Skip to content

P
Paddle

Crayon鑫 / Paddle
与 Fork 源项目一致

Fork自 PaddlePaddle / Paddle

通知 1
Star 1
Fork 0
P
Paddle
提交 469b3ad5 编写于 作者: X xzl
浏览文件
操作

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into add_dilation

上级 83818cd4 2f3d1c34
展开全部 隐藏空白更改
内联 并排
Showing with 21413 addition and 4671 deletion
.travis.yml
浏览文件 @ 469b3ad5
......@@ -30,6 +30,7 @@ addons:
- automake
- libtool
- ccache
ssh_known_hosts: 52.76.173.135
before_install:
- if [[ "$JOB" == "check_style" ]]; then sudo ln -s /usr/bin/clang-format-3.8 /usr/bin/clang-format; fi
# Paddle is using protobuf 3.1 currently. Protobuf 3.2 breaks the compatibility. So we specify the python
......@@ -42,6 +43,14 @@ script:
- |
timeout 2580 paddle/scripts/travis/${JOB}.sh # 43min timeout
RESULT=$?; if [ $RESULT -eq 0 ] || [ $RESULT -eq 142 ]; then true; else false; fi;
- |
if [[ "$JOB" != "build_doc" ]]; then exit 0; fi;
if [[ "$TRAVIS_PULL_REQUEST" != "false" ]]; then exit 0; fi;
if [[ "$TRAVIS_BRANCH" != "develop" && ! "$TRAVIS_BRANCH" =~ ^v[[:digit:]]+\.[[:digit:]]+(\.[[:digit:]]+)?(-\S*)?$ ]]; then exit 0; fi;
export DEPLOY_DOCS_SH=https://raw.githubusercontent.com/PaddlePaddle/PaddlePaddle.org/master/scripts/deploy/deploy_docs.sh
export DOCS_DIR=`pwd`
cd ..
curl $DEPLOY_DOCS_SH | bash -s $CONTENT_DEC_PASSWD $TRAVIS_BRANCH $DOCS_DIR $DOCS_DIR/build/doc
notifications:
email:
on_success: change
......
CMakeLists.txt
浏览文件 @ 469b3ad5
......@@ -126,7 +126,7 @@ include(external/swig) # download, build, install swig
include(external/warpctc) # download, build, install warpctc
include(external/any) # download libn::any
include(external/eigen) # download eigen3
include(external/pybind11) # download pybind11
include(external/pybind11) # download pybind11
include(external/nccl)
include(cudnn) # set cudnn libraries, must before configure
......
benchmark/IntelOptimizedPaddle.md 0 → 100644
浏览文件 @ 469b3ad5
# Benchmark
Machine:
- Server
- Intel(R) Xeon(R) Gold 6148 CPU @ 2.40GHz, 2 Sockets, 20 Cores per socket
- Laptop
- DELL XPS15-9560-R1745: i7-7700HQ 8G 256GSSD
- i5 MacBook Pro (Retina, 13-inch, Early 2015)
- Desktop
- i7-6700k
System: CentOS release 6.3 (Final), Docker 1.12.1.
PaddlePaddle: paddlepaddle/paddle:latest (TODO: will rerun after 0.11.0)
- MKL-DNN tag v0.10
- MKLML 2018.0.20170720
- OpenBLAS v0.2.20
On each machine, we will test and compare the performance of training on single node using MKL-DNN / MKLML / OpenBLAS respectively.
## Benchmark Model
### Server
Test on batch size 64, 128, 256 on Intel(R) Xeon(R) Gold 6148 CPU @ 2.40GHz
Input image size - 3 * 224 * 224, Time: images/second
- VGG-19
| BatchSize | 64 | 128 | 256 |
|--------------|-------| -----| --------|
| OpenBLAS | 7.82 | 8.62 | 10.34 |
| MKLML | 11.02 | 12.86 | 15.33 |
| MKL-DNN | 27.69 | 28.8 | 29.27 |
chart on batch size 128
TBD
- ResNet
- GoogLeNet
### Laptop
TBD
### Desktop
TBD
benchmark/paddle/image/resnet.py 0 → 100644
浏览文件 @ 469b3ad5
#!/usr/bin/env python
from paddle.trainer_config_helpers import *
height = 224
width = 224
num_class = 1000
batch_size = get_config_arg('batch_size', int, 64)
layer_num = get_config_arg("layer_num", int, 50)
is_test = get_config_arg("is_test", bool, False)
args = {'height': height, 'width': width, 'color': True, 'num_class': num_class}
define_py_data_sources2(
"train.list", None, module="provider", obj="process", args=args)
settings(
batch_size=batch_size,
learning_rate=0.01 / batch_size,
learning_method=MomentumOptimizer(0.9),
regularization=L2Regularization(0.0005 * batch_size))
#######################Network Configuration #############
def conv_bn_layer(name,
input,
filter_size,
num_filters,
stride,
padding,
channels=None,
active_type=ReluActivation()):
"""
A wrapper for conv layer with batch normalization layers.
Note:
conv layer has no activation.
"""
tmp = img_conv_layer(
name=name + "_conv",
input=input,
filter_size=filter_size,
num_channels=channels,
num_filters=num_filters,
stride=stride,
padding=padding,
act=LinearActivation(),
bias_attr=False)
return batch_norm_layer(
name=name + "_bn", input=tmp, act=active_type, use_global_stats=is_test)
def bottleneck_block(name, input, num_filters1, num_filters2):
"""
A wrapper for bottlenect building block in ResNet.
Last conv_bn_layer has no activation.
Addto layer has activation of relu.
"""
last_name = conv_bn_layer(
name=name + '_branch2a',
input=input,
filter_size=1,
num_filters=num_filters1,
stride=1,
padding=0)
last_name = conv_bn_layer(
name=name + '_branch2b',
input=last_name,
filter_size=3,
num_filters=num_filters1,
stride=1,
padding=1)
last_name = conv_bn_layer(
name=name + '_branch2c',
input=last_name,
filter_size=1,
num_filters=num_filters2,
stride=1,
padding=0,
active_type=LinearActivation())
return addto_layer(
name=name + "_addto", input=[input, last_name], act=ReluActivation())
def mid_projection(name, input, num_filters1, num_filters2, stride=2):
"""
A wrapper for middile projection in ResNet.
projection shortcuts are used for increasing dimensions,
and other shortcuts are identity
branch1: projection shortcuts are used for increasing
dimensions, has no activation.
branch2x: bottleneck building block, shortcuts are identity.
"""
# stride = 2
branch1 = conv_bn_layer(
name=name + '_branch1',
input=input,
filter_size=1,
num_filters=num_filters2,
stride=stride,
padding=0,
active_type=LinearActivation())
last_name = conv_bn_layer(
name=name + '_branch2a',
input=input,
filter_size=1,
num_filters=num_filters1,
stride=stride,
padding=0)
last_name = conv_bn_layer(
name=name + '_branch2b',
input=last_name,
filter_size=3,
num_filters=num_filters1,
stride=1,
padding=1)
last_name = conv_bn_layer(
name=name + '_branch2c',
input=last_name,
filter_size=1,
num_filters=num_filters2,
stride=1,
padding=0,
active_type=LinearActivation())
return addto_layer(
name=name + "_addto", input=[branch1, last_name], act=ReluActivation())
img = data_layer(name='image', size=height * width * 3)
def deep_res_net(res2_num=3, res3_num=4, res4_num=6, res5_num=3):
"""
A wrapper for 50,101,152 layers of ResNet.
res2_num: number of blocks stacked in conv2_x
res3_num: number of blocks stacked in conv3_x
res4_num: number of blocks stacked in conv4_x
res5_num: number of blocks stacked in conv5_x
"""
# For ImageNet
# conv1: 112x112
tmp = conv_bn_layer(
"conv1",
input=img,
filter_size=7,
channels=3,
num_filters=64,
stride=2,
padding=3)
tmp = img_pool_layer(name="pool1", input=tmp, pool_size=3, stride=2)
# conv2_x: 56x56
tmp = mid_projection(
name="res2_1", input=tmp, num_filters1=64, num_filters2=256, stride=1)
for i in xrange(2, res2_num + 1, 1):
tmp = bottleneck_block(
name="res2_" + str(i), input=tmp, num_filters1=64, num_filters2=256)
# conv3_x: 28x28
tmp = mid_projection(
name="res3_1", input=tmp, num_filters1=128, num_filters2=512)
for i in xrange(2, res3_num + 1, 1):
tmp = bottleneck_block(
name="res3_" + str(i),
input=tmp,
num_filters1=128,
num_filters2=512)
# conv4_x: 14x14
tmp = mid_projection(
name="res4_1", input=tmp, num_filters1=256, num_filters2=1024)
for i in xrange(2, res4_num + 1, 1):
tmp = bottleneck_block(
name="res4_" + str(i),
input=tmp,
num_filters1=256,
num_filters2=1024)
# conv5_x: 7x7
tmp = mid_projection(
name="res5_1", input=tmp, num_filters1=512, num_filters2=2048)
for i in xrange(2, res5_num + 1, 1):
tmp = bottleneck_block(
name="res5_" + str(i),
input=tmp,
num_filters1=512,
num_filters2=2048)
tmp = img_pool_layer(
name='avgpool',
input=tmp,
pool_size=7,
stride=1,
pool_type=AvgPooling())
return fc_layer(input=tmp, size=num_class, act=SoftmaxActivation())
if layer_num == 50:
resnet = deep_res_net(3, 4, 6, 3)
elif layer_num == 101:
resnet = deep_res_net(3, 4, 23, 3)
elif layer_num == 152:
resnet = deep_res_net(3, 8, 36, 3)
else:
print("Wrong layer number.")
lbl = data_layer(name="label", size=num_class)
loss = cross_entropy(name='loss', input=resnet, label=lbl)
inputs(img, lbl)
outputs(loss)
benchmark/paddle/image/run_mkldnn.sh
浏览文件 @ 469b3ad5
......@@ -5,22 +5,23 @@ function train() {
export OMP_DYNAMIC="FALSE"
export KMP_AFFINITY="granularity=fine,compact,0,0"
topology=$1
bs=$2
use_mkldnn=$3
if [ $3 == "True" ]; then
layer_num=$2
bs=$3
use_mkldnn=$4
if [ $4 == "True" ]; then
thread=1
log="logs/${topology}-mkldnn-${bs}.log"
elif [ $3 == "False" ]; then
log="logs/${topology}-${layer_num}-mkldnn-${bs}.log"
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}-${thread}mklml-${bs}.log"
log="logs/${topology}-${layer_num}-${thread}mklml-${bs}.log"
else
echo "Wrong input $3, use True or False."
exit 0
fi
args="batch_size=${bs}"
args="batch_size=${bs},layer_num=${layer_num}"
config="${topology}.py"
paddle train --job=time \
--config=$config \
......@@ -40,12 +41,9 @@ if [ ! -d "logs" ]; then
mkdir logs
fi
#========== mkldnn ==========#
train vgg 64 True
train vgg 128 True
train vgg 256 True
#========== mklml ===========#
train vgg 64 False
train vgg 128 False
train vgg 256 False
for use_mkldnn in True False; do
for batchsize in 64 128 256; do
train vgg 19 $batchsize $use_mkldnn
train resnet 50 $batchsize $use_mkldnn
done
done
benchmark/paddle/image/vgg.py
浏览文件 @ 469b3ad5
......@@ -13,7 +13,7 @@ define_py_data_sources2(
settings(
batch_size=batch_size,
learning_rate=0.01 / batch_size,
learning_rate=0.001 / batch_size,
learning_method=MomentumOptimizer(0.9),
regularization=L2Regularization(0.0005 * batch_size))
......
cmake/cblas.cmake
浏览文件 @ 469b3ad5
# Find the CBlas and lapack libraries
#
# It will search MKL, atlas, OpenBlas, reference-cblas in order.
# It will search MKLML, atlas, OpenBlas, reference-cblas in order.
#
# If any cblas implementation found, the following variable will be set.
# CBLAS_PROVIDER # one of MKL, ATLAS, OPENBLAS, REFERENCE
# CBLAS_PROVIDER # one of MKLML, ATLAS, OPENBLAS, REFERENCE
# CBLAS_INC_DIR # the include directory for cblas.
# CBLAS_LIBS # a list of libraries should be linked by paddle.
# # Each library should be full path to object file.
#
# User should set one of MKL_ROOT, ATLAS_ROOT, OPENBLAS_ROOT, REFERENCE_CBLAS_ROOT
# during cmake. If none of them set, it will try to find cblas implementation in
# system paths.
#
set(CBLAS_FOUND OFF)
......@@ -30,44 +25,6 @@ if(WITH_MKLML AND MKLML_INC_DIR AND MKLML_LIB)
return()
endif()
## Then find MKL.
set(INTEL_MKL_ROOT "/opt/intel/mkl" CACHE PATH "Folder contains intel mkl libs")
set(MKL_ROOT $ENV{MKL_ROOT} CACHE PATH "Folder contains env MKL")
set(MKL_INCLUDE_SEARCH_PATHS
${MKL_ROOT}/include
${INTEL_MKL_ROOT}/include)
set(MKL_LIB_SEARCH_PATHS
${MKL_ROOT}/lib
${MKL_ROOT}/lib/intel64
${INTEL_MKL_ROOT}/lib
${INTEL_MKL_ROOT}/lib/intel64)
find_path(MKL_INC_DIR mkl.h PATHS
${MKL_INCLUDE_SEARCH_PATHS})
find_path(MKL_LAPACK_INC_DIR mkl_lapacke.h PATHS
${MKL_INCLUDE_SEARCH_PATHS})
find_library(MKL_CORE_LIB NAMES mkl_core PATHS
${MKL_LIB_SEARCH_PATHS})
find_library(MKL_SEQUENTIAL_LIB NAMES mkl_sequential PATHS
${MKL_LIB_SEARCH_PATHS})
find_library(MKL_INTEL_LP64 NAMES mkl_intel_lp64 PATHS
${MKL_LIB_SEARCH_PATHS})
if(MKL_LAPACK_INC_DIR AND MKL_INC_DIR AND MKL_CORE_LIB AND MKL_SEQUENTIAL_LIB AND MKL_INTEL_LP64)
set(CBLAS_FOUND ON)
set(CBLAS_PROVIDER MKL)
set(CBLAS_INC_DIR ${MKL_INC_DIR} ${MKL_LAPACK_INC_DIR})
set(CBLAS_LIBRARIES ${MKL_INTEL_LP64} ${MKL_SEQUENTIAL_LIB} ${MKL_CORE_LIB})
add_definitions(-DPADDLE_USE_MKL)
add_definitions(-DLAPACK_FOUND)
message(STATUS "Found MKL (include: ${MKL_INC_DIR}, library: ${CBLAS_LIBRARIES})")
message(STATUS "Found lapack in MKL (include: ${MKL_LAPACK_INC_DIR})")
return()
endif()
## Then find atlas.
set(ATLAS_ROOT $ENV{ATLAS_ROOT} CACHE PATH "Folder contains Atlas")
set(ATLAS_INCLUDE_SEARCH_PATHS
......
cmake/cross_compiling/ios.cmake
浏览文件 @ 469b3ad5
......@@ -79,9 +79,8 @@ if(NOT DEFINED IOS_ARCH)
# FIXME(liuyiqun): support "armv7;armv7s;arm64" future
set(IOS_ARCH "arm64")
elseif(IOS_PLATFORM STREQUAL "SIMULATOR")
set(IOS_ARCH "i386;x86_64")
elseif(IOS_PLATFORM STREQUAL "WATCHOS")
set(IOS_ARCH armv7k)
# FIXME(liuyiqun): support "i386;x86_64" future
set(IOS_ARCH "x86_64")
endif()
endif()
set(CMAKE_OSX_ARCHITECTURES ${IOS_ARCH} CACHE string "Build architecture for iOS")
......
cmake/external/mkldnn.cmake
浏览文件 @ 469b3ad5
......@@ -46,16 +46,20 @@ IF(${CBLAS_PROVIDER} STREQUAL "MKLML")
MESSAGE(STATUS "Build MKLDNN with ${MKLDNN_MKLROOT}")
ENDIF()
SET(MKLDNN_CFLAG "${CMAKE_C_FLAGS} -Wno-error=strict-overflow")
SET(MKLDNN_CXXFLAG "${CMAKE_CXX_FLAGS} -Wno-error=strict-overflow")
ExternalProject_Add(
${MKLDNN_PROJECT}
${EXTERNAL_PROJECT_LOG_ARGS}
DEPENDS ${MKLDNN_DEPENDS}
GIT_REPOSITORY "https://github.com/01org/mkl-dnn.git"
GIT_TAG "v0.10"
GIT_TAG "v0.11"
PREFIX ${MKLDNN_SOURCES_DIR}
UPDATE_COMMAND ""
CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${MKLDNN_INSTALL_DIR}
CMAKE_ARGS -DMKLROOT=${MKLDNN_MKLROOT}
CMAKE_ARGS -DCMAKE_C_FLAGS=${MKLDNN_CFLAG}
CMAKE_ARGS -DCMAKE_CXX_FLAGS=${MKLDNN_CXXFLAG}
CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${MKLDNN_INSTALL_DIR}
-DMKLROOT:PATH=${MKLDNN_MKLROOT}
)
......
cmake/external/mklml.cmake
浏览文件 @ 469b3ad5
......@@ -27,8 +27,8 @@ ENDIF()
INCLUDE(ExternalProject)
SET(MKLML_PROJECT "extern_mklml")
SET(MKLML_VER "mklml_lnx_2018.0.20170720")
SET(MKLML_URL "https://github.com/01org/mkl-dnn/releases/download/v0.10/${MKLML_VER}.tgz")
SET(MKLML_VER "mklml_lnx_2018.0.1.20171007")
SET(MKLML_URL "https://github.com/01org/mkl-dnn/releases/download/v0.11/${MKLML_VER}.tgz")
SET(MKLML_SOURCE_DIR "${THIRD_PARTY_PATH}/mklml")
SET(MKLML_DOWNLOAD_DIR "${MKLML_SOURCE_DIR}/src/${MKLML_PROJECT}")
SET(MKLML_DST_DIR "mklml")
......
cmake/external/nccl.cmake
浏览文件 @ 469b3ad5
# 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.
if(NOT WITH_GPU)
return()
endif()
include(ExternalProject)
set(NCCL_SOURCE_DIR ${THIRD_PARTY_PATH}/nccl)
......
cmake/external/openblas.cmake
浏览文件 @ 469b3ad5
# 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.
......@@ -86,7 +86,7 @@ IF(NOT ${CBLAS_FOUND})
UPDATE_COMMAND ""
CONFIGURE_COMMAND ""
)
SET(CBLAS_PROVIDER openblas)
IF(WITH_C_API)
INSTALL(DIRECTORY ${CBLAS_INC_DIR} DESTINATION third_party/openblas)
# Because libopenblas.a is a symbolic link of another library, thus need to
......@@ -115,7 +115,7 @@ INCLUDE_DIRECTORIES(${CBLAS_INC_DIR})
# linear algebra libraries for cc_library(xxx SRCS xxx.c DEPS cblas)
SET(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/cblas_dummy.c)
FILE(WRITE ${dummyfile} "const char * dummy = \"${dummyfile}\";")
IF(${CBLAS_PROVIDER} MATCHES MKL)
IF("${CBLAS_PROVIDER}" STREQUAL "MKLML")
ADD_LIBRARY(cblas SHARED ${dummyfile})
ELSE()
ADD_LIBRARY(cblas STATIC ${dummyfile})
......
cmake/external/pybind11.cmake
浏览文件 @ 469b3ad5
INCLUDE(ExternalProject)
# 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.
SET(PYBIND_SOURCE_DIR ${THIRD_PARTY_PATH}/pybind)
if(NOT WITH_PYTHON)
return()
endif()
include(ExternalProject)
INCLUDE_DIRECTORIES(${PYBIND_SOURCE_DIR}/src/extern_pybind/include)
set(PYBIND_SOURCE_DIR ${THIRD_PARTY_PATH}/pybind)
include_directories(${PYBIND_SOURCE_DIR}/src/extern_pybind/include)
ExternalProject_Add(
extern_pybind
......@@ -17,14 +35,12 @@ ExternalProject_Add(
TEST_COMMAND ""
)
if (${CMAKE_VERSION} VERSION_LESS "3.3.0")
if(${CMAKE_VERSION} VERSION_LESS "3.3.0")
set(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/pybind_dummy.c)
file(WRITE ${dummyfile} "const char * dummy_any = \"${dummyfile}\";")
file(WRITE ${dummyfile} "const char * dummy_pybind = \"${dummyfile}\";")
add_library(pybind STATIC ${dummyfile})
else()
add_library(pybind INTERFACE)
endif()
add_dependencies(pybind extern_pybind)
LIST(APPEND external_project_dependencies pybind)
cmake/external/swig.cmake
浏览文件 @ 469b3ad5
# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
#
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#
# http://www.apache.org/licenses/LICENSE-2.0
#
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
......
cmake/external/zlib.cmake
浏览文件 @ 469b3ad5
# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
#
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#
# http://www.apache.org/licenses/LICENSE-2.0
#
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
......
