diff --git a/CMakeLists.txt b/CMakeLists.txt
index fd3582a1bca199d62d19550ffdd1efe9db520fa7..9e30dff70fed51b604059610b22057349f22db58 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -36,8 +36,7 @@ include(simd)
################################ Configurations #######################################
option(WITH_GPU "Compile PaddlePaddle with NVIDIA GPU" ${CUDA_FOUND})
option(WITH_AVX "Compile PaddlePaddle with AVX intrinsics" ${AVX_FOUND})
-option(WITH_MKLDNN "Compile PaddlePaddle with mkl-dnn support." ${AVX_FOUND})
-option(WITH_MKLML "Compile PaddlePaddle with mklml package." ${AVX_FOUND})
+option(WITH_MKL "Compile PaddlePaddle with MKL support." ${AVX_FOUND})
option(WITH_DSO "Compile PaddlePaddle with dynamic linked CUDA" ON)
option(WITH_TESTING "Compile PaddlePaddle with unit testing" ON)
option(WITH_SWIG_PY "Compile PaddlePaddle with inference api" ON)
@@ -82,10 +81,8 @@ if(ANDROID OR IOS)
"Disable PYTHON when cross-compiling for Android and iOS" FORCE)
set(WITH_RDMA OFF CACHE STRING
"Disable RDMA when cross-compiling for Android and iOS" FORCE)
- set(WITH_MKLDNN OFF CACHE STRING
- "Disable MKLDNN when cross-compiling for Android and iOS" FORCE)
- set(WITH_MKLML OFF CACHE STRING
- "Disable MKLML package when cross-compiling for Android and iOS" FORCE)
+ set(WITH_MKL OFF CACHE STRING
+ "Disable MKL when cross-compiling for Android and iOS" FORCE)
# Compile PaddlePaddle mobile inference library
if (NOT WITH_C_API)
@@ -111,6 +108,17 @@ else()
set(THIRD_PARTY_BUILD_TYPE Release)
endif()
+if(WITH_MKL)
+ set(WITH_MKLML ON)
+ set(WITH_MKLDNN ${AVX2_FOUND})
+ if(NOT WITH_MKLDNN)
+ message(WARNING "Do not have AVX2 intrinsics and disabled MKL-DNN")
+ endif()
+else()
+ set(WITH_MKLML OFF)
+ set(WITH_MKLDNN OFF)
+endif()
+
########################################################################################
include(external/mklml) # download mklml package
@@ -164,8 +172,12 @@ if(WITH_GPU)
endif(NOT WITH_DSO)
endif(WITH_GPU)
+if(WITH_MKLML)
+ list(APPEND EXTERNAL_LIBS ${MKLML_IOMP_LIB})
+endif()
+
if(WITH_MKLDNN)
- list(APPEND EXTERNAL_LIBS ${MKLDNN_LIB} ${MKLDNN_IOMP_LIB})
+ list(APPEND EXTERNAL_LIBS ${MKLDNN_LIB})
endif()
if(USE_NNPACK)
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 $3, use True or False."
diff --git a/cmake/configure.cmake b/cmake/configure.cmake
index 24ddb24399dabeec9b8e5faf36be3eb21f420111..e550ec285668ea25757eeee9e7c5dc48fc9d339d 100644
--- a/cmake/configure.cmake
+++ b/cmake/configure.cmake
@@ -76,27 +76,14 @@ else()
include_directories(${CUDA_TOOLKIT_INCLUDE})
endif(NOT WITH_GPU)
-if(WITH_MKLDNN)
- add_definitions(-DPADDLE_USE_MKLDNN)
- if (WITH_MKLML AND MKLDNN_IOMP_DIR)
- message(STATUS "Enable Intel OpenMP at ${MKLDNN_IOMP_DIR}")
- set(OPENMP_FLAGS "-fopenmp")
- set(CMAKE_C_CREATE_SHARED_LIBRARY_FORBIDDEN_FLAGS ${OPENMP_FLAGS})
- set(CMAKE_CXX_CREATE_SHARED_LIBRARY_FORBIDDEN_FLAGS ${OPENMP_FLAGS})
- set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${OPENMP_FLAGS}")
- set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${OPENMP_FLAGS}")
- else()
- find_package(OpenMP)
- if(OPENMP_FOUND)
- set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${OpenMP_C_FLAGS}")
- set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${OpenMP_CXX_FLAGS}")
- else()
- message(WARNING "Can not find OpenMP."
- "Some performance features in MKLDNN may not be available")
- endif()
- endif()
-
-endif(WITH_MKLDNN)
+if (WITH_MKLML AND MKLML_IOMP_LIB)
+ message(STATUS "Enable Intel OpenMP with ${MKLML_IOMP_LIB}")
+ set(OPENMP_FLAGS "-fopenmp")
+ set(CMAKE_C_CREATE_SHARED_LIBRARY_FORBIDDEN_FLAGS ${OPENMP_FLAGS})
+ set(CMAKE_CXX_CREATE_SHARED_LIBRARY_FORBIDDEN_FLAGS ${OPENMP_FLAGS})
+ set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${OPENMP_FLAGS}")
+ set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${OPENMP_FLAGS}")
+endif()
set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${SIMD_FLAG}")
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${SIMD_FLAG}")
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/mkldnn.cmake b/cmake/external/mkldnn.cmake
index 5a06825beb73e85d8a55b7b578b187bee2c4340c..fc52d339d7a336b44c97f2e0a9fc8d6604854365 100644
--- a/cmake/external/mkldnn.cmake
+++ b/cmake/external/mkldnn.cmake
@@ -40,10 +40,9 @@ INCLUDE_DIRECTORIES(${MKLDNN_INC_DIR})
IF(${CBLAS_PROVIDER} STREQUAL "MKLML")
SET(MKLDNN_DEPENDS ${MKLML_PROJECT})
- SET(MKLDNN_MKLROOT ${MKLML_ROOT})
- SET(MKLDNN_IOMP_LIB ${MKLML_IOMP_LIB})
- SET(MKLDNN_IOMP_DIR ${MKLML_LIB_DIR})
- MESSAGE(STATUS "Build MKLDNN with ${MKLDNN_MKLROOT}")
+ MESSAGE(STATUS "Build MKLDNN with MKLML ${MKLML_ROOT}")
+ELSE()
+ MESSAGE(FATAL_ERROR "Should enable MKLML when build MKLDNN")
ENDIF()
SET(MKLDNN_CFLAG "${CMAKE_C_FLAGS} -Wno-error=strict-overflow")
@@ -57,15 +56,16 @@ ExternalProject_Add(
PREFIX ${MKLDNN_SOURCES_DIR}
UPDATE_COMMAND ""
CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${MKLDNN_INSTALL_DIR}
- CMAKE_ARGS -DMKLROOT=${MKLDNN_MKLROOT}
+ CMAKE_ARGS -DMKLROOT=${MKLML_ROOT}
CMAKE_ARGS -DCMAKE_C_FLAGS=${MKLDNN_CFLAG}
CMAKE_ARGS -DCMAKE_CXX_FLAGS=${MKLDNN_CXXFLAG}
CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${MKLDNN_INSTALL_DIR}
- -DMKLROOT:PATH=${MKLDNN_MKLROOT}
+ -DMKLROOT:PATH=${MKLML_ROOT}
)
ADD_LIBRARY(mkldnn SHARED IMPORTED GLOBAL)
SET_PROPERTY(TARGET mkldnn PROPERTY IMPORTED_LOCATION ${MKLDNN_LIB})
ADD_DEPENDENCIES(mkldnn ${MKLDNN_PROJECT})
-MESSAGE(STATUS "Mkldnn library: ${MKLDNN_LIB}")
+MESSAGE(STATUS "MKLDNN library: ${MKLDNN_LIB}")
+add_definitions(-DPADDLE_USE_MKLDNN)
LIST(APPEND external_project_dependencies mkldnn)
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 ${THIRD_PARTY_PATH}/warpctc)
diff --git a/cmake/util.cmake b/cmake/util.cmake
index 117ab7f49cdf4a568cd203b2b17767643d0b2d50..ad905ab55ba3537054fa5b30b5fca4d83c406702 100644
--- a/cmake/util.cmake
+++ b/cmake/util.cmake
@@ -115,8 +115,8 @@ function(link_paddle_exe TARGET_NAME)
target_link_libraries(${TARGET_NAME} log)
endif(ANDROID)
- if(WITH_MKLDNN AND WITH_MKLML AND MKLDNN_IOMP_DIR)
- target_link_libraries(${TARGET_NAME} "-L${MKLDNN_IOMP_DIR} -liomp5 -Wl,--as-needed")
+ if(WITH_MKLML AND MKLML_LIB_DIR AND MKLML_IOMP_LIB)
+ target_link_libraries(${TARGET_NAME} "-L${MKLML_LIB_DIR} -liomp5 -Wl,--as-needed")
endif()
add_dependencies(${TARGET_NAME} ${external_project_dependencies})
diff --git a/doc/design/mkldnn/README.MD b/doc/design/mkldnn/README.MD
index 16236763a73770f3fe5eadf67645765d0456f875..ec6d4681836e189f46dbb9b915a237dc15cda7cf 100644
--- a/doc/design/mkldnn/README.MD
+++ b/doc/design/mkldnn/README.MD
@@ -36,13 +36,13 @@ Figure 1. PaddlePaddle on IA.
我们把集成方案大致分为了如下几个方面。
### CMake
-我们会在`CMakeLists.txt`中会添加`WITH_MKLDNN`的选项，当设置这个值为`ON`的时候会启用编译MKL-DNN功能。同时会自动开启OpenMP用于提高MKL-DNN的性能。
+我们会在`CMakeLists.txt`中会给用户添加一个`WITH_MKL`的开关，他是负责`WITH_MKLML`和`WITH_MKLDNN`的总开关。
-同时，我们会引入`WITH_MKLML`选项，用于选择是否使用MKL-DNN自带的MKLML安装包。这个安装包可以独立于MKL-DNN使用，但是建议在开启MKL-DNN的同时也打开MKLML的开关，这样才能发挥最好的性能。
+当打开`WITH_MKL`时，会开启MKLML的功能，作为PaddlePaddle的CBLAS和LAPACK库，同时会开启Intel OpenMP用于提高MKLML的性能。 如果系统支持AVX2指令集及以上，同时会开启MKL-DNN功能。
-所以，我们会在`cmake/external`目录新建`mkldnn.cmake`和`mklml.cmake`文件，它们会在编译PaddlePaddle的时候下载对应的软件包，并放到PaddlePaddle的third party目录中。
+当关闭`WITH_MKL`时，MKLML和MKL-DNN功能会同时关闭。
-**备注**：当`WITH_MKLML=ON`的时候，会优先使用这个包作为PaddlePaddle的CBLAS和LAPACK库，所以会稍微改动`cmake/cblas.cmake`中的逻辑。
+所以，我们会在`cmake/external`目录新建`mkldnn.cmake`和`mklml.cmake`文件，它们会在编译PaddlePaddle的时候下载对应的软件包，并放到PaddlePaddle的third party目录中。
### Layers
所有MKL-DNN相关的C++ layers，都会按照PaddlePaddle的目录结构存放在
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/howto/dev/write_docs_cn.rst b/doc/howto/dev/write_docs_cn.rst
index 731a63f945c29ba78538b3d71289b234e569354d..61f3a223547b352cf7929615cf3682b29b9a738f 100644
--- a/doc/howto/dev/write_docs_cn.rst
+++ b/doc/howto/dev/write_docs_cn.rst
@@ -34,7 +34,7 @@ PaddlePaddle的文档构建有两种方式。
cd TO_YOUR_PADDLE_CLONE_PATH
mkdir -p build
cd build
- cmake .. -DCMAKE_BUILD_TYPE=Debug -DWITH_GPU=OFF -DWITH_MKLDNN=OFF -DWITH_MKLML=OFF -DWITH_DOC=ON
+ cmake .. -DCMAKE_BUILD_TYPE=Debug -DWITH_GPU=OFF -DWITH_MKL=OFF -DWITH_DOC=ON
make gen_proto_py
make paddle_docs paddle_docs_cn
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：
| OS |
- armv7, armv7s, arm64 (默认) |
+ armv7, armv7s, arm64 |
| SIMULATOR |
- i386, 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/gserver/CMakeLists.txt b/paddle/gserver/CMakeLists.txt
index 91d732641a4a5eed050841b59fd10da397eb732f..41ead3c5ecef248830cfb0f8be360f21dcd58e7b 100644
--- a/paddle/gserver/CMakeLists.txt
+++ b/paddle/gserver/CMakeLists.txt
@@ -73,7 +73,6 @@ if(MOBILE_INFERENCE)
list(REMOVE_ITEM GSERVER_SOURCES
dataproviders/DataProvider.cpp
dataproviders/MultiDataProvider.cpp
- dataproviders/ProtoDataProvider.cpp
dataproviders/PyDataProvider2.cpp
dataproviders/PyDataProvider.cpp)
diff --git a/paddle/gserver/dataproviders/DataProvider.cpp b/paddle/gserver/dataproviders/DataProvider.cpp
index 0478256f9cd81f4a99eb0cbcbd1a5a21de5cf14b..106cf5b6228e636026ded558d0f591022f1ae586 100644
--- a/paddle/gserver/dataproviders/DataProvider.cpp
+++ b/paddle/gserver/dataproviders/DataProvider.cpp
@@ -16,8 +16,8 @@ limitations under the License. */
#include
#include
-#include "ProtoDataProvider.h"
#include "paddle/utils/Logging.h"
+#include "paddle/utils/Stat.h"
#include "paddle/utils/StringUtil.h"
#include "paddle/utils/Util.h"
@@ -164,8 +164,6 @@ DataProvider* DataProvider::create(const DataConfig& config,
REGISTER_DATA_PROVIDER(simple, SimpleDataProvider);
REGISTER_DATA_PROVIDER(dummy, DummyDataProvider);
-REGISTER_DATA_PROVIDER(proto, ProtoDataProvider);
-REGISTER_DATA_PROVIDER(proto_sequence, ProtoSequenceDataProvider);
int64_t DataProvider::getNextBatch(int64_t size, DataBatch* batch) {
int64_t batchSize = doubleBuffer_ ? getNextBatchFromBuffer(size, batch)
diff --git a/paddle/gserver/dataproviders/ProtoDataProvider.cpp b/paddle/gserver/dataproviders/ProtoDataProvider.cpp
deleted file mode 100644
index c6f5cab1915b7f41d505c37a7fef762a392bad7f..0000000000000000000000000000000000000000
--- a/paddle/gserver/dataproviders/ProtoDataProvider.cpp
+++ /dev/null
@@ -1,932 +0,0 @@
-/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
- http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License. */
-
-#include "ProtoDataProvider.h"
-#include
-#include
-#include
-#include "paddle/utils/StringUtil.h"
-#include "paddle/utils/Util.h"
-
-#include "DataProviderGroup.h"
-#include "paddle/utils/Logging.h"
-
-DEFINE_double(memory_threshold_on_load_data,
- 1.0,
- "stop loading data when memory is not sufficient");
-
-namespace paddle {
-
-REGISTER_DATA_PROVIDER(proto_group, DataProviderGroup);
-REGISTER_DATA_PROVIDER(proto_sequence_group,
- DataProviderGroup);
-
-ProtoDataProvider::ProtoDataProvider(const DataConfig& config,
- bool useGpu,
- bool loadDataAll)
- : DataProvider(config, useGpu), sampleNums_(0), currentSequenceIndex_(0) {
- if (loadDataAll) {
- loadData(config_.files());
- }
-}
-
-void ProtoDataProvider::loadData(const std::vector& fileList) {
- for (auto& file : fileList) {
- if (FLAGS_memory_threshold_on_load_data < 1.0) {
- double memUsage = getMemoryUsage();
- if (memUsage > FLAGS_memory_threshold_on_load_data) {
- LOG(INFO) << "memUsage is " << memUsage << ", > "
- << FLAGS_memory_threshold_on_load_data
- << " therefore SKIP ALL REMAINING file.";
- break;
- }
- }
- LOG(INFO) << "load data file " << file;
- loadDataFile(file);
- }
-
- if (sequenceStartPositions_.size() == sampleNums_) {
- // This means that each sample is one sequence
- shuffledSequenceIds_.swap(sequenceStartPositions_);
- } else {
- sequenceStartPositions_.push_back(sampleNums_);
- shuffledSequenceIds_.reserve(sequenceStartPositions_.size() - 1);
- for (size_t i = 0; i < sequenceStartPositions_.size() - 1; ++i) {
- shuffledSequenceIds_.push_back(i);
- }
- }
-
- LOG(INFO) << "read done, num of instance=" << sampleNums_;
- showDataStats();
-}
-
-void ProtoDataProvider::loadData(const std::string& fileName) {
- std::vector fileList;
- loadFileList(fileName, fileList);
- loadData(fileList);
-}
-
-void ProtoDataProvider::checkDataHeader(const DataHeader& header) {
- if (header_.slot_defs_size()) {
- // header_ is already set. Need to check consistency.
