACHE FILEPATH "openblas library." FORCE) - SET(OPENBLAS_CC "${CMAKE_C_COMPILER}") + SET(OPENBLAS_CC "${CMAKE_C_COMPILER} -Wno-unused-but-set-variable -Wno-unused-variable") IF(CMAKE_CROSSCOMPILING) SET(OPTIONAL_ARGS HOSTCC=${HOST_C_COMPILER}) @@ -45,15 +45,14 @@ IF(NOT ${CBLAS_FOUND}) SET(OPTIONAL_ARGS ${OPTIONAL_ARGS} TARGET=ARMV8 BINARY=64 USE_THREAD=0) ENDIF() ELSEIF(IOS) - # FIXME(liuyiqun): support multiple architectures - SET(OPENBLAS_COMMIT "b5c96fcfcdc82945502a2303116a64d89985daf5") - SET(OPENBLAS_CC "${OPENBLAS_CC} ${CMAKE_C_FLAGS} -isysroot ${CMAKE_OSX_SYSROOT}") - IF(CMAKE_OSX_ARCHITECTURES MATCHES "armv7") - SET(OPENBLAS_CC "${OPENBLAS_CC} -arch armv7") - SET(OPTIONAL_ARGS ${OPTIONAL_ARGS} TARGET=ARMV7 ARM_SOFTFP_ABI=1 USE_THREAD=0) - ELSEIF(CMAKE_OSX_ARCHITECTURES MATCHES "arm64") + IF(CMAKE_OSX_ARCHITECTURES MATCHES "arm64") + SET(OPENBLAS_COMMIT "b5c96fcfcdc82945502a2303116a64d89985daf5") + SET(OPENBLAS_CC "${OPENBLAS_CC} ${CMAKE_C_FLAGS} -isysroot ${CMAKE_OSX_SYSROOT}") SET(OPENBLAS_CC "${OPENBLAS_CC} -arch arm64") SET(OPTIONAL_ARGS ${OPTIONAL_ARGS} TARGET=ARMV8 BINARY=64 USE_THREAD=0 CROSS_SUFFIX=${CROSS_SUFFIX}) + ELSE() + MESSAGE(FATAL_ERROR "OpenBLAS only support arm64 architectures on iOS. " + "You can set IOS_USE_VECLIB_FOR_BLAS=ON or USE_EIGEN_FOR_BLAS=ON to use other blas library instead.") ENDIF() ELSEIF(RPI) # use hardfp diff --git a/cmake/external/warpctc.cmake b/cmake/external/warpctc.cmake index 8bd058222880b4df3b08da09c02f9fe7f1d0ee66..a8e1aca49c97df256b1269c286b0bce7732fa932 100644 --- a/cmake/external/warpctc.cmake +++ b/cmake/external/warpctc.cmake @@ -12,6 +12,10 @@ # See the License for the specific language governing permissions and # limitations under the License. +IF(MOBILE_INFERENCE) + return() +ENDIF() + INCLUDE(ExternalProject) SET(WARPCTC_SOURCES_DIR ${THIRD_PARTY_PATH}/warpctc) diff --git a/doc/design/ops/images/2_level_rnn.dot b/doc/design/ops/images/2_level_rnn.dot index a498e882a3d85a33d44dbad7474fa2a340e33976..5d77865061ca7bbbfcf254dd938f09aef5553505 100644 --- a/doc/design/ops/images/2_level_rnn.dot +++ b/doc/design/ops/images/2_level_rnn.dot @@ -1,6 +1,6 @@ digraph G { - rnn [label="1-th level RNN" shape=box] + rnn [label="1st level RNN" shape=box] subgraph cluster0 { label = "time step 0" @@ -8,7 +8,7 @@ digraph G { sent0 [label="sentence"] sent1 [label="sentence"] - rnn1 [label="2-th level RNN" shape=box] + rnn1 [label="2nd level RNN" shape=box] sent0 -> rnn1 sent1 -> rnn1 @@ -20,7 +20,7 @@ digraph G { sent2 [label="sentence"] sent3 [label="sentence"] - rnn2 [label="2-th level RNN" shape=box] + rnn2 [label="2nd level RNN" shape=box] sent2 -> rnn2 sent3 -> rnn2 @@ -32,7 +32,7 @@ digraph G { sent4 [label="sentence"] sent5 [label="sentence"] - rnn3 [label="2-th level RNN" shape=box] + rnn3 [label="2nd level RNN" shape=box] sent4 -> rnn3 sent5 -> rnn3 diff --git a/doc/design/ops/rnn.md b/doc/design/ops/rnn.md index a78eea7d45e9e9553d153170aa31da55ec6e8289..2f4854793fa1f0b02e4dc17b51a48a972be61c06 100644 --- a/doc/design/ops/rnn.md +++ b/doc/design/ops/rnn.md @@ -1,62 +1,62 @@ # RNNOp design -This document is about an RNN operator which requires that instances in a mini-batch have the same length. We will have a more flexible RNN operator. +This document describes the RNN (Recurrent Neural Network) operator and how it is implemented in PaddlePaddle. The RNN op requires that all instances in a mini-batch have the same length. We will have a more flexible dynamic RNN operator in the future. ## RNN Algorithm Implementation -

+

The above diagram shows an RNN unrolled into a full network. -There are several important concepts: +There are several important concepts here: -- *step-net*: the sub-graph to run at each step, -- *memory*, $h_t$, the state of the current step, -- *ex-memory*, $h_{t-1}$, the state of the previous step, -- *initial memory value*, the ex-memory of the first step. +- *step-net*: the sub-graph that runs at each step. +- *memory*, $h_t$, the state of the current step. +- *ex-memory*, $h_{t-1}$, the state of the previous step. +- *initial memory value*, the memory of the first (initial) step. ### Step-scope -There could be local variables defined in step-nets. PaddlePaddle runtime realizes these variables in *step-scopes* -- scopes created for each step. +There could be local variables defined in each step-net. PaddlePaddle runtime realizes these variables in *step-scopes* which are created for each step. -

+


-Figure 2 the RNN's data flow +Figure 2 illustrates the RNN's data flow

-Please be aware that all steps run the same step-net. Each step +Please be aware that every step runs the same step-net. Each step does the following: -1. creates the step-scope, -2. realizes local variables, including step-outputs, in the step-scope, and -3. runs the step-net, which could use these variables. +1. Creates the step-scope. +2. Initializes the local variables including step-outputs, in the step-scope. +3. Runs the step-net, which uses the above mentioned variables. -The RNN operator will compose its output from step outputs in step scopes. +The RNN operator will compose its output from step outputs in each of the step scopes. ### Memory and Ex-memory -Let's give more details about memory and ex-memory via a simply example: +Let's give more details about memory and ex-memory using a simple example: $$ h_t = U h_{t-1} + W x_t $$, -where $h_t$ and $h_{t-1}$ are the memory and ex-memory of step $t$'s respectively. +where $h_t$ and $h_{t-1}$ are the memory and ex-memory (previous memory) of step $t$ respectively. -In the implementation, we can make an ex-memory variable either "refers to" the memory variable of the previous step, -or copy the value of the previous memory value to the current ex-memory variable. +In the implementation, we can make an ex-memory variable either "refer to" the memory variable of the previous step, +or copy the memory value of the previous step to the current ex-memory variable. ### Usage in Python For more information on Block, please refer to the [design doc](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/block.md). -We can define an RNN's step-net using Block: +We can define an RNN's step-net using a Block: ```python import paddle as pd -X = some_op() # x is some operator's output, and is a LoDTensor +X = some_op() # x is some operator's output and is a LoDTensor a = some_op() # declare parameters @@ -68,7 +68,7 @@ with rnn.stepnet(): x = rnn.add_input(X) # declare a memory (rnn's step) h = rnn.add_memory(init=a) - # h.pre_state() means previous memory of rnn + # h.pre_state(), the previous memory of rnn new_state = pd.add_two( pd.matmul(W, x) + pd.matmul(U, h.pre_state())) # update current memory h.update(new_state) @@ -80,19 +80,19 @@ out = rnn() Python API functions in above example: -- `rnn.add_input` indicates the parameter is a variable that will be segmented into step-inputs. -- `rnn.add_memory` creates a variable used as the memory. -- `rnn.add_outputs` mark the variables that will be concatenated across steps into the RNN output. +- `rnn.add_input`: indicates that the parameter is a variable that will be segmented into step-inputs. +- `rnn.add_memory`: creates a variable used as the memory. +- `rnn.add_outputs`: marks the variables that will be concatenated across steps into the RNN output. ### Nested RNN and LoDTensor An RNN whose step-net includes other RNN operators is known as an *nested RNN*. -For example, we could have a 2-level RNN, where the top level corresponds to paragraphs, and the lower level corresponds to sentences. +For example, we could have a 2-level RNN, where the top level corresponds to paragraphs, and the lower level corresponds to sentences. Each step of the higher level RNN also receives an input from the corresponding step of the lower level, and additionally the output from the previous time step at the same level. -The following figure illustrates the feeding of text into the lower level, one sentence each step, and the feeding of step outputs to the top level. The final top level output is about the whole text. +The following figure illustrates feeding in text into the lower level, one sentence at a step, and the feeding in step outputs to the top level. The final top level output is about the whole text. -

+

@@ -110,7 +110,7 @@ a = some_op() # chapter_data is a set of 128-dim word vectors # the first level of LoD is sentence -# the second level of LoD is chapter +# the second level of LoD is a chapter chapter_data = pd.Variable(shape=[None, 128], type=pd.lod_tensor, level=2) def lower_level_rnn(paragraph): @@ -138,14 +138,14 @@ with top_level_rnn.stepnet(): pd.matmul(W0, paragraph_data) + pd.matmul(U0, h.pre_state())) top_level_rnn.add_outputs(h) -# just output the last step +# output the last step chapter_out = top_level_rnn(output_all_steps=False) ``` -in above example, the construction of the `top_level_rnn` calls `lower_level_rnn`. The input is a LoD Tensor. The top level RNN segments input text data into paragraphs, and the lower level RNN segments each paragraph into sentences. +In the above example, the construction of the `top_level_rnn` calls `lower_level_rnn`. The input is an LoD Tensor. The top level RNN segments input text data into paragraphs, and the lower level RNN segments each paragraph into sentences. -By default, the `RNNOp` will concatenate the outputs from all the time steps, -if the `output_all_steps` set to False, it will only output the final time step. +By default, the `RNNOp` will concatenate the outputs from all the time steps. +If the `output_all_steps` is set to False, it will only output the final time step.

diff --git a/doc/design/ops/sequence_decoder.md b/doc/design/ops/sequence_decoder.md index 9007aae7a8355ed06c6720a921351f81b859c1fe..9db5fb8e9a9f89b004bf71ddc064cd976c0d0bee 100644 --- a/doc/design/ops/sequence_decoder.md +++ b/doc/design/ops/sequence_decoder.md @@ -1,35 +1,28 @@ # Design: Sequence Decoder Generating LoDTensors -In tasks such as machine translation and image to text, -a [sequence decoder](https://github.com/PaddlePaddle/book/blob/develop/08.machine_translation/README.md) is necessary to generate sequences. +In tasks such as machine translation and visual captioning, +a [sequence decoder](https://github.com/PaddlePaddle/book/blob/develop/08.machine_translation/README.md) is necessary to generate sequences, one word at a time. This documentation describes how to implement the sequence decoder as an operator. ## Beam Search based Decoder -The [beam search algorithm](https://en.wikipedia.org/wiki/Beam_search) is necessary when generating sequences, -it is a heuristic search algorithm that explores the paths by expanding the most promising node in a limited set. +The [beam search algorithm](https://en.wikipedia.org/wiki/Beam_search) is necessary when generating sequences. It is a heuristic search algorithm that explores the paths by expanding the most promising node in a limited set. -In the old version of PaddlePaddle, a C++ class `RecurrentGradientMachine` implements the general sequence decoder based on beam search, -due to the complexity, the implementation relays on a lot of special data structures, -quite trivial and hard to be customized by users. +In the old version of PaddlePaddle, the C++ class `RecurrentGradientMachine` implements the general sequence decoder based on beam search, due to the complexity involved, the implementation relies on a lot of special data structures that are quite trivial and hard to be customized by users. -There are a lot of heuristic tricks in the sequence generation tasks, -so the flexibility of sequence decoder is very important to users. +There are a lot of heuristic tricks in the sequence generation tasks, so the flexibility of sequence decoder is very important to users. -During PaddlePaddle's refactoring work, -some new concept is proposed such as [LoDTensor](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/lod_tensor.md) and [TensorArray](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/tensor_array.md) that can better support sequence usage, -and they can help to make the implementation of beam search based sequence decoder **more transparent and modular** . +During the refactoring of PaddlePaddle, some new concepts are proposed such as: [LoDTensor](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/lod_tensor.md) and [TensorArray](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/tensor_array.md) that can better support the sequence usage, and they can also help make the implementation of beam search based sequence decoder **more transparent and modular** . -For example, the RNN sates, candidates IDs and probabilities of beam search can be represented as `LoDTensors`; +For example, the RNN states, candidates IDs and probabilities of beam search can be represented all as `LoDTensors`; the selected candidate's IDs in each time step can be stored in a `TensorArray`, and `Packed` to the sentences translated. ## Changing LoD's absolute offset to relative offsets -The current `LoDTensor` is designed to store levels of variable-length sequences, -it stores several arrays of integers each represents a level. +The current `LoDTensor` is designed to store levels of variable-length sequences. It stores several arrays of integers where each represents a level. -The integers in each level represents the begin and end (not inclusive) offset of a sequence **in the underlying tensor**, -let's call this format the **absolute-offset LoD** for clear. +The integers in each level represent the begin and end (not inclusive) offset of a sequence **in the underlying tensor**, +let's call this format the **absolute-offset LoD** for clarity. -The relative-offset LoD can fast retrieve any sequence but fails to represent empty sequences, for example, a two-level LoD is as follows +The relative-offset LoD can retrieve any sequence very quickly but fails to represent empty sequences, for example, a two-level LoD is as follows ```python [[0, 3, 9] [0, 2, 3, 3, 3, 9]] @@ -41,10 +34,9 @@ The first level tells that there are two sequences: while on the second level, there are several empty sequences that both begin and end at `3`. It is impossible to tell how many empty second-level sequences exist in the first-level sequences. -There are many scenarios that relay on empty sequence representation, -such as machine translation or image to text, one instance has no translations or the empty candidate set for a prefix. +There are many scenarios that rely on empty sequence representation, for example in machine translation or visual captioning, one instance has no translation or the empty candidate set for a prefix. -So let's introduce another format of LoD, +So let's introduce another format of LoD, it stores **the offsets of the lower level sequences** and is called **relative-offset** LoD. For example, to represent the same sequences of the above data @@ -54,19 +46,18 @@ For example, to represent the same sequences of the above data [0, 2, 3, 3, 3, 9]] ``` -the first level represents that there are two sequences, +the first level represents that there are two sequences, their offsets in the second-level LoD is `[0, 3)` and `[3, 5)`. The second level is the same with the relative offset example because the lower level is a tensor. It is easy to find out the second sequence in the first-level LoD has two empty sequences. -The following demos are based on relative-offset LoD. +The following examples are based on relative-offset LoD. ## Usage in a simple machine translation model -Let's start from a simple machine translation model that is simplified from [machine translation chapter](https://github.com/PaddlePaddle/book/tree/develop/08.machine_translation) to draw a simple blueprint of what a sequence decoder can do and how to use it. +Let's start from a simple machine translation model that is simplified from the [machine translation chapter](https://github.com/PaddlePaddle/book/tree/develop/08.machine_translation) to draw a blueprint of what a sequence decoder can do and how to use it. -The model has an encoder that learns the semantic vector from a sequence, -and a decoder which uses the sequence decoder to generate new sentences. +The model has an encoder that learns the semantic vector from a sequence, and a decoder which uses the sequence encoder to generate new sentences. **Encoder** ```python @@ -117,7 +108,7 @@ def generate(): # which means there are 2 sentences to translate # - the first sentence has 1 translation prefixes, the offsets are [0, 1) # - the second sentence has 2 translation prefixes, the offsets are [1, 3) and [3, 6) - # the target_word.lod is + # the target_word.lod is # [[0, 1, 6] # [0, 2, 4, 7, 9 12]] # which means 2 sentences to translate, each has 1 and 5 prefixes @@ -154,37 +145,36 @@ def generate(): translation_ids, translation_scores = decoder() ``` -The `decoder.beam_search` is a operator that given the candidates and the scores of translations including the candidates, -return the result of the beam search algorithm. +The `decoder.beam_search` is an operator that, given the candidates and the scores of translations including the candidates, +returns the result of the beam search algorithm. -In this way, users can customize anything on the inputs or outputs of beam search, for example, two ways to prune some translation prefixes +In this way, users can customize anything on the input or output of beam search, for example: -1. meke the correspondind elements in `topk_generated_scores` zero or some small values, beam_search will discard this candidate. -2. remove some specific candidate in `selected_ids` -3. get the final `translation_ids`, remove the translation sequence in it. +1. Make the corresponding elements in `topk_generated_scores` zero or some small values, beam_search will discard this candidate. +2. Remove some specific candidate in `selected_ids`. +3. Get the final `translation_ids`, remove the translation sequence in it. -The implementation of sequence decoder can reuse the C++ class [RNNAlgorithm](https://github.com/Superjom/Paddle/blob/68cac3c0f8451fe62a4cdf156747d6dc0ee000b3/paddle/operators/dynamic_recurrent_op.h#L30), -so the python syntax is quite similar to a [RNN](https://github.com/Superjom/Paddle/blob/68cac3c0f8451fe62a4cdf156747d6dc0ee000b3/doc/design/block.md#blocks-with-for-and-rnnop). +The implementation of sequence decoder can reuse the C++ class: [RNNAlgorithm](https://github.com/Superjom/Paddle/blob/68cac3c0f8451fe62a4cdf156747d6dc0ee000b3/paddle/operators/dynamic_recurrent_op.h#L30), +so the python syntax is quite similar to that of an [RNN](https://github.com/Superjom/Paddle/blob/68cac3c0f8451fe62a4cdf156747d6dc0ee000b3/doc/design/block.md#blocks-with-for-and-rnnop). -Both of them are two-level `LoDTensors` +Both of them are two-level `LoDTensors`: -- the first level represents `batch_size` of (source) sentences; -- the second level represents the candidate ID sets for translation prefix. +- The first level represents `batch_size` of (source) sentences. +- The second level represents the candidate ID sets for translation prefix. -for example, 3 source sentences to translate, and has 2, 3, 1 candidates. +For example, 3 source sentences to translate, and has 2, 3, 1 candidates. -Unlike an RNN, in sequence decoder, the previous state and the current state have different LoD and shape, -a `lod_expand` operator is used to expand the LoD of the previous state to fit the current state. +Unlike an RNN, in sequence decoder, the previous state and the current state have different LoD and shape, and an `lod_expand` operator is used to expand the LoD of the previous state to fit the current state. -For example, the previous state +For example, the previous state: * LoD is `[0, 1, 3][0, 2, 5, 6]` * content of tensor is `a1 a2 b1 b2 b3 c1` -the current state stored in `encoder_ctx_expanded` +the current state is stored in `encoder_ctx_expanded`: * LoD is `[0, 2, 7][0 3 5 8 9 11 11]` -* the content is +* the content is - a1 a1 a1 (a1 has 3 candidates, so the state should be copied 3 times for each candidates) - a2 a2 - b1 b1 b1 @@ -192,54 +182,48 @@ the current state stored in `encoder_ctx_expanded` - b3 b3 - None (c1 has 0 candidates, so c1 is dropped) -Benefit from the relative offset LoD, empty candidate set can be represented naturally. +The benefit from the relative offset LoD is that the empty candidate set can be represented naturally. -the status in each time step can be stored in `TensorArray`, and `Pack`ed to a final LoDTensor, the corresponding syntax is +The status in each time step can be stored in `TensorArray`, and `Pack`ed to a final LoDTensor. The corresponding syntax is: ```python decoder.output(selected_ids) decoder.output(selected_generation_scores) ``` -the `selected_ids` is the candidate ids for the prefixes, -it will be `Packed` by `TensorArray` to a two-level `LoDTensor`, -the first level represents the source sequences, -the second level represents generated sequences. +The `selected_ids` are the candidate ids for the prefixes, and will be `Packed` by `TensorArray` to a two-level `LoDTensor`, where the first level represents the source sequences and the second level represents generated sequences. -Pack the `selected_scores` will get a `LoDTensor` that stores scores of each candidate of translations. +Packing the `selected_scores` will get a `LoDTensor` that stores scores of each translation candidate. -Pack the `selected_generation_scores` will get a `LoDTensor`, and each tail is the probability of the translation. +Packing the `selected_generation_scores` will get a `LoDTensor`, and each tail is the probability of the translation. ## LoD and shape changes during decoding

-According the image above, the only phrase to change LoD is beam search. +According to the image above, the only phase that changes the LoD is beam search. ## Beam search design -The beam search algorthm will be implemented as one method of the sequence decoder, it has 3 inputs +The beam search algorithm will be implemented as one method of the sequence decoder and has 3 inputs: -1. `topk_ids`, top K candidate ids for each prefix. +1. `topk_ids`, the top K candidate ids for each prefix. 2. `topk_scores`, the corresponding scores for `topk_ids` 3. `generated_scores`, the score of the prefixes. -All of the are LoDTensors, so that the sequence affilication is clear. -Beam search will keep a beam for each prefix and select a smaller candidate set for each prefix. +All of these are LoDTensors, so that the sequence affiliation is clear. Beam search will keep a beam for each prefix and select a smaller candidate set for each prefix. -It will return three variables +It will return three variables: 1. `selected_ids`, the final candidate beam search function selected for the next step. 2. `selected_scores`, the scores for the candidates. -3. `generated_scores`, the updated scores for each prefixes (with the new candidates appended). +3. `generated_scores`, the updated scores for each prefix (with the new candidates appended). ## Introducing the LoD-based `Pack` and `Unpack` methods in `TensorArray` -The `selected_ids`, `selected_scores` and `generated_scores` are LoDTensors, -and they exist in each time step, +The `selected_ids`, `selected_scores` and `generated_scores` are LoDTensors that exist at each time step, so it is natural to store them in arrays. -Currently, PaddlePaddle has a module called `TensorArray` which can store an array of tensors, -the results of beam search are better to store in a `TensorArray`. +Currently, PaddlePaddle has a module called `TensorArray` which can store an array of tensors. It is better to store the results of beam search in a `TensorArray`. -The `Pack` and `UnPack` in `TensorArray` are used to package tensors in the array to a `LoDTensor` or split the `LoDTensor` to an array of tensors. -It needs some extensions to support pack or unpack an array of `LoDTensors`. +The `Pack` and `UnPack` in `TensorArray` are used to pack tensors in the array to an `LoDTensor` or split the `LoDTensor` to an array of tensors. +It needs some extensions to support the packing or unpacking an array of `LoDTensors`. diff --git a/doc/mobile/cross_compiling_for_android_cn.md b/doc/mobile/cross_compiling_for_android_cn.md index 882066f23714f7ab3bba9199b5fa5ff2325ce849..424d7718c64438496cf0895397babd5408e1ca02 100644 --- a/doc/mobile/cross_compiling_for_android_cn.md +++ b/doc/mobile/cross_compiling_for_android_cn.md @@ -1,4 +1,4 @@ -# 构建Android平台上的PaddlePaddle库 +# Android平台编译指南 用户可通过如下两种方式，交叉编译Android平台上适用的PaddlePaddle库： - 基于Docker容器的编译方式 diff --git a/doc/mobile/cross_compiling_for_ios_cn.md b/doc/mobile/cross_compiling_for_ios_cn.md index cda636a67de712e072f4cc7ad859dda75211eaa8..9da48e7f2119ce901fbb3abab73400df27be16d2 100644 --- a/doc/mobile/cross_compiling_for_ios_cn.md +++ b/doc/mobile/cross_compiling_for_ios_cn.md @@ -1,4 +1,4 @@ -# 构建iOS平台上的PaddlePaddle库 +# iOS平台编译指南 交叉编译iOS平台上适用的PaddlePaddle库，需要在MacOS系统上进行。本文的将介绍在MacOS上，从源码交叉编译iOS平台上适用的PaddlePaddle库。 ## 准备交叉编译环境 @@ -25,7 +25,7 @@ iOS平台可选配置参数： - `IOS_PLATFORM`，可设置为`OS/SIMULATOR`，默认值为`OS`。 - `OS`，构建目标为`arm`架构的iPhone或者iPad等物理设备。 - `SIMULATOR`，构建目标为`x86`架构的模拟器平台。 -- `IOS_ARCH`，目标架构。针对不同的`IOS_PLATFORM`，可设置的目标架构如下表所示： +- `IOS_ARCH`，目标架构。针对不同的`IOS_PLATFORM`，可设置的目标架构如下表所示，默认编译所有架构： @@ -41,11 +41,11 @@ iOS平台可选配置参数： - + - +
OSarmv7, armv7s, arm64 (默认)armv7, armv7s, arm64
SIMULATORi386, x86_64 (默认)i386, x86_64
@@ -66,7 +66,7 @@ iOS平台可选配置参数： ```bash cmake -DCMAKE_SYSTEM_NAME=iOS \ -DIOS_PLATFORM=OS \ - -DIOS_ARCH="arm64" \ + -DIOS_ARCH="armv7;arm64" \ -DIOS_ENABLE_BITCODE=ON \ -DIOS_USE_VECLIB_FOR_BLAS=ON \ -DCMAKE_INSTALL_PREFIX=your/path/to/install \ @@ -112,6 +112,6 @@ $ make install - `lib`目录，其中包含PaddlePaddle的C-API静态库 - `third_party`目录，其中包含所依赖的所有第三方库 -注意，不同架构的PaddlePaddle库建议安装到不同的目录下，然后使用`lipo`工具将多个静态库合并成一个支持多个架构的fat库。 +注意，如果PaddlePaddle库需要同时支持真机和模拟器，则需要分别编译真机和模拟器版本，然后使用`lipo`工具合并fat库。 自此，PaddlePaddle库已经安装完成，用户可将合成的fat库用于深度学习相关的iOS App中，调用方法见C-API文档。 diff --git a/doc/mobile/cross_compiling_for_raspberry_cn.md b/doc/mobile/cross_compiling_for_raspberry_cn.md index 6e983645faaed1f67edaeeb82ddbef9cef6bb85f..f8ef9dc8031613831437745995268f3abc392f5b 100644 --- a/doc/mobile/cross_compiling_for_raspberry_cn.md +++ b/doc/mobile/cross_compiling_for_raspberry_cn.md @@ -1,4 +1,4 @@ -# 构建Raspberry Pi平台上的PaddlePaddle库 +# Raspberry Pi平台编译指南 通常有两个方法来构建基于 Rasspberry Pi 的版本： diff --git a/paddle/capi/Main.cpp b/paddle/capi/Main.cpp index 78c43949dfe325d0e1a6ba10ae51cb7b858f6c52..bb8249a5511c089ec2f2263ff4cc290f0a5a8fce 100644 --- a/paddle/capi/Main.cpp +++ b/paddle/capi/Main.cpp @@ -29,6 +29,9 @@ static void initPaddle(int argc, char** argv) { extern "C" { paddle_error paddle_init(int argc, char** argv) { + static bool isInit = false; + if (isInit) return kPD_NO_ERROR; + std::vector realArgv; realArgv.reserve(argc + 1); realArgv.push_back(strdup("")); @@ -37,6 +40,7 @@ paddle_error paddle_init(int argc, char** argv) { } initPaddle(argc + 1, realArgv.data()); free(realArgv[0]); + isInit = true; return kPD_NO_ERROR; } } diff --git a/paddle/cuda/include/hl_gpu.h b/paddle/cuda/include/hl_gpu.h index ede2670882ee2b93f610a2261a4ecc1784bc2d0c..4ab8de80d1c7be0f8e3eb848955373dd5e21bc18 100644 --- a/paddle/cuda/include/hl_gpu.h +++ b/paddle/cuda/include/hl_gpu.h @@ -25,7 +25,9 @@ limitations under the License. */ #include "hl_matrix.h" #include "hl_sequence.h" #include "hl_sparse.h" +#ifndef PADDLE_MOBILE_INFERENCE #include "hl_warpctc_wrap.h" +#endif #ifdef HPPL_STUB_FUNC #include "stub/hl_aggregate_stub.h" diff --git a/paddle/framework/backward.cc b/paddle/framework/backward.cc index b3b9c45ded95ce2e735b8898d47760956dcacdce..00d9dd238ec5328be28f58f8118daad3a039e08c 100644 --- a/paddle/framework/backward.cc +++ b/paddle/framework/backward.cc @@ -270,6 +270,19 @@ static bool AllGradInSet(const std::vector& names, return false; } } + if (VLOG_IS_ON(10)) { + std::ostringstream sout; + sout << "All input {"; + for (auto& name : names) { + sout << name << ","; + } + sout << "} is in {"; + for (auto& name : set) { + sout << name << ","; + } + sout << "}"; + VLOG(10) << sout.str(); + } return true; } @@ -290,14 +303,12 @@ static void CreateGradVarInBlock( auto ops = block_desc->AllOps(); for (size_t op_index = grad_op_start_index; op_index < ops.size(); ++op_index) { - bool need_infer_shape = false; std::unordered_set new_vars; ForEachVarName(ops[op_index]->Outputs(), [&](const std::string& grad_var_name) { if (block_desc->HasVar(grad_var_name)) { return false; } - need_infer_shape = true; auto var = block_desc->Var(grad_var_name); new_vars.insert(var->Name()); auto it = param_name_map.find(grad_var_name); @@ -311,23 +322,21 @@ static void CreateGradVarInBlock( grad_record.op_idx_ = static_cast(op_index); return false; /* not break */ }); - if (need_infer_shape) { - ops[op_index]->InferVarType(block_desc); - for (auto& arg : ops[op_index]->OutputArgumentNames()) { - if (new_vars.find(arg) == new_vars.end()) { - continue; - } - auto pname = FwdName(arg); - auto* param = block_desc->FindVarRecursive(pname); - auto* grad = block_desc->FindVar(arg); - if (param == nullptr) { - grad->SetDataType(DataType::FP32); - } else { - grad->SetDataType(param->GetDataType()); - } + ops[op_index]->InferVarType(block_desc); + for (auto& arg : ops[op_index]->OutputArgumentNames()) { + if (new_vars.find(arg) == new_vars.end()) { + continue; + } + auto pname = FwdName(arg); + auto* param = block_desc->FindVarRecursive(pname); + auto* grad = block_desc->FindVar(arg); + if (param == nullptr) { + grad->SetDataType(DataType::FP32); + } else { + grad->SetDataType(param->GetDataType()); } - ops[op_index]->InferShape(*block_desc); } + ops[op_index]->InferShape(*block_desc); } } @@ -387,6 +396,7 @@ std::vector> MakeBlockBackward( ProgramDescBind& program_desc, int block_idx, std::unordered_set* no_grad_vars, std::unordered_map* grad_to_var) { + VLOG(5) << "MakeBlockBackward"; BlockDescBind* cur_block = program_desc.MutableBlock(block_idx); std::vector op_descs = cur_block->AllOps(); std::unordered_map> dup_out_ops; @@ -394,9 +404,10 @@ std::vector> MakeBlockBackward( std::vector> backward_descs; for (auto it = op_descs.rbegin(); it != op_descs.rend(); ++it) { + VLOG(5) << "Making backward " << (*it)->Type() << " op"; std::vector> op_grads; - if ((*it)->Type() == "recurrent") { + if ((*it)->Type() == "recurrent" || (*it)->Type() == "while") { int step_block_idx = (*it)->GetBlockAttr("step_block"); BlockDescBind* backward_block = CreateStepBlock( program_desc, no_grad_vars, grad_to_var, step_block_idx); @@ -410,6 +421,15 @@ std::vector> MakeBlockBackward( op_grads = MakeOpGrad(*it, no_grad_vars, grad_to_var); } + if (VLOG_IS_ON(10)) { + std::ostringstream sout; + sout << "Made "; + for (auto& op_grad : op_grads) { + sout << op_grad->Type() << " "; + } + VLOG(10) << sout.str(); + } + for (const auto& desc : op_grads) { for (const std::string& out_name : desc->OutputArgumentNames()) { if (out_name.find("@GRAD") == std::string::npos) { @@ -425,6 +445,8 @@ std::vector> MakeBlockBackward( op_grads.begin(), op_grads.end(), std::back_inserter(backward_descs), [](std::unique_ptr& ptr) { return std::move(ptr); }); } + + VLOG(5) << "Appending Sums"; // Check whether some variables are written more than once std::list>> pending_sum_ops; for (const auto& dup : dup_out_ops) { @@ -432,16 +454,22 @@ std::vector> MakeBlockBackward( const std::vector dup_op = dup.second; if (out_name != kEmptyVarName && dup_op.size() > 1) { std::vector sum_op_inputs; + std::string next_g_name = out_name; for (size_t i = 0; i < dup_op.size(); ++i) { + VLOG(10) << backward_descs[dup_op[i]]->Type() << " has " << out_name + << " duplicated"; std::string new_name = out_name + "@RENAME@" + std::to_string(i); - backward_descs[dup_op[i]]->Rename(out_name, new_name); + backward_descs[dup_op[i]]->RenameOutput(out_name, new_name); + backward_descs[dup_op[i]]->RenameInput(out_name, next_g_name); sum_op_inputs.emplace_back(new_name); + next_g_name = sum_op_inputs.back(); } 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) { @@ -452,6 +480,8 @@ std::vector> MakeBlockBackward( std::move(p.second)); } + VLOG(5) << "MakeBlockBackward Finished"; + return backward_descs; } diff --git a/paddle/framework/data_type.h b/paddle/framework/data_type.h index 3ec88d7a72c3339bf5e7d0ca3957a3f608f039b7..be144d8fc0104fccc08006532a85906ade25c2a1 100644 --- a/paddle/framework/data_type.h +++ b/paddle/framework/data_type.h @@ -29,6 +29,8 @@ inline DataType ToDataType(std::type_index type) { return DataType::INT32; } else if (typeid(int64_t).hash_code() == type.hash_code()) { return DataType::INT64; + } else if (typeid(bool).hash_code() == type.hash_code()) { + return DataType::BOOL; } else { PADDLE_THROW("Not supported"); } diff --git a/paddle/framework/ddim.cc b/paddle/framework/ddim.cc index 53b899a23997b71e723a298ec360a4e018d89878..8b6f42b82df14bfcd25f33ef16b5903fb965a8ba 100644 --- a/paddle/framework/ddim.cc +++ b/paddle/framework/ddim.cc @@ -60,8 +60,7 @@ void make_ddim(DDim& ddim, const int64_t* dims, int n) { ddim = make_dim<9>(dims); break; default: - throw std::invalid_argument( - "Dynamic dimensions must have between [1, 9] dimensions."); + PADDLE_THROW("Dynamic dimensions must have between [1, 9] dimensions."); } } diff --git a/paddle/framework/executor.cc b/paddle/framework/executor.cc index 2fcf41d69f0011b0d9a3d89c97fcebacb0703e97..adedd8cb0e8504fd6fc924e62a2ede3c1c7ce698 100644 --- a/paddle/framework/executor.cc +++ b/paddle/framework/executor.cc @@ -120,6 +120,7 @@ void Executor::Run(const ProgramDescBind& pdesc, Scope* scope, int block_id, for (auto& op_desc : block.AllOps()) { auto op = paddle::framework::OpRegistry::CreateOp(*op_desc); + VLOG(10) << op->DebugString(); op->Run(*local_scope, *device); } if (create_local_scope) { diff --git a/paddle/framework/op_desc.cc b/paddle/framework/op_desc.cc index 39c8def82e1ebb10a0e357a648af760099020c32..48cd131550dea5ad3f368b25c31d753efbe0dff9 100644 --- a/paddle/framework/op_desc.cc +++ b/paddle/framework/op_desc.cc @@ -235,6 +235,23 @@ void OpDescBind::Rename(const std::string &old_name, need_update_ = true; } +void OpDescBind::RenameOutput(const std::string &old_name, + const std::string &new_name) { + for (auto &output : outputs_) { + std::replace(output.second.begin(), output.second.end(), old_name, + new_name); + } + need_update_ = true; +} + +void OpDescBind::RenameInput(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); + } + need_update_ = true; +} + struct SetAttrDescVisitor : public boost::static_visitor { explicit SetAttrDescVisitor(OpDesc::Attr *attr) : attr_(attr) {} mutable OpDesc::Attr *attr_; @@ -448,7 +465,12 @@ const std::vector &CompileTimeInferShapeContext::Outputs( DDim CompileTimeInferShapeContext::GetDim(const std::string &name) const { auto var = block_.FindVarRecursive(name); PADDLE_ENFORCE(var != nullptr, "Cannot find variable %s", name); - return framework::make_ddim(var->Shape()); + try { + return framework::make_ddim(var->Shape()); + } catch (...) { + VLOG(5) << "GetDim of variable " << name << " error"; + std::rethrow_exception(std::current_exception()); + } } void CompileTimeInferShapeContext::SetDim(const std::string &name, diff --git a/paddle/framework/op_desc.h b/paddle/framework/op_desc.h index e3e96441bbf51729f2ba69c9257e6961b1de0d5c..da032319afa775571d3942bf6ae415db7d233735 100644 --- a/paddle/framework/op_desc.h +++ b/paddle/framework/op_desc.h @@ -73,6 +73,10 @@ class OpDescBind { void Rename(const std::string &old_name, const std::string &new_name); + void RenameOutput(const std::string &old_name, const std::string &new_name); + + void RenameInput(const std::string &old_name, const std::string &new_name); + // Only be used in C++ const AttributeMap &GetAttrMap() const; diff --git a/paddle/framework/operator.cc b/paddle/framework/operator.cc index 3276f8af396fe58450a8dc6713fe61e49d5ca708..93467ab8ac796277b47a861a427de2837fb2d3d4 100644 --- a/paddle/framework/operator.cc +++ b/paddle/framework/operator.cc @@ -403,19 +403,6 @@ class RuntimeInferShapeContext : public InferShapeContext { void OperatorWithKernel::Run(const Scope& scope, const platform::DeviceContext& dev_ctx) const { - if (VLOG_IS_ON(1)) { - auto inputs = this->InputVars(); - auto outputs = this->OutputVars(true); - std::ostringstream sout; - sout << "Run operator " << this->Type() << " From ["; - std::ostream_iterator out_it(sout, ","); - std::copy(inputs.begin(), inputs.end(), out_it); - sout << "] to ["; - std::copy(outputs.begin(), outputs.end(), out_it); - sout << "]"; - VLOG(1) << sout.str(); - } - RuntimeInferShapeContext infer_shape_ctx(*this, scope); this->InferShape(&infer_shape_ctx); diff --git a/paddle/framework/scope.cc b/paddle/framework/scope.cc index 9428b8a07ea0af005f6e960ddaa02da624ad9d97..9ad6272c99dd6a85520ae44c1331ac232bc6a9a2 100644 --- a/paddle/framework/scope.cc +++ b/paddle/framework/scope.cc @@ -38,11 +38,12 @@ Scope& Scope::NewScope() const { Variable* Scope::Var(const std::string& name) { auto iter = vars_.find(name); if (iter != vars_.end()) { + VLOG(3) << "Get existing variable " << name; return iter->second; } Variable* v = new Variable(); vars_[name] = v; - VLOG(3) << "Create variable " << name << " on scope"; + VLOG(3) << "Create variable " << name; v->name_ = &(vars_.find(name)->first); return v; } diff --git a/paddle/framework/shape_inference.h b/paddle/framework/shape_inference.h index 7d36ead2ca85328c7843b3b5d423cf8e921d1c93..05dc47f06ac81f0acb6d0317cbecb3009c7dd7f0 100644 --- a/paddle/framework/shape_inference.h +++ b/paddle/framework/shape_inference.h @@ -53,6 +53,10 @@ class InferShapeContext { virtual bool IsRuntime() const = 0; + // Note: In while op, we need this to be public + void SetDims(const std::vector &names, + const std::vector &dims); + protected: virtual framework::DDim GetDim(const std::string &name) const = 0; virtual void SetDim(const std::string &name, const framework::DDim &dim) = 0; @@ -60,9 +64,6 @@ class InferShapeContext { std::vector GetDims( const std::vector &names) const; - void SetDims(const std::vector &names, - const std::vector &dims); - std::vector GetVarTypes( const std::vector &names) const; diff --git a/paddle/math/Storage.cpp b/paddle/math/Storage.cpp index 4adaaef9838f0d178468af3af142031325bfc11d..a2ef731ecbcd18ca4bd0b2381de04650a2686c2d 100644 --- a/paddle/math/Storage.cpp +++ b/paddle/math/Storage.cpp @@ -17,9 +17,13 @@ limitations under the License. */ #include "paddle/utils/StringUtil.h" #include "paddle/utils/Util.h" +#ifndef PADDLE_MOBILE_INFERENCE DEFINE_int32(pool_limit_size, 536870912, "maximum memory size managed by a memory pool, default is 512M"); +#else +DEFINE_int32(pool_limit_size, 0, "default is 0"); +#endif namespace paddle { diff --git a/paddle/memory/README.md b/paddle/memory/README.md index 7f95e80f980b0c0b93ecb418e6b923045313eaa5..6cb003c50bc7d142d65b0591e7e5235431d2ea42 100644 --- a/paddle/memory/README.md +++ b/paddle/memory/README.md @@ -1,4 +1,141 @@ # Region-based Heterogeneous Memory Management +## Design -Please check out the [design documentation](http://gangliao.me) to find out more details about -buddy memory allocator for both CPU and GPU. +### Usage + +To allocate 4KB CPU memory: + +```cpp +p = memory::Alloc(platform::CPUPlace(), 4*1024); +``` + +To allocate 4KB memory on the 3rd GPU: + +```cpp +p = memory::Alloc(platform::GPUPlace(2), 4*1024); +``` + +To free memory and check the so-far used amount of memory on a place: + +```cpp +auto pl = platform::GPUPlace(0); +p = memory::Alloc(pl, 4*1024); +cout << memory::Used(pl); +memory::Free(pl, p); +``` + +### API + +In `paddle/memory/memory.h` we have: + +```cpp +namespace memory { +template void* Alloc(Place, size_t); +template void Free(Place, void*); +template size_t Used(Place); +} // namespace memory +``` + +These function templates have specializations on either `platform::CPUPlace` or `platform::GPUPlace`: + +```cpp +template<> +void* Alloc(CPUPlace p, size_t size) { + return GetCPUBuddyAllocator()->Alloc(size); +} +``` + +and + +```cpp +template<> +void Alloc(GPUPlace p, size_t size) { + return GetGPUBuddyAllocator(p.id)->Alloc(size); +} +``` + +Similar specializations exist for `Free` and `Used`. + +### Implementation + +`GetCPUBuddyAllocator` and `GetGPUBuddyAllocator` are singletions. + +```cpp +BuddyAllocator* GetCPUBuddyAllocator() { + static BuddyAllocator* a = NULL; + if (a == NULL) { + a = new BuddyAllocator(new CPUAllocator /*backup allocator*/, ...); + } + return a; +} + +BuddyAllocator* GetGPUBuddyAllocator(int gpu_id) { + static BuddyAllocator* as = NULL; + if (as == NULL) { + as = new BuddyAllocator*[platform::NumGPUs()]; + for (int gpu = 0; gpu < platform::NumGPUs(); gpu++) { + as[gpu] = new BuddyAllocator(new GPUAllocator(gpu) /* backup allocator */, ...); + } + } + return as[gpu_id); +``` + +#### `BuddyAllocator` + +`BuddyAllocator` implements the buddy allocation algorithm. Its constructor takes parameters only related with the algorithm: + +```cpp +BuddyAllocator::BuddyAllocator(initial_pool_size, max_pool_size) { + ... +} +``` + +Please be aware that **`BuddyAllocator` always allocate aligned memory**, aligned on 32-bytes, which can hold a `BuddyAllocator::Block` object: + +```cpp +class BuddyAllocator { + private: + struct Block { + size_t size; + Block* left, right; + size_t index; // allocator id + }; + ... +}; +``` + +Because BuddyAllocator has the meta-data of each block, it can trace the used memory -- record the amount returned by `Alloc` freed in `Free`. Instead, `CPUAllocator` and `GPUAllocator` doesn't know the size of freed memory block and cannot do the trace. + +#### System Allocators + +The `GPUAllocator` and `CPUAllocator` are calls *system allocators*. They work as the fallback allocators of `BuddyAllocator`. + +## Justification + +I got inspiration from Majel and Caffe2, though above design look different from both. + +### Caffe2 + +In Caffe2, `Tensor::mutable_data()` allocates the memroy. In particular, [`Tensor::mutable_data`](https://github.com/caffe2/caffe2/blob/v0.7.0/caffe2/core/tensor.h#L523) calls [`Tensor::raw_mutable_data`](https://github.com/caffe2/caffe2/blob/v0.7.0/caffe2/core/tensor.h#L459), which in turn calls [`Context::New`](https://github.com/caffe2/caffe2/blob/v0.7.0/caffe2/core/tensor.h#L479). + +There are two implementations of `Context`: + +1. [`CPUContext`](https://github.com/caffe2/caffe2/blob/v0.7.0/caffe2/core/context.h#L105), whose [`New` method](https://github.com/caffe2/caffe2/blob/v0.7.0/caffe2/core/context.h#L131) calls [`g_cpu_allocator.get()->New(size_t)`](https://github.com/caffe2/caffe2/blob/v0.7.0/caffe2/core/context.cc#L15) to allocate the memory. + +1. [`CUDAContext`](https://github.com/caffe2/caffe2/blob/v0.7.0/caffe2/core/context_gpu.h#L99), which has a data member [`int gpu_id_`](https://github.com/caffe2/caffe2/blob/v0.7.0/caffe2/core/context_gpu.h#L202). This looks very similar to class `majel::GPUPlace`, who also has an `int id_` data member. `CUDAContext::New(size_t)` calls [`g_cub_allocator->DeviceAllocate(&ptr, nbytes)`](https://github.com/caffe2/caffe2/blob/v0.7.0/caffe2/core/context_gpu.cu#L355) to allocate the memory. + +### Majel + +In Majel, there are basically two allocator types: + +1. `cpu::SystemAllocator`, which has similar functionality to `caffe2::CPUContext::New/Delete`. +1. `gpu::SystemAllocator`, which has similar functionality to `caffe2::CUDAContext::New/Delete`. + +However, memory allocation is not via these two allocators. Instead, these two allocators are defined in hidden namespaces. + +In Majel there are hidden global variables like: + +1. `cpu::SystemAllocator g_cpu_allocator`, and +1. `vector g_gpu_allocators(NUM_GPUS)`. + +Programs allocate memory via a BuddyAllocator, which can take the `g_cpu_allocator` or a `g_gpu_allocators[gpu_id]` as its *fallback allocator*, so that if BuddyAllocator cannot find a block in its memory pool, it extends its memory pool by calling the fallback allocator's `New(size_t)`. diff --git a/paddle/operators/CMakeLists.txt b/paddle/operators/CMakeLists.txt index 709f7de2e43093114d096cbfca5b5d49293a6d3e..a719da2560291dbc7e98aadfae41d4692d8afcad 100644 --- a/paddle/operators/CMakeLists.txt +++ b/paddle/operators/CMakeLists.txt @@ -9,6 +9,7 @@ function(op_library TARGET) set(OP_LIBRARY ${TARGET} ${OP_LIBRARY} PARENT_SCOPE) set(cc_srcs) set(cu_srcs) + set(cu_cc_srcs) set(op_common_deps operator op_registry math_function) set(options "") set(oneValueArgs "") @@ -22,6 +23,9 @@ function(op_library TARGET) if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.cc) list(APPEND cc_srcs ${TARGET}.cc) endif() + if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.cu.cc) + list(APPEND cu_cc_srcs ${TARGET}.cu.cc) + endif() if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.cu) list(APPEND cu_srcs ${TARGET}.cu) endif() @@ -29,6 +33,8 @@ function(op_library TARGET) foreach(src ${op_library_SRCS}) if (${src} MATCHES ".*\\.cu$") list(APPEND cu_srcs ${src}) + elseif(${src} MATCHES ".*\\.cu.cc$") + list(APPEND cu_cc_srcs ${src}) elseif(${src} MATCHES ".*\\.cc$") list(APPEND cc_srcs ${src}) else() @@ -43,7 +49,7 @@ function(op_library TARGET) endif() if (WITH_GPU) - nv_library(${TARGET} SRCS ${cc_srcs} ${cu_srcs} DEPS ${op_library_DEPS} + nv_library(${TARGET} SRCS ${cc_srcs} ${cu_cc_srcs} ${cu_srcs} DEPS ${op_library_DEPS} ${op_common_deps}) else() cc_library(${TARGET} SRCS ${cc_srcs} DEPS ${op_library_DEPS} @@ -140,7 +146,9 @@ function(op_library TARGET) # pybind USE_CPU_ONLY_OP list(LENGTH cu_srcs cu_srcs_len) - if (${pybind_flag} EQUAL 0 AND ${cu_srcs_len} EQUAL 0) + list(LENGTH cu_cc_srcs cu_cc_srcs_len) + + if (${pybind_flag} EQUAL 0 AND ${cu_srcs_len} EQUAL 0 AND ${cu_cc_srcs_len} EQUAL 0) file(APPEND ${pybind_file} "USE_CPU_ONLY_OP(${TARGET});\n") set(pybind_flag 1) endif() @@ -160,11 +168,12 @@ set(DEPS_OPS recurrent_op dynamic_recurrent_op softmax_with_cross_entropy_op + softmax_op + sequence_softmax_op sum_op pool_op pool_with_index_op conv_op - lstm_op conv_transpose_op nccl_op sequence_conv_op @@ -174,13 +183,20 @@ set(DEPS_OPS array_to_lod_tensor_op lstm_op tensor_array_read_write_op - gru_op) + gru_op + adagrad_op + sgd_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(softmax_op DEPS softmax) +op_library(sequence_softmax_op DEPS softmax) +op_library(sum_op DEPS selected_rows_functor) +op_library(sgd_op DEPS selected_rows_functor) +op_library(adagrad_op DEPS selected_rows_functor) op_library(conv_op DEPS vol2col) -op_library(sum_op DEPS net_op selected_rows_functor) op_library(pool_op DEPS pooling) op_library(pool_with_index_op DEPS pooling) op_library(lod_rank_table_op SRCS lod_rank_table_op.cc DEPS lod_rank_table) @@ -220,6 +236,6 @@ cc_test(dynamic_recurrent_op_test SRCS dynamic_recurrent_op_test.cc rnn/recurrent_op_utils.cc DEPS dynamic_recurrent_op) if(WITH_GPU) - nv_test(nccl_op_test SRCS nccl_op_test.cu DEPS nccl_op gpu_info device_context) + cc_test(nccl_op_test SRCS nccl_op_test.cu.cc DEPS nccl_op gpu_info device_context) endif() cc_test(save_load_op_test SRCS save_load_op_test.cc DEPS save_op load_op) diff --git a/paddle/operators/accuracy_op.cu b/paddle/operators/accuracy_op.cu index b575c682f0d30678a72a33040cce6cc799da26cb..d2dcab4e548b99c6beecfaa570ac31804fd07d82 100644 --- a/paddle/operators/accuracy_op.cu +++ b/paddle/operators/accuracy_op.cu @@ -16,6 +16,7 @@ limitations under the License. */ #include #include "paddle/operators/accuracy_op.h" #include "paddle/platform/cuda_helper.h" +#include "paddle/platform/gpu_info.h" namespace paddle { namespace operators { @@ -73,26 +74,28 @@ class AccuracyOpCUDAKernel : public framework::OpKernel { int num_samples = static_cast(inference->dims()[0]); size_t infer_width = inference->dims()[1]; - PADDLE_ENFORCE(cudaMemset(accuracy_data, 0, sizeof(float))); - // cudaMemset((void**)&correct_data, 0, sizeof(float)); + auto stream = ctx.cuda_device_context().stream(); + platform::GpuMemsetAsync(accuracy_data, 0, sizeof(float), stream); if (num_samples == 0) { return; } - cudaMemcpy(total_data, &num_samples, sizeof(int), cudaMemcpyHostToDevice); + platform::GpuMemcpyAsync(total_data, &num_samples, sizeof(int), + cudaMemcpyHostToDevice, stream); - AccuracyCudaKernel<<< - 1, PADDLE_CUDA_NUM_THREADS, 0, ctx.cuda_device_context().stream()>>>( + AccuracyCudaKernel< + PADDLE_CUDA_NUM_THREADS><<<1, PADDLE_CUDA_NUM_THREADS, 0, stream>>>( num_samples, infer_width, indices_data, label_data, correct_data, accuracy_data); int d_num_samples, d_num_correct; float d_accuracy; - cudaMemcpy(&d_num_correct, correct_data, sizeof(int), - cudaMemcpyDeviceToHost); - cudaMemcpy(&d_num_samples, total_data, sizeof(int), cudaMemcpyDeviceToHost); - cudaMemcpy(&d_accuracy, accuracy_data, sizeof(float), - cudaMemcpyDeviceToHost); + platform::GpuMemcpyAsync(&d_num_correct, correct_data, sizeof(int), + cudaMemcpyDeviceToHost, stream); + platform::GpuMemcpyAsync(&d_num_samples, total_data, sizeof(int), + cudaMemcpyDeviceToHost, stream); + platform::GpuMemcpyAsync(&d_accuracy, accuracy_data, sizeof(float), + cudaMemcpyDeviceToHost, stream); } }; diff --git a/paddle/operators/adagrad_op.cc b/paddle/operators/adagrad_op.cc index 8d1a2b7938d2c6607cbeb3cecb72d1d5b83dd8b9..d6686e3ef3165976cf4c077a7a0f213082aa7716 100644 --- a/paddle/operators/adagrad_op.cc +++ b/paddle/operators/adagrad_op.cc @@ -14,6 +14,11 @@ limitations under the License. */ #include "paddle/operators/adagrad_op.h" +#include + +#include "paddle/operators/math/math_function.h" +#include "paddle/operators/math/selected_rows_functor.h" + namespace paddle { namespace operators { @@ -21,7 +26,7 @@ class AdagradOp : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; - void InferShape(framework::InferShapeContext *ctx) const override { + 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"), @@ -54,8 +59,8 @@ class AdagradOp : public framework::OperatorWithKernel { class AdagradOpMaker : public framework::OpProtoAndCheckerMaker { public: - AdagradOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + AdagradOpMaker(framework::OpProto* proto, + framework::OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("Param", "(Tensor) Input parameter"); AddInput("Grad", "(Tensor) Input gradient"); @@ -87,10 +92,85 @@ for numerical stability to avoid the division by zero error. )DOC"); } }; + +namespace { +size_t FindPos(const std::vector& rows, int64_t value) { + return std::find(rows.begin(), rows.end(), value) - rows.begin(); +} +} // namespace + +template +struct SparseAdagradFunctor { + void operator()(const platform::DeviceContext& context, + const framework::SelectedRows& grad, + const framework::Tensor& learning_rate, T epsilon, + framework::Tensor* moment, framework::Tensor* param) { + // 1. g_m.rows = set(g.rows) + auto grad_rows = grad.rows(); + std::set row_set(grad_rows.begin(), grad_rows.end()); + std::vector merge_rows(row_set.begin(), row_set.end()); + + auto grad_width = grad.value().dims()[1]; + std::unique_ptr grad_merge{ + new framework::SelectedRows()}; + grad_merge->set_rows(merge_rows); + grad_merge->set_height(grad.height()); + grad_merge->mutable_value()->mutable_data( + framework::make_ddim( + {static_cast(merge_rows.size()), grad_width}), + context.GetPlace()); + + math::SetConstant constant_functor; + constant_functor(context, grad_merge->mutable_value(), 0.0); + + auto* grad_merge_data = grad_merge->mutable_value()->data(); + auto* grad_data = grad.value().data(); + + for (size_t i = 0; i < grad_rows.size(); i++) { + size_t grad_merge_i = FindPos(merge_rows, grad_rows[i]); + for (int64_t j = 0; j < grad_width; j++) { + grad_merge_data[grad_merge_i * grad_width + j] += + grad_data[i * grad_width + j]; + } + } + + // 2. m += g_m * g_m + std::unique_ptr grad_square{ + new framework::SelectedRows()}; + grad_square->set_rows(grad_merge->rows()); + grad_square->set_height(grad_merge->height()); + grad_square->mutable_value()->mutable_data(grad_merge->value().dims(), + context.GetPlace()); + auto gs = + framework::EigenVector::Flatten(*(grad_square->mutable_value())); + auto gm = framework::EigenVector::Flatten(grad_merge->value()); + gs.device(*context.GetEigenDevice()) = gm * gm; + + math::SelectedRowsAddToTensor functor; + functor(context, *grad_square, moment); + + // 3. update parameter + auto* lr = learning_rate.data(); + auto* param_data = param->data(); + auto* moment_data = moment->data(); + + for (size_t i = 0; i < merge_rows.size(); i++) { + for (int64_t j = 0; j < grad_width; j++) { + param_data[merge_rows[i] * grad_width + j] -= + lr[0] * grad_merge_data[i * grad_width + j] / + (std::sqrt(moment_data[merge_rows[i] * grad_width + j]) + epsilon); + } + } + } +}; + +template struct SparseAdagradFunctor; +template struct SparseAdagradFunctor; } // namespace operators } // namespace paddle namespace ops = paddle::operators; REGISTER_OP_WITHOUT_GRADIENT(adagrad, ops::AdagradOp, ops::AdagradOpMaker); -REGISTER_OP_CPU_KERNEL(adagrad, - ops::AdagradOpKernel); +REGISTER_OP_CPU_KERNEL( + adagrad, ops::AdagradOpKernel, + ops::AdagradOpKernel); diff --git a/paddle/operators/adagrad_op.cu b/paddle/operators/adagrad_op.cu index a5b7951121360f78612f9008a522235104708112..5b869e6bc5f4604ba6055ffd62fa21e4a1f41b93 100644 --- a/paddle/operators/adagrad_op.cu +++ b/paddle/operators/adagrad_op.cu @@ -14,7 +14,138 @@ #define EIGEN_USE_GPU #include "paddle/operators/adagrad_op.h" +#include "paddle/operators/math/selected_rows_functor.h" +#include "paddle/operators/math/math_function.h" +#include "paddle/platform/cuda_helper.h" + +namespace paddle { +namespace operators { + +namespace { + +template +__global__ void MergeGradKernel(const T* grad, const int64_t* grad_rows, + T* grad_merge, const int64_t* grad_merge_rows, + size_t grad_merge_rows_size, + int64_t row_numel) { + const int ty = blockIdx.y; + int tid = threadIdx.x; + __shared__ size_t grad_merge_idx; + + if (tid == 0) { + for (size_t i = 0; i < grad_merge_rows_size; i++) { + if (grad_rows[ty] == grad_merge_rows[i]) { + grad_merge_idx = i; + } + } + } + + __syncthreads(); + + grad += ty * row_numel; + grad_merge += grad_merge_idx * row_numel; + for (int index = tid; index < row_numel; index += block_size) { + paddle::platform::CudaAtomicAdd(grad_merge + index, grad[index]); + } +} + +template +__global__ void SparseAdagradFunctorKernel(const T* grad, const int64_t* rows, + const T* learning_rate, T* param, + T* moment, int64_t row_numel, + T epsilon) { + const int ty = blockIdx.y; + int tid = threadIdx.x; + + grad += ty * row_numel; + param += rows[ty] * row_numel; + moment += rows[ty] * row_numel; + + for (int index = tid; index < row_numel; index += block_size) { + // Since index in rows of SelectedRows can be duplicate, we have to use + // Atomic Operation to avoid concurrent write error. + paddle::platform::CudaAtomicAdd(param + index, + -1.0 * learning_rate[0] * grad[index] / + (sqrt(moment[index]) + epsilon)); + } +} +} // namespace + +template +struct SparseAdagradFunctor { + void operator()(const platform::DeviceContext& context, + const framework::SelectedRows& grad, + const framework::Tensor& learning_rate, T epsilon, + framework::Tensor* moment, framework::Tensor* param) { + // 1. g_m.rows = set(g.rows) + auto grad_rows = grad.rows(); + std::set row_set(grad_rows.begin(), grad_rows.end()); + std::vector merge_rows(row_set.begin(), row_set.end()); + + auto grad_width = grad.value().dims()[1]; + std::unique_ptr grad_merge{ + new framework::SelectedRows()}; + grad_merge->set_rows(merge_rows); + grad_merge->set_height(grad.height()); + grad_merge->mutable_value()->mutable_data( + framework::make_ddim( + {static_cast(merge_rows.size()), grad_width}), + context.GetPlace()); + + math::SetConstant constant_functor; + constant_functor(context, grad_merge->mutable_value(), 0.0); + + auto* grad_merge_data = grad_merge->mutable_value()->data(); + auto* grad_data = grad.value().data(); + + const int block_size = 256; + dim3 threads(block_size, 1); + dim3 grid1(1, grad_rows.size()); + + MergeGradKernel< + T, 256><<(context) + .stream()>>>(grad_data, grad.rows().data(), + grad_merge_data, grad_merge->rows().data(), + grad_merge->rows().size(), grad_width); + + // 2. m += g_m * g_m + std::unique_ptr grad_square{ + new framework::SelectedRows()}; + grad_square->set_rows(grad_merge->rows()); + grad_square->set_height(grad_merge->height()); + grad_square->mutable_value()->mutable_data(grad_merge->value().dims(), + context.GetPlace()); + auto gs = + framework::EigenVector::Flatten(*(grad_square->mutable_value())); + auto gm = framework::EigenVector::Flatten(grad_merge->value()); + gs.device(*context.GetEigenDevice()) = gm * gm; + + math::SelectedRowsAddToTensor functor; + functor(context, *grad_square, moment); + + // 3. update parameter + auto* lr = learning_rate.data(); + auto* param_data = param->data(); + auto* moment_data = moment->data(); + + dim3 grid2(1, merge_rows.size()); + SparseAdagradFunctorKernel< + T, 256><<(context) + .stream()>>>(grad_merge_data, grad_merge->rows().data(), + lr, param_data, + moment_data, grad_width, epsilon); + } +}; + +template struct SparseAdagradFunctor; +template struct SparseAdagradFunctor; + +} // namespace operators +} // namespace paddle namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL(adagrad, - ops::AdagradOpKernel); +REGISTER_OP_GPU_KERNEL( + adagrad, ops::AdagradOpKernel, + ops::AdagradOpKernel); diff --git a/paddle/operators/adagrad_op.h b/paddle/operators/adagrad_op.h index c5d8f751d3527f89b96d4274328ba0bb5f6efa44..4d4a6434c7c472d8ceb01edfc4050fbb009d6c9f 100644 --- a/paddle/operators/adagrad_op.h +++ b/paddle/operators/adagrad_op.h @@ -19,35 +19,59 @@ limitations under the License. */ namespace paddle { namespace operators { +template +struct SparseAdagradFunctor { + void operator()(const platform::DeviceContext& context, + const framework::SelectedRows& grad, + const framework::Tensor& learning_rate, T epsilon, + framework::Tensor* moment, framework::Tensor* param); +}; + 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"); + 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); + T epsilon = static_cast(ctx.Attr("epsilon")); + + auto* grad_var = ctx.InputVar("Grad"); + if (grad_var->IsType()) { + 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); + } else if (grad_var->IsType()) { + auto* param_tensor = ctx.Input("Param"); + PADDLE_ENFORCE_EQ(param_tensor, param_out_tensor); + + auto* moment_tensor = ctx.Input("Moment"); + PADDLE_ENFORCE_EQ(moment_tensor, moment_out_tensor); + + SparseAdagradFunctor functor; + functor(ctx.device_context(), *ctx.Input("Grad"), + *ctx.Input("LearningRate"), epsilon, + moment_out_tensor, param_out_tensor); + } else { + PADDLE_THROW("Unsupported Variable Type of Grad"); + } } }; diff --git a/paddle/operators/array_operator.h b/paddle/operators/array_operator.h index 666043e824f885e9c0e79e319d0a38ba108c209a..233a81198e336d3190565fb18556f96979cec0ce 100644 --- a/paddle/operators/array_operator.h +++ b/paddle/operators/array_operator.h @@ -42,6 +42,7 @@ class ArrayOp : public framework::OperatorBase { } else { offset = static_cast(*i_tensor.data()); } + VLOG(10) << " Offset = " << offset; return offset; } }; diff --git a/paddle/operators/batch_norm_op.cu b/paddle/operators/batch_norm_op.cu.cc similarity index 100% rename from paddle/operators/batch_norm_op.cu rename to paddle/operators/batch_norm_op.cu.cc diff --git a/paddle/operators/beam_search_op.cc b/paddle/operators/beam_search_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..17926a813d5b0b8ace6a1b20066cd0007703c696 --- /dev/null +++ b/paddle/operators/beam_search_op.cc @@ -0,0 +1,185 @@ +/* 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/beam_search_op.h" + +#include +#include "paddle/framework/lod_tensor.h" +#include "paddle/framework/op_registry.h" + +namespace paddle { +namespace operators { + +void BeamSearch::operator()(const framework::LoDTensor &pre_ids, + framework::LoDTensor *selected_ids, + framework::LoDTensor *selected_scores) { + auto items = SelectTopBeamSizeItems(); + auto selected_items = ToMap(items); + PruneEndidCandidates(pre_ids, &selected_items); + // calculate the output tensor's height + size_t num_instances = std::accumulate( + std::begin(items), std::end(items), 0, + [](size_t a, std::vector &b) { return a + b.size(); }); + // the output tensor shape should be [num_instances, 1] + auto dims = framework::make_ddim( + std::vector({static_cast(num_instances), 1})); + selected_ids->Resize(dims); + selected_scores->Resize(dims); + + std::map> hash; + framework::LoD new_lod; + auto *ids_data = selected_ids->mutable_data(platform::CPUPlace()); + auto *scores_data = + selected_scores->mutable_data(platform::CPUPlace()); + + // fill in data + std::vector low_level; + size_t low_offset = 0; + for (auto &items : selected_items) { + low_level.push_back(low_offset); + for (auto &item : items) { + ids_data[low_offset] = item.id; + scores_data[low_offset] = item.score; + low_offset++; + } + } + // fill lod + auto abs_lod = framework::ToAbsOffset(ids_->lod()); + auto &high_level = abs_lod[lod_level_]; + framework::LoD lod(2); + lod[0].assign(high_level.begin(), high_level.end()); + lod[1].assign(low_level.begin(), low_level.end()); + selected_ids->set_lod(lod); + selected_scores->set_lod(lod); +} + +void BeamSearch::PruneEndidCandidates(const framework::LoDTensor &pre_ids, + std::vector> *items) { + auto *pre_ids_data = pre_ids.data(); + + for (size_t offset = 0; offset < items->size(); offset++) { + auto prefix_id = pre_ids_data[offset]; + if (prefix_id == end_id_) { + items->at(offset).clear(); + } + } +} + +std::vector> BeamSearch::ToMap( + const std::vector> &items) { + std::vector> result; + for (auto &entries : items) { + for (const auto &item : entries) { + if (item.offset >= result.size()) { + result.resize(item.offset + 1); + } + result[item.offset].push_back(item); + } + } + return result; +} + +std::vector> +BeamSearch::SelectTopBeamSizeItems() { + std::vector> result; + std::vector items; + // for each source sentence, select the top beam_size items across all + // candidate sets. + while (NextItemSet(&items)) { + std::nth_element(std::begin(items), std::begin(items) + beam_size_, + std::end(items), [](const Item &a, const Item &b) { + // TODO(superjom) make score's comparation customizable. + // partial sort in descending order + return a.score > b.score; + }); + // prune the top beam_size items. + if (items.size() > beam_size_) { + items.resize(beam_size_); + } + result.emplace_back(items); + } + return result; +} + +// the candidates of a source +bool BeamSearch::NextItemSet(std::vector *items) { + if (sent_offset_ >= ids_->NumElements(lod_level_)) { + return false; + } + // find the current candidates + auto ids = *ids_; + auto scores = *scores_; + + auto source_abs_two_level_lod = framework::SliceInLevel( + ids.lod(), lod_level_, sent_offset_, sent_offset_ + 1); + source_abs_two_level_lod = framework::ToAbsOffset(source_abs_two_level_lod); + auto abs_lod = framework::ToAbsOffset(ids.lod()); + PADDLE_ENFORCE_GE(source_abs_two_level_lod.size(), 2UL); + + auto *ids_data = ids.data(); + auto *scores_data = scores.data(); + + size_t instance_dim = 1; + for (int i = 1; i < ids.dims().size(); i++) { + instance_dim *= ids.dims()[i]; + } + + items->clear(); + items->reserve(framework::product(ids.dims())); + for (size_t offset = abs_lod[lod_level_][sent_offset_]; + offset < abs_lod[lod_level_][sent_offset_ + 1]; offset++) { + for (int d = 0; d < instance_dim; d++) { + const size_t dim_offset = offset * instance_dim + d; + items->emplace_back(offset, ids_data[dim_offset], + scores_data[dim_offset]); + } + } + + sent_offset_++; + return true; +} + +class BeamSearchProtoAndCheckerMaker + : public framework::OpProtoAndCheckerMaker { + public: + BeamSearchProtoAndCheckerMaker(framework::OpProto *proto, + framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + // inputs and outputs stored in proto + AddInput("pre_ids", "ids in previous step"); + AddInput("ids", "a LoDTensor of shape of [None,k]"); + AddInput("scores", + "a LoDTensor that has the same shape and LoD with `ids`"); + AddOutput("selected_ids", + "a LoDTensor that stores the IDs selected by beam search"); + AddOutput( + "selected_scores", + "a LoDTensor that has the same shape and LoD with `selected_ids`"); + + // Attributes stored in AttributeMap + AddAttr("level", "the level of LoDTensor"); + AddAttr("beam_size", "beam size for beam search"); + AddAttr("end_id", + "the token id which indicates the end of a sequence"); + + AddComment( + "This is a beam search operator that help to generate sequences."); + } +}; + +} // namespace operators +} // namespace paddle + +REGISTER_OP_WITHOUT_GRADIENT(beam_search, paddle::operators::BeamSearchOp, + paddle::operators::BeamSearchProtoAndCheckerMaker); diff --git a/paddle/operators/beam_search_op.h b/paddle/operators/beam_search_op.h new file mode 100644 index 0000000000000000000000000000000000000000..cc556bfe42ab12d73c0eb503d033efc272b5dd68 --- /dev/null +++ b/paddle/operators/beam_search_op.h @@ -0,0 +1,226 @@ +/* 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 + +#ifdef PADDLE_WITH_TESTING +#include "gtest/gtest.h" +#endif + +#include "paddle/framework/lod_tensor.h" +#include "paddle/framework/operator.h" + +namespace paddle { +namespace operators { + +/* + * This is an implementation of beam search. + * + * To explain the details, lets take machine translation task for example, in + * this task, one source sentence is translated to multiple target sentences, + * during this period, one sentence will be translated to multiple translation + * prefixes(target sentence that have not ended), in each time step a prefix + * will have some candidates, input the candidate ids and their corresponding + * scores (probabilities), it will sort and select the top beam_size candidates + * for each source sentence, and store the selected candidates's score and their + * corresponding ids to LoDTensors. + * + * A detailed example: + * + * Input + * + * ids: + * LoD (should have 2 levels) + * first level: [0, 1, 4] + * second level: [0, 1, 2, 3, 4] + * + * tensor's data + * [ + * [4, 2, 5] + * [2, 1, 3] + * [3, 5, 2] + * [8, 2, 1] + * ] + * + * scores: + * LoD same as `ids` + * tensor's data + * [ + * [0.5, 0.3, 0.2] + * [0.6, 0.3, 0.1] + * [0.9, 0.5, 0.1] + * [0.7, 0.5, 0.1] + * ] + * + * the inputs means that there are 2 source sentences to translate, and the + * first source has 1 prefix, the second source has 2 prefix. + * + * lets assume beam size is 2, and the beam search's output should be + * LoD + * first level: + * [0, 1, 2] + * second level: + * [0, 2, 4] + * + * tensor's data + * [[ + * 0.5, + * 0.3, + * 0.9, + * 0.7 + * ]] + * + * TODO all the prune operations should be in the beam search, so it is better + * to split the beam search algorithm into a sequence of smaller operators, and + * the prune operators can be inserted in this sequence. + */ +class BeamSearch { + public: + // TODO(superjom) make type customizable + using id_t = size_t; + using score_t = float; + /* + * Input the arguments that needed by this class. + */ + BeamSearch(const framework::LoDTensor& ids, + const framework::LoDTensor& scores, size_t level, size_t beam_size, + int end_id) + : beam_size_(beam_size), + ids_(&ids), + scores_(&scores), + lod_level_(level), + end_id_(end_id) {} + + /* + * The main function of beam search. + * + * @selected_ids: a [None, 1]-shaped tensor with LoD. + * In a machine translation model, it might be the candidate term id sets, + * each set stored as a varience-length sequence. + * The format might be described with a two-level LoD + * - [[0 1] + * - [0 1 2]] + * - [[] + * - [0 1]] + * the first level of LoD tells that there are two source sentences. The + * second level describes the details of the candidate id set's offsets in + * the + * source sentences. + * + * @selected_scores: a LoD tensor with the same shape and LoD with + * selected_ids. + * It stores the corresponding scores of candidate ids in selected_ids. + * + * Return false if all the input tensor is empty, in machine translation task + * that means no candidates is provided, and the task will stop running. + */ + void operator()(const framework::LoDTensor& pre_ids, + framework::LoDTensor* selected_ids, + framework::LoDTensor* selected_scores); + + protected: + /* + * The basic items help to sort. + */ + struct Item { + Item() {} + Item(size_t offset, size_t id, float score) + : offset(offset), id(id), score(score) {} + // offset in the lod_level_+1 + size_t offset; + // the candidate id + id_t id; + // the corresponding score + score_t score; + }; + + void