Skip to content

P
Paddle

PaddlePaddle / Paddle
9 个月 前同步成功

通知 2282
Star 20931
Fork 5422
P
Paddle

前往新版Gitcode，体验更适合开发者的 AI 搜索 >>

提交 fb3bfc3d 编写于 作者: X Xin Pan
浏览文件
操作

Merge remote-tracking branch 'upstream/release/0.13.0' into new_master

上级 652332cd 9d40eb3d
展开全部 隐藏空白更改
内联 并排
Showing with 21370 addition and 9708 deletion

要显示的变更太多。

To preserve performance only 1000 of 1000+ files are displayed.
文本差异 电子邮件补丁
.clang-format
浏览文件 @ fb3bfc3d
......@@ -19,7 +19,7 @@ BasedOnStyle: Google
IndentWidth: 2
TabWidth: 2
ContinuationIndentWidth: 4
AccessModifierOffset: -2 # The private/protected/public has no indent in class
AccessModifierOffset: -1 # The private/protected/public has no indent in class
Standard: Cpp11
AllowAllParametersOfDeclarationOnNextLine: true
BinPackParameters: false
......
.copyright.hook
浏览文件 @ fb3bfc3d
......@@ -9,7 +9,7 @@ import subprocess
import platform
COPYRIGHT = '''
Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
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.
......
.pre-commit-config.yaml
浏览文件 @ fb3bfc3d
......@@ -34,6 +34,14 @@ repos:
entry: bash ./tools/codestyle/cpplint_pre_commit.hook
language: system
files: \.(c|cc|cxx|cpp|cu|h|hpp|hxx)$
- repo: local
hooks:
- id: pylint-doc-string
name: pylint
description: Check python docstring style using docstring_checker.
entry: bash ./tools/codestyle/pylint_pre_commit.hook
language: system
files: \.(py)$
- repo: https://github.com/PaddlePaddle/pre-commit-golang
sha: 8337620115c25ff8333f1b1a493bd031049bd7c0
hooks:
......
.travis.yml
浏览文件 @ fb3bfc3d
......@@ -12,45 +12,24 @@ services:
os:
- linux
env:
- JOB=build_doc
- JOB=doc
- JOB=check_style
- JOB=build_android
addons:
apt:
packages:
- gcc-4.8
- g++-4.8
- git
- build-essential
- python
- python-pip
- python2.7-dev
- python-wheel
- libboost-dev
- curl
- swig
- graphviz
- clang-format-3.8
- automake
- libtool
- ccache
ssh_known_hosts: 13.229.163.131
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
# protobuf version.
- sudo pip install -r $TRAVIS_BUILD_DIR/python/requirements.txt
- sudo pip install wheel sphinx==1.5.6 recommonmark sphinx-rtd-theme==0.1.9 virtualenv pre-commit LinkChecker
# For pylint dockstring checker
- sudo pip install pylint pytest astroid isort
- |
function timeout() { perl -e 'alarm shift; exec @ARGV' "$@"; }
script:
- |
# 43min timeout
if [[ "$JOB" == "build_android" ]]; then timeout 2580 docker run -it --rm -v "$TRAVIS_BUILD_DIR:/paddle" paddlepaddle/paddle:latest-dev-android;
else timeout 2580 paddle/scripts/travis/${JOB}.sh; fi;
RESULT=$?; if [ $RESULT -eq 0 ] || [ $RESULT -eq 142 ]; then true; else exit 1; fi;
paddle/scripts/paddle_docker_build.sh ${JOB}
if [ $? -eq 0 ] || [ $? -eq 142 ]; then true; else exit 1; fi;
- |
if [[ "$JOB" != "build_doc" ]]; then exit 0; fi;
if [[ "$JOB" != "doc" ]]; then exit 0; fi;
# For document only
if [[ "$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
......
AUTHORS.md
浏览文件 @ fb3bfc3d
......@@ -2,12 +2,14 @@
|---|---|
| abhinavarora | Abhinav Arora |
| backyes | Yan-Fei Wang |
| baiyfbupt | Yi-Fan Bai |
| beckett1124 | Bin Qi |
| JiayiFeng | Jia-Yi Feng |
| chengxiaohua1105 | Xiao-Hua Cheng |
| cxwangyi, yiwangbaidu, wangkuiyi | Yi Wang |
| cxysteven | Xing-Yi Cheng |
| dzhwinter | Zhi-Hong Dong |
| dragonwarrior | Long Wang |
| dyning | Yuning Du |
| emailweixu | Wei Xu |
| gangliao | Gang Liao |
| gongweibao | Wei-Bao Gong |
......@@ -16,6 +18,9 @@
| hedaoyuan | Dao-Yuan He |
| helinwang | He-Lin Wang |
| jacquesqiao | Long-Fei Qiao |
| jczaja | Jacek Czaja |
| JiayiFeng | Jia-Yi Feng |
| kbinias | Krzysztof Binias |
| kuke | Yi-Bing Liu |
| lcy-seso | Ying Cao |
| lipeng-unisound | Peng Li |
......@@ -24,15 +29,20 @@
| llxxxll | Yong-Feng Liu |
| luotao01 | Tao Luo |
| lzhao4ever | Liang Zhao |
| mozga-intel | Mateusz Ozga |
| NHZlX | Zhao-Long Xing |
| Noplz | Yuan Gao |
| pakchoi | Chuan-Jiang Song |
| panyx0718 | Xin Pan |
| pengli09 | Peng Li |
| pkuyym | Ya-Ming Yang |
| pzelazko-intel | Pawel Zelazko |
| QiJune | Jun Qi |
| qingqing01 | Qing-Qing Dang |
| reyoung | Yang Yu |
| Superjom | Chun-Wei Yan |
| tianbingsz | Tian-Bing Xu |
| tpatejko | Tomasz Patejko |
| typhoonzero | Yi Wu |
| wanghaoshuang | Hao-Shuang Wang |
| wangyang59 | Yang Wang |
......