cmake/generic.cmake
浏览文件 @ 469b3ad5
......@@ -93,7 +93,7 @@ include_directories(${CMAKE_CURRENT_BINARY_DIR})
if(NOT APPLE AND NOT ANDROID)
find_package(Threads REQUIRED)
link_libraries(${CMAKE_THREAD_LIBS_INIT})
set(CMAKE_CXX_LINK_EXECUTABLE "${CMAKE_CXX_LINK_EXECUTABLE} -ldl -lrt")
set(CMAKE_CXX_LINK_EXECUTABLE "${CMAKE_CXX_LINK_EXECUTABLE} -pthread -ldl -lrt")
endif(NOT APPLE AND NOT ANDROID)
function(merge_static_libs TARGET_NAME)
......
cmake/simd.cmake
浏览文件 @ 469b3ad5
# This file is use to check all support level of AVX on your machine
# so that PaddlePaddle can unleash the vectorization power of muticore.
INCLUDE(CheckCXXSourceRuns)
INCLUDE(CheckCXXSourceCompiles)
include(CheckCXXSourceRuns)
include(CheckCXXSourceCompiles)
IF(CMAKE_COMPILER_IS_GNUCC OR CMAKE_COMPILER_IS_GNUCXX OR CMAKE_CXX_COMPILER_ID MATCHES "Clang")
if(CMAKE_COMPILER_IS_GNUCC OR CMAKE_COMPILER_IS_GNUCXX OR CMAKE_CXX_COMPILER_ID MATCHES "Clang")
set(MMX_FLAG "-mmmx")
set(SSE2_FLAG "-msse2")
set(SSE3_FLAG "-msse3")
SET(AVX_FLAG "-mavx")
SET(AVX2_FLAG "-mavx2")
ELSEIF(MSVC)
set(AVX_FLAG "-mavx")
set(AVX2_FLAG "-mavx2")
elseif(MSVC)
set(MMX_FLAG "/arch:MMX")
set(SSE2_FLAG "/arch:SSE2")
set(SSE3_FLAG "/arch:SSE3")
SET(AVX_FLAG "/arch:AVX")
SET(AVX2_FLAG "/arch:AVX2")
ENDIF()
endif()
set(CMAKE_REQUIRED_FLAGS_RETAINED ${CMAKE_REQUIRED_FLAGS})
# Check MMX
set(CMAKE_REQUIRED_FLAGS ${MMX_FLAG})
set(MMX_FOUND_EXITCODE 1 CACHE STRING "Result from TRY_RUN" FORCE)
CHECK_CXX_SOURCE_RUNS("
#include <mmintrin.h>
int main()
......@@ -32,6 +33,7 @@ int main()
# Check SSE2
set(CMAKE_REQUIRED_FLAGS ${SSE2_FLAG})
set(SSE2_FOUND_EXITCODE 1 CACHE STRING "Result from TRY_RUN" FORCE)
CHECK_CXX_SOURCE_RUNS("
#include <emmintrin.h>
int main()
......@@ -42,6 +44,7 @@ int main()
# Check SSE3
set(CMAKE_REQUIRED_FLAGS ${SSE3_FLAG})
set(SSE3_FOUND_EXITCODE 1 CACHE STRING "Result from TRY_RUN" FORCE)
CHECK_CXX_SOURCE_RUNS("
#include <pmmintrin.h>
int main()
......@@ -55,6 +58,7 @@ int main()
# Check AVX
set(CMAKE_REQUIRED_FLAGS ${AVX_FLAG})
set(AVX_FOUND_EXITCODE 1 CACHE STRING "Result from TRY_RUN" FORCE)
CHECK_CXX_SOURCE_RUNS("
#include <immintrin.h>
int main()
......@@ -67,6 +71,7 @@ int main()
# Check AVX 2
set(CMAKE_REQUIRED_FLAGS ${AVX2_FLAG})
set(AVX2_FOUND_EXITCODE 1 CACHE STRING "Result from TRY_RUN" FORCE)
CHECK_CXX_SOURCE_RUNS("
#include <immintrin.h>
int main()
......
doc/api/v2/config/layer.rst
浏览文件 @ 469b3ad5
......@@ -82,6 +82,11 @@ maxout
.. autoclass:: paddle.v2.layer.maxout
:noindex:
roi_pool
--------
.. autoclass:: paddle.v2.layer.roi_pool
:noindex:
Norm Layer
==========
......
doc/api/v2/data.rst
浏览文件 @ 469b3ad5
......@@ -2,112 +2,9 @@
Data Reader Interface and DataSets
==================================
.. toctree::
:maxdepth: 1
DataTypes
=========
.. automodule:: paddle.v2.data_type
:members:
:noindex:
DataFeeder
==========
.. automodule:: paddle.v2.data_feeder
:members:
:noindex:
Reader
======
.. automodule:: paddle.v2.reader
:members:
:noindex:
.. automodule:: paddle.v2.reader.creator
:members:
:noindex:
minibatch
=========
.. automodule:: paddle.v2.minibatch
:members:
:noindex:
Dataset
=======
.. automodule:: paddle.v2.dataset
:members:
:noindex:
mnist
+++++
.. automodule:: paddle.v2.dataset.mnist
:members:
:noindex:
cifar
+++++
.. automodule:: paddle.v2.dataset.cifar
:members:
:noindex:
conll05
+++++++
.. automodule:: paddle.v2.dataset.conll05
:members: get_dict,get_embedding,test
:noindex:
imdb
++++
.. automodule:: paddle.v2.dataset.imdb
:members:
:noindex:
imikolov
++++++++
.. automodule:: paddle.v2.dataset.imikolov
:members:
:noindex:
movielens
+++++++++
.. automodule:: paddle.v2.dataset.movielens
:members:
:noindex:
.. autoclass:: paddle.v2.dataset.movielens.MovieInfo
:noindex:
.. autoclass:: paddle.v2.dataset.movielens.UserInfo
:noindex:
sentiment
+++++++++
.. automodule:: paddle.v2.dataset.sentiment
:members:
:noindex:
uci_housing
+++++++++++
.. automodule:: paddle.v2.dataset.uci_housing
:members:
:noindex:
wmt14
+++++
.. automodule:: paddle.v2.dataset.wmt14
:members:
:noindex:
data/data_reader.rst
data/image.rst
data/dataset.rst
doc/api/v2/data/data_reader.rst 0 → 100644
浏览文件 @ 469b3ad5
=====================
Data Reader Interface
=====================
DataTypes
=========
.. automodule:: paddle.v2.data_type
:members:
:noindex:
DataFeeder
==========
.. automodule:: paddle.v2.data_feeder
:members:
:noindex:
Reader
======
.. automodule:: paddle.v2.reader
:members:
:noindex:
.. automodule:: paddle.v2.reader.creator
:members:
:noindex:
minibatch
=========
.. automodule:: paddle.v2.minibatch
:members:
:noindex:
doc/api/v2/data/dataset.rst 0 → 100644
浏览文件 @ 469b3ad5
Dataset
=======
.. automodule:: paddle.v2.dataset
:members:
:noindex:
mnist
+++++
.. automodule:: paddle.v2.dataset.mnist
:members:
:noindex:
cifar
+++++
.. automodule:: paddle.v2.dataset.cifar
:members:
:noindex:
conll05
+++++++
.. automodule:: paddle.v2.dataset.conll05
:members: get_dict,get_embedding,test
:noindex:
imdb
++++
.. automodule:: paddle.v2.dataset.imdb
:members:
:noindex:
imikolov
++++++++
.. automodule:: paddle.v2.dataset.imikolov
:members:
:noindex:
movielens
+++++++++
.. automodule:: paddle.v2.dataset.movielens
:members:
:noindex:
.. autoclass:: paddle.v2.dataset.movielens.MovieInfo
:noindex:
.. autoclass:: paddle.v2.dataset.movielens.UserInfo
:noindex:
sentiment
+++++++++
.. automodule:: paddle.v2.dataset.sentiment
:members:
:noindex:
uci_housing
+++++++++++
.. automodule:: paddle.v2.dataset.uci_housing
:members:
:noindex:
wmt14
+++++
.. automodule:: paddle.v2.dataset.wmt14
:members:
:noindex:
doc/api/v2/data/image.rst 0 → 100644
浏览文件 @ 469b3ad5
Image Interface
===============
.. automodule:: paddle.v2.image
:members:
doc/design/float16.md 0 → 100644
浏览文件 @ 469b3ad5
# Design Doc: float16
## Why float16
Half precision (float16) is a binary floating-point format that occupies 16 bits in memory. float16 is half the size of traditional 32-bit single precision format (float) and has lower precision and smaller range.
When high precision computation is not required, using float16 data type could potentially
- reduce storage space, memory bandwidth, and power usages;
- increase the chance of data fitting into a smaller cache of lower latency;
- provide arithmetic speed up if supported by hardware.
## Survey of current float16 support
A brief survey of float16 support on different compilers, hardwares, and libraries can be found below. Interested readers can refer to [link1](https://github.com/PaddlePaddle/Paddle/issues/4853) and [link2](https://github.com/Xreki/Xreki.github.io/blob/master/multi_data_types_in_dl_framework/ppt/float16_and_quantized_type.md) for more info.
The goal of float16 is to serve as a key for the executor to find and run the correct version of compute method specialized for float16 in operator kernel. It should be compatible with various natively supported float16 implementations including `__half` for cuda, `float16_t` for ARM, and `Eigen::half` for Eigen to make writing customized float16 kernels easier.
### Compiler
- nvcc supports `__half` data type after CUDA 7.5.
- `__fp16` or `float16_t` is supported as storage type for gcc >= 6.1 and clang >= 3.4.
- `__fp16` or `float16_t` is supported as arithmetic type for gcc >= 7.1 and clang >= 3.9.
### Hardware
- `__half` is supported on GPU with compute capability >= 5.3.
- `__fp16` is supported as storage type for ARMv7-A, ARMv8-A, and above.
- `__fp16` is supported as arithmetic type after ARMv8.2-A (currently, the only microarchitecture implementing ARMv8.2-A is ARM Cortex-A75, which is announced in May 2017. There seems to be no application processors currently available on market that adopts this architecture. It is reported that Qualcomm Snapdragon 845 uses Cortex-A75 design and will be available in mobile devices in early 2018).
### Libraries
- [Eigen](https://github.com/RLovelett/eigen) >= 3.3 supports float16 calculation on both GPU and CPU using the `Eigen::half` class. It is mostly useful for Nvidia GPUs because of the overloaded arithmetic operators using cuda intrinsics. It falls back to using software emulation on CPU for calculation and there is no special treatment to ARM processors.
- [ARM compute library](https://github.com/ARM-software/ComputeLibrary) >= 17.02.01 supports NEON FP16 kernels (requires ARMv8.2-A CPU).
## Implementation
The float16 class holds a 16-bit `uint16_t` data internally.
```
struct float16 {
uint16_t x;
};
```
float16 supports the following features:
- constructors / assignment operators that take input from primitive data types including bool, integers of various length, float, and double.
- constructors / assignment operators that take input from `__half` on cuda, `float16_t` on ARM, and `Eigen::half` on Eigen.
- conversion operators to primitive data types and half precision data types on cuda, ARM and Eigen.
- overloaded arithmetic operators for cuda, arm, and non-arm cpu, respectively. These operators will take advantage of the cuda and ARM intrinsics on the corresponding hardware.
To support the above features, two fundamental conversion functions are provided:
```
float16 float_to_half_rn(float f); // convert to half precision in round-to-nearest-even mode
float half_to_float(float16 h);
```
which provides one-to-one conversion between float32 and float16. These twos functions will do different conversion routines based on the current hardware. CUDA/ARM instrinsics will be used when the corresonding hardware is available. If the hardware or compiler level does not support float32 to float16 conversion, software emulation will be performed to do the conversion.
## To do
After float16 class is available, some of the future items are below:
- Update pybind/tensor_py.h to bind c++ float16 with numpy float16.
- Modify `GetKernelType()` method in `framework/operator.h` to make it compatible with float16.
- Create a type-casting operator that can convert the data type in tensor between float16 and other types.
doc/design/mkldnn/README.MD
浏览文件 @ 469b3ad5
......@@ -15,6 +15,7 @@
- [CMake](#cmake)
- [Layers](#layers)
- [Activations](#activations)
- [Weights](#weights)
- [Unit Tests](#unit-tests)
- [Protobuf Messages](#protobuf-messages)
- [Python API](#python-api)
......@@ -45,17 +46,23 @@ Figure 1. PaddlePaddle on IA.
### Layers
所有MKL-DNN相关的C++ layers，都会按照PaddlePaddle的目录结构存放在
`paddle/gserver/layers`中，并且文件名都会一以*Mkldnn*开头。
`paddle/gserver/layers`中，并且文件名都会一以*MKLDNN*开头。
所有MKL-DNN的layers都会继承于一个叫做`MkldnnLayer`的父类，该父类继承于PaddlePaddle的基类`Layer`。
所有MKL-DNN的layers都会继承于一个叫做`MKLDNNLayer`的父类，该父类继承于PaddlePaddle的基类`Layer`。
在`MKLDNNLayer`中会提供一些必要的接口和函数，并且会写好`forward`和`backward`的基本逻辑。部分函数定义为纯虚函数，子类只需要实现这些函数即可。
### Activations
由于在PaddlePaddle中，激活函数是独立于layer概念的，所以会在`paddle/gserver/activations`目录下添加一个`MkldnnActivation.h`文件定义一些用于MKL-DNN的接口，实现方法还是会在`ActivationFunction.cpp`文件。
由于在PaddlePaddle中，激活函数是独立于layer概念的，所以会在`paddle/gserver/activations`目录下添加`MKLDNNActivation.h`和`MKLDNNActivation.cpp`文件用于定义和使用MKL-DNN的接口。
### Unit Tests
会在`paddle/gserver/test`目录下添加`test_Mkldnn.cpp`和`MkldnnTester.*`用于MKL-DNN的测试。
### Weights
由于有些layer是含有参数的，我们会尽量让MKL-DNN的参数与PaddlePaddle中`parameter`共享一块内存。
同时，由于MKL-DNN在训练时使用的参数layout可能与PaddlePaddle默认的`nchw`不一致，我们会在网络训练的开始和结束时分别转换这个layout，使得最终保存的参数格式与PaddlePaddle一致。
Activation的测试，计划在PaddlePaddle原有的测试文件上直接添加新的测试type。
### Unit Tests
会在`paddle/gserver/test`目录下添加`test_MKLDNN.cpp`和`MKLDNNTester.*`用于MKL-DNN的测试。
测试分为每个layer(或activation)的单元测试和简单网络的整体测试。
每个测试会对比PaddlePaddle中CPU算出的结果与MKL-DNN的结果，小于某个比较小的阈值认为通过。
### Protobuf Messages
根据具体layer的需求可能会在`proto/ModelConfig.proto`里面添加必要的选项。
......@@ -82,7 +89,7 @@ if use_mkldnn
会在`v1_api_demo`目录下添加一个`mkldnn`的文件夹，里面放入一些用于MKL-DNN测试的demo脚本。
### Benchmarking
会考虑添加部分逻辑在`benchmark/paddle/image/run.sh`，添加使用MKL-DNN的测试。
会添加`benchmark/paddle/image/run_mkldnn.sh`，用于测试使用MKL-DNN之后的性能。
### Others
1. 如果在使用MKL-DNN的情况下，会把CPU的Buffer对齐为64。
......@@ -94,14 +101,16 @@ if use_mkldnn
我们总结出一些特别需要注意的点：
1. 使用**deviceId_**。为了尽可能少的在父类Layer中添加变量或者函数，我们决定使用已有的`deviceId_`变量来区分layer的属性，定义`-2`为`MkldnnLayer`特有的设备ID。
1. 使用**deviceId_**。为了尽可能少的在父类Layer中添加变量或者函数，我们决定使用已有的`deviceId_`变量来区分layer的属性，定义`-2`为`MKLDNNLayer`特有的设备ID。
2. 重写父类Layer的**init**函数，修改`deviceId_`为`-2`，代表这个layer是用于跑在MKL-DNN的环境下。
3. 创建`MkldnnMatrix`，用于管理MKL-DNN会用到的相关memory函数、接口以及会用的到格式信息。
4. 创建`MkldnnBase`，定义一些除了layer和memory相关的类和函数。包括MKL-DNN会用到`MkldnnStream`和`CpuEngine`，和未来可能还会用到`FPGAEngine`等。
5. 在**Argument**里添加两个`MkldnnMatrixPtr`，取名为`mkldnnValue`和`mkldnnGrad`，用于存放`MkldnnLayer`会用到的memory buffer。 并且添加函数cvt(会修改为一个更加合适的函数名)，用于处理"CPU device"和"MKL-DNN device"之间memory的相互转化。
6. 在父类`Layer`中的`getOutput`函数中添加一段逻辑，用于判断`deviceId`，并针对device在MKL-DNN和CPU之间不统一的情况，做一个前期转换。 也就是调用`Argument`的cvt函数把output统一到需要的device上。
7. 在原来的`FLAGS`中添加一个`use_mkldnn`的flag，用于选择是否使用MKL-DNN的相关功能。
8. 关于MKLDNN参数的保存。由于MKLDNN参数的格式与PaddlePaddle原有的格式存在不一样的情况，所以需要在保存参数时同时保存该格式信息。目前准备扩展[Header](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/parameter/Parameter.h#L247)里面的`int32_t version`。这个值不管是在v1还是在v2里面，一直保存的是0，所以可以充分利用这个信息，定义一个枚举处理所有MKLDNN的参数格式，从而`MKLDNNLayer`就可以从输入的参数中获取需要的格式信息。
3. 创建`MKLDNNMatrix`，同时继承`CpuMatrix`和`mkldnn::memory`。用于管理MKL-DNN会用到的相关memory函数、接口以及会用的到格式信息。
4. 创建`MKLDNNBase`，定义一些除了layer和memory相关的类和函数。包括MKL-DNN会用到`MKLDNNStream`和`CPUEngine`，和未来可能还会用到`FPGAEngine`等。
5. 每个`MKLDNNlayer`都会有`inVal_`,`inGrad_`,`outVal_`和`outGrad_`，分别代表input value， input gradient，output value和output gradient。他们会存放MKL-DNN用到的internal memory。同时还会定义以*ext*开头的`MKLDNNMatrix`(表示external的memory)，主要是在格式与PaddlePaddle默认的`nchw`格式不匹配时，用于转换内存的工作。必要的转换函数也会在`MKLDNNLayer`中提前定义好，每个子类只需要调用定义好的reset buffer函数即可。
6. 每个`MKLDNNlayer`的resetbuffer相关的函数（包括reset input、output的Value和grad），他们会根据输入参数reset internal和external的memory，当然这两者也可以相等，即表示不需要转换。只需要把握一个原则，每个`MKLDNNlayer`的子类，只需要使用internal的memory就可以了，所有external的转换工作在父类的reset函数中都提前准备好了。
7. 一般来说，external的memory会尽量与PaddlePaddle中的`value`和`grad`共享内存。同时每个`MKLDNNLayer`中的external output value和gradient(也就是`extOutVal_`和`extOutGrad_`)必须分别与`output_.value`和`output_.grad`共享内存，因为PaddlePaddle的activation会直接使用`output_.value`和`output_.grad`。如果不需要external的buffer用于转换，那么internal的buffer也会与他们共享内存。
8. 如果MKL-DNN layer的后面接有cpu device，那么就会使`output_.value`与`extOutVal_`共享内存，同时数据格式就是`nchw`，这样下一个cpu device就能拿到正确的数据。在有cpu device的时候，external的memory的格式始终是`nchw`或者`nc`。
9. 由于MKL-DNN的输出操作都是覆盖data的，不是在原来的数据上累加，所以当网络出现分支时，在`backward`时会需要merge不同layer的梯度。`MKLDNNlayer`中会实现merge的方法，此时每个小分支的input gradient会先临时保存在一个`MKLDNNMatrix`中，由分支处的layer负责求和，并把结果放到这个layer的`output_.grad`中。所以整体上，每个子类并不会需要关心分支的事情，也是在父类都实现好了。
10. 在原来的`FLAGS`中添加一个`use_mkldnn`的flag，用于选择是否使用MKL-DNN的相关功能。
## References
......
doc/design/ops/sequence_decoder.md 0 → 100644
浏览文件 @ 469b3ad5
# 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.
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.
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.
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** .
For example, the RNN sates, candidates IDs and probabilities of beam search can be represented 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 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 relative-offset LoD can fast retrieve any sequence 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]]
```
The first level tells that there are two sequences:
- the first's offset is `[0, 3)`
- the second's offset is `[3, 9)`
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.
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
```python
[[0, 3, 6]
[0, 2, 3, 3, 3, 9]]
```
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.
## 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.
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.