PruneEndidCandidates(const framework::LoDTensor& pre_ids, + std::vector>* items); + + /* + * Transform the items into a map whose key is offset, value is the items. + * NOTE low performance + */ + std::vector> ToMap( + const std::vector>& inputs); + + /* + * For each source, select top beam_size records. + */ + std::vector> SelectTopBeamSizeItems(); + + /* + * Get the items of next source sequence, return false if no remaining items. + */ + bool NextItemSet(std::vector* items); + + private: + size_t beam_size_; + const framework::LoDTensor* ids_; + const framework::LoDTensor* scores_; + size_t lod_level_{0}; + size_t sent_offset_{0}; + int end_id_{0}; +}; + +class BeamSearchOp : public framework::OperatorBase { + public: + BeamSearchOp(const std::string& type, + const framework::VariableNameMap& inputs, + const framework::VariableNameMap& outputs, + const framework::AttributeMap& attrs) + : OperatorBase(type, inputs, outputs, attrs) {} + + BeamSearchOp(const BeamSearchOp& o) + : framework::OperatorBase( + static_cast(o)) { + PADDLE_THROW("Not Implemented"); + } + + void Run(const framework::Scope& scope, + const platform::DeviceContext& dev_ctx) const override { + LOG(INFO) << "run beam search op"; + auto ids_var = scope.FindVar(Input("ids")); + auto scores_var = scope.FindVar(Input("scores")); + auto pre_ids_var = scope.FindVar(Input("pre_ids")); + PADDLE_ENFORCE_NOT_NULL(ids_var); + PADDLE_ENFORCE_NOT_NULL(scores_var); + PADDLE_ENFORCE_NOT_NULL(pre_ids_var); + + auto& ids = ids_var->Get(); + auto& scores = scores_var->Get(); + auto& pre_ids = pre_ids_var->Get(); + size_t level = Attr("level"); + size_t beam_size = Attr("beam_size"); + int end_id = Attr("end_id"); + LOG(INFO) << "init beam search"; + BeamSearch alg(ids, scores, level, beam_size, end_id); + + LOG(INFO) << "after beam search"; + auto selected_ids_var = scope.FindVar(Output("selected_ids")); + auto selected_scores_var = scope.FindVar(Output("selected_scores")); + PADDLE_ENFORCE_NOT_NULL(selected_ids_var); + PADDLE_ENFORCE_NOT_NULL(selected_scores_var); + auto& selected_ids_tensor = + *selected_ids_var->GetMutable(); + auto& selected_scores_tensor = + *selected_scores_var->GetMutable(); + LOG(INFO) << "run beam search"; + alg(pre_ids, &selected_ids_tensor, &selected_scores_tensor); + LOG(INFO) << "finish beam search"; + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/bilinear_tensor_product_op.h b/paddle/operators/bilinear_tensor_product_op.h index ffa4f43a327418498c1f110504127e7d2878409d..1113a4c6f357edb4f6b14b73c6eec9c6cca24ce5 100644 --- a/paddle/operators/bilinear_tensor_product_op.h +++ b/paddle/operators/bilinear_tensor_product_op.h @@ -174,7 +174,7 @@ class BilinearTensorProductGradKernel : public framework::OpKernel { // Caculate the gradient of Input(Bias). if (d_bias) { d_bias->mutable_data(ctx.GetPlace()); - auto d_bias_mat = EigenMatrix::From(*d_bias); + auto d_bias_mat = framework::EigenVector::Flatten(*d_bias); d_bias_mat.device(place) = d_out_mat.sum(Eigen::DSizes(0)); } } diff --git a/paddle/operators/concat_op.cu b/paddle/operators/concat_op.cu.cc similarity index 100% rename from paddle/operators/concat_op.cu rename to paddle/operators/concat_op.cu.cc diff --git a/paddle/operators/conv2d_transpose_cudnn_op.cu b/paddle/operators/conv2d_transpose_cudnn_op.cu.cc similarity index 96% rename from paddle/operators/conv2d_transpose_cudnn_op.cu rename to paddle/operators/conv2d_transpose_cudnn_op.cu.cc index 694526ec01214acf2ec6a3d68d3cf072739ac185..eff058afc6cc5dacf2a054a33f352824865c1924 100644 --- a/paddle/operators/conv2d_transpose_cudnn_op.cu +++ b/paddle/operators/conv2d_transpose_cudnn_op.cu.cc @@ -200,9 +200,7 @@ class CudnnConvTransposeGradOpKernel : public framework::OpKernel { T alpha = 1.0f, beta = 0.0f; if (input_grad) { T* input_grad_data = input_grad->mutable_data(ctx.GetPlace()); - auto t = framework::EigenVector::Flatten(*input_grad); - t.device(ctx.GetEigenDevice()) = - t.constant(static_cast(0)); + math::set_constant(ctx.device_context(), input_grad, 0); PADDLE_ENFORCE(platform::dynload::cudnnConvolutionForward( handle, &alpha, cudnn_output_desc, output_grad_data, @@ -214,9 +212,8 @@ class CudnnConvTransposeGradOpKernel : public framework::OpKernel { // ------------------- cudnn conv backward filter --------------------- if (filter_grad) { T* filter_grad_data = filter_grad->mutable_data(ctx.GetPlace()); - auto t = framework::EigenVector::Flatten(*filter_grad); - t.device(ctx.GetEigenDevice()) = - t.constant(static_cast(0)); + math::set_constant(ctx.device_context(), filter_grad, 0); + // Gradient with respect to the filter PADDLE_ENFORCE(platform::dynload::cudnnConvolutionBackwardFilter( handle, &alpha, cudnn_output_desc, output_grad_data, cudnn_input_desc, diff --git a/paddle/operators/conv_cudnn_op.cc b/paddle/operators/conv_cudnn_op.cc index 97f31bf22d7072d89bd043045045dcb5bb5518b8..4c65b60d2349d2989128f4b1da705ea18391b8a3 100644 --- a/paddle/operators/conv_cudnn_op.cc +++ b/paddle/operators/conv_cudnn_op.cc @@ -22,8 +22,6 @@ class CudnnConvOpMaker : public Conv2DOpMaker { CudnnConvOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker) : Conv2DOpMaker(proto, op_checker) { - AddAttr>("dilations", "dilations of convolution operator.") - .SetDefault(std::vector{1, 1}); AddAttr("workspace_size_MB", "workspace size for cudnn, in MB, " "workspace is a section of GPU memory which will be " diff --git a/paddle/operators/conv_cudnn_op.cu b/paddle/operators/conv_cudnn_op.cu.cc similarity index 100% rename from paddle/operators/conv_cudnn_op.cu rename to paddle/operators/conv_cudnn_op.cu.cc diff --git a/paddle/operators/conv_op.cc b/paddle/operators/conv_op.cc index a6f65f10165929316f971d195f3790fd9e7ed376..687d741cb22a081eab18c61752200b9fd48f68a7 100644 --- a/paddle/operators/conv_op.cc +++ b/paddle/operators/conv_op.cc @@ -30,6 +30,7 @@ void ConvOp::InferShape(framework::InferShapeContext* ctx) const { std::vector strides = ctx->Attrs().Get>("strides"); std::vector paddings = ctx->Attrs().Get>("paddings"); int groups = ctx->Attrs().Get("groups"); + std::vector dilations = ctx->Attrs().Get>("dilations"); int input_channels = in_dims[1]; int output_channels = filter_dims[0]; @@ -52,9 +53,15 @@ void ConvOp::InferShape(framework::InferShapeContext* ctx) const { "The number of output channels should be divided by groups."); std::vector output_shape({in_dims[0], filter_dims[0]}); - for (size_t i = 0; i < paddings.size(); ++i) { + for (size_t i = 0; i < strides.size(); ++i) { + PADDLE_ENFORCE(in_dims[i + 2] + 2 * paddings[i] - + (dilations[i] * (filter_dims[i + 2] - 1) + 1) > + 0, + "Due to the settings of paddings, filter_dims and " + "dilations, the output size is less than 0, please check " + "again."); output_shape.push_back(OutputSize(in_dims[i + 2], filter_dims[i + 2], - paddings[i], strides[i])); + dilations[i], paddings[i], strides[i])); } ctx->SetOutputDim("Output", framework::make_ddim(output_shape)); } @@ -78,9 +85,15 @@ Conv2DOpMaker::Conv2DOpMaker(framework::OpProto* proto, AddOutput("Output", "(Tensor) The output tensor of convolution operator. " "The format of output tensor is also NCHW."); - AddAttr>("strides", "strides of convolution operator.") + AddAttr>("strides", + "(vector default:{1, 1}), the " + "strides(h_stride, w_stride) of " + "convolution operator.") .SetDefault({1, 1}); - AddAttr>("paddings", "paddings of convolution operator.") + AddAttr>("paddings", + "(vector default:{0, 0}), the " + "paddings(h_pad, w_pad) of " + "convolution operator.") .SetDefault({0, 0}); AddAttr( "groups", @@ -90,15 +103,20 @@ Conv2DOpMaker::Conv2DOpMaker(framework::OpProto* proto, "first half of the input channels, while the second half of the filters " "is only connected to the second half of the input channels.") .SetDefault(1); + AddAttr>("dilations", + "(vector default:{1, 1}), the " + "dilations(h_dilation, w_dilation) of " + "convolution operator.") + .SetDefault({1, 1}); AddComment(R"DOC( Convolution Operator. The convolution operation calculates the output based on the input, filter -and strides, paddings, groups parameters. The size of each dimension of the +and strides, paddings, groups, dilations parameters. The size of each dimension of the parameters is checked in the infer-shape. Input(Input, Filter) and output(Output) are in NCHW format. Where N is batch size, C is the number of channels, H is the height of the feature, and W is -the width of the feature. Parameters(ksize, strides, paddings) are two elements. +the width of the feature. Parameters(ksize, strides, paddings, dilations) are two elements. These two elements represent height and width, respectively. The input(X) size and output(Out) size may be different. @@ -109,8 +127,8 @@ Example: Output: Output shape: (N, C_out, H_out, W_out) where - H_out = (H_in - filter_size[0] + 2 * paddings[0]) / strides[0] + 1; - W_out = (W_in - filter_size[1] + 2 * paddings[1]) / strides[1] + 1; + H_out = (H_in + 2 * paddings[0] - (dilations[0]*(filter_size[0] - 1) + 1)) / strides[0] + 1; + W_out = (W_in + 2 * paddings[1] - (dilations[1]*(filter_size[1] - 1) + 1)) / strides[1] + 1; )DOC"); } @@ -135,13 +153,15 @@ Conv3DOpMaker::Conv3DOpMaker(framework::OpProto* proto, AddOutput("Output", "(Tensor) The output tensor of convolution operator." "The format of output tensor is also NCDHW."); - AddAttr>( - "strides", - "(vector, default:{0, 0, 0}), the strides of convolution operator.") + AddAttr>("strides", + "(vector, default:{1, 1, 1}), the " + "strides(d_stride, h_stride, w_stride) of " + "convolution operator.") .SetDefault({1, 1, 1}); - AddAttr>( - "paddings", - "(vector, default:{0, 0, 0}), the paddings of convolution operator.") + AddAttr>("paddings", + "(vector, default:{0, 0, 0}), the " + "paddings(d_pad, h_pad, w_pad) of convolution " + "operator.") .SetDefault({0, 0, 0}); AddAttr( "groups", @@ -151,6 +171,12 @@ Conv3DOpMaker::Conv3DOpMaker(framework::OpProto* proto, "first half of the input channels, while the second half of the filters " "is only connected to the second half of the input channels.") .SetDefault(1); + AddAttr>("dilations", + "(vector default:{1, 1, 1}), the " + "dilations(d_dilation, h_dilation, w_dilation) of " + "convolution operator. Currently, conv3d doesn't " + "support dilation.") + .SetDefault({1, 1, 1}); AddComment(R"DOC( Convolution3D Operator. diff --git a/paddle/operators/conv_op.cu b/paddle/operators/conv_op.cu.cc similarity index 100% rename from paddle/operators/conv_op.cu rename to paddle/operators/conv_op.cu.cc diff --git a/paddle/operators/conv_op.h b/paddle/operators/conv_op.h index 7c1729213bf3f5f3987afbf2d51d5b5339ae521d..fac5f1d0e25fe205f89fc7eeb9fadfd8431517d5 100644 --- a/paddle/operators/conv_op.h +++ b/paddle/operators/conv_op.h @@ -27,11 +27,24 @@ using Tensor = framework::Tensor; // Base convolution operator definations for other conv // like operators to reuse the implementation. -inline int OutputSize(int input_size, int filter_size, int padding, - int stride) { - int output_size = (input_size - filter_size + 2 * padding) / stride + 1; +inline int OutputSize(int input_size, int filter_size, int dilation, + int padding, int stride) { + const int dkernel = dilation * (filter_size - 1) + 1; + const int output_size = (input_size + 2 * padding - dkernel) / stride + 1; return output_size; } +inline bool IsExpand(std::vector& filter_dim, + std::vector& strides, std::vector& paddings, + std::vector& dilations) { + bool filter_1 = true, strides_1 = true, padding_0 = true, dilation_1 = true; + for (size_t j = 0; j < strides.size(); ++j) { + filter_1 = filter_1 && (static_cast(filter_dim[j]) == 1); + strides_1 = strides_1 && (strides[j] == 1); + padding_0 = padding_0 && (paddings[j] == 0); + dilation_1 = dilation_1 && (dilations[j] == 1); + } + return !(filter_1 && strides_1 && padding_0 && dilation_1); +} // Define Op classes in .h file so that other conv // operator implementations can reuse the code. @@ -50,14 +63,12 @@ class Conv3DOpMaker : public framework::OpProtoAndCheckerMaker { class ConvOp : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; - void InferShape(framework::InferShapeContext* ctx) const override; }; class ConvOpGrad : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; - void InferShape(framework::InferShapeContext* ctx) const override; }; @@ -73,9 +84,10 @@ class GemmConvKernel : public framework::OpKernel { Tensor* output = context.Output("Output"); output->mutable_data(context.GetPlace()); + int groups = context.Attr("groups"); std::vector strides = context.Attr>("strides"); std::vector paddings = context.Attr>("paddings"); - int groups = context.Attr("groups"); + std::vector dilations = context.Attr>("dilations"); const int batch_size = static_cast(input->dims()[0]); @@ -106,14 +118,17 @@ class GemmConvKernel : public framework::OpKernel { framework::DDim col_matrix_shape = framework::flatten_to_2d(col_shape, filter_shape_vec.size() + 1); + bool is_expand = IsExpand(filter_shape_vec, strides, paddings, dilations); Tensor col; - col.mutable_data(col_shape, context.GetPlace()); // col_matrix shares the same piece of data with col, // but will be reshaped into a two-dimensional matrix shape // to call the matrix multiplication interface. Tensor col_matrix; - col_matrix.ShareDataWith(col); - col_matrix.Resize(col_matrix_shape); + if (is_expand) { + col.mutable_data(col_shape, context.GetPlace()); + col_matrix.ShareDataWith(col); + col_matrix.Resize(col_matrix_shape); + } framework::DDim input_shape = framework::slice_ddim( input->dims(), 1, static_cast(input->dims().size())); @@ -130,24 +145,30 @@ class GemmConvKernel : public framework::OpKernel { int in_step = static_cast(input->dims()[1]) / groups; int out_step = static_cast(output->dims()[1]) / groups; + math::Vol2ColFunctor vol2col; + math::Im2ColFunctor im2col; + for (int i = 0; i < batch_size; i++) { Tensor in_batch = input->Slice(i, i + 1).Resize(input_shape); Tensor out_batch = output->Slice(i, i + 1).Resize(output_matrix_shape); + for (int g = 0; g < groups; g++) { Tensor in_slice = in_batch.Slice(g * in_step, (g + 1) * in_step); - if (filter_shape_vec.size() == 2) { + if (!is_expand) { + col.ShareDataWith(in_slice); + col_matrix.ShareDataWith(col); + col_matrix.Resize(col_matrix_shape); + } else if (filter_shape_vec.size() == 2) { // im2col - math::Im2ColFunctor im2col; - im2col(context.device_context(), in_slice, col, strides[0], - strides[1], paddings[0], paddings[0], paddings[1], - paddings[1]); + im2col(context.device_context(), in_slice, dilations, strides, + std::vector{paddings[0], paddings[1], paddings[0], + paddings[1]}, + &col); } else if (filter_shape_vec.size() == 3) { // vol2col - math::Vol2ColFunctor vol2col; - vol2col(context.device_context(), in_slice, col, strides[0], - strides[1], strides[2], paddings[0], paddings[1], - paddings[2]); + vol2col(context.device_context(), in_slice, dilations, strides, + paddings, &col); } // gemm @@ -178,9 +199,10 @@ class GemmConvGradKernel : public framework::OpKernel { if (!input_grad && !filter_grad) return; + int groups = context.Attr("groups"); std::vector strides = context.Attr>("strides"); std::vector paddings = context.Attr>("paddings"); - int groups = context.Attr("groups"); + std::vector dilations = context.Attr>("dilations"); const int batch_size = static_cast(input->dims()[0]); @@ -230,14 +252,17 @@ class GemmConvGradKernel : public framework::OpKernel { int in_step = static_cast(input->dims()[1]) / groups; int out_step = static_cast(output_grad->dims()[1]) / groups; + bool is_expand = IsExpand(filter_shape_vec, strides, paddings, dilations); Tensor col; // col_matrix shares the same piece of data with col, // but will be reshaped into a two-dimensional matrix shape // to call the matrix multiplication interface. Tensor col_matrix; - col.mutable_data(col_shape, context.GetPlace()); - col_matrix.ShareDataWith(col); - col_matrix.Resize(col_matrix_shape); + if (is_expand) { + col.mutable_data(col_shape, context.GetPlace()); + col_matrix.ShareDataWith(col); + col_matrix.Resize(col_matrix_shape); + } math::SetConstant set_zero; @@ -245,6 +270,9 @@ class GemmConvGradKernel : public framework::OpKernel { input_grad->mutable_data(context.GetPlace()); set_zero(context.device_context(), input_grad, static_cast(0)); + math::Col2VolFunctor col2vol; + math::Col2ImFunctor col2im; + for (int i = 0; i < batch_size; i++) { Tensor out_grad_batch = output_grad->Slice(i, i + 1).Resize(output_matrix_shape); @@ -254,24 +282,26 @@ class GemmConvGradKernel : public framework::OpKernel { Tensor out_grad_slice = out_grad_batch.Slice(g * out_step, (g + 1) * out_step); Tensor filter_slice = filter.Slice(g * out_step, (g + 1) * out_step); - math::matmul(context.device_context(), filter_slice, true, - out_grad_slice, false, T(1.0), &col_matrix, - T(0.0)); - // col2im + Tensor in_grad_slice = in_grad_batch.Slice(g * in_step, (g + 1) * in_step); - if (filter_shape_vec.size() == 2) { - math::Col2ImFunctor col2im; - col2im(context.device_context(), in_grad_slice, col, strides[0], - strides[1], paddings[0], paddings[0], paddings[1], - paddings[1]); + if (!is_expand) { + col_matrix.ShareDataWith(in_grad_slice); + col_matrix.Resize(col_matrix_shape); + } + math::matmul(context.device_context(), filter_slice, true, + out_grad_slice, false, T(1.0), &col_matrix, + T(0.0)); - } else if (filter_shape_vec.size() == 3) { - math::Col2VolFunctor col2vol; - col2vol(context.device_context(), in_grad_slice, col, strides[0], - strides[1], strides[2], paddings[0], paddings[1], - paddings[2]); + if (is_expand && filter_shape_vec.size() == 2) { + col2im(context.device_context(), col, dilations, strides, + std::vector{paddings[0], paddings[1], paddings[0], + paddings[1]}, + &in_grad_slice); + } else if (is_expand && filter_shape_vec.size() == 3) { + col2vol(context.device_context(), col, dilations, strides, paddings, + &in_grad_slice); } } } @@ -282,7 +312,8 @@ class GemmConvGradKernel : public framework::OpKernel { Tensor filter_grad_ = *filter_grad; filter_grad_.Resize(filter_matrix_shape); set_zero(context.device_context(), filter_grad, static_cast(0)); - + math::Im2ColFunctor im2col; + math::Vol2ColFunctor vol2col; for (int i = 0; i < batch_size; i++) { Tensor out_grad_batch = output_grad->Slice(i, i + 1).Resize(output_matrix_shape); @@ -293,16 +324,18 @@ class GemmConvGradKernel : public framework::OpKernel { out_grad_batch.Slice(g * out_step, (g + 1) * out_step); Tensor in_slice = in_batch.Slice(g * in_step, (g + 1) * in_step); - if (filter_shape_vec.size() == 2) { - math::Im2ColFunctor im2col; - im2col(context.device_context(), in_slice, col, strides[0], - strides[1], paddings[0], paddings[0], paddings[1], - paddings[1]); + if (!is_expand) { + col.ShareDataWith(in_slice); + col_matrix.ShareDataWith(col); + col_matrix.Resize(col_matrix_shape); + } else if (filter_shape_vec.size() == 2) { + im2col(context.device_context(), in_slice, dilations, strides, + std::vector{paddings[0], paddings[1], paddings[0], + paddings[1]}, + &col); } else if (filter_shape_vec.size() == 3) { - math::Vol2ColFunctor vol2col; - vol2col(context.device_context(), in_slice, col, strides[0], - strides[1], strides[2], paddings[0], paddings[1], - paddings[2]); + vol2col(context.device_context(), in_slice, dilations, strides, + paddings, &col); } // gemm diff --git a/paddle/operators/conv_shift_op.cu b/paddle/operators/conv_shift_op.cu index 74ed1b0ed358afc4f1a4e6a0c322eb032029d551..95e13c38a8dd234f49393d2d4808607a447b0d4c 100644 --- a/paddle/operators/conv_shift_op.cu +++ b/paddle/operators/conv_shift_op.cu @@ -13,6 +13,7 @@ limitations under the License. */ #include "paddle/operators/conv_shift_op.h" +#include "paddle/operators/math/math_function.h" #include "paddle/platform/cuda_helper.h" namespace paddle { @@ -22,7 +23,7 @@ using framework::Tensor; namespace { -inline int div_up(int x, int y) { return (x + y - 1) / y; } +inline int DivUp(int x, int y) { return (x + y - 1) / y; } // Some notes on the design: // @@ -33,9 +34,9 @@ inline int div_up(int x, int y) { return (x + y - 1) / y; } // 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) { +__global__ void ConvShiftForward(const T *x, const T *y, int x_width, + int y_width, int y_half_width, int batch_size, + T *out) { extern __shared__ T mem[]; int tx = threadIdx.x; @@ -62,25 +63,26 @@ __global__ void conv_shift_forward(const T *x, const T *y, T *out, int x_width, 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]; - } + if (tx < num_x) { + // 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; + // 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) { +__global__ void ConvShiftGradX(const T *dout, const T *y, int x_width, + int y_width, int y_half_width, int batch_size, + T *dx) { int i = blockIdx.x * blockDim.x + threadIdx.x; // x index int j = blockIdx.y; // y index int k = blockIdx.z; // batch index @@ -94,8 +96,8 @@ __global__ void conv_shift_dx(const T *dout, const T *y, T *dx, int x_width, // 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) { +__global__ void ConvShiftDy(const T *x, const T *dout, int x_width, int y_width, + int y_half_width, int batch_size, T *dy) { int i = blockIdx.x * blockDim.x + threadIdx.x; // x index int j = blockIdx.y; // y index int k = blockIdx.z; // batch index @@ -125,15 +127,15 @@ class ConvShiftKernel : public framework::OpKernel { 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 num_x_blocks = DivUp(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 = context.cuda_device_context().stream(); - conv_shift_forward<<>>( - x_data, y_data, out_data, x_width, y_width, y_half_width, batch_size); + ConvShiftForward<<>>( + x_data, y_data, x_width, y_width, y_half_width, batch_size, out_data); } }; @@ -157,25 +159,26 @@ class ConvShiftGradKernel int y_width = Y->dims()[1]; int y_half_width = (y_width - 1) / 2; - auto stream = context.cuda_device_context().stream(); + auto &device_ctx = context.cuda_device_context(); + math::SetConstant zero; const int x_per_block = 256; - int num_x_blocks = div_up(x_width, x_per_block); + int num_x_blocks = DivUp(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); + zero(device_ctx, dX, static_cast(0.0)); + ConvShiftGradX<<>>( + dout_data, y_data, x_width, y_width, y_half_width, batch_size, + dx_data); } 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); + zero(device_ctx, dY, static_cast(0.0)); + ConvShiftDy<<>>( + x_data, dout_data, x_width, y_width, y_half_width, batch_size, + dy_data); } } }; diff --git a/paddle/operators/conv_transpose_op.cc b/paddle/operators/conv_transpose_op.cc index 50081779a5ea3c81884007d4e4b7832dc4ea2bdd..310e3f5c937bd1345663b2c2307610a485a027ef 100644 --- a/paddle/operators/conv_transpose_op.cc +++ b/paddle/operators/conv_transpose_op.cc @@ -30,11 +30,6 @@ void ConvTransposeOp::InferShape(framework::InferShapeContext* ctx) const { std::vector strides = ctx->Attrs().Get>("strides"); std::vector paddings = ctx->Attrs().Get>("paddings"); - for (size_t i = 0; i < paddings.size(); ++i) { - PADDLE_ENFORCE_EQ(paddings[i], 0, - "No Padding allowed in conv transpose op."); - } - PADDLE_ENFORCE(in_dims.size() == 4 || in_dims.size() == 5, "ConvTransposeOp intput should be 4-D or 5-D tensor."); PADDLE_ENFORCE_EQ(in_dims.size(), filter_dims.size(), @@ -51,8 +46,8 @@ void ConvTransposeOp::InferShape(framework::InferShapeContext* ctx) const { "as the number of filters."); std::vector output_shape({in_dims[0], filter_dims[1]}); - for (size_t i = 0; i < paddings.size(); ++i) { - output_shape.push_back((in_dims[i + 2] - 1) * strides[i] + + for (size_t i = 0; i < strides.size(); ++i) { + output_shape.push_back((in_dims[i + 2] - 1) * strides[i] - 2 * paddings[i] + filter_dims[i + 2]); } ctx->SetOutputDim("Output", framework::make_ddim(output_shape)); @@ -79,11 +74,13 @@ Conv2DTransposeOpMaker::Conv2DTransposeOpMaker( "The format of output tensor is also NCHW."); AddAttr>( "strides", - "(vector defalut:{1, 1}), strides of convolution transpose operator.") + "(vector defalut:{1, 1}), the strides(h_stride, w_stride) of " + "convolution transpose operator.") .SetDefault({1, 1}); AddAttr>( "paddings", - "(vector defalut:{0, 0}), paddings of convolution transpose operator.") + "(vector defalut:{0, 0}), the paddings(h_pad, w_pad) of convolution " + "transpose operator.") .SetDefault({0, 0}); AddComment(R"DOC( Convolution2D Transpose Operator. @@ -132,13 +129,14 @@ Conv3DTransposeOpMaker::Conv3DTransposeOpMaker( "Where N is batch size, C is " "the number of channels, D is the depth of the feature, H is the " "height of the feature, and W is the width of the feature."); - AddAttr>( - "strides", - "(vector defalut:{1, 1, 1}), strides of convolution transpose operator.") + AddAttr>("strides", + "(vector defalut:{1, 1, 1}), the " + "strides{d_stride, h_stride, w_stride} of " + "convolution transpose operator.") .SetDefault({1, 1, 1}); - AddAttr>( - "paddings", - "(vector defalut:{0, 0, 0}), paddings of convolution transpose operator.") + AddAttr>("paddings", + "(vector defalut:{0, 0, 0}), paddings(d_pad, " + "h_pad, w_pad) of convolution transpose operator.") .SetDefault({0, 0, 0}); AddComment(R"DOC( Convolution3D Transpose Operator. diff --git a/paddle/operators/conv_transpose_op.cu b/paddle/operators/conv_transpose_op.cu.cc similarity index 100% rename from paddle/operators/conv_transpose_op.cu rename to paddle/operators/conv_transpose_op.cu.cc diff --git a/paddle/operators/conv_transpose_op.h b/paddle/operators/conv_transpose_op.h index 6c1a6220d784abf89ec789f94d9cff9e5414db04..ab336ad23ce1c180b68d04e4c85b299e301d5376 100644 --- a/paddle/operators/conv_transpose_op.h +++ b/paddle/operators/conv_transpose_op.h @@ -43,16 +43,12 @@ class Conv3DTransposeOpMaker : public framework::OpProtoAndCheckerMaker { class ConvTransposeOp : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; - - protected: void InferShape(framework::InferShapeContext* ctx) const override; }; class ConvTransposeOpGrad : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; - - protected: void InferShape(framework::InferShapeContext* ctx) const override; }; @@ -66,6 +62,7 @@ class GemmConvTransposeKernel : public framework::OpKernel { Tensor* output = context.Output("Output"); std::vector strides = context.Attr>("strides"); + std::vector paddings = context.Attr>("paddings"); // TODO(Zhuoyuan): Paddings can be added in future. // groups will alway be disabled in conv2dtranspose. @@ -120,6 +117,10 @@ class GemmConvTransposeKernel : public framework::OpKernel { math::SetConstant set_zero; set_zero(context.device_context(), output, static_cast(0)); + math::Col2ImFunctor col2im; + math::Col2VolFunctor col2vol; + std::vector dilations({1, 1, 1}); + // convolution transpose: gemm + col2im or col2vol (similar to conv-backward // on input) for (int i = 0; i < batch_size; i++) { @@ -138,16 +139,16 @@ class GemmConvTransposeKernel : public framework::OpKernel { if (filter_shape_vec.size() == 2) { // col2im: col_matrix -> dy // from (c * k_h * k_w, h * w) to (c, o_h, o_w) - math::Col2ImFunctor col2im; - - col2im(context.device_context(), output_batch, col, strides[0], - strides[1], 0, 0, 0, 0); + col2im(context.device_context(), col, + std::vector{dilations[0], dilations[1]}, strides, + std::vector{paddings[0], paddings[1], paddings[0], + paddings[1]}, + &output_batch); } else if (filter_shape_vec.size() == 3) { // col2vol: col_matrix -> dy // from (c * k_d * k_h * k_w, d * h * w) to (c, o_d, o_h, o_w) - math::Col2VolFunctor col2vol; - col2vol(context.device_context(), output_batch, col, strides[0], - strides[1], strides[2], 0, 0, 0); + col2vol(context.device_context(), col, dilations, strides, paddings, + &output_batch); } } } @@ -171,7 +172,6 @@ class GemmConvTransposeGradKernel : public framework::OpKernel { if ((!input_grad) && (!filter_grad)) return; std::vector strides = context.Attr>("strides"); - // Actually, no paddings and groups allowed in conv transpose. std::vector paddings = context.Attr>("paddings"); const int batch_size = static_cast(input->dims()[0]); @@ -228,6 +228,10 @@ class GemmConvTransposeGradKernel : public framework::OpKernel { Tensor filter_grad_; math::SetConstant set_zero; + math::Im2ColFunctor im2col; + math::Vol2ColFunctor vol2col; + std::vector dilations({1, 1, 1}); + if (input_grad) { input_grad->mutable_data(context.GetPlace()); set_zero(context.device_context(), input_grad, static_cast(0)); @@ -247,17 +251,16 @@ class GemmConvTransposeGradKernel : public framework::OpKernel { if (filter_shape_vec.size() == 2) { // im2col: dy -> col matrix // from (c, o_h, o_w) to (c * k_h * k_w, h * w) - math::Im2ColFunctor im2col; - im2col(context.device_context(), output_grad_batch, col, strides[0], - strides[1], paddings[0], paddings[0], paddings[1], - paddings[1]); + im2col(context.device_context(), output_grad_batch, + std::vector{dilations[0], dilations[1]}, strides, + std::vector{paddings[0], paddings[1], paddings[0], + paddings[1]}, + &col); } else if (filter_shape_vec.size() == 3) { // vol2col: dy -> col_matrix // from (c, o_d, o_h, o_w) to (c * k_d * k_h * k_w, d * h * w) - math::Vol2ColFunctor vol2col; - vol2col(context.device_context(), output_grad_batch, col, strides[0], - strides[1], strides[2], paddings[0], paddings[1], - paddings[2]); + vol2col(context.device_context(), output_grad_batch, dilations, + strides, paddings, &col); } if (input_grad) { diff --git a/paddle/operators/cos_sim_op.h b/paddle/operators/cos_sim_op.h index 68c56f531f941e1b8f66ac7ba6bf318881642c4f..62a4e484eceeabc4cc26e68ac54a50be1ac95df7 100644 --- a/paddle/operators/cos_sim_op.h +++ b/paddle/operators/cos_sim_op.h @@ -132,7 +132,7 @@ class CosSimGradKernel : public framework::OpKernel { // compute dy if (out_grad_y) { out_grad_y->mutable_data(context.GetPlace()); - auto dy = EigenMatrix::Reshape(*out_grad_y, 1); + auto dy = EigenVector::Flatten(*out_grad_y); auto grad = x / norm_prod_bcast - z_bcast * y_bcast / y_snorm_bcast; dy.device(place) = (dz_bcast * grad).sum(Eigen::array({{0}})); } diff --git a/paddle/operators/cross_entropy_op.cu b/paddle/operators/cross_entropy_op.cu index 530b319a44eac915f0d49eb55bfe5929908eab26..6212e39dfde33c5943958adbd1a0a052262e119e 100644 --- a/paddle/operators/cross_entropy_op.cu +++ b/paddle/operators/cross_entropy_op.cu @@ -23,8 +23,6 @@ template __global__ void CrossEntropyGradientKernel(T* dX, const T* dY, const T* X, const int64_t* 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) { int idx = i * D + label[i]; diff --git a/paddle/operators/detail/safe_ref.h b/paddle/operators/detail/safe_ref.h new file mode 100644 index 0000000000000000000000000000000000000000..b71af17309f9f46b5c87f0f479d4e03443fa7f93 --- /dev/null +++ b/paddle/operators/detail/safe_ref.h @@ -0,0 +1,31 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + + 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 + +namespace paddle { +namespace operators { +namespace detail { +/** + * Get Reference From Pointer with check. The error message is printf format, + * and passed by `args` + */ +template +inline T &Ref(T *ptr, ARGS &&... args) { + PADDLE_ENFORCE(ptr != nullptr, args...); + return *ptr; +} +} // namespace detail +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/fill_constant_batch_size_like_op.cc b/paddle/operators/fill_constant_batch_size_like_op.cc index 85871ebbfcd8ee38ef5e8078d1d6cb6bdda46a7b..985b5d1e865e513d833bff72dcd20a8f20851d8c 100644 --- a/paddle/operators/fill_constant_batch_size_like_op.cc +++ b/paddle/operators/fill_constant_batch_size_like_op.cc @@ -101,4 +101,7 @@ REGISTER_OPERATOR(fill_constant_batch_size_like, REGISTER_OP_CPU_KERNEL( fill_constant_batch_size_like, ops::FillConstantBatchSizeLikeOpKernel, - ops::FillConstantBatchSizeLikeOpKernel); + ops::FillConstantBatchSizeLikeOpKernel, + ops::FillConstantBatchSizeLikeOpKernel, + ops::FillConstantBatchSizeLikeOpKernel); diff --git a/paddle/operators/fill_constant_batch_size_like_op.cu b/paddle/operators/fill_constant_batch_size_like_op.cu.cc similarity index 81% rename from paddle/operators/fill_constant_batch_size_like_op.cu rename to paddle/operators/fill_constant_batch_size_like_op.cu.cc index 298c196f1dfef388640e34153264986bd518a11a..9e7a1eeab863c962ca72908e561e12a04d5021c5 100644 --- a/paddle/operators/fill_constant_batch_size_like_op.cu +++ b/paddle/operators/fill_constant_batch_size_like_op.cu.cc @@ -12,11 +12,14 @@ See the License for the specific language governing permissions and limitations under the License. */ -#include "paddle/framework/op_registry.h" #include "paddle/operators/fill_constant_batch_size_like_op.h" +#include "paddle/framework/op_registry.h" namespace ops = paddle::operators; REGISTER_OP_GPU_KERNEL( fill_constant_batch_size_like, ops::FillConstantBatchSizeLikeOpKernel, - ops::FillConstantBatchSizeLikeOpKernel); + ops::FillConstantBatchSizeLikeOpKernel, + ops::FillConstantBatchSizeLikeOpKernel, + ops::FillConstantBatchSizeLikeOpKernel); diff --git a/paddle/operators/fill_zeros_like_op.cc b/paddle/operators/fill_zeros_like_op.cc index 8ab39d4fb012b8fa3883f33e4d15be7918500354..95fb5932b8b555e1357adc9fdfb7b6e6db7da71d 100644 --- a/paddle/operators/fill_zeros_like_op.cc +++ b/paddle/operators/fill_zeros_like_op.cc @@ -54,5 +54,8 @@ namespace ops = paddle::operators; REGISTER_OP_WITHOUT_GRADIENT(fill_zeros_like, ops::FillZerosLikeOp, ops::FillZerosLikeOpMaker); REGISTER_OP_CPU_KERNEL( - fill_zeros_like, - ops::FillZerosLikeKernel); + fill_zeros_like, ops::FillZerosLikeKernel, + ops::FillZerosLikeKernel, + ops::FillZerosLikeKernel, + ops::FillZerosLikeKernel, + ops::FillZerosLikeKernel); diff --git a/paddle/operators/fill_zeros_like_op.cu b/paddle/operators/fill_zeros_like_op.cu.cc similarity index 69% rename from paddle/operators/fill_zeros_like_op.cu rename to paddle/operators/fill_zeros_like_op.cu.cc index a6d4ba64bde534ea76867c456537b130a45b9496..1501a17441072223ba0e8cf5b6c8cdd5e903a467 100644 --- a/paddle/operators/fill_zeros_like_op.cu +++ b/paddle/operators/fill_zeros_like_op.cu.cc @@ -12,10 +12,13 @@ See the License for the specific language governing permissions and limitations under the License. */ -#include "paddle/framework/op_registry.h" #include "paddle/operators/fill_zeros_like_op.h" +#include "paddle/framework/op_registry.h" namespace ops = paddle::operators; REGISTER_OP_GPU_KERNEL( - fill_zeros_like, - ops::FillZerosLikeKernel); + fill_zeros_like, ops::FillZerosLikeKernel, + ops::FillZerosLikeKernel, + ops::FillZerosLikeKernel, + ops::FillZerosLikeKernel, + ops::FillZerosLikeKernel); diff --git a/paddle/operators/gru_op.cu b/paddle/operators/gru_op.cu.cc similarity index 97% rename from paddle/operators/gru_op.cu rename to paddle/operators/gru_op.cu.cc index 35538c74b4bf678f8068999bfadb2589a1671be0..0ceff94ec3ddaadbd5f0ca4f5a4eebe6cb8ee3a9 100644 --- a/paddle/operators/gru_op.cu +++ b/paddle/operators/gru_op.cu.cc @@ -12,7 +12,6 @@ See the License for the specific language governing permissions and limitations under the License. */ -#define EIGEN_USE_GPU #include "paddle/operators/gru_op.h" namespace ops = paddle::operators; diff --git a/paddle/operators/gru_op.h b/paddle/operators/gru_op.h index ba90ec9816c40a6a49065ac6efcee6b93dffce90..55e9cc4a98bd6d36ce5d6bb4116039d0ec18b485 100644 --- a/paddle/operators/gru_op.h +++ b/paddle/operators/gru_op.h @@ -27,10 +27,6 @@ namespace operators { using Tensor = framework::Tensor; using LoDTensor = framework::LoDTensor; -template -using EigenMatrix = framework::EigenMatrix; - template class GRUKernel : public framework::OpKernel { public: @@ -57,19 +53,15 @@ class GRUKernel : public framework::OpKernel { bool is_reverse = context.Attr("is_reverse"); math::LoDTensor2BatchFunctor to_batch; - to_batch(context.device_context(), *input, *batch_gate, true, is_reverse); + auto& dev_ctx = context.device_context(); + to_batch(dev_ctx, *input, *batch_gate, true, is_reverse); - int frame_size = hidden_dims[1]; - int batch_size = hidden_dims[0]; - auto g = EigenMatrix::From(*batch_gate); - auto place = context.GetEigenDevice(); if (bias) { - auto b = EigenMatrix::From(*bias); - g.device(place) = g + - b.reshape(Eigen::array({{1, frame_size * 3}})) - .broadcast(Eigen::array({{batch_size, 1}})); + math::RowwiseAdd add_bias; + add_bias(dev_ctx, *batch_gate, *bias, batch_gate); } + int frame_size = hidden_dims[1]; math::hl_gru_value gru_value; gru_value.gateWeight = const_cast(weight_data); gru_value.stateWeight = @@ -89,7 +81,7 @@ class GRUKernel : public framework::OpKernel { gru_value.gateValue = gate_t.data(); gru_value.resetOutputValue = reset_hidden_prev_t.data(); math::GRUUnitFunctor::compute( - context.device_context(), gru_value, frame_size, cur_batch_size, + dev_ctx, gru_value, frame_size, cur_batch_size, math::ActiveType(context.Attr("activation")), math::ActiveType(context.Attr("gate_activation"))); gru_value.prevOutValue = gru_value.outputValue; @@ -97,7 +89,7 @@ class GRUKernel : public framework::OpKernel { math::Batch2LoDTensorFunctor to_seq; batch_hidden->set_lod(batch_gate->lod()); - to_seq(context.device_context(), *batch_hidden, *hidden); + to_seq(dev_ctx, *batch_hidden, *hidden); } void Compute(const framework::ExecutionContext& context) const override { @@ -138,15 +130,14 @@ class GRUGradKernel : public framework::OpKernel { batch_reset_hidden_prev_grad.mutable_data(hidden_dims, context.GetPlace()); math::SetConstant zero; - zero(context.device_context(), &batch_hidden_grad, static_cast(0.0)); - zero(context.device_context(), &batch_gate_grad, static_cast(0.0)); - zero(context.device_context(), &batch_reset_hidden_prev_grad, - static_cast(0.0)); + auto& dev_ctx = context.device_context(); + zero(dev_ctx, &batch_hidden_grad, static_cast(0.0)); + zero(dev_ctx, &batch_gate_grad, static_cast(0.0)); + zero(dev_ctx, &batch_reset_hidden_prev_grad, static_cast(0.0)); bool is_reverse = context.Attr("is_reverse"); batch_hidden_grad.set_lod(batch_hidden->lod()); - to_batch(context.device_context(), *hidden_grad, batch_hidden_grad, false, - is_reverse); + to_batch(dev_ctx, *hidden_grad, batch_hidden_grad, false, is_reverse); math::hl_gru_value gru_value; gru_value.gateWeight = const_cast(weight_data); @@ -157,7 +148,7 @@ class GRUGradKernel : public framework::OpKernel { if (weight_grad) { gru_grad.gateWeightGrad = weight_grad->mutable_data(context.GetPlace()); - zero(context.device_context(), weight_grad, static_cast(0.0)); + zero(dev_ctx, weight_grad, static_cast(0.0)); gru_grad.stateWeightGrad = weight_grad->data() + 2 * frame_size * frame_size; } else { @@ -188,7 +179,7 @@ class GRUGradKernel : public framework::OpKernel { gru_value.prevOutValue = const_cast(h0_data); if (h0_grad) { T* h0_grad_data = h0_grad->mutable_data(context.GetPlace()); - zero(context.device_context(), h0_grad, static_cast(0.0)); + zero(dev_ctx, h0_grad, static_cast(0.0)); gru_grad.prevOutGrad = h0_grad_data; } else { gru_grad.prevOutGrad = nullptr; @@ -202,8 +193,7 @@ class GRUGradKernel : public framework::OpKernel { } math::GRUUnitGradFunctor::compute( - context.device_context(), gru_value, gru_grad, frame_size, - cur_batch_size, + dev_ctx, gru_value, gru_grad, frame_size, cur_batch_size, math::ActiveType(context.Attr("activation")), math::ActiveType(context.Attr("gate_activation"))); } @@ -211,14 +201,12 @@ class GRUGradKernel : public framework::OpKernel { input_grad->mutable_data(context.GetPlace()); math::Batch2LoDTensorFunctor to_seq; batch_gate_grad.set_lod(batch_gate->lod()); - to_seq(context.device_context(), batch_gate_grad, *input_grad); + to_seq(dev_ctx, batch_gate_grad, *input_grad); } if (bias_grad) { bias_grad->mutable_data(context.GetPlace()); - auto d_b = EigenMatrix::From(*bias_grad); - auto d_g = EigenMatrix::From(batch_gate_grad); - auto place = context.GetEigenDevice(); - d_b.device(place) = d_g.sum(Eigen::array({{0}})); + math::ColwiseSum col_sum; + col_sum(dev_ctx, batch_gate_grad, bias_grad); } } diff --git a/paddle/operators/is_empty_op.cc b/paddle/operators/is_empty_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..54fecf44e881b5c283c81580fd161da9808d253e --- /dev/null +++ b/paddle/operators/is_empty_op.cc @@ -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. */ + +#include "paddle/framework/op_registry.h" +#include "paddle/framework/operator.h" + +namespace paddle { +namespace operators { + +constexpr char kInput[] = "X"; +constexpr char kOutput[] = "Out"; + +class IsEmptyOp : public framework::OperatorBase { + public: + IsEmptyOp(const std::string &type, const framework::VariableNameMap &inputs, + const framework::VariableNameMap &outputs, + const framework::AttributeMap &attrs) + : OperatorBase(type, inputs, outputs, attrs) {} + + void Run(const framework::Scope &scope, + const platform::DeviceContext &dev_ctx) const override { + // get input + auto *var = scope.FindVar(Input(kInput)); + PADDLE_ENFORCE_NOT_NULL(var); + auto &tensor = var->Get(); + // get output + auto *out = scope.FindVar(Output(kOutput)); + PADDLE_ENFORCE_NOT_NULL(out); + auto *out_tensor = out->GetMutable(); + + out_tensor->Resize({1}); + out_tensor->mutable_data(platform::CPUPlace())[0] = + framework::product(tensor.dims()) == 0; + } +}; + +class IsEmptyOpProtoMaker : public framework::OpProtoAndCheckerMaker { + public: + IsEmptyOpProtoMaker(framework::OpProto *proto, + framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput(kInput, "(Tensor) Tensor which is to be checked."); + AddOutput(kOutput, "(Tensor) a boolean Tensor that indicate empty or not."); + AddComment(R"DOC( +IsEmpty Operator which checks whether a tensor is empty. + +It will just return product(tensor.ddims()) > 0; + )DOC"); + } +}; + +} // namespace operators +} // namespace paddle + +REGISTER_OP_WITHOUT_GRADIENT(is_empty, paddle::operators::IsEmptyOp, + paddle::operators::IsEmptyOpProtoMaker); diff --git a/paddle/operators/lstm_op.cu b/paddle/operators/lstm_op.cu.cc similarity index 97% rename from paddle/operators/lstm_op.cu rename to paddle/operators/lstm_op.cu.cc index 9ad56941553bf19a56c25f41f76fe20dfa3a106f..610cbb03e890203407b1489800bc17f1a196d12c 100644 --- a/paddle/operators/lstm_op.cu +++ b/paddle/operators/lstm_op.cu.cc @@ -12,7 +12,6 @@ See the License for the specific language governing permissions and limitations under the License. */ -#define EIGEN_USE_GPU #include "paddle/operators/lstm_op.h" namespace ops = paddle::operators; diff --git a/paddle/operators/lstm_op.h b/paddle/operators/lstm_op.h index fca84e2d8fa832a3780eab7e0fa2facceb4d613b..721aa42c92f2926aabbc13d0a9027b2b4e573225 100644 --- a/paddle/operators/lstm_op.h +++ b/paddle/operators/lstm_op.h @@ -24,10 +24,6 @@ namespace operators { using LoDTensor = framework::LoDTensor; using Tensor = framework::Tensor; -template -using EigenMatrix = framework::EigenMatrix; - template inline void ReorderInitState(const platform::DeviceContext& ctx, const framework::Tensor& src, const size_t* index, @@ -65,16 +61,11 @@ class LSTMKernel : public framework::OpKernel { framework::DDim dims({in_dims[0], frame_size}); if (bias) { - Eigen::array extents({{1, 4 * frame_size}}); - Eigen::array offsets({{0, 0}}); - auto b = EigenMatrix::From(*bias); - auto gate = EigenMatrix::From(*batch_gate); - gate.device(ctx.GetEigenDevice()) = - gate + - b.slice(offsets, extents) - .reshape(Eigen::array({{1, frame_size * 4}})) - .broadcast( - Eigen::array({{static_cast(in_dims[0]), 1}})); + Tensor b = *bias; + b.Resize({bias->numel(), 1}); + Tensor gate_bias = b.Slice(0, 4 * frame_size); + math::RowwiseAdd add_bias; + add_bias(device_ctx, *batch_gate, gate_bias, batch_gate); } math::LstmMetaValue lstm_value; @@ -350,16 +341,11 @@ class LSTMGradKernel : public framework::OpKernel { } if (bias && bias_g) { /* backward bias */ - int m = static_cast(batch_gate_g.dims()[0]); - int n = static_cast(batch_gate_g.dims()[1]); - - Tensor ones; - ones.mutable_data({m}, ctx.GetPlace()); - math::SetConstant set; - set(device_ctx, &ones, static_cast(1.0)); - - math::gemv(device_ctx, true, m, n, 1., batch_gate_g.data(), - ones.data(), 0., bias_g->data()); + Tensor b_g = *bias_g; + b_g.Resize({bias_g->numel(), 1}); + Tensor gate_bias_g = b_g.Slice(0, 4 * frame_size); + math::ColwiseSum col_sum; + col_sum(device_ctx, batch_gate_g, &gate_bias_g); } if (h0 && h0_g) { diff --git a/paddle/operators/math/CMakeLists.txt b/paddle/operators/math/CMakeLists.txt index ab7f23f57043844d45c36acc475422613164bee1..002b68fecf4f1e294387357f0346d9926a2b2b5a 100644 --- a/paddle/operators/math/CMakeLists.txt +++ b/paddle/operators/math/CMakeLists.txt @@ -1,28 +1,28 @@ add_subdirectory(detail) if(WITH_GPU) - nv_library(math_function SRCS math_function.cc math_function.cu im2col.cc im2col.cu DEPS cblas device_context operator) + nv_library(math_function SRCS math_function.cc math_function.cu im2col.cc im2col.cu DEPS cblas device_context framework_proto) nv_test(math_function_gpu_test SRCS math_function_test.cu DEPS math_function tensor) nv_library(selected_rows_functor SRCS selected_rows_functor.cc selected_rows_functor.cu DEPS selected_rows math_function) nv_test(selected_rows_functor_gpu_test SRCS selected_rows_functor_test.cu DEPS selected_rows_functor) - nv_library(softmax SRCS softmax.cc softmax.cu DEPS operator) - nv_library(cross_entropy SRCS cross_entropy.cc cross_entropy.cu DEPS operator) + nv_library(softmax SRCS softmax.cc softmax.cu DEPS device_context) + nv_library(cross_entropy SRCS cross_entropy.cc cross_entropy.cu DEPS device_context) nv_library(pooling SRCS pooling.cc pooling.cu DEPS device_context) nv_library(sequence_pooling SRCS sequence_pooling.cc sequence_pooling.cu DEPS device_context math_function) nv_library(vol2col SRCS vol2col.cc vol2col.cu DEPS device_context) - nv_library(context_project SRCS context_project.cc context_project.cu DEPS device_context) + nv_library(context_project SRCS context_project.cc context_project.cu DEPS device_context math_function) nv_library(sequence2batch SRCS sequence2batch.cc sequence2batch.cu DEPS device_context) nv_library(lstm_compute SRCS lstm_compute.cc lstm_compute.cu DEPS device_context activation_functions) nv_library(gru_compute SRCS gru_compute.cc gru_compute.cu DEPS device_context activation_functions math_function) else() - cc_library(math_function SRCS math_function.cc im2col.cc DEPS cblas device_context operator) + cc_library(math_function SRCS math_function.cc im2col.cc DEPS cblas device_context framework_proto) cc_library(selected_rows_functor SRCS selected_rows_functor.cc DEPS selected_rows math_function) - cc_library(softmax SRCS softmax.cc DEPS operator) - cc_library(cross_entropy SRCS cross_entropy.cc DEPS operator) + cc_library(softmax SRCS softmax.cc DEPS device_context) + cc_library(cross_entropy SRCS cross_entropy.cc DEPS device_context) cc_library(pooling SRCS pooling.cc DEPS device_context) cc_library(sequence_pooling SRCS sequence_pooling.cc DEPS device_context math_function) cc_library(vol2col SRCS vol2col.cc DEPS device_context) - cc_library(context_project SRCS context_project.cc DEPS device_context) + cc_library(context_project SRCS context_project.cc DEPS device_context math_function) cc_library(sequence2batch SRCS sequence2batch.cc DEPS device_context) cc_library(lstm_compute SRCS lstm_compute.cc DEPS device_context activation_functions) cc_library(gru_compute SRCS gru_compute.cc DEPS device_context activation_functions math_function) diff --git a/paddle/operators/math/context_project.h b/paddle/operators/math/context_project.h index e0283360414fbdfb3dae2e94b45c9c8daeed3c74..72f4202bace4461d2597204feaa2a21e355bd1ac 100644 --- a/paddle/operators/math/context_project.h +++ b/paddle/operators/math/context_project.h @@ -14,9 +14,9 @@ limitations under the License. */ #pragma once -#include "paddle/framework/eigen.h" #include "paddle/framework/lod_tensor.h" #include "paddle/operators/math/im2col.h" +#include "paddle/operators/math/math_function.h" namespace paddle { namespace operators { @@ -24,9 +24,6 @@ namespace math { using Tensor = framework::Tensor; using LoDTensor = framework::LoDTensor; -template -using EigenMatrix = framework::EigenMatrix; /* * \brief Context projection concatenates features in adjacent time-steps in @@ -88,13 +85,18 @@ template class ContextProjectFunctor { public: void operator()(const platform::DeviceContext& context, const LoDTensor& in, - const Tensor& padding_data, Tensor& col, - bool padding_trainable, int context_start, int context_length, - int context_stride, int up_pad, int down_pad) { + const Tensor& padding_data, bool padding_trainable, + const int context_start, const int context_length, + const int context_stride, const int up_pad, + const int down_pad, Tensor* col) { auto lod_level_0 = in.lod()[0]; math::Im2ColFunctor im2col_ocf; + std::vector dilation({1, 1}); + std::vector padding({up_pad, 0, down_pad, 0}); + std::vector stride({context_stride, 1}); + int input_row_begin, input_row_end; int sequence_height, sequence_width; sequence_width = in.dims()[1]; @@ -105,8 +107,8 @@ class ContextProjectFunctor { : static_cast(lod_level_0[i]); input_row_end = static_cast(lod_level_0[i + 1]); - Tensor out_t = col.Slice(static_cast(lod_level_0[i]), - static_cast(lod_level_0[i + 1])); + Tensor out_t = col->Slice(static_cast(lod_level_0[i]), + static_cast(lod_level_0[i + 1])); sequence_height = static_cast(out_t.dims()[0]); @@ -123,17 +125,14 @@ class ContextProjectFunctor { {1, input_row_end - input_row_begin, sequence_width}); // input_channels, input_height, input_width in_t.Resize(framework::make_ddim(input_shape)); - - im2col_ocf(context, in_t, out_t, - /*stride_height*/ context_stride, /*stride_width*/ 1, up_pad, - down_pad, 0, 0); + im2col_ocf(context, in_t, dilation, stride, padding, &out_t); out_t.Resize({sequence_height, context_length * sequence_width}); } } if (padding_trainable) { for (int i = 0; i < static_cast(lod_level_0.size()) - 1; ++i) { - Tensor out_t = col.Slice(static_cast(lod_level_0[i]), - static_cast(lod_level_0[i + 1])); + Tensor out_t = col->Slice(static_cast(lod_level_0[i]), + static_cast(lod_level_0[i + 1])); sequence_height = static_cast(out_t.dims()[0]); @@ -150,9 +149,7 @@ class ContextProjectFunctor { Tensor out_t_sub = out_t.Slice(k * context_length, k * context_length + padding_size); Tensor w_sub = padding_data.Slice(k, k + padding_size); - auto out_t_sub_e = EigenMatrix::From(out_t_sub); - auto w_sub_e = EigenMatrix::From(w_sub); - out_t_sub_e.device(*context.GetEigenDevice()) = w_sub_e; + out_t_sub.CopyFrom(w_sub, context.GetPlace(), context); } } if (down_pad > 0) { // add down pad @@ -182,9 +179,7 @@ class ContextProjectFunctor { (down_pad_begin_row + t) * context_length); Tensor w_sub = padding_data.Slice( up_pad + padding_idx, up_pad + padding_idx + padding_size); - auto out_t_sub_e = EigenMatrix::From(out_t_sub); - auto w_sub_e = EigenMatrix::From(w_sub); - out_t_sub_e.device(*context.GetEigenDevice()) = w_sub_e; + out_t_sub.CopyFrom(w_sub, context.GetPlace(), context); } } out_t.Resize({sequence_height, context_length * sequence_width}); @@ -196,14 +191,19 @@ class ContextProjectFunctor { template class ContextProjectGradFunctor { public: - void operator()(const platform::DeviceContext& context, LoDTensor& in, - Tensor& padding_data, Tensor& col, bool padding_trainable, - int context_start, int context_length, int context_stride, - int up_pad, int down_pad, bool input_grad, bool pad_grad) { + void operator()(const platform::DeviceContext& context, const LoDTensor& in, + bool padding_trainable, const int context_start, + const int context_length, const int context_stride, + const int up_pad, const int down_pad, bool pad_grad, + bool input_grad, Tensor* padding_data, Tensor* col) { auto lod_level_0 = in.lod()[0]; math::Col2ImFunctor col2im_ocf; + std::vector dilation({1, 1}); + std::vector padding({up_pad, 0, down_pad, 0}); + std::vector stride({context_stride, 1}); + int input_row_begin, input_row_end; int sequence_height, sequence_width; sequence_width = in.dims()[1]; @@ -215,8 +215,8 @@ class ContextProjectGradFunctor { : static_cast(lod_level_0[i]); input_row_end = static_cast(lod_level_0[i + 1]); - Tensor out_t = col.Slice(static_cast(lod_level_0[i]), - static_cast(lod_level_0[i + 1])); + Tensor out_t = col->Slice(static_cast(lod_level_0[i]), + static_cast(lod_level_0[i + 1])); sequence_height = static_cast(out_t.dims()[0]); @@ -234,9 +234,7 @@ class ContextProjectGradFunctor { sequence_width}); // input_channels, input_height, input_width in_t.Resize(framework::make_ddim(input_shape)); - col2im_ocf(context, in_t, out_t, - /*stride_height*/ context_stride, /*stride_width*/ 1, - up_pad, down_pad, 0, 0); + col2im_ocf(context, out_t, dilation, stride, padding, &in_t); out_t.Resize({sequence_height, context_length * sequence_width}); } } @@ -244,8 +242,8 @@ class ContextProjectGradFunctor { if (pad_grad) { if (padding_trainable) { for (int i = 0; i < static_cast(lod_level_0.size()) - 1; ++i) { - Tensor out_t = col.Slice(static_cast(lod_level_0[i]), - static_cast(lod_level_0[i + 1])); + Tensor out_t = col->Slice(static_cast(lod_level_0[i]), + static_cast(lod_level_0[i + 1])); sequence_height = static_cast(out_t.dims()[0]); out_t.Resize({sequence_height * context_length, sequence_width}); @@ -259,11 +257,9 @@ class ContextProjectGradFunctor { k + context_length < up_pad ? context_length : up_pad - k; Tensor out_t_sub = out_t.Slice(k * context_length, k * context_length + padding_size); - Tensor w_sub = padding_data.Slice(k, k + padding_size); - auto out_t_sub_e = EigenMatrix::From(out_t_sub); - auto w_sub_e = EigenMatrix::From(w_sub); - w_sub_e.device(*context.GetEigenDevice()) = - w_sub_e + out_t_sub_e; + Tensor w_sub = padding_data->Slice(k, k + padding_size); + axpy(context, w_sub.numel(), static_cast(1), + out_t_sub.data(), w_sub.data()); } } if (down_pad > 0) { @@ -292,12 +288,10 @@ class ContextProjectGradFunctor { Tensor out_t_sub = out_t.Slice( (down_pad_begin_row + t) * context_length - padding_size, (down_pad_begin_row + t) * context_length); - Tensor w_sub = padding_data.Slice( + Tensor w_sub = padding_data->Slice( up_pad + padding_idx, up_pad + padding_idx + padding_size); - auto out_t_sub_e = EigenMatrix::From(out_t_sub); - auto w_sub_e = EigenMatrix::From(w_sub); - w_sub_e.device(*context.GetEigenDevice()) = - w_sub_e + out_t_sub_e; + axpy(context, w_sub.numel(), static_cast(1), + out_t_sub.data(), w_sub.data()); } } out_t.Resize({sequence_height, context_length * sequence_width}); diff --git a/paddle/operators/math/cross_entropy.h b/paddle/operators/math/cross_entropy.h index 0ab6827ffa8f8b90b432a801607a97206e010cf4..70ed9ddd551bb8cb7989727c02fea870186c9f2e 100644 --- a/paddle/operators/math/cross_entropy.h +++ b/paddle/operators/math/cross_entropy.h @@ -14,7 +14,6 @@ #pragma once #include "paddle/framework/eigen.h" -#include "paddle/framework/operator.h" #include "paddle/framework/tensor.h" #include "paddle/platform/hostdevice.h" diff --git a/paddle/operators/math/im2col.cc b/paddle/operators/math/im2col.cc index 3b1b0bd71dd3768b932864e185af8dc839b4653e..c10c44c52076c8ee56eee3a0d82c31df70a1c9c7 100644 --- a/paddle/operators/math/im2col.cc +++ b/paddle/operators/math/im2col.cc @@ -28,57 +28,55 @@ class Im2ColFunctor { public: void operator()(const platform::DeviceContext& context, - const framework::Tensor& im, framework::Tensor& col, - int stride_height, int stride_width, int padding_up, - int padding_down, int padding_left, int padding_right) { + const framework::Tensor& im, const std::vector& dilation, + const std::vector& stride, + const std::vector& padding, framework::Tensor* col) { PADDLE_ENFORCE(im.dims().size() == 3); - PADDLE_ENFORCE(col.dims().size() == 5); + PADDLE_ENFORCE(col->dims().size() == 5); - int input_channels = im.dims()[0]; - int input_height = im.dims()[1]; - int input_width = im.dims()[2]; - int filter_height = col.dims()[1]; - int filter_width = col.dims()[2]; - int output_height = col.dims()[3]; - int output_width = col.dims()[4]; + int im_channels = im.dims()[0]; + int im_height = im.dims()[1]; + int im_width = im.dims()[2]; + int filter_height = col->dims()[1]; + int filter_width = col->dims()[2]; + int col_height = col->dims()[3]; + int col_width = col->dims()[4]; - PADDLE_ENFORCE_EQ( - (input_height + padding_up + padding_down - filter_height) / - stride_height + - 1, - output_height, - "Output_height and padding(padding_up, padding_down) are " - "inconsistent."); - PADDLE_ENFORCE_EQ( - (input_width + padding_left + padding_right - filter_width) / - stride_width + - 1, - output_width, - "output_width and padding(padding_left, padding_right) are " - "inconsistent."); + PADDLE_ENFORCE_EQ((im_height + padding[0] + padding[2] - + ((dilation[0] * (filter_height - 1) + 1))) / + stride[0] + + 1, + col_height, + "Output_height and padding(padding_up, padding_down) are " + "inconsistent."); + PADDLE_ENFORCE_EQ((im_width + padding[1] + padding[3] - + ((dilation[1] * (filter_width - 1) + 1))) / + stride[1] + + 1, + col_width, + "Output_height and padding(padding_up, padding_down) are " + "inconsistent."); - int channels_col = input_channels * filter_height * filter_width; + int channels_col = im_channels * filter_height * filter_width; const T* im_data = im.data(); - T* col_data = col.data(); + T* col_data = col->data(); for (int c = 0; c < channels_col; ++c) { int w_offset = c % filter_width; int h_offset = (c / filter_width) % filter_height; int c_im = c / filter_width / filter_height; - for (int h = 0; h < output_height; ++h) { - for (int w = 0; w < output_width; ++w) { - int im_row_idx = h * stride_height + h_offset - padding_up; - int im_col_idx = w * stride_width + w_offset - padding_left; + for (int h = 0; h < col_height; ++h) { + for (int w = 0; w < col_width; ++w) { + int im_row_idx = h * stride[0] - padding[0] + h_offset * dilation[0]; + int im_col_idx = w * stride[1] - padding[1] + w_offset * dilation[1]; + int col_idx = (c * col_height + h) * col_width + w; + int im_idx = (im_row_idx + c_im * im_height) * im_width + im_col_idx; - if (im_row_idx < 0 || im_row_idx >= input_height || im_col_idx < 0 || - im_col_idx >= input_width) { - col_data[(c * output_height + h) * output_width + w] = T(0); - } else { - im_row_idx += c_im * input_height; - col_data[(c * output_height + h) * output_width + w] = - im_data[im_row_idx * input_width + im_col_idx]; - } + col_data[col_idx] = (im_row_idx < 0 || im_row_idx >= im_height || + im_col_idx < 0 || im_col_idx >= im_width) + ? static_cast(0) + : im_data[im_idx]; } } } @@ -94,54 +92,55 @@ template class Col2ImFunctor { public: - void operator()(const platform::DeviceContext& context, framework::Tensor& im, - const framework::Tensor& col, int stride_height, - int stride_width, int padding_up, int padding_down, - int padding_left, int padding_right) { - PADDLE_ENFORCE(im.dims().size() == 3); + void operator()(const platform::DeviceContext& context, + const framework::Tensor& col, + const std::vector& dilation, + const std::vector& stride, + const std::vector& padding, framework::Tensor* im) { + PADDLE_ENFORCE(im->dims().size() == 3); PADDLE_ENFORCE(col.dims().size() == 5); - int input_channels = im.dims()[0]; - int input_height = im.dims()[1]; - int input_width = im.dims()[2]; + int im_channels = im->dims()[0]; + int im_height = im->dims()[1]; + int im_width = im->dims()[2]; int filter_height = col.dims()[1]; int filter_width = col.dims()[2]; - int output_height = col.dims()[3]; - int output_width = col.dims()[4]; + int col_height = col.dims()[3]; + int col_width = col.dims()[4]; - PADDLE_ENFORCE_EQ( - (input_height + padding_up + padding_down - filter_height) / - stride_height + - 1, - output_height, - "Output_height and padding(padding_up, padding_down) are " - "inconsistent."); - PADDLE_ENFORCE_EQ( - (input_width + padding_left + padding_right - filter_width) / - stride_width + - 1, - output_width, - "output_width and padding(padding_left, padding_right) are " - "inconsistent."); + PADDLE_ENFORCE_EQ((im_height + padding[0] + padding[2] - + ((dilation[0] * (filter_height - 1) + 1))) / + stride[0] + + 1, + col_height, + "Output_height and padding(padding_up, padding_down) are " + "inconsistent."); + PADDLE_ENFORCE_EQ((im_width + padding[1] + padding[3] - + ((dilation[1] * (filter_width - 1) + 1))) / + stride[1] + + 1, + col_width, + "Output_height and padding(padding_up, padding_down) are " + "inconsistent."); - int channels_col = input_channels * filter_height * filter_width; + int channels_col = im_channels * filter_height * filter_width; - T* im_data = im.data(); + T* im_data = im->data(); const T* col_data = col.data(); for (int c = 0; c < channels_col; ++c) { int w_offset = c % filter_width; int h_offset = (c / filter_width) % filter_height; int c_im = c / filter_width / filter_height; - for (int h = 0; h < output_height; ++h) { - for (int w = 0; w < output_width; ++w) { - int im_row_idx = h * stride_height + h_offset - padding_up; - int im_col_idx = w * stride_width + w_offset - padding_left; + for (int h = 0; h < col_height; ++h) { + for (int w = 0; w < col_width; ++w) { + int im_row_idx = h * stride[0] - padding[0] + h_offset * dilation[0]; + int im_col_idx = w * stride[1] - padding[1] + w_offset * dilation[1]; - if ((im_row_idx) >= 0 && (im_row_idx) < input_height && - (im_col_idx) >= 0 && (im_col_idx) < input_width) { - im_row_idx += c_im * input_height; - im_data[im_row_idx * input_width + im_col_idx] += - col_data[(c * output_height + h) * output_width + w]; + if ((im_row_idx) >= 0 && (im_row_idx) < im_height && + (im_col_idx) >= 0 && (im_col_idx) < im_width) { + im_row_idx += c_im * im_height; + im_data[im_row_idx * im_width + im_col_idx] += + col_data[(c * col_height + h) * col_width + w]; } } } @@ -168,64 +167,59 @@ class Im2ColFunctor { public: void operator()(const platform::DeviceContext& context, - const framework::Tensor& im, framework::Tensor& col, - int stride_height, int stride_width, int padding_up, - int padding_down, int padding_left, int padding_right) { + const framework::Tensor& im, const std::vector& dilation, + const std::vector& stride, + const std::vector& padding, framework::Tensor* col) { PADDLE_ENFORCE(im.dims().size() == 3); - PADDLE_ENFORCE(col.dims().size() == 5); - int input_channels = im.dims()[0]; - int input_height = im.dims()[1]; - int input_width = im.dims()[2]; - int filter_height = col.dims()[3]; - int filter_width = col.dims()[4]; - int output_height = col.dims()[0]; - int output_width = col.dims()[1]; + PADDLE_ENFORCE(col->dims().size() == 5); + int im_channels = im.dims()[0]; + int im_height = im.dims()[1]; + int im_width = im.dims()[2]; + int filter_height = col->dims()[3]; + int filter_width = col->dims()[4]; + int col_height = col->dims()[0]; + int col_width = col->dims()[1]; PADDLE_ENFORCE_EQ( - (input_height + padding_up + padding_down - filter_height) / - stride_height + - 1, - output_height, + (im_height + padding[0] + padding[2] - filter_height) / stride[0] + 1, + col_height, "Output_height and padding(padding_up, padding_down) are " "inconsistent."); PADDLE_ENFORCE_EQ( - (input_width + padding_left + padding_right - filter_width) / - stride_width + - 1, - output_width, - "output_width and padding(padding_left, padding_right) are " + (im_width + padding[1] + padding[3] - filter_width) / stride[1] + 1, + col_width, + "col_width and padding(padding_left, padding_right) are " "inconsistent."); const T* im_data = im.data(); - T* col_data = col.data(); + T* col_data = col->data(); - for (int col_row_idx = 0; col_row_idx < output_height; ++col_row_idx) { - for (int col_col_idx = 0; col_col_idx < output_width; ++col_col_idx) { - for (int channel = 0; channel < input_channels; ++channel) { + for (int col_row_idx = 0; col_row_idx < col_height; ++col_row_idx) { + for (int col_col_idx = 0; col_col_idx < col_width; ++col_col_idx) { + for (int channel = 0; channel < im_channels; ++channel) { for (int filter_row_idx = 0; filter_row_idx < filter_height; ++filter_row_idx) { for (int filter_col_idx = 0; filter_col_idx < filter_width; ++filter_col_idx) { int im_row_offset = - col_row_idx * stride_height + filter_row_idx - padding_up; + col_row_idx * stride[0] + filter_row_idx - padding[0]; int im_col_offset = - col_col_idx * stride_width + filter_col_idx - padding_left; - int col_offset = ((((col_row_idx)*output_width + col_col_idx) * - input_channels + - channel) * - filter_height + - filter_row_idx) * - filter_width + - filter_col_idx; - if (im_row_offset < 0 || im_row_offset >= input_height || - im_col_offset < 0 || im_col_offset >= input_width) { - col_data[col_offset] = T(0); - } else { - int im_offset = - (channel * input_height + im_row_offset) * input_width + - im_col_offset; - col_data[col_offset] = im_data[im_offset]; - } + col_col_idx * stride[1] + filter_col_idx - padding[1]; + int col_offset = + ((((col_row_idx)*col_width + col_col_idx) * im_channels + + channel) * + filter_height + + filter_row_idx) * + filter_width + + filter_col_idx; + + int im_offset = (channel * im_height + im_row_offset) * im_width + + im_col_offset; + col_data[col_offset] = + (im_row_offset < 0 || im_row_offset >= im_height || + im_col_offset < 0 || im_col_offset >= im_width) + ? static_cast(0) + : im_data[im_offset]; } } } @@ -243,60 +237,57 @@ template class Col2ImFunctor { public: - void operator()(const platform::DeviceContext& context, framework::Tensor& im, - const framework::Tensor& col, int stride_height, - int stride_width, int padding_up, int padding_down, - int padding_left, int padding_right) { - PADDLE_ENFORCE(im.dims().size() == 3); + void operator()(const platform::DeviceContext& context, + const framework::Tensor& col, + const std::vector& dilation, + const std::vector& stride, + const std::vector& padding, framework::Tensor* im) { + PADDLE_ENFORCE(im->dims().size() == 3); PADDLE_ENFORCE(col.dims().size() == 5); - int input_channels = im.dims()[0]; - int input_height = im.dims()[1]; - int input_width = im.dims()[2]; + int im_channels = im->dims()[0]; + int im_height = im->dims()[1]; + int im_width = im->dims()[2]; int filter_height = col.dims()[3]; int filter_width = col.dims()[4]; - int output_height = col.dims()[0]; - int output_width = col.dims()[1]; + int col_height = col.dims()[0]; + int col_width = col.dims()[1]; PADDLE_ENFORCE_EQ( - (input_height + padding_up + padding_down - filter_height) / - stride_height + - 1, - output_height, + (im_height + padding[0] + padding[2] - filter_height) / stride[0] + 1, + col_height, "Output_height and padding(padding_up, padding_down) are " "inconsistent."); PADDLE_ENFORCE_EQ( - (input_width + padding_left + padding_right - filter_width) / - stride_width + - 1, - output_width, - "output_width and padding(padding_left, padding_right) are " + (im_width + padding[1] + padding[3] - filter_width) / stride[1] + 1, + col_width, + "col_width and padding(padding_left, padding_right) are " "inconsistent."); - T* im_data = im.data(); + T* im_data = im->data(); const T* col_data = col.data(); - for (int col_row_idx = 0; col_row_idx < output_height; ++col_row_idx) { - for (int col_col_idx = 0; col_col_idx < output_width; ++col_col_idx) { - for (int channel = 0; channel < input_channels; ++channel) { + for (int col_row_idx = 0; col_row_idx < col_height; ++col_row_idx) { + for (int col_col_idx = 0; col_col_idx < col_width; ++col_col_idx) { + for (int channel = 0; channel < im_channels; ++channel) { for (int filter_row_idx = 0; filter_row_idx < filter_height; ++filter_row_idx) { for (int filter_col_idx = 0; filter_col_idx < filter_width; ++filter_col_idx) { int im_row_offset = - col_row_idx * stride_height + filter_row_idx - padding_up; + col_row_idx * stride[0] + filter_row_idx - padding[0]; int im_col_offset = - col_col_idx * stride_width + filter_col_idx - padding_left; - int col_offset = (((col_row_idx * output_width + col_col_idx) * - input_channels + - channel) * - filter_height + - filter_row_idx) * - filter_width + - filter_col_idx; - if (im_row_offset >= 0 && im_row_offset < input_height && - im_col_offset >= 0 && im_col_offset < input_width) { + col_col_idx * stride[1] + filter_col_idx - padding[1]; + int col_offset = + (((col_row_idx * col_width + col_col_idx) * im_channels + + channel) * + filter_height + + filter_row_idx) * + filter_width + + filter_col_idx; + if (im_row_offset >= 0 && im_row_offset < im_height && + im_col_offset >= 0 && im_col_offset < im_width) { int im_offset = - (channel * input_height + im_row_offset) * input_width + + (channel * im_height + im_row_offset) * im_width + im_col_offset; im_data[im_offset] += col_data[col_offset]; } diff --git a/paddle/operators/math/im2col.cu b/paddle/operators/math/im2col.cu index 7b201fdbf3c5dd7d336d359e00b7323cecc0231a..bf7894243919571c2ab15d53690b1ef05bfcc6ee 100644 --- a/paddle/operators/math/im2col.cu +++ b/paddle/operators/math/im2col.cu @@ -20,36 +20,32 @@ namespace operators { namespace math { template -__global__ void im2col(const T* data_im, int num_outs, int height, int width, +__global__ void im2col(const T* data_im, int num_outs, int im_height, + int im_width, int dilation_h, int dilation_w, int filter_height, int filter_width, int stride_height, int stride_width, int padding_height, int padding_width, - int output_height, int output_width, T* data_col) { - int index = (blockIdx.x * gridDim.y + blockIdx.y) * blockDim.x + threadIdx.x; + int col_height, int col_width, T* data_col) { + const int index = + (blockIdx.x * gridDim.y + blockIdx.y) * blockDim.x + threadIdx.x; if (index < num_outs) { - int w_out = index % output_width; - index /= output_width; - int h_out = index % output_height; - int channel_in = index / output_height; + int w_out = index % col_width; + int h_out = (index / col_width) % col_height; + int channel_in = index / col_width / col_height; int channel_out = channel_in * filter_height * filter_width; - int h_in = h_out * stride_height; - int w_in = w_out * stride_width; + int h_in = h_out * stride_height - padding_height; + int w_in = w_out * stride_width - padding_width; - data_col += (channel_out * output_height + h_out) * output_width + w_out; + data_col += (channel_out * col_height + h_out) * col_width + w_out; + data_im += (channel_in * im_height + h_in) * im_width + w_in; for (int i = 0; i < filter_height; ++i) { for (int j = 0; j < filter_width; ++j) { - int rIdx = int(h_in + i); - int cIdx = int(w_in + j); - if ((rIdx - (int)padding_height) >= (int)height || - (rIdx - (int)padding_height) < 0 || - (cIdx - (int)padding_width) >= (int)width || - (cIdx - (int)padding_width) < 0) { - *data_col = 0; - } else { - rIdx = rIdx + channel_in * height - padding_height; - cIdx = cIdx - padding_width; - *data_col = data_im[rIdx * width + cIdx]; - } - data_col += output_height * output_width; + int rIdx = h_in + i * dilation_h; + int cIdx = w_in + j * dilation_w; + *data_col = + (rIdx >= im_height || rIdx < 0 || cIdx >= im_width || cIdx < 0) + ? 