- CHECK_EQ(header_.slot_defs_size(), header.slot_defs_size())
- << "Different header";
- for (int i = 0; i < header.slot_defs_size(); ++i) {
- CHECK_EQ(header_.slot_defs(i).type(), header.slot_defs(i).type());
- CHECK_EQ(header_.slot_defs(i).dim(), header.slot_defs(i).dim());
- }
- return;
- }
-
- // header_ is not set before
- CHECK(header.slot_defs_size()) << "Invalid header: no slot is defined";
- int i;
- for (i = 0; i < header.slot_defs_size(); ++i) {
- if (header.slot_defs(i).type() == SlotDef::INDEX ||
- header.slot_defs(i).type() == SlotDef::VAR_MDIM_INDEX) {
- break;
- }
- constexpr int kBufLen = 100;
- char buf[kBufLen];
- snprintf(buf, kBufLen, "slot%d_nnz", i);
- nnzStats_.push_back(getStat(buf));
- }
- numVecSlots_ = i;
-
- // Check that INDEX slots are after VECTOR slots
- for (int i = numVecSlots_; i < header.slot_defs_size(); ++i) {
- CHECK(header.slot_defs(i).type() == SlotDef::INDEX ||
- header.slot_defs(i).type() == SlotDef::VAR_MDIM_INDEX);
- }
-
- slots_.clear();
- slots_.reserve(header.slot_defs_size());
- for (int i = 0; i < header.slot_defs_size(); ++i) {
- slots_.emplace_back();
- slots_.back().type = header.slot_defs(i).type();
- slots_.back().dim = header.slot_defs(i).dim();
- if (SlotDef::VECTOR_SPARSE_NON_VALUE == header.slot_defs(i).type() ||
- SlotDef::VECTOR_SPARSE_VALUE == header.slot_defs(i).type()) {
- slots_.back().indices.push_back(0);
- }
- }
-
- header_ = header;
-}
-
-void ProtoDataProvider::checkSample(const DataSample& sample) {
- CHECK_EQ(numVecSlots_, sample.vector_slots_size());
- CHECK(header_.slot_defs_size() == numVecSlots_ + sample.id_slots_size() ||
- header_.slot_defs_size() == numVecSlots_ + sample.var_id_slots_size());
- for (int i = 0; i < numVecSlots_; ++i) {
- uint32_t dim = header_.slot_defs(i).dim();
- switch (header_.slot_defs(i).type()) {
- case SlotDef::VECTOR_DENSE: {
- CHECK_EQ(static_cast(dim), sample.vector_slots(i).values_size());
- CHECK_EQ(0, sample.vector_slots(i).ids_size());
- break;
- }
- case SlotDef::VECTOR_SPARSE_NON_VALUE: {
- if (0 == sample.vector_slots(i).ids_size()) {
- break;
- }
- CHECK_LT(0, sample.vector_slots(i).ids_size());
- CHECK_EQ(0, sample.vector_slots(i).values_size());
- auto maxId = *std::max_element(sample.vector_slots(i).ids().begin(),
- sample.vector_slots(i).ids().end());
- CHECK_GT(dim, maxId);
- break;
- }
- case SlotDef::VECTOR_SPARSE_VALUE: {
- if (0 == sample.vector_slots(i).ids_size()) {
- CHECK_EQ(0, sample.vector_slots(i).values_size());
- break;
- }
- CHECK_LT(0, sample.vector_slots(i).values_size());
- CHECK_GE(static_cast(dim), sample.vector_slots(i).values_size());
- CHECK_EQ(sample.vector_slots(i).values_size(),
- sample.vector_slots(i).ids_size());
- auto maxId = *std::max_element(sample.vector_slots(i).ids().begin(),
- sample.vector_slots(i).ids().end());
- CHECK_GT(dim, maxId);
- break;
- }
- case SlotDef::VAR_MDIM_DENSE: {
- if (static_cast(dim) != 0) {
- CHECK_EQ(static_cast(dim), sample.vector_slots(i).values_size());
- if (sample.vector_slots(i).dims_size() != 0) {
- int totalDim = sample.vector_slots(i).dims(0);
- for (int j = 1; j < sample.vector_slots(i).dims_size(); ++j) {
- totalDim *= sample.vector_slots(i).dims(j);
- }
- CHECK_EQ(static_cast(dim), totalDim);
- }
- } else {
- CHECK_NE(sample.vector_slots(i).dims_size(), 0);
- int totalDim = sample.vector_slots(i).dims(0);
- for (int j = 1; j < sample.vector_slots(i).dims_size(); ++j) {
- totalDim *= sample.vector_slots(i).dims(j);
- }
- CHECK_EQ(totalDim, sample.vector_slots(i).values_size());
- }
- break;
- }
- case SlotDef::STRING: {
- CHECK_EQ(static_cast(1), sample.vector_slots(i).strs_size());
- CHECK_EQ(0, sample.vector_slots(i).ids_size());
- CHECK_EQ(0, sample.vector_slots(i).values_size());
- break;
- }
- default:
- LOG(FATAL) << "BUG: Should not reach here";
- }
- }
- for (int i = numVecSlots_; i < header_.slot_defs_size(); ++i) {
- if (header_.slot_defs(i).type() != SlotDef::VAR_MDIM_INDEX) {
- uint32_t id = sample.id_slots(i - numVecSlots_);
- if (id == -1U) continue;
- CHECK_LT(id, header_.slot_defs(i).dim());
- } else {
- for (int j = 0; j < sample.var_id_slots(i - numVecSlots_).ids_size();
- ++j) {
- uint32_t id = sample.var_id_slots(i - numVecSlots_).ids(j);
- CHECK_LT(id, header_.slot_defs(i).dim());
- }
- }
- }
-}
-
-void ProtoDataProvider::loadDataFile(const std::string& fileName) {
- std::ifstream is(fileName);
- CHECK(is) << "Fail to open " << fileName;
- bool dataCompression = str::endsWith(fileName, ".gz");
- std::unique_ptr reader(new ProtoReader(&is, dataCompression));
- CHECK(reader) << "Fail to create proto data input stream";
-
- DataHeader header;
- CHECK(reader->read(&header));
- checkDataHeader(header);
-
- DataSample sample;
- do {
- if (!reader->read(&sample)) {
- break;
- }
- checkSample(sample);
- if (sample.is_beginning()) {
- sequenceStartPositions_.push_back(sampleNums_);
- }
- fillSlots(sample);
- ++sampleNums_;
- } while (true);
-
- CHECK(is.eof()) << "Fail to read file";
- reader.reset(nullptr);
- is.close();
-}
-
-// checkSample has done before, no check here
-void ProtoDataProvider::fillSlots(const DataSample& sample) {
- for (size_t i = 0; i < slots_.size(); ++i) {
- auto& slot = slots_[i];
- int dim = slot.dim;
- switch (slot.type) {
- case SlotDef::VECTOR_DENSE: {
- size_t oldSize = slot.denseData.size();
- slot.denseData.resize(oldSize + dim);
- const float* values = sample.vector_slots(i).values().data();
-#ifdef PADDLE_TYPE_DOUBLE
- std::copy(values, values + dim, slot.denseData.begin() + oldSize);
-#else
- memcpy(slot.denseData.data() + oldSize, values, sizeof(real) * dim);
-#endif
- break;
- }
- case SlotDef::VECTOR_SPARSE_NON_VALUE: {
- int slotSize = sample.vector_slots(i).ids_size();
- int subSlotSize = 0;
- int id = 0; // the slot id
- // find whether this vector_slots has subseq. If not has subseq,
- // subSlotSize = 0.
- for (id = 0; id < sample.subseq_slots_size(); id++) {
- if (sample.subseq_slots(id).slot_id() == i) {
- subSlotSize = sample.subseq_slots(id).lens_size();
- break;
- }
- }
- if (subSlotSize && slot.subIndices.size() == 0UL) {
- // If has subSeq, the first element of subIndices = 0.
- slot.subIndices.push_back(0);
- }
- if (slotSize == 0UL) {
- // if has no id, new indices = old indices.
- slot.indices.push_back(slot.indices.back());
- // if has subSeq, new subIndices = old subIndices.
- if (slot.subIndices.size()) {
- slot.subIndices.push_back(slot.subIndices.back());
- }
- break;
- }
- slot.sparseNonValueData.resize(slot.indices.back() + slotSize);
- const unsigned int* ids = sample.vector_slots(i).ids().data();
- memcpy(slot.sparseNonValueData.data() + slot.indices.back(),
- ids,
- sizeof(*ids) * slotSize);
- slot.indices.push_back(slot.indices.back() + slotSize);
- if (subSlotSize) {
- for (int ii = 0; ii < subSlotSize; ++ii) {
- slot.subIndices.push_back(slot.subIndices.back() +
- sample.subseq_slots(id).lens(ii));
- }
- }
- break;
- }
- case SlotDef::VECTOR_SPARSE_VALUE: {
- if (0 == sample.vector_slots(i).ids_size()) {
- slot.indices.push_back(slot.indices.back());
- break;
- }
- int slotSize = sample.vector_slots(i).ids_size();
- slot.sparseFloatValueData.resize(slot.indices.back() + slotSize);
- const unsigned int* ids = sample.vector_slots(i).ids().data();
- const float* values = sample.vector_slots(i).values().data();
- for (int ii = 0; ii < slotSize; ++ii) {
- slot.sparseFloatValueData[slot.indices.back() + ii].col = ids[ii];
- slot.sparseFloatValueData[slot.indices.back() + ii].value =
- values[ii];
- }
- slot.indices.push_back(slot.indices.back() + slotSize);
- break;
- }
- case SlotDef::INDEX: {
- slot.indexData.push_back(sample.id_slots(i - numVecSlots_));
- break;
- }
- case SlotDef::VAR_MDIM_DENSE: {
- size_t oldSize = slot.varDenseData.size();
- slot.varDenseData.resize(oldSize + 1);
- size_t varDim = sample.vector_slots(i).values_size();
- slot.varDenseData[oldSize].data.resize(varDim);
- const float* values = sample.vector_slots(i).values().data();
-#ifdef PADDLE_TYPE_DOUBLE
- std::copy(
- values, values + varDim, slot.varDenseData[oldSize].data.data());
-#else
- memcpy(slot.varDenseData[oldSize].data.data(),
- values,
- sizeof(real) * varDim);
-#endif
- slot.varDenseData[oldSize].dims.resize(
- sample.vector_slots(i).dims_size());
- memcpy(slot.varDenseData[oldSize].dims.data(),
- sample.vector_slots(i).dims().data(),
- sizeof(uint32_t) * sample.vector_slots(i).dims_size());
- break;
- }
- case SlotDef::VAR_MDIM_INDEX: {
- size_t oldSize = slot.varIndices.size();
- slot.varIndices.resize(oldSize + 1);
- size_t varDim = sample.var_id_slots(i - numVecSlots_).ids_size();
- slot.varIndices[oldSize].resize(varDim);
- memcpy(slot.varIndices[oldSize].data(),
- sample.var_id_slots(i - numVecSlots_).ids().data(),
- sizeof(uint32_t) * varDim);
- break;
- }
- case SlotDef::STRING: {
- slot.strData.push_back(sample.vector_slots(i).strs(0));
- break;
- }
- }
- }
-}
-
-void ProtoDataProvider::showDataStats() {
- std::ostringstream oss;
- for (size_t i = 0; i < slots_.size(); ++i) {
- auto& slot = slots_[i];
- if (slot.type == SlotDef::VECTOR_SPARSE_NON_VALUE) {
- size_t nnz = slot.sparseNonValueData.size();
- oss << "slot" << i << ":avgNNZ=" << ((double)nnz / sampleNums_) << "; ";
- } else if (slot.type == SlotDef::VECTOR_SPARSE_VALUE) {
- size_t nnz = slot.sparseFloatValueData.size();
- oss << "slot" << i << ":avgNNZ=" << ((double)nnz / sampleNums_) << "; ";
- }
- }
- LOG(INFO) << oss.str();
-}
-
-void ProtoDataProvider::reset() {
- currentSequenceIndex_ = 0;
- if (!skipShuffle_) {
- shuffle();
- }
-
- DataProvider::reset();
-}
-
-void ProtoDataProvider::shuffle() {
- std::shuffle(shuffledSequenceIds_.begin(),
- shuffledSequenceIds_.end(),
- ThreadLocalRandomEngine::get());
-}
-
-/*
- Loop through sequences starting from currentSequenceIndex_
- for at most size samples. For each sequence ranging from [begin, end),
- op(begin, end) will be called.