CMakeLists.txt
浏览文件 @ fb3bfc3d
......@@ -25,7 +25,6 @@ message(STATUS "CXX compiler: ${CMAKE_CXX_COMPILER}, version: "
message(STATUS "C compiler: ${CMAKE_C_COMPILER}, version: "
"${CMAKE_C_COMPILER_ID} ${CMAKE_C_COMPILER_VERSION}")
find_package(Sphinx)
if(NOT CMAKE_CROSSCOMPILING)
find_package(CUDA QUIET)
endif(NOT CMAKE_CROSSCOMPILING)
......@@ -42,7 +41,6 @@ option(WITH_MKL "Compile PaddlePaddle with MKL support." ${AVX_FO
option(WITH_DSO "Compile PaddlePaddle with dynamic linked CUDA" ON)
option(WITH_TESTING "Compile PaddlePaddle with unit testing" OFF)
option(WITH_SWIG_PY "Compile PaddlePaddle with inference api" ON)
option(WITH_STYLE_CHECK "Compile PaddlePaddle with style check" ON)
option(WITH_PYTHON "Compile PaddlePaddle with python interpreter" ON)
option(WITH_DOUBLE "Compile PaddlePaddle with double precision" OFF)
option(WITH_RDMA "Compile PaddlePaddle with RDMA support" OFF)
......@@ -59,8 +57,10 @@ option(GLIDE_INSTALL "Download and install go dependencies " ON)
option(USE_NNPACK "Compile PaddlePaddle with NNPACK library" OFF)
option(WITH_DISTRIBUTE "Compile with grpc distributed support" OFF)
option(USE_EIGEN_FOR_BLAS "Use matrix multiplication in Eigen" OFF)
option(EIGEN_USE_THREADS "Compile with multi-threaded Eigen" OFF)
option(WITH_ARM_FP16 "Use half precision support on armv8.2-a cpu" OFF)
option(WITH_FAST_BUNDLE_TEST "Bundle tests that can be run in a single process together to reduce launch overhead" OFF)
option(WITH_CONTRIB "Compile the third-party contributation" OFF)
# CMAKE_BUILD_TYPE
if(NOT CMAKE_BUILD_TYPE)
......@@ -101,6 +101,9 @@ endif()
set(THIRD_PARTY_PATH "${CMAKE_BINARY_DIR}/third_party" CACHE STRING
"A path setting third party libraries download & build directories.")
set(FLUID_INSTALL_DIR "${CMAKE_BINARY_DIR}/fluid_install_dir" CACHE STRING
"A path setting fluid shared and static libraries")
if (WITH_C_API AND WITH_PYTHON)
message(WARNING "It is suggest not embedded a python interpreter in Paddle "
"when using C-API. It will give an unpredictable behavior when using a "
......@@ -118,13 +121,14 @@ else()
endif()
set(WITH_MKLML ${WITH_MKL})
if (WITH_MKL AND AVX2_FOUND)
set(WITH_MKLDNN ON)
else()
message(STATUS "Do not have AVX2 intrinsics and disabled MKL-DNN")
set(WITH_MKLDNN OFF)
if (NOT DEFINED WITH_MKLDNN)
if (WITH_MKL AND AVX2_FOUND)
set(WITH_MKLDNN ON)
else()
message(STATUS "Do not have AVX2 intrinsics and disabled MKL-DNN")
set(WITH_MKLDNN OFF)
endif()
endif()
########################################################################################
include(external/mklml) # download mklml package
......@@ -153,7 +157,6 @@ include(cupti)
include(configure) # add paddle env configuration
include(generic) # simplify cmake module
include(package) # set paddle packages
include(cpplint) # set paddle c++ style
include(ccache) # set ccache for compilation
include(util) # set unittest and link libs
include(rdma) # set rdma libraries
......@@ -202,7 +205,7 @@ endif(USE_NNPACK)
add_subdirectory(proto)
if(NOT MOBILE_INFERENCE)
if(NOT MOBILE_INFERENCE AND NOT WITH_FLUID_ONLY)
# "add_subdirectory(go)" should be placed after the following loine,
# because it depends on paddle/optimizer.
add_subdirectory(paddle/optimizer)
......@@ -226,5 +229,11 @@ if(WITH_PYTHON)
endif()
if(WITH_DOC)
find_package(Sphinx REQUIRED)
find_python_module(recommonmark REQUIRED)
add_subdirectory(doc)
endif()
if (WITH_CONTRIB)
add_subdirectory(paddle/contrib)
endif()
CONTRIBUTING.md
浏览文件 @ fb3bfc3d
......@@ -58,6 +58,8 @@ PaddlePaddle uses this [Git branching model](http://nvie.com/posts/a-successful-
create mode 100644 233
```
NOTE: The `yapf` installed by `pip install pre-commit` and `conda install -c conda-forge pre-commit` is slightly different. Paddle developers use `pip install pre-commit`.
1. Build and test
Users can build PaddlePaddle natively on Linux and Mac OS X. But to unify the building environment and to make it easy for debugging, the recommended way is [using Docker](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/dev/build_en.md).
......