0 + : data_im[i * dilation_h * im_width + j * dilation_w]; + data_col += col_height * col_width; } } } @@ -65,30 +61,36 @@ class Im2ColFunctor { public: void operator()(const platform::DeviceContext& context, - const framework::Tensor& im, framework::Tensor& col, - int stride_height, int stride_width, int padding_up, - int padding_down, int padding_left, int padding_right) { + const framework::Tensor& im, const std::vector& dilation, + const std::vector& stride, + const std::vector& padding, framework::Tensor* col) { PADDLE_ENFORCE(im.dims().size() == 3); - PADDLE_ENFORCE(col.dims().size() == 5); - - int input_channels = im.dims()[0]; - int input_height = im.dims()[1]; - int input_width = im.dims()[2]; - int filter_height = col.dims()[1]; - int filter_width = col.dims()[2]; - int output_height = col.dims()[3]; - int output_width = col.dims()[4]; - - PADDLE_ENFORCE((input_height + padding_up + padding_down - filter_height) / - stride_height + - 1 == - output_height); - PADDLE_ENFORCE((input_width + padding_left + padding_right - filter_width) / - stride_width + - 1 == - output_width); - - int num_outputs = input_channels * output_height * output_width; + PADDLE_ENFORCE(col->dims().size() == 5); + + int im_channels = im.dims()[0]; + int im_height = im.dims()[1]; + int im_width = im.dims()[2]; + int filter_height = col->dims()[1]; + int filter_width = col->dims()[2]; + int col_height = col->dims()[3]; + int col_width = col->dims()[4]; + + PADDLE_ENFORCE_EQ((im_height + padding[0] + padding[2] - + (dilation[0] * (filter_height - 1) + 1)) / + stride[0] + + 1, + col_height, + "Output_height and padding(padding_up, padding_down) are " + "inconsistent."); + PADDLE_ENFORCE_EQ((im_width + padding[1] + padding[3] - + (dilation[1] * (filter_width - 1) + 1)) / + stride[1] + + 1, + col_width, + "col_width and padding(padding_left, padding_right) are " + "inconsistent."); + + int num_outputs = im_channels * col_height * col_width; int blocks = (num_outputs + 1024 - 1) / 1024; int block_x = 512; int block_y = (blocks + 512 - 1) / 512; @@ -97,56 +99,57 @@ class Im2ColFunctor<<(context) .stream()>>>( - im.data(), num_outputs, input_height, input_width, filter_height, - filter_width, stride_height, stride_width, padding_up, padding_left, - output_height, output_width, col.data()); + im.data(), num_outputs, im_height, im_width, dilation[0], + dilation[1], filter_height, filter_width, stride[0], stride[1], + padding[0], padding[1], col_height, col_width, col->data()); } }; template -__global__ void col2im(size_t n, const T* data_col, size_t height, size_t width, - size_t channels, size_t filter_height, - size_t filter_width, size_t stride_height, - size_t stride_width, size_t padding_height, - size_t padding_width, size_t output_height, - size_t output_width, T* data_im) { - size_t index = +__global__ void col2im(int n, const T* data_col, int im_height, int im_width, + int dilation_h, int dilation_w, int filter_height, + int filter_width, int stride_height, int stride_width, + int padding_height, int padding_width, int col_height, + int col_width, T* data_im) { + const int index = (blockIdx.x * gridDim.y + blockIdx.y) * blockDim.x + threadIdx.x; + + const int d_filter_height = dilation_h * (filter_height - 1) + 1; + const int d_filter_width = dilation_w * (filter_width - 1) + 1; + if (index < n) { T val = 0; - int w = int(index % width); - int h = int((index / width) % height); - int c = int(index / (width * height)); - if ((w - (int)padding_width) >= 0 && - (w - (int)padding_width) < (width - 2 * padding_width) && - (h - (int)padding_height) >= 0 && - (h - padding_height) < (height - 2 * padding_height)) { - // compute the start and end of the output - int w_col_start = (w < (int)filter_width) - ? 0 - : (w - int(filter_width)) / (int)stride_width + 1; - int w_col_end = - min((int)(w / (int)stride_width + 1), (int)(output_width)); - int h_col_start = (h < (int)filter_height) - ? 0 - : (h - (int)filter_height) / (int)stride_height + 1; - int h_col_end = min(int(h / stride_height + 1), int(output_height)); - for (int h_col = h_col_start; h_col < h_col_end; ++h_col) { - for (int w_col = w_col_start; w_col < w_col_end; ++w_col) { - // the col location: [c * width * height + h_out, w_out] - int c_col = int(c * filter_height * filter_width) + - (h - h_col * (int)stride_height) * (int)filter_width + - (w - w_col * (int)stride_width); - val += - data_col[(c_col * output_height + h_col) * output_width + w_col]; + int w = index % im_width + padding_width; + int h = (index / im_width) % im_height + padding_height; + int c = index / (im_width * im_height); + + // compute the start and end of the output + int w_col_start = + (w < d_filter_width) ? 0 : (w - d_filter_width) / stride_width + 1; + int w_col_end = min(w / stride_width + 1, col_width); + int h_col_start = + (h < d_filter_height) ? 0 : (h - d_filter_height) / stride_height + 1; + int h_col_end = min(h / stride_height + 1, col_height); + + for (int h_col = h_col_start; h_col < h_col_end; ++h_col) { + for (int w_col = w_col_start; w_col < w_col_end; ++w_col) { + int h_off = (h - h_col * stride_height); + int w_off = (w - w_col * stride_width); + if (h_off % dilation_h == 0 && w_off % dilation_w == 0) { + h_off /= dilation_h; + w_off /= dilation_w; + int data_col_index = + (((c * filter_height + h_off) * filter_width + w_off) * + col_height + + h_col) * + col_width + + w_col; + + val += data_col[data_col_index]; } } - h -= padding_height; - w -= padding_width; - data_im[c * ((width - 2 * padding_width) * - (height - 2 * padding_height)) + - h * (width - 2 * padding_width) + w] += val; } + data_im[index] = val; } } @@ -159,33 +162,38 @@ template class Col2ImFunctor { public: - void operator()(const platform::DeviceContext& context, framework::Tensor& im, - const framework::Tensor& col, int stride_height, - int stride_width, int padding_up, int padding_down, - int padding_left, int padding_right) { - PADDLE_ENFORCE(im.dims().size() == 3); + void operator()(const platform::DeviceContext& context, + const framework::Tensor& col, + const std::vector& dilation, + const std::vector& stride, + const std::vector& padding, framework::Tensor* im) { + PADDLE_ENFORCE(im->dims().size() == 3); PADDLE_ENFORCE(col.dims().size() == 5); - int input_channels = im.dims()[0]; - int input_height = im.dims()[1]; - int input_width = im.dims()[2]; + int im_channels = im->dims()[0]; + int im_height = im->dims()[1]; + int im_width = im->dims()[2]; int filter_height = col.dims()[1]; int filter_width = col.dims()[2]; - int output_height = col.dims()[3]; - int output_width = col.dims()[4]; - - PADDLE_ENFORCE((input_height + padding_up + padding_down - filter_height) / - stride_height + - 1 == - output_height); - PADDLE_ENFORCE((input_width + padding_left + padding_right - filter_width) / - stride_width + - 1 == - output_width); - - size_t num_kernels = input_channels * - (input_height + padding_up + padding_down) * - (input_width + padding_left + padding_right); + int col_height = col.dims()[3]; + int col_width = col.dims()[4]; + + PADDLE_ENFORCE_EQ((im_height + padding[0] + padding[2] - + (dilation[0] * (filter_height - 1) + 1)) / + stride[0] + + 1, + col_height, + "Output_height and padding(padding_up, padding_down) are " + "inconsistent."); + PADDLE_ENFORCE_EQ((im_width + padding[1] + padding[3] - + (dilation[1] * (filter_width - 1) + 1)) / + stride[1] + + 1, + col_width, + "col_width and padding(padding_left, padding_right) are " + "inconsistent."); + + size_t num_kernels = im_channels * im_height * im_width; size_t blocks = (num_kernels + 1024 - 1) / 1024; size_t block_x = 512; @@ -198,10 +206,9 @@ class Col2ImFunctor<<(context) .stream()>>>( - num_kernels, col.data(), input_height + padding_up + padding_down, - input_width + padding_left + padding_left, input_channels, - filter_height, filter_width, stride_height, stride_width, padding_up, - padding_left, output_height, output_width, im.data()); + num_kernels, col.data(), im_height, im_width, dilation[0], + dilation[1], filter_height, filter_width, stride[0], stride[1], + padding[0], padding[2], col_height, col_width, im->data()); } }; @@ -215,33 +222,32 @@ template class Col2ImFunctor; template -__global__ void im2colOCF(const T* im_data, T* col_data, int input_channels, - int input_height, int input_width, int filter_height, - int filter_width, int stride_height, int stride_width, - int padding_height, int padding_width, - int output_height, int output_width) { +__global__ void im2colOCF(const T* im_data, int im_channels, int im_height, + int im_width, int filter_height, int filter_width, + int stride_height, int stride_width, + int padding_height, int padding_width, int col_height, + int col_width, T* col_data) { int swid = blockIdx.x; int shid = blockIdx.y; - for (int channelid = threadIdx.z; channelid < input_channels; + for (int channelid = threadIdx.z; channelid < im_channels; channelid += blockDim.z) { for (int idy = threadIdx.y; idy < filter_height; idy += blockDim.y) { for (int idx = threadIdx.x; idx < filter_width; idx += blockDim.x) { int width_offset = idx + swid * stride_width - padding_width; int height_offset = idy + shid * stride_height - padding_height; - int im_offset = width_offset + height_offset * input_width + - channelid * input_height * input_width; + int im_offset = width_offset + height_offset * im_width + + channelid * im_height * im_width; int col_offset = idx + idy * filter_width + channelid * filter_height * filter_width + - (shid * output_width + swid) * - (input_channels * filter_height * filter_width); - - if (height_offset >= input_height || height_offset < 0 || - width_offset >= input_width || width_offset < 0) { - col_data[col_offset] = T(0); - } else { - col_data[col_offset] = im_data[im_offset]; - } + (shid * col_width + swid) * + (im_channels * filter_height * filter_width); + + col_data[col_offset] = + (height_offset >= im_height || height_offset < 0 || + width_offset >= im_width || width_offset < 0) + ? T(0) + : im_data[im_offset]; } } } @@ -257,27 +263,33 @@ class Im2ColFunctor { public: void operator()(const platform::DeviceContext& context, - const framework::Tensor& im, framework::Tensor& col, - int stride_height, int stride_width, int padding_up, - int padding_down, int padding_left, int padding_right) { + const framework::Tensor& im, const std::vector& dilation, + const std::vector& stride, + const std::vector& padding, framework::Tensor* col) { PADDLE_ENFORCE(im.dims().size() == 3); - PADDLE_ENFORCE(col.dims().size() == 5); - int input_channels = im.dims()[0]; - int input_height = im.dims()[1]; - int input_width = im.dims()[2]; - int filter_height = col.dims()[3]; - int filter_width = col.dims()[4]; - int output_height = col.dims()[0]; - int output_width = col.dims()[1]; - - PADDLE_ENFORCE((input_height + padding_up + padding_down - filter_height) / - stride_height + - 1 == - output_height); - PADDLE_ENFORCE((input_width + padding_left + padding_right - filter_width) / - stride_width + - 1 == - output_width); + PADDLE_ENFORCE(col->dims().size() == 5); + int im_channels = im.dims()[0]; + int im_height = im.dims()[1]; + int im_width = im.dims()[2]; + int filter_height = col->dims()[3]; + int filter_width = col->dims()[4]; + int col_height = col->dims()[0]; + int col_width = col->dims()[1]; + + PADDLE_ENFORCE_EQ((im_height + padding[0] + padding[2] - + (dilation[0] * (filter_height - 1) + 1)) / + stride[0] + + 1, + col_height, + "Output_height and padding(padding_up, padding_down) are " + "inconsistent."); + PADDLE_ENFORCE_EQ((im_width + padding[1] + padding[3] - + (dilation[1] * (filter_width - 1) + 1)) / + stride[1] + + 1, + col_width, + "col_width and padding(padding_left, padding_right) are " + "inconsistent."); int block_dim_x = 0; int block_dim_y = 0; @@ -296,42 +308,41 @@ class Im2ColFunctor<<(context) .stream()>>>( - im.data(), col.data(), input_channels, input_height, input_width, - filter_height, filter_width, stride_height, stride_width, padding_up, - padding_left, output_height, output_width); + im.data(), im_channels, im_height, im_width, filter_height, + filter_width, stride[0], stride[1], padding[0], padding[1], col_height, + col_width, col->data()); } }; template -__global__ void col2imOCF(T* im_data, const T* col_data, int input_channels, - int input_height, int input_width, int filter_height, - int filter_width, int stride_height, int stride_width, - int padding_height, int padding_width, - int output_height, int output_width) { +__global__ void col2imOCF(const T* col_data, int im_channels, int im_height, + int im_width, int filter_height, int filter_width, + int stride_height, int stride_width, + int padding_height, int padding_width, int col_height, + int col_width, T* im_data) { int swid = blockIdx.x; int shid = blockIdx.y; - for (int channelid = threadIdx.z; channelid < input_channels; + for (int channelid = threadIdx.z; channelid < im_channels; channelid += blockDim.z) { for (int idy = threadIdx.y; idy < filter_height; idy += blockDim.y) { for (int idx = threadIdx.x; idx < filter_width; idx += blockDim.x) { int width_offset = idx + swid * stride_width - padding_width; int height_offset = idy + shid * stride_height - padding_height; - int im_offset = width_offset + height_offset * input_width + - channelid * input_height * input_width; + int im_offset = width_offset + height_offset * im_width + + channelid * im_height * im_width; int col_offset = idx + idy * filter_width + channelid * filter_height * filter_width + - (shid * output_width + swid) * - (input_channels * filter_height * filter_width); + (shid * col_width + swid) * + (im_channels * filter_height * filter_width); - if (height_offset >= 0 && height_offset < input_height && - width_offset >= 0 && width_offset < input_width) { + if (height_offset >= 0 && height_offset < im_height && + width_offset >= 0 && width_offset < im_width) { paddle::platform::CudaAtomicAdd(im_data + im_offset, col_data[col_offset]); } @@ -349,28 +360,35 @@ template class Col2ImFunctor { public: - void operator()(const platform::DeviceContext& context, framework::Tensor& im, - const framework::Tensor& col, int stride_height, - int stride_width, int padding_up, int padding_down, - int padding_left, int padding_right) { - PADDLE_ENFORCE(im.dims().size() == 3); + void operator()(const platform::DeviceContext& context, + const framework::Tensor& col, + const std::vector& dilation, + const std::vector& stride, + const std::vector& padding, framework::Tensor* im) { + PADDLE_ENFORCE(im->dims().size() == 3); PADDLE_ENFORCE(col.dims().size() == 5); - int input_channels = im.dims()[0]; - int input_height = im.dims()[1]; - int input_width = im.dims()[2]; + int im_channels = im->dims()[0]; + int im_height = im->dims()[1]; + int im_width = im->dims()[2]; int filter_height = col.dims()[3]; int filter_width = col.dims()[4]; - int output_height = col.dims()[0]; - int output_width = col.dims()[1]; - - PADDLE_ENFORCE((input_height + padding_up + padding_down - filter_height) / - stride_height + - 1 == - output_height); - PADDLE_ENFORCE((input_width + padding_left + padding_right - filter_width) / - stride_width + - 1 == - output_width); + int col_height = col.dims()[0]; + int col_width = col.dims()[1]; + + PADDLE_ENFORCE_EQ((im_height + padding[0] + padding[2] - + (dilation[0] * (filter_height - 1) + 1)) / + stride[0] + + 1, + col_height, + "Output_height and padding(padding_up, padding_down) are " + "inconsistent."); + PADDLE_ENFORCE_EQ((im_width + padding[1] + padding[3] - + (dilation[1] * (filter_width - 1) + 1)) / + stride[1] + + 1, + col_width, + "col_width and padding(padding_left, padding_right) are " + "inconsistent."); int block_dim_x = 0; int block_dim_y = 0; @@ -389,15 +407,14 @@ class Col2ImFunctor<<(context) .stream()>>>( - im.data(), col.data(), input_channels, input_height, input_width, - filter_height, filter_width, stride_height, stride_width, padding_up, - padding_left, output_height, output_width); + col.data(), im_channels, im_height, im_width, filter_height, + filter_width, stride[0], stride[1], padding[0], padding[1], col_height, + col_width, im->data()); } }; diff --git a/paddle/operators/math/im2col.h b/paddle/operators/math/im2col.h index c736d4fa523c2af3e3dd7a11114d7f84021bc5c1..deb60051beef56437cf75f0fa2cef90bbc0a209a 100644 --- a/paddle/operators/math/im2col.h +++ b/paddle/operators/math/im2col.h @@ -35,6 +35,15 @@ enum class ColFormat { kCFO = 0, kOCF = 1 }; * \param colData Column data. * \param colShape The shape of colData. * + * \param dilations dilation data. + * \param 2-dimension [dilation_height, dilation_width]. + * + * \param strides stride data. + * \param 2-dimension [stride_height, stride_width]. + * + * \param paddings padding data. + * \param 4-dimension [up_pad, left_pad, down_pad, right_pad]. + * * If the template argument Format is kCFO, the shape of colData is: * [input_channels, filter_height, filter_width, output_height, output_width] * So, it is easy to reshape into a convolution matrix for convolution @@ -73,18 +82,19 @@ template class Im2ColFunctor { public: void operator()(const platform::DeviceContext& context, - const framework::Tensor& im, framework::Tensor& col, - int stride_height, int stride_width, int padding_up, - int padding_down, int padding_left, int padding_right); + const framework::Tensor& im, const std::vector& dilation, + const std::vector& stride, + const std::vector& padding, framework::Tensor* col); }; template class Col2ImFunctor { public: - void operator()(const platform::DeviceContext& context, framework::Tensor& im, - const framework::Tensor& col, int stride_height, - int stride_width, int padding_up, int padding_down, - int padding_left, int padding_right); + void operator()(const platform::DeviceContext& context, + const framework::Tensor& col, + const std::vector& dilation, + const std::vector& stride, + const std::vector& padding, framework::Tensor* im); }; } // namespace math diff --git a/paddle/operators/math/im2col_test.cc b/paddle/operators/math/im2col_test.cc index 5763782c4edec87f44dabef2ccffe3097eeb2421..10c28da72ba9d3b94bb59c5cf00e7f5a2f28fd06 100644 --- a/paddle/operators/math/im2col_test.cc +++ b/paddle/operators/math/im2col_test.cc @@ -45,10 +45,14 @@ void testIm2col() { int input_height = 2; int input_width = 3; int filter_size = 2; - int stride = 1; - int padding = 0; - int output_height = (input_height - filter_size + 2 * padding) / stride + 1; - int output_width = (input_width - filter_size + 2 * padding) / stride + 1; + std::vector stride({1, 1}); // stride_y, stride_x + std::vector padding( + {0, 0, 0, 0}); // up_pad, left_pad, down_pad, right_pad + std::vector dilation({1, 1}); // dilation_y, dilation_x + int output_height = + (input_height - filter_size + padding[0] + padding[1]) / stride[0] + 1; + int output_width = + (input_width - filter_size + padding[2] + padding[3]) / stride[1] + 1; float* input_ptr = input_tmp.mutable_data( {1, input_height, input_width}, paddle::platform::CPUPlace()); float arr[6] = {0, 1, 2, 3, 4, 5}; @@ -85,10 +89,8 @@ void testIm2col() { paddle::operators::math::ColFormat::kOCF, Place, float> im2col_ocf; - im2col(*context, input, output_cfo, stride, stride, padding, padding, padding, - padding); - im2col_ocf(*context, input, output_ocf, stride, stride, padding, padding, - padding, padding); + im2col(*context, input, dilation, stride, padding, &output_cfo); + im2col_ocf(*context, input, dilation, stride, padding, &output_ocf); float out_cfo_data[] = {0, 1, 1, 2, 3, 4, 4, 5}; float out_ocf_data[] = {0, 1, 3, 4, 1, 2, 4, 5}; @@ -131,8 +133,7 @@ void testIm2col() { input.CopyFrom(input_tmp, *place, *context); } - col2im(*context, input, output_cfo, stride, stride, padding, padding, padding, - padding); + col2im(*context, output_cfo, dilation, stride, padding, &input); float* in_ptr; if (paddle::platform::is_cpu_place(*place)) { @@ -153,8 +154,7 @@ void testIm2col() { input.CopyFrom(input_tmp, *place, *context); } - col2im_ocf(*context, input, output_ocf, stride, stride, padding, padding, - padding, padding); + col2im_ocf(*context, output_ocf, dilation, stride, padding, &input); if (paddle::platform::is_cpu_place(*place)) { in_ptr = input.data(); diff --git a/paddle/operators/math/math_function.cc b/paddle/operators/math/math_function.cc index 1b0d4c8bdc683b5203a4bc4b3838560cffe00bc8..2e333a8cde721f8e65dbf2cf5e3aac6272172cc0 100644 --- a/paddle/operators/math/math_function.cc +++ b/paddle/operators/math/math_function.cc @@ -14,6 +14,7 @@ limitations under the License. */ #include "paddle/operators/math/math_function.h" #include "paddle/framework/data_type.h" +#include "paddle/operators/math/math_function_impl.h" namespace paddle { namespace operators { @@ -232,7 +233,36 @@ void gemv(const platform::DeviceContext& context, cblas_dgemv(CblasRowMajor, transA, M, N, alpha, A, N, B, 1, beta, C, 1); } +template <> +void axpy(const platform::DeviceContext& context, + const int n, const float alpha, + const float* x, float* y) { + cblas_saxpy(n, alpha, x, 1, y, 1); +} + +template <> +void axpy(const platform::DeviceContext& context, + const int n, const double alpha, + const double* x, double* y) { + cblas_daxpy(n, alpha, x, 1, y, 1); +} + template struct SetConstant; +template struct SetConstant; +template struct SetConstant; +template struct SetConstant; +template struct SetConstant; + +#define DEFINE_CPU_TRANS(RANK) \ + template struct Transpose; \ + template struct Transpose; + +DEFINE_CPU_TRANS(1); +DEFINE_CPU_TRANS(2); +DEFINE_CPU_TRANS(3); +DEFINE_CPU_TRANS(4); +DEFINE_CPU_TRANS(5); +DEFINE_CPU_TRANS(6); struct TensorSetConstantCPU { TensorSetConstantCPU(framework::Tensor* tensor, float value) @@ -280,6 +310,11 @@ void set_constant(const platform::DeviceContext& context, #endif } +template struct RowwiseAdd; +template struct RowwiseAdd; +template struct ColwiseSum; +template struct ColwiseSum; + } // namespace math } // namespace operators } // namespace paddle diff --git a/paddle/operators/math/math_function.cu b/paddle/operators/math/math_function.cu index 817deec94314bdfd2ed7e4b0ba5212c72b813455..58356a4b7783241ca0292829bf05dc1a8ed80c6c 100644 --- a/paddle/operators/math/math_function.cu +++ b/paddle/operators/math/math_function.cu @@ -12,8 +12,10 @@ 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/framework/data_type.h" #include "paddle/operators/math/math_function.h" +#include "paddle/operators/math/math_function_impl.h" namespace paddle { namespace operators { @@ -231,11 +233,46 @@ void gemv(const platform::DeviceContext& context, cuTransA, N, M, &alpha, A, N, B, 1, &beta, C, 1)); } +template <> +void axpy(const platform::DeviceContext& context, + const int n, const float alpha, + const float* x, float* y) { + PADDLE_ENFORCE(platform::dynload::cublasSaxpy( + reinterpret_cast(context) + .cublas_handle(), + n, &alpha, x, 1, y, 1)); +} + +template <> +void axpy(const platform::DeviceContext& context, + const int n, const double alpha, + const double* x, double* y) { + PADDLE_ENFORCE(platform::dynload::cublasDaxpy( + reinterpret_cast(context) + .cublas_handle(), + n, &alpha, x, 1, y, 1)); +} + template struct SetConstant; +template struct SetConstant; +template struct SetConstant; +template struct SetConstant; +template struct SetConstant; + +#define DEFINE_GPU_TRANS(RANK) \ + template struct Transpose; \ + template struct Transpose; + +DEFINE_GPU_TRANS(1); +DEFINE_GPU_TRANS(2); +DEFINE_GPU_TRANS(3); +DEFINE_GPU_TRANS(4); +DEFINE_GPU_TRANS(5); +DEFINE_GPU_TRANS(6); struct TensorSetConstantGPU { TensorSetConstantGPU(const platform::DeviceContext& context, - framework::Tensor* tensor, float value) + framework::Tensor* tensor, float value) : context_(context), tensor_(tensor), value_(value) {} template @@ -257,6 +294,11 @@ void set_constant_with_place( TensorSetConstantGPU(context, tensor, value)); } +template struct RowwiseAdd; +template struct RowwiseAdd; +template struct ColwiseSum; +template struct ColwiseSum; + } // namespace math } // namespace operators } // namespace paddle diff --git a/paddle/operators/math/math_function.h b/paddle/operators/math/math_function.h index c2aaa1d7b7e920c3e6fd9ae4424eae725c3b7c0e..ffb99f53808c4316ede96b04e57aec4dae4134de 100644 --- a/paddle/operators/math/math_function.h +++ b/paddle/operators/math/math_function.h @@ -93,14 +93,21 @@ void gemv(const platform::DeviceContext& context, const bool trans_a, const int M, const int N, const T alpha, const T* A, const T* B, const T beta, T* C); +template +void axpy(const platform::DeviceContext& context, const int n, const T alpha, + const T* x, T* y); + +template +struct Transpose { + void operator()(const platform::DeviceContext& context, + const framework::Tensor& in, framework::Tensor* out, + const std::vector& axis); +}; + template struct SetConstant { void operator()(const platform::DeviceContext& context, - framework::Tensor* tensor, T num) { - auto t = framework::EigenVector::Flatten(*tensor); - t.device(*context.GetEigenDevice()) = - t.constant(static_cast(num)); - } + framework::Tensor* tensor, T num); }; template @@ -110,6 +117,19 @@ void set_constant_with_place(const platform::DeviceContext& context, void set_constant(const platform::DeviceContext& context, framework::Tensor* tensor, float value); +template +struct RowwiseAdd { + void operator()(const platform::DeviceContext& context, + const framework::Tensor& input, const framework::Tensor& vec, + framework::Tensor* output); +}; + +template +struct ColwiseSum { + void operator()(const platform::DeviceContext& context, + const framework::Tensor& input, framework::Tensor* vec); +}; + } // namespace math } // namespace operators } // namespace paddle diff --git a/paddle/operators/math/math_function_impl.h b/paddle/operators/math/math_function_impl.h new file mode 100644 index 0000000000000000000000000000000000000000..4dc17a4e525c52b8f696277274a7ad00a6b00a08 --- /dev/null +++ b/paddle/operators/math/math_function_impl.h @@ -0,0 +1,83 @@ +/* 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/data_type.h" +#include "paddle/operators/math/math_function.h" + +namespace paddle { +namespace operators { +namespace math { + +template +void SetConstant::operator()(const platform::DeviceContext& context, + framework::Tensor* tensor, T num) { + auto t = framework::EigenVector::Flatten(*tensor); + t.device(*context.GetEigenDevice()) = t.constant(static_cast(num)); +} + +template +void Transpose::operator()( + const platform::DeviceContext& context, const framework::Tensor& in, + framework::Tensor* out, const std::vector& axis) { + Eigen::array permute; + for (int i = 0; i < Rank; i++) { + permute[i] = axis[i]; + } + auto in_dim = in.dims(); + auto out_dim = out->dims(); + + auto eigen_in = framework::EigenTensor::From(in); + auto eigen_out = framework::EigenTensor::From(*out); + auto* dev = context.GetEigenDevice(); + eigen_out.device(*dev) = eigen_in.shuffle(permute); +} + +template +void RowwiseAdd::operator()(const platform::DeviceContext& context, + const framework::Tensor& input, + const framework::Tensor& vector, + framework::Tensor* output) { + auto in_dims = input.dims(); + auto size = input.numel() / in_dims[0]; + PADDLE_ENFORCE_EQ(vector.numel(), size); + PADDLE_ENFORCE_EQ(output->dims(), in_dims); + + auto in = framework::EigenMatrix::From(input); + auto vec = framework::EigenMatrix::From(vector); + auto out = framework::EigenMatrix::From(*output); + Eigen::array shape({{1, static_cast(size)}}); + Eigen::array bcast({{static_cast(in_dims[0]), 1}}); + out.device(*context.GetEigenDevice()) = + in + vec.reshape(shape).broadcast(bcast); +} + +template +void ColwiseSum::operator()(const platform::DeviceContext& context, + const framework::Tensor& input, + framework::Tensor* vector) { + auto in_dims = input.dims(); + auto size = input.numel() / in_dims[0]; + PADDLE_ENFORCE_EQ(vector->numel(), size); + + auto vec = framework::EigenMatrix::From(*vector); + auto in = framework::EigenMatrix::From(input); + Eigen::array shape({{1, static_cast(size)}}); + vec.reshape(shape).device(*context.GetEigenDevice()) = + in.sum(Eigen::array({{0}})).reshape(shape); +} + +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/math/sequence2batch.cu b/paddle/operators/math/sequence2batch.cu index 8d04653832d58aa048f73e53b8349a08da3145a4..c5d968aeb216bbb3e0e17f138b9e891494d99f75 100644 --- a/paddle/operators/math/sequence2batch.cu +++ b/paddle/operators/math/sequence2batch.cu @@ -12,6 +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. */ +#define EIGEN_USE_GPU #include "paddle/operators/math/sequence2batch.h" namespace paddle { diff --git a/paddle/operators/math/sequence2batch.h b/paddle/operators/math/sequence2batch.h index 794c7d43973924d470124baf8c0c3de66e4ba087..73295ddbcb73fe80be08e732790f0ec75e94b415 100644 --- a/paddle/operators/math/sequence2batch.h +++ b/paddle/operators/math/sequence2batch.h @@ -13,6 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once +#include "paddle/framework/eigen.h" #include "paddle/framework/lod_tensor.h" #include "paddle/framework/tensor.h" #include "paddle/platform/device_context.h" @@ -21,6 +22,10 @@ namespace paddle { namespace operators { namespace math { +template +using EigenMatrix = framework::EigenMatrix; + template class CopyMatrixRowsFunctor { public: diff --git a/paddle/operators/math/softmax.cc b/paddle/operators/math/softmax.cc index 0ba8197ab8b64649c8adcf67771ba01eca7f1d10..3e2f15d6c27f58818128f32fab0bd4c5f36b0050 100644 --- a/paddle/operators/math/softmax.cc +++ b/paddle/operators/math/softmax.cc @@ -13,13 +13,16 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/operators/math/softmax.h" +#include "paddle/operators/math/softmax_impl.h" namespace paddle { namespace operators { namespace math { template class SoftmaxFunctor; +template class SoftmaxFunctor; template class SoftmaxGradFunctor; +template class SoftmaxGradFunctor; } // namespace math } // namespace operators diff --git a/paddle/operators/math/softmax.cu b/paddle/operators/math/softmax.cu index 99f988d51e4b16c3f3bfd9c76b411bb53619603e..4dbab51d46bdaaa506a6c242d0958c73687f4eb9 100644 --- a/paddle/operators/math/softmax.cu +++ b/paddle/operators/math/softmax.cu @@ -15,13 +15,16 @@ limitations under the License. */ #define EIGEN_USE_GPU #include "paddle/operators/math/softmax.h" +#include "paddle/operators/math/softmax_impl.h" namespace paddle { namespace operators { namespace math { template class SoftmaxFunctor; +template class SoftmaxFunctor; template class SoftmaxGradFunctor; +template class SoftmaxGradFunctor; } // namespace math } // namespace operators diff --git a/paddle/operators/math/softmax.h b/paddle/operators/math/softmax.h index b7f627eee7f8fe68a83595a3390a55d438c97afb..fe1074650234c5beb5889e7efd713164769ad740 100644 --- a/paddle/operators/math/softmax.h +++ b/paddle/operators/math/softmax.h @@ -13,60 +13,17 @@ 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)); - } + const framework::Tensor* X, framework::Tensor* Y); }; template @@ -74,29 +31,7 @@ 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; - } + framework::Tensor* x_grad); }; } // namespace math diff --git a/paddle/operators/math/softmax_impl.h b/paddle/operators/math/softmax_impl.h new file mode 100644 index 0000000000000000000000000000000000000000..05793eeb3eeafaf36c301236197555b7b35e5803 --- /dev/null +++ b/paddle/operators/math/softmax_impl.h @@ -0,0 +1,98 @@ +/* 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" + +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 +void SoftmaxFunctor::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 +void SoftmaxGradFunctor::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/math/vol2col.cc b/paddle/operators/math/vol2col.cc index e9718a047381596a1570b4b00546622968b70227..99eb7fd46de42400a915d86706580d15b08a74a2 100644 --- a/paddle/operators/math/vol2col.cc +++ b/paddle/operators/math/vol2col.cc @@ -28,28 +28,51 @@ template class Vol2ColFunctor { public: void operator()(const platform::DeviceContext& context, - const framework::Tensor& vol, framework::Tensor& col, - int stride_depth, int stride_height, int stride_width, - int padding_depth, int padding_height, - int padding_width) const { + const framework::Tensor& vol, + const std::vector& dilations, + const std::vector& strides, + const std::vector& paddings, + framework::Tensor* col) const { PADDLE_ENFORCE(vol.dims().size() == 4); - PADDLE_ENFORCE(col.dims().size() == 