-
- return the number of sequences scanned
-*/
-template
-int64_t ProtoDataProvider::sequenceLoop(Op op, int64_t size) {
- int64_t sz = 0;
- size_t i;
- size_t sequenceCount = shuffledSequenceIds_.size();
- if (usageRatio_ < 1.0f) {
- sequenceCount = static_cast(sequenceCount * usageRatio_);
- }
- for (i = currentSequenceIndex_; i < sequenceCount; ++i) {
- size_t id = shuffledSequenceIds_[i];
- int64_t begin = sequenceStartPositions_[id];
- int64_t end = sequenceStartPositions_[id + 1];
- int64_t len = end - begin;
- if (sz + len > size && sz > 0) break;
- sz += len;
- op(begin, end);
- }
- return i - currentSequenceIndex_;
-}
-
-/*
- Loop through sequences starting from currentSequenceIndex_
- for at most size samples. For each sample of each sequence at position
- pos, op(pos) will be called.
-
- return the number of sequences scanned
-*/
-template
-int64_t ProtoDataProvider::sampleLoop(Op op, int64_t size) {
- if (iidData()) {
- size = std::min(sampleNums_ - currentSequenceIndex_, size);
- for (int64_t i = currentSequenceIndex_; i < currentSequenceIndex_ + size;
- ++i) {
- size_t pos = shuffledSequenceIds_[i];
- op(pos);
- }
- return size;
- } else {
- auto f = [op](int64_t begin, int64_t end) {
- for (int64_t pos = begin; pos < end; ++pos) {
- op(pos);
- }
- };
- return sequenceLoop(f, size);
- }
-}
-
-/*
- Loop through sub-sequences starting from currentSequenceIndex_
- for at most size samples. For each sample of each sub-sequence at position
- pos, op(pos) will be called.
-
- return the number of sub-sequences scanned
-*/
-template
-int64_t ProtoDataProvider::subSampleLoop(Op op, int64_t size, int slot) {
- CHECK(iidData()) << "subSampleLoop only accepts iid data";
- size = std::min(sampleNums_ - currentSequenceIndex_, size);
- int subSize = 0;
- for (int64_t i = currentSequenceIndex_; i < currentSequenceIndex_ + size;
- ++i) {
- size_t pos = shuffledSequenceIds_[i];
- int64_t* indexs = slots_[slot].indices.data();
- int64_t* subIndexs = slots_[slot].subIndices.data();
- int64_t subSeqStart = 0;
- int64_t subSeqEnd = 0;
- for (int j = 0; j < (int)slots_[slot].subIndices.size(); j++) {
- if (subIndexs[j] == indexs[pos]) {
- subSeqStart = j;
- if (subIndexs[pos] == subIndexs[pos + 1]) {
- subSeqEnd = j + 1;
- break;
- }
- } else if (subIndexs[j] == indexs[pos + 1]) {
- subSeqEnd = j;
- break;
- }
- }
- for (int j = subSeqStart; j < subSeqEnd; j++) {
- op(j);
- }
- subSize += subSeqEnd - subSeqStart;
- }
- return subSize;
-}
-
-int64_t ProtoDataProvider::getNextBatchInternal(int64_t size,
- DataBatch* batch) {
- int64_t numSequences = 0; // actual number of sequences in the batch
-
- // the number of sequences scanned, including those skipped because too long
- int64_t numScannedSeqs = 0;
- std::lock_guard guard(lock_);
- if (iidData()) {
- size = std::min(getSize() - currentSequenceIndex_, size);
- numScannedSeqs = numSequences = size;
- } else {
- int64_t sz = 0;
- auto op = [&sz, &numSequences](int64_t begin, int64_t end) {
- ++numSequences;
- sz += end - begin;
- };
- numScannedSeqs = sequenceLoop(op, size);
- VLOG_IF(1, numScannedSeqs > numSequences)
- << numScannedSeqs - numSequences
- << " sequences are skipped because longer than " << size;
- size = sz;
- }
- if (size <= 0) return 0;
-
- DataBatch& cpuBatch = *cpuBatch_;
- std::vector& cpuArguments = cpuBatch.getStreams();
- cpuBatch.setSize(size);
- cpuArguments.resize(header_.slot_defs_size());
-
- if (!iidData()) {
- ICpuGpuVector::resizeOrCreate(cpuArguments[0].sequenceStartPositions,
- numSequences + 1,
- /* useGpu= */ false);
- int* buf = cpuArguments[0].sequenceStartPositions->getMutableData(false);
- int pos = 0;
- int i = 0;
- auto op = [buf, &pos, &i](int64_t begin, int64_t end) {
- buf[i] = pos;
- pos += end - begin;
- ++i;
- };
- sequenceLoop(op, size);
- buf[i] = size;
- for (size_t slot = 1; slot < cpuArguments.size(); ++slot) {
- cpuArguments[slot].sequenceStartPositions =
- cpuArguments[0].sequenceStartPositions;
- }
- }
-
- for (int slot = 0; slot < header_.slot_defs_size(); ++slot) {
- size_t dim = header_.slot_defs(slot).dim();
- SlotDef::SlotType slotType = header_.slot_defs(slot).type();
-
- std::vector dataPos;
- dataPos.reserve(size);
- auto op = [this, &dataPos](int64_t pos) { dataPos.push_back(pos); };
- sampleLoop(op, size);
-
- switch (slotType) {
- case SlotDef::VECTOR_DENSE: {
- Matrix::resizeOrCreate(cpuArguments[slot].value,
- size,
- dim,
- false, // trans = false
- false); // useGpu = false
- real* buf = cpuArguments[slot].value->getData();
- for (int i = 0; i < size; ++i) {
- memcpy(buf + i * dim,
- slots_[slot].denseData.data() + dataPos[i] * dim,
- sizeof(real) * dim);
- }
- break;
- }
- case SlotDef::VECTOR_SPARSE_NON_VALUE: {
- if (!(cpuArguments[slot].value)) {
- cpuArguments[slot].value =
- Matrix::createSparseMatrix(size,
- dim,
- size /*DEFAULT_AVG_WIDTH = 1*/,
- NO_VALUE,
- SPARSE_CSR,
- false,
- useGpu_);
- }
- auto mat = cpuArguments[slot].value;
- mat->resize(size, dim);
- if (std::dynamic_pointer_cast(mat)) {
- std::dynamic_pointer_cast(mat)->copyFrom(
- dataPos.data(),
- slots_[slot].indices.data(),
- slots_[slot].sparseNonValueData.data(),
- HPPL_STREAM_1);
- } else if (std::dynamic_pointer_cast(mat)) {
- std::dynamic_pointer_cast(mat)->copyFrom(
- dataPos.data(),
- slots_[slot].indices.data(),
- slots_[slot].sparseNonValueData.data());
- } else {
- LOG(FATAL) << "Not Supported";
- }
- size_t numElements = 0;
- for (auto pos : dataPos) {
- numElements +=
- slots_[slot].indices[pos + 1] - slots_[slot].indices[pos];
- }
- nnzStats_[slot]->addSample(numElements);
-
- break;
- }
- case SlotDef::VECTOR_SPARSE_VALUE: {
- if (!(cpuArguments[slot].value)) {
- cpuArguments[slot].value =
- Matrix::createSparseMatrix(size,
- dim,
- size /*DEFAULT_AVG_WIDTH = 1*/,
- FLOAT_VALUE,
- SPARSE_CSR,
- false,
- useGpu_);
- }
- auto mat = cpuArguments[slot].value;
- mat->resize(size, dim);
- if (std::dynamic_pointer_cast(mat)) {
- std::dynamic_pointer_cast(mat)->copyFrom(
- dataPos.data(),
- slots_[slot].indices.data(),
- slots_[slot].sparseFloatValueData.data(),
- HPPL_STREAM_1);
- } else if (std::dynamic_pointer_cast(mat)) {
- std::dynamic_pointer_cast(mat)->copyFrom(
- dataPos.data(),
- slots_[slot].indices.data(),
- slots_[slot].sparseFloatValueData.data());
- } else {
- LOG(FATAL) << "Not Supported";
- }
- break;
- }
- case SlotDef::INDEX: {
- IVector::resizeOrCreate(cpuArguments[slot].ids,
- size,
- /* useGpu= */ false);
- int* buf = cpuArguments[slot].ids->getData();
- for (int i = 0; i < size; ++i) {
- buf[i] = slots_[slot].indexData[dataPos[i]];
- }
- break;
- }
- case SlotDef::VAR_MDIM_DENSE: {
- CHECK_EQ(size, 1);
- auto mat = cpuArguments[slot].value;
- size_t totalDim = slots_[slot].varDenseData[dataPos[0]].data.size();
-
- CHECK_EQ(slots_[slot].varDenseData[dataPos[0]].dims.size(), size_t(3));
- size_t height, width, depth, oldWidth;
- /* dims[2] is depth, will be changed to dims[0] in future */
- depth = slots_[slot].varDenseData[dataPos[0]].dims[2];
- height = slots_[slot].varDenseData[dataPos[0]].dims[1];
- width = slots_[slot].varDenseData[dataPos[0]].dims[0];
- oldWidth = width;
- /* process the undesirable sample */
- if (oldWidth < height) {
- width = height;
- }
- cpuArguments[slot].setFrameHeight(height);
- cpuArguments[slot].setFrameWidth(width);
-
- if (oldWidth < height) {
- totalDim = width * height * depth;
- }
- Matrix::resizeOrCreate(cpuArguments[slot].value,
- size,
- totalDim,
- false, // trans = false
- false); // useGpu = false
- real* buf = cpuArguments[slot].value->getData();
- cpuArguments[slot].value->zeroMem();
- if (oldWidth < height) {
- real* srcBuf = slots_[slot].varDenseData[dataPos[0]].data.data();
- for (size_t i = 0; i < depth; i++) {
- for (size_t j = 0; j < height; j++) {
- for (size_t k = 0; k < oldWidth; k++) {
- buf[i * height * width + j * width + k] =
- srcBuf[i * height * oldWidth + j * oldWidth + k];
- }
- }
- }
- } else {
- memcpy(buf,
- slots_[slot].varDenseData[dataPos[0]].data.data(),
- sizeof(real) * totalDim);
- }
- ICpuGpuVector::resizeOrCreate(cpuArguments[slot].sequenceStartPositions,
- size + 1, /* size == 1 currently */
- /* useGpu= */ false);
- int* bufStarts =
- cpuArguments[slot].sequenceStartPositions->getMutableData(false);
- bufStarts[0] = 0;
- bufStarts[1] = 1;
- break;
- }
- case SlotDef::VAR_MDIM_INDEX: {
- CHECK_EQ(size, 1);
- size_t totalDim = slots_[slot].varIndices[dataPos[0]].size();
- IVector::resizeOrCreate(cpuArguments[slot].ids,
- totalDim,
- /* useGpu= */ false);
- int* buf = cpuArguments[slot].ids->getData();
- memcpy(buf,
- slots_[slot].varIndices[dataPos[0]].data(),
- sizeof(int) * totalDim);
-
- ICpuGpuVector::resizeOrCreate(cpuArguments[slot].sequenceStartPositions,
- size + 1, /* size == 1 currently */
- /* useGpu= */ false);
- int* bufStarts =
- cpuArguments[slot].sequenceStartPositions->getMutableData(false);
- bufStarts[0] = 0;
- /* we expand the convolutinal feature map to a sequence data,
- * so there should be a corresponding sequence labels */
- bufStarts[1] = totalDim;
- break;
- }
- case SlotDef::STRING: {
- if (cpuArguments[slot].strs) {
- cpuArguments[slot].strs->resize(size);
- } else {
- cpuArguments[slot].strs =
- std::make_shared>(size);
- }
- for (int i = 0; i < size; ++i) {
- (*cpuArguments[slot].strs)[i] = slots_[slot].strData[dataPos[i]];
- }
- break;
- }
- }
- }
-
- if (useGpu_) {
- std::vector& cpuArguments = cpuBatch.getStreams();
- DataBatch& gpuBatch = *gpuBatch_;
- std::vector& gpuArguments = gpuBatch.getStreams();
- gpuArguments.resize(cpuArguments.size());
- gpuBatch.setSize(size);
- for (int i = 0; i < header_.slot_defs_size(); ++i) {
- SlotDef::SlotType slotType = header_.slot_defs(i).type();
- if (SlotDef::VECTOR_SPARSE_VALUE == slotType ||
- SlotDef::VECTOR_SPARSE_NON_VALUE == slotType) {
- gpuArguments[i] = cpuArguments[i];
- gpuArguments[i].sequenceStartPositions =
- cpuArguments[i].sequenceStartPositions;
- } else {
- gpuArguments[i].resizeAndCopyFrom(
- cpuArguments[i], useGpu_, HPPL_STREAM_1);
- }
- }
- hl_stream_synchronize(HPPL_STREAM_1);
- *batch = gpuBatch;
- } else {
- *batch = cpuBatch;
- }
-
- currentSequenceIndex_ += numScannedSeqs;
-
- return batch->getSize();
-}
-
-ProtoSequenceDataProvider::ProtoSequenceDataProvider(const DataConfig& config,
- bool useGpu,
- bool loadDataAll)
- : ProtoDataProvider(config, useGpu, loadDataAll) {}
-
-int64_t ProtoSequenceDataProvider::getNextBatchInternal(int64_t size,
- DataBatch* batch) {
- CHECK(iidData()) << "ProtoSequenceDataProvider only accepts iid data";
- int64_t numSequences = 0; // actual number of sequences in the batch
-
- // the number of sequences scanned, including those skipped because too long
- int64_t numScannedSeqs = 0;
- std::lock_guard guard(lock_);
- size = std::min(getSize() - currentSequenceIndex_, size);
- numScannedSeqs = numSequences = size;
- if (size <= 0) return 0;
-
- DataBatch& cpuBatch = *cpuBatch_;
- std::vector& cpuArguments = cpuBatch.getStreams();
- cpuBatch.setSize(size);
- cpuArguments.resize(header_.slot_defs_size());
-
- for (int slot = 0; slot < header_.slot_defs_size(); ++slot) {
- SlotDef::SlotType slotType = header_.slot_defs(slot).type();
-
- std::vector dataPos;
- dataPos.reserve(size);
- auto op = [this, &dataPos](int64_t pos) { dataPos.push_back(pos); };
- sampleLoop(op, size);
-
- // current slot: sequenceStartPositions
- ICpuGpuVector::resizeOrCreate(cpuArguments[slot].sequenceStartPositions,
- size + 1,
- /* useGpu= */ false);
-
- switch (slotType) {
- case SlotDef::VECTOR_SPARSE_VALUE:
- case SlotDef::VAR_MDIM_DENSE:
- case SlotDef::VAR_MDIM_INDEX: {
- LOG(FATAL) << "ProtoSequenceDataProvider only support"
- << " VECTOR_DENSE, VECTOR_SPARSE_NON_VALUE and INDEX slots";
- break;
- }
- case SlotDef::VECTOR_SPARSE_NON_VALUE: {
- // copy to IDS, not value
- // pointers used in current slot
- sparse_non_value_t* data = slots_[slot].sparseNonValueData.data();
- int64_t* indexs = slots_[slot].indices.data();
- int64_t* seqs = dataPos.data();
-
- // current slot: i need size instances. what is the total length?