Dockerfile
浏览文件 @ fb3bfc3d
# A image for building paddle binaries
# Use cuda devel base image for both cpu and gpu environment
# When you modify it, please be aware of cudnn-runtime version
# When you modify it, please be aware of cudnn-runtime version
# and libcudnn.so.x in paddle/scripts/docker/build.sh
FROM nvidia/cuda:8.0-cudnn7-devel-ubuntu16.04
MAINTAINER PaddlePaddle Authors <paddle-dev@baidu.com>
......@@ -24,16 +23,16 @@ ENV HOME /root
COPY ./paddle/scripts/docker/root/ /root/
RUN apt-get update && \
apt-get install -y \
apt-get install -y --allow-downgrades \
git python-pip python-dev openssh-server bison \
libnccl2=2.1.2-1+cuda8.0 libnccl-dev=2.1.2-1+cuda8.0 \
wget unzip unrar tar xz-utils bzip2 gzip coreutils ntp \
curl sed grep graphviz libjpeg-dev zlib1g-dev \
python-matplotlib gcc-4.8 g++-4.8 \
automake locales clang-format swig doxygen cmake \
automake locales clang-format swig cmake \
liblapack-dev liblapacke-dev \
clang-3.8 llvm-3.8 libclang-3.8-dev \
net-tools libtool && \
net-tools libtool ccache && \
apt-get clean -y
# Install Go and glide
......@@ -71,7 +70,7 @@ RUN localedef -i en_US -f UTF-8 en_US.UTF-8
# specify sphinx version as 1.5.6 and remove -U option for [pip install -U
# sphinx-rtd-theme] since -U option will cause sphinx being updated to newest
# version(1.7.1 for now), which causes building documentation failed.
RUN pip install --upgrade pip==9.0.3 && \
RUN easy_install -U pip && \
pip install -U wheel && \
pip install -U docopt PyYAML sphinx==1.5.6 && \
pip install sphinx-rtd-theme==0.1.9 recommonmark
......@@ -80,6 +79,9 @@ RUN pip install pre-commit 'ipython==5.3.0' && \
pip install 'ipykernel==4.6.0' 'jupyter==1.0.0' && \
pip install opencv-python
#For docstring checker
RUN pip install pylint pytest astroid isort
COPY ./python/requirements.txt /root/
RUN pip install -r /root/requirements.txt
......@@ -102,6 +104,3 @@ RUN echo 'root:root' | chpasswd
RUN sed -ri 's/^PermitRootLogin\s+.*/PermitRootLogin yes/' /etc/ssh/sshd_config
RUN sed -ri 's/UsePAM yes/#UsePAM yes/g' /etc/ssh/sshd_config
EXPOSE 22
# development image default do build work
CMD ["bash", "/paddle/paddle/scripts/docker/build.sh"]
Dockerfile.android
浏览文件 @ fb3bfc3d
......@@ -40,5 +40,3 @@ RUN mkdir -p ${ANDROID_TOOLCHAINS_DIR} && \
unzip -q android-ndk-r14b-linux-x86_64.zip && \
mv android-ndk-r14b ${ANDROID_NDK_HOME} && \
rm -rf /opt/android-ndk-tmp
CMD ["bash", "/paddle/paddle/scripts/docker/build_android.sh"]
README.md
浏览文件 @ fb3bfc3d
......@@ -62,9 +62,9 @@ Please refer to our [release announcement](https://github.com/PaddlePaddle/Paddl
## Installation
It is recommended to check out the
[Docker installation guide](http://www.paddlepaddle.org/docs/develop/documentation/en/getstarted/build_and_install/docker_install_en.html)
[Docker installation guide](http://www.paddlepaddle.org/docs/develop/documentation/fluid/en/build_and_install/docker_install_en.html)
before looking into the
[build from source guide](http://www.paddlepaddle.org/docs/develop/documentation/en/getstarted/build_and_install/build_from_source_en.html).
[build from source guide](http://www.paddlepaddle.org/docs/develop/documentation/fluid/en/build_and_install/build_from_source_en.html).
## Documentation
......@@ -75,19 +75,19 @@ We provide [English](http://www.paddlepaddle.org/docs/develop/documentation/en/g
You might want to start from this online interactive book that can run in a Jupyter Notebook.
- [Distributed Training](http://www.paddlepaddle.org/docs/develop/documentation/en/howto/usage/cluster/cluster_train_en.html)
- [Distributed Training](http://www.paddlepaddle.org/docs/develop/documentation/en/howto/cluster/index_en.html)
You can run distributed training jobs on MPI clusters.
- [Distributed Training on Kubernetes](http://www.paddlepaddle.org/docs/develop/documentation/en/howto/usage/cluster/k8s_en.html)
- [Distributed Training on Kubernetes](http://www.paddlepaddle.org/docs/develop/documentation/en/howto/cluster/multi_cluster/k8s_en.html)
You can also run distributed training jobs on Kubernetes clusters.
- [Python API](http://www.paddlepaddle.org/docs/develop/documentation/en/api/index_en.html)
- [Python API](http://www.paddlepaddle.org/docs/develop/api/en/overview.html)
Our new API enables much shorter programs.
- [How to Contribute](http://www.paddlepaddle.org/docs/develop/documentation/en/howto/dev/contribute_to_paddle_en.html)
- [How to Contribute](http://www.paddlepaddle.org/docs/develop/documentation/fluid/en/dev/contribute_to_paddle_en.html)
We appreciate your contributions!
......
benchmark/cluster/README.md 已删除 100644 → 0
浏览文件 @ 652332cd
# Cluster Training Benchmark
## Setup
- Platform
- Kubernetes: v1.6.2
- Linux Kernel: v3.10.0
- Resource
- CPU: 10 Cores per Pod
- Memory: 5GB per Pod
- Docker Image
We use different base Docker Image to run the benchmark on Kubernetes:
- PaddlePaddle v2: paddlepaddle/paddle:0.11.0
- PaddlePaddle Fluid: paddlepaddle/paddle:[commit-id]
- TensorFlow: tensorflow/tensorflow:1.5.0-rc0
- Model
vgg16 is used in this benchmark.
## Cases
- Variable
- Batch Size of training data.
- PServer count of the training job.
- The number of trainers.
- Invariant
- The resource of trainer/pserver Pod.
### Measure the Performance for Different Batch Size
- PServer Count: 40
- Trainer Count: 100
- Metrics: mini-batch / sec
<table>
<thead>
<tr>
<th>Batch Size </th>
<th> 32</th>
<th>64</th>
<th>128 </th>
<th>256</th>
</tr>
</thead>
<tbody>
<tr>
<td> PaddlePaddle Fluid</td>
<td>-</td>
<td>- </td>
<td>- </td>
<td>- </td>
</tr>
<tr>
<td>PaddlePaddle v2 </td>
<td>- </td>
<td>- </td>
<td>- </td>
<td>- </td>
</tr>
<tr>
<td>TensorFlow </td>
<td>- </td>
<td>- </td>
<td>- </td>
<td>- </td>
</tr>
</tbody>
</table>
### Measure the Performance for Different PServer Count
- Trainer Count: 100
- Batch Size: 64
- Metrics: mini-batch / sec
<table>
<thead>
<tr>
<th>PServer Count </th>
<th>10</th>
<th>20</th>
<th>40 </th>
<th>60</th>
</tr>
</thead>
<tbody>
<tr>
<td> PaddlePaddle Fluid</td>
<td>-</td>
<td>- </td>
<td>- </td>
<td>- </td>
</tr>
<tr>
<td>PaddlePaddle v2 </td>
<td>- </td>
<td>- </td>
<td>- </td>
<td>- </td>
</tr>
<tr>
<td>TensorFlow </td>
<td>- </td>
<td>- </td>
<td>- </td>
<td>- </td>
</tr>
</tbody>
</table>
### Measure Parallel Efficiency By Increasing Trainer Count
- PServer Count: 20
- Batch Size: 64
- Metrics:
$S = \div(T1, TN)$
which S is the ratio of T1 over TN, training time of 1 and N trainers.
The parallel efficiency is:
$E = \div(S, N)$
<table>
<thead>
<tr>
<th>Trainer Counter </th>
<th>1</th>
<th>10</th>
<th>20 </th>
<th>30</th>
<th>40</th>
<th>50</th>
<th>60 </th>
<th>70</th>
<th>80</th>
<th>90</th>
<th>100 </th>
</tr>
</thead>
<tbody>
<tr>
<td> PaddlePaddle Fluid</td>
<td>-</td>
<td>- </td>
<td>- </td>
<td>- </td>
<td>-</td>
<td>- </td>
<td>- </td>
<td>- </td>
<td>-</td>
<td>- </td>
<td>- </td>
</tr>
<tr>
<td>PaddlePaddle v2 </td>
<td>- </td>
<td>- </td>
<td>- </td>
<td>- </td>
<td>-</td>
<td>- </td>
<td>- </td>
<td>- </td>
<td>-</td>
<td>- </td>
<td>- </td>
</tr>
<tr>
<td>TensorFlow </td>
<td>- </td>
<td>- </td>
<td>- </td>
<td>- </td>
<td>-</td>
<td>- </td>
<td>- </td>
<td>- </td>
<td>-</td>
<td>- </td>
<td>- </td>
</tr>
</tbody>
</table>
## Reproduce the benchmark
TODO
benchmark/cluster/vgg16/Dockerfile 已删除 100644 → 0
浏览文件 @ 652332cd
FROM nvidia/cuda:8.0-cudnn5-runtime-ubuntu16.04
# you can get mirror list here:
# https://launchpad.net/ubuntu/+archivemirrors
ARG UBUNTU_MIRROR
RUN /bin/bash -c 'if [[ -n ${UBUNTU_MIRROR} ]]; then sed -i 's#http://archive.ubuntu.com/ubuntu#${UBUNTU_MIRROR}#g' /etc/apt/sources.list; fi'
RUN apt-get update && apt-get install -y python python-dev python-pip iputils-ping libgtk2.0-dev