7); + PADDLE_ENFORCE(col->dims().size() == 7); int input_channels = vol.dims()[0]; int input_depth = vol.dims()[1]; int input_height = vol.dims()[2]; int input_width = vol.dims()[3]; - int filter_depth = col.dims()[1]; - int filter_height = col.dims()[2]; - int filter_width = col.dims()[3]; - int output_depth = col.dims()[4]; - int output_height = col.dims()[5]; - int output_width = col.dims()[6]; + int filter_depth = col->dims()[1]; + int filter_height = col->dims()[2]; + int filter_width = col->dims()[3]; + int output_depth = col->dims()[4]; + int output_height = col->dims()[5]; + int output_width = col->dims()[6]; int channels_col = input_channels * filter_depth * filter_height * filter_width; + PADDLE_ENFORCE_EQ((input_depth + 2 * paddings[0] - + ((dilations[0] * (filter_depth - 1) + 1))) / + strides[0] + + 1, + output_depth, + "input_depth and output_depth are " + "mismatching."); + PADDLE_ENFORCE_EQ((input_height + 2 * paddings[1] - + ((dilations[1] * (filter_height - 1) + 1))) / + strides[1] + + 1, + output_height, + "input_height and output_height are " + "mismatching."); + PADDLE_ENFORCE_EQ((input_width + 2 * paddings[2] - + ((dilations[2] * (filter_width - 1) + 1))) / + strides[2] + + 1, + output_width, + "input_width and output_width are " + "mismatching."); + const T* vol_data = vol.data(); - T* col_data = col.data(); + T* col_data = col->data(); for (int c = 0; c < channels_col; ++c) { int w_offset = c % filter_width; @@ -57,24 +80,23 @@ class Vol2ColFunctor { int d_offset = (c / filter_width / filter_height) % filter_depth; int c_in = c / filter_width / filter_height / filter_depth; for (int d = 0; d < output_depth; ++d) { - int d_pad = d * stride_depth - padding_depth + d_offset; + int d_pad = d * strides[0] - paddings[0] + d_offset * dilations[0]; for (int h = 0; h < output_height; ++h) { - int h_pad = h * stride_height - padding_height + h_offset; + int h_pad = h * strides[1] - paddings[1] + h_offset * dilations[1]; for (int w = 0; w < output_width; ++w) { - int w_pad = w * stride_width - padding_width + w_offset; + int w_pad = w * strides[2] - paddings[2] + w_offset * dilations[2]; int col_idx = ((c * output_depth + d) * output_height + h) * output_width + w; - if (h_pad < 0 || h_pad >= input_height || w_pad < 0 || - w_pad >= input_width || d_pad < 0 || d_pad >= input_depth) { - col_data[col_idx] = static_cast(0); - } else { - int vol_idx = - ((c_in * input_depth + d_pad) * input_height + h_pad) * - input_width + - w_pad; - col_data[col_idx] = vol_data[vol_idx]; - } + int vol_idx = + ((c_in * input_depth + d_pad) * input_height + h_pad) * + input_width + + w_pad; + col_data[col_idx] = + (h_pad < 0 || h_pad >= input_height || w_pad < 0 || + w_pad >= input_width || d_pad < 0 || d_pad >= input_depth) + ? static_cast(0) + : vol_data[vol_idx]; } } } @@ -92,17 +114,18 @@ template class Col2VolFunctor { public: void operator()(const platform::DeviceContext& context, - framework::Tensor& vol, const framework::Tensor& col, - int stride_depth, int stride_height, int stride_width, - int padding_depth, int padding_height, - int padding_width) const { - PADDLE_ENFORCE(vol.dims().size() == 4); + const framework::Tensor& col, + const std::vector& dilations, + const std::vector& strides, + const std::vector& paddings, + framework::Tensor* vol) const { + PADDLE_ENFORCE(vol->dims().size() == 4); PADDLE_ENFORCE(col.dims().size() == 7); - int input_channels = vol.dims()[0]; - int input_depth = vol.dims()[1]; - int input_height = vol.dims()[2]; - int input_width = vol.dims()[3]; + int input_channels = vol->dims()[0]; + int input_depth = vol->dims()[1]; + int input_height = vol->dims()[2]; + int input_width = vol->dims()[3]; int filter_depth = col.dims()[1]; int filter_height = col.dims()[2]; int filter_width = col.dims()[3]; @@ -112,7 +135,28 @@ class Col2VolFunctor { int channels_col = input_channels * filter_depth * filter_height * filter_width; - T* vol_data = vol.data(); + PADDLE_ENFORCE_EQ((input_depth + 2 * paddings[0] - + ((dilations[0] * (filter_depth - 1) + 1))) / + strides[0] + + 1, + output_depth, + "input_depth and output_depth are " + "mismatching."); + PADDLE_ENFORCE_EQ((input_height + 2 * paddings[1] - + ((dilations[1] * (filter_height - 1) + 1))) / + strides[1] + + 1, + output_height, + "input_height and output_height are " + "mismatching."); + PADDLE_ENFORCE_EQ((input_width + 2 * paddings[2] - + ((dilations[2] * (filter_width - 1) + 1))) / + strides[2] + + 1, + output_width, + "input_width and output_width are " + "mismatching."); + T* vol_data = vol->data(); const T* col_data = col.data(); for (int c = 0; c < channels_col; ++c) { @@ -121,11 +165,11 @@ class Col2VolFunctor { int d_offset = (c / filter_width / filter_height) % filter_depth; int cIm = c / filter_width / filter_height / filter_depth; for (int d = 0; d < output_depth; ++d) { - int d_pad = d * stride_depth - padding_depth + d_offset; + int d_pad = d * strides[0] - paddings[0] + d_offset * dilations[0]; for (int h = 0; h < output_height; ++h) { - int h_pad = h * stride_height - padding_height + h_offset; + int h_pad = h * strides[1] - paddings[1] + h_offset * dilations[1]; for (int w = 0; w < output_width; ++w) { - int w_pad = w * stride_width - padding_width + w_offset; + int w_pad = w * strides[2] - paddings[2] + w_offset * dilations[2]; if (h_pad >= 0 && h_pad < input_height && w_pad >= 0 && w_pad < input_width && d_pad >= 0 && d_pad < input_depth) { @@ -133,6 +177,7 @@ class Col2VolFunctor { ((cIm * input_depth + d_pad) * input_height + h_pad) * input_width + w_pad; + int col_idx = ((c * output_depth + d) * output_height + h) * output_width + w; diff --git a/paddle/operators/math/vol2col.cu b/paddle/operators/math/vol2col.cu index 27b11fb237575fd25a789a5fcc24ed4e30607009..dae3be858e9f47d0133aa37e8a5f90a0addf1dfd 100644 --- a/paddle/operators/math/vol2col.cu +++ b/paddle/operators/math/vol2col.cu @@ -21,11 +21,12 @@ namespace math { template __global__ void vol2col(int num_kernels, const T* data_vol, int depth, - int height, int width, int filter_depth, - int filter_height, int filter_width, int stride_depth, - int stride_height, int stride_width, int padding_depth, - int padding_height, int padding_width, int output_detph, - int output_height, int output_width, T* data_col) { + int height, int width, int dilation_d, int dilation_h, + int dilation_w, int filter_depth, int filter_height, + int filter_width, int stride_depth, int stride_height, + int stride_width, int padding_depth, int padding_height, + int padding_width, int output_detph, int output_height, + int output_width, T* data_col) { for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < num_kernels; index += blockDim.x * gridDim.x) { int w_out = index % output_width; @@ -44,12 +45,14 @@ __global__ void vol2col(int num_kernels, const T* data_vol, int depth, for (int k = 0; k < filter_depth; ++k) { for (int i = 0; i < filter_height; ++i) { for (int j = 0; j < filter_width; ++j) { - int d = d_in + k; - int h = h_in + i; - int w = w_in + j; + int d = d_in + k * dilation_d; + int h = h_in + i * dilation_h; + int w = w_in + j * dilation_w; + int col_idx = (k * dilation_d * height + i * dilation_h) * width + + j * dilation_w; *data_col = (d >= 0 && d < depth && h >= 0 && h < height && w >= 0 && w < width) - ? data_vol[(k * height + i) * width + j] + ? data_vol[col_idx] : 0; data_col += output_detph * output_height * output_width; } @@ -68,23 +71,46 @@ template class Vol2ColFunctor { public: void operator()(const platform::DeviceContext& context, - const framework::Tensor& vol, framework::Tensor& col, - int stride_depth, int stride_height, int stride_width, - int padding_depth, int padding_height, - int padding_width) const { + const framework::Tensor& vol, + const std::vector& dilations, + const std::vector& strides, + const std::vector& paddings, + framework::Tensor* col) const { PADDLE_ENFORCE(vol.dims().size() == 4); - PADDLE_ENFORCE(col.dims().size() == 7); + PADDLE_ENFORCE(col->dims().size() == 7); int input_channels = vol.dims()[0]; int input_depth = vol.dims()[1]; int input_height = vol.dims()[2]; int input_width = vol.dims()[3]; - int filter_depth = col.dims()[1]; - int filter_height = col.dims()[2]; - int filter_width = col.dims()[3]; - int output_depth = col.dims()[4]; - int output_height = col.dims()[5]; - int output_width = col.dims()[6]; + int filter_depth = col->dims()[1]; + int filter_height = col->dims()[2]; + int filter_width = col->dims()[3]; + int output_depth = col->dims()[4]; + int output_height = col->dims()[5]; + int output_width = col->dims()[6]; + + PADDLE_ENFORCE_EQ((input_depth + 2 * paddings[0] - + ((dilations[0] * (filter_depth - 1) + 1))) / + strides[0] + + 1, + output_depth, + "input_depth and output_depth are " + "Mismatching."); + PADDLE_ENFORCE_EQ((input_height + 2 * paddings[1] - + ((dilations[1] * (filter_height - 1) + 1))) / + strides[1] + + 1, + output_height, + "input_height and output_height are " + "Mismatching."); + PADDLE_ENFORCE_EQ((input_width + 2 * paddings[2] - + ((dilations[2] * (filter_width - 1) + 1))) / + strides[2] + + 1, + output_width, + "input_width and output_width are " + "Mismatching."); int num_outputs = input_channels * output_depth * output_height * output_width; @@ -95,19 +121,25 @@ class Vol2ColFunctor { reinterpret_cast(context) .stream()>>>( num_outputs, vol.data(), input_depth, input_height, input_width, - filter_depth, filter_height, filter_width, stride_depth, stride_height, - stride_width, padding_depth, padding_height, padding_width, - output_depth, output_height, output_width, col.data()); + dilations[0], dilations[1], dilations[2], filter_depth, filter_height, + filter_width, strides[0], strides[1], strides[2], paddings[0], + paddings[1], paddings[2], output_depth, output_height, output_width, + col->data()); } }; template __global__ void col2vol(int num_kernels, const T* data_col, int depth, - int height, int width, int filter_depth, - int filter_height, int filter_width, int stride_depth, - int stride_height, int stride_width, int padding_depth, - int padding_height, int padding_width, int output_detph, - int output_height, int output_width, T* data_vol) { + int height, int width, int dilation_d, int dilation_h, + int dilation_w, int filter_depth, int filter_height, + int filter_width, int stride_depth, int stride_height, + int stride_width, int padding_depth, int padding_height, + int padding_width, int output_detph, int output_height, + int output_width, T* data_vol) { + const int d_filter_depth = dilation_d * (filter_depth - 1) + 1; + const int d_filter_height = dilation_h * (filter_height - 1) + 1; + const int d_filter_width = dilation_w * (filter_width - 1) + 1; + for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < num_kernels; index += blockDim.x * gridDim.x) { T src_val = 0; @@ -115,35 +147,41 @@ __global__ void col2vol(int num_kernels, const T* data_col, int depth, int h = (index / width) % height + padding_height; int d = (index / width / height) % depth + padding_depth; int c = index / width / height / depth; + // compute the start and end of the output int w_col_start = - (w < filter_width) ? 0 : (w - filter_width) / stride_width + 1; + (w < d_filter_width) ? 0 : (w - d_filter_width) / stride_width + 1; int w_col_end = min(w / stride_width + 1, output_width); int h_col_start = - (h < filter_height) ? 0 : (h - filter_height) / stride_height + 1; + (h < d_filter_height) ? 0 : (h - d_filter_height) / stride_height + 1; int h_col_end = min(h / stride_height + 1, output_height); int d_col_start = - (d < filter_depth) ? 0 : (d - filter_depth) / stride_depth + 1; + (d < d_filter_depth) ? 0 : (d - d_filter_depth) / stride_depth + 1; int d_col_end = min(d / stride_depth + 1, output_detph); - int offset = (c * filter_depth * filter_height * filter_width + - d * filter_width * filter_height + h * filter_width + w) * - output_detph * output_height * output_width; - - int coeff_d_col = - (1 - stride_depth * filter_width * filter_height * output_detph) * - output_height * output_width; - int coeff_h_col = - (1 - stride_height * filter_width * output_detph * output_height) * - output_width; - int coeff_w_col = - (1 - stride_width * output_detph * output_height * output_width); - for (int d_col = d_col_start; d_col < d_col_end; ++d_col) { for (int h_col = h_col_start; h_col < h_col_end; ++h_col) { for (int w_col = w_col_start; w_col < w_col_end; ++w_col) { - src_val += data_col[offset + d_col * coeff_d_col + - h_col * coeff_h_col + w_col * coeff_w_col]; + int d_off = (d - d_col * stride_depth); + int h_off = (h - h_col * stride_height); + int w_off = (w - w_col * stride_width); + if (d_off % dilation_d == 0 && h_off % dilation_h == 0 && + w_off % dilation_w == 0) { + d_off /= dilation_d; + h_off /= dilation_h; + w_off /= dilation_w; + + int data_col_index = + (((((c * filter_depth + d_off) * filter_height + h_off) * + filter_width + + w_off))); + data_col_index = + ((data_col_index * output_detph + d_col) * output_height + + h_col) * + output_width + + w_col; + src_val += data_col[data_col_index]; + } } } } @@ -161,17 +199,18 @@ template class Col2VolFunctor { public: void operator()(const platform::DeviceContext& context, - framework::Tensor& vol, const framework::Tensor& col, - int stride_depth, int stride_height, int stride_width, - int padding_depth, int padding_height, - int padding_width) const { - PADDLE_ENFORCE(vol.dims().size() == 4); + const framework::Tensor& col, + const std::vector& dilations, + const std::vector& strides, + const std::vector& paddings, + framework::Tensor* vol) const { + PADDLE_ENFORCE(vol->dims().size() == 4); PADDLE_ENFORCE(col.dims().size() == 7); - int input_channels = vol.dims()[0]; - int input_depth = vol.dims()[1]; - int input_height = vol.dims()[2]; - int input_width = vol.dims()[3]; + int input_channels = vol->dims()[0]; + int input_depth = vol->dims()[1]; + int input_height = vol->dims()[2]; + int input_width = vol->dims()[3]; int filter_depth = col.dims()[1]; int filter_height = col.dims()[2]; int filter_width = col.dims()[3]; @@ -179,6 +218,28 @@ class Col2VolFunctor { int output_height = col.dims()[5]; int output_width = col.dims()[6]; + PADDLE_ENFORCE_EQ((input_depth + 2 * paddings[0] - + ((dilations[0] * (filter_depth - 1) + 1))) / + strides[0] + + 1, + output_depth, + "input_depth and output_depth are " + "Mismatching."); + PADDLE_ENFORCE_EQ((input_height + 2 * paddings[1] - + ((dilations[1] * (filter_height - 1) + 1))) / + strides[1] + + 1, + output_height, + "input_height and output_height are " + "Mismatching."); + PADDLE_ENFORCE_EQ((input_width + 2 * paddings[2] - + ((dilations[2] * (filter_width - 1) + 1))) / + strides[2] + + 1, + output_width, + "input_width and output_width are " + "Mismatching."); + int num_kernels = input_channels * input_depth * input_height * input_width; const int threads = 1024; @@ -188,9 +249,10 @@ class Col2VolFunctor { reinterpret_cast(context) .stream()>>>( num_kernels, col.data(), input_depth, input_height, input_width, - filter_depth, filter_height, filter_width, stride_depth, stride_height, - stride_width, padding_depth, padding_height, padding_width, - output_depth, output_height, output_width, vol.data()); + dilations[0], dilations[1], dilations[2], filter_depth, filter_height, + filter_width, strides[0], strides[1], strides[2], paddings[0], + paddings[1], paddings[2], output_depth, output_height, output_width, + vol->data()); } }; diff --git a/paddle/operators/math/vol2col.h b/paddle/operators/math/vol2col.h index f022365a16fbf61981e94bedbd8b21a32887b235..cbc30bd754608dd6e6def1a4097d69bdf0c942c3 100644 --- a/paddle/operators/math/vol2col.h +++ b/paddle/operators/math/vol2col.h @@ -31,6 +31,15 @@ namespace math { * \param colData Column data. * \param colShape The shape of colData. * + * \param dilations dilation data. + * \param 3-dimension [dilation_depth, dilation_height, dilation_width]. + * + * \param strides stride data. + * \param 3-dimension [stride_depth, stride_height, stride_width]. + * + * \param paddings padding data. + * \param 3-dimension [d_pad, h_pad, w_pad]. + * * The shape of colData is: * [input_channels, filter_depth, filter_height, filter_width, output_depth, * output_height, output_width] @@ -57,20 +66,22 @@ template class Vol2ColFunctor { public: void operator()(const platform::DeviceContext& context, - const framework::Tensor& vol, framework::Tensor& col, - int stride_depth, int stride_height, int stride_width, - int padding_depth, int padding_height, - int padding_width) const; + const framework::Tensor& vol, + const std::vector& dilations, + const std::vector& strides, + const std::vector& paddings, + framework::Tensor* col) const; }; template class Col2VolFunctor { public: void operator()(const platform::DeviceContext& context, - framework::Tensor& vol, const framework::Tensor& col, - int stride_depth, int stride_height, int stride_width, - int padding_depth, int padding_height, - int padding_width) const; + const framework::Tensor& col, + const std::vector& dilations, + const std::vector& strides, + const std::vector& paddings, + framework::Tensor* vol) const; }; } // namespace math diff --git a/paddle/operators/math/vol2col_test.cc b/paddle/operators/math/vol2col_test.cc index 74590d17cd0f974f830e760d85daef8ab5318a43..c31c716842f30de67c29b803866b8c82ddcf4a41 100644 --- a/paddle/operators/math/vol2col_test.cc +++ b/paddle/operators/math/vol2col_test.cc @@ -62,11 +62,15 @@ void testVol2col() { int input_height = 2; int input_width = 3; int filter_size = 2; - int stride = 1; - int padding = 0; - int output_depth = (input_depth - filter_size + 2 * padding) / stride + 1; - int output_height = (input_height - filter_size + 2 * padding) / stride + 1; - int output_width = (input_width - filter_size + 2 * padding) / stride + 1; + std::vector strides({1, 1, 1}); + std::vector paddings({0, 0, 0}); + std::vector dilations({1, 1, 1}); + int output_depth = + (input_depth - filter_size + 2 * paddings[0]) / strides[0] + 1; + int output_height = + (input_height - filter_size + 2 * paddings[1]) / strides[1] + 1; + int output_width = + (input_width - filter_size + 2 * paddings[2]) / strides[2] + 1; // Vol2Col test float* input_ptr = @@ -85,8 +89,7 @@ void testVol2col() { *place); paddle::operators::math::Vol2ColFunctor vol2col; - vol2col(*context, input, output, stride, stride, stride, padding, padding, - padding); + vol2col(*context, input, dilations, strides, paddings, &output); float vol_2_col[] = {0, 1, 1, 2, 3, 4, 4, 5, 6, 7, 7, 8, 9, 10, 10, 11}; float* out_cfo_ptr; @@ -111,8 +114,7 @@ void testVol2col() { } paddle::operators::math::Col2VolFunctor col2vol; - col2vol(*context, input, output, stride, stride, stride, padding, padding, - padding); + col2vol(*context, output, dilations, strides, paddings, &input); float* in_ptr; if (paddle::platform::is_cpu_place(*place)) { diff --git a/paddle/operators/matmul_op.cu b/paddle/operators/matmul_op.cu.cc similarity index 100% rename from paddle/operators/matmul_op.cu rename to paddle/operators/matmul_op.cu.cc diff --git a/paddle/operators/matmul_op.h b/paddle/operators/matmul_op.h index 4f565946d596b5e5fbf90f16c0c13c780c36886c..1e4aa48b7018d8e3d6f02591fbca2877ddbd3c5d 100644 --- a/paddle/operators/matmul_op.h +++ b/paddle/operators/matmul_op.h @@ -15,8 +15,8 @@ #pragma once #include "paddle/framework/op_registry.h" +#include "paddle/operators/math/math_function.h" #include "paddle/operators/math/matmul.h" -#include "paddle/operators/transpose_op.h" namespace paddle { namespace operators { @@ -76,7 +76,10 @@ Tensor CombineBatchAndN(const framework::ExecutionContext& context, if (in_dims.size() == 3) { output.Resize({in_dims[1], in_dims[0], in_dims[2]}); output.mutable_data(context.GetPlace()); - EigenTranspose(context, input, output, {1, 0, 2}); + std::vector axis = {1, 0, 2}; + math::Transpose trans; + trans(context.device_context(), input, &output, axis); + std::vector out_dims = {in_dims[1], in_dims[0] * in_dims[2]}; output.Resize({in_dims[1], in_dims[0] * in_dims[2]}); } else { output.ShareDataWith(input); diff --git a/paddle/operators/mul_op.cu b/paddle/operators/mul_op.cu.cc similarity index 100% rename from paddle/operators/mul_op.cu rename to paddle/operators/mul_op.cu.cc diff --git a/paddle/operators/nccl_op.cu b/paddle/operators/nccl_op.cu.cc similarity index 100% rename from paddle/operators/nccl_op.cu rename to paddle/operators/nccl_op.cu.cc diff --git a/paddle/operators/nccl_op_test.cu b/paddle/operators/nccl_op_test.cu.cc similarity index 100% rename from paddle/operators/nccl_op_test.cu rename to paddle/operators/nccl_op_test.cu.cc diff --git a/paddle/operators/pool_cudnn_op.cu b/paddle/operators/pool_cudnn_op.cu.cc similarity index 100% rename from paddle/operators/pool_cudnn_op.cu rename to paddle/operators/pool_cudnn_op.cu.cc diff --git a/paddle/operators/pool_op.cu b/paddle/operators/pool_op.cu.cc similarity index 100% rename from paddle/operators/pool_op.cu rename to paddle/operators/pool_op.cu.cc diff --git a/paddle/operators/pool_with_index_op.cu b/paddle/operators/pool_with_index_op.cu.cc similarity index 100% rename from paddle/operators/pool_with_index_op.cu rename to paddle/operators/pool_with_index_op.cu.cc diff --git a/paddle/operators/pool_with_index_op.h b/paddle/operators/pool_with_index_op.h index c0e3b117dc3ea351b9edfed4d1823de0db27d30a..a081607edce335f0265388ab01238d584bcf3ead 100644 --- a/paddle/operators/pool_with_index_op.h +++ b/paddle/operators/pool_with_index_op.h @@ -81,22 +81,21 @@ class MaxPoolWithIndexGradKernel : public framework::OpKernel { 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)); + auto& device_ctx = context.device_context(); + math::set_constant(device_ctx, in_x_grad, 0); switch (ksize.size()) { case 2: { paddle::operators::math::MaxPool2dWithIndexGradFunctor pool2d_backward; - pool2d_backward(context.device_context(), *out_grad, *mask, ksize, - strides, paddings, in_x_grad); + pool2d_backward(device_ctx, *out_grad, *mask, ksize, strides, + paddings, in_x_grad); } break; case 3: { paddle::operators::math::MaxPool3dWithIndexGradFunctor pool3d_backward; - pool3d_backward(context.device_context(), *out_grad, *mask, ksize, - strides, paddings, in_x_grad); + pool3d_backward(device_ctx, *out_grad, *mask, ksize, strides, + paddings, in_x_grad); } break; default: { PADDLE_THROW("Pool op only supports 2D and 3D input."); } } diff --git a/paddle/operators/reshape_op.cu b/paddle/operators/reshape_op.cu.cc similarity index 100% rename from paddle/operators/reshape_op.cu rename to paddle/operators/reshape_op.cu.cc diff --git a/paddle/operators/sequence_concat_op.cu b/paddle/operators/sequence_concat_op.cu.cc similarity index 100% rename from paddle/operators/sequence_concat_op.cu rename to paddle/operators/sequence_concat_op.cu.cc diff --git a/paddle/operators/sequence_conv_op.cu b/paddle/operators/sequence_conv_op.cu.cc similarity index 97% rename from paddle/operators/sequence_conv_op.cu rename to paddle/operators/sequence_conv_op.cu.cc index 4c0c673a517c4b05c3abd8bf6b5cf5bbb19cfae0..6106b0e46c0ab96e01dfc344055f23dbf4a1a2c3 100644 --- a/paddle/operators/sequence_conv_op.cu +++ b/paddle/operators/sequence_conv_op.cu.cc @@ -12,8 +12,6 @@ See the License for the specific language governing permissions and limitations under the License. */ -#define EIGEN_USE_GPU - #include "paddle/operators/sequence_conv_op.h" namespace ops = paddle::operators; diff --git a/paddle/operators/sequence_conv_op.h b/paddle/operators/sequence_conv_op.h index a57e1752bb8ed4844423f752bf0ad9f8e114486a..b8fbe2647c4338a2fa16aa655ebab64dd8d5417d 100644 --- a/paddle/operators/sequence_conv_op.h +++ b/paddle/operators/sequence_conv_op.h @@ -13,7 +13,6 @@ 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/context_project.h" #include "paddle/operators/math/math_function.h" @@ -62,9 +61,9 @@ class SequenceConvKernel : public framework::OpKernel { math::ContextProjectFunctor seq_project_functor; - seq_project_functor(context.device_context(), *in, *padding_data, col, + seq_project_functor(context.device_context(), *in, *padding_data, padding_trainable, context_start, context_length, - context_stride, up_pad, down_pad); + context_stride, up_pad, down_pad, &col); math::matmul(context.device_context(), col, false, filter, false, static_cast(1.0), out, static_cast(0.0)); @@ -117,10 +116,10 @@ class SequenceConvGradKernel : public framework::OpKernel { in_g->set_lod(in->lod()); set_zero(context.device_context(), in_g, static_cast(0)); - seq_project_grad_functor(context.device_context(), *in_g, *padding_data_g, - col, padding_trainable, context_start, - context_length, context_stride, up_pad, down_pad, - true, false); + seq_project_grad_functor(context.device_context(), *in_g, + padding_trainable, context_start, context_length, + context_stride, up_pad, down_pad, false, true, + padding_data_g, &col); } if (padding_trainable && padding_data_g) { @@ -129,9 +128,9 @@ class SequenceConvGradKernel : public framework::OpKernel { LoDTensor* input = const_cast(in); seq_project_grad_functor(context.device_context(), *input, - *padding_data_g, col, padding_trainable, - context_start, context_length, context_stride, - up_pad, down_pad, false, true); + padding_trainable, context_start, context_length, + context_stride, up_pad, down_pad, true, false, + padding_data_g, &col); } if (filter_g) { @@ -146,9 +145,9 @@ class SequenceConvGradKernel : public framework::OpKernel { padding_data = context.Input("PaddingData"); } - seq_project_functor(context.device_context(), *in, *padding_data, col, + seq_project_functor(context.device_context(), *in, *padding_data, padding_trainable, context_start, context_length, - context_stride, up_pad, down_pad); + context_stride, up_pad, down_pad, &col); math::matmul(context.device_context(), col, true, out_grad, false, T(1.0), &filter_grad, T(1.0)); diff --git a/paddle/operators/sequence_softmax_op.cu b/paddle/operators/sequence_softmax_op.cu.cc similarity index 100% rename from paddle/operators/sequence_softmax_op.cu rename to paddle/operators/sequence_softmax_op.cu.cc diff --git a/paddle/operators/sgd_op.cu b/paddle/operators/sgd_op.cu index 2f41c7fc121950926f6e8d842eb629d59738f321..7b6c5ec30628b521b594ceaa3b7f1e0e03e497e4 100644 --- a/paddle/operators/sgd_op.cu +++ b/paddle/operators/sgd_op.cu @@ -20,11 +20,11 @@ namespace paddle { namespace operators { namespace { -template +template __global__ void SparseSGDFunctorKernel(const T* selected_rows, const int64_t* rows, const T* learning_rate, T* tensor_out, - int64_t row_numel, int block_size) { + int64_t row_numel) { const int ty = blockIdx.y; int tid = threadIdx.x; @@ -59,14 +59,15 @@ struct SparseSGDFunctor { auto* in_data = in_value.data(); auto* out_data = output->data(); - int block_size = 256; + const int block_size = 256; dim3 threads(block_size, 1); dim3 grid(1, in_rows.size()); SparseSGDFunctorKernel< - T><<(context) - .stream()>>>(in_data, in_rows.data(), learning_rate.data(), - out_data, in_row_numel, block_size); + T, 256><<(context) + .stream()>>>(in_data, in_rows.data(), + learning_rate.data(), out_data, + in_row_numel); } }; diff --git a/paddle/operators/softmax_op.cu b/paddle/operators/softmax_op.cu.cc similarity index 100% rename from paddle/operators/softmax_op.cu rename to paddle/operators/softmax_op.cu.cc diff --git a/paddle/operators/softmax_with_cross_entropy_op.cc b/paddle/operators/softmax_with_cross_entropy_op.cc index ed96e8cee5a78e63ea29ed383d06c1258abdc328..3dbb62d2e571eb92025c1b3fc0a6653c7cda007a 100644 --- a/paddle/operators/softmax_with_cross_entropy_op.cc +++ b/paddle/operators/softmax_with_cross_entropy_op.cc @@ -14,7 +14,6 @@ limitations under the License. */ #include "paddle/operators/softmax_with_cross_entropy_op.h" #include -#include namespace paddle { namespace operators { diff --git a/paddle/operators/split_op.cu b/paddle/operators/split_op.cu.cc similarity index 100% rename from paddle/operators/split_op.cu rename to paddle/operators/split_op.cu.cc diff --git a/paddle/operators/sum_op.cc b/paddle/operators/sum_op.cc index 57b99bdb3a9359bbfdbe62a6fc9afca6c4d5df9e..c2b7632b2865a3ef66051d815d7722a08c6a8cbd 100644 --- a/paddle/operators/sum_op.cc +++ b/paddle/operators/sum_op.cc @@ -12,7 +12,7 @@ limitations under the License. */ #include "paddle/operators/sum_op.h" #include #include "paddle/framework/var_type_inference.h" -#include "paddle/operators/net_op.h" +#include "paddle/operators/detail/safe_ref.h" namespace paddle { namespace operators { @@ -60,13 +60,16 @@ class SumOp : public framework::OperatorWithKernel { x_vars[0]->Get().value().type()), ctx.device_context()); } else if (x_vars[0]->IsType()) { - auto& array = x_vars[0]->Get(); - for (auto& each : array) { - if (each.numel() != 0) { - return framework::OpKernelType(framework::ToDataType(each.type()), - ctx.device_context()); + for (auto& x_var : x_vars) { + auto& array = x_var->Get(); + for (auto& each : array) { + if (each.numel() != 0) { + return framework::OpKernelType(framework::ToDataType(each.type()), + ctx.device_context()); + } } } + PADDLE_THROW("Cannot find the input data type by all input data"); } PADDLE_THROW("Unexpected branch. Input type is %s", x_vars[0]->Type().name()); @@ -97,6 +100,11 @@ class SumOpVarTypeInference : public framework::VarTypeInference { auto& inputs = op_desc.Input("X"); auto var_type = framework::VarDesc::SELECTED_ROWS; + for (auto& name : op_desc.Input("X")) { + VLOG(10) << name << " " + << block->FindRecursiveOrCreateVar(name)->GetType(); + } + bool any_input_is_lod_tensor = std::any_of( inputs.begin(), inputs.end(), [block](const std::string& name) { return block->FindRecursiveOrCreateVar(name)->GetType() == @@ -104,7 +112,7 @@ class SumOpVarTypeInference : public framework::VarTypeInference { }); auto is_tensor_array = [block](const std::string& name) { - return block->FindRecursiveOrCreateVar(name)->GetType() == + return detail::Ref(block->FindRecursiveOrCreateVar(name)).GetType() == framework::VarDesc::LOD_TENSOR_ARRAY; }; @@ -114,14 +122,26 @@ class SumOpVarTypeInference : public framework::VarTypeInference { std::all_of(inputs.begin(), inputs.end(), is_tensor_array); if (any_input_is_tensor_array) { - PADDLE_ENFORCE(all_inputs_are_tensor_array); + if (!all_inputs_are_tensor_array) { + std::ostringstream os; + for (auto& each : inputs) { + os << " " << each << " type is " + << detail::Ref(block->FindRecursiveOrCreateVar(each)).GetType() + << "\n"; + } + PADDLE_ENFORCE(all_inputs_are_tensor_array, + "Not all inputs are tensor array:\n%s", os.str()); + } var_type = framework::VarDesc::LOD_TENSOR_ARRAY; } else if (any_input_is_lod_tensor) { var_type = framework::VarDesc::LOD_TENSOR; } auto out_var_name = op_desc.Output("Out").front(); - block->FindRecursiveOrCreateVar(out_var_name)->SetType(var_type); + auto& out_var = detail::Ref(block->FindRecursiveOrCreateVar(out_var_name)); + out_var.SetType(var_type); + auto& in_var = detail::Ref(block->FindVarRecursive(inputs.front())); + out_var.SetDataType(in_var.GetDataType()); } }; diff --git a/paddle/operators/tensor_array_read_write_op.cc b/paddle/operators/tensor_array_read_write_op.cc index 62e15604c47f25c458abc69ecd1cabf964de39bb..ae1b48d7a8e3d573a5134a822a2ed5ef70511077 100644 --- a/paddle/operators/tensor_array_read_write_op.cc +++ b/paddle/operators/tensor_array_read_write_op.cc @@ -12,7 +12,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/operators/array_operator.h" - +#include "paddle/operators/detail/safe_ref.h" namespace paddle { namespace operators { @@ -33,6 +33,8 @@ class WriteToArrayOp : public ArrayOp { auto *out = scope.FindVar(Output("Out"))->GetMutable(); if (offset >= out->size()) { + VLOG(10) << "Resize " << Output("Out") << " from " << out->size() + << " to " << offset + 1; out->resize(offset + 1); } auto *out_tensor = &out->at(offset); @@ -85,11 +87,15 @@ class WriteToArrayInferVarType : public framework::VarTypeInference { public: void operator()(const framework::OpDescBind &op_desc, framework::BlockDescBind *block) const override { - for (auto &out_var : op_desc.OutputArgumentNames()) { - VLOG(10) << "Set Variable " << out_var << " as LOD_TENSOR_ARRAY"; - block->FindRecursiveOrCreateVar(out_var)->SetType( - framework::VarDesc::LOD_TENSOR_ARRAY); - } + auto x_name = op_desc.Input("X")[0]; + auto out_name = op_desc.Output("Out")[0]; + VLOG(10) << "Set Variable " << out_name << " as LOD_TENSOR_ARRAY"; + auto &out = detail::Ref(block->FindRecursiveOrCreateVar(out_name), + "Cannot found %s", out_name); + out.SetType(framework::VarDesc::LOD_TENSOR_ARRAY); + auto &x = + detail::Ref(block->FindVarRecursive(x_name), "Cannot found %s", x_name); + out.SetDataType(x.GetDataType()); } }; @@ -107,11 +113,11 @@ class ReadFromArrayOp : public ArrayOp { auto &x_array = x->Get(); auto *out = scope.FindVar(Output("Out")); PADDLE_ENFORCE(out != nullptr, "Out must be set"); - auto *out_tesnor = out->GetMutable(); + auto *out_tensor = out->GetMutable(); size_t offset = GetOffset(scope, dev_ctx); PADDLE_ENFORCE_LT(offset, x_array.size()); - out_tesnor->CopyFrom(x_array[offset], dev_ctx.GetPlace(), dev_ctx); - out_tesnor->set_lod(x_array[offset].lod()); + out_tensor->CopyFrom(x_array[offset], dev_ctx.GetPlace(), dev_ctx); + out_tensor->set_lod(x_array[offset].lod()); } }; diff --git a/paddle/operators/transpose_op.cu b/paddle/operators/transpose_op.cu.cc similarity index 100% rename from paddle/operators/transpose_op.cu rename to paddle/operators/transpose_op.cu.cc diff --git a/paddle/operators/transpose_op.h b/paddle/operators/transpose_op.h index aaa3f47ab5545accd4d1108e0ad6f5a3062186d0..e296032f4147f9f8338148f9e4fef100c7cf816f 100644 --- a/paddle/operators/transpose_op.h +++ b/paddle/operators/transpose_op.h @@ -14,27 +14,44 @@ #pragma once -#include "paddle/framework/eigen.h" #include "paddle/framework/op_registry.h" +#include "paddle/operators/math/math_function.h" namespace paddle { namespace operators { -template -void EigenTranspose(const framework::ExecutionContext& context, - const framework::Tensor& in, framework::Tensor& out, - std::vector axis) { - Eigen::array permute; - for (int i = 0; i < Rank; i++) { - permute[i] = axis[i]; +template +inline void TransCompute(const int dim, const platform::DeviceContext& dev_ctx, + const framework::Tensor& in, framework::Tensor* out, + const std::vector& axis) { + switch (dim) { + case 1: + math::Transpose