- int totalFeatureInCurrentSlot = 0;
- for (int ins = 0; ins < size; ins++) {
- int64_t currInsId = seqs[ins];
- totalFeatureInCurrentSlot +=
- indexs[currInsId + 1] - indexs[currInsId];
- // special: if current instance has NO feature in current slot
- if (indexs[currInsId + 1] == indexs[currInsId]) {
- totalFeatureInCurrentSlot++;
- }
- }
- // done
-
- // current slot: ids
- IVector::resizeOrCreate(cpuArguments[slot].ids,
- totalFeatureInCurrentSlot,
- /* useGpu= */ false);
-
- // where to write
- int* currPosOfArgumentId = cpuArguments[slot].ids->getData();
- int* currPosOfArgumentSeqStart =
- cpuArguments[slot].sequenceStartPositions->getMutableData(false);
- int allSequenceLength = 0;
- currPosOfArgumentSeqStart[0] = 0;
- // for each instance, copy data and fill sequence positions
- for (int instance = 0; instance < size; instance++) {
- int64_t currInstanceId = seqs[instance];
- int64_t currInstanceLength =
- indexs[currInstanceId + 1] - indexs[currInstanceId];
- sparse_non_value_t* currInstanceData = data + indexs[currInstanceId];
- // write sequenceStartPositions
- allSequenceLength += currInstanceLength;
- currPosOfArgumentSeqStart[instance + 1] = allSequenceLength;
- // copy features
- for (int featCopier = 0; featCopier < currInstanceLength;
- featCopier++) {
- currPosOfArgumentId[featCopier] = currInstanceData[featCopier].col;
- }
- currPosOfArgumentId += currInstanceLength;
- // special: if current instance has NO feature in current slot
- if (currInstanceLength == 0) {
- allSequenceLength++;
- currPosOfArgumentSeqStart[instance + 1] = allSequenceLength;
- currPosOfArgumentId[0] = -1;
- currPosOfArgumentId++;
- }
- // done
- }
- if (slots_[slot].subIndices.size()) {
- std::vector dataSubPos;
- auto op = [this, &dataSubPos](int64_t pos) {
- dataSubPos.push_back(pos);
- };
- int subSize = subSampleLoop(op, size, slot);
- ICpuGpuVector::resizeOrCreate(
- cpuArguments[slot].subSequenceStartPositions, subSize + 1, false);
- int* currPosOfArgumentSubSeqStart =
- cpuArguments[slot].subSequenceStartPositions->getMutableData(
- false);
- int64_t* subSeqs = dataSubPos.data();
- int64_t* subIndexs = slots_[slot].subIndices.data();
- int allSubSequenceLength = 0;
- currPosOfArgumentSubSeqStart[0] = 0;
- // for each instance, compute sub-sequence number
- for (int instance = 0; instance < subSize; instance++) {
- int64_t currSubInstanceId = subSeqs[instance];
- int64_t currSubInstanceLength =
- subIndexs[currSubInstanceId + 1] - subIndexs[currSubInstanceId];
- // write subSequenceStartPositions
- allSubSequenceLength += currSubInstanceLength;
- currPosOfArgumentSubSeqStart[instance + 1] = allSubSequenceLength;
- // special: if current instance has NO feature in current slot
- if (currSubInstanceLength == 0) {
- allSubSequenceLength++;
- currPosOfArgumentSubSeqStart[instance + 1] = allSubSequenceLength;
- }
- }
- cpuArguments[slot].checkSubset();
- }
- break;
- }
- case SlotDef::INDEX: {
- // label slot
- IVector::resizeOrCreate(cpuArguments[slot].ids,
- size,
- /* useGpu= */ false);
- // fill labels
- int* buf = cpuArguments[slot].ids->getData();
- for (int i = 0; i < size; ++i) {
- buf[i] = slots_[slot].indexData[dataPos[i]];
- }
- // label HAS sequence structure
- cpuArguments[slot].sequenceStartPositions->fillSequence(false);
- break;
- }
- case SlotDef::VECTOR_DENSE: {
- // copy values
- size_t dim = header_.slot_defs(slot).dim();
- Matrix::resizeOrCreate(cpuArguments[slot].value,
- size,
- dim,
- false, // trans = false
- false); // useGpu = false
- real* buf = cpuArguments[slot].value->getData();
- for (int i = 0; i < size; ++i) {
- memcpy(buf + i * dim,
- slots_[slot].denseData.data() + dataPos[i] * dim,
- sizeof(real) * dim);
- }
- // sequence structure
- cpuArguments[slot].sequenceStartPositions->fillSequence(false);
- break;
- }
- default: { LOG(FATAL) << "should not reach here"; }
- }
- }
-
- if (useGpu_) {
- std::vector& cpuArguments = cpuBatch.getStreams();
- DataBatch& gpuBatch = *gpuBatch_;
- std::vector& gpuArguments = gpuBatch.getStreams();
- gpuArguments.resize(cpuArguments.size());
- gpuBatch.setSize(size);
- for (size_t i = 0; i < cpuArguments.size(); ++i) {
- gpuArguments[i].resizeAndCopyFrom(
- cpuArguments[i], useGpu_, HPPL_STREAM_1);
- }
- hl_stream_synchronize(HPPL_STREAM_1);
- *batch = gpuBatch;
- } else {
- *batch = cpuBatch;
- }
-
- currentSequenceIndex_ += numScannedSeqs;
- return batch->getSize();
-}
-
-} // namespace paddle
diff --git a/paddle/gserver/dataproviders/ProtoDataProvider.h b/paddle/gserver/dataproviders/ProtoDataProvider.h
deleted file mode 100644
index 7dd45e062248f20d24c633dd4e1c8b7eebcbfa1b..0000000000000000000000000000000000000000
--- a/paddle/gserver/dataproviders/ProtoDataProvider.h
+++ /dev/null
@@ -1,179 +0,0 @@
-/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
- http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License. */
-
-#pragma once
-
-#include
-
-#include "DataFormat.pb.h"
-#include "paddle/utils/Stat.h"
-
-#include "DataProvider.h"
-#include "ProtoReader.h"
-
-namespace paddle {
-
-/**
- * @brief Provider data from protobuf data file with each sample
- * specified by proto message
- *
- * DataSample defined in DataFormat.proto.
- *
- * The file format is
- *
- * header
- *
- * sample1
- *
- * sample2
- *
- * ...
- *
- * sampleN
- *
- * @note: In the data file, each message is prefixed with its length.
- * The read/write of the protbuf are implemented in ProtoReader.h
- */
-class ProtoDataProvider : public DataProvider {
-public:
- ProtoDataProvider(const DataConfig& config,
- bool useGpu,
- bool loadDataAll = true);
- virtual void reset();
-
- /**
- * @note this size includes the sequences which are skipped because they
- * are longer than the batch size.
- */
- virtual int64_t getSize() {
- int64_t size = sampleNums_;
- if (usageRatio_ < 1.0f) {
- size = static_cast(size * usageRatio_);
- }
- return size;
- }
- virtual void shuffle();
-
- void loadData(const std::vector& fileList);
-
- virtual int64_t getNextBatchInternal(int64_t size, DataBatch* batch);
-
-protected:
- /**
- * @brief load protobuf data from a list of file
- * @param[in] fileName file name of a file which contains
- * a list of file names
- */
- void loadData(const std::string& fileName);
-
- /**
- * @brief load protobuf data from file
- * @param[in] fileName data file name
- */
- void loadDataFile(const std::string& fileName);
- /** @brief check data header of each data sample
- * @param[in] header data header read from protobuf data
- */
- void checkDataHeader(const DataHeader& header);
- /**
- * @brief fill protobuf data into slot_,
- * slot_ is a vector of ProtoSlot in memory.
- * @param[in] sample data sample read from protobuf data
- */
- void fillSlots(const DataSample& sample);
-
- /**
- * @brief return true if each sample is one sequence, i.e., independent
- * of other samples.
- */
- inline bool iidData() const { return sequenceStartPositions_.empty(); }
-
- /**
- * @brief check that sample is consistent with header_
- */
- void checkSample(const DataSample& sample);
-
- template
- int64_t sequenceLoop(Op op, int64_t size);
-
- template
- int64_t sampleLoop(Op op, int64_t size);
-
- template
- int64_t subSampleLoop(Op op, int64_t size, int slot);
-
- void showDataStats();
-
-protected:
- struct ProtoVarSlot {
- std::vector data;
- std::vector dims;
- };
-
- struct ProtoSlot {
- SlotDef::SlotType type;
- int dim;
- std::vector indexData;
- std::vector denseData;
- std::vector sparseNonValueData;
- std::vector sparseFloatValueData;
- std::vector indices;
- std::vector subIndices;
-
- std::vector varDenseData;
- std::vector> varIndices;
- std::vector strData;
- };
- DataHeader header_;
- int numVecSlots_;
-
- std::vector slots_;
- size_t sampleNums_;
-
- /**
- * The starting position of each sequence in samples.
- * The last element should be num of samples.
- * If empty, each sample is one sequence.
- */
- std::vector sequenceStartPositions_;
-
- int64_t currentSequenceIndex_;
-
- // The size should be the number of sequences.
- std::vector shuffledSequenceIds_;
-
- ThreadLocalD cpuBatch_;
- ThreadLocalD gpuBatch_;
-
- RWLock lock_;
- std::vector nnzStats_; // stats for number of none-zeros entries
-};
-
-/**
- * @brief Special use for Proto data: instances should contain sparse-non-value
- * slots
- * and label.