**Encoder**
```python
import paddle as pd
dict_size = 8000
source_dict_size = dict_size
target_dict_size = dict_size
word_vector_dim = 128
encoder_dim = 128
decoder_dim = 128
beam_size = 5
max_length = 120
# encoder
src_word_id = pd.data(
name='source_language_word',
type=pd.data.integer_value_sequence(source_dict_dim))
src_embedding = pd.embedding(size=source_dict_size, size=word_vector_dim)
src_word_vec = pd.lookup(src_embedding, src_word_id)
encoder_out_seq = pd.gru(input=src_word_vec, size=encoder_dim)
encoder_ctx = pd.last_seq(encoder_out_seq)
# encoder_ctx_proj is the learned semantic vector
encoder_ctx_proj = pd.fc(
encoder_ctx, size=decoder_dim, act=pd.activation.Tanh(), bias=None)
```
**Decoder**
```python
def generate():
decoder = pd.while_loop()
with decoder.step():
decoder_mem = decoder.memory(init=encoder_ctx) # mark the memory
generated_ids = decoder.memory() # TODO init to batch_size <s>s
generated_scores = decoder.memory() # TODO init to batch_size 1s or 0s
target_word = pd.lookup(trg_embedding, gendrated_ids)
# expand encoder_ctx's batch to fit target_word's lod
# for example
# decoder_mem.lod is
# [[0 1 3],
# [0 1 3 6]]
# its tensor content is [a1 a2 a3 a4 a5]
# 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
# [[0, 1, 6]
# [0, 2, 4, 7, 9 12]]
# which means 2 sentences to translate, each has 1 and 5 prefixes
# the first prefix has 2 candidates
# the following has 2, 3, 2, 3 candidates
# the encoder_ctx_expanded's content will be
# [a1 a1 a2 a2 a3 a3 a3 a4 a4 a5 a5 a5]
encoder_ctx_expanded = pd.lod_expand(encoder_ctx, target_word)
decoder_input = pd.fc(
act=pd.activation.Linear(),
input=[target_word, encoder_ctx],
size=3 * decoder_dim)
gru_out, cur_mem = pd.gru_step(
decoder_input, mem=decoder_mem, size=decoder_dim)
scores = pd.fc(
gru_out,
size=trg_dic_size,
bias=None,
act=pd.activation.Softmax())
# K is an config
topk_scores, topk_ids = pd.top_k(scores, K)
topk_generated_scores = pd.add_scalar(topk_scores, generated_scores)
selected_ids, selected_generation_scores = decoder.beam_search(
topk_ids, topk_generated_scores)
# update the states
decoder_mem.update(cur_mem) # tells how to update state
generated_ids.update(selected_ids)
generated_scores.update(selected_generation_scores)
decoder.output(selected_ids)
decoder.output(selected_generation_scores)
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.
In this way, users can customize anything on the inputs or outputs of beam search, for example, two ways to prune some translation prefixes
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.
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).
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.
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.
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`
* LoD is `[0, 2, 7][0 3 5 8 9 11 11]`
* 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
- b2
- 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 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.
Pack the `selected_scores` will get a `LoDTensor` that stores scores of each candidate of translations.
Pack the `selected_generation_scores` will get a `LoDTensor`, and each tail is the probability of the translation.
## LoD and shape changes during decoding
<p align="center">
<img src="./images/LOD-and-shape-changes-during-decoding.jpg"/>
</p>
According the image above, the only phrase to change 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
1. `topk_ids`, 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.
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).
## 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,
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`.
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`.
doc/design/parameter_average.md 0 → 100644
浏览文件 @ 469b3ad5
# Averaging Parameter in PaddlePaddle
## Why Averaging
In a large scale machine learning setup where the size of the training data is huge, it could take us a large number of iterations over the training data before we can achieve the optimal values of parameters of our model. Looking at the problem setup, it is desirable if we can obtain the optimal values of parameters by going through the data in as few passes as we can.
Polyak and Juditsky (1992) showed that the test performance of simple average of parameters obtained by Stochastic Gradient Descent (SGD) is as good as that of parameter values that are obtained by training the model over and over again, over the training dataset.
Hence, to accelerate the speed of Stochastic Gradient Descent, Averaged Stochastic Gradient Descent (ASGD) was proposed in Polyak and Juditsky (1992). For ASGD, the running average of parameters obtained by SGD, is used as the estimator for <img src="./images/theta_star.gif"/><br/> . The averaging is done as follows:
<img src="./images/asgd.gif" align="center"/><br/>
We propose averaging for any optimizer similar to how ASGD performs it, as mentioned above.
### How to perform Parameter Averaging in PaddlePaddle
Parameter Averaging in PaddlePaddle works in the following way during training :
1. It will take in an instance of a normal optimizer as an input, e.g. RMSPropOptimizer
2. The optimizer itself is responsible for updating the parameters.
3. The ParameterAverageOptimizer maintains a separate copy of the parameters for itself:
1. In concept, the values of this copy are the average of the values of the parameters in the most recent N batches.
2. However, saving all the N instances of the parameters in memory is not feasible.
3. Therefore, an approximation algorithm is used.
Hence, overall we have have two copies of the parameters: one for the optimizer itself, and one for the ParameterAverageOptimizer. The former should be used in back propagation, while the latter should be used during testing and should be saved.
During the testing/ saving the model phase, we perform the following steps:
1. Perform the delayed operations.
2. Save current values of the parameters to a temporary variable.
3. Replace the values of the parameters with the averaged values.
4. Perform testing and/or save the parameters.
5. Restore the values of the parameters once done.
### How to implement Averaging of Parameter in PaddlePaddle
We can add the ParameterAverageOptimizer op to the graph through Python API. Using this approach, we manually add this op to the graph and direct the output of the optimizer op to this op during training.
**Advantages**:
- Allows for greater flexibility to the users of PaddlePaddle. Using this approach, the users can plug different optimizers into ParameterAverageOptimizer by passing in the optimizer to the op.
- Makes it easy for the users to customize and extend the framework.
**Disadvantages**:
- Implementation requires re-writing the averaging methodology in Python.
### Low-Level implementation
In the new design, we propose to create a new operation for averaging parameter updates (ParameterAverageOptimizer). For now, we can add an op that takes in the following as input:
- the optimizer
- the window_size to keep the updates
The ParameterAverageOptimizer op can be like any other operator with its own CPU/GPU implementation either using Eigen or separate CPU and GPU kernels. As the initial implementation, we can implement the kernel using Eigen following the abstraction pattern implemented for [Operators](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/rmsprop_op.h). We also want to support the case when the Trainer/Optimizer runs on the GPU while ParameterAverageOptimizer runs on a CPU.
The idea of building an op for averaging is in sync with the refactored PaddlePaddle philosophy of using operators to represent any computation unit. The way the op will be added to the computation graph will be decided by the [layer functions](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/python_api.md#layer-function) in Python API.
### Python API implementation for ParameterAverageOptimizer
Based on Polyak and Juditsky (1992), we can generalize the averaging of updates to any optimizer. The input to the op would be the following:
- Any optimizer (RMSProp , AdaGrad etc.)
- A window size. The op keeps accumulating updated parameter values over a window of N batches and takes an average. Move the averaged value to a buffer when window is full to avoid loss of precision.
Using the ParameterAverageOptimizer op, any user can add the operation to their computation graphs. However, this will require a lot of lines of code and we should design Python APIs that support averaging. As per the PaddlePaddle [Python API design](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/python_api.md), the layer functions are responsible for creating operators, operator parameters and variables. Since ParameterAverageOptimizer will be an operator, it makes sense to create it in the layer functions.
We will have a wrapper written in Python that will support the functionality and implement the actual core computation in C++ core as we have done for other [Optimizers](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/rmsprop_op.cc)
#### Creation of the ParameterAverageOptimizer operator
There are two ways for creating the ParameterAverageOptimizer op:
1. We create the op immediately while building the computation graph.
2. We add the op in a lazy manner, just before the backward pass, similar to the way the optimization ops are added.
The proposal is to add the op immediately while building the computation graph.
#### High-level API
In PaddlePaddle Python API, users will primarily rely on [layer functions](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/python_api.md#layer-function) to create neural network layers. Hence, we also need to provide parameter average functionality in layer functions.
doc/faq/local/index_cn.rst
浏览文件 @ 469b3ad5
......@@ -99,7 +99,7 @@ PaddlePaddle支持Sparse的训练，sparse训练需要训练特征是 :code:`spa
利用更多的计算资源
++++++++++++++++++
利用更多的计算资源可以分为一下几个方式来进行\:
利用更多的计算资源可以分为以下几个方式来进行\:
* 单机CPU训练
......
doc/faq/parameter/index_cn.rst
浏览文件 @ 469b3ad5
......@@ -75,7 +75,7 @@ PaddlePaddle目前支持8种learning_rate_schedule，这8种learning_rate_schedu
optimizer = paddle.optimizer.Adam(
learning_rate=1e-3,
learning_rate_schedule="manual",
learning_rate_schedule="pass_manual",
learning_rate_args="1:1.0,2:0.9,3:0.8",)
在该示例中，当已训练pass数小于等于1时，学习率为 :code:`1e-3 * 1.0`；当已训练pass数大于1小于等于2时，学习率为 :code:`1e-3 * 0.9`；当已训练pass数大于2时，学习率为 :code:`1e-3 * 0.8`。
......
doc/getstarted/build_and_install/docker_install_cn.rst
浏览文件 @ 469b3ad5
......@@ -145,7 +145,7 @@ PaddlePaddle发布新版本的时候都会发布对应版本的生产镜像以
Jupyter Notebook是一个开源的web程序，大家可以通过它制作和分享带有代码、公式、图表、文字的交互式文档。用户可以通过网页浏览文档。
PaddlePaddle Book是为用户和开发者制作的一个交互式的Jupyter Nodebook。
PaddlePaddle Book是为用户和开发者制作的一个交互式的Jupyter Notebook。
如果您想要更深入了解deep learning，PaddlePaddle Book一定是您最好的选择。
我们提供可以直接运行PaddlePaddle Book的Docker镜像，直接运行：
......
doc/howto/dev/new_op_cn.md
浏览文件 @ 469b3ad5
......@@ -214,7 +214,7 @@ MulOp(const std::string &type, const framework::VariableNameMap &inputs,
```cpp
// if use Eigen unsupported module before include head files
#define EIGEN_USE_GPU
// #define EIGEN_USE_GPU
namespace ops = paddle::operators;
REGISTER_OP_GPU_KERNEL(mul, ops::MulKernel<paddle::platform::GPUPlace, float>);
......
doc/howto/usage/cmd_parameter/arguments_cn.md
浏览文件 @ 469b3ad5
......@@ -63,7 +63,7 @@
</tr>
<tr>
<td class="left" rowspan="15">训练</td><td class="left">dot_period</td>
<td class="left" rowspan="14">训练</td><td class="left">dot_period</td>
<td class="left">√</td><td class="left">√</td><td class="left"></td><td class="left"></td>
</tr>
......
doc/index_cn.rst
浏览文件 @ 469b3ad5
......@@ -8,3 +8,4 @@ PaddlePaddle 文档
howto/index_cn.rst
api/index_cn.rst
faq/index_cn.rst
mobile/index_cn.rst
doc/index_en.rst
浏览文件 @ 469b3ad5
......@@ -7,3 +7,4 @@ PaddlePaddle Documentation
getstarted/index_en.rst
howto/index_en.rst
api/index_en.rst
mobile/index_en.rst
doc/howto/cross_compiling/cross_compiling_for_android_cn.md → doc/mobile/cross_compiling_for_android_cn.md
浏览文件 @ 469b3ad5
# 构建Android平台上的PaddlePaddle库
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
- 基于Docker容器的编译方式
- 基于Linux交叉编译环境的编译方式
## 基于Docker容器的编译方式
......@@ -20,20 +20,42 @@ $ docker build -t username/paddle-android:dev . -f Dockerfile.android
构建好开发镜像后，即可使用开发镜像来编译Android版PaddlePaddle C-API库。
Android的Docker开发镜像向用户提供两个可配置的参数：
| Argument | Optional Values | Default |
|-----------------|-------------------------|---------|
|`ANDROID_ABI` |`armeabi-v7a, arm64-v8a` | `armeabi-v7a` |
|`ANDROID_API` |`>= 21` | `21` |
<table class="docutils">
<colgroup>
<col width="25%" />
<col width="50%" />
<col width="25%" />
</colgroup>
<thead valign="bottom">
<tr class="row-odd">
<th class="head">Argument</th>
<th class="head">Optional Values</th>
<th class="head">Default</th>
</tr>
</thead>
<tbody valign="top">
<tr class="row-even">
<td>ANDROID_ABI</td>
<td>armeabi-v7a, arm64-v8a</td>
<td>armeabi-v7a</td>
</tr>
<tr class="row-odd">
<td>ANDROID_API</td>
<td>>= 21</td>
<td>21</td>
</tr>
</tbody>
</table>
- 编译`armeabi-v7a`，`Android API 21`的PaddlePaddle库
```bash
$ docker run -it --rm -v $PWD:/paddle -e "ANDROID_ABI=armeabi-v7a" -e "ANDROID_API=21" username/paddle-android:dev
```
```bash
$ docker run -it --rm -v $PWD:/paddle -e "ANDROID_ABI=armeabi-v7a" -e "ANDROID_API=21" username/paddle-android:dev
```
- 编译`arm64-v8a`，`Android API 21`的PaddlePaddle库
```bash
$ docker run -it --rm -v $PWD:/paddle -e "ANDROID_ABI=arm64-v8a" -e "ANDROID_API=21" username/paddle-android:dev
```
- 编译`arm64-v8a`，`Android API 21`的PaddlePaddle库
```bash
$ docker run -it --rm -v $PWD:/paddle -e "ANDROID_ABI=arm64-v8a" -e "ANDROID_API=21" username/paddle-android:dev
```
执行上述`docker run`命令时，容器默认执行[paddle/scripts/docker/build_android.sh](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/scripts/docker/build_android.sh)脚本。该脚本中记录了交叉编译Android版PaddlePaddle库常用的CMake配置，并且会根据`ANDROID_ABI`和`ANDROID_API`自动构建独立工具链、进行编译和安装。由于arm64架构要求Android API不小于21。因此当`ANDROID_ABI=arm64-v8a`，`ANDROID_API<21`时，Docker容器中将默认使用`Android API 21`的编译工具链。用户可以参考下文**配置交叉编译参数**章节，根据个人的需求修改定制Docker容器所执行的脚本。编译安装结束之后，PaddlePaddle的C-API库将被安装到`$PWD/install_android`目录，所依赖的第三方库同时也被安装到`$PWD/install_android/third_party`目录。
......@@ -82,16 +104,16 @@ CMake系统对交叉编译提供了支持[cmake-toolchains](https://cmake.org/cm
Android平台可选配置参数：
- `ANDROID_STANDALONE_TOOLCHAIN`，独立工具链所在的绝对路径，或者相对于构建目录的相对路径。PaddlePaddle的CMake系统将根据该值自动推导和设置需要使用的交叉编译器、sysroot、以及Android API级别；否则，用户需要在cmake时手动设置这些值。无默认值。
- `ANDROID_TOOLCHAIN`，目标工具链。可设置`gcc/clang`，默认值为`clang`。
- CMake 3.7以上，将会始终使用`clang`工具链；CMake 3.7以下，可设置`ANDROID_TOOLCHAIN=gcc`以使用`gcc`工具链。
- `ANDROID_TOOLCHAIN`，目标工具链。可设置`gcc/clang`，默认值为`clang`。
- CMake 3.7以上，将会始终使用`clang`工具链；CMake 3.7以下，可设置`ANDROID_TOOLCHAIN=gcc`以使用`gcc`工具链。
- Android官方提供的`clang`编译器要求系统支持`GLIBC 2.15`以上。
- `ANDROID_ABI`，目标架构ABI。目前支持`armeabi-v7a`和`arm64-v8a`，默认值为`armeabi-v7a`。
- `ANDROID_NATIVE_API_LEVEL`，工具链的Android API级别。若没有显式设置，PaddlePaddle将根据`ANDROID_STANDALONE_TOOLCHAIN`的值自动推导得到。
- `ANROID_ARM_MODE`，是否使用ARM模式。
- `ANDROID_ABI=armeabi-v7a`时，可设置`ON/OFF`，默认值为`ON`；
- `ANROID_ARM_MODE`，是否使用ARM模式。
- `ANDROID_ABI=armeabi-v7a`时，可设置`ON/OFF`，默认值为`ON`；
- `ANDROID_ABI=arm64-v8a`时，不需要设置。
- `ANDROID_ARM_NEON`，是否使用NEON指令。
- `ANDROID_ABI=armeabi-v7a`时，可设置`ON/OFF`，默认值为`ON`；
- `ANDROID_ARM_NEON`，是否使用NEON指令。
- `ANDROID_ABI=armeabi-v7a`时，可设置`ON/OFF`，默认值为`ON`；
- `ANDROID_ABI=arm64-v8a`时，不需要设置。
其他配置参数：
......@@ -119,7 +141,7 @@ cmake -DCMAKE_SYSTEM_NAME=Android \
-DANDROID_STANDALONE_TOOLCHAIN=your/path/to/arm64_standalone_toolchain \
-DANDROID_ABI=arm64-v8a \
-DUSE_EIGEN_FOR_BLAS=OFF \
-DCMAKE_INSTALL_PREFIX=your/path/to/install \
-DCMAKE_INSTALL_PREFIX=your/path/to/install \
-DWITH_C_API=ON \
-DWITH_SWIG_PY=OFF \
..
......@@ -128,8 +150,8 @@ cmake -DCMAKE_SYSTEM_NAME=Android \
用户还可根据自己的需求设置其他编译参数。比如希望最小化生成的库的大小，可以设置`CMAKE_BUILD_TYPE`为`MinSizeRel`；若希望最快的执行速度，则可设置`CMAKE_BUILD_TYPE`为`Release`。亦可以通过手动设置`CMAKE_C/CXX_FLAGS_MINSIZEREL/RELEASE`来影响PaddlePaddle的编译过程。
**性能TIPS**，为了达到最快的计算速度，在CMake参数配置上，有以下建议：
- 设置`CMAKE_BUILD_TYPE`为`Release`
- 使用`clang`编译工具链
- 设置`CMAKE_BUILD_TYPE`为`Release`
- 使用`clang`编译工具链
- `armeabi-v7a`时，设置`USE_EIGEN_BLAS=ON`，使用Eigen进行矩阵计算；`arm64-v8a`时，设置`USE_EIGEN_FOR_BLAS=OFF`，使用OpenBLAS进行矩阵计算
### 编译和安装
......
doc/mobile/cross_compiling_for_android_en.md 0 → 100644
浏览文件 @ 469b3ad5
# Build PaddlePaddle for Android
There are two approaches to build PaddlePaddle for Android: using Docker and on Linux without Docker.
## Cross-Compiling Using Docker
Docker-based cross-compiling is the recommended approach because Docker runs on all major operating systems, including Linux, Mac OS X, and Windows.
### Build the Docker Image
The following steps pack all the tools that we need to build PaddlePaddle into a Docker image.
```bash
$ git clone https://github.com/PaddlePaddle/Paddle.git
$ cd Paddle
$ docker build -t paddle:dev-android . -f Dockerfile.android
```
### Build the Inference Library
We can run the Docker image we just created to build the inference library of PaddlePaddle for Android using the command below:
```bash
$ docker run -it --rm -v $PWD:/paddle -e "ANDROID_ABI=armeabi-v7a" -e "ANDROID_API=21" paddle:dev-android
```
The Docker image accepts two arguments `ANDROID_ABI` and `ANDROID_API`:
<table class="docutils">
<colgroup>
<col width="25%" />
<col width="50%" />
<col width="25%" />
</colgroup>
<thead valign="bottom">
<tr class="row-odd">
<th class="head">Argument</th>
<th class="head">Optional Values</th>
<th class="head">Default</th>
</tr>
</thead>
<tbody valign="top">
<tr class="row-even">
<td>ANDROID_ABI</td>
<td>armeabi-v7a, arm64-v8a</td>
<td>armeabi-v7a</td>
</tr>
<tr class="row-odd">
<td>ANDROID_API</td>
<td>>= 21</td>
<td>21</td>
</tr>
</tbody>
</table>
The ARM-64 architecture (`arm64-v8a`) requires at least level 21 of Android API.
The default entry-point of the Docker image, [`paddle/scripts/docker/build_android.sh`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/scripts/docker/build_android.sh) generates the [Android cross-compiling standalone toolchain](https://developer.android.com/ndk/guides/standalone_toolchain.html) based on the argument: `ANDROID_ABI` or `ANDROID_API`. For information about other configuration arguments, please continue reading.
The above command generates and outputs the inference library in `$PWD/install_android` and puts third-party libraries in `$PWD/install_android/third_party`.
## Cross-Compiling on Linux
The Linux-base approach to cross-compile is to run steps in `Dockerfile.android` manually on a Linux x64 computer.
### Setup the Environment
To build for Android's, we need [Android NDK](
https://developer.android.com/ndk/downloads/index.html):
```bash
wget -q https://dl.google.com/android/repository/android-ndk-r14b-linux-x86_64.zip
unzip -q android-ndk-r14b-linux-x86_64.zip
```
Android NDK includes everything we need to build the [*standalone toolchain*](https://developer.android.com/ndk/guides/standalone_toolchain.html), which in then used to build PaddlePaddle for Android. (We plan to remove the intermediate stage of building the standalone toolchain in the near future.)