trans1; + trans1(dev_ctx, in, out, axis); + break; + case 2: + math::Transpose trans2; + trans2(dev_ctx, in, out, axis); + break; + case 3: + math::Transpose trans3; + trans3(dev_ctx, in, out, axis); + break; + case 4: + math::Transpose trans4; + trans4(dev_ctx, in, out, axis); + break; + case 5: + math::Transpose trans5; + trans5(dev_ctx, in, out, axis); + break; + case 6: + math::Transpose trans6; + trans6(dev_ctx, in, out, axis); + break; + default: + PADDLE_THROW("Tensors with rank at most 6 are supported"); } - auto in_dim = in.dims(); - auto out_dim = out.dims(); - - auto eigen_in = framework::EigenTensor::From(in); - auto eigen_out = framework::EigenTensor::From(out); - auto& dev = context.GetEigenDevice(); - eigen_out.device(dev) = eigen_in.shuffle(permute); } template @@ -47,28 +64,8 @@ class TransposeKernel : public framework::OpKernel { std::vector axis = context.Attr>("axis"); int ndims = axis.size(); - switch (ndims) { - case 1: - EigenTranspose(context, *x, *out, axis); - break; - case 2: - EigenTranspose(context, *x, *out, axis); - break; - case 3: - EigenTranspose(context, *x, *out, axis); - break; - case 4: - EigenTranspose(context, *x, *out, axis); - break; - case 5: - EigenTranspose(context, *x, *out, axis); - break; - case 6: - EigenTranspose(context, *x, *out, axis); - break; - default: - PADDLE_THROW("Tensors with rank at most 6 are supported"); - } + auto& dev_ctx = context.device_context(); + TransCompute(ndims, dev_ctx, *x, out, axis); } }; @@ -80,47 +77,19 @@ class TransposeGradKernel : public framework::OpKernel { context.Input(framework::GradVarName("Out")); auto* x_grad = context.Output(framework::GradVarName("X")); - if (x_grad) { - x_grad->mutable_data(context.GetPlace()); - - std::vector axis = context.Attr>("axis"); - std::vector reversed_axis(axis); + if (!x_grad) return; - for (size_t i = 0; i < axis.size(); i++) { - reversed_axis[axis[i]] = i; - } - - int ndims = axis.size(); + x_grad->mutable_data(context.GetPlace()); + std::vector axis = context.Attr>("axis"); + std::vector reversed_axis(axis); - switch (ndims) { - case 1: - EigenTranspose(context, *out_grad, *x_grad, - reversed_axis); - break; - case 2: - EigenTranspose(context, *out_grad, *x_grad, - reversed_axis); - break; - case 3: - EigenTranspose(context, *out_grad, *x_grad, - reversed_axis); - break; - case 4: - EigenTranspose(context, *out_grad, *x_grad, - reversed_axis); - break; - case 5: - EigenTranspose(context, *out_grad, *x_grad, - reversed_axis); - break; - case 6: - EigenTranspose(context, *out_grad, *x_grad, - reversed_axis); - break; - default: - PADDLE_THROW("Tensors with rank at most 6 are supported"); - } + for (size_t i = 0; i < axis.size(); i++) { + reversed_axis[axis[i]] = i; } + + int ndims = axis.size(); + auto& dev_ctx = context.device_context(); + TransCompute(ndims, dev_ctx, *out_grad, x_grad, reversed_axis); } }; diff --git a/paddle/operators/while_op.cc b/paddle/operators/while_op.cc index 4ca6c8507a48507fd29a9c9acae2bdf36ed936ee..dcc59f5ff2ae3a8ca999d72a20cfd5c759987d89 100644 --- a/paddle/operators/while_op.cc +++ b/paddle/operators/while_op.cc @@ -14,8 +14,10 @@ #include #include "paddle/framework/executor.h" +#include "paddle/framework/lod_tensor_array.h" #include "paddle/framework/op_registry.h" #include "paddle/framework/operator.h" +#include "paddle/operators/detail/safe_ref.h" namespace paddle { namespace operators { @@ -26,8 +28,9 @@ using LoDTensor = framework::LoDTensor; constexpr char kStepBlock[] = "step_block"; constexpr char kCondition[] = "Condition"; constexpr char kStepScopes[] = "StepScopes"; -constexpr char kParamGrads[] = "X@Grad"; constexpr char kParameters[] = "X"; +constexpr char kParamGrads[] = "X@GRAD"; +constexpr char kOutputs[] = "Out"; class WhileOp : public framework::OperatorBase { public: @@ -71,9 +74,9 @@ class WhileOpMaker : public framework::OpProtoAndCheckerMaker { kCondition, "(Bool) An scalar. When it's False, the While Op will be terminated.") .AsDuplicable(); - AddOutput("Out", + AddOutput(kOutputs, "A set of variables, which will be assigned with values " - "generated by perators inside the block of While Op.") + "generated by the operators inside the block of While Op.") .AsDuplicable(); AddOutput(kStepScopes, "(StepScopeVar) A vector of local scope, which size equals the " @@ -104,17 +107,64 @@ class WhileGradOp : public framework::OperatorBase { auto *step_scopes = scope.FindVar(Input(kStepScopes))->GetMutable(); + auto outside_og_names = Inputs(framework::GradVarName(kOutputs)); + auto inside_og_names = + Attr>("original_output_grad"); + + PADDLE_ENFORCE_EQ(outside_og_names.size(), inside_og_names.size()); + for (auto cur_scope_iter = step_scopes->rbegin(); cur_scope_iter != step_scopes->rend(); ++cur_scope_iter) { + VLOG(3) << "Start backward at time_step " + << cur_scope_iter - step_scopes->rbegin(); + framework::Scope &cur_scope = **cur_scope_iter; + // Link OG from outside to inside + for (size_t i = 0; i < outside_og_names.size(); ++i) { + auto outside_og_name = outside_og_names[i]; + auto inside_og_name = inside_og_names[i]; + VLOG(10) << "Linking outside " << outside_og_name << " --> inside " + << inside_og_name; + auto &og_outside = detail::Ref(scope.FindVar(outside_og_name)); + auto &og_inside = detail::Ref(cur_scope.Var(inside_og_name)); + if (og_outside.Type().hash_code() == + typeid(framework::LoDTensor).hash_code()) { + auto &outside_tensor = og_outside.Get(); + auto &inside_tensor = + detail::Ref(og_inside.GetMutable()); + inside_tensor.set_lod(outside_tensor.lod()); + inside_tensor.ShareDataWith(outside_tensor); + } else if (og_outside.Type().hash_code() == + typeid(framework::LoDTensorArray).hash_code()) { + auto &outside_array = og_outside.Get(); + auto &inside_array = + detail::Ref(og_inside.GetMutable()); + VLOG(10) << outside_og_name << " size = " << outside_array.size(); + inside_array.resize(outside_array.size()); + + for (size_t j = 0; j < inside_array.size(); ++j) { + VLOG(10) << j << " " << outside_array[j].numel(); + if (outside_array[j].numel() != 0) { + inside_array[j].set_lod(outside_array[j].lod()); + inside_array[j].ShareDataWith(outside_array[j]); + } else { + PADDLE_ENFORCE_EQ(inside_array[j].numel(), 0); + } + } + } + } + executor.Run(*program, *cur_scope_iter, block->ID(), false); auto &pg_names = Outputs(kParamGrads); auto &p_names = Inputs(kParameters); PADDLE_ENFORCE_EQ(pg_names.size(), p_names.size()); - for (size_t prog_id = 0; prog_id < pg_names.size(); ++prog_id) { - auto inside_grad_name = framework::GradVarName(p_names[prog_id]); + for (size_t param_id = 0; param_id < pg_names.size(); ++param_id) { + if (pg_names[param_id] == framework::kEmptyVarName) { + continue; // iterator doesn't have gradient + } + auto inside_grad_name = framework::GradVarName(p_names[param_id]); - // // TODO(tonyyang-savil: Not sure we need the following + // // TODO(tonyyang-svail): Not sure we need the following // // If does not compute gradient of that variable inside rnn, // just // // continue @@ -126,7 +176,7 @@ class WhileGradOp : public framework::OperatorBase { // zero gradient variable in step 0 if (cur_scope_iter == step_scopes->rbegin()) { auto *var = (*cur_scope_iter)->FindVar(inside_grad_name); - PADDLE_ENFORCE_NOT_NULL(var); + PADDLE_ENFORCE_NOT_NULL(var, "Can not find var %s", inside_grad_name); if (var->IsType()) { auto &inside_tensor = var->Get(); framework::AttributeMap attrs; @@ -135,27 +185,18 @@ class WhileGradOp : public framework::OperatorBase { attrs["value"] = 0.0f; auto zero_op = framework::OpRegistry::CreateOp( - "fill_constant", {}, {{"Out", {pg_names[prog_id]}}}, attrs); + "fill_constant", {}, {{"Out", {pg_names[param_id]}}}, attrs); zero_op->Run(scope, dev_ctx); } } // sum gradient - auto *outside_var = scope.FindVar(pg_names[prog_id]); - PADDLE_ENFORCE_NOT_NULL(outside_var); - auto &outside_tensor = *outside_var->GetMutable(); - - std::string result_var_name; - auto *local_result_var = (*cur_scope_iter)->Var(&result_var_name); - auto &local_result_tensor = - *local_result_var->GetMutable(); - - local_result_tensor.ShareDataWith(outside_tensor); - + auto new_inside_name = cur_scope.Rename(inside_grad_name); auto sum_op = framework::OpRegistry::CreateOp( - "sum", {{"X", {result_var_name, inside_grad_name}}}, - {{"Out", {result_var_name}}}, {}); - sum_op->Run(**cur_scope_iter, dev_ctx); + "sum", {{"X", {pg_names[param_id], new_inside_name}}}, + {{"Out", {pg_names[param_id]}}}, {}); + sum_op->Run(cur_scope, dev_ctx); + cur_scope.Rename(new_inside_name, inside_grad_name); } } } @@ -169,29 +210,110 @@ class WhileGradOpDescMaker : public framework::SingleGradOpDescMaker { virtual std::unique_ptr Apply() const { auto *grad = new framework::OpDescBind(); grad->SetType("while_grad"); - for (auto &input_param : this->InputNames()) { - grad->SetInput(input_param, this->Input(input_param)); - grad->SetOutput(framework::GradVarName(input_param), - this->InputGrad(input_param)); + grad->SetInput(kParameters, Input(kParameters)); + grad->SetOutput( + framework::GradVarName(kParameters), + InputGrad(kParameters, /*do not drop empty gradient*/ false)); + grad->SetInput(kOutputs, Output(kOutputs)); + + // OG should be re-calculated by step blocks, since many outputs of while op + // do not need to calculate gradients. + std::unordered_set block_ins; + { + for (auto &p : Input(kParameters)) { + block_ins.insert(p); + } + for (auto &o : Output(kOutputs)) { + block_ins.insert(o); + } } + std::unordered_set extra_inputs; + for (size_t i = 0; i < grad_block_[0]->OpSize(); ++i) { + for (auto &input_name : grad_block_[0]->Op(i)->InputArgumentNames()) { + if (block_ins.find(input_name) != block_ins.end()) { + continue; + } + extra_inputs.insert(input_name); + } - for (auto &output_param : this->OutputNames()) { - grad->SetInput(output_param, this->Output(output_param)); - if (output_param != kStepScopes) { - grad->SetInput(framework::GradVarName(output_param), - this->OutputGrad(output_param)); + for (auto &output_name : grad_block_[0]->Op(i)->OutputArgumentNames()) { + block_ins.insert(output_name); } } + + std::vector extra_inputs_list; + extra_inputs_list.resize(extra_inputs.size()); + std::copy(extra_inputs.begin(), extra_inputs.end(), + extra_inputs_list.begin()); + grad->SetInput(framework::GradVarName(kOutputs), extra_inputs_list); + grad->SetInput(kStepScopes, Output(kStepScopes)); grad->SetAttrMap(this->Attrs()); grad->SetBlockAttr(kStepBlock, *grad_block_[0]); + // record the original output gradient names, since the gradient name of + // while operator could be renamed. + grad->SetAttr("original_output_grad", extra_inputs_list); return std::unique_ptr(grad); } }; +class WhileGradOpVarTypeInference : public framework::VarTypeInference { + public: + void operator()(const framework::OpDescBind &op_desc, + framework::BlockDescBind *block) const override { + auto p_names = op_desc.Input(kParameters); + auto pg_names = op_desc.Output(framework::GradVarName(kParameters)); + + for (size_t i = 0; i < p_names.size(); ++i) { + auto &p_var = detail::Ref(block->FindVarRecursive(p_names[i])); + auto *g_var = block->FindVarRecursive(pg_names[i]); + if (g_var != nullptr) { // Gradient could be @EMPTY@ + VLOG(5) << "Setting " << pg_names[i] << " following " << p_names[i] + << " type: " << p_var.GetType(); + g_var->SetType(p_var.GetType()); + g_var->SetDataType(p_var.GetDataType()); + } + } + } +}; + +class WhileGradOpShapeInference : public framework::InferShapeBase { + public: + void operator()(framework::InferShapeContext *ctx) const override { + ctx->HasInputs(kParameters); + ctx->HasOutputs(framework::GradVarName(kParameters)); + ctx->HasInputs(kOutputs); + ctx->HasInputs(framework::GradVarName(kOutputs)); + + auto p_names = ctx->Inputs(kParameters); + auto pg_names = ctx->Outputs(kParamGrads); + auto dims = ctx->GetInputsDim(kParameters); + auto var_types = ctx->GetInputsVarType(kParameters); + std::vector names_to_set; + std::vector dims_to_set; + for (size_t i = 0; i < p_names.size(); ++i) { + if (pg_names[i] == framework::kEmptyVarName) { + continue; + } + if (var_types[i] == framework::VarDesc::LOD_TENSOR) { + names_to_set.push_back(pg_names[i]); + dims_to_set.push_back(dims[i]); + } else if (var_types[i] == framework::VarDesc::LOD_TENSOR_ARRAY) { + // not sure how to set the dim of LOD_TENSOR_ARRAY + names_to_set.push_back(pg_names[i]); + dims_to_set.push_back(dims[i]); + } + } + ctx->SetDims(names_to_set, dims_to_set); + } +}; + } // namespace operators } // namespace paddle REGISTER_OPERATOR(while, paddle::operators::WhileOp, paddle::operators::WhileOpMaker, paddle::operators::WhileGradOpDescMaker); +REGISTER_OPERATOR(while_grad, paddle::operators::WhileGradOp, + paddle::operators::WhileGradOpShapeInference, + paddle::operators::WhileGradOpVarTypeInference); diff --git a/paddle/platform/dynload/cublas.h b/paddle/platform/dynload/cublas.h index 6b64539b0a9a4d535a53447fbcc0e458f3ac9129..61a22d9db3e07cbe6fbca0e0b09fedcba232ff6c 100644 --- a/paddle/platform/dynload/cublas.h +++ b/paddle/platform/dynload/cublas.h @@ -62,6 +62,8 @@ extern void *cublas_dso_handle; DECLARE_DYNAMIC_LOAD_CUBLAS_WRAP(__name) #define CUBLAS_BLAS_ROUTINE_EACH(__macro) \ + __macro(cublasSaxpy_v2); \ + __macro(cublasDaxpy_v2); \ __macro(cublasSgemv_v2); \ __macro(cublasDgemv_v2); \ __macro(cublasSgemm_v2); \ diff --git a/paddle/platform/gpu_info.cc b/paddle/platform/gpu_info.cc index f3455a8733862c91eaece629b6684d446672336c..36b216d872138d49bfd5ab6e3499d15d49ebd0ca 100644 --- a/paddle/platform/gpu_info.cc +++ b/paddle/platform/gpu_info.cc @@ -109,5 +109,10 @@ void GpuMemcpyPeer(void *dst, int dst_device, const void *src, int src_device, cudaMemcpyPeerAsync(dst, dst_device, src, src_device, count, stream), "cudaMemcpyPeerAsync failed in paddle::platform::GpuMemcpyPeer"); } + +void GpuMemsetAsync(void *dst, int value, size_t count, cudaStream_t stream) { + PADDLE_ENFORCE(cudaMemsetAsync(dst, value, count, stream), + "cudaMemsetAsync failed in paddle::platform::GpuMemsetAsync"); +} } // namespace platform } // namespace paddle diff --git a/paddle/platform/gpu_info.h b/paddle/platform/gpu_info.h index 37665b97d764fbcfe0964127d230b1d28d90b687..db961f3838af73855312d4cf6a80e2355306e08f 100644 --- a/paddle/platform/gpu_info.h +++ b/paddle/platform/gpu_info.h @@ -60,6 +60,9 @@ void GpuMemcpySync(void *dst, const void *src, size_t count, void GpuMemcpyPeer(void *dst, int dst_device, const void *src, int src_device, size_t count, cudaStream_t stream); +//! Set memory dst with value count size asynchronously +void GpuMemsetAsync(void *dst, int value, size_t count, cudaStream_t stream); + } // namespace platform } // namespace paddle diff --git a/paddle/scripts/submit_local.sh.in b/paddle/scripts/submit_local.sh.in index 5c4b5a2495182ea5d2b3341cff650dfb4d8b0c0f..b9a49526a7e02131767a4e9b26cd0b53278176d0 100755 --- a/paddle/scripts/submit_local.sh.in +++ b/paddle/scripts/submit_local.sh.in @@ -43,6 +43,54 @@ function ver2num() { set +e } +function cpu_config() { + # auto set KMP_AFFINITY and OMP_DYNAMIC from Hyper Threading Status + # only when MKLDNN or MKLML enabled + if [ "@WITH_MKLDNN@" == "OFF" ] && [ "@WITH_MKLML@" == "OFF"]; then + return 0 + fi + ht=`lscpu |grep "per core"|awk -F':' '{print $2}'|xargs` + if [ $ht -eq 1 ]; then # HT is OFF + if [ -z "$KMP_AFFINITY" ]; then + export KMP_AFFINITY="granularity=fine,compact,0,0" + fi + if [ -z "$OMP_DYNAMIC" ]; then + export OMP_DYNAMIC="FALSE" + fi + else # HT is ON + if [ -z "$KMP_AFFINITY" ]; then + export KMP_AFFINITY="granularity=fine,compact,1,0" + fi + if [ -z "$OMP_DYNAMIC" ]; then + export OMP_DYNAMIC="True" + fi + fi +} + +function threads_config() { + # auto set OMP_NUM_THREADS and MKL_NUM_THREADS + # according to trainer_count and total processors + # only when MKLDNN or MKLML enabled + if [ "@WITH_MKLDNN@" == "OFF" ] && [ "@WITH_MKLML@" == "OFF"]; then + return 0 + fi + processors=`grep "processor" /proc/cpuinfo|sort -u|wc -l` + trainers=`grep -Eo 'trainer_count.[0-9]+' <<< "$@" |grep -Eo '[0-9]+'|xargs` + if [ -z $trainers ]; then + trainers=1 + fi + threads=$((processors / trainers)) + if [ $threads -eq 0 ]; then + threads=1 + fi + if [ -z "$OMP_NUM_THREADS" ]; then + export OMP_NUM_THREADS=$threads + fi + if [ -z "$MKL_NUM_THREADS" ]; then + export MKL_NUM_THREADS=$threads + fi +} + PADDLE_CONF_HOME="$HOME/.config/paddle" mkdir -p ${PADDLE_CONF_HOME} @@ -92,9 +140,13 @@ else: sys.exit(0) EOF +cpu_config +# echo $KMP_AFFINITY $OMP_DYNAMIC case "$1" in "train") + threads_config $@ + # echo $OMP_NUM_THREADS $MKL_NUM_THREADS ${DEBUGGER} $PADDLE_BIN_PATH/paddle_trainer ${@:2} ;; "merge_model") diff --git a/paddle/trainer/Trainer.cpp b/paddle/trainer/Trainer.cpp index b68e29cd5ea223272151e7a8b52d998832f47103..88e684849df6fbfe4042b92bdb76ef98159eecea 100644 --- a/paddle/trainer/Trainer.cpp +++ b/paddle/trainer/Trainer.cpp @@ -137,6 +137,10 @@ void Trainer::init(const std::shared_ptr& config, } } + if (FLAGS_use_mkldnn) { + CHECK_EQ(FLAGS_trainer_count, 1UL) << "MKLDNN only need 1 trainer"; + } + if (testing) { LOG(INFO) << "trainer: in testing mode"; if (config_->getOptConfig().use_sparse_remote_updater() || diff --git a/python/paddle/trainer_config_helpers/layers.py b/python/paddle/trainer_config_helpers/layers.py index 5de1c18950a3236faa91edabf0119b590b22c6d9..a02eba007ddf929ff92df995df253f5a386bac7b 100644 --- a/python/paddle/trainer_config_helpers/layers.py +++ b/python/paddle/trainer_config_helpers/layers.py @@ -2987,8 +2987,10 @@ def img_cmrnorm_layer(input, layer_attr=None): """ Response normalization across feature maps. - The details please refer to - `Alex's paper `_. + + Reference: + ImageNet Classification with Deep Convolutional Neural Networks + http://www.cs.toronto.edu/~fritz/absps/imagenet.pdf The example usage is: @@ -2997,7 +2999,7 @@ def img_cmrnorm_layer(input, norm = img_cmrnorm_layer(input=net, size=5) :param name: The name of this layer. It is optional. - :type name: None | basestring + :type name: basestring :param input: The input of this layer. :type input: LayerOutput :param size: Normalize in number of :math:`size` feature maps. @@ -3006,9 +3008,11 @@ def img_cmrnorm_layer(input, :type scale: float :param power: The hyper-parameter. :type power: float - :param num_channels: input layer's filers number or channels. If - num_channels is None, it will be set automatically. - :param layer_attr: Extra Layer Attribute. + :param num_channels: The number of input channels. If the parameter is not set or + set to None, its actual value will be automatically set to + the channels number of the input. + :param layer_attr: The extra layer attributes. See ExtraLayerAttribute for + details. :type layer_attr: ExtraLayerAttribute :return: LayerOutput object. :rtype: LayerOutput @@ -3036,7 +3040,7 @@ def batch_norm_layer(input, use_global_stats=None, mean_var_names=None): """ - Batch Normalization Layer. The notation of this layer as follow. + Batch Normalization Layer. The notation of this layer is as follows. :math:`x` is the input features over a mini-batch. @@ -3050,8 +3054,10 @@ def batch_norm_layer(input, \\sigma_{\\beta}^{2} + \\epsilon}} \\qquad &//\ normalize \\\\ y_i &\\gets \\gamma \\hat{x_i} + \\beta \\qquad &//\ scale\ and\ shift - The details of batch normalization please refer to this - `paper `_. + Reference: + Batch Normalization: Accelerating Deep Network Training by Reducing + Internal Covariate Shift + http://arxiv.org/abs/1502.03167 The example usage is: @@ -3061,48 +3067,47 @@ def batch_norm_layer(input, :param name: The name of this layer. It is optional. :type name: basestring - :param input: batch normalization input. Better be linear activation. - Because there is an activation inside batch_normalization. + :param input: This layer's input which is to be performed batch normalization on. :type input: LayerOutput :param batch_norm_type: We have batch_norm, mkldnn_batch_norm and cudnn_batch_norm. batch_norm supports CPU, MKLDNN and GPU. cudnn_batch_norm requires cuDNN version greater or equal to v4 (>=v4). But cudnn_batch_norm is faster and needs less memory than batch_norm. mkldnn_batch_norm requires - enable use_mkldnn. By default (None), we will - automaticly select cudnn_batch_norm for GPU, + use_mkldnn is enabled. By default (None), we will + automatically select cudnn_batch_norm for GPU, mkldnn_batch_norm for MKLDNN and batch_norm for CPU. - Otherwise, select batch norm type based on the - specified type. If you use cudnn_batch_norm, - we suggested you use latest version, such as v5.1. + Users can specify the batch norm type. If you use + cudnn_batch_norm, we suggested you use latest version, + such as v5.1. :type batch_norm_type: None | string, None or "batch_norm" or "cudnn_batch_norm" or "mkldnn_batch_norm" - :param act: Activation Type. Better be relu. Because batch - normalization will normalize input near zero. + :param act: Activation type. ReluActivation is the default activation. :type act: BaseActivation - :param num_channels: num of image channels or previous layer's number of - filters. None will automatically get from layer's - input. + :param num_channels: The number of input channels. If the parameter is not set or + set to None, its actual value will be automatically set to + the channels number of the input. :type num_channels: int - :param bias_attr: :math:`\\beta`, better be zero when initialize. So the - initial_std=0, initial_mean=1 is best practice. + :param bias_attr: :math:`\\beta`. The bias attribute. If the parameter is set to + False or an object whose type is not ParameterAttribute, no + bias is defined. If the parameter is set to True, the bias is + initialized to zero. :type bias_attr: ParameterAttribute | None | bool | Any - :param param_attr: :math:`\\gamma`, better be one when initialize. So the - initial_std=0, initial_mean=1 is best practice. + :param param_attr: :math:`\\gamma`. The parameter attribute. See ParameterAttribute + for details. :type param_attr: ParameterAttribute - :param layer_attr: Extra Layer Attribute. + :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for + details. :type layer_attr: ExtraLayerAttribute - :param use_global_stats: whether use moving mean/variance statistics - during testing peroid. If None or True, - it will use moving mean/variance statistics during - testing. If False, it will use the mean - and variance of current batch of test data for - testing. + :param use_global_stats: Whether use moving mean/variance statistics during + testing peroid. If the parameter is set to None or + True, it will use moving mean/variance statistics + during testing. If the parameter is set to False, it + will use the mean and variance of the current batch + of test data. :type use_global_stats: bool | None. - :param moving_average_fraction: Factor used in the moving average - computation, referred to as facotr, - :math:`runningMean = newMean*(1-factor) - + runningMean*factor` + :param moving_average_fraction: Factor used in the moving average computation. + :math:`runningMean = newMean*(1-factor) + runningMean*factor` :type moving_average_fraction: float. :param mean_var_names: [mean name, variance name] :type mean_var_names: string list @@ -3164,8 +3169,9 @@ def sum_to_one_norm_layer(input, name=None, layer_attr=None): :type input: LayerOutput :param name: The name of this layer. It is optional. :type name: basestring - :param layer_attr: extra layer attributes. - :type layer_attr: ExtraLayerAttribute. + :param layer_attr: The extra layer attribute. See ExtraLayerAttribute + for details. + :type layer_attr: ExtraLayerAttribute :return: LayerOutput object. :rtype: LayerOutput """ @@ -3200,7 +3206,8 @@ def row_l2_norm_layer(input, name=None, layer_attr=None): :type input: LayerOutput :param name: The name of this layer. It is optional. :type name: basestring - :param layer_attr: extra layer attributes. + :param layer_attr: The extra layer attribute. See ExtraLayerAttribute + for details. :type layer_attr: ExtraLayerAttribute. :return: LayerOutput object. :rtype: LayerOutput @@ -3237,22 +3244,17 @@ def addto_layer(input, act=None, name=None, bias_attr=None, layer_attr=None): act=ReluActivation(), bias_attr=False) - This layer just simply add all input layers together, then activate the sum - inputs. Each input of this layer should be the same size, which is also the - output size of this layer. + This layer just simply adds all input layers together, then activates the + sum. All inputs should share the same dimension, which is also the dimension + of this layer's output. There is no weight matrix for each input, because it just a simple add operation. If you want a complicated operation before add, please use mixed_layer. - It is a very good way to set dropout outside the layers. Since not all - PaddlePaddle layer support dropout, you can add an add_to layer, set - dropout here. - Please refer to dropout_layer for details. - :param name: The name of this layer. It is optional. :type name: basestring - :param input: Input layers. It could be a LayerOutput or list/tuple of + :param input: The input layers. It could be a LayerOutput or list/tuple of LayerOutput. :type input: LayerOutput | list | tuple :param act: Activation Type. LinearActivation is the default activation. @@ -3261,7 +3263,8 @@ def addto_layer(input, act=None, name=None, bias_attr=None, layer_attr=None): whose type is not ParameterAttribute, no bias is defined. If the parameter is set to True, the bias is initialized to zero. :type bias_attr: ParameterAttribute | None | bool | Any - :param layer_attr: Extra Layer attribute. + :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for + details. :type layer_attr: ExtraLayerAttribute :return: LayerOutput object. :rtype: LayerOutput @@ -3300,8 +3303,8 @@ def addto_layer(input, act=None, name=None, bias_attr=None, layer_attr=None): @layer_support(DROPOUT, ERROR_CLIPPING) def concat_layer(input, act=None, name=None, layer_attr=None, bias_attr=None): """ - Concat all input vector into one huge vector. - Inputs can be list of LayerOutput or list of projection. + Concatenate all input vectors to one vector. + Inputs can be a list of LayerOutput or a list of projection. The example usage is: @@ -3311,11 +3314,12 @@ def concat_layer(input, act=None, name=None, layer_attr=None, bias_attr=None): :param name: The name of this layer. It is optional. :type name: basestring - :param input: input layers or projections + :param input: The input layers or projections :type input: list | tuple | collections.Sequence :param act: Activation type. IdentityActivation is the default activation. :type act: BaseActivation - :param layer_attr: Extra Layer Attribute. + :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for + details. :type layer_attr: ExtraLayerAttribute :return: LayerOutput object. :rtype: LayerOutput @@ -3385,7 +3389,7 @@ def concat_layer(input, act=None, name=None, layer_attr=None, bias_attr=None): def seq_concat_layer(a, b, act=None, name=None, layer_attr=None, bias_attr=None): """ - Concat sequence a with sequence b. + Concatenate sequence a and sequence b. Inputs: - a = [a1, a2, ..., am] @@ -3404,13 +3408,14 @@ def seq_concat_layer(a, b, act=None, name=None, layer_attr=None, :param name: The name of this layer. It is optional. :type name: basestring - :param a: input sequence layer + :param a: The first input sequence layer :type a: LayerOutput - :param b: input sequence layer + :param b: The second input sequence layer :type b: LayerOutput :param act: Activation type. IdentityActivation is the default activation. :type act: BaseActivation - :param layer_attr: Extra Layer Attribute. + :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for + details. :type layer_attr: ExtraLayerAttribute :param bias_attr: The bias attribute. If the parameter is set to False or an object whose type is not ParameterAttribute, no bias is defined. If the @@ -3447,31 +3452,25 @@ def memory(name, boot_bias_active_type=None, boot_with_const_id=None): """ - The memory layers is a layer cross each time step. Reference this output - as previous time step layer :code:`name` 's output. + The memory takes a layer's output at previous time step as its own output. - The default memory is zero in first time step, previous time step's - output in the rest time steps. + If boot_bias, the activation of the bias is the initial value of the memory. - If boot_bias, the first time step value is this bias and - with activation. + If boot_with_const_id is set, then the memory's output at the first time step + is a IndexSlot, the Arguments.ids()[0] is this :code:`cost_id`. - If boot_with_const_id, then the first time stop is a IndexSlot, the - Arguments.ids()[0] is this :code:`cost_id`. + If boot_layer is specified, the memory's output at the first time step will + be the boot_layer's output. - If boot_layer is not null, the memory is just the boot_layer's output. - Set :code:`is_seq` is true boot layer is sequence. - - The same name layer in recurrent group will set memory on each time - step. + In other case, the default memory's output at the first time step is zero. .. code-block:: python mem = memory(size=256, name='state') state = fc_layer(input=mem, size=256, name='state') - If you do not want to specify the name, you can equivalently use set_input() - to specify the layer needs to be remembered as the following: + If you do not want to specify the name, you can also use set_input() + to specify the layer to be remembered as the following: .. code-block:: python @@ -3479,26 +3478,31 @@ def memory(name, state = fc_layer(input=mem, size=256) mem.set_input(mem) - :param name: the name of the layer which this memory remembers. + :param name: The name of the layer which this memory remembers. If name is None, user should call set_input() to specify the name of the layer which this memory remembers. :type name: basestring - :param size: size of memory. + :param size: The dimensionality of memory. :type size: int - :param memory_name: the name of the memory. - It is ignored when name is provided. + :param memory_name: The name of the memory. It is ignored when name is provided. :type memory_name: basestring :param is_seq: DEPRECATED. is sequence for boot_layer :type is_seq: bool - :param boot_layer: boot layer of memory. + :param boot_layer: This parameter specifies memory's output at the first time + step and the output is boot_layer's output. :type boot_layer: LayerOutput | None - :param boot_bias: boot layer's bias + :param boot_bias: The bias attribute of memory's output at the first time step. + If the parameter is set to False or an object whose type is not + ParameterAttribute, no bias is defined. If the parameter is set + to True, the bias is initialized to zero. :type boot_bias: ParameterAttribute | None - :param boot_bias_active_type: boot layer's active type. + :param boot_bias_active_type: Activation type for memory's bias at the first time + step. LinearActivation is the default activation. :type boot_bias_active_type: BaseActivation - :param boot_with_const_id: boot layer's id. + :param boot_with_const_id: This parameter specifies memory's output at the first + time step and the output is an index. :type boot_with_const_id: int - :return: LayerOutput object which is a memory. + :return: LayerOutput object. :rtype: LayerOutput """ if boot_bias_active_type is None: diff --git a/python/paddle/v2/__init__.py b/python/paddle/v2/__init__.py index 3d7051384346e92f2776511799d83b5d7f6ce62d..7bbe3eaaa67a117bc53571e6571365c3a26814c1 100644 --- a/python/paddle/v2/__init__.py +++ b/python/paddle/v2/__init__.py @@ -33,7 +33,6 @@ import networks import minibatch import plot import image -import model import paddle.trainer.config_parser as cp __all__ = [ @@ -58,7 +57,6 @@ __all__ = [ 'evaluator', 'image', 'master', - 'model', ] cp.begin_parse() @@ -78,6 +76,31 @@ def init(**kwargs): for key in args_dict.keys(): args.append('--%s=%s' % (key, str(args_dict[key]))) + # auto set cpu environment + def set_env(key, value): + '''If the key has not been set in the environment, set it with value.''' + assert isinstance(key, str) + assert isinstance(value, str) + envset = os.environ.get(key) + if envset is None: + os.environ[key] = value + + ht = os.popen("lscpu |grep \"per core\"|awk -F':' '{print $2}'|xargs") + ht = int(ht.read()) + if ht == 1: # ht is off + set_env("OMP_DYNAMIC", "false") + set_env("KMP_AFFINITY", "granularity=fine,compact,0,0") + else: + set_env("OMP_DYNAMIC", "true") + set_env("KMP_AFFINITY", "granularity=fine,compact,1,0") + processors = os.popen("grep \"processor\" /proc/cpuinfo|sort -u|wc -l") + processors = int(processors.read()) + trainers = kwargs.get('trainer_count', 1) + threads = processors / trainers + threads = '1' if threads < 1 else str(threads) + set_env("OMP_NUM_THREADS", threads) + set_env("MKL_NUM_THREADS", threads) + if 'use_gpu' in kwargs: cp.g_command_config_args['use_gpu'] = kwargs['use_gpu'] if 'use_mkldnn' in kwargs: diff --git a/python/paddle/v2/fluid/framework.py b/python/paddle/v2/fluid/framework.py index f20567243ae67baecbdbac13f879f4cf2f66d298..a6eca2d7194c30aabeafc34de0957792feeebbec 100644 --- a/python/paddle/v2/fluid/framework.py +++ b/python/paddle/v2/fluid/framework.py @@ -12,9 +12,9 @@ def unique_name(prefix): return "_".join([prefix, str(uid)]) -def _debug_string_(proto): +def _debug_string_(proto, throw_on_error=True): error_fields = list() - if not proto.IsInitialized(error_fields): + if not proto.IsInitialized(error_fields) and throw_on_error: raise ValueError("{0} are not initialized\nThe message is {1}".format( error_fields, proto)) return proto.