- *
- * @note ProtoSequenceDataProvider treats each SPARSE SLOT as a SEQUENCE
- */
-class ProtoSequenceDataProvider : public ProtoDataProvider {
-public:
- ProtoSequenceDataProvider(const DataConfig& config,
- bool useGpu,
- bool loadDataAll = true);
- ~ProtoSequenceDataProvider() {}
- virtual int64_t getNextBatchInternal(int64_t size, DataBatch* batch);
-};
-
-} // namespace paddle
diff --git a/paddle/gserver/layers/MKLDNNConcatLayer.cpp b/paddle/gserver/layers/MKLDNNConcatLayer.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..c9099297cc5c741fbae0b42f21b988e6c561ef11
--- /dev/null
+++ b/paddle/gserver/layers/MKLDNNConcatLayer.cpp
@@ -0,0 +1,202 @@
+/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "MKLDNNConcatLayer.h"
+
+using namespace mkldnn; // NOLINT
+typedef memory::format format;
+
+namespace paddle {
+
+REGISTER_LAYER(mkldnn_concat, MKLDNNConcatLayer);
+
+bool MKLDNNConcatLayer::init(const LayerMap& layerMap,
+ const ParameterMap& parameterMap) {
+ if (!MKLDNNLayer::init(layerMap, parameterMap)) {
+ return false;
+ }
+ CHECK_GT(inputLayers_.size(), 1UL);
+ CHECK(!biasParameter_);
+ return true;
+}
+
+void MKLDNNConcatLayer::reshape(
+ int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) {
+ reshapeInput(bs, ih, iw);
+ ic = inputLayers_[0]->getSize() / ih / iw;
+ CHECK_EQ((size_t)ic * ih * iw, inputLayers_[0]->getSize());
+ CHECK_EQ(inputElemenCnt_, (size_t)bs * ic * ih * iw);
+ CHECK_GT(inputLayers_.size(), 1UL);
+ channels_.resize(inputLayers_.size());
+ channels_[0] = ic;
+ // need change the output channel, so use oc_ instead
+ // TODO(TJ): change API, use &oc
+ oc_ = ic;
+ for (size_t i = 1; i < inputLayers_.size(); i++) {
+ int batchsize, height, witdh;
+ reshapeInput(batchsize, height, witdh, i);
+ CHECK_EQ(bs, batchsize);
+ CHECK_EQ(ih, height);
+ CHECK_EQ(iw, witdh);
+
+ channels_[i] = inputLayers_[i]->getSize() / height / witdh;
+ CHECK_EQ((size_t)channels_[i] * height * witdh, inputLayers_[i]->getSize());
+ oc_ += channels_[i];
+ }
+ oh = ih;
+ ow = iw;
+ reshapeOutput(oh, ow);
+ resizeOutput(bs, oc_ * oh * ow);
+}
+
+void MKLDNNConcatLayer::resetFwd(std::vector& pipeline,
+ MKLDNNMatrixPtr& in,
+ MKLDNNMatrixPtr& wgt,
+ MKLDNNMatrixPtr& bias,
+ MKLDNNMatrixPtr& out) {
+ resetFwdBuffers(inVals_, out);
+ in = inVals_[0];
+
+ std::shared_ptr fwdPD;
+ resetFwdPD(fwdPD, inVals_, out);
+
+ resetFwdPipeline(pipeline, fwdPD, inVals_, out);
+}
+
+void MKLDNNConcatLayer::resetBwd(std::vector& pipeline,
+ MKLDNNMatrixPtr& in,
+ MKLDNNMatrixPtr& wgt,
+ MKLDNNMatrixPtr& bias,
+ MKLDNNMatrixPtr& out) {
+ resetBwdBuffers(inGrads_, out);
+ in = inGrads_[0];
+
+ resetBwdPipeline(pipeline, bwds_, inGrads_, out);
+}
+
+void MKLDNNConcatLayer::resetFwdBuffers(std::vector& inputs,
+ MKLDNNMatrixPtr& out) {
+ inputs.resize(inputLayers_.size());
+ bool has8c = false, has16c = false, hasnc = false;
+ for (size_t i = 0; i < inputs.size(); i++) {
+ // resetInValue will use ic_ so temporary change as current input's channel
+ // TODO(TJ): change ic_ as vector then can remove channels_
+ ic_ = channels_[i];
+ resetInValue(inputs[i], nullptr, i);
+ CHECK(inputs[i]);
+ auto dm = inputs[i]->getDims();
+ // inputs format can be different, but ndims must equal
+ CHECK(i == 0 || dm.size() == inputs[0]->getDims().size());
+ CHECK_EQ(bs_, dm[0]);
+ CHECK_EQ(channels_[i], dm[1]);
+ if (dm.size() > 2) {
+ CHECK_EQ(ih_, dm[2]);
+ CHECK_EQ(iw_, dm[3]);
+ }
+ if (inputs[i]->getFormat() == format::nc) {
+ hasnc = true;
+ }
+ if (inputs[i]->getFormat() == format::nChw8c) {
+ has8c = true;
+ }
+ if (inputs[i]->getFormat() == format::nChw16c) {
+ has16c = true;
+ }
+ }
+ // change back, ic_ always save the input 0 size
+ ic_ = channels_[0];
+
+ format outFmt;
+ if (has16c && oc_ % 16 == 0) {
+ outFmt = format::nChw16c;
+ } else if (has8c && oc_ % 8 == 0) {
+ outFmt = format::nChw8c;
+ } else if (hasnc) {
+ CHECK(oh_ == 1 && ow_ == 1);
+ outFmt = format::nc;
+ } else {
+ outFmt = format::nchw;
+ }
+ memory::dims outDims =
+ hasnc ? memory::dims{bs_, oc_} : memory::dims{bs_, oc_, oh_, ow_};
+ auto outPD = MKLDNNMatrix::createPrimitiveDesc(outDims, outFmt, engine_);
+ resetOutValue(out, outPD);
+}
+
+void MKLDNNConcatLayer::resetFwdPD(std::shared_ptr& pd,
+ std::vector& inputs,
+ MKLDNNMatrixPtr out) {
+ std::vector srcPDs;
+ for (size_t i = 0; i < inputs.size(); i++) {
+ srcPDs.push_back(inputs[i]->getPrimitiveDesc());
+ }
+ CHECK(out);
+ pd.reset(new concat::primitive_desc(out->getMemoryDesc(), axis_, srcPDs));
+ CHECK_PRIMITIVE_DESC_EQ(out, pd->dst_primitive_desc());
+}
+
+void MKLDNNConcatLayer::resetFwdPipeline(
+ std::vector& pipeline,
+ std::shared_ptr& pd,
+ std::vector& inputs,
+ MKLDNNMatrixPtr& out) {
+ std::vector srcs;
+ for (size_t i = 0; i < inputs.size(); i++) {
+ srcs.push_back(*(inputs[i]));
+ }
+ fwd_.reset(new concat(*pd, srcs, *out));
+ pipeline.push_back(*fwd_);
+}
+
+void MKLDNNConcatLayer::resetBwdBuffers(std::vector& inputs,
+ MKLDNNMatrixPtr& out) {
+ CHECK(outVal_);
+ resetOutGrad(out, outVal_->getPrimitiveDesc());
+ CHECK(out);
+
+ inputs.resize(inputLayers_.size());
+ for (size_t i = 0; i < inputs.size(); i++) {
+ CHECK(inVals_[i]);
+ // resetInGrad will use inVal_
+ // TODO(TJ): change move inVals_ to MKLDNNLayer ans remove inVal_
+ inVal_ = inVals_[i];
+ resetInGrad(inputs[i], inVals_[i]->getPrimitiveDesc(), i);
+ CHECK_PRIMITIVE_DESC_EQ(inputs[i], inVals_[i]->getPrimitiveDesc());
+ }
+ // change back, inVal_ always save the input 0
+ inVal_ = inVals_[0];
+}
+
+void MKLDNNConcatLayer::resetBwdPipeline(
+ std::vector& pipeline,
+ std::vector>& prims,
+ std::vector& inputs,
+ MKLDNNMatrixPtr& out) {
+ // reset the backward primitives
+ memory::dims offsets = {0, 0, 0, 0};
+ prims.resize(inputs.size());
+ CHECK_EQ(inputs.size(), channels_.size());
+ for (size_t i = 0; i < inputs.size(); i++) {
+ auto viewPD = view::primitive_desc(
+ out->getPrimitiveDesc(), inputs[i]->getDims(), offsets);
+ auto bwdPD = reorder::primitive_desc(viewPD.dst_primitive_desc(),
+ inputs[i]->getPrimitiveDesc());
+ prims[i].reset(new reorder(bwdPD, *out, *(inputs[i])));
+ offsets[axis_] += channels_[i];
+ // push to pipeline
+ pipeline.push_back(*prims[i]);
+ }
+}
+
+} // namespace paddle
diff --git a/paddle/gserver/layers/MKLDNNConcatLayer.h b/paddle/gserver/layers/MKLDNNConcatLayer.h
new file mode 100644
index 0000000000000000000000000000000000000000..d5749d327e4259b81541a234f48a4538ab035fe4
--- /dev/null
+++ b/paddle/gserver/layers/MKLDNNConcatLayer.h
@@ -0,0 +1,129 @@
+/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+#include "MKLDNNLayer.h"
+#include "mkldnn.hpp"
+
+namespace paddle {
+
+/**
+ * @brief A subclass of MKLDNNLayer Concatenate layer.
+ *
+ * The config file api is mkldnn_concat
+ */
+class MKLDNNConcatLayer : public MKLDNNLayer {
+protected:
+ std::vector inVals_;
+ std::vector inGrads_;
+ std::vector> bwds_;
+ // input channel numbers
+ std::vector channels_;
+
+ // concat_dimension in MKLDNN
+ // if axis_ == 0, concat batchsize
+ // if axis_ == 1, concat channel (default)
+ int axis_;
+
+public:
+ explicit MKLDNNConcatLayer(const LayerConfig& config)
+ : MKLDNNLayer(config), axis_(1) {}
+
+ ~MKLDNNConcatLayer() {}
+
+ bool init(const LayerMap& layerMap,
+ const ParameterMap& parameterMap) override;
+
+ void reshape(
+ int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) override;
+
+ void resetFwd(std::vector& pipeline,
+ MKLDNNMatrixPtr& in,
+ MKLDNNMatrixPtr& wgt,
+ MKLDNNMatrixPtr& bias,
+ MKLDNNMatrixPtr& out) override;
+
+ void resetBwd(std::vector& pipeline,
+ MKLDNNMatrixPtr& in,
+ MKLDNNMatrixPtr& wgt,
+ MKLDNNMatrixPtr& bias,
+ MKLDNNMatrixPtr& out) override;
+
+ void printSizeInfo() override {
+ CHECK_EQ(channels_.size(), inputLayers_.size());
+ for (size_t i = 0; i < channels_.size(); ++i) {
+ VLOG(MKLDNN_SIZES) << "Input " << i << ", " << inputLayers_[i]->getName()
+ << ": " << bs_ << ", " << channels_[i] << ", " << ih_
+ << ", " << iw_;
+ }
+ VLOG(MKLDNN_SIZES) << "Output: " << bs_ << ", " << oc_ << ", " << oh_
+ << ", " << ow_;
+ }
+
+ void printValueFormat() override {
+ for (size_t i = 0; i < inVals_.size(); ++i) {
+ VLOG(MKLDNN_FMTS) << "Input " << i << ", " << inputLayers_[i]->getName()
+ << ": " << inVals_[i]->getFormat() << " >>>";
+ }
+ if (outVal_) {
+ VLOG(MKLDNN_FMTS) << outVal_->getFormat() << " >>> ";
+ }
+ if (extOutVal_) {
+ VLOG(MKLDNN_FMTS) << extOutVal_->getFormat();
+ }
+ }
+
+ void printGradFormat() override {
+ if (extOutGrad_) {
+ VLOG(MKLDNN_FMTS) << extOutGrad_->getFormat();
+ }
+ if (outGrad_) {
+ VLOG(MKLDNN_FMTS) << outGrad_->getFormat() << " <<< ";
+ }
+ for (size_t i = 0; i < inGrads_.size(); ++i) {
+ VLOG(MKLDNN_FMTS) << "Input " << i << ", " << inputLayers_[i]->getName()
+ << ": " << inGrads_[i]->getFormat() << "<<<";
+ }
+ }
+
+protected:
+ /**
+ * Forward functions: reset buffers(inputs, output, bias),
+ * reset primitive descriptor,
+ * reset pipeline.
+ */
+ void resetFwdBuffers(std::vector& inputs,
+ MKLDNNMatrixPtr& out);
+ void resetFwdPD(std::shared_ptr& pd,
+ std::vector& inputs,
+ MKLDNNMatrixPtr out);
+ void resetFwdPipeline(std::vector& pipeline,
+ std::shared_ptr& pd,
+ std::vector& inputs,
+ MKLDNNMatrixPtr& out);
+
+ /**
+ * Backward functions: reset buffers(inputs, output, bias)
+ * reset primitives and pipeline
+ */
+ void resetBwdBuffers(std::vector& inputs,
+ MKLDNNMatrixPtr& out);
+ void resetBwdPipeline(std::vector& pipeline,
+ std::vector>& prims,
+ std::vector& inputs,
+ MKLDNNMatrixPtr& out);
+};
+
+} // namespace paddle
diff --git a/paddle/gserver/layers/MKLDNNLayer.cpp b/paddle/gserver/layers/MKLDNNLayer.cpp
index e75ac5ba4647a8267b7bc189893bd7adb5c3053f..cf42da0735282d667d6b87061c8c59bf2f96e0be 100644
--- a/paddle/gserver/layers/MKLDNNLayer.cpp
+++ b/paddle/gserver/layers/MKLDNNLayer.cpp
@@ -21,8 +21,8 @@ namespace paddle {
bool MKLDNNLayer::init(const LayerMap& layerMap,
const ParameterMap& parameterMap) {
- CHECK(FLAGS_use_mkldnn) << "MkldnnLayers only support use_mkldnn."
- << "Please set WITH_MKLDNN=ON "
+ CHECK(FLAGS_use_mkldnn) << "MKLDNNLayers only support use_mkldnn."
+ << "Please set WITH_MKL=ON "
<< "and set use_mkldnn=True";
CHECK(!useGpu_) << "Do not support GPU yet";
@@ -138,8 +138,11 @@ void MKLDNNLayer::backward(const UpdateCallback& callback) {
}
}
-void MKLDNNLayer::reshapeInput(int& batchsize, int& height, int& width) {
- const Argument& input = inputLayers_[0]->getOutput();
+void MKLDNNLayer::reshapeInput(int& batchsize,
+ int& height,
+ int& width,
+ size_t inputIdx) {
+ const Argument& input = inputLayers_[inputIdx]->getOutput();
batchsize = input.getBatchSize();
int h = input.getFrameHeight();
int w = input.getFrameWidth();
diff --git a/paddle/gserver/layers/MKLDNNLayer.h b/paddle/gserver/layers/MKLDNNLayer.h
index 7479c34c92b5231b2521493bc631474d4efd4224..4c42df1bee75fa7b28c2001c30797cc0df7c5554 100644
--- a/paddle/gserver/layers/MKLDNNLayer.h
+++ b/paddle/gserver/layers/MKLDNNLayer.h
@@ -178,7 +178,10 @@ protected:
/**
* reshape the input image sizes and input batchsize
*/
- void reshapeInput(int& batchsize, int& height, int& width);
+ void reshapeInput(int& batchsize,
+ int& height,
+ int& width,
+ size_t inputIdx = 0);
/**
* reshape output image sizes
diff --git a/paddle/gserver/tests/CMakeLists.txt b/paddle/gserver/tests/CMakeLists.txt
index 4bea348f637f39444e8aad89278e6366ecd73b1d..c295ea19c9ccb3d05c509a41925d2c36efdba8ef 100644
--- a/paddle/gserver/tests/CMakeLists.txt
+++ b/paddle/gserver/tests/CMakeLists.txt
@@ -29,7 +29,7 @@ gserver_test(test_KmaxSeqScore)
gserver_test(test_Expand)
gserver_test(test_MaxPoolingWithMaskOutput)
-########## test_Mkldnn layers and activations ##########
+########## test_MKLDNN layers and activations ##########
if(WITH_MKLDNN)
add_unittest_without_exec(test_MKLDNN
test_MKLDNN.cpp
@@ -62,17 +62,6 @@ if(NOT WITH_DOUBLE AND NOT MOBILE_INFERENCE)
endif()
if(NOT MOBILE_INFERENCE)
-################### test_ProtoDataProvider ############
- add_unittest_without_exec(test_ProtoDataProvider
- test_ProtoDataProvider.cpp)
-
- # test_ProtoDataProvider will mkdir as same name,
- # so if WORKING_DIRECTORY is default directory, then
- # mkdir will get error.