- To build the standalone toolchain for `armeabi-v7a` and Android API level 21:
```bash
your/path/to/android-ndk-r14b-linux-x86_64/build/tools/make-standalone-toolchain.sh \
--arch=arm --platform=android-21 --install-dir=your/path/to/arm_standalone_toolchain
```
The generated standalone toolchain will be in `your/path/to/arm_standalone_toolchain`.
- To build the standalone toolchain for `arm64-v8a` and Android API level 21:
```bash
your/path/to/android-ndk-r14b-linux-x86_64/build/tools/make-standalone-toolchain.sh \
--arch=arm64 --platform=android-21 --install-dir=your/path/to/arm64_standalone_toolchain
```
The generated standalone toolchain will be in `your/path/to/arm64_standalone_toolchain`.
**Please be aware that the minimum level of Android API required by PaddlePaddle is 21.**
### Cross-Compiling Arguments
CMake supports [choosing the toolchain](https://cmake.org/cmake/help/v3.0/manual/cmake-toolchains.7.html#cross-compiling). PaddlePaddle provides [`android.cmake`](https://github.com/PaddlePaddle/Paddle/blob/develop/cmake/cross_compiling/android.cmake), which configures the Android cross-compiling toolchain for CMake. `android.cmake` is not required for CMake >= 3.7, which support Android cross-compiling. PaddlePaddle detects the CMake version, for those newer than 3.7, it uses [the official version](https://cmake.org/cmake/help/v3.7/manual/cmake-toolchains.7.html#cross-compiling).
Some other CMake arguments you need to know:
- `CMAKE_SYSTEM_NAME` must be `Android`. This tells PaddlePaddle's CMake system to cross-compile third-party dependencies. This also changes some other CMake arguments like `WITH_GPU=OFF`, `WITH_AVX=OFF`, `WITH_PYTHON=OFF`, and `WITH_RDMA=OFF`.
- `WITH_C_API` must be `ON`, to build the C-based inference library for Android.
- `WITH_SWIG_PY` must be `OFF` because the Android platform doesn't support SWIG-based API.
Some Android-specific arguments:
- `ANDROID_STANDALONE_TOOLCHAIN`: the absolute path of the Android standalone toolchain, or the path relative to the CMake build directory. PaddlePaddle's CMake extensions would derive the cross-compiler, sysroot and Android API level from this argument.
- `ANDROID_TOOLCHAIN`: could be `gcc` or `clang`. The default value is `clang`.
- For CMake >= 3.7, it should anyway be `clang`. For older versions, it could be `gcc`.
- Android's official `clang` requires `glibc` >= 2.15.
- `ANDROID_ABI`: could be `armeabi-v7a` or `arm64-v8a`. The default value is `armeabi-v7a`.
- `ANDROID_NATIVE_API_LEVEL`: could be derived from the value of `ANDROID_STANDALONE_TOOLCHAIN`.
- `ANROID_ARM_MODE`:
- could be `ON` or `OFF`, and defaults to `ON`, when `ANDROID_ABI=armeabi-v7a`;
- no need to specify when `ANDROID_ABI=arm64-v8a`.
- `ANDROID_ARM_NEON`: indicates if to use NEON instructions.
- could be `ON` or `OFF`, and defaults to `ON`, when `ANDROID_ABI=armeabi-v7a`;
- no need to specify when `ANDROID_ABI=arm64-v8a`.
Other useful arguments:
- `USE_EIGEN_FOR_BLAS`: indicates if using Eigen. Could be `ON` or `OFF`, defaults to `OFF`.
- `HOST_C/CXX_COMPILER`: specifies the host compiler, which is used to build the host-specific protoc and target-specific OpenBLAS. It defaults to the value of the environment variable `CC`, or `cc`.
Some frequent configurations for your reference:
```bash
cmake -DCMAKE_SYSTEM_NAME=Android \
-DANDROID_STANDALONE_TOOLCHAIN=your/path/to/arm_standalone_toolchain \
-DANDROID_ABI=armeabi-v7a \
-DANDROID_ARM_NEON=ON \
-DANDROID_ARM_MODE=ON \
-DUSE_EIGEN_FOR_BLAS=ON \
-DCMAKE_INSTALL_PREFIX=your/path/to/install \
-DWITH_C_API=ON \
-DWITH_SWIG_PY=OFF \
..
```
```
cmake -DCMAKE_SYSTEM_NAME=Android \
-DANDROID_STANDALONE_TOOLCHAIN=your/path/to/arm64_standalone_toolchain \
-DANDROID_ABI=arm64-v8a \
-DUSE_EIGEN_FOR_BLAS=OFF \
-DCMAKE_INSTALL_PREFIX=your/path/to/install \
-DWITH_C_API=ON \
-DWITH_SWIG_PY=OFF \
..
```
There are some other arguments you might want to configure.
- `CMAKE_BUILD_TYPE=MinSizeRel` minimizes the size of library.
- `CMAKE_BUILD_TYPE-Release` optimizes the runtime performance.
Our own tip for performance optimization to use clang and Eigen or OpenBLAS:
- `CMAKE_BUILD_TYPE=Release`
- `ANDROID_TOOLCHAIN=clang`
- `USE_EIGEN_BLAS=ON` for `armeabi-v7a`, or `USE_EIGEN_FOR_BLAS=OFF` for `arm64-v8a`.
### Build and Install
After running `cmake`, we can run `make; make install` to build and install.
Before building, you might want to remove the `third_party` and `build` directories including pre-built libraries for other architectures.
After building，in the directory `CMAKE_INSTALL_PREFIX`, you will find three sub-directories:
- `include`: the header file of the inference library,
- `lib`: the inference library built for various Android ABIs,
- `third_party`: dependent third-party libraries built for Android.
doc/howto/cross_compiling/cross_compiling_for_ios_cn.md → doc/mobile/cross_compiling_for_ios_cn.md
浏览文件 @ 469b3ad5
......@@ -27,10 +27,28 @@ iOS平台可选配置参数：
- `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_PLATFORM | IOS_ARCH |
|--------------|----------------------|
| OS | armv7, armv7s, arm64 (默认) |
| SIMULATOR | i386, x86_64 (默认) |
<table class="docutils">
<colgroup>
<col width="35%" />
<col width="65%" />
</colgroup>
<thead valign="bottom">
<tr class="row-odd">
<th class="head">IOS_PLATFORM</th>
<th class="head">IOS_ARCH</th>
</tr>
</thead>
<tbody valign="top">
<tr class="row-even">
<td>OS</td>
<td>armv7, armv7s, arm64 (默认)</td>
</tr>
<tr class="row-odd">
<td>SIMULATOR</td>
<td>i386, x86_64 (默认)</td>
</tr>
</tbody>
</table>
- `IOS_DEPLOYMENT_TARGET`，最小的iOS部署版本，默认值为`7.0`。
- `IOS_ENABLE_BITCODE`，是否使能[Bitcode](https://developer.apple.com/library/content/documentation/IDEs/Conceptual/AppDistributionGuide/AppThinning/AppThinning.html#//apple_ref/doc/uid/TP40012582-CH35-SW3)，可设置`ON/OFF`，默认值为`ON`。
......
doc/mobile/index_cn.rst 0 → 100644
浏览文件 @ 469b3ad5
MOBILE
======
.. toctree::
:maxdepth: 1
cross_compiling_for_android_cn.md
cross_compiling_for_ios_cn.md
cross_compiling_for_raspberry_cn.md
doc/mobile/index_en.rst 0 → 100644
浏览文件 @ 469b3ad5
MOBILE
======
.. toctree::
:maxdepth: 1
cross_compiling_for_android_en.md
cross_compiling_for_raspberry_en.md
paddle/capi/CMakeLists.txt
浏览文件 @ 469b3ad5
......@@ -29,32 +29,32 @@ add_style_check_target(paddle_capi ${CAPI_SOURCES} ${CAPI_HEADER}
add_dependencies(paddle_capi paddle_proto)
# TODO: paddle_capi_whole will be removed.
set(PADDLE_CAPI_LAYERS_LIBS
paddle_function
paddle_gserver)
if(MOBILE_INFERENCE)
set(PADDLE_CAPI_INFER_LIBS
paddle_utils
paddle_parameter
paddle_math
paddle_cuda
paddle_function
paddle_gserver
paddle_proto)
set(PADDLE_CAPI_ENGINE_LIBS
paddle_utils
paddle_parameter
paddle_math
paddle_cuda
paddle_proto)
else()
set(PADDLE_CAPI_INFER_LIBS
paddle_utils
paddle_parameter
paddle_math
paddle_cuda
paddle_function
paddle_gserver
paddle_proto
paddle_pserver
paddle_network)
set(PADDLE_CAPI_ENGINE_LIBS
paddle_utils
paddle_parameter
paddle_math
paddle_cuda
paddle_proto
paddle_pserver
paddle_network)
endif()
set(PADDLE_CAPI_INFER_LIBS ${PADDLE_CAPI_LAYERS_LIBS} ${PADDLE_CAPI_ENGINE_LIBS})
cc_library(paddle_capi_whole DEPS paddle_capi ${PADDLE_CAPI_INFER_LIBS})
# Link the static library for inference
cc_library(paddle_capi_engine DEPS paddle_capi paddle_utils paddle_parameter paddle_math paddle_cuda paddle_proto)
cc_library(paddle_capi_layers DEPS paddle_function paddle_gserver)
cc_library(paddle_capi_engine DEPS paddle_capi ${PADDLE_CAPI_ENGINE_LIBS})
cc_library(paddle_capi_layers DEPS ${PADDLE_CAPI_LAYERS_LIBS})
# Link the shared library for inference
if(NOT IOS)
......
paddle/capi/Matrix.cpp
浏览文件 @ 469b3ad5
......@@ -54,6 +54,46 @@ paddle_error paddle_matrix_set_row(paddle_matrix mat,
return kPD_NO_ERROR;
}
PD_API paddle_error paddle_matrix_set_value(paddle_matrix mat,
paddle_real* value) {
if (mat == nullptr || value == nullptr) return kPD_NULLPTR;
auto ptr = cast(mat);
if (ptr->mat == nullptr) return kPD_NULLPTR;
paddle::real* buf = ptr->mat->getRowBuf(0);
size_t width = ptr->mat->getWidth();
size_t height = ptr->mat->getHeight();
if (ptr->mat->useGpu()) {
#ifdef PADDLE_WITH_CUDA
hl_memcpy(buf, value, sizeof(paddle::real) * width * height);
#else
return kPD_NOT_SUPPORTED;
#endif
} else {
std::copy(value, value + width * height, buf);
}
return kPD_NO_ERROR;
}
PD_API paddle_error paddle_matrix_get_value(paddle_matrix mat,
paddle_real* result) {
if (mat == nullptr || result == nullptr) return kPD_NULLPTR;
auto ptr = cast(mat);
if (ptr->mat == nullptr) return kPD_NULLPTR;
paddle::real* buf = ptr->mat->getRowBuf(0);
size_t width = ptr->mat->getWidth();
size_t height = ptr->mat->getHeight();
if (ptr->mat->useGpu()) {
#ifdef PADDLE_WITH_CUDA
hl_memcpy(result, buf, width * height * sizeof(paddle::real));
#else
return kPD_NOT_SUPPORTED;
#endif
} else {
std::copy(buf, buf + width * height, result);
}
return kPD_NO_ERROR;
}
paddle_error paddle_matrix_get_row(paddle_matrix mat,
uint64_t rowID,
paddle_real** rawRowBuffer) {
......
paddle/capi/examples/model_inference/dense/main.c
浏览文件 @ 469b3ad5
......@@ -27,18 +27,20 @@ int main() {
CHECK(paddle_arguments_resize(in_args, 1));
// Create input matrix.
paddle_matrix mat = paddle_matrix_create(/* sample_num */ 1,
paddle_matrix mat = paddle_matrix_create(/* sample_num */ 10,
/* size */ 784,
/* useGPU */ false);
srand(time(0));
paddle_real* array;
// Get First row.
CHECK(paddle_matrix_get_row(mat, 0, &array));
std::vector<paddle_real> input;
input.resize(784 * 10);
for (int i = 0; i < 784; ++i) {
array[i] = rand() / ((float)RAND_MAX);
for (int i = 0; i < input.size(); ++i) {
input[i] = rand() / ((float)RAND_MAX);
}
// Set value for the input matrix
CHECK(paddle_matrix_set_value(mat, input.data()));
CHECK(paddle_arguments_set_value(in_args, 0, mat));
......@@ -51,11 +53,17 @@ int main() {
CHECK(paddle_arguments_get_value(out_args, 0, prob));
CHECK(paddle_matrix_get_row(prob, 0, &array));
std::std::vector<paddle_real> result;
int height;
int width;
CHECK(paddle_matrix_get_shape(prob, &height, &width);
result.resize(height * width);
CHECK(paddle_matrix_get_value(prob, result.data()));
printf("Prob: ");
for (int i = 0; i < 10; ++i) {
printf("%.2f ", array[i]);
for (int i = 0; i < height * width; ++i) {
printf("%.2f ", result[i]);
}
printf("\n");
......
paddle/capi/matrix.h
浏览文件 @ 469b3ad5
......@@ -70,6 +70,16 @@ PD_API paddle_error paddle_matrix_set_row(paddle_matrix mat,
uint64_t rowID,
paddle_real* rowArray);
/**
* @brief paddle_matrix_set_value Set value to matrix.
* @param mat Target Matrix
* @param value Row data.
* @return paddle_error
* @note value should contain enough element of data to init the mat
*/
PD_API paddle_error paddle_matrix_set_value(paddle_matrix mat,
paddle_real* value);
/**
* @brief PDMatGetRow Get raw row buffer from matrix
* @param [in] mat Target matrix
......@@ -81,6 +91,15 @@ PD_API paddle_error paddle_matrix_get_row(paddle_matrix mat,
uint64_t rowID,
paddle_real** rawRowBuffer);
/**
* @brief copy data from the matrix
* @param [in] mat Target matrix
* @param [out] result pointer to store the matrix data
* @return paddle_error
* @note the space of the result should allocated before invoke this API
*/
PD_API paddle_error paddle_matrix_get_value(paddle_matrix mat,
paddle_real* result);
/**
* @brief PDMatCreateNone Create None Matrix
* @return
......
paddle/capi/tests/test_Matrix.cpp
浏览文件 @ 469b3ad5
......@@ -45,3 +45,49 @@ TEST(CAPIMatrix, createNone) {
paddle_matrix mat = paddle_matrix_create_none();
ASSERT_EQ(kPD_NO_ERROR, paddle_matrix_destroy(mat));
}
TEST(CAPIMatrix, cpu_get_set_value) {
paddle_matrix mat = paddle_matrix_create(128, 32, false);
std::vector<paddle_real> sample;
std::vector<paddle_real> result;
sample.resize(128 * 32);
result.resize(128 * 32);
for (size_t i = 0; i < sample.size(); ++i) {
sample[i] = 1.0 / (i + 1.0);
}
ASSERT_EQ(kPD_NO_ERROR, paddle_matrix_set_value(mat, sample.data()));
ASSERT_EQ(kPD_NO_ERROR, paddle_matrix_get_value(mat, result.data()));
for (size_t i = 0; i < sample.size(); ++i) {
ASSERT_NEAR(sample[i], result[i], 1e-5);
}
uint64_t height, width;
ASSERT_EQ(kPD_NO_ERROR, paddle_matrix_get_shape(mat, &height, &width));
ASSERT_EQ(128UL, height);
ASSERT_EQ(32UL, width);
ASSERT_EQ(kPD_NO_ERROR, paddle_matrix_destroy(mat));
}
#ifdef PADDLE_WITH_CUDA
TEST(CAPIMatrix, gpu_get_set_value) {
paddle_matrix mat = paddle_matrix_create(128, 32, true);
std::vector<paddle_real> sample;
std::vector<paddle_real> result;
sample.resize(128 * 32);
result.resize(128 * 32);
for (size_t i = 0; i < sample.size(); ++i) {
sample[i] = 1.0 / (i + 1.0);
}
ASSERT_EQ(kPD_NO_ERROR, paddle_matrix_set_value(mat, sample.data()));
ASSERT_EQ(kPD_NO_ERROR, paddle_matrix_get_value(mat, result.data()));
for (size_t i = 0; i < sample.size(); ++i) {
ASSERT_NEAR(sample[i], result[i], 1e-5);
}
uint64_t height, width;
ASSERT_EQ(kPD_NO_ERROR, paddle_matrix_get_shape(mat, &height, &width));
ASSERT_EQ(128UL, height);
ASSERT_EQ(32UL, width);
ASSERT_EQ(kPD_NO_ERROR, paddle_matrix_destroy(mat));
}
#endif
paddle/cuda/include/hl_matrix.h
浏览文件 @ 469b3ad5
......@@ -300,4 +300,12 @@ extern void hl_matrix_col2Vol(real* dataDst,
real alpha,
real beta);
/**
* @brief Matrix col2Vol: Convert col matrix into 3D volume
* @param[out] out output int vector.
* @param[in] vec input float vector.
* @param[in] size size of the vector.
*/
extern void hl_vector_cast2int(int* out, real* vec, int size);
#endif /* HL_MATRIX_H_ */
paddle/cuda/include/stub/hl_matrix_stub.h
浏览文件 @ 469b3ad5
......@@ -133,4 +133,6 @@ inline void hl_matrix_col2Vol(real* dataDst,
real alpha,
real beta) {}
inline void hl_vector_cast2int(int* out, real* vec, int size) {}
#endif // HL_MATRIX_STUB_H_
paddle/cuda/src/hl_cuda_matrix.cu
浏览文件 @ 469b3ad5
......@@ -793,3 +793,14 @@ void hl_matrix_col2Vol(real* dataDst,
CHECK_SYNC("hl_matrix_col2Vol failed");
}
__global__ void keVectorCast2Int(int* out, real* vec, int size) {
for (int i = threadIdx.x; i < (size); i += blockDim.x) {
out[i] = int(vec[i]);
}
}
void hl_vector_cast2int(int* out, real* vec, int size) {
keVectorCast2Int<<<1, 512, 0, STREAM_DEFAULT>>>(out, vec, size);
CHECK_SYNC("hl_vector_cast2int failed");
}
paddle/framework/CMakeLists.txt
浏览文件 @ 469b3ad5
......@@ -20,7 +20,8 @@ cc_test(scope_test SRCS scope_test.cc DEPS scope)
cc_library(attribute SRCS attribute.cc DEPS framework_proto)
cc_test(program_desc_test SRCS program_desc_test.cc DEPS proto_desc)
cc_test(program_desc_test SRCS program_desc_test.cc DEPS proto_desc
device_context)
cc_library(op_proto_maker SRCS op_proto_maker.cc DEPS framework_proto attribute)
cc_test(op_proto_maker_test SRCS op_proto_maker_test.cc DEPS op_proto_maker)
cc_library(op_info SRCS op_info.cc DEPS attribute framework_proto)
......@@ -44,8 +45,9 @@ add_custom_command(TARGET framework_py_proto POST_BUILD
cc_library(backward SRCS backward.cc DEPS net_op)
cc_test(backward_test SRCS backward_test.cc DEPS backward recurrent_op device_context fill_constant_op)
cc_library(lod_rank_table SRCS lod_rank_table.cc DEPS lod_tensor)
cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto backward glog)
cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto backward glog lod_rank_table)
cc_library(prune SRCS prune.cc DEPS framework_proto)
cc_test(prune_test SRCS prune_test.cc DEPS op_info prune recurrent_op device_context)
......
paddle/framework/backward.cc
浏览文件 @ 469b3ad5
......@@ -24,7 +24,6 @@
#include "paddle/framework/op_registry.h"
#include "paddle/operators/dynamic_recurrent_op.h"
#include "paddle/operators/net_op.h"
#include "paddle/operators/recurrent_op.h"
namespace paddle {
namespace framework {
......@@ -38,7 +37,7 @@ static inline std::unique_ptr<OperatorBase> CreateGradOp(
op_desc.SetType(op.Type());
op_desc.SetAttrMap(op.Attrs());
auto& info = OpInfoMap::Instance().Get(op.Type());
auto grad_descs = info.GradOpMaker()(op_desc, no_grad_set, grad_to_var);
auto grad_descs = info.GradOpMaker()(op_desc, no_grad_set, grad_to_var, {});
std::vector<std::unique_ptr<OperatorBase>> grad_ops;
grad_ops.reserve(grad_descs.size());
std::transform(grad_descs.begin(), grad_descs.end(),
......@@ -220,19 +219,7 @@ static std::unique_ptr<OperatorBase> BackwardRecursive(
});
// process recurrent gradient op as a special operator.
if (forwardOp.Type() == "recurrent") {
// NOTE clean up cycle call somewhere (RNN's stepnet constains itself),
// or this will result in infinite loop.
const auto& rnnop =
*static_cast<const operators::RecurrentOp*>(&forwardOp);
auto rnn_grad_op =
static_cast<operators::RecurrentGradientOp*>(grad_op.get());
const auto& stepnet_op =
*static_cast<const OperatorBase*>(&rnnop.stepnet());
// create stepnet's gradient op
rnn_grad_op->set_stepnet(
BackwardRecursive(stepnet_op, no_grad_names, grad_to_var, uniq_id));
} else if (forwardOp.Type() == "dynamic_recurrent") {
if (forwardOp.Type() == "dynamic_recurrent") {
// NOTE clean up cycle call somewhere (RNN's stepnet constains itself),
// or this will result in infinite loop.