RUN pip install -U kubernetes opencv-python
RUN pip install paddlepaddle
# if network is slowly, you may need to add proxy here.
# ENV https_proxy=
RUN sh -c 'echo "import paddle.v2 as paddle\npaddle.dataset.cifar.train10()" | python'
RUN pip uninstall -y paddlepaddle
# unset proxy if it is setted.
# ENV https_proxy=""
# NOTE: By default CI built wheel packages turn WITH_DISTRIBUTE=OFF,
# so we must build one with distribute support to install in this image.
ADD *.whl /
RUN pip install /*.whl && rm -f /*.whl
ENV LD_LIBRARY_PATH=/usr/local/lib
# tf k8s
RUN pip install tensorflow==1.4.0
ADD tf_k8s /usr/bin
RUN chmod +x /usr/bin/tf_k8s
ADD vgg16_tf.py /workspace/
# below lines may change a lot for debugging
ADD https://raw.githubusercontent.com/PaddlePaddle/cloud/develop/docker/paddle_k8s /usr/bin
ADD https://raw.githubusercontent.com/PaddlePaddle/cloud/develop/docker/k8s_tools.py /root
RUN chmod +x /usr/bin/paddle_k8s
ADD vgg16_fluid.py vgg16_v2.py /workspace/
benchmark/cluster/vgg16/README.md 已删除 100644 → 0
浏览文件 @ 652332cd
# Performance for Distributed vgg16
## Test Result
### Hardware Infomation
- CPU: Intel(R) Xeon(R) CPU E5-2620 v4 @ 2.10GHz
- cpu MHz : 2101.000
- cache size : 20480 KB
### Blas settings
Setting environment variable: `MKL_NUM_THREADS=1`.
### Single Node Single Thread
- Metrics: samples / sec
<table>
<thead>
<tr>
<th>Batch Size </th>
<th> 32</th>
<th>64</th>
<th>128 </th>
<th>256</th>
</tr>
</thead>
<tbody>
<tr>
<td> PaddlePaddle Fluid</td>
<td> 15.44 </td>
<td> 16.32 </td>
<td> 16.74 </td>
<td> 16.79 </td>
</tr>
<tr>
<td>PaddlePaddle v2 </td>
<td> 15.97 </td>
<td> 17.04 </td>
<td> 17.60 </td>
<td> 17.83 </td>
</tr>
<tr>
<td>TensorFlow </td>
<td> 9.09 </td>
<td> 9.10 </td>
<td> 9.24 </td>
<td> 8.66 </td>
</tr>
</tbody>
</table>
### Different Batch Size
- PServer Count: 10
- Trainer Count: 20
- Metrics: samples / sec
<table>
<thead>
<tr>
<th>Batch Size </th>
<th> 32</th>
<th>64</th>
<th>128 </th>
<th>256</th>
</tr>
</thead>
<tbody>
<tr>
<td> PaddlePaddle Fluid</td>
<td> 190.20 </td>
<td> 222.15 </td>
<td> 247.40 </td>
<td> 258.18 </td>
</tr>
<tr>
<td>PaddlePaddle v2 </td>
<td> 170.96 </td>
<td> 233.71 </td>
<td> 256.14 </td>
<td> 329.23 </td>
</tr>
<tr>
<td>TensorFlow </td>
<td> - </td>
<td> - </td>
<td> - </td>
<td> - </td>
</tr>
</tbody>
</table>
### Accelerate Rate
- Pserver Count: 20
- Batch Size: 128
- Metrics: samples / sec
<table>
<thead>
<tr>
<th>Trainer Count </th>
<th>20</th>
<th>40</th>
<th>80</th>
<th>100</th>
</tr>
</thead>
<tbody>
<tr>
<td> PaddlePaddle Fluid</td>
<td> 263.29 (78.64%) </td>
<td> 518.80 (77.47%) </td>
<td> 836.26 (62.44%) </td>
<td> 1019.29 (60.89%) </td>
</tr>
<tr>
<td>PaddlePaddle v2 (need more tests) </td>
<td> 326.85 (92.85%) </td>
<td> 534.58 (75.93%) </td>
<td> 853.30 (60.60%) </td>
<td> 1041.99 (59.20%) </td>
</tr>
<tr>
<td>TensorFlow </td>
<td> - </td>
<td> - </td>
<td> - </td>
<td> - </td>
</tr>
</tbody>
</table>
### Different Pserver Count
- Trainer Count: 60
- Batch Size: 128
- Metrics: samples/ sec
<table>
<thead>
<tr>
<th>PServer Count </th>
<th>3</th>
<th>6</th>
<th>10</th>
<th>20</th>
</tr>
</thead>
<tbody>
<tr>
<td> PaddlePaddle Fluid(should fix in next PR) </td>
<td> 589.1 </td>
<td> 592.6 </td>
<td> 656.4 </td>
<td> 655.8 </td>
</tr>
<tr>
<td>PaddlePaddle v2 (need more tests) </td>
<td> 593.4 </td>
<td> 791.3 </td>
<td> 729.7 </td>
<td> 821.7 </td>
</tr>
<tr>
<td>TensorFlow </td>
<td> - </td>
<td> - </td>
<td> - </td>
<td> - </td>
</tr>
</tbody>
</table>
*The performance gap between Fuild and v2 comes from the network interference.*
## Steps to Run the Performance Test
1. You must re-compile PaddlePaddle and enable `-DWITH_DISTRIBUTE` to build PaddlePaddle with distributed support.
1. When the build finishes, copy the output `whl` package located under `build/python/dist` to current directory.
1. Run `docker build -t [image:tag] .` to build the docker image and run `docker push [image:tag]` to push the image to reponsitory so kubernetes can find it.
1. Run `kubectl create -f pserver.yaml && kubectl create -f trainer.yaml` to start the job on your kubernetes cluster (you must configure the `kubectl` client before this step).
1. Run `kubectl get po` to get running pods, and run `kubectl logs [podID]` to fetch the pod log of pservers and trainers.