__str__() @@ -101,9 +101,12 @@ class Variable(object): self.stop_gradient = stop_gradient def __str__(self): + return self.to_string(True) + + def to_string(self, throw_on_error): protostr = self.desc.serialize_to_string() proto = framework_pb2.VarDesc.FromString(str(protostr)) - return _debug_string_(proto) + return _debug_string_(proto, throw_on_error) __repr__ = __str__ @@ -291,10 +294,13 @@ class Operator(object): self.desc.infer_var_type(self.block.desc) self.desc.infer_shape(self.block.desc) - def __str__(self): + def to_string(self, throw_on_error): protostr = self.desc.serialize_to_string() proto = framework_pb2.OpDesc.FromString(str(protostr)) - return _debug_string_(proto) + return _debug_string_(proto, throw_on_error) + + def __str__(self): + return self.to_string(True) __repr__ = __str__ @@ -349,9 +355,12 @@ class Block(object): self.program = program def __str__(self): + return self.to_string(True) + + def to_string(self, throw_on_error): protostr = self.desc.serialize_to_string() proto = framework_pb2.BlockDesc.FromString(str(protostr)) - return _debug_string_(proto) + return _debug_string_(proto, throw_on_error) __repr__ = __str__ @@ -454,9 +463,12 @@ class Program(object): self.current_block_idx = 0 def __str__(self): + return self.to_string(True) + + def to_string(self, throw_on_error): protostr = self.desc.serialize_to_string() proto = framework_pb2.ProgramDesc.FromString(str(protostr)) - return _debug_string_(proto) + return _debug_string_(proto, throw_on_error) def clone(self): p = Program() @@ -512,7 +524,14 @@ class Program(object): assert isinstance(target, Variable) if no_grad_set is None: no_grad_set = set() - param_to_grad_info = self.desc.append_backward(target.desc, no_grad_set) + try: + param_to_grad_info = self.desc.append_backward(target.desc, + no_grad_set) + except Exception as e: + raise core.EnforceNotMet( + str(e) + "\nCurrent protobuf is\n{0}".format( + self.to_string(False))) + self.sync_with_cpp() return param_to_grad_info diff --git a/python/paddle/v2/fluid/io.py b/python/paddle/v2/fluid/io.py index 394a171c67a99ffb0c7caaf71e850fe541f8286e..2d070814eef0b099ba71bef223596e30388ac48a 100644 --- a/python/paddle/v2/fluid/io.py +++ b/python/paddle/v2/fluid/io.py @@ -35,7 +35,7 @@ def save_vars(executor, dirname, main_program=None, vars=None, predicate=None): :param executor: executor that save variable :param dirname: directory path - :param main_program: program. If vars is None, then filter all variables in this + :param main_program: program. If vars is None, then filter all variables in this program which fit `predicate`. Default g_program. :param predicate: The Predicate describes a callable that returns a variable as a bool. If it returns true, the variables will be saved. @@ -96,11 +96,11 @@ def load_vars(executor, dirname, main_program=None, vars=None, predicate=None): :param executor: executor that save variable :param dirname: directory path - :param main_program: program. If vars is None, then filter all variables in this + :param main_program: program. If vars is None, then filter all variables in this program which fit `predicate`. Default g_program. :param predicate: The Predicate describes a callable that returns a variable as a bool. If it returns true, the variables will be loaded. - :param vars: variables need to be loaded. If specify vars, program & + :param vars: variables need to be loaded. If specify vars, program & predicate will be ignored :return: None """ @@ -157,15 +157,15 @@ def save_inference_model(dirname, executor, main_program=None): """ - Build a model especially for inference, + Build a model especially for inference, and save it to directory by the executor. :param dirname: directory path :param feeded_var_names: Names of variables that need to be feeded data during inference :param target_vars: Variables from which we can get inference results. :param executor: executor that save inference model - :param main_program: original program, which will be pruned to build the inference model. - Default g_program. + :param main_program: original program, which will be pruned to build the inference model. + Default g_main_program. :return: None """ @@ -234,3 +234,35 @@ def load_inference_model(dirname, executor): fetch_vars = [program.global_block().var(name) for name in fetch_var_names] return [program, feed_var_names, fetch_vars] + + +def get_parameter_value(para, executor): + """ + Get the LoDTensor for the parameter + + :param executor: executor for retrieving the value + :param para: the given parameter + :return: the LoDTensor for the parameter + """ + assert is_parameter(para) + + get_program = Program() + block = get_program.global_block() + new_var = _clone_var_in_block_(block, para) + return executor.run(get_program, feed={}, fetch_list=[new_var])[0] + + +def get_parameter_value_by_name(name, executor, program=None): + """ + Get the LoDTensor for paramter with the given name + + :param executor: executor for retrieving the value + :param name: the name of the parameter + :param program: the program where the variable is found + Default g_main_program. + :return: the LoDTensor for the variable + """ + if program is None: + program = g_main_program + var = program.global_block().var(name) + return get_parameter_value(var, executor) diff --git a/python/paddle/v2/fluid/layer_helper.py b/python/paddle/v2/fluid/layer_helper.py index 9dc3c119ea47ca11956d85119ce1ec6d3d6bb7e8..a97e07982bd89be72386970f28a0dd049f82372d 100644 --- a/python/paddle/v2/fluid/layer_helper.py +++ b/python/paddle/v2/fluid/layer_helper.py @@ -72,7 +72,7 @@ class LayerHelper(object): @property def bias_attr(self): - default = {'name': None, 'initializer': XavierInitializer()} + default = {'name': None, 'initializer': ConstantInitializer()} bias_attr = self.kwargs.get('bias_attr', None) if bias_attr is None: bias_attr = default @@ -149,24 +149,19 @@ class LayerHelper(object): persistable=True, initializer=initializer) - def append_bias_op(self, input_var, num_flatten_dims=None): + def append_bias_op(self, input_var, dim_start=1, dim_end=None): """ - Append bias operator and return its output. If the user does not set + Append bias operator and return its output. If the user does not set bias_attr, append_bias_op will return input_var - + :param input_var: the input variable. The len(input_var.shape) is larger or equal than 2. - :param num_flatten_dims: The input tensor will be flatten as a matrix - when adding bias. - `matrix.shape = product(input_var.shape[0:num_flatten_dims]), product( - input_var.shape[num_flatten_dims:])` + :param dim_start: + :param dim_end: the shape of the bias will be + input_var.shape[dim_start:dim_end]. The bias is broadcasted to other + dimensions and added to input_var to get the output """ - if num_flatten_dims is None: - num_flatten_dims = self.kwargs.get('num_flatten_dims', None) - if num_flatten_dims is None: - num_flatten_dims = 1 - - size = list(input_var.shape[num_flatten_dims:]) + size = list(input_var.shape[dim_start:dim_end]) bias_attr = self.bias_attr if not bias_attr: return input_var @@ -178,7 +173,8 @@ class LayerHelper(object): type='elementwise_add', inputs={'X': [input_var], 'Y': [b]}, - outputs={'Out': [tmp]}) + outputs={'Out': [tmp]}, + attrs={'axis': dim_start}) return tmp def append_activation(self, input_var): diff --git a/python/paddle/v2/fluid/layers.py b/python/paddle/v2/fluid/layers.py index b582f2ef6df4ca38e77f69ea49f8e8bc2c91f23e..1789d2f82a8813331b3610fc69f8447925cd7501 100644 --- a/python/paddle/v2/fluid/layers.py +++ b/python/paddle/v2/fluid/layers.py @@ -250,7 +250,7 @@ def _convert_(name): def _generate_doc_string_(op_proto): """ Generate docstring by OpProto - + Args: op_proto (framework_pb2.OpProto): a protobuf message typed OpProto @@ -694,7 +694,7 @@ def conv2d(input, 'paddings': padding, 'groups': groups}) - pre_act = helper.append_bias_op(pre_bias, 1) + pre_act = helper.append_bias_op(pre_bias, dim_start=1, dim_end=2) return helper.append_activation(pre_act) diff --git a/python/paddle/v2/fluid/net_drawer.py b/python/paddle/v2/fluid/net_drawer.py index 17ad547c2bb5b79ef8225dd1a8f1ef49a6572508..94fdd5e38970b309580de6fc934b158a3c46e464 100644 --- a/python/paddle/v2/fluid/net_drawer.py +++ b/python/paddle/v2/fluid/net_drawer.py @@ -66,10 +66,13 @@ def parse_graph(program, graph, var_dict, **kwargs): if not var_dict.has_key(var): var_dict[var] = "Feed" + temp_id = 0 proto = framework_pb2.ProgramDesc.FromString( program.desc.serialize_to_string()) for block in proto.blocks: for op in block.ops: + op.type = op.type + "_" + str(temp_id) + temp_id += 1 graph.node(**draw_node(op)) for o in op.outputs: for arg in o.arguments: @@ -78,6 +81,7 @@ def parse_graph(program, graph, var_dict, **kwargs): for arg in e.arguments: if var_dict.has_key(arg): graph.edge(**draw_edge(var_dict, op, e, arg)) + break # only plot the first block def draw_graph(startup_program, main_program, **kwargs): diff --git a/python/paddle/v2/fluid/tests/test_adagrad_op.py b/python/paddle/v2/fluid/tests/test_adagrad_op.py index 66bad349e59b608cb3cc965401c81ef4c716b318..903e84c32887100bbeef6ebf81f66f06f084fab5 100644 --- a/python/paddle/v2/fluid/tests/test_adagrad_op.py +++ b/python/paddle/v2/fluid/tests/test_adagrad_op.py @@ -1,6 +1,9 @@ import unittest import numpy as np +import paddle.v2.fluid.core as core +from paddle.v2.fluid.op import Operator from op_test import OpTest +import math class TestAdagradOp1(OpTest): @@ -65,5 +68,110 @@ class TestAdagradOp2(OpTest): self.check_output() +class TestSparseAdagradOp(unittest.TestCase): + def check_with_place(self, place): + scope = core.Scope() + + # create and initialize Grad Variable + height = 10 + rows = [0, 4, 7, 4] + row_numel = 12 + + grad_selected_rows = scope.var('Grad').get_selected_rows() + grad_selected_rows.set_height(height) + grad_selected_rows.set_rows(rows) + np_array = np.ones((len(rows), row_numel)).astype("float32") + np_array[0, 0] = 2.0 + np_array[2, 8] = 4.0 + + grad_tensor = grad_selected_rows.get_tensor() + grad_tensor.set(np_array, place) + + # create and initialize Param Variable + param = scope.var('Param').get_tensor() + param_array = np.full((height, row_numel), 5.0).astype("float32") + param.set(param_array, place) + + # create and initialize LeraningRate Variable + lr = scope.var('LearningRate').get_tensor() + lr_array = np.full((1), 2.0).astype("float32") + lr.set(lr_array, place) + + # create and initialize moment Variable + moment = scope.var('Moment').get_tensor() + moment_np_array = np.full((height, row_numel), 2.0).astype("float32") + moment.set(moment_np_array, place) + + # create and run sgd operator + adagrad_op = Operator( + "adagrad", + Param='Param', + Grad='Grad', + ParamOut='Param', + Moment='Moment', + MomentOut='Moment', + LearningRate='LearningRate', + epsilon=2.0) + + ctx = core.DeviceContext.create(place) + adagrad_op.run(scope, ctx) + + # get and compare moment result + moment_result_array = np.array(moment) + + self.assertAlmostEqual(6.0, moment_result_array[rows[0], 0]) + self.assertAlmostEqual(3.0, moment_result_array[rows[0], 2]) + self.assertAlmostEqual(2.0, moment_result_array[1, 0]) + # 2.0 + (1.0 + 1.0)^2 + self.assertAlmostEqual(6.0, moment_result_array[rows[1], 10]) + self.assertAlmostEqual(6.0, moment_result_array[rows[3], 4]) + + self.assertAlmostEqual(2.0, moment_result_array[5, 8]) + self.assertAlmostEqual(3.0, moment_result_array[rows[2], 1]) + self.assertAlmostEqual(18.0, moment_result_array[rows[2], 8]) + + # get and compare param result + result_array = np.array(param) + + def get_out(param, lr, grad, m, epsilon): + return param - lr * grad / (math.sqrt(m) + epsilon) + + self.assertAlmostEqual( + get_out(5.0, 2.0, 2.0, 6.0, 2.0), + result_array[rows[0], 0], + places=5) + self.assertAlmostEqual( + get_out(5.0, 2.0, 1.0, 3.0, 2.0), + result_array[rows[0], 2], + places=5) + self.assertAlmostEqual( + get_out(5.0, 2.0, 0.0, 2.0, 2.0), result_array[1, 0], places=5) + + # grad_merge = 1.0 + 1.0 + # m = 6.0 + self.assertAlmostEqual( + get_out(5.0, 2.0, 2.0, 6.0, 2.0), + result_array[rows[1], 10], + places=5) + + self.assertAlmostEqual( + get_out(5.0, 2.0, 0.0, 2.0, 2.0), result_array[5, 8], places=5) + self.assertAlmostEqual( + get_out(5.0, 2.0, 1.0, 3.0, 2.0), + result_array[rows[2], 1], + places=5) + self.assertAlmostEqual( + get_out(5.0, 2.0, 4.0, 18.0, 2.0), + result_array[rows[2], 8], + places=5) + + def test_sparse_adagrad(self): + places = [core.CPUPlace()] + if core.is_compile_gpu(): + places.append(core.GPUPlace(0)) + for place in places: + self.check_with_place(place) + + if __name__ == "__main__": unittest.main() diff --git a/python/paddle/v2/fluid/tests/test_beam_search_op.py b/python/paddle/v2/fluid/tests/test_beam_search_op.py new file mode 100644 index 0000000000000000000000000000000000000000..cc7c09bb59de3f83e47b4d95c1203f7f050c5132 --- /dev/null +++ b/python/paddle/v2/fluid/tests/test_beam_search_op.py @@ -0,0 +1,65 @@ +import logging +from paddle.v2.fluid.op import Operator, DynamicRecurrentOp +import paddle.v2.fluid.core as core +import unittest +import numpy as np + + +def create_tensor(scope, name, np_data): + tensor = scope.var(name).get_tensor() + tensor.set(np_data, core.CPUPlace()) + return tensor + + +class BeamSearchOpTester(unittest.TestCase): + def setUp(self): + self.scope = core.Scope() + self.ctx = core.DeviceContext.create(core.CPUPlace()) + self._create_ids() + self._create_scores() + self._create_pre_ids() + self.scope.var('selected_ids') + self.scope.var('selected_scores') + + def test_run(self): + op = Operator( + 'beam_search', + pre_ids="pre_ids", + ids='ids', + scores='scores', + selected_ids='selected_ids', + selected_scores='selected_scores', + level=0, + beam_size=2, + end_id=0, ) + op.run(self.scope, self.ctx) + selected_ids = self.scope.find_var("selected_ids").get_tensor() + print 'selected_ids', np.array(selected_ids) + print 'lod', selected_ids.lod() + + def _create_pre_ids(self): + np_data = np.array([[1, 2, 3, 4]], dtype='int32') + tensor = create_tensor(self.scope, "pre_ids", np_data) + + def _create_ids(self): + self.lod = [[0, 1, 4], [0, 1, 2, 3, 4]] + np_data = np.array( + [[4, 2, 5], [2, 1, 3], [3, 5, 2], [8, 2, 1]], dtype='int32') + tensor = create_tensor(self.scope, "ids", np_data) + tensor.set_lod(self.lod) + + def _create_scores(self): + np_data = np.array( + [ + [0.5, 0.3, 0.2], + [0.6, 0.3, 0.1], + [0.9, 0.5, 0.1], + [0.7, 0.5, 0.1], + ], + dtype='float32') + tensor = create_tensor(self.scope, "scores", np_data) + tensor.set_lod(self.lod) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/v2/fluid/tests/test_conv2d_op.py b/python/paddle/v2/fluid/tests/test_conv2d_op.py index 04ae7f294c27fdceaaff2e9a7ed854213e643945..2240dc73cdd31f320fed174dd811e93c6640137f 100644 --- a/python/paddle/v2/fluid/tests/test_conv2d_op.py +++ b/python/paddle/v2/fluid/tests/test_conv2d_op.py @@ -10,23 +10,33 @@ def conv2d_forward_naive(input, filter, group, conv_param): assert np.mod(out_c, group) == 0 sub_out_c = out_c / group - stride, pad = conv_param['stride'], conv_param['pad'] - out_h = 1 + (in_h + 2 * pad[0] - f_h) / stride[0] - out_w = 1 + (in_w + 2 * pad[1] - f_w) / stride[1] + stride, pad, dilation = conv_param['stride'], conv_param['pad'], conv_param[ + 'dilation'] + out_h = 1 + (in_h + 2 * pad[0] - (dilation[0] * (f_h - 1) + 1)) / stride[0] + out_w = 1 + (in_w + 2 * pad[1] - (dilation[1] * (f_w - 1) + 1)) / stride[1] out = np.zeros((in_n, out_c, out_h, out_w)) + d_bolck_w = (dilation[0] * (f_h - 1) + 1) + d_bolck_h = (dilation[1] * (f_w - 1) + 1) + input_pad = np.pad(input, ((0, ), (0, ), (pad[0], ), (pad[1], )), mode='constant', constant_values=0) + + filter_dilation = np.zeros((out_c, f_c, d_bolck_h, d_bolck_w)) + filter_dilation[:, :, 0:d_bolck_h:dilation[0], 0:d_bolck_w:dilation[ + 1]] = filter + for i in range(out_h): for j in range(out_w): for g in range(group): input_pad_masked = \ input_pad[:, g * f_c:(g + 1) * f_c, - i * stride[0]:i * stride[0] + f_h, - j * stride[1]:j * stride[1] + f_w] + i * stride[0]:i * stride[0] + d_bolck_h, + j * stride[1]:j * stride[1] + d_bolck_w] - f_sub = filter[g * sub_out_c:(g + 1) * sub_out_c, :, :, :] + f_sub = filter_dilation[g * sub_out_c:(g + 1) * + sub_out_c, :, :, :] for k in range(sub_out_c): out[:, g * sub_out_c + k, i, j] = \ np.sum(input_pad_masked * f_sub[k, :, :, :], @@ -39,9 +49,14 @@ class TestConv2dOp(OpTest): def setUp(self): self.init_op_type() self.init_group() + self.init_dilation() self.init_test_case() - conv2d_param = {'stride': self.stride, 'pad': self.pad} + conv2d_param = { + 'stride': self.stride, + 'pad': self.pad, + 'dilation': self.dilations + } input = np.random.random(self.input_size).astype("float32") filter = np.random.random(self.filter_size).astype("float32") output = conv2d_forward_naive(input, filter, self.groups, @@ -80,12 +95,14 @@ class TestConv2dOp(OpTest): def init_test_case(self): self.pad = [0, 0] self.stride = [1, 1] - self.dilations = [1, 1] self.input_size = [2, 3, 5, 5] # NCHW assert np.mod(self.input_size[1], self.groups) == 0 f_c = self.input_size[1] / self.groups self.filter_size = [6, f_c, 3, 3] + def init_dilation(self): + self.dilations = [1, 1] + def init_group(self): self.groups = 1 @@ -93,32 +110,90 @@ class TestConv2dOp(OpTest): self.op_type = "conv2d" +class TestWithPad(TestConv2dOp): + def init_test_case(self): + self.pad = [1, 1] + self.stride = [1, 1] + self.input_size = [2, 3, 5, 5] # NCHW + assert np.mod(self.input_size[1], self.groups) == 0 + f_c = self.input_size[1] / self.groups + self.filter_size = [6, f_c, 3, 3] + + +class TestWithStride(TestConv2dOp): + def init_test_case(self): + self.pad = [1, 1] + self.stride = [2, 2] + self.input_size = [2, 3, 6, 6] # NCHW + assert np.mod(self.input_size[1], self.groups) == 0 + f_c = self.input_size[1] / self.groups + self.filter_size = [6, f_c, 3, 3] + + class TestWithGroup(TestConv2dOp): def init_group(self): self.groups = 3 - def init_op_type(self): - self.op_type = "conv2d" +class TestWith1x1(TestConv2dOp): + def init_test_case(self): + self.pad = [0, 0] + self.stride = [1, 1] + self.input_size = [2, 3, 5, 5] # NCHW + assert np.mod(self.input_size[1], self.groups) == 0 + f_c = self.input_size[1] / self.groups + self.filter_size = [6, f_c, 1, 1] -#----------------Conv2dCudnn---------------- + def init_group(self): + self.groups = 3 -class TestCudnn(TestConv2dOp): +class TestWithDilation(TestConv2dOp): + def init_test_case(self): + self.pad = [0, 0] + self.stride = [1, 1] + self.input_size = [2, 3, 10, 10] # NCHW + assert np.mod(self.input_size[1], self.groups) == 0 + f_c = self.input_size[1] / self.groups + self.filter_size = [6, f_c, 3, 3] + + def init_dilation(self): + self.dilations = [2, 2] + def init_group(self): - self.groups = 1 + self.groups = 3 + + +#----------------Conv2dCudnn---------------- +class TestCudnn(TestConv2dOp): + def init_op_type(self): + self.op_type = "conv_cudnn" + +class TestCudnnWithPad(TestWithPad): def init_op_type(self): self.op_type = "conv_cudnn" -class TestCudnnWithGroup(TestConv2dOp): - def init_group(self): - self.groups = 3 +class TestCudnnWithStride(TestWithStride): + def init_op_type(self): + self.op_type = "conv_cudnn" + +class TestCudnnWithGroup(TestWithGroup): def init_op_type(self): self.op_type = "conv_cudnn" +class TestCudnnWith1x1(TestWith1x1): + def init_op_type(self): + self.op_type = "conv_cudnn" + + +# cudnn v5 does not support dilation conv. +# class TestCudnnWithDilation(TestWithDilation): +# def init_op_type(self): +# self.op_type = "conv_cudnn" + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/v2/fluid/tests/test_conv2d_transpose_op.py b/python/paddle/v2/fluid/tests/test_conv2d_transpose_op.py index 54349c018c4a53b8767d6cd4f94d99c719dc0237..d7b1f2f2a3abf6335998742dbbef8e17794170fa 100644 --- a/python/paddle/v2/fluid/tests/test_conv2d_transpose_op.py +++ b/python/paddle/v2/fluid/tests/test_conv2d_transpose_op.py @@ -4,9 +4,7 @@ from op_test import OpTest def conv2dtranspose_forward_naive(input_, filter_, conv2dtranspose_param): - # [2, 3, 5, 5] in_n, in_c, in_h, in_w = input_.shape - # [3, 6, 3, 3] f_c, out_c, f_h, f_w = filter_.shape assert in_c == f_c @@ -29,6 +27,7 @@ def conv2dtranspose_forward_naive(input_, filter_, conv2dtranspose_param): j1, j2 = j * stride[0], j * stride[0] + f_w out[n, k, i1:i2, j1:j2] += tmp_out + out = out[:, :, pad[0]:out_h - pad[0], pad[1]:out_w - pad[1]] return out @@ -36,8 +35,6 @@ class TestConv2dTransposeOp(OpTest): def setUp(self): # init as conv transpose self.init_op_type() - - # [2, 3, 5, 5] -> kernel [3, 6, 3, 3] -> output [2, 6, 7, 7] self.init_test_case() conv2dtranspose_param = {'stride': self.stride, 'pad': self.pad} @@ -55,7 +52,6 @@ class TestConv2dTransposeOp(OpTest): self.outputs = {'Output': output} def test_check_output(self): - print 'check output here for', self.op_type self.check_output() def test_check_grad_no_input(self): @@ -88,6 +84,26 @@ class TestConv2dTransposeOp(OpTest): self.op_type = "conv2d_transpose" +class TestWithPad(TestConv2dTransposeOp): + def init_test_case(self): + self.pad = [1, 1] + self.stride = [1, 1] + self.dilations = [1, 1] + self.input_size = [2, 3, 5, 5] # NCHW + f_c = self.input_size[1] + self.filter_size = [f_c, 6, 3, 3] + + +class TestWithStride(TestConv2dTransposeOp): + def init_test_case(self): + self.pad = [1, 1] + self.stride = [2, 2] + self.dilations = [1, 1] + self.input_size = [2, 3, 5, 5] # NCHW + f_c = self.input_size[1] + self.filter_size = [f_c, 6, 3, 3] + + # ------------ test_cudnn ------------ class TestCudnn(TestConv2dTransposeOp): def init_op_type(self): diff --git a/python/paddle/v2/fluid/tests/test_conv3d_op.py b/python/paddle/v2/fluid/tests/test_conv3d_op.py index 44c192f58d25f8ddaa38d2ba7c7c19b9a5bd7dc1..934ea46437d67b78309a86a2779e0c6577399136 100644 --- a/python/paddle/v2/fluid/tests/test_conv3d_op.py +++ b/python/paddle/v2/fluid/tests/test_conv3d_op.py @@ -10,27 +10,40 @@ def conv3d_forward_naive(input, filter, group, conv_param): assert np.mod(out_c, group) == 0 sub_out_c = out_c / group - stride, pad = conv_param['stride'], conv_param['pad'] - out_d = 1 + (in_d + 2 * pad[0] - f_h) / stride[0] - out_h = 1 + (in_h + 2 * pad[1] - f_h) / stride[1] - out_w = 1 + (in_w + 2 * pad[2] - f_w) / stride[2] + stride, pad, dilation = conv_param['stride'], conv_param['pad'], conv_param[ + 'dilations'] + + out_d = 1 + (in_d + 2 * pad[0] - (dilation[0] * (f_d - 1) + 1)) / stride[0] + out_h = 1 + (in_h + 2 * pad[1] - (dilation[1] * (f_h - 1) + 1)) / stride[1] + out_w = 1 + (in_w + 2 * pad[2] - (dilation[2] * (f_w - 1) + 1)) / stride[2] + out = np.zeros((in_n, out_c, out_d, out_h, out_w)) + d_bolck_d = (dilation[0] * (f_d - 1) + 1) + d_bolck_h = (dilation[1] * (f_h - 1) + 1) + d_bolck_w = (dilation[2] * (f_w - 1) + 1) + input_pad = np.pad(input, ((0, ), (0, ), (pad[0], ), (pad[1], ), (pad[2], )), mode='constant', constant_values=0) + + filter_dilation = np.zeros((out_c, f_c, d_bolck_d, d_bolck_h, d_bolck_w)) + filter_dilation[:, :, 0:d_bolck_d:dilation[0], 0:d_bolck_h:dilation[1], 0: + d_bolck_w:dilation[2]] = filter + for d in range(out_d): for i in range(out_h): for j in range(out_w): for g in range(group): input_pad_masked = \ input_pad[:, g * f_c:(g + 1) * f_c, - d * stride[0]:d * stride[0] + f_d, - i * stride[1]:i * stride[1] + f_h, - j * stride[2]:j * stride[2] + f_w] - f_sub = filter[g * sub_out_c:(g + 1) * - sub_out_c, :, :, :, :] + d * stride[0]:d * stride[0] + d_bolck_d, + i * stride[1]:i * stride[1] + d_bolck_h, + j * stride[2]:j * stride[2] + d_bolck_w] + + f_sub = filter_dilation[g * sub_out_c:(g + 1) * + sub_out_c, :, :, :, :] for k in range(sub_out_c): out[:, g * sub_out_c + k, d, i, j] = \ np.sum(input_pad_masked * f_sub[k, :, :, :, :], @@ -43,9 +56,14 @@ class TestConv3dOp(OpTest): def setUp(self): self.init_group() self.init_op_type() + self.init_dilation() self.init_test_case() - conv3d_param = {'stride': self.stride, 'pad': self.pad} + conv3d_param = { + 'stride': self.stride, + 'pad': self.pad, + 'dilations': self.dilations + } input = np.random.random(self.input_size).astype("float32") filter = np.random.random(self.filter_size).astype("float32") output = conv3d_forward_naive(input, filter, self.groups, @@ -55,7 +73,8 @@ class TestConv3dOp(OpTest): self.attrs = { 'strides': self.stride, 'paddings': self.pad, - 'groups': self.groups + 'groups': self.groups, + 'dilations': self.dilations } self.outputs = {'Output': output} @@ -88,6 +107,9 @@ class TestConv3dOp(OpTest): f_c = self.input_size[1] / self.groups self.filter_size = [6, f_c, 3, 3, 3] + def init_dilation(self): + self.dilations = [1, 1, 1] + def init_group(self): self.groups = 1 @@ -104,27 +126,47 @@ class TestCase1(TestConv3dOp): f_c = self.input_size[1] / self.groups self.filter_size = [6, f_c, 3, 3, 3] - def init_group(self): - self.groups = 1 - def init_op_type(self): - self.op_type = "conv3d" +class TestWithGroup1(TestConv3dOp): + def init_group(self): + self.groups = 3 -class TestWithGroup1(TestConv3dOp): +class TestWithGroup2(TestCase1): def init_group(self): self.groups = 3 - def init_op_type(self): - self.op_type = "conv3d" +class TestWith1x1(TestConv3dOp): + def init_test_case(self): + self.pad = [0, 0, 0] + self.stride = [1, 1, 1] + self.input_size = [2, 3, 4, 4, 4] # NCHW + assert np.mod(self.input_size[1], self.groups) == 0 + f_c = self.input_size[1] / self.groups + self.filter_size = [6, f_c, 1, 1, 1] + + def init_dilation(self): + self.dilations = [1, 1, 1] -class TestWithGroup2(TestCase1): def init_group(self): self.groups = 3 - def init_op_type(self): - self.op_type = "conv3d" + +class TestWithDilation(TestConv3dOp): + def init_test_case(self): + self.pad = [0, 0, 0] + self.stride = [1, 1, 1] + self.input_size = [2, 3, 6, 6, 6] # NCDHW + assert np.mod(self.input_size[1], self.groups) == 0 + f_c = self.input_size[1] / self.groups + self.filter_size = [6, f_c, 2, 2, 2] + + def init_dilation(self): + self.dilations = [2, 2, 2] + + def init_group(self): + self.groups = 3 if __name__ == '__main__': diff --git a/python/paddle/v2/fluid/tests/test_conv3d_transpose_op.py b/python/paddle/v2/fluid/tests/test_conv3d_transpose_op.py index 132fe7931438a30cf02e4ad2894c0838e48ffc9f..59a32c40821f2109306e898a6a798fea52b1e0ca 100644 --- a/python/paddle/v2/fluid/tests/test_conv3d_transpose_op.py +++ b/python/paddle/v2/fluid/tests/test_conv3d_transpose_op.py @@ -4,9 +4,7 @@ from op_test import OpTest def conv3dtranspose_forward_naive(input_, filter_, conv3dtranspose_param): - # [2, 3, 5, 5, 5] in_n, in_c, in_d, in_h, in_w = input_.shape - # [3, 6, 3, 3, 3] f_c, out_c, f_d, f_h, f_w = filter_.shape assert in_c == f_c @@ -14,7 +12,6 @@ def conv3dtranspose_forward_naive(input_, filter_, conv3dtranspose_param): out_d = (in_d - 1) * stride[0] + f_d out_h = (in_h - 1) * stride[1] + f_h out_w = (in_w - 1) * stride[2] + f_w - out = np.zeros((in_n, out_c, out_d, out_h, out_w)) for n in range(in_n): @@ -33,6 +30,8 @@ def conv3dtranspose_forward_naive(input_, filter_, conv3dtranspose_param): j1, j2 = j * stride[2], j * stride[2] + f_w out[n, k, d1:d2, i1:i2, j1:j2] += tmp_out + out = out[:, :, pad[0]:out_d - pad[0], pad[1]:out_h - pad[1], pad[2]:out_w - + pad[2]] return out @@ -40,8 +39,6 @@ class TestConv3dTransposeOp(OpTest): def setUp(self): # init as conv transpose self.init_op_type() - - # [2, 3, 5, 5, 5] -> kernel [3, 6, 3, 3, 3] -> output [2, 6, 7, 7, 7] self.init_test_case() conv3dtranspose_param = {'stride': self.stride, 'pad': self.pad} @@ -49,7 +46,6 @@ class TestConv3dTransposeOp(OpTest): filter_ = np.random.random(self.filter_size).astype("float32") output = conv3dtranspose_forward_naive( input_, filter_, conv3dtranspose_param).astype("float32") - # print 'deconv output py', output, output.shape self.inputs = {'Input': input_, 'Filter': filter_} self.attrs = { @@ -60,7 +56,6 @@ class TestConv3dTransposeOp(OpTest): self.outputs = {'Output': output} def test_check_output(self): - print 'check output here' self.check_output() def test_check_grad(self): @@ -85,7 +80,7 @@ class TestConv3dTransposeOp(OpTest): self.pad = [0, 0, 0] self.stride = [1, 1, 1] self.dilations = [1, 1, 1] - self.input_size = [2, 3, 5, 5, 5] # NCHW + self.input_size = [2, 3, 5, 5, 5] # NCDHW f_c = self.input_size[1] self.filter_size = [f_c, 6, 3, 3, 3] @@ -93,5 +88,25 @@ class TestConv3dTransposeOp(OpTest): self.op_type = "conv3d_transpose" +class TestWithPad(TestConv3dTransposeOp): + def init_test_case(self): + self.pad = [1, 1, 1] + self.stride = [1, 1, 1] + self.dilations = [1, 1, 1] + self.input_size = [2, 3, 5, 5, 5] # NCDHW + f_c = self.input_size[1] + self.filter_size = [f_c, 6, 3, 3, 3] + + +class TestWithStride(TestConv3dTransposeOp): + def init_test_case(self): + self.pad = [1, 1, 1] + self.stride = [2, 2, 2] + self.dilations = [1, 1, 1] + self.input_size = [2, 3, 5, 5, 5] # NCDHW + f_c = self.input_size[1] + self.filter_size = [f_c, 6, 3, 3, 3] + + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/v2/fluid/tests/test_parameter.py b/python/paddle/v2/fluid/tests/test_parameter.py index 71a1bd2aaf5a9c6362ce0d35c256ed228e942fce..a633d22c2b1db2728b6eb767078ce4aec6cce163 100644 --- a/python/paddle/v2/fluid/tests/test_parameter.py +++ b/python/paddle/v2/fluid/tests/test_parameter.py @@ -1,26 +1,32 @@ import unittest from paddle.v2.fluid.framework import g_main_program import paddle.v2.fluid.core as core +from paddle.v2.fluid.executor import Executor +import paddle.v2.fluid.io as io +from paddle.v2.fluid.initializer import ConstantInitializer +import numpy as np class TestParameter(unittest.TestCase): def test_param(self): - b = g_main_program.create_block() + shape = [784, 100] + val = 1.0625 + b = g_main_program.global_block() param = b.create_parameter( name='fc.w', - shape=[784, 100], + shape=shape, dtype='float32', - initialize_attr={ - 'type': 'uniform_random', - 'seed': 13, - 'min': -5.0, - 'max': 5.0 - }) + initializer=ConstantInitializer(val)) self.assertIsNotNone(param) self.assertEqual('fc.w', param.name) self.assertEqual((784, 100), param.shape) self.assertEqual(core.DataType.FP32, param.data_type) self.assertEqual(0, param.block.idx) + exe = Executor(core.CPUPlace()) + p = exe.run(g_main_program, fetch_list=[param])[0] + self.assertTrue(np.allclose(np.array(p), np.ones(shape) * val)) + p = io.get_parameter_value_by_name('fc.w', exe, g_main_program) + self.assertTrue(np.allclose(np.array(p), np.ones(shape) * val)) if __name__ == '__main__': diff --git a/python/paddle/v2/fluid/tests/test_while_op.py b/python/paddle/v2/fluid/tests/test_while_op.py index 0f01acb3b94dc55a3536e751108e785ddc6e47bb..84b432333f950f754a97bc1a051b59c16fb22aed 100644 --- a/python/paddle/v2/fluid/tests/test_while_op.py +++ b/python/paddle/v2/fluid/tests/test_while_op.py @@ -2,6 +2,7 @@ import unittest import paddle.v2.fluid.layers as layers from paddle.v2.fluid.executor import Executor import paddle.v2.fluid.core as core +from paddle.v2.fluid.backward import append_backward_ops import numpy @@ -16,7 +17,7 @@ class TestWhileOp(unittest.TestCase): i = layers.zeros(shape=[1], dtype='int64') i.stop_gradient = True init = layers.zeros(shape=[10], dtype='float32') - mem_array = layers.array_write(init, i=i) + mem_array = layers.array_write(x=init, i=i) data_array = layers.array_write(x=d0, i=i) i = layers.increment(i) @@ -29,17 +30,23 @@ class TestWhileOp(unittest.TestCase): i.stop_gradient = True array_len = layers.fill_constant(shape=[1], dtype='int64', value=3) + array_len.stop_gradient = True cond = layers.less_than(x=i, y=array_len) while_op = layers.While(cond=cond) with while_op.block(): d = layers.array_read(array=data_array, i=i) prev = layers.array_read(array=mem_array, i=i) - i = layers.increment(x=i, in_place=True) result = layers.sums(input=[d, prev]) + + i = layers.increment(x=i, in_place=True) layers.array_write(result, i=i, array=mem_array) layers.less_than(x=i, y=array_len, cond=cond) - sum_result = layers.array_read(mem_array, i=array_len) + + sum_result = layers.array_read(array=mem_array, i=i) + loss = layers.mean(x=sum_result) + + append_backward_ops(loss) cpu = core.CPUPlace() exe = Executor(cpu) diff --git a/python/paddle/v2/framework/tests/test_is_empty_op.py b/python/paddle/v2/framework/tests/test_is_empty_op.py new file mode 100644 index 0000000000000000000000000000000000000000..129d1c19447990fb0affa8fb10bc7156ec5c8cc3 --- /dev/null +++ b/python/paddle/v2/framework/tests/test_is_empty_op.py @@ -0,0 +1,43 @@ +import unittest +import numpy as np +from paddle.v2.framework.op import Operator +import paddle.v2.framework.core as core + + +def create_tensor(scope, name, np_data): + tensor = scope.var(name).get_tensor() + tensor.set_dims(np_data.shape) + tensor.set(np_data, core.CPUPlace()) + return tensor + + +class TestIsEmptyOp(unittest.TestCase): + def setUp(self): + self.scope = core.Scope() + # create input variables + np_data0 = np.array([0, 1, 2]) + create_tensor(self.scope, "X0", np_data0) + + np_data1 = np.array([1]) + t = create_tensor(self.scope, "X1", np_data1) + t.set_dims([0]) + + # create output variables + self.scope.var("out") + + def test_no_empty(self): + self.one_case("X0", False) + + def test_empty(self): + self.one_case("X1", True) + + def one_case(self, input, target): + op = Operator(type="is_empty", X=input, Out="out") + ctx = core.DeviceContext.create(core.CPUPlace()) + op.run(self.scope, ctx) + out = self.scope.var("out").get_tensor() + self.assertEqual(np.array(out)[0], target) + + +if __name__ == "__main__": + unittest.main() diff --git a/python/paddle/v2/model.py b/python/paddle/v2/model.py deleted file mode 100644 index 4634db55a919584db91e456e61d393b9e15129ac..0000000000000000000000000000000000000000 --- a/python/paddle/v2/model.py +++ /dev/null @@ -1,73 +0,0 @@ -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# 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. - -import os -import errno -import uuid - -import paddle.v2.master - -__all__ = ["save_model", "load_model"] - -trainer_id = str(uuid.uuid4()) - - -def mkdir_p(path): - try: - os.makedirs(path) - except OSError as exc: - if exc.errno == errno.EEXIST and os.path.isdir(path): - pass - else: - raise - - -def save_model(parameters, path): - need_request = "KUBERNETES_SERVICE_HOST" in os.environ.keys() - - if need_request: - # TODO(helin): figure out how MPI trains, since MPI only save - # model when trainer_id == "0", we can consolidate the logic - # here. - - # TODO(helin): change this environment variable name from - # MASTER_IP to ETCD_IP - etcd_name = "MASTER_IP" - if etcd_name not in os.environ.keys(): - raise Exception('not find ' + etcd_name + - ' in environment variable.') - - etcd_ip = os.environ.get(etcd_name) - client = paddle.v2.master.client("http://" + etcd_ip + ":2379", 5, 0) - r = client.request_save_model(trainer_id, 5000) - if r == 0: - # do not need to save - return - elif r < 0: - # error - return - else: - # save model - path = os.path.join(path, trainer_id) - path = os.path.join(path, "model.tar") - - mkdir_p(path) - - with open(path, 'wb') as f: - parameters.to_tar(f) - - -def load_model(parameters, path): - with open(path, 'rb') as f: - parameters.from_tar(f)