const auto& rnnop =
......@@ -331,11 +318,9 @@ static void CreateGradVarInBlock(
continue;
}
auto pname = FwdName(arg);
auto* param = block_desc->FindVar(pname);
auto* param = block_desc->FindVarRecursive(pname);
auto* grad = block_desc->FindVar(arg);
if (param == nullptr) {
LOG(WARNING) << "Cannot find forward variable of " << arg
<< ". Set its gradient to FP32";
grad->SetDataType(DataType::FP32);
} else {
grad->SetDataType(param->GetDataType());
......@@ -348,7 +333,9 @@ static void CreateGradVarInBlock(
std::vector<std::unique_ptr<OpDescBind>> MakeOpGrad(
const OpDescBind* op_desc, std::unordered_set<std::string>* no_grad_vars,
std::unordered_map<std::string, std::string>* grad_to_var) {
std::unordered_map<std::string, std::string>* grad_to_var,
const std::vector<BlockDescBind*>& grad_block =
std::vector<BlockDescBind*>()) {
std::vector<std::unique_ptr<OpDescBind>> grad_op_descs;
// All input gradients of forwarding operator do not need to calculate.
const std::vector<std::string>& inputs = op_desc->InputArgumentNames();
......@@ -364,9 +351,10 @@ std::vector<std::unique_ptr<OpDescBind>> MakeOpGrad(
return grad_op_descs; // empty vector
}
grad_op_descs = OpInfoMap::Instance()
.Get(op_desc->Type())
.GradOpMaker()(*op_desc, *no_grad_vars, grad_to_var);
grad_op_descs =
OpInfoMap::Instance()
.Get(op_desc->Type())
.GradOpMaker()(*op_desc, *no_grad_vars, grad_to_var, grad_block);
std::list<std::unique_ptr<OpDescBind>> pending_fill_zeros_ops;
for (auto& desc : grad_op_descs) {
......@@ -389,6 +377,12 @@ std::vector<std::unique_ptr<OpDescBind>> MakeOpGrad(
return grad_op_descs;
}
static BlockDescBind* CreateStepBlock(
ProgramDescBind& program_desc,
std::unordered_set<std::string>* no_grad_vars,
std::unordered_map<std::string, std::string>* grad_to_var,
int step_block_idx);
std::vector<std::unique_ptr<OpDescBind>> MakeBlockBackward(
ProgramDescBind& program_desc, int block_idx,
std::unordered_set<std::string>* no_grad_vars,
......@@ -400,25 +394,29 @@ std::vector<std::unique_ptr<OpDescBind>> MakeBlockBackward(
std::vector<std::unique_ptr<OpDescBind>> backward_descs;
for (auto it = op_descs.rbegin(); it != op_descs.rend(); ++it) {
std::vector<std::unique_ptr<OpDescBind>> op_grads =
MakeOpGrad(*it, no_grad_vars, grad_to_var);
std::vector<std::unique_ptr<OpDescBind>> op_grads;
if ((*it)->Type() == "recurrent") {
PADDLE_ENFORCE_EQ(
op_grads.size(), static_cast<size_t>(1),
"rnn_op's gradient process should contain only one op.");
int step_block_idx = (*it)->GetBlockAttr("step_block");
auto backward_block_op_descs = MakeBlockBackward(
program_desc, step_block_idx, no_grad_vars, grad_to_var);
BlockDescBind* backward_block = program_desc.AppendBlock(*cur_block);
for (auto& ptr : backward_block_op_descs) {
backward_block->AppendAllocatedOp(std::move(ptr));
}
op_grads[0]->SetBlockAttr("step_block", *backward_block);
BlockDescBind* backward_block = CreateStepBlock(
program_desc, no_grad_vars, grad_to_var, step_block_idx);
op_grads = MakeOpGrad(*it, no_grad_vars, grad_to_var, {backward_block});
} else if ((*it)->Type() == "conditional_block") {
BlockDescBind* backward_block =
CreateStepBlock(program_desc, no_grad_vars, grad_to_var,
(*it)->GetBlockAttr("block"));
op_grads = MakeOpGrad(*it, no_grad_vars, grad_to_var, {backward_block});
} else {
op_grads = MakeOpGrad(*it, no_grad_vars, grad_to_var);
}
for (const auto& desc : op_grads) {
for (const std::string& out_name : desc->OutputArgumentNames()) {
if (out_name.find("@GRAD") == std::string::npos) {
// Not all outputs of a backward operator is a gradient. Only gradient
// need to be sum. Skip variables are not gradient.
continue;
}
dup_out_ops[out_name].emplace_back(grad_desc_idx);
}
++grad_desc_idx;
......@@ -457,6 +455,21 @@ std::vector<std::unique_ptr<OpDescBind>> MakeBlockBackward(
return backward_descs;
}
static BlockDescBind* CreateStepBlock(
ProgramDescBind& program_desc,
std::unordered_set<std::string>* no_grad_vars,
std::unordered_map<std::string, std::string>* grad_to_var,
int step_block_idx) {
auto backward_block_op_descs = MakeBlockBackward(program_desc, step_block_idx,
no_grad_vars, grad_to_var);
BlockDescBind* backward_block =
program_desc.AppendBlock(*program_desc.MutableBlock(step_block_idx));
for (auto& ptr : backward_block_op_descs) {
backward_block->AppendAllocatedOp(move(ptr));
}
return backward_block;
}
ParamGradInfoMap AppendBackward(
ProgramDescBind& program_desc, const VarDescBind& target,
const std::unordered_set<std::string>& no_grad_vars) {
......
paddle/framework/backward_test.cc
浏览文件 @ 469b3ad5
......@@ -21,7 +21,7 @@
#include "paddle/framework/var_desc.h"
#include "paddle/operators/net_op.h"
USE_OP(fill_constant);
USE_NO_KERNEL_OP(fill_constant);
namespace paddle {
namespace framework {
......
paddle/framework/block_desc.cc
浏览文件 @ 469b3ad5
......@@ -50,6 +50,15 @@ VarDescBind *BlockDescBind::FindVarRecursive(const std::string &name) const {
return it->second.get();
}
VarDescBind *BlockDescBind::FindRecursiveOrCreateVar(
const std::string &name_bytes) {
VarDescBind *res = FindVarRecursive(name_bytes);
if (res == nullptr) {
res = Var(name_bytes);
}
return res;
}
bool BlockDescBind::HasVarRecursive(const std::string &name) const {
return FindVarRecursive(name) != nullptr;
}
......
paddle/framework/block_desc.h
浏览文件 @ 469b3ad5
......@@ -58,6 +58,8 @@ class BlockDescBind {
VarDescBind *FindVarRecursive(const std::string &name_bytes) const;
VarDescBind *FindRecursiveOrCreateVar(const std::string &name_bytes);
bool HasVarRecursive(const std::string &var_name) const;
std::set<std::string> LocalVarNames() const {
......@@ -88,6 +90,8 @@ class BlockDescBind {
BlockDesc *Proto();
ProgramDescBind *Program() { return this->prog_; }
private:
void ClearPBOps();
void ClearPBVars();
......
paddle/framework/data_type.h
浏览文件 @ 469b3ad5
......@@ -34,6 +34,21 @@ inline DataType ToDataType(std::type_index type) {
}
}
inline std::type_index ToTypeIndex(DataType type) {
switch (type) {
case DataType::FP32:
return typeid(float);
case DataType::FP64:
return typeid(double);
case DataType::INT32:
return typeid(int);
case DataType::INT64:
return typeid(int64_t);
default:
PADDLE_THROW("Not support type %d", type);
}
}
template <typename Visitor>
inline void VisitDataType(DataType type, Visitor visitor) {
switch (type) {
......
paddle/framework/ddim.cc
浏览文件 @ 469b3ad5
......@@ -79,6 +79,13 @@ DDim make_ddim(const std::vector<int64_t>& dims) {
return result;
}
DDim make_ddim(const std::vector<int>& dims) {
std::vector<int64_t> res(dims.size());
std::transform(dims.begin(), dims.end(), res.begin(),
[](int d) { return static_cast<int64_t>(d); });
return make_ddim(res);
}
/// @cond HIDDEN
// XXX For some reason, putting this in an anonymous namespace causes errors
class DynamicMutableIndexer : public boost::static_visitor<int64_t&> {
......@@ -117,7 +124,7 @@ int64_t DDim::operator[](int idx) const {
return boost::apply_visitor(DynamicConstIndexer(idx), var);
}
int64_t DDim::size() const { return arity(*this); }
int DDim::size() const { return arity(*this); }
bool DDim::operator==(DDim d) const {
if (var.which() != d.getVar().which()) {
......
paddle/framework/ddim.h
浏览文件 @ 469b3ad5
......@@ -71,7 +71,7 @@ struct DDim {
DDim operator*(DDim d) const;
int64_t size() const;
int size() const;
};
/**
......@@ -81,6 +81,8 @@ struct DDim {
*/
DDim make_ddim(const std::vector<int64_t>& dims);
DDim make_ddim(const std::vector<int>& dims);
/**
* \brief Make a DDim from an initializer list
*
......
paddle/framework/details/op_registry.h
浏览文件 @ 469b3ad5
......@@ -108,8 +108,9 @@ struct OpInfoFiller<T, kGradOpDescMaker> {
info->grad_op_maker_ = [](
const OpDescBind& fwd_op,
const std::unordered_set<std::string>& no_grad_set,
std::unordered_map<std::string, std::string>* grad_to_var) {
T maker(fwd_op, no_grad_set, grad_to_var);
std::unordered_map<std::string, std::string>* grad_to_var,
const std::vector<BlockDescBind*>& grad_block) {
T maker(fwd_op, no_grad_set, grad_to_var, grad_block);
return maker();
};
}
......
paddle/framework/executor.cc
浏览文件 @ 469b3ad5
......@@ -21,7 +21,9 @@ limitations under the License. */
#include <vector>
#include "paddle/framework/feed_fetch_type.h"
#include "paddle/framework/lod_rank_table.h"
#include "paddle/framework/lod_tensor.h"
#include "paddle/framework/lod_tensor_array.h"
#include "paddle/framework/op_registry.h"
#include "paddle/framework/scope.h"
......@@ -31,7 +33,7 @@ namespace framework {
const std::string kFeedOpType = "feed";
const std::string kFetchOpType = "fetch";
Executor::Executor(const std::vector<platform::Place>& places) {
Executor::Executor(const std::vector<platform::Place>& places) : own_(true) {
PADDLE_ENFORCE_GT(places.size(), 0);
device_contexts_.resize(places.size());
for (size_t i = 0; i < places.size(); i++) {
......@@ -52,8 +54,10 @@ Executor::Executor(const std::vector<platform::Place>& places) {
}
Executor::~Executor() {
for (auto& device_context : device_contexts_) {
delete device_context;
if (own_) {
for (auto& device_context : device_contexts_) {
delete device_context;
}
}
}
......@@ -66,44 +70,65 @@ static void CreateTensor(Variable* var, VarDesc::VarType var_type) {
var->GetMutable<FeedFetchList>();
} else if (var_type == VarDesc::FETCH_LIST) {
var->GetMutable<FeedFetchList>();
} else if (var_type == VarDesc::STEP_SCOPES) {
var->GetMutable<std::vector<framework::Scope>>();
} else if (var_type == VarDesc::LOD_RANK_TABLE) {
var->GetMutable<LoDRankTable>();
} else if (var_type == VarDesc::LOD_TENSOR_ARRAY) {
var->GetMutable<LoDTensorArray>();
} else {
PADDLE_THROW(
"Variable type must be "
"LoDTensor/SelectedRows/FEED_MINIBATCH/FETCH_LIST.");
"Variable type %d is not in "
"[LoDTensor, SelectedRows, FEED_MINIBATCH, FETCH_LIST, LOD_RANK_TABLE]",
var_type);
}
}
void Executor::Run(const ProgramDescBind& pdesc, Scope* scope, int block_id) {
void Executor::Run(const ProgramDescBind& pdesc, Scope* scope, int block_id,
bool create_local_scope) {
// TODO(tonyyang-svail):
// - only runs on the first device (i.e. no interdevice communication)
// - will change to use multiple blocks for RNN op and Cond Op
PADDLE_ENFORCE_LT(block_id, pdesc.Size());
PADDLE_ENFORCE_LT(static_cast<size_t>(block_id), pdesc.Size());
auto& block = pdesc.Block(block_id);
auto& device = device_contexts_[0];
Scope& local_scope = scope->NewScope();
for (auto& var : block.AllVars()) {
if (var->Persistable()) {
auto* ptr = scope->Var(var->Name());
CreateTensor(ptr, var->GetType());
VLOG(3) << "Create Variable " << var->Name()
<< " global, which pointer is " << ptr;
} else {
auto* ptr = local_scope.Var(var->Name());
Scope* local_scope = scope;
if (create_local_scope) {
local_scope = &scope->NewScope();
for (auto& var : block.AllVars()) {
if (var->Persistable()) {
auto* ptr = scope->Var(var->Name());
CreateTensor(ptr, var->GetType());
VLOG(3) << "Create Variable " << var->Name()
<< " global, which pointer is " << ptr;
} else {
auto* ptr = local_scope->Var(var->Name());
CreateTensor(ptr, var->GetType());
VLOG(3) << "Create Variable " << var->Name()
<< " locally, which pointer is " << ptr;
}
}
} else {
for (auto& var : block.AllVars()) {
auto* ptr = local_scope->Var(var->Name());
CreateTensor(ptr, var->GetType());
VLOG(3) << "Create Variable " << var->Name()
<< " locally, which pointer is " << ptr;
VLOG(3) << "Create variable " << var->Name() << ", which pointer is "
<< ptr;
}
}
for (auto& op_desc : block.AllOps()) {
auto op = paddle::framework::OpRegistry::CreateOp(*op_desc);
op->Run(local_scope, *device);
op->Run(*local_scope, *device);
}
if (create_local_scope) {
scope->DeleteScope(local_scope);
}
scope->DeleteScope(&local_scope);
}
Executor::Executor(const platform::DeviceContext& device)
: device_contexts_({&device}), own_(false) {}
} // namespace framework
} // namespace paddle
paddle/framework/executor.h
浏览文件 @ 469b3ad5
......@@ -25,6 +25,7 @@ namespace framework {
class Executor {
public:
explicit Executor(const std::vector<platform::Place>& places);
explicit Executor(const platform::DeviceContext& devices);
~Executor();
/* @Brief
......@@ -34,10 +35,11 @@ class Executor {
* ProgramDesc
* Scope
*/
void Run(const ProgramDescBind&, Scope*, int);
void Run(const ProgramDescBind&, Scope*, int, bool create_local_scope = true);
private:
std::vector<platform::DeviceContext*> device_contexts_;
std::vector<const platform::DeviceContext*> device_contexts_;
bool own_;
};
} // namespace framework
......
paddle/framework/framework.proto
浏览文件 @ 469b3ad5
......@@ -109,6 +109,11 @@ message LoDTensorDesc {
optional int32 lod_level = 2 [ default = 0 ];
}
message LoDTensorArrayDesc {
required TensorDesc tensor = 1;
optional int32 lod_level = 2 [ default = 0 ];
}
message VarDesc {
enum VarType {
LOD_TENSOR = 1;
......@@ -116,11 +121,14 @@ message VarDesc {
FEED_MINIBATCH = 3;
FETCH_LIST = 4;
STEP_SCOPES = 5;
LOD_RANK_TABLE = 6;
LOD_TENSOR_ARRAY = 7;
}
required string name = 1;
required VarType type = 2;
optional LoDTensorDesc lod_tensor = 3;
optional TensorDesc selected_rows = 4;
optional LoDTensorArrayDesc tensor_array = 6;
optional bool persistable = 5 [ default = false ];
}
......
paddle/framework/grad_op_desc_maker.h
浏览文件 @ 469b3ad5
......@@ -15,6 +15,7 @@
#pragma once
#include <string>
#include <unordered_set>
#include <vector>
#include "paddle/framework/op_desc.h"
#include "paddle/framework/operator.h"
......@@ -26,8 +27,13 @@ class GradOpDescMakerBase {
explicit GradOpDescMakerBase(
const OpDescBind& fwd_op,
const std::unordered_set<std::string>& no_grad_set,
std::unordered_map<std::string, std::string>* grad_to_var)
: fwd_op_(fwd_op), no_grad_set_(no_grad_set), grad_to_var_(grad_to_var) {}
std::unordered_map<std::string, std::string>* grad_to_var,
const std::vector<BlockDescBind*>& grad_block =
std::vector<BlockDescBind*>())
: fwd_op_(fwd_op),
no_grad_set_(no_grad_set),
grad_to_var_(grad_to_var),
grad_block_(grad_block) {}
virtual ~GradOpDescMakerBase() = default;
virtual std::vector<std::unique_ptr<OpDescBind>> operator()() const = 0;
......@@ -102,6 +108,9 @@ class GradOpDescMakerBase {
const OpDescBind& fwd_op_;
const std::unordered_set<std::string>& no_grad_set_;
std::unordered_map<std::string, std::string>* grad_to_var_;
protected:
std::vector<BlockDescBind*> grad_block_;
};
class SingleGradOpDescMaker : public GradOpDescMakerBase {
......
paddle/framework/lod_rank_table.cc 0 → 100644
浏览文件 @ 469b3ad5
/* 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. */
#include "paddle/framework/lod_rank_table.h"
namespace paddle {
namespace framework {
void LoDRankTable::Reset(const LoD& lod, size_t level) {
this->coarse_lod_.clear();
this->items_.clear();
PADDLE_ENFORCE(level < lod.size(),
"Cannot rank lod since the level %d is less than lod size %d",
level, lod.size());
coarse_lod_.reserve(level);
for (size_t i = 0; i < level; ++i) {
coarse_lod_.push_back(lod[i]);
}
auto& vec = lod[level];
for (size_t i = 0; i < vec.size() - 1; ++i) {
TableItem item;
item.index = i;
item.length = vec[i + 1] - vec[i];
VLOG(10) << "Add item to rank table " << item.index << " " << item.length;
items_.emplace_back(item);
}
// NOTE(yuyang18):
//
// The time complexity of stable_sort is O(N*log(N)) if additional memory is
// available. It is easy to debug and unit test when using `stable_sort`
// instead of `sort`. Also, the items of a rank table will not be too large.
std::stable_sort(items_.begin(), items_.end(),
[](const TableItem& a, const TableItem& b) {
return a.length > b.length;
});
}
} // namespace framework
} // namespace paddle
paddle/framework/lod_rank_table.h 0 → 100644
浏览文件 @ 469b3ad5
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include "paddle/framework/lod_tensor.h"
namespace paddle {
namespace framework {
// LoD Rank Table stores the `level` of `lod` which is ordered by sequence
// length in descending order. It is useful when implement dynamic RNN and is
// shared by dynamic RNN memory, dynamic RNN slice input and dynamic RNN slice
// output operators.
//
// The table item contains two element. The length of sequence and the index of
// sequence in that level.
//
// LoDRankTable also stores the coarse_lod, which is the lod information whose
// level is less than input level, in order to restore the output LoD
// information.
class LoDRankTable {
public:
struct TableItem {
size_t index;
size_t length;
};
LoDRankTable() {}
void Reset(const LoD& lod, size_t level);
const std::vector<TableItem>& items() const { return this->items_; }
const LoD& coarse_lod() const { return this->coarse_lod_; }
size_t level() const { return coarse_lod_.size(); }
private:
LoD coarse_lod_;
std::vector<TableItem> items_;
};
} // namespace framework
} // namespace paddle
paddle/framework/lod_tensor.cc
浏览文件 @ 469b3ad5
......@@ -27,6 +27,20 @@
namespace paddle {
namespace framework {
std::ostream& operator<<(std::ostream& os, const LoD& lod) {
os << "{";
for (auto& v : lod) {
os << "{";
for (auto& i : v) {
os << i << ",";
}
os << "}";
}
os << "}";
return os;
}
LoD SliceLevels(const LoD& in, size_t level_begin, size_t level_end) {
LoD new_lod;
new_lod.reserve(level_end - level_begin);
......@@ -135,5 +149,41 @@ void LoDTensor::ShrinkInLevel(size_t level, size_t elem_begin,
PADDLE_ENFORCE_LT(begin, end, "Cannot shrink, the result tensor is empty.");
ShareDataWith(Slice(begin, end));
}
using LoDAndOffset = std::pair<LoD, std::pair<size_t, size_t>>;
LoDAndOffset GetSubLoDAndAbsoluteOffset(const LoD& lod, size_t start_idx,
size_t end_idx, size_t start_level) {
LoD sub_lod;
for (size_t level_idx = start_level; level_idx < lod.size(); ++level_idx) {
PADDLE_ENFORCE_LE(start_idx, end_idx);
PADDLE_ENFORCE_LT(end_idx, lod[level_idx].size());
std::vector<size_t> level_lens;
for (size_t i = start_idx; i < end_idx; ++i) {
level_lens.push_back(lod[level_idx][i + 1] - lod[level_idx][i]);
}
sub_lod.emplace_back(level_lens);
start_idx = lod[level_idx][start_idx];
end_idx = lod[level_idx][end_idx];
}
return LoDAndOffset{sub_lod, {start_idx, end_idx}};
}
void AppendLoD(LoD* lod, const LoD& lod_length) {
PADDLE_ENFORCE(
lod->empty() || lod->size() == lod_length.size(),
"The lod_length should has the same size with the appended lod.");
if (lod->empty()) {
*lod = LoD(lod_length.size(), std::vector<size_t>({0}));
}
for (size_t i = 0; i < lod->size(); ++i) {
auto& level = (*lod)[i];
for (size_t len : lod_length[i]) {
level.push_back(level.back() + len);
}
}
}
} // namespace framework
} // namespace paddle
paddle/framework/lod_tensor.h
浏览文件 @ 469b3ad5
......@@ -56,6 +56,8 @@ using Vector = thrust::host_vector<
*/
using LoD = std::vector<Vector<size_t>>;
std::ostream& operator<<(std::ostream& os, const LoD& lod);
/*
* Slice levels from a LoD.