Check the logs for the distributed training progress and analyze the performance.
## Enable Verbos Logs
Edit `pserver.yaml` and `trainer.yaml` and add an environment variable `GLOG_v=3` and `GLOG_logtostderr=1` to see what happend in detail.
benchmark/cluster/vgg16/fluid_pserver.yaml 已删除 100644 → 0
浏览文件 @ 652332cd
apiVersion: extensions/v1beta1
kind: ReplicaSet
metadata:
name: vgg16job-pserver
spec:
replicas: 10
template:
metadata:
labels:
paddle-job-pserver: vgg16job
spec:
hostNetwork: true
imagePullSecrets:
- name: job-registry-secret
containers:
- name: pserver
image: "registry.baidu.com/paddlepaddle/fluid_benchmark:vgg16"
imagePullPolicy: Always
ports:
- name: jobport-30236
containerPort: 30236
env:
- name: PADDLE_JOB_NAME
value: vgg16job
- name: MKL_NUM_THREADS
value: "1"
- name: TRAINING_ROLE
value: "PSERVER"
- name: TRAINERS
value: "20"
- name: PSERVERS
value: "10"
- name: TOPOLOGY
value: ""
- name: ENTRY
value: "MKL_NUM_THREADS=1 python /workspace/vgg16_fluid.py --local 0"
- name: TRAINER_PACKAGE
value: "/workspace"
- name: PADDLE_INIT_PORT
value: "30236"
- name: PADDLE_INIT_NICS
value: "xgbe0"
- name: PADDLE_INIT_TRAINER_COUNT
value: "1"
- name: PADDLE_INIT_PORTS_NUM
value: "1"
- name: PADDLE_INIT_PORTS_NUM_FOR_SPARSE
value: "1"
- name: PADDLE_INIT_NUM_GRADIENT_SERVERS
value: "20"
- name: PADDLE_INIT_NUM_PASSES
value: "1"
- name: PADDLE_INIT_USE_GPU
value: "0"
- name: LD_LIBRARY_PATH
value: "/usr/local/lib:/usr/local/nvidia/lib64"
- name: NAMESPACE
valueFrom:
fieldRef:
fieldPath: "metadata.namespace"
- name: POD_IP
valueFrom:
fieldRef:
fieldPath: "status.podIP"
command: ["paddle_k8s", "start_fluid"]
resources:
requests:
memory: 10Gi
cpu: 4
limits:
memory: 10Gi
cpu: 4
benchmark/cluster/vgg16/fluid_trainer.yaml 已删除 100644 → 0
浏览文件 @ 652332cd
apiVersion: batch/v1
kind: Job
metadata:
name: vgg16job-trainer
spec:
parallelism: 20
completions: 20
template:
metadata:
labels:
paddle-job: vgg16job
spec:
imagePullSecrets:
- name: job-registry-secret
hostNetwork: true
containers:
- name: trainer
image: "registry.baidu.com/paddlepaddle/fluid_benchmark:vgg16"
imagePullPolicy: Always
command: ["paddle_k8s", "start_fluid"]
env:
- name: PADDLE_JOB_NAME
value: vgg16job
- name: TRAINING_ROLE
value: "TRAINER"
- name: TRAINERS
value: "20"
- name: PSERVERS
value: "10"
- name: TOPOLOGY
value: ""
- name: ENTRY
value: "MKL_NUM_THREADS=1 python /workspace/vgg16_fluid.py --local 0 --batch_size 128"
- name: TRAINER_PACKAGE
value: "/workspace"
- name: PADDLE_INIT_PORT
value: "30236"
- name: PADDLE_INIT_NICS
value: "xgbe0"
- name: PADDLE_INIT_TRAINER_COUNT
value: "1"
- name: PADDLE_INIT_PORTS_NUM
value: "1"
- name: PADDLE_INIT_PORTS_NUM_FOR_SPARSE
value: "1"
- name: PADDLE_INIT_NUM_GRADIENT_SERVERS
value: "20"
- name: PADDLE_INIT_NUM_PASSES
value: "1"
- name: PADDLE_INIT_USE_GPU
value: "0"
- name: LD_LIBRARY_PATH
value: "/usr/local/lib:/usr/local/nvidia/lib64"
- name: NAMESPACE
valueFrom:
fieldRef:
fieldPath: "metadata.namespace"
- name: POD_IP
valueFrom:
fieldRef:
fieldPath: "status.podIP"
resources:
requests:
memory: 40Gi
cpu: 2
limits:
memory: 40Gi
cpu: 2
restartPolicy: Never
benchmark/cluster/vgg16/tf_k8s 已删除 100644 → 0
浏览文件 @ 652332cd
#!/bin/bash
check_trainer_ret() {
ret=$1
stdbuf -oL echo "job returned $ret...setting pod return message..."
stdbuf -oL echo "==============================="
if [ $ret -eq 136 ] ; then
echo "Error Arithmetic Operation(Floating Point Exception)" > /dev/termination-log
elif [ $ret -eq 139 ] ; then
echo "Segmentation Fault" > /dev/termination-log
elif [ $ret -eq 1 ] ; then
echo "General Error" > /dev/termination-log
elif [ $ret -eq 134 ] ; then
echo "Program Abort" > /dev/termination-log
fi
stdbuf -oL echo "termination log wroted..."
exit $ret
}
g_pservers=""
g_trainers=""
wait_running_pods(){
pserver_label="tf-job-pserver=${JOB_NAME}"
trainer_label="tf-job-trainer=${JOB_NAME}"
stdbuf -oL python /root/k8s_tools.py wait_pods_running ${pserver_label} ${PSERVERS_NUM}
stdbuf -oL python /root/k8s_tools.py wait_pods_running ${trainer_label} ${TRAINERS_NUM}
g_pservers=$(python /root/k8s_tools.py fetch_endpoints ${pserver_label} ${PORT})
g_trainers=$(python /root/k8s_tools.py fetch_endpoints ${trainer_label} ${PORT})
}
start_tf_pserver(){
wait_running_pods
label="tf-job-pserver=${JOB_NAME}"
pserver_id=$(python /root/k8s_tools.py fetch_id ${label})
cmd="${ENTRY} --ps_hosts=${g_pservers} --worker_hosts=${g_trainers} \
--job_name=${TF_JOB_NAME} --task_index=${pserver_id}"
stdbuf -oL sh -c "cd ${TRAINER_PACKAGE} && ${cmd}"
}
start_tf_trainer(){
wait_running_pods
label="tf-job-trainer=${JOB_NAME}"
trainer_id=$(python /root/k8s_tools.py fetch_id ${label})
cmd="${ENTRY} --ps_hosts=${g_pservers} --worker_hosts=${g_trainers} \
--job_name=${TF_JOB_NAME} --task_index=${trainer_id} --batch_size=${BATCH_SIZE}"
stdbuf -oL sh -c "cd ${TRAINER_PACKAGE} && ${cmd}"
check_trainer_ret $?