- add_test(NAME test_ProtoDataProvider
- COMMAND ${CMAKE_CURRENT_BINARY_DIR}/test_ProtoDataProvider
- WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle)
-
################## test_Evaluator #######################
add_unittest(test_Evaluator
test_Evaluator.cpp)
@@ -110,3 +99,24 @@ add_test(NAME test_PyDataProvider2
COMMAND .set_python_path.sh -d ${PADDLE_SOURCE_DIR}/paddle/gserver/tests:${PADDLE_SOURCE_DIR}/python ${CMAKE_CURRENT_BINARY_DIR}/test_PyDataProvider2
WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle
)
+
+################# test_CompareSparse ##################
+add_unittest_without_exec(test_CompareSparse
+ test_CompareSparse.cpp)
+if(NOT ON_TRAVIS)
+ add_test(NAME test_CompareSparse
+ COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d
+ ${PADDLE_SOURCE_DIR}/python:${PADDLE_SOURCE_DIR}/paddle/gserver/tests
+ ./.set_port.sh -p port -n 6
+ ${CMAKE_CURRENT_BINARY_DIR}/test_CompareSparse
+ WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/)
+endif()
+
+################ test_CompareTwoNets ######################
+add_unittest_without_exec(test_CompareTwoNets
+ test_CompareTwoNets.cpp)
+add_test(NAME test_CompareTwoNets
+ COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d
+ ${PADDLE_SOURCE_DIR}/python:${PADDLE_SOURCE_DIR}/paddle/gserver/tests
+ ${CMAKE_CURRENT_BINARY_DIR}/test_CompareTwoNets
+ WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/)
diff --git a/paddle/gserver/tests/MKLDNNTester.h b/paddle/gserver/tests/MKLDNNTester.h
index ca55a45bc77b4e171619ab788d7c7dfeefcd036a..9d61533c0b6f20c41130d7b7c15ad93392b2d24c 100644
--- a/paddle/gserver/tests/MKLDNNTester.h
+++ b/paddle/gserver/tests/MKLDNNTester.h
@@ -23,7 +23,7 @@ limitations under the License. */
namespace paddle {
/**
- * @brief test the functionality of Mkldnnlayers
+ * @brief test the functionality of MKLDNNlayers and MKLDNNActivations
* refer to paddle original function
*/
class MKLDNNTester {
diff --git a/paddle/gserver/tests/proto_files.txt b/paddle/gserver/tests/proto_files.txt
deleted file mode 100644
index 691b38c7940bd21360eb00384e060554aa4b3e22..0000000000000000000000000000000000000000
--- a/paddle/gserver/tests/proto_files.txt
+++ /dev/null
@@ -1,2 +0,0 @@
-./test_ProtoDataProvider/data1.bin
-./test_ProtoDataProvider/data2.bin
diff --git a/paddle/gserver/tests/proto_files_compressed.txt b/paddle/gserver/tests/proto_files_compressed.txt
deleted file mode 100644
index 7413c81e185d02e0d03aefa06480b9722357c5eb..0000000000000000000000000000000000000000
--- a/paddle/gserver/tests/proto_files_compressed.txt
+++ /dev/null
@@ -1,2 +0,0 @@
-./test_ProtoDataProvider/data1.bin.gz
-./test_ProtoDataProvider/data2.bin.gz
diff --git a/paddle/gserver/tests/sequence_lstm.conf b/paddle/gserver/tests/sequence_lstm.conf
new file mode 100644
index 0000000000000000000000000000000000000000..f49a827f22edce056eaf9903e99b732cab7f3784
--- /dev/null
+++ b/paddle/gserver/tests/sequence_lstm.conf
@@ -0,0 +1,64 @@
+#!/usr/bin/env python
+# 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.
+
+from paddle.trainer_config_helpers import *
+
+######################## data source ################################
+dict_path = 'gserver/tests/Sequence/tour_dict_phrase.dict'
+dict_file = dict()
+for line_count, line in enumerate(open(dict_path, "r")):
+ dict_file[line.strip()] = line_count
+
+define_py_data_sources2(
+ train_list='gserver/tests/Sequence/train.list',
+ test_list=None,
+ module='sequenceGen',
+ obj='process',
+ args={"dict_file": dict_file})
+
+settings(batch_size=5)
+######################## network configure ################################
+dict_dim = len(open(dict_path, 'r').readlines())
+word_dim = 128
+hidden_dim = 256
+label_dim = 3
+sparse_update = get_config_arg("sparse_update", bool, False)
+
+data = data_layer(name="word", size=dict_dim)
+
+emb = embedding_layer(
+ input=data,
+ size=word_dim,
+ param_attr=ParamAttr(sparse_update=sparse_update))
+
+with mixed_layer(size=hidden_dim * 4) as lstm_input:
+ lstm_input += full_matrix_projection(input=emb)
+
+lstm = lstmemory(
+ input=lstm_input,
+ act=TanhActivation(),
+ gate_act=SigmoidActivation(),
+ state_act=TanhActivation())
+
+lstm_last = last_seq(input=lstm)
+
+with mixed_layer(
+ size=label_dim, act=SoftmaxActivation(), bias_attr=True) as output:
+ output += full_matrix_projection(input=lstm_last)
+
+outputs(
+ classification_cost(
+ input=output, label=data_layer(
+ name="label", size=1)))
diff --git a/paddle/gserver/tests/sequence_recurrent.py b/paddle/gserver/tests/sequence_recurrent.py
new file mode 100644
index 0000000000000000000000000000000000000000..4895df186bfecc5cb5263676a9cd5bac5039d565
--- /dev/null
+++ b/paddle/gserver/tests/sequence_recurrent.py
@@ -0,0 +1,56 @@
+#!/usr/bin/env python
+# 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.
+
+from paddle.trainer_config_helpers import *
+
+######################## data source ################################
+dict_path = 'gserver/tests/Sequence/tour_dict_phrase.dict'
+dict_file = dict()
+for line_count, line in enumerate(open(dict_path, "r")):
+ dict_file[line.strip()] = line_count
+
+define_py_data_sources2(
+ train_list='gserver/tests/Sequence/train.list',
+ test_list=None,
+ module='sequenceGen',
+ obj='process',
+ args={"dict_file": dict_file})
+
+settings(batch_size=5)
+######################## network configure ################################
+dict_dim = len(open(dict_path, 'r').readlines())
+word_dim = 128
+hidden_dim = 128
+label_dim = 3
+
+# This config is designed to be equivalent with sequence_recurrent_group.py
+
+data = data_layer(name="word", size=dict_dim)
+
+emb = embedding_layer(
+ input=data, size=word_dim, param_attr=ParamAttr(name="emb"))
+
+recurrent = recurrent_layer(input=emb, bias_attr=False, act=SoftmaxActivation())
+
+recurrent_last = last_seq(input=recurrent)
+
+with mixed_layer(
+ size=label_dim, act=SoftmaxActivation(), bias_attr=True) as output:
+ output += full_matrix_projection(input=recurrent_last)
+
+outputs(
+ classification_cost(
+ input=output, label=data_layer(
+ name="label", size=1)))
diff --git a/paddle/gserver/tests/sequence_recurrent_group.py b/paddle/gserver/tests/sequence_recurrent_group.py
new file mode 100644
index 0000000000000000000000000000000000000000..a1d54542e3bc4e89f70d31d5e89c0f44953c9f90
--- /dev/null
+++ b/paddle/gserver/tests/sequence_recurrent_group.py
@@ -0,0 +1,70 @@
+#!/usr/bin/env python
+# 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.
+
+from paddle.trainer_config_helpers import *
+
+######################## data source ################################
+dict_path = 'gserver/tests/Sequence/tour_dict_phrase.dict'
+dict_file = dict()
+for line_count, line in enumerate(open(dict_path, "r")):
+ dict_file[line.strip()] = line_count
+
+define_py_data_sources2(
+ train_list='gserver/tests/Sequence/train.list',
+ test_list=None,
+ module='sequenceGen',
+ obj='process',
+ args={"dict_file": dict_file})
+
+settings(batch_size=5)
+######################## network configure ################################
+dict_dim = len(open(dict_path, 'r').readlines())
+word_dim = 128
+hidden_dim = 128
+label_dim = 3
+
+# This config is designed to be equivalent with sequence_recurrent.py
+
+data = data_layer(name="word", size=dict_dim)
+
+emb = embedding_layer(
+ input=data, size=word_dim, param_attr=ParamAttr(name="emb"))
+
+
+def step(y):
+ mem = memory(name="rnn_state", size=hidden_dim)
+ with mixed_layer(
+ name="rnn_state",
+ size=hidden_dim,
+ bias_attr=False,
+ act=SoftmaxActivation()) as out:
+ out += identity_projection(input=y)
+ out += full_matrix_projection(
+ input=mem, param_attr=ParamAttr(name="___recurrent_layer_0__"))
+ return out
+
+
+recurrent = recurrent_group(name="rnn", step=step, input=emb)
+
+recurrent_last = last_seq(input=recurrent)
+
+with mixed_layer(
+ size=label_dim, act=SoftmaxActivation(), bias_attr=True) as output:
+ output += full_matrix_projection(input=recurrent_last)
+
+outputs(
+ classification_cost(
+ input=output, label=data_layer(
+ name="label", size=1)))
diff --git a/paddle/trainer/tests/test_CompareSparse.cpp b/paddle/gserver/tests/test_CompareSparse.cpp
similarity index 98%
rename from paddle/trainer/tests/test_CompareSparse.cpp
rename to paddle/gserver/tests/test_CompareSparse.cpp
index 5f1834bd730375fc10762fc19788d0c693f8e752..c6e07650fc4805a25baf38b9059f6c996d00cafc 100644
--- a/paddle/trainer/tests/test_CompareSparse.cpp
+++ b/paddle/gserver/tests/test_CompareSparse.cpp
@@ -22,8 +22,7 @@ limitations under the License. */
using namespace paddle; // NOLINT
using namespace std; // NOLINT
-static const string& configFile1 =
- "trainer/tests/sample_trainer_config_compare_sparse.conf";
+static const string& configFile1 = "gserver/tests/sequence_lstm.conf";
DECLARE_bool(use_gpu);
DECLARE_string(config);
diff --git a/paddle/trainer/tests/test_CompareTwoNets.cpp b/paddle/gserver/tests/test_CompareTwoNets.cpp
similarity index 95%
rename from paddle/trainer/tests/test_CompareTwoNets.cpp
rename to paddle/gserver/tests/test_CompareTwoNets.cpp
index 94f65e545d116c802fb4877dc14f07aaaf83a4fb..801d9607565910b1f7f68a9c4532de5877e44f30 100644
--- a/paddle/trainer/tests/test_CompareTwoNets.cpp
+++ b/paddle/gserver/tests/test_CompareTwoNets.cpp
@@ -30,8 +30,6 @@ DECLARE_bool(use_gpu);
DECLARE_string(config);
DECLARE_string(nics);
-DEFINE_string(config_file_a, "", "config of one network to compare");
-DEFINE_string(config_file_b, "", "config of another network to compare");
DEFINE_bool(need_high_accuracy,
false,
"whether need to run in double accuracy");
@@ -42,6 +40,10 @@ DEFINE_double(
DECLARE_bool(thread_local_rand_use_global_seed);
DECLARE_int32(seed);
+static const string& config_file_a = "gserver/tests/sequence_recurrent.py";
+static const string& config_file_b =
+ "gserver/tests/sequence_recurrent_group.py";
+
struct ComData {
vector outArgs;
vector parameters;
@@ -66,6 +68,7 @@ void calcGradient(ComData& data, const string configFile) {
DataBatch dataBatch;
int32_t batchSize = trainer.getConfig().opt_config().batch_size();
+ trainer.getDataProvider()->reset();
trainer.getDataProvider()->setSkipShuffle();
trainer.getDataProvider()->getNextBatch(batchSize, &dataBatch);
@@ -167,11 +170,11 @@ void compareGradient(ComData& comDataA, ComData& comDataB) {
TEST(Trainer, create) {
ComData dataA;
- calcGradient(dataA, FLAGS_config_file_a);
+ calcGradient(dataA, config_file_a);
LOG(INFO) << "\n\nforwardBackward of Network A is finished\n\n";
ComData dataB;
- calcGradient(dataB, FLAGS_config_file_b);
+ calcGradient(dataB, config_file_b);
LOG(INFO) << "\n\nforwardBackward of the Network B is finished\n\n";
compareGradient(dataA, dataB);
diff --git a/paddle/gserver/tests/test_MKLDNN.cpp b/paddle/gserver/tests/test_MKLDNN.cpp
index a859e34c8996d81f14bf1edcb6e23d5a4f687e6b..42644e9601a82ea81c417adc6441edeb036998e2 100644
--- a/paddle/gserver/tests/test_MKLDNN.cpp
+++ b/paddle/gserver/tests/test_MKLDNN.cpp
@@ -313,6 +313,47 @@ TEST(MKLDNNLayer, AddtoLayer) {
testAddtoLayer({4, 12, 1, 1}, 3);
}
+static void getMKLDNNConcatConfig(TestConfig& cfg,
+ const std::vector& inputs) {
+ CHECK_GE(inputs.size(), 2) << "at least two inputs";
+ int oc = inputs[0].ic;
+ for (size_t i = 1; i < inputs.size(); ++i) {
+ CHECK_EQ(inputs[i].bs, inputs[0].bs);
+ CHECK_EQ(inputs[i].ih, inputs[0].ih);
+ CHECK_EQ(inputs[i].iw, inputs[0].iw);
+ oc += inputs[i].ic;
+ }
+ cfg.biasSize = 0;
+ cfg.layerConfig.set_type("mkldnn_concat");
+ cfg.layerConfig.set_size(oc * inputs[0].ih * inputs[0].iw);
+ cfg.layerConfig.set_active_type("relu");
+ for (size_t i = 0; i < inputs.size(); ++i) {
+ std::stringstream ss;