* NOTE the lowest level should always be the absolute offsets of the underlying
......@@ -181,5 +183,10 @@ LoDTensor LodExpand(const LoDTensor& source, const LoD& lod, size_t level,
return tensor;
}
std::pair<LoD, std::pair<size_t, size_t>> GetSubLoDAndAbsoluteOffset(
const LoD& lod, size_t start_idx, size_t end_idx, size_t start_level);
void AppendLoD(LoD* lod, const LoD& lod_length);
} // namespace framework
} // namespace paddle
paddle/framework/lod_tensor.md
浏览文件 @ 469b3ad5
......@@ -140,19 +140,9 @@ Similarly, the lengths in the top level LoD
are transformed into offsets of elements/words as follows:
```
0 9 10 15
= = =
3+2+4 1+9 2+3+10
```
so we can tell that the first article is from word 0 to word 9, and the second article is from word 9 to word 10.
The complete offset representation is as follows:
```
0 9 10 15
0 3 5 9 10 12 15
||| || |||| | || |||
0 3 4 6
= = =
3 3+1 4+2
```
## Slicing of LoD Tensors
......
paddle/framework/lod_tensor_array.h 0 → 100644
浏览文件 @ 469b3ad5
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include <vector>
#include "paddle/framework/lod_tensor.h"
namespace paddle {
namespace framework {
using LoDTensorArray = std::vector<LoDTensor>;
}
} // namespace paddle
paddle/framework/lod_tensor_test.cc
浏览文件 @ 469b3ad5
......@@ -144,5 +144,48 @@ TEST(LodExpand, test) {
}
}
TEST(LoD, GetFineGrainedLoDLength) {
LoD lod;
lod.push_back(std::vector<size_t>({0, 2, 4, 5}));
lod.push_back(std::vector<size_t>({0, 1, 6, 8, 10, 11}));
lod.push_back(
std::vector<size_t>({0, 2, 5, 7, 10, 12, 15, 17, 20, 24, 26, 29}));
auto lod_and_offset =
paddle::framework::GetSubLoDAndAbsoluteOffset(lod, 1, 2, 0);
LoD lod_length = lod_and_offset.first;
size_t start_offset = lod_and_offset.second.first;
size_t end_offset = lod_and_offset.second.second;
LoD expected;
expected.push_back(std::vector<size_t>{2});
expected.push_back(std::vector<size_t>{2, 2});
expected.push_back(std::vector<size_t>{2, 3, 4, 2});
EXPECT_EQ(lod_length, expected);
EXPECT_EQ(start_offset, 15UL);
EXPECT_EQ(end_offset, 26UL);
}
TEST(LoD, AppendLoD) {
LoD lod_lens;
lod_lens.push_back(std::vector<size_t>({2}));
lod_lens.push_back(std::vector<size_t>({2, 2}));
lod_lens.push_back(std::vector<size_t>({2, 3, 4, 2}));
LoD origin;
origin.push_back(std::vector<size_t>({0, 2}));
origin.push_back(std::vector<size_t>({0, 1, 6}));
origin.push_back(std::vector<size_t>({0, 2, 5, 7, 10, 12, 15}));
paddle::framework::AppendLoD(&origin, lod_lens);
LoD expected;
expected.push_back(std::vector<size_t>({0, 2, 4}));
expected.push_back(std::vector<size_t>({0, 1, 6, 8, 10}));
expected.push_back(
std::vector<size_t>({0, 2, 5, 7, 10, 12, 15, 17, 20, 24, 26}));
EXPECT_EQ(origin, expected);
}
} // namespace framework
} // namespace paddle
paddle/framework/op_desc.cc
浏览文件 @ 469b3ad5
......@@ -67,8 +67,11 @@ class CompileTimeInferShapeContext : public InferShapeContext {
out);
in_var->SetLoDLevel(out_var->GetLodLevel());
}
bool IsRuntime() const override;
protected:
VarDesc::VarType GetVarType(const std::string &name) const override;
private:
DDim GetDim(const std::string &name) const override;
void SetDim(const std::string &name, const DDim &dim) override;
......@@ -327,6 +330,19 @@ void OpDescBind::InferShape(const BlockDescBind &block) const {
PADDLE_ENFORCE(static_cast<bool>(infer_shape),
"%s's infer_shape has not been registered", this->Type());
CompileTimeInferShapeContext ctx(*this, block);
if (VLOG_IS_ON(10)) {
std::ostringstream sout;
auto inames = this->InputArgumentNames();
sout << " From [";
std::copy(inames.begin(), inames.end(),
std::ostream_iterator<std::string>(sout, ", "));
sout << "] to [";
auto onames = this->OutputArgumentNames();
std::copy(onames.begin(), onames.end(),
std::ostream_iterator<std::string>(sout, ", "));
sout << "]";
VLOG(10) << sout.str();
}
infer_shape(&ctx);
}
......@@ -336,9 +352,13 @@ void OpDescBind::InferVarType(BlockDescBind *block) const {
info.infer_var_type_(*this, block);
} else {
// all output type is LoDTensor by default
VLOG(10) << this->Type()
<< " has not registered InferVarType. Set output variables to "
"LOD_TENSOR";
for (auto &out_pair : this->outputs_) {
for (auto &out_var_name : out_pair.second) {
block->Var(out_var_name)->SetType(VarDesc::LOD_TENSOR);
block->FindRecursiveOrCreateVar(out_var_name)
->SetType(VarDesc::LOD_TENSOR);
}
}
}
......@@ -435,6 +455,12 @@ void CompileTimeInferShapeContext::SetDim(const std::string &name,
const DDim &dim) {
block_.FindVarRecursive(name)->SetShape(framework::vectorize(dim));
}
bool CompileTimeInferShapeContext::IsRuntime() const { return false; }
VarDesc::VarType CompileTimeInferShapeContext::GetVarType(
const std::string &name) const {
return block_.FindVarRecursive(name)->GetType();
}
} // namespace framework
} // namespace paddle
paddle/framework/op_registry.h
浏览文件 @ 469b3ad5
......@@ -92,8 +92,7 @@ struct OpKernelRegistrarFunctor<PlaceType, false, I, KernelTypes...> {
void operator()(const char* op_type) const {
using T = typename KERNEL_TYPE::ELEMENT_TYPE;
OperatorWithKernel::OpKernelKey key(ToDataType(std::type_index(typeid(T))),
PlaceType());
OpKernelType key(ToDataType(std::type_index(typeid(T))), PlaceType());
OperatorWithKernel::AllOpKernels()[op_type][key].reset(new KERNEL_TYPE);
constexpr auto size = std::tuple_size<std::tuple<KernelTypes...>>::value;
......
paddle/framework/operator.cc
浏览文件 @ 469b3ad5
......@@ -15,7 +15,9 @@ limitations under the License. */
#include "paddle/framework/operator.h"
#include <algorithm>
#include <atomic>
#include "paddle/framework/lod_tensor_array.h"
#include "paddle/framework/shape_inference.h"
#include "paddle/framework/var_type.h"
namespace paddle {
namespace framework {
......@@ -126,7 +128,7 @@ OperatorBase::OperatorBase(const std::string& type,
std::vector<std::string> OperatorBase::InputVars() const {
std::vector<std::string> ret_val;
for (auto& o : outputs_) {
for (auto& o : inputs_) {
ret_val.reserve(ret_val.size() + o.second.size());
ret_val.insert(ret_val.end(), o.second.begin(), o.second.end());
}
......@@ -252,8 +254,7 @@ std::vector<Tensor*> ExecutionContext::MultiOutput<Tensor>(
return res;
}
std::ostream& operator<<(std::ostream& os,
const OperatorWithKernel::OpKernelKey& kernel_key) {
std::ostream& operator<<(std::ostream& os, const OpKernelType& kernel_key) {
os << "place[" << kernel_key.place_ << "]:data_type[" << kernel_key.data_type_
<< "]";
return os;
......@@ -365,7 +366,9 @@ class RuntimeInferShapeContext : public InferShapeContext {
out_tensor->set_lod(in_tensor.lod());
}
private:
bool IsRuntime() const override { return true; }
protected:
DDim GetDim(const std::string& name) const override {
Variable* var = scope_.FindVar(name);
if (var->IsType<LoDTensor>()) {
......@@ -388,13 +391,31 @@ class RuntimeInferShapeContext : public InferShapeContext {
}
}
VarDesc::VarType GetVarType(const std::string& name) const override {
auto* var = scope_.FindVar(name);
return ToVarType(var->Type());
}
private:
const OperatorBase& op_;
const Scope& scope_;
};
void OperatorWithKernel::Run(const Scope& scope,
const platform::DeviceContext& dev_ctx) const {
VLOG(3) << "Running operator " << this->Type();
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<std::string> 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);
......@@ -410,7 +431,7 @@ void OperatorWithKernel::Run(const Scope& scope,
// check if op[type] have kernel for kernel_key
OpKernelMap& kernels = kernels_iter->second;
auto kernel_key = OpKernelKey(IndicateDataType(ctx), dev_ctx);
auto kernel_key = GetKernelType(ctx);
auto kernel_iter = kernels.find(kernel_key);
if (kernel_iter == kernels.end()) {
......@@ -418,6 +439,41 @@ void OperatorWithKernel::Run(const Scope& scope,
}
kernel_iter->second->Compute(ctx);
// throws errors if have.
dev_ctx.Finish();
}
OpKernelType OperatorWithKernel::GetKernelType(
const ExecutionContext& ctx) const {
return OpKernelType(IndicateDataType(ctx), ctx.device_context());
}
DataType OperatorWithKernel::IndicateDataType(
const ExecutionContext& ctx) const {
auto& scope = ctx.scope();
int data_type = -1;
for (auto& input : this->inputs_) {
for (auto& ipt_name : input.second) {
auto* var = scope.FindVar(ipt_name);
if (var != nullptr) {
const Tensor* t = nullptr;
if (var->IsType<Tensor>()) {
t = &var->Get<Tensor>();
} else if (var->IsType<LoDTensor>()) {
t = &var->Get<LoDTensor>();
} else if (var->IsType<SelectedRows>()) {
t = &(var->Get<SelectedRows>().value());
}
if (t != nullptr) {
int tmp = static_cast<int>(ToDataType(t->type()));
PADDLE_ENFORCE(tmp == data_type || data_type == -1,
"DataType of Paddle Op %s must be the same.", Type());
data_type = tmp;
}
}
}
}
PADDLE_ENFORCE(data_type != -1, "DataType should be indicated by input");
return static_cast<DataType>(data_type);
}
} // namespace framework
......
paddle/framework/operator.h
浏览文件 @ 469b3ad5
......@@ -298,11 +298,10 @@ class ExecutionContext {
}
#ifdef PADDLE_WITH_CUDA
const platform::CUDADeviceContext& cuda_device_context() const {
const inline platform::CUDADeviceContext& cuda_device_context() const {
PADDLE_ENFORCE(platform::is_gpu_place(device_context_.GetPlace()));
auto cuda_ctx =
reinterpret_cast<const platform::CUDADeviceContext*>(&device_context_);
return *cuda_ctx;
return *reinterpret_cast<const platform::CUDADeviceContext*>(
&device_context_);
}
#endif
......@@ -346,27 +345,10 @@ class OpKernel : public OpKernelBase {
using ELEMENT_TYPE = T;
};
class OperatorWithKernel : public OperatorBase {
public:
struct OpKernelKey {
platform::Place place_;
DataType data_type_;
OpKernelKey(DataType data_type, platform::Place place)
: place_(place), data_type_(data_type) {}
OpKernelKey(DataType data_type, const platform::DeviceContext& dev_ctx)
: place_(dev_ctx.GetPlace()), data_type_(data_type) {}
bool operator==(const OpKernelKey& o) const {
return platform::places_are_same_class(place_, o.place_) &&
data_type_ == o.data_type_;
}
};
struct OpKernelHash {
struct OpKernelType {
struct Hash {
std::hash<int> hash_;
size_t operator()(const OpKernelKey& key) const {
size_t operator()(const OpKernelType& key) const {
int place = key.place_.which();
int data_type = static_cast<int>(key.data_type_);
int pre_hash = data_type << NUM_PLACE_TYPE_LIMIT_IN_BIT |
......@@ -375,9 +357,26 @@ class OperatorWithKernel : public OperatorBase {
}
};
platform::Place place_;
DataType data_type_;
OpKernelType(DataType data_type, platform::Place place)
: place_(place), data_type_(data_type) {}
OpKernelType(DataType data_type, const platform::DeviceContext& dev_ctx)
: place_(dev_ctx.GetPlace()), data_type_(data_type) {}
bool operator==(const OpKernelType& o) const {
return platform::places_are_same_class(place_, o.place_) &&
data_type_ == o.data_type_;
}
};
class OperatorWithKernel : public OperatorBase {
public:
using OpKernelMap =
std::unordered_map<OpKernelKey, std::unique_ptr<OpKernelBase>,
OpKernelHash>;
std::unordered_map<OpKernelType, std::unique_ptr<OpKernelBase>,
OpKernelType::Hash>;
OperatorWithKernel(const std::string& type, const VariableNameMap& inputs,
const VariableNameMap& outputs, const AttributeMap& attrs)
......@@ -405,42 +404,15 @@ class OperatorWithKernel : public OperatorBase {
}
protected:
virtual OpKernelType GetKernelType(const ExecutionContext& ctx) const;
private:
// indicate kernel DataType by input data. Defaultly all input data must be
// same.
virtual DataType IndicateDataType(const ExecutionContext& ctx) const {
VLOG(3) << "Default IndicateDataType " << this->Type();
auto& scope = ctx.scope();
int data_type = -1;
for (auto& input : this->inputs_) {
for (auto& ipt_name : input.second) {
auto* var = scope.FindVar(ipt_name);
if (var != nullptr) {
const Tensor* t = nullptr;
if (var->IsType<Tensor>()) {
t = &var->Get<Tensor>();
} else if (var->IsType<LoDTensor>()) {
t = &var->Get<LoDTensor>();
} else if (var->IsType<SelectedRows>()) {
t = &(var->Get<SelectedRows>().value());
}
if (t != nullptr) {
int tmp = static_cast<int>(ToDataType(t->type()));
VLOG(3) << "Input " << ipt_name << " with data_type " << tmp;
PADDLE_ENFORCE(tmp == data_type || data_type == -1,
"DataType of Paddle Op %s must be the same.",
Type());
data_type = tmp;
}
}
}
}
PADDLE_ENFORCE(data_type != -1, "DataType should be indicated by input");
return static_cast<DataType>(data_type);
}
DataType IndicateDataType(const ExecutionContext& ctx) const;
};
std::ostream& operator<<(std::ostream& os,
const OperatorWithKernel::OpKernelKey& kernel_key);
std::ostream& operator<<(std::ostream& os, const OpKernelType& kernel_key);
extern bool OpSupportGPU(const std::string& op_type);
......
paddle/framework/operator_test.cc
浏览文件 @ 469b3ad5
......@@ -114,8 +114,8 @@ class OpWithKernelTest : public OperatorWithKernel {
protected:
void InferShape(framework::InferShapeContext* ctx) const override {}
DataType IndicateDataType(const ExecutionContext& ctx) const override {
return DataType::FP32;
OpKernelType GetKernelType(const ExecutionContext& ctx) const override {
return OpKernelType(DataType::FP32, ctx.device_context());
}
};
......
paddle/framework/scope.cc
浏览文件 @ 469b3ad5
......@@ -47,8 +47,12 @@ Variable* Scope::Var(const std::string& name) {
return v;
}
Variable* Scope::Var() {
return Var(string::Sprintf("%p.%d", this, vars_.size()));
Variable* Scope::Var(std::string* name) {
auto var_name = string::Sprintf("%p.%d", this, vars_.size());
if (name != nullptr) {
*name = var_name;
}
return Var(var_name);
}
Variable* Scope::FindVar(const std::string& name) const {
......@@ -94,5 +98,23 @@ void Scope::DeleteScope(Scope* scope) {
delete scope;
}
void Scope::Rename(const std::string& origin_name,
const std::string& new_name) const {
auto origin_it = vars_.find(origin_name);
PADDLE_ENFORCE(origin_it != vars_.end(),
"Cannot find original variable with name %s", origin_name);
auto new_it = vars_.find(new_name);
PADDLE_ENFORCE(new_it == vars_.end(),
"The variable with name %s is already in the scope", new_name);
vars_[new_name] = origin_it->second;
vars_.erase(origin_it);
}
std::string Scope::Rename(const std::string& origin_name) const {
auto var_name = string::Sprintf("%p.%d", this, vars_.size());
Rename(origin_name, var_name);
return var_name;
}
} // namespace framework
} // namespace paddle
paddle/framework/scope.h
浏览文件 @ 469b3ad5
......@@ -49,7 +49,7 @@ class Scope {
Variable* Var(const std::string& name);
/// Create a variable with a scope-unique name.
Variable* Var();
Variable* Var(std::string* name = nullptr);
/// Find a variable in the scope or any of its ancestors. Returns
/// nullptr if cannot find.
......@@ -68,11 +68,18 @@ class Scope {
// enumerate all the variables current contains.
std::vector<std::string> GetAllNames(bool recursive = false) const;
// Rename variable to a new name
void Rename(const std::string& origin_name,
const std::string& new_name) const;
// Rename variable to a new name and return the new name
std::string Rename(const std::string& origin_name) const;
private:
// Call Scope::NewScope for a sub-scope.
explicit Scope(Scope const* parent) : parent_(parent) {}
std::unordered_map<std::string, Variable*> vars_;
mutable std::unordered_map<std::string, Variable*> vars_;
mutable std::list<Scope*> kids_;
Scope const* parent_{nullptr};
......
paddle/framework/shape_inference.cc
浏览文件 @ 469b3ad5
......@@ -46,6 +46,23 @@ void InferShapeContext::SetDims(const std::vector<std::string> &names,
SetDim(names[i], dims[i]);
}
}
std::vector<VarDesc::VarType> InferShapeContext::GetInputsVarType(
const std::string &name) const {
return GetVarTypes(Inputs(name));
}
std::vector<VarDesc::VarType> InferShapeContext::GetOutputsVarType(
const std::string &name) const {
return GetVarTypes(Outputs(name));
}
std::vector<VarDesc::VarType> InferShapeContext::GetVarTypes(
const std::vector<std::string> &names) const {
std::vector<VarDesc::VarType> retv;
retv.resize(names.size());
std::transform(names.begin(), names.end(), retv.begin(),
std::bind(std::mem_fn(&InferShapeContext::GetVarType), this,
std::placeholders::_1));
return retv;
}
} // namespace framework
} // namespace paddle
paddle/framework/shape_inference.h
浏览文件 @ 469b3ad5
......@@ -16,6 +16,7 @@ limitations under the License. */
#include "paddle/framework/attribute.h"
#include "paddle/framework/ddim.h"
#include "paddle/framework/framework.pb.h"
namespace paddle {
namespace framework {
......@@ -26,6 +27,10 @@ class InferShapeContext {
virtual bool HasInput(const std::string &name) const = 0;
virtual bool HasOutput(const std::string &name) const = 0;
std::vector<VarDesc::VarType> GetInputsVarType(const std::string &name) const;
std::vector<VarDesc::VarType> GetOutputsVarType(
const std::string &name) const;
virtual bool HasInputs(const std::string &name) const = 0;
virtual bool HasOutputs(const std::string &name) const = 0;
......@@ -46,6 +51,8 @@ class InferShapeContext {
virtual void ShareLoD(const std::string &in, const std::string &out,
size_t i = 0, size_t j = 0) const = 0;
virtual bool IsRuntime() const = 0;
protected:
virtual framework::DDim GetDim(const std::string &name) const = 0;
virtual void SetDim(const std::string &name, const framework::DDim &dim) = 0;
......@@ -55,6 +62,11 @@ class InferShapeContext {
void SetDims(const std::vector<std::string> &names,
const std::vector<framework::DDim> &dims);
std::vector<VarDesc::VarType> GetVarTypes(
const std::vector<std::string> &names) const;
virtual VarDesc::VarType GetVarType(const std::string &name) const = 0;
};
} // namespace framework
......
paddle/framework/tensor.h
浏览文件 @ 469b3ad5
......@@ -125,7 +125,7 @@ class Tensor {
* @param[in] end_idx The index of the end row(exclusive) to slice.