}
start_tf(){
if [[ "${TF_JOB_NAME}" == "worker" ]]; then
start_tf_trainer
else
start_tf_pserver
fi
}
usage() {
echo "usage: tf_k8s [<args>]:"
echo " start_tf Start tensorflow jobs"
}
case "$1" in
start_tf)
start_tf
;;
--help)
usage
;;
*)
usage
;;
esac
benchmark/cluster/vgg16/tf_pserver.yaml 已删除 100644 → 0
浏览文件 @ 652332cd
apiVersion: extensions/v1beta1
kind: ReplicaSet
metadata:
name: vgg16job-tf-pserver
spec:
replicas: 10
template:
metadata:
labels:
tf-job-pserver: vgg16job-tf
spec:
hostNetwork: true
imagePullSecrets:
- name: job-registry-secret
containers:
- name: pserver
image: "registry.baidu.com/paddlepaddle/fluid_benchmark_tf:vgg16"
imagePullPolicy: Always
command: ["tf_k8s", "start_tf"]
ports:
- name: jobport-30236
containerPort: 30236
env:
- name: PORT
value: "32036"
- name: ENTRY
value: "python vgg16_tf.py"
- name: JOB_NAME
value: vgg16job-tf
- name: PSERVERS_NUM
value: "10"
- name: TF_JOB_NAME
value: "ps"
- name: TRAINERS_NUM
value: "20"
- name: BATCH_SIZE
value: "128"
- name: TRAINER_PACKAGE
value: "/workspace"
- name: NUM_PASSES
value: "1"
- name: NAMESPACE
valueFrom:
fieldRef:
fieldPath: "metadata.namespace"
- name: POD_IP
valueFrom:
fieldRef:
fieldPath: "status.podIP"
resources:
requests:
memory: 10Gi
cpu: 4
limits:
memory: 10Gi
cpu: 4
benchmark/cluster/vgg16/tf_trainer.yaml 已删除 100644 → 0
浏览文件 @ 652332cd
apiVersion: batch/v1
kind: Job
metadata:
name: vgg16job-tf-trainer
spec:
parallelism: 20
completions: 20
template:
metadata:
labels:
tf-job-trainer: vgg16job-tf
spec:
imagePullSecrets:
- name: job-registry-secret
hostNetwork: true
containers:
- name: trainer
image: "registry.baidu.com/paddlepaddle/fluid_benchmark_tf:vgg16"
imagePullPolicy: Always
command: ["tf_k8s", "start_tf"]
ports:
- name: jobport-30236
containerPort: 30236
env:
- name: PORT
value: "32036"
- name: JOB_NAME
value: vgg16job-tf
- name: TF_JOB_NAME
value: "worker"
- name: ENTRY
value: "python vgg16_tf.py"
- name: PSERVERS_NUM
value: "10"
- name: BATCH_SIZE
value: "128"
- name: TRAINERS_NUM
value: "20"
- name: TRAINER_PACKAGE
value: "/workspace"
- name: NUM_PASSES
value: "1"
- name: NAMESPACE
valueFrom:
fieldRef:
fieldPath: "metadata.namespace"
- name: POD_IP
valueFrom:
fieldRef:
fieldPath: "status.podIP"
resources:
requests:
memory: 40Gi
cpu: 2
limits:
memory: 40Gi
cpu: 2
restartPolicy: Never
benchmark/cluster/vgg16/v2_pserver.yaml 已删除 100644 → 0
浏览文件 @ 652332cd
apiVersion: extensions/v1beta1
kind: ReplicaSet
metadata:
name: vgg16v2job-pserver
spec:
replicas: 10
template:
metadata:
labels:
paddle-job-pserver: vgg16v2job
spec:
hostNetwork: true
imagePullSecrets:
- name: job-registry-secret
containers:
- name: pserver
image: "registry.baidu.com/paddlepaddle/fluid_benchmark:vgg16"
imagePullPolicy: Always
ports:
- name: jobport-30236
containerPort: 30236
env:
- name: PADDLE_JOB_NAME
value: vgg16v2job
- name: TRAINERS
value: "20"
- name: PSERVERS
value: "10"
- name: TOPOLOGY
value: ""
- name: ENTRY
value: "python train.py"
- name: TRAINER_PACKAGE
value: "/workspace"
- name: PADDLE_INIT_PORT
value: "30236"
- name: PADDLE_INIT_NICS
value: "xgbe0"
- name: PADDLE_INIT_TRAINER_COUNT
value: "1"
- name: PADDLE_INIT_PORTS_NUM
value: "1"
- name: PADDLE_INIT_PORTS_NUM_FOR_SPARSE
value: "1"
- name: PADDLE_INIT_NUM_GRADIENT_SERVERS
value: "20"
- name: PADDLE_INIT_NUM_PASSES
value: "1"
- name: PADDLE_INIT_USE_GPU
value: "0"
- name: LD_LIBRARY_PATH
value: "/usr/local/lib:/usr/local/nvidia/lib64"
- name: NAMESPACE
valueFrom:
fieldRef:
fieldPath: "metadata.namespace"
command: ["paddle_k8s", "start_pserver"]
resources:
requests:
memory: 10Gi
cpu: 4
limits:
memory: 10Gi
cpu: 4
benchmark/cluster/vgg16/v2_trainer.yaml 已删除 100644 → 0
浏览文件 @ 652332cd
apiVersion: batch/v1
kind: Job
metadata:
name: vgg16v2job-trainer
spec:
parallelism: 20
completions: 20
template:
metadata:
labels:
paddle-job: vgg16v2job
spec:
imagePullSecrets:
- name: job-registry-secret
hostNetwork: true
containers:
- name: trainer
image: "registry.baidu.com/paddlepaddle/fluid_benchmark:vgg16"
imagePullPolicy: Always
command: ["paddle_k8s", "start_trainer", "v2"]
env:
- name: PADDLE_JOB_NAME
value: vgg16v2job
- name: BATCH_SIZE
value: "256"
- name: TRAINERS
value: "20"
- name: PSERVERS
value: "10"
- name: TOPOLOGY
value: ""
- name: ENTRY
value: "cd /workspace && MKL_NUM_THREADS=1 python /workspace/vgg16_v2.py"
- name: TRAINER_PACKAGE
value: "/workspace"
- name: PADDLE_INIT_PORT
value: "30236"
- name: PADDLE_INIT_NICS
value: "xgbe0"
- name: PADDLE_INIT_TRAINER_COUNT
value: "1"
- name: PADDLE_INIT_PORTS_NUM
value: "1"
- name: PADDLE_INIT_PORTS_NUM_FOR_SPARSE
value: "1"
- name: PADDLE_INIT_NUM_GRADIENT_SERVERS
value: "20"
- name: PADDLE_INIT_NUM_PASSES
value: "2"
- name: PADDLE_INIT_USE_GPU
value: "0"
- name: LD_LIBRARY_PATH
value: "/usr/local/lib:/usr/local/nvidia/lib64"
- name: NAMESPACE
valueFrom:
fieldRef:
fieldPath: "metadata.namespace"
resources:
requests:
memory: 40Gi
cpu: 2
limits:
memory: 40Gi
cpu: 2
restartPolicy: Never
benchmark/cluster/vgg16/vgg16_fluid.py 已删除 100644 → 0
浏览文件 @ 652332cd
# Copyright (c) 2018 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.
"""VGG16 benchmark in Fluid"""
from __future__ import print_function
import sys
import time
import numpy as np
import paddle.v2 as paddle
import paddle.fluid as fluid
import paddle.fluid.core as core
import paddle.fluid.profiler as profiler
import argparse
import functools
import os
from paddle.fluid import debuger
def str2bool(v):
if v.lower() in ('yes', 'true', 't', 'y', '1'):
return True
elif v.lower() in ('no', 'false', 'f', 'n', '0'):
return False
else:
raise argparse.ArgumentTypeError('Boolean value expected.')
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument(
'--batch_size', type=int, default=128, help="Batch size for training.")
parser.add_argument(
'--learning_rate',
type=float,
default=1e-3,
help="Learning rate for training.")
parser.add_argument('--num_passes', type=int, default=50, help="No. of passes.")
parser.add_argument(
'--device',
type=str,
default='CPU',
choices=['CPU', 'GPU'],
help="The device type.")
parser.add_argument('--device_id', type=int, default=0, help="The device id.")
parser.add_argument(
'--data_format',
type=str,
default='NCHW',
choices=['NCHW', 'NHWC'],
help='The data order, now only support NCHW.')
parser.add_argument(
'--data_set',
type=str,
default='cifar10',
choices=['cifar10', 'flowers'],
help='Optional dataset for benchmark.')
parser.add_argument(
'--local',
type=str2bool,
default=True,
help='Whether to run as local mode.')
parser.add_argument(
"--ps_hosts",
type=str,
default="",
help="Comma-separated list of hostname:port pairs")
parser.add_argument(
"--trainer_hosts",
type=str,
default="",
help="Comma-separated list of hostname:port pairs")
# Flags for defining the tf.train.Server
parser.add_argument(
"--task_index", type=int, default=0, help="Index of task within the job")
args = parser.parse_args()
def vgg16_bn_drop(input):
def conv_block(input, num_filter, groups, dropouts):
return fluid.nets.img_conv_group(
input=input,
pool_size=2,
pool_stride=2,
conv_num_filter=[num_filter] * groups,
conv_filter_size=3,
conv_act='relu',
conv_with_batchnorm=True,
conv_batchnorm_drop_rate=dropouts,
pool_type='max')
conv1 = conv_block(input, 64, 2, [0.3, 0])
conv2 = conv_block(conv1, 128, 2, [0.4, 0])
conv3 = conv_block(conv2, 256, 3, [0.4, 0.4, 0])
conv4 = conv_block(conv3, 512, 3, [0.4, 0.4, 0])
conv5 = conv_block(conv4, 512, 3, [0.4, 0.4, 0])
drop = fluid.layers.dropout(x=conv5, dropout_prob=0.5)
fc1 = fluid.layers.fc(input=drop, size=4096, act=None)
bn = fluid.layers.batch_norm(input=fc1, act='relu')
drop2 = fluid.layers.dropout(x=bn, dropout_prob=0.5)
fc2 = fluid.layers.fc(input=drop2, size=4096, act=None)
return fc2
def main():
if args.data_set == "cifar10":
classdim = 10
if args.data_format == 'NCHW':
data_shape = [3, 32, 32]
else:
data_shape = [32, 32, 3]
else:
classdim = 102
if args.data_format == 'NCHW':
data_shape = [3, 224, 224]
else:
data_shape = [224, 224, 3]
# Input data
images = fluid.layers.data(name='pixel', shape=data_shape, dtype='float32')
label = fluid.layers.data(name='label', shape=[1], dtype='int64')
# Train program
net = vgg16_bn_drop(images)
predict = fluid.layers.fc(input=net, size=classdim, act='softmax')
cost = fluid.layers.cross_entropy(input=predict, label=label)
avg_cost = fluid.layers.mean(x=cost)
# Evaluator
batch_size = fluid.layers.create_tensor(dtype='int64')
batch_acc = fluid.layers.accuracy(
input=predict, label=label, total=batch_size)
# inference program
inference_program = fluid.default_main_program().clone()
with fluid.program_guard(inference_program):
inference_program = fluid.io.get_inference_program(batch_acc)
# Optimization
optimizer = fluid.optimizer.Adam(learning_rate=args.learning_rate)
optimize_ops, params_grads = optimizer.minimize(avg_cost)
# Initialize executor
place = core.CPUPlace() if args.device == 'CPU' else core.CUDAPlace(
args.device_id)
exe = fluid.Executor(place)
# test
def test(exe):
test_pass_acc = fluid.average.WeightedAverage()
for batch_id, data in enumerate(test_reader()):
img_data = np.array(map(lambda x: x[0].reshape(data_shape),
data)).astype("float32")
y_data = np.array(map(lambda x: x[1], data)).astype("int64")
y_data = y_data.reshape([-1, 1])
outs = exe.run(inference_program,
feed={"pixel": img_data,
"label": y_data},
fetch_list=[batch_acc, batch_size])
test_pass_acc.add(value=np.array(outs[0]), weight=np.array(outs[1]))
return test_pass_acc.eval()
def train_loop(exe, trainer_prog):
iters = 0
ts = time.time()
train_pass_acc = fluid.average.WeightedAverage()
for pass_id in range(args.num_passes):
# train
start_time = time.time()
num_samples = 0
train_pass_acc.reset()
for batch_id, data in enumerate(train_reader()):
ts = time.time()
img_data = np.array(
map(lambda x: x[0].reshape(data_shape), data)).astype(
"float32")
y_data = np.array(map(lambda x: x[1], data)).astype("int64")
y_data = y_data.reshape([-1, 1])
loss, acc, b_size = exe.run(
trainer_prog,
feed={"pixel": img_data,
"label": y_data},
fetch_list=[avg_cost, batch_acc, batch_size])
iters += 1
num_samples += len(data)
train_pass_acc.add(value=acc, weight=b_size)
print(
"Pass = %d, Iters = %d, Loss = %f, Accuracy = %f, Speed = %.2f img/s"
% (pass_id, iters, loss, acc,
len(data) / (time.time() - ts))
) # The accuracy is the accumulation of batches, but not the current batch.