+ ss << "layer_" << i;
+ cfg.inputDefs.push_back(
+ {INPUT_DATA,
+ ss.str(),
+ (size_t)(inputs[i].ic) * inputs[i].ih * inputs[i].iw,
+ 0});
+ LayerInputConfig* input = cfg.layerConfig.add_inputs();
+ ImageConfig* img_conf = input->mutable_image_conf();
+ img_conf->set_channels(inputs[i].ic);
+ img_conf->set_img_size_y(inputs[i].ih);
+ img_conf->set_img_size(inputs[i].iw);
+ }
+}
+
+void testConcatLayer(const std::vector& inputs) {
+ TestConfig dnnConfig;
+ getMKLDNNConcatConfig(dnnConfig, inputs);
+ RUN_MKLDNN_TEST_LAYER(dnnConfig, "concat", inputs[0])
+}
+
+TEST(MKLDNNLayer, ConcatLayer) {
+ testConcatLayer({{64, 128, 1, 1}, {64, 32, 1, 1}, {64, 64, 1, 1}});
+ testConcatLayer({{32, 100, 8, 8}, {32, 10, 8, 8}});
+}
+
void testActivation(std::string actType, const testImageDesc& pm) {
// TODO(TJ): remove me when paddle support elu activation
if (actType == "mkldnn_elu") {
diff --git a/paddle/gserver/tests/test_ProtoDataProvider.cpp b/paddle/gserver/tests/test_ProtoDataProvider.cpp
deleted file mode 100644
index af6472619d1840e82787974d265d601b4a406c09..0000000000000000000000000000000000000000
--- a/paddle/gserver/tests/test_ProtoDataProvider.cpp
+++ /dev/null
@@ -1,732 +0,0 @@
-/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
- http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License. */
-
-#include
-#include
-
-#include
-
-#include "paddle/gserver/dataproviders/ProtoDataProvider.h"
-#include "paddle/utils/Util.h"
-
-#include "paddle/testing/TestUtil.h"
-
-using namespace std; // NOLINT
-
-std::vector protoFiles{
- "./test_ProtoDataProvider/data1.bin", "./test_ProtoDataProvider/data2.bin",
-};
-std::vector protoFilesCompressed{
- "./test_ProtoDataProvider/data1.bin.gz",
- "./test_ProtoDataProvider/data2.bin.gz",
-};
-
-const char* kTestDir = "./test_ProtoDataProvider";
-const char kProtoFileList[] = "gserver/tests/proto_files.txt";
-const char kProtoFileListCompressed[] =
- "gserver/tests/proto_files_compressed.txt";
-const int kSpraseMatrixDim = 1024;
-
-using namespace paddle; // NOLINT
-
-void prepareData(DataBatch* batch,
- const int* numPerSlotType,
- bool iid,
- bool useGpu) {
- batch->clear();
- int64_t size = uniformRandom(100) + 10;
- batch->setSize(size);
-
- ICpuGpuVectorPtr sequenceStartPositions;
- ICpuGpuVectorPtr subSequenceStartPositions;
- if (!iid) {
- int numSeqs = uniformRandom(10) + 1;
- sequenceStartPositions =
- ICpuGpuVector::create(numSeqs + 1, /* useGpu= */ false);
- int* buf = sequenceStartPositions->getMutableData(false);
- subSequenceStartPositions =
- ICpuGpuVector::create(numSeqs + 1, /* useGpu= */ false);
- int* subBuf = subSequenceStartPositions->getMutableData(false);
- int64_t pos = 0;
- int maxLen = 2 * size / numSeqs;
- for (int i = 0; i < numSeqs; ++i) {
- int len =
- uniformRandom(min(maxLen, size - pos - numSeqs + i)) + 1;
- buf[i] = pos;
- subBuf[i] = pos;
- pos += len;
- VLOG(1) << " len=" << len;
- }
- buf[numSeqs] = size;
- subBuf[numSeqs] = size;
- }
-
- vector& arguments = batch->getStreams();
- for (int i = 0; i < numPerSlotType[SlotDef::VECTOR_DENSE]; ++i) {
- int64_t dim = rand() % 10 + 4; // NOLINT rand_r
- MatrixPtr mat = Matrix::create(size, dim, /* trans= */ false, false);
- mat->randomizeUniform();
- Argument arg;
- arg.value = mat;
- arg.sequenceStartPositions = sequenceStartPositions;
- arguments.push_back(arg);
- }
- for (int i = 0; i < numPerSlotType[SlotDef::VECTOR_SPARSE_NON_VALUE]; ++i) {
- MatrixPtr mat =
- makeRandomSparseMatrix(size, kSpraseMatrixDim, false, useGpu);
- Argument arg;
- arg.value = mat;
- arg.sequenceStartPositions = sequenceStartPositions;
- arg.subSequenceStartPositions = subSequenceStartPositions;
- arguments.push_back(arg);
- }
- for (int i = 0; i < numPerSlotType[SlotDef::VECTOR_SPARSE_VALUE]; ++i) {
- MatrixPtr mat =
- makeRandomSparseMatrix(size, kSpraseMatrixDim, true, useGpu);
- Argument arg;
- arg.value = mat;
- arg.sequenceStartPositions = sequenceStartPositions;
- arguments.push_back(arg);
- }
- for (int i = 0; i < numPerSlotType[SlotDef::STRING]; ++i) {
- int64_t dim = rand() % 10 + 4; // NOLINT rand_r
- SVectorPtr vec = std::make_shared>();
- for (int j = 0; j < size; ++j) {
- vec->push_back(randStr(dim));
- }
- Argument arg;
- arg.strs = vec;
- arg.sequenceStartPositions = sequenceStartPositions;
- arguments.push_back(arg);
- }
- for (int i = 0; i < numPerSlotType[SlotDef::INDEX]; ++i) {
- int64_t dim = rand() % 10 + 4; // NOLINT rand_r
- IVectorPtr vec = IVector::create(size, /* useGpu= */ false);
- int* buf = vec->getData();
- for (int j = 0; j < size; ++j) {
- buf[j] = uniformRandom(dim);
- }
- Argument arg;
- arg.ids = vec;
- arg.sequenceStartPositions = sequenceStartPositions;
- arguments.push_back(arg);
- }
-}
-
-inline int getSlotDim(const Argument& arg) {
- if (arg.value) {
- return arg.value->getWidth();
- } else if (arg.ids) {
- return arg.ids->getMax() + 1;
- } else if (arg.strs) {
- return 1;
- }
- LOG(FATAL) << "Invalid argument";
- return 0;
-}
-
-inline SlotDef::SlotType getSlotType(const Argument& arg) {
- if (arg.value) {
- auto& m = *arg.value;
- auto& type = typeid(m);
- if (type == typeid(CpuMatrix) || type == typeid(GpuMatrix)) {
- return SlotDef::VECTOR_DENSE;
- }
- if (type == typeid(CpuSparseMatrix)) {
- auto valueType =
- std::dynamic_pointer_cast(arg.value)->getValueType();
- if (NO_VALUE == valueType) {
- return SlotDef::VECTOR_SPARSE_NON_VALUE;
- } else {
- return SlotDef::VECTOR_SPARSE_VALUE;
- }
- }
- if (type == typeid(GpuSparseMatrix)) {
- auto valueType =
- std::dynamic_pointer_cast(arg.value)->getValueType();
- if (NO_VALUE == valueType) {
- return SlotDef::VECTOR_SPARSE_NON_VALUE;
- } else {
- return SlotDef::VECTOR_SPARSE_VALUE;
- }
- }
-
- LOG(FATAL) << "Unknown matrix type";
- }
- if (arg.ids) return SlotDef::INDEX;
- if (arg.strs) return SlotDef::STRING;
- LOG(FATAL) << "Invalid argument";
- return SlotDef::VECTOR_DENSE;
-}
-
-void getColRow(const Argument& arg,
- int64_t pos,
- bool useGpu,
- int* colNum,
- const int** rowCols,
- const real** rowValues) {
- SlotDef::SlotType type = getSlotType(arg);
- GpuSparseMatrixPtr matGpu;
- CpuSparseMatrixPtr matCpu;
- if (useGpu) {
- matGpu = dynamic_pointer_cast(arg.value);
- ASSERT_TRUE(matGpu != NULL);
- } else {
- matCpu = dynamic_pointer_cast(arg.value);
- ASSERT_TRUE(matCpu != NULL);
- }
- *colNum = useGpu ? matGpu->getColNum(pos) : matCpu->getColNum(pos);
- *rowCols = useGpu ? matGpu->getRowCols(pos) : matCpu->getRowCols(pos);
- if (type == SlotDef::VECTOR_SPARSE_VALUE) {
- *rowValues = useGpu ? matGpu->getRowValues(pos) : matCpu->getRowValues(pos);
- } else {
- *rowValues = NULL;
- }
-}
-
-void makeSample(const vector& arguments,
- int64_t pos,
- bool isBeginning,
- DataSample* sample,
- bool useGpu) {
- sample->set_is_beginning(isBeginning);
- int slotid = 0;
- for (auto& arg : arguments) {
- SlotDef::SlotType type = getSlotType(arg);
- int64_t dim = getSlotDim(arg);
- switch (type) {
- case SlotDef::VECTOR_DENSE: {
- VectorSlot* vecSlot = sample->add_vector_slots();
- auto values = vecSlot->mutable_values();
- values->Reserve(dim);
- for (int i = 0; i < dim; ++i) {
- values->AddAlreadyReserved(
- static_cast(arg.value->getElement(pos, i)));
- }
- break;
- }
- case SlotDef::INDEX: {
- sample->add_id_slots(arg.ids->get(pos));
- break;
- }
- case SlotDef::VECTOR_SPARSE_NON_VALUE: {
- VectorSlot* vecSlot = sample->add_vector_slots();
- auto ids = vecSlot->mutable_ids();
- int colNum;
- const int* rowCols;
- const real* rowValues; // nullptr
- getColRow(arg, pos, useGpu, &colNum, &rowCols, &rowValues);
- ids->Reserve(colNum);
- for (int i = 0; i < colNum; ++i) {
- ids->AddAlreadyReserved(rowCols[i]);
- }
- SubseqSlot* subseqSlot = sample->add_subseq_slots(); // subseq
- subseqSlot->set_slot_id(slotid);
- auto lens = subseqSlot->mutable_lens();
- lens->Add(colNum);
- break;
- }
- case SlotDef::VECTOR_SPARSE_VALUE: {
- VectorSlot* vecSlot = sample->add_vector_slots();
- auto values = vecSlot->mutable_values();
- auto ids = vecSlot->mutable_ids();
- int colNum;
- const int* rowCols;
- const real* rowValues;
- getColRow(arg, pos, useGpu, &colNum, &rowCols, &rowValues);
- ids->Reserve(colNum);
- values->Reserve(colNum);
- for (int i = 0; i < colNum; ++i) {
- ids->AddAlreadyReserved(rowCols[i]);
- values->AddAlreadyReserved(rowValues[i]);
- }
- break;
- }
- case SlotDef::VAR_MDIM_DENSE:
- case SlotDef::VAR_MDIM_INDEX: {
- LOG(FATAL) << "Not implemented";
- break;
- }
- case SlotDef::STRING: {
- VectorSlot* vecSlot = sample->add_vector_slots();
- vecSlot->add_strs((*arg.strs)[pos]);
- break;
- }
- }
- slotid++;
- }
-}
-
-void writeData(const DataBatch& batch, bool useGpu, bool dataCompression) {
- DataHeader header;
- const vector& arguments = batch.getStreams();
- for (auto& argument : arguments) {
- SlotDef* slotDef = header.add_slot_defs();
- slotDef->set_type(getSlotType(argument));
- slotDef->set_dim(getSlotDim(argument));
- }
- VLOG(1) << "header=" << header.DebugString();
-
- int64_t totalSeqs = batch.getNumSequences();
- int64_t seq = 0;
- ICpuGpuVectorPtr sequenceStartPositions = arguments[0].sequenceStartPositions;
- int64_t numWritten = 0;
- vector curProtoFiles =
- dataCompression ? protoFilesCompressed : protoFiles;
- for (size_t i = 0; i < curProtoFiles.size(); ++i) {
- int64_t numSeqs = totalSeqs * (i + 1) / curProtoFiles.size() -
- totalSeqs * i / curProtoFiles.size();
- ofstream os(curProtoFiles[i]);
- CHECK(os) << "Fail to open " << curProtoFiles[i];
- unique_ptr writer(new ProtoWriter(&os, dataCompression));
- CHECK(writer->write(header));
- for (int j = 0; j < numSeqs; ++j, ++seq) {
- int64_t begin = seq;
- int64_t end = seq + 1;
- if (sequenceStartPositions) {
- begin = sequenceStartPositions->getElement(seq);
- end = sequenceStartPositions->getElement(seq + 1);
- }
- for (int pos = begin; pos < end; ++pos) {
- DataSample sample;
- makeSample(arguments, pos, pos == begin, &sample, useGpu);
- CHECK(writer->write(sample));
- ++numWritten;
- }
- }
-
- writer.reset(nullptr);
- os.close();
- }
- CHECK_EQ(arguments[0].getBatchSize(), numWritten);
-}
-
-// check that the sample at pos1 in args1 is same as the sample at pos2 in args2
-void checkSample(const vector& args1,
- int64_t pos1,
- const vector& args2,
- int64_t pos2,
- bool useGpu) {
- EXPECT_EQ(args1.size(), args2.size());
- VLOG(1) << " pos1=" << pos1 << " pos2=" << pos2;
-
- for (size_t i = 0; i < args1.size(); ++i) {
- auto type = getSlotType(args1[i]);
- int dim = getSlotDim(args1[i]);
- EXPECT_EQ(type, getSlotType(args2[i]));
- if (type == SlotDef::INDEX) {
- EXPECT_GE(dim, getSlotDim(args2[i]));
- } else {
- EXPECT_EQ(dim, getSlotDim(args2[i]));
- }
- switch (type) {
- case SlotDef::VECTOR_DENSE: {
- for (int j = 0; j < dim; ++j) {
- EXPECT_EQ(static_cast(args1[i].value->getElement(pos1, j)),
- static_cast(args2[i].value->getElement(pos2, j)));
- }
- break;
- }
- case SlotDef::INDEX: {
- EXPECT_EQ(args1[i].ids->get(pos1), args2[i].ids->get(pos2));
- break;
- }
- case SlotDef::VECTOR_SPARSE_NON_VALUE:
- case SlotDef::VECTOR_SPARSE_VALUE: {
- int colNum1, colNum2;
- const int *rowCols1, *rowCols2;
- const real *rowValues1, *rowValues2;
- getColRow(args1[i], pos1, useGpu, &colNum1, &rowCols1, &rowValues1);
- getColRow(args2[i], pos2, useGpu, &colNum2, &rowCols2, &rowValues2);
- EXPECT_EQ(colNum1, colNum2);
- for (int j = 0; j < colNum1; ++j) {
- EXPECT_EQ(rowCols1[j], rowCols2[j]);
- if (type == SlotDef::VECTOR_SPARSE_VALUE) {
- EXPECT_EQ(rowValues1[j], rowValues2[j]);
- }
- }
- break;
- }
- case SlotDef::VAR_MDIM_DENSE:
- case SlotDef::VAR_MDIM_INDEX: {
- LOG(FATAL) << "Not implemented";
- break;