* The index number begins from 0.
*/
inline Tensor Slice(const int& begin_idx, const int& end_idx) const;
inline Tensor Slice(int begin_idx, int end_idx) const;
platform::Place place() const {
PADDLE_ENFORCE_NOT_NULL(
......
paddle/framework/tensor_impl.h
浏览文件 @ 469b3ad5
......@@ -52,7 +52,7 @@ struct SizeOfTypeFunctor<HEAD, TAIL...> {
};
static inline size_t SizeOfType(std::type_index type) {
SizeOfTypeFunctor<int, float, double, int16_t, int64_t> functor;
SizeOfTypeFunctor<int, float, double, int16_t, int64_t, bool> functor;
size_t size = functor(type);
PADDLE_ENFORCE(size != 0UL, "Cannot get size of type %s", type.name());
return size;
......@@ -228,7 +228,7 @@ inline void Tensor::CopyFromVector(const std::vector<T>& src,
#endif
}
inline Tensor Tensor::Slice(const int& begin_idx, const int& end_idx) const {
inline Tensor Tensor::Slice(int begin_idx, int end_idx) const {
check_memory_size();
PADDLE_ENFORCE_GE(begin_idx, 0,
"The start row index must be greater than 0.");
......
paddle/framework/type_defs.h
浏览文件 @ 469b3ad5
......@@ -29,6 +29,7 @@ class OpDescBind;
class BlockDescBind;
class BlockDesc;
class InferShapeContext;
class BlockDescBind;
using VariableNameMap = std::map<std::string, std::vector<std::string>>;
......@@ -46,7 +47,8 @@ using OpCreator = std::function<OperatorBase*(
using GradOpMakerFN = std::function<std::vector<std::unique_ptr<OpDescBind>>(
const OpDescBind&, const std::unordered_set<std::string>& /*no_grad_set*/,
std::unordered_map<std::string, std::string>* /*grad_to_var*/)>;
std::unordered_map<std::string, std::string>* /*grad_to_var*/,
const std::vector<BlockDescBind*>& grad_block)>;
using InferVarTypeFN = std::function<void(const OpDescBind& /*op_desc*/,
BlockDescBind* /*block*/)>;
......
paddle/framework/var_desc.cc
浏览文件 @ 469b3ad5
......@@ -37,13 +37,29 @@ std::vector<int64_t> VarDescBind::Shape() const {
DataType VarDescBind::GetDataType() const { return tensor_desc().data_type(); }
void VarDescBind::SetLoDLevel(int32_t lod_level) {
PADDLE_ENFORCE(desc_.type() == VarDesc::LOD_TENSOR);
desc_.mutable_lod_tensor()->set_lod_level(lod_level);
switch (desc_.type()) {
case VarDesc::LOD_TENSOR:
desc_.mutable_lod_tensor()->set_lod_level(lod_level);
break;
case VarDesc::LOD_TENSOR_ARRAY:
desc_.mutable_tensor_array()->set_lod_level(lod_level);
break;
default:
PADDLE_THROW("Tensor type=%d does not support LoDLevel",
desc_.tensor_array().lod_level());
}
}
int32_t VarDescBind::GetLodLevel() const {
PADDLE_ENFORCE(desc_.type() == VarDesc::LOD_TENSOR);
return desc_.lod_tensor().lod_level();
switch (desc_.type()) {
case VarDesc::LOD_TENSOR:
return desc_.lod_tensor().lod_level();
case VarDesc::LOD_TENSOR_ARRAY:
return desc_.tensor_array().lod_level();
default:
PADDLE_THROW("Tensor type=%d does not support LoDLevel",
desc_.tensor_array().lod_level());
}
}
const TensorDesc &VarDescBind::tensor_desc() const {
......@@ -53,6 +69,8 @@ const TensorDesc &VarDescBind::tensor_desc() const {
return desc_.selected_rows();
case VarDesc::LOD_TENSOR:
return desc_.lod_tensor().tensor();
case VarDesc::LOD_TENSOR_ARRAY:
return desc_.tensor_array().tensor();
default:
PADDLE_THROW("Unexpected branch.");
}
......@@ -66,6 +84,8 @@ TensorDesc *VarDescBind::mutable_tensor_desc() {
return desc_.mutable_selected_rows();
case VarDesc::LOD_TENSOR:
return desc_.mutable_lod_tensor()->mutable_tensor();
case VarDesc::LOD_TENSOR_ARRAY:
return desc_.mutable_tensor_array()->mutable_tensor();
default:
PADDLE_THROW("Unexpected branch.");
}
......
paddle/framework/var_desc.h
浏览文件 @ 469b3ad5
......@@ -15,6 +15,7 @@ limitations under the License. */
#pragma once
#include <vector>
#include "glog/logging.h"
#include "paddle/framework/framework.pb.h"
namespace paddle {
......
paddle/framework/var_type.h 0 → 100644
浏览文件 @ 469b3ad5
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include "paddle/framework/framework.pb.h"
#include "paddle/framework/lod_rank_table.h"
#include "paddle/framework/lod_tensor.h"
#include "paddle/framework/lod_tensor_array.h"
namespace paddle {
namespace framework {
inline VarDesc::VarType ToVarType(std::type_index type) {
if (type.hash_code() == typeid(LoDTensor).hash_code()) {
return VarDesc_VarType_LOD_TENSOR;
} else if (type.hash_code() == typeid(LoDRankTable).hash_code()) {
return VarDesc_VarType_LOD_RANK_TABLE;
} else if (type.hash_code() == typeid(LoDTensorArray).hash_code()) {
return VarDesc_VarType_LOD_TENSOR_ARRAY;
} else if (type.hash_code() == typeid(SelectedRows).hash_code()) {
return VarDesc_VarType_SELECTED_ROWS;
} else {
PADDLE_THROW("ToVarType:Unsupported type %s", type.name());
}
}
template <typename Visitor>
inline void VisitVarType(const Variable& var, Visitor visitor) {
switch (ToVarType(var.Type())) {
case VarDesc_VarType_LOD_TENSOR:
visitor(var.Get<framework::LoDTensor>());
return;
case VarDesc_VarType_LOD_RANK_TABLE:
visitor(var.Get<LoDRankTable>());
return;
case VarDesc_VarType_LOD_TENSOR_ARRAY:
visitor(var.Get<LoDTensorArray>());
return;
case VarDesc_VarType_SELECTED_ROWS:
visitor(var.Get<SelectedRows>());
return;
default:
PADDLE_THROW("Not supported visit type, %d", ToVarType(var.Type()));
}
}
} // namespace framework
} // namespace paddle
paddle/framework/variable.h
浏览文件 @ 469b3ad5
......@@ -48,6 +48,11 @@ class Variable {
void Clear() { holder_.reset(); }
std::type_index Type() const {
PADDLE_ENFORCE(holder_ != nullptr, "Must hold memory");
return holder_->Type();
}
private:
struct Placeholder {
virtual ~Placeholder() {}
......
paddle/function/CMakeLists.txt
浏览文件 @ 469b3ad5
......@@ -45,6 +45,7 @@ if(WITH_GPU)
add_simple_unittest(BlockExpandOpTest)
add_simple_unittest(CropOpTest)
add_simple_unittest(SwitchOpTest)
add_simple_unittest(ScaleSubRegionOpTest)
endif()
add_simple_unittest(Im2ColTest)
......
paddle/function/FunctionTest.h
浏览文件 @ 469b3ad5
......@@ -110,6 +110,7 @@ public:
function2_(FunctionBase::funcRegistrar_.createByType(name2)) {
function1_->init(config);
function2_->init(config);
initArgsCallback_ = nullptr;
}
~Compare2Function() {}
......@@ -170,6 +171,10 @@ public:
*seq2_));
}
void registerInitCallback(std::function<void(BufferArg&, size_t)> callback) {
initArgsCallback_ = callback;
}
// output need only contains shape, do not contains data.
void addOutputs(const BufferArg& output, ArgType argType = ASSIGN_TO) {
size_t size =
......@@ -340,6 +345,10 @@ protected:
initArg(*func1Inputs_[i]);
}
if (initArgsCallback_ != nullptr) {
initArgsCallback_(*func1Inputs_[i], i);
}
copyArg_(*func1Inputs_[i], *func2Inputs_[i]);
}
}
......@@ -386,6 +395,7 @@ protected:
std::shared_ptr<SequenceIdArg> seq1_;
std::shared_ptr<SequenceIdArg> seq2_;
test::CopyArgument<DType1, DType2> copyArg_;
std::function<void(BufferArg&, size_t)> initArgsCallback_;
};
class CpuGpuFuncCompare
......
paddle/function/ScaleSubRegionOp.cpp 0 → 100644
浏览文件 @ 469b3ad5
/* 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 "ScaleSubRegionOp.h"
#include "paddle/function/TensorShape.h"
namespace paddle {
template <>
void ScaleSubRegion<DEVICE_TYPE_CPU>(real* outputs,
const real* inputs,
const real* indices,
const TensorShape shape,
const FuncConfig& conf) {
real value = conf.get<real>("value");
int number = shape[0];
int channel = shape[1];
int height = shape[2];
int width = shape[3];
memcpy(outputs, inputs, number * channel * height * width * sizeof(real));
for (int n = 0; n < number; ++n) {
// indices start from 1
int offset = n * 6;
for (int c = indices[offset] - 1; c < indices[offset + 1]; ++c) {
for (int h = indices[offset + 2] - 1; h < indices[offset + 3]; ++h) {
for (int w = indices[offset + 4] - 1; w < indices[offset + 5]; ++w) {
int idx = ((n * channel + c) * height + h) * width + w;
outputs[idx] *= value;
}
}
}
}
}
template <>
void ScaleSubRegionGrad<DEVICE_TYPE_CPU>(const real* inGrad,
real* outGrad,
const real* indices,
const TensorShape shape,
const FuncConfig& conf) {
real value = conf.get<real>("value");
int number = shape[0];
int channel = shape[1];
int height = shape[2];
int width = shape[3];
for (int n = 0; n < number; ++n) {
for (int c = 0; c < channel; ++c) {
for (int h = 0; h < height; ++h) {
for (int w = 0; w < width; ++w) {
int idx = ((n * channel + c) * height + h) * width + w;
int offset = n * 6;
if (c >= (indices[offset] - 1) && c <= (indices[offset + 1] - 1) &&
h >= (indices[offset + 2] - 1) &&
h <= (indices[offset + 3] - 1) &&
w >= (indices[offset + 4] - 1) &&
w <= (indices[offset + 5] - 1)) {
outGrad[idx] += inGrad[idx] * value;
} else {
outGrad[idx] += inGrad[idx];
}
}
}
}
}
}
/**
* \brief For each instance, ScaleSubRegion can be used to multiply a value to
* a specified sub continuous region. By providing start index and end
* index for C/H/W, you can specify the location and shape of the region.
*
* Argument in this Function:
* \param inputs A 4-D tensor with shape [N, C, H, W], only one input.
* \param indices A 2-D tensor with shape [N, 6], indicates the sub region.
* \param outputs A 4-D tensor with same shape as inputs, output value.
*/
template <DeviceType Device>
class ScaleSubRegionFunc : public FunctionBase {
public:
void init(const FuncConfig& config) override { conf_ = config; }
void calc(const BufferArgs& inputs, const BufferArgs& outputs) override {
CHECK_EQ(2UL, inputs.size());
CHECK_EQ(1UL, outputs.size());
CHECK_EQ(outputs[0].getArgType(), ASSIGN_TO);
TensorShape shape = inputs[0].shape();
ScaleSubRegion<Device>(outputs[0].data<real>(),
inputs[0].data<real>(),
inputs[1].data<real>(),
shape,
conf_);
}
private:
FuncConfig conf_;
};
/**
* \brief The backward propagation of ScaleSubRegion Function.
*
* Argument in this Function:
* \param inputs A 4-D tensor with shape [N, C, H, W], output gradient.
* \param indices A 2-D tensor with shape [N, 6], indicates the sub region.
* \param outputs A 4-D tensor with shape [N, C, H, W], gradient of input value.
*/
template <DeviceType Device>
class ScaleSubRegionGradFunc : public FunctionBase {
public:
void init(const FuncConfig& config) override { conf_ = config; }
void calc(const BufferArgs& inputs, const BufferArgs& outputs) override {
CHECK_EQ(2UL, inputs.size());
CHECK_EQ(1UL, outputs.size());
CHECK_EQ(outputs[0].getArgType(), ADD_TO);
TensorShape shape = inputs[0].shape();
ScaleSubRegionGrad<Device>(inputs[0].data<real>(),
outputs[0].data<real>(),
inputs[1].data<real>(),
shape,
conf_);
}
private:
FuncConfig conf_;
};
REGISTER_TYPED_FUNC(ScaleSubRegion, CPU, ScaleSubRegionFunc);
REGISTER_TYPED_FUNC(ScaleSubRegionGrad, CPU, ScaleSubRegionGradFunc);
#ifdef PADDLE_WITH_CUDA
REGISTER_TYPED_FUNC(ScaleSubRegion, GPU, ScaleSubRegionFunc);
REGISTER_TYPED_FUNC(ScaleSubRegionGrad, GPU, ScaleSubRegionGradFunc);
#endif
} // namespace paddle
paddle/function/ScaleSubRegionOp.h 0 → 100644
浏览文件 @ 469b3ad5
/* 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 "Function.h"
namespace paddle {
/**
* \brief Function to multiply a value to values in specified sub continuous
* region. Indices must be provided to indcate the location and shape of
* the region and the multiplied value is passed by configure variable.
*
*
* \param[out] outputs Output value.
* \param[in] inputs Input data which contains NCHW information.
* \param[in] indices Indices data to indcate the sub region.
* \param[in] shape Tensor shape of input value.
* \param[in] conf Configure variable which contains the multiplied value.
*/
template <DeviceType Device>
void ScaleSubRegion(real* outputs,
const real* inputs,
const real* indices,
const TensorShape shape,
const FuncConfig& conf);
/**
* \brief Backward propagation function of ScaleSubRegion.
*
* \param[out] inGrad Gradients of previous layer.
* \param[in] outGrad Output gradient.
* \param[in] indices Indices data.
* \param[in] shape The Shape of input tensor.
* \param[in] conf Configure variable.
*/
template <DeviceType Device>
void ScaleSubRegionGrad(const real* inGrad,
real* outGrad,
const real* indices,
const TensorShape shape,
const FuncConfig& conf);
} // namespace paddle
paddle/function/ScaleSubRegionOpGpu.cu 0 → 100644
浏览文件 @ 469b3ad5
/* 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 "ScaleSubRegionOp.h"
#include "hl_base.h"
namespace paddle {
__global__ void KeScaleSubRegion(real* outputs,
const real* inputs,
const real* indices,
real value,
int channel,
int height,
int width,
int nthreads) {
const int idx = threadIdx.x + blockIdx.x * blockDim.x;
if (idx < nthreads) {
const int w = idx % width;
const int h = (idx / width) % height;
const int c = (idx / width / height) % channel;
const int n = idx / width / height / channel;
const int offset = n * 6;
if (c >= (indices[offset] - 1) && c <= (indices[offset + 1] - 1) &&
h >= (indices[offset + 2] - 1) && h <= (indices[offset + 3] - 1) &&
w >= (indices[offset + 4] - 1) && w <= (indices[offset + 5] - 1)) {
outputs[idx] = inputs[idx] * value;
} else {
outputs[idx] = inputs[idx];
}
}
}
template <>
void ScaleSubRegion<DEVICE_TYPE_GPU>(real* outputs,
const real* inputs,
const real* indices,
const TensorShape shape,
const FuncConfig& conf) {
real value = conf.get<real>("value");
int number = shape[0];
int channel = shape[1];
int height = shape[2];
int width = shape[3];
size_t nth = number * channel * height * width;
int blockSize = 1024;
int gridSize = (nth + blockSize - 1) / blockSize;
KeScaleSubRegion<<<gridSize, blockSize, 0, STREAM_DEFAULT>>>(
outputs, inputs, indices, value, channel, height, width, nth);
CHECK_SYNC("ScaleSubRegion");
}
__global__ void KeScaleSubRegionDiff(const real* inGrad,
real* outGrad,
const real* indices,
real value,
int channel,
int height,
int width,
int nthreads) {
const int idx = threadIdx.x + blockIdx.x * blockDim.x;
if (idx < nthreads) {
const int w = idx % width;
const int h = (idx / width) % height;
const int c = (idx / width / height) % channel;
const int n = idx / width / height / channel;
const int offset = n * 6;
if (c >= (indices[offset] - 1) && c <= (indices[offset + 1] - 1) &&
h >= (indices[offset + 2] - 1) && h <= (indices[offset + 3] - 1) &&
w >= (indices[offset + 4] - 1) && w <= (indices[offset + 5] - 1)) {
outGrad[idx] += inGrad[idx] * value;
} else {
outGrad[idx] += inGrad[idx];
}
}
}
template <>
void ScaleSubRegionGrad<DEVICE_TYPE_GPU>(const real* inGrad,
real* outGrad,
const real* indices,
const TensorShape shape,
const FuncConfig& conf) {
real value = conf.get<real>("value");
int number = shape[0];
int channel = shape[1];
int height = shape[2];
int width = shape[3];
size_t nth = number * channel * height * width;
int blockSize = 1024;
int gridSize = (nth + blockSize - 1) / blockSize;
KeScaleSubRegionDiff<<<gridSize, blockSize, 0, STREAM_DEFAULT>>>(
inGrad, outGrad, indices, value, channel, height, width, nth);
CHECK_SYNC("ScaleSubRegionGrad");
}
} // namespace paddle
paddle/function/ScaleSubRegionOpTest.cpp 0 → 100644
浏览文件 @ 469b3ad5
/* 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 <gtest/gtest.h>
#include "FunctionTest.h"
namespace paddle {
TEST(ScaleSubRegion, real) {
for (size_t numSamples : {5, 32}) {
for (size_t channels : {5, 32}) {
for (size_t imgSizeH : {5, 33}) {
for (size_t imgSizeW : {5, 32}) {
for (real value : {-0.5, 0.0, 0.5}) {
for (bool firstHalf : {false, true}) {
VLOG(3) << " numSamples=" << numSamples
<< " channels=" << channels << " imgSizeH=" << imgSizeH
<< " imgSizeW=" << imgSizeW;
for (bool testGrad : {false, true}) {
CpuGpuFuncCompare compare(
testGrad ? "ScaleSubRegionGrad" : "ScaleSubRegion",
FuncConfig().set<real>("value", value));
TensorShape shape{numSamples, channels, imgSizeH, imgSizeW};
TensorShape indicesShape{numSamples, 6};
compare.addInputs(BufferArg(VALUE_TYPE_FLOAT, shape));
compare.addInputs(BufferArg(VALUE_TYPE_FLOAT, indicesShape));
compare.registerInitCallback([=](BufferArg& arg, size_t index) {
if (index == 1) {
real* data = (real*)arg.data();
for (size_t i = 0; i < numSamples; ++i) {
size_t offset = i * 6;
data[offset] = firstHalf ? 1 : channels / 2;
data[offset + 1] = firstHalf ? channels / 2 : channels;
data[offset + 2] = firstHalf ? 1 : imgSizeH / 2;
data[offset + 3] = firstHalf ? imgSizeH / 2 : imgSizeH;
data[offset + 4] = firstHalf ? 1 : imgSizeW / 2;
data[offset + 5] = firstHalf ? imgSizeW / 2 : imgSizeW;
}
}
});
compare.addOutputs(
BufferArg(
VALUE_TYPE_FLOAT, shape, testGrad ? ADD_TO : ASSIGN_TO),
testGrad ? ADD_TO : ASSIGN_TO);
compare.run();
}
}
}
}
}
}
}
}
} // namespace paddle
paddle/gserver/evaluators/Evaluator.cpp
浏览文件 @ 469b3ad5
......@@ -395,14 +395,24 @@ real AucEvaluator::evalImp(std::vector<Argument>& arguments) {
CHECK_LE(arguments.size(), (size_t)3);
MatrixPtr output = arguments[0].value;
IVectorPtr label = arguments[1].ids;
MatrixPtr labelval = arguments[1].value;
bool supportWeight = (3 == arguments.size()) ? true : false;
MatrixPtr weight = supportWeight ? arguments[2].value : nullptr;
if (nullptr == output || nullptr == label ||
(supportWeight && nullptr == weight)) {
if (nullptr == output || (supportWeight && nullptr == weight)) {
return 0;
}
size_t insNum = output->getHeight();
size_t outputDim = output->getWidth();
// Copy label from value to a vector.