pass_elapsed = time.time() - start_time
pass_train_acc = train_pass_acc.eval()
pass_test_acc = test(exe)
print(
"Pass = %d, Training performance = %f imgs/s, Train accuracy = %f, Test accuracy = %f\n"
% (pass_id, num_samples / pass_elapsed, pass_train_acc,
pass_test_acc))
if args.local:
# Parameter initialization
exe.run(fluid.default_startup_program())
# data reader
train_reader = paddle.batch(
paddle.reader.shuffle(
paddle.dataset.cifar.train10() if args.data_set == 'cifar10'
else paddle.dataset.flowers.train(),
buf_size=5120),
batch_size=args.batch_size)
test_reader = paddle.batch(
paddle.dataset.cifar.test10()
if args.data_set == 'cifar10' else paddle.dataset.flowers.test(),
batch_size=args.batch_size)
train_loop(exe, fluid.default_main_program())
else:
trainers = int(os.getenv("TRAINERS")) # total trainer count
print("trainers total: ", trainers)
training_role = os.getenv(
"TRAINING_ROLE",
"TRAINER") # get the training role: trainer/pserver
t = fluid.DistributeTranspiler()
t.transpile(
optimize_ops,
params_grads,
trainer_id=args.task_index,
pservers=args.ps_hosts,
trainers=trainers)
if training_role == "PSERVER":
current_endpoint = os.getenv("POD_IP") + ":" + os.getenv(
"PADDLE_INIT_PORT")
if not current_endpoint:
print("need env SERVER_ENDPOINT")
exit(1)
pserver_prog = t.get_pserver_program(current_endpoint)
pserver_startup = t.get_startup_program(current_endpoint,
pserver_prog)
exe.run(pserver_startup)
exe.run(pserver_prog)
elif training_role == "TRAINER":
# Parameter initialization
exe.run(fluid.default_startup_program())
# data reader
train_reader = paddle.batch(
paddle.reader.shuffle(
paddle.dataset.cifar.train10() if args.data_set == 'cifar10'
else paddle.dataset.flowers.train(),
buf_size=5120),
batch_size=args.batch_size)
test_reader = paddle.batch(
paddle.dataset.cifar.test10() if args.data_set == 'cifar10' else
paddle.dataset.flowers.test(),
batch_size=args.batch_size)
trainer_prog = t.get_trainer_program()
feeder = fluid.DataFeeder(feed_list=[images, label], place=place)
# TODO(typhoonzero): change trainer startup program to fetch parameters from pserver
exe.run(fluid.default_startup_program())
train_loop(exe, trainer_prog)
else:
print("environment var TRAINER_ROLE should be TRAINER os PSERVER")
def print_arguments():
print('----------- Configuration Arguments -----------')
for arg, value in sorted(vars(args).iteritems()):
print('%s: %s' % (arg, value))
print('------------------------------------------------')
if __name__ == "__main__":
print_arguments()
main()
benchmark/cluster/vgg16/vgg16_tf.py 已删除 100644 → 0
浏览文件 @ 652332cd
# Copyright (c) 2018 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.
"""VGG16 benchmark in TensorFlow
You can get distribution example template structure here:
https://medium.com/clusterone/how-to-write-distributed-tensorflow-code-with-an-example-on-tensorport-70bf3306adcb
https://www.tensorflow.org/deploy/distributed
"""
import tensorflow as tf
import paddle.v2 as paddle
import numpy as np
import argparse
import time
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument(
'--batch_size', type=int, default=128, help="Batch size for training.")
parser.add_argument(
'--learning_rate',
type=float,
default=1e-3,
help="Learning rate for training.")
parser.add_argument('--num_passes', type=int, default=50, help="No. of passes.")
parser.add_argument(
'--device',
type=str,
default='CPU',
choices=['CPU', 'GPU'],
help="The device type.")
parser.add_argument(
'--data_format',
type=str,
default='NHWC',
choices=['NCHW', 'NHWC'],
help='The data order, NCHW=[batch, channels, height, width].'
'Only support NHWC right now.')
parser.add_argument(
'--data_set',
type=str,
default='cifar10',
choices=['cifar10', 'flowers'],
help='Optional dataset for benchmark.')
parser.add_argument(
"--ps_hosts",
type=str,
default="",
help="Comma-separated list of hostname:port pairs")
parser.add_argument(
"--worker_hosts",
type=str,
default="",
help="Comma-separated list of hostname:port pairs")
parser.add_argument(
"--job_name", type=str, default="", help="One of 'worker', 'ps'")
# Flags for defining the tf.train.Server
parser.add_argument(
"--task_index", type=int, default=0, help="Index of task within the job")
args = parser.parse_args()
class VGG16Model(object):
def __init__(self):
self.parameters = []
def batch_norm_relu(self, inputs, is_training):
"""Performs a batch normalization followed by a ReLU."""
# We set fused=True for a significant speed boost. See
# https://www.tensorflow.org/speed/speed_guide#common_fused_ops
inputs = tf.layers.batch_normalization(
inputs=inputs,
axis=1 if args.data_format == 'NCHW' else -1,
momentum=0.9,
epsilon=1e-05,
center=True,
scale=True,
training=is_training,
fused=True)
inputs = tf.nn.relu(inputs)
return inputs
def conv_bn_layer(self,
name,
images,
kernel_shape,
is_training,
drop_rate=0.0):
with tf.name_scope(name) as scope:
kernel = tf.Variable(
tf.truncated_normal(
kernel_shape, dtype=tf.float32, stddev=1e-1),
name='weights')
conv = tf.nn.conv2d(
images,
kernel, [1, 1, 1, 1],
data_format=args.data_format,
padding='SAME')
biases = tf.Variable(
tf.constant(
0.0, shape=[kernel_shape[-1]], dtype=tf.float32),
trainable=True,
name='biases')
out = tf.nn.bias_add(conv, biases)
out = self.batch_norm_relu(out, is_training)
out = tf.layers.dropout(out, rate=drop_rate, training=is_training)
return out
def fc_layer(self, name, inputs, shape):
with tf.name_scope(name) as scope:
fc_w = tf.Variable(
tf.truncated_normal(
shape, dtype=tf.float32, stddev=1e-1),
name='weights')
fc_b = tf.Variable(
tf.constant(
0.0, shape=[shape[-1]], dtype=tf.float32),
trainable=True,
name='biases')
out = tf.nn.bias_add(tf.matmul(inputs, fc_w), fc_b)
return out
def network(self, images, class_dim, is_training):
""" VGG16 model structure.
TODO(kuke): enable this network to support the 