- }
- case SlotDef::STRING: {
- EXPECT_EQ((*args1[i].strs)[pos1], (*args2[i].strs)[pos2]);
- break;
- }
- }
- }
-}
-
-void testProtoDataProvider(int* numPerSlotType,
- bool iid,
- bool async,
- bool useGpu,
- bool dataCompression,
- int numConstantSlots = 0) {
- mkDir(kTestDir);
- DataBatch data;
-
- prepareData(&data, numPerSlotType, iid, useGpu);
- writeData(data, useGpu, dataCompression);
-
- DataConfig config;
- config.set_type("proto");
- config.set_files(dataCompression ? kProtoFileListCompressed : kProtoFileList);
- config.set_async_load_data(async);
-
- for (int i = 0; i < numConstantSlots; ++i) {
- config.add_constant_slots(i + 11);
- MatrixPtr w = Matrix::create(data.getSize(),
- 1,
- /* trans= */ false,
- /* useGpu= */ false);
- w->assign(config.constant_slots(i));
- data.appendData(w);
- }
-
- unique_ptr dataProvider(DataProvider::create(config, useGpu));
- dataProvider->setSkipShuffle();
-
- EXPECT_EQ(data.getSize(), dataProvider->getSize());
-
- int64_t batchSize = 10;
- DataBatch batch;
-
- size_t seq1 = 0;
- vector& args1 = data.getStreams();
- ICpuGpuVectorPtr sequenceStartPositions1 = args1[0].sequenceStartPositions;
-
- dataProvider->reset();
-
- while (dataProvider->getNextBatch(batchSize, &batch) > 0) {
- CHECK_EQ(data.getNumStreams(), batch.getNumStreams());
- vector& args2 = batch.getStreams();
- ICpuGpuVectorPtr sequenceStartPositions2 = args2[0].sequenceStartPositions;
- for (auto& arg : args2) {
- EXPECT_EQ(iid, !arg.sequenceStartPositions);
- }
- size_t numSeqs = batch.getNumSequences();
- VLOG(1) << "numSeqs=" << numSeqs;
- for (size_t seq2 = 0; seq2 < numSeqs; ++seq1, ++seq2) {
- int64_t begin1 = seq1;
- int64_t end1 = seq1 + 1;
- if (sequenceStartPositions1) {
- begin1 = sequenceStartPositions1->getElement(seq1);
- end1 = sequenceStartPositions1->getElement(seq1 + 1);
- EXPECT_LT(seq1, sequenceStartPositions1->getSize() - 1);
- }
-
- int64_t begin2 = seq2;
- int64_t end2 = seq2 + 1;
- if (sequenceStartPositions2) {
- begin2 = sequenceStartPositions2->getElement(seq2);
- end2 = sequenceStartPositions2->getElement(seq2 + 1);
- }
- VLOG(1) << " begin1=" << begin1 << " end1=" << end1
- << " begin2=" << begin2 << " end2=" << end2;
- EXPECT_EQ(end1 - begin1, end2 - begin2);
- for (int i = 0; i < end1 - begin1; ++i) {
- checkSample(args1, begin1 + i, args2, begin2 + i, useGpu);
- }
- }
- }
-
- EXPECT_EQ(seq1, (size_t)data.getNumSequences());
- rmDir(kTestDir);
-}
-
-TEST(ProtoDataProvider, test) {
- int numSlotsArray[] = {0, 3};
- int numTwoArray[] = {0, 1};
- int numSlotsArraySize = sizeof(numSlotsArray) / sizeof(numSlotsArray[0]);
- const int numSlot = 5;
- int combination[numSlot] = {0};
- int k = numSlot - 1;
- while (k >= 0) {
- int numDenseVecSlots = numSlotsArray[combination[0]];
- int numSparseNonValueVecSlots = numSlotsArray[combination[1]];
- int numSparseValueVectorSlots = numSlotsArray[combination[2]];
- int numStrSlots = numSlotsArray[combination[3]];
- int numIdSlots = numSlotsArray[combination[4]];
- // while loop : traverse all cases
- k = numSlot - 1;
- while (k >= 0) {
- if (combination[k] < (numSlotsArraySize - 1)) {
- ++combination[k];
- break;
- } else {
- combination[k] = 0;
- --k;
- }
- }
- if (numDenseVecSlots + numSparseNonValueVecSlots +
- numSparseValueVectorSlots + numStrSlots + numIdSlots <
- 1)
- continue;
- for (int iid : numTwoArray) {
- for (int async : numTwoArray) {
- for (int useGpu : numTwoArray) {
- for (int dataCompression : numTwoArray) {
- if (async && useGpu) {
- // Currently in async mode, useGpu is not supported
- continue;
- }
-#ifndef PADDLE_WITH_CUDA
- if (useGpu) {
- continue;
- }
-#endif
- LOG(INFO) << " numDenseVecSlots=" << numDenseVecSlots
- << " numSparseNonValueVecSlots="
- << numSparseNonValueVecSlots
- << " numSparseValueVectorSlots="
- << numSparseValueVectorSlots
- << " numStrSlots=" << numStrSlots
- << " numIdSlots=" << numIdSlots << " iid=" << iid
- << " async=" << async << " useGpu=" << useGpu
- << " dataCompression=" << dataCompression;
- int numPerSlotType[SlotDef::SlotType_ARRAYSIZE] = {0};
- numPerSlotType[SlotDef::VECTOR_DENSE] = numDenseVecSlots;
- numPerSlotType[SlotDef::VECTOR_SPARSE_NON_VALUE] =
- numSparseNonValueVecSlots;
- numPerSlotType[SlotDef::VECTOR_SPARSE_VALUE] =
- numSparseValueVectorSlots;
- numPerSlotType[SlotDef::INDEX] = numIdSlots;
- numPerSlotType[SlotDef::STRING] = numStrSlots;
- testProtoDataProvider(
- numPerSlotType, iid, async, useGpu, dataCompression);
- } // end for (int dataCompression : numTwoArray)
- } // end for (int useGpu : numTwoArray)
- } // end for (int async : numTwoArray)
- } // end for (int iid : numTwoArray)
- } // end for (while, traverse all slots)
-}
-
-TEST(ProtoDataProvider, constant_slots) {
- int numSlotsArray[] = {0, 3};
- int numTwoArray[] = {0, 1};
- for (int numDenseVecSlots : numSlotsArray) {
- for (int numSparseNonValueVecSlots : numSlotsArray) {
- if (numDenseVecSlots + numSparseNonValueVecSlots < 1) continue;
- for (int numConstantSlots : {1, 2}) {
- for (int useGpu : numTwoArray) {
- for (int dataCompression : numTwoArray) {
-#ifndef PADDLE_WITH_CUDA
- if (useGpu) {
- continue;
- }
-#endif
- LOG(INFO) << " numDenseVecSlots=" << numDenseVecSlots
- << " numSparseNonValueVecSlots="
- << numSparseNonValueVecSlots
- << " numConstantSlogs=" << numConstantSlots
- << " useGpu=" << useGpu
- << " dataCompression=" << dataCompression;
- int numPerSlotType[SlotDef::SlotType_ARRAYSIZE] = {0};
- numPerSlotType[SlotDef::VECTOR_DENSE] = numDenseVecSlots;
- numPerSlotType[SlotDef::VECTOR_SPARSE_NON_VALUE] =
- numSparseNonValueVecSlots;
- numPerSlotType[SlotDef::VECTOR_SPARSE_VALUE] = 1;
- numPerSlotType[SlotDef::INDEX] = 1;
- testProtoDataProvider(numPerSlotType,
- /* iid= */ true,
- /* async= */ false,
- useGpu,
- dataCompression,
- numConstantSlots);
- } // end for (int dataCompression : numTwoArray)
- } // end for (int useGpu : numTwoArray)
- } // end for (int numConstantSlots : {1, 2})
- } // end for (int numSparseNonValueVecSlots : numSlotsArray)
- } // end for (int numDenseVecSlots : numSlotsArray)
-}
-
-void checkSampleSequence(const vector& args1,
- const vector& args2,
- int64_t offset,
- int64_t numSeqs,
- bool useGpu) {
- // check slot num are equal
- EXPECT_EQ(args1.size(), args2.size());
- for (size_t i = 0; i < args1.size(); i++) {
- auto type = getSlotType(args1[i]);
- // check for args2: sequenceStartPositions vs numSeqs
- // (1) size
- EXPECT_EQ(args2[i].sequenceStartPositions->getSize(), (size_t)numSeqs + 1);
- // (2) content
- auto checkArgContent = [&](const Argument& args, int numSeqs) {
- for (int j = 0; j <= numSeqs; j++) {
- int start_pos = args.sequenceStartPositions->getElement(j);
- EXPECT_EQ(start_pos, j);
- }
- };
- switch (type) {
- case SlotDef::INDEX: {
- // args1: for label
- checkArgContent(args2[i], numSeqs);
- // check for args2: ids are equal to args1[offset]
- // (1) size
- EXPECT_EQ(args2[i].ids->getSize(), (size_t)numSeqs);
- // (2) content
- for (int j = 0; j < numSeqs; j++) {
- EXPECT_EQ(args2[i].ids->get(j), args1[i].ids->get(offset + j));
- }
- break;
- }
- case SlotDef::VECTOR_SPARSE_NON_VALUE: {
- // args1: for sparse_non_value
- // args2 should put sparse indexes in ids
- int colNum1;
- const int* rowCols1;
- const real* rowValues1; // nullptr
- int totalLength = 0;
- for (int j = 0; j < numSeqs; j++) {
- getColRow(
- args1[i], offset + j, useGpu, &colNum1, &rowCols1, &rowValues1);
- // (1) lengths
- EXPECT_EQ(totalLength,
- args2[i].sequenceStartPositions->getElement(j));
- EXPECT_EQ(totalLength,
- args2[i].subSequenceStartPositions->getElement(j));
- // (2) content
- for (int k = 0; k < colNum1; k++) {
- EXPECT_EQ(rowCols1[k], args2[i].ids->get(totalLength + k));
- }
- totalLength += colNum1;
- if (colNum1 == 0) {
- // special case here: we will put a "-1" into ids when column num is
- // zero. see ProtoSequenceDataProvider::getNextBatchInternal.
- EXPECT_EQ(-1, args2[i].ids->get(totalLength));
- totalLength++;
- }
- }
- EXPECT_EQ(totalLength,
- args2[i].sequenceStartPositions->getElement(numSeqs));
- EXPECT_EQ(totalLength,
- args2[i].subSequenceStartPositions->getElement(numSeqs));
- break;
- }
- case SlotDef::VECTOR_DENSE: {
- // args1: for dense vector
- checkArgContent(args2[i], numSeqs);
- // check for args2: values are equal to args1[offset]
- // (1) size
- EXPECT_EQ(args2[i].value->getHeight(), (size_t)numSeqs);
- EXPECT_EQ(args2[i].value->getWidth(), (size_t)getSlotDim(args1[i]));
- // (2) content
- for (int j = 0; j < numSeqs; j++) {
- for (size_t k = 0; k < args2[i].value->getWidth(); k++) {
- EXPECT_EQ(
- static_cast(args1[i].value->getElement(j + offset, k)),
- static_cast(args2[i].value->getElement(j, k)));
- }
- }
- break;
- }
- default: { EXPECT_EQ(true, false) << "should not reach here"; }
- }
- }
-}
-
-void testProtoSequenceDataProvider(int* numPerSlotType,
- bool async,
- bool useGpu) {
- mkDir(kTestDir);
- DataBatch data;
-
- prepareData(&data,
- numPerSlotType,
- /* iid */ true,
- useGpu);
- writeData(data, useGpu, /* dataCompression */ false);
-
- DataConfig config;
- config.set_type("proto_sequence");
- config.set_files(kProtoFileList);
- config.set_async_load_data(async);
-
- unique_ptr dataProvider(DataProvider::create(config, useGpu));
- dataProvider->setSkipShuffle();
-
- EXPECT_EQ(data.getSize(), dataProvider->getSize());
-
- int64_t batchSize = 10;
- DataBatch batch;
-
- vector& args1 = data.getStreams();
- ICpuGpuVectorPtr sequenceStartPositions1 = args1[0].sequenceStartPositions;
-
- dataProvider->reset();
-
- size_t args1Offset = 0;
- while (dataProvider->getNextBatch(batchSize, &batch) > 0) {
- CHECK_EQ(data.getNumStreams(), batch.getNumStreams());
- vector& args2 = batch.getStreams();
- ICpuGpuVectorPtr sequenceStartPositions2 = args2[0].sequenceStartPositions;
- for (auto& arg : args1) {
- // args1 should not has sequence
- EXPECT_EQ(true, !arg.sequenceStartPositions);
- }
- for (auto& arg : args2) {
- // args2 should has sequence
- EXPECT_NE(true, !arg.sequenceStartPositions);
- }
- size_t numSeqs = batch.getNumSequences();
- checkSampleSequence(args1, args2, args1Offset, numSeqs, useGpu);
- args1Offset += numSeqs;
- }
-
- EXPECT_EQ(args1Offset, (size_t)data.getNumSequences());
- rmDir(kTestDir);
-}
-
-TEST(ProtoSequenceDataProvider, test) {
- int numSlotsArray[] = {0, 3};
- int numTwoArray[] = {0, 1};
- for (int numSparseNonValueVecSlots : numSlotsArray) {
- for (int numIdSlots : numSlotsArray) {
- for (int numDenseVecSlots : numSlotsArray) {
- if (numDenseVecSlots + numSparseNonValueVecSlots + numIdSlots < 1)
- continue;
- for (int async : numTwoArray) {
- for (int useGpu : numTwoArray) {
- if (async && useGpu) {
- // Currently in async mode, useGpu is not supported
- continue;
- }
-#ifndef PADDLE_WITH_CUDA
- if (useGpu) {
- continue;
- }
-#endif
- LOG(INFO) << " numDenseVecSlots=" << numDenseVecSlots
- << " numSparseNonValueVecSlots="
- << numSparseNonValueVecSlots
- << " numIdSlots=" << numIdSlots << " async=" << async
- << " useGpu=" << useGpu;
- int numPerSlotType[SlotDef::SlotType_ARRAYSIZE] = {0};
- numPerSlotType[SlotDef::VECTOR_DENSE] = numDenseVecSlots;
- numPerSlotType[SlotDef::VECTOR_SPARSE_NON_VALUE] =
- numSparseNonValueVecSlots;
- numPerSlotType[SlotDef::INDEX] = numIdSlots;
- testProtoSequenceDataProvider(numPerSlotType, async, useGpu);
- } // end for (int useGpu : numTwoArray)
- } // end for (int async : numTwoArray)
- } // end for (int numDenseVecSlots : numSlotsArray)
- } // end for (int numIdSlots : numSlotsArray)
- } // end for (int numSparseNonValueVecSlots : numSlotsArray)
-}
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