if (nullptr == label && nullptr != labelval) {
// label width is 1
CHECK_EQ(1U, labelval->getWidth());
VectorPtr vec =
Vector::create(labelval->getData(), insNum, output->useGpu());
label = vec->castToInt();
}
CHECK_EQ(insNum, label->getSize());
if (supportWeight) {
CHECK_EQ(insNum, weight->getHeight());
......@@ -443,6 +453,7 @@ real AucEvaluator::evalImp(std::vector<Argument>& arguments) {
int* labelD = label->getData();
real* weightD = supportWeight ? weight->getData() : nullptr;
size_t pos = realColumnIdx_;
for (size_t i = 0; i < insNum; ++i) {
real value = outputD[pos];
uint32_t binIdx = static_cast<uint32_t>(value * kBinNum_);
......
paddle/gserver/layers/ConvBaseProjection.cpp
浏览文件 @ 469b3ad5
......@@ -17,7 +17,7 @@ limitations under the License. */
namespace paddle {
ThreadLocalD<std::vector<MemoryHandle *>> ConvBaseProjection::convMem_;
ThreadLocalD<std::vector<MemoryHandlePtr>> ConvBaseProjection::convMem_;
ConvBaseProjection::ConvBaseProjection(const ProjectionConfig &config,
ParameterPtr parameter,
......@@ -175,18 +175,18 @@ void ConvBaseProjection::reshape(int batchSize) {
}
void *ConvBaseProjection::getSpaceBytes(size_t size) {
std::vector<MemoryHandle *> &convMem = *convMem_;
std::vector<MemoryHandlePtr> &convMem = *convMem_;
if (convMem.empty()) {
int numDevices = hl_get_device_count();
convMem.resize(numDevices);
}
int devId = hl_get_device();
MemoryHandle **localMem = &(convMem[devId]);
if (NULL == *localMem || size > (*localMem)->getAllocSize()) {
*localMem = new GpuMemoryHandle(size);
MemoryHandlePtr localMem = convMem[devId];
if (NULL == localMem || size > localMem->getAllocSize()) {
localMem = std::make_shared<GpuMemoryHandle>(size);
}
return (*localMem)->getBuf();
return localMem->getBuf();
}
ConvBaseProjection::~ConvBaseProjection() {
......
paddle/gserver/layers/ConvBaseProjection.h
浏览文件 @ 469b3ad5
......@@ -105,7 +105,7 @@ protected:
bool bias_;
std::unique_ptr<Weight> weight_;
static ThreadLocalD<std::vector<MemoryHandle*>> convMem_;
static ThreadLocalD<std::vector<MemoryHandlePtr>> convMem_;
};
} // namespace paddle
paddle/gserver/layers/MKLDNNAddtoLayer.cpp 0 → 100644
浏览文件 @ 469b3ad5
/* 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 "MKLDNNAddtoLayer.h"
using namespace mkldnn; // NOLINT
namespace paddle {
REGISTER_LAYER(mkldnn_addto, MKLDNNAddtoLayer);
bool MKLDNNAddtoLayer::init(const LayerMap& layerMap,
const ParameterMap& parameterMap) {
if (!MKLDNNLayer::init(layerMap, parameterMap)) {
return false;
}
layerSize_ = getSize();
for (size_t i = 0; i < inputLayers_.size(); i++) {
CHECK_EQ(layerSize_, inputLayers_[i]->getSize()) << "input size must equal";
}
if (biasParameter_.get() != NULL) {
biases_ =
std::unique_ptr<Weight>(new Weight(1, layerSize_, biasParameter_, 0));
}
return true;
}
void MKLDNNAddtoLayer::reshape(
int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) {
CHECK_EQ(layerSize_, getSize()) << "this layer size can not be changed";
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);
for (size_t i = 0; i < inputLayers_.size(); i++) {
CHECK_EQ(int64_t(bs), inputLayers_[i]->getOutput().getBatchSize());
CHECK_EQ(layerSize_, inputLayers_[i]->getSize());
}
oc = ic;
oh = ih;
ow = iw;
reshapeOutput(oh, ow);
resizeOutput(bs, oc * oh * ow);
}
void MKLDNNAddtoLayer::resetFwd(std::vector<primitive>& pipeline,
MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& wgt,
MKLDNNMatrixPtr& bias,
MKLDNNMatrixPtr& out) {
resetFwdBuffers(inVals_, bias, out);
in = inVals_[0];
std::shared_ptr<sum::primitive_desc> fwdPD;
std::shared_ptr<sum::primitive_desc> biasPD;
resetFwdPD(fwdPD, biasPD, inVals_, bias, out);
resetFwdPipeline(pipeline, fwdPD, biasPD, inVals_, bias, out);
}
void MKLDNNAddtoLayer::resetBwd(std::vector<primitive>& pipeline,
MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& wgt,
MKLDNNMatrixPtr& bias,
MKLDNNMatrixPtr& out) {
resetBwdBuffers(inGrads_, bias, out);
in = inGrads_[0];
// backward only need share output grad to input grad
for (size_t i = 0; i < inGrads_.size(); i++) {
if (inGrads_[i] != nullptr) {
inGrads_[i] = out;
inputLayers_[i]->getOutputGrad()->setData(inGrads_[i]->getData());
}
}
// backward bias
bwdBias_ = nullptr;
if (bias) {
std::vector<float> scales(bs_, 1.0);
std::vector<memory::primitive_desc> srcPDs(bs_, bias->getPrimitiveDesc());
auto biasPD = sum::primitive_desc(bias->getMemoryDesc(), scales, srcPDs);
std::vector<primitive::at> srcs;
for (size_t i = 0; i < grads_.size(); ++i) {
srcs.push_back(*(grads_[i]));
}
bwdBias_.reset(new sum(biasPD, srcs, *bias));
pipeline.push_back(*bwdBias_);
}
}
void MKLDNNAddtoLayer::updateWeights(const UpdateCallback& callback) {
if (biases_ && biases_->getWGrad()) {
biases_->getParameterPtr()->incUpdate(callback);
}
}
void MKLDNNAddtoLayer::prepareBias(MKLDNNMatrixPtr& bias,
const MatrixPtr& biasMat,
const MKLDNNMatrixPtr& out,
std::vector<MKLDNNMatrixPtr>& outs) {
auto pd = MKLDNNMatrix::createPrimitiveDesc(
{(int)layerSize_}, memory::format::x, engine_);
bias = MKLDNNMatrix::create(pd, biasMat);
outs.clear();
real* data = out->getData();
CHECK_EQ(bs_ * layerSize_, out->getElementCnt());
for (int i = 0; i < bs_; ++i) {
MatrixPtr tmp =
Matrix::create(data + i * layerSize_, 1, layerSize_, false, false);
outs.push_back(MKLDNNMatrix::create(bias->getPrimitiveDesc(), tmp));
}
}
void MKLDNNAddtoLayer::resetFwdBuffers(std::vector<MKLDNNMatrixPtr>& inputs,
MKLDNNMatrixPtr& bias,
MKLDNNMatrixPtr& out) {
inputs.resize(inputLayers_.size());
for (size_t i = 0; i < inputs.size(); i++) {
resetInValue(inputs[i], nullptr, i);
CHECK(inputs[i]);
inputs[i]->downSpatial();
}
for (size_t i = 1; i < inputs.size(); i++) {
CHECK_PRIMITIVE_DESC_EQ(inputs[i], inputs[0]->getPrimitiveDesc());
}
resetOutValue(out, inputs[0]->getPrimitiveDesc());
if (biases_ && biases_->getW()) {
prepareBias(bias, biases_->getW(), out, vals_);
} else {
bias = nullptr;
}
}
void MKLDNNAddtoLayer::resetFwdPD(std::shared_ptr<sum::primitive_desc>& pd,
std::shared_ptr<sum::primitive_desc>& biasPD,
std::vector<MKLDNNMatrixPtr>& inputs,
MKLDNNMatrixPtr bias,
MKLDNNMatrixPtr out) {
std::vector<float> scales(inputs.size(), 1.0);
std::vector<memory::primitive_desc> srcPDs;
for (size_t i = 0; i < inputs.size(); i++) {
srcPDs.push_back(inputs[i]->getPrimitiveDesc());
}
CHECK(out);
pd.reset(new sum::primitive_desc(out->getMemoryDesc(), scales, srcPDs));
CHECK_PRIMITIVE_DESC_EQ(out, pd->dst_primitive_desc());
biasPD = nullptr;
if (bias) {
std::vector<float> scales(2, 1.0);
std::vector<memory::primitive_desc> srcPDs(2, bias->getPrimitiveDesc());
biasPD.reset(
new sum::primitive_desc(bias->getMemoryDesc(), scales, srcPDs));
CHECK_PRIMITIVE_DESC_EQ(bias, biasPD->dst_primitive_desc());
}
}
void MKLDNNAddtoLayer::resetFwdPipeline(
std::vector<primitive>& pipeline,
std::shared_ptr<sum::primitive_desc>& pd,
std::shared_ptr<sum::primitive_desc>& biasPD,
std::vector<MKLDNNMatrixPtr>& inputs,
MKLDNNMatrixPtr& bias,
MKLDNNMatrixPtr& out) {
std::vector<primitive::at> srcs;
for (size_t i = 0; i < inputs.size(); i++) {
srcs.push_back(*(inputs[i]));
}
fwd_.reset(new sum(*pd, srcs, *out));
pipeline.push_back(*fwd_);
fwdBias_.clear();
if (biasPD == nullptr || bias == nullptr) {
return;
}
fwdBias_.resize(vals_.size());
for (size_t i = 0; i < vals_.size(); ++i) {
std::vector<primitive::at> srcs;
srcs.push_back(*(vals_[i]));
srcs.push_back(*bias);
fwdBias_[i].reset(new sum(*biasPD, srcs, *vals_[i]));
pipeline.push_back(*fwdBias_[i]);
}
}
void MKLDNNAddtoLayer::resetBwdBuffers(std::vector<MKLDNNMatrixPtr>& inputs,
MKLDNNMatrixPtr& bias,
MKLDNNMatrixPtr& out) {
CHECK(outVal_);
resetOutGrad(out, outVal_->getPrimitiveDesc());
CHECK(out);
inputs.resize(inputLayers_.size());
for (size_t i = 0; i < inputs.size(); i++) {
resetInGrad(inputs[i], inVal_->getPrimitiveDesc(), i);
CHECK_PRIMITIVE_DESC_EQ(inputs[i], out->getPrimitiveDesc());
}
if (biases_ && biases_->getWGrad()) {
prepareBias(bias, biases_->getWGrad(), out, grads_);
} else {
bias = nullptr;
}
}
} // namespace paddle
paddle/gserver/layers/MKLDNNAddtoLayer.h 0 → 100644
浏览文件 @ 469b3ad5
/* 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 Addto layer.
*
* The config file api is mkldnn_addto
*/
class MKLDNNAddtoLayer : public MKLDNNLayer {
protected:
std::vector<MKLDNNMatrixPtr> inVals_;
std::vector<MKLDNNMatrixPtr> inGrads_;
// layer size == ic * ih * iw == oc * oh *ow, and can not be changed
size_t layerSize_;
std::unique_ptr<Weight> biases_;
// buffers for adding bias
std::vector<MKLDNNMatrixPtr> vals_;
std::vector<MKLDNNMatrixPtr> grads_;
// primitives for adding bias
std::vector<std::shared_ptr<mkldnn::primitive>> fwdBias_;
std::shared_ptr<mkldnn::primitive> bwdBias_;
public:
explicit MKLDNNAddtoLayer(const LayerConfig& config) : MKLDNNLayer(config) {}
~MKLDNNAddtoLayer() {}
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<mkldnn::primitive>& pipeline,
MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& wgt,
MKLDNNMatrixPtr& bias,
MKLDNNMatrixPtr& out) override;
void resetBwd(std::vector<mkldnn::primitive>& pipeline,
MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& wgt,
MKLDNNMatrixPtr& bias,
MKLDNNMatrixPtr& out) override;
void updateWeights(const UpdateCallback& callback) override;
void printValueFormat() override {
for (size_t i = 0; i < inVals_.size(); ++i) {
VLOG(MKLDNN_FMTS) << i << " input: " << 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) << i << " input: " << inGrads_[i]->getFormat() << "<<<";
}
}
protected:
/**
* Forward functions: reset buffers(inputs, output, bias),
* reset primitive descriptor,
* reset pipeline.
*/
void resetFwdBuffers(std::vector<MKLDNNMatrixPtr>& inputs,
MKLDNNMatrixPtr& bias,
MKLDNNMatrixPtr& out);
void resetFwdPD(std::shared_ptr<mkldnn::sum::primitive_desc>& pd,
std::shared_ptr<mkldnn::sum::primitive_desc>& biasPD,
std::vector<MKLDNNMatrixPtr>& inputs,
MKLDNNMatrixPtr bias,
MKLDNNMatrixPtr out);
void resetFwdPipeline(std::vector<mkldnn::primitive>& pipeline,
std::shared_ptr<mkldnn::sum::primitive_desc>& pd,
std::shared_ptr<mkldnn::sum::primitive_desc>& biasPD,
std::vector<MKLDNNMatrixPtr>& inputs,
MKLDNNMatrixPtr& bias,
MKLDNNMatrixPtr& out);
/**
* Backward functions: reset buffers(inputs, output, bias)
*/
void resetBwdBuffers(std::vector<MKLDNNMatrixPtr>& inputs,
MKLDNNMatrixPtr& bias,
MKLDNNMatrixPtr& out);
/**
* prepare for bias
*/
void prepareBias(MKLDNNMatrixPtr& bias,
const MatrixPtr& biasMat,
const MKLDNNMatrixPtr& out,
std::vector<MKLDNNMatrixPtr>& outs);
};
} // namespace paddle
paddle/gserver/layers/MKLDNNBatchNormLayer.cpp
浏览文件 @ 469b3ad5
......@@ -119,13 +119,12 @@ void MKLDNNBatchNormLayer::reshape(
int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) {
reshapeInput(bs, ih, iw);
oh = ih;
ow = ow;
ow = iw;
// ic_ and oc can not be changed
CHECK_EQ(inputElemenCnt_ / bs / ih / iw, (size_t)ic)
<< "Input channel can not be changed";
reshapeOutput(oh, ow);
resizeOutput(bs, oc * oh * ow);
printSizeInfo();
}
void MKLDNNBatchNormLayer::resetFwd(std::vector<primitive>& pipeline,
......
paddle/gserver/layers/MKLDNNConvLayer.cpp
浏览文件 @ 469b3ad5
......@@ -102,8 +102,6 @@ void MKLDNNConvLayer::reshape(
reshapeOutput(oh, ow);
resizeOutput(bs, oc * oh * ow);
printSizeInfo();
}
void MKLDNNConvLayer::resetFwd(std::vector<primitive>& pipeline,
......
paddle/gserver/layers/MKLDNNLayer.h
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/gserver/layers/ROIPoolLayer.cpp 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/gserver/layers/ROIPoolLayer.h 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/gserver/layers/ScaleSubRegionLayer.cpp 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/gserver/tests/MKLDNNTester.h
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/gserver/tests/mkldnn_branches_fc.conf 已删除 100644 → 0
浏览文件 @ 83818cd4
此差异已折叠。
paddle/math/MKLDNNMatrix.cpp
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/math/MKLDNNMatrix.h
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/math/MathFunctions.cpp
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/math/MathFunctions.h
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/math/Vector.cpp
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/math/Vector.h
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/math/tests/TensorCheck.h
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/CMakeLists.txt
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/accuracy_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/accuracy_op.cu
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/accuracy_op.h
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/activation_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/activation_op.h
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/adadelta_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/adagrad_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/adam_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/adamax_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/array_operator.h 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/array_to_lod_tensor_op.cc 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/assign_op.cc 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/auc_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/batch_norm_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/beam_search_decode_op.cc 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/beam_search_decode_op.h 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/beam_search_decode_op_test.cc 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/bilinear_tensor_product_op.cc 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/bilinear_tensor_product_op.h 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/cast_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/chunk_eval_op.cc 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/chunk_eval_op.h 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/clip_by_norm_op.cc 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/clip_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/compare_op.cc 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/compare_op.cu 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/compare_op.h 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/concat_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/cond_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/conditional_block_op.cc 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/conv2d_transpose_cudnn_op.cc 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/conv2d_transpose_cudnn_op.cu 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/conv_cudnn_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/conv_cudnn_op.cu
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/conv_shift_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/conv_shift_op.cu
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/cos_sim_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/crf_decoding_op.cc 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/crf_decoding_op.h 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/crop_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/dropout_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/elementwise_op.h
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/expand_op.cc 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/expand_op.h 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/feed_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/fetch_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/gather_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/gru_op.cc 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/gru_op.cu 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/gru_op.h 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/gru_unit_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/huber_loss_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/increment_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/l1_norm_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/load_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/lod_array_length_op.cc 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/lod_rank_table_op.cc 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/lod_reset_op.cc 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/lod_reset_op.cu 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/lod_reset_op.h 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/lod_tensor_to_array_op.cc 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/lookup_table_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/lookup_table_op.cu
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/lookup_table_op.h
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/lrn_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/lstm_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/lstm_op.h
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/lstm_unit_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/lstm_unit_op.cu
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/lstm_unit_op.h
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/math/detail/activation_functions.h 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/math/detail/gru_cpu_kernel.h 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/math/detail/gru_gpu_kernel.h 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/math/detail/gru_kernel.h 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/math/gru_compute.cc 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/math/gru_compute.cu 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/math/gru_compute.h 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/math/sequence_pooling.cc 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/math/sequence_pooling.cu 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/matmul_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/matmul_op.h
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/mean_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/merge_lod_tensor_op.cc 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/minus_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/momentum_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/momentum_op.h
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/mul_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/mul_op.cu
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/mul_op.h
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/multiplex_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/multiplex_op.cu
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/name_convention.md
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/nccl_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/nccl_op.cu
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/nccl_op_test.cu
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/pad_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/pool_cudnn_op.cu
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/pool_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/pool_op.h
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/positive_negative_pair_op.cc 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/positive_negative_pair_op.h 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/precision_recall_op.cc 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/precision_recall_op.h 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/prelu_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/proximal_gd_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/rank_loss_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/recurrent_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/recurrent_op.h 已删除 100644 → 0
浏览文件 @ 83818cd4
此差异已折叠。
paddle/operators/reduce_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/reshape_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/rmsprop_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/rnn_memory_helper_op.cc 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/save_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/scale_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/scatter_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/seq_expand_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/seq_expand_op.h
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/sequence_pool_op.h
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/sgd_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/shrink_rnn_memory_op.cc 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/sign_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/softmax_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/softmax_op.cu
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/softmax_op.h
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/split_lod_tensor_op.cc 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/split_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/sum_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/sum_op.h
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/tensor_array_read_write_op.cc 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/top_k_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/transpose_op.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/operators/while_op.cc 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/optimizer/CMakeLists.txt
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/optimizer/adam_optimizer.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/optimizer/adam_optimizer.h
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/optimizer/optimizer.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/optimizer/optimizer.h
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/optimizer/sgd_optimizer.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/optimizer/sgd_optimizer.h
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/platform/call_once.h 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/platform/device_context.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/platform/device_context.h
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/platform/dynload/nccl.h
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/platform/transform.h
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/pybind/protobuf.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/pybind/pybind.cc
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/pybind/tensor_py.h
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/scripts/docker/build.sh
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/scripts/travis/build_doc.sh
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/trainer/tests/CMakeLists.txt
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/utils/Excepts.h
浏览文件 @ 469b3ad5
此差异已折叠。
paddle/utils/arch/osx/Excepts.cpp
浏览文件 @ 469b3ad5
此差异已折叠。
proto/ModelConfig.proto
浏览文件 @ 469b3ad5
此差异已折叠。
python/CMakeLists.txt
浏览文件 @ 469b3ad5
此差异已折叠。
python/paddle/utils/merge_model.py
浏览文件 @ 469b3ad5
此差异已折叠。
python/paddle/v2/framework/evaluator.py 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
python/paddle/v2/framework/io.py
浏览文件 @ 469b3ad5
此差异已折叠。
python/paddle/v2/framework/net_drawer.py 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
python/paddle/v2/framework/nets.py
浏览文件 @ 469b3ad5
此差异已折叠。
python/paddle/v2/framework/tests/book/CMakeLists.txt 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
python/paddle/v2/framework/tests/test_assign_op.py 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
python/paddle/v2/framework/tests/test_chunk_eval_op.py 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
python/paddle/v2/framework/tests/test_compare_op.py 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
python/paddle/v2/framework/tests/test_conv3d_op.py 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
python/paddle/v2/framework/tests/test_evaluator.py 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
python/paddle/v2/framework/tests/test_expand_op.py 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
python/paddle/v2/framework/tests/test_gru_op.py 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
python/paddle/v2/framework/tests/test_lod_rank_table.py 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
python/paddle/v2/framework/tests/test_lod_reset_op.py 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
python/paddle/v2/framework/tests/test_while_op.py 0 → 100644
浏览文件 @ 469b3ad5
此差异已折叠。
python/paddle/v2/image.py
浏览文件 @ 469b3ad5
此差异已折叠。
python/paddle/v2/optimizer.py
浏览文件 @ 469b3ad5
此差异已折叠。
python/paddle/v2/plot/plot.py
浏览文件 @ 469b3ad5
此差异已折叠。
python/requirements.txt
浏览文件 @ 469b3ad5
此差异已折叠。
python/setup.py.in
浏览文件 @ 469b3ad5
此差异已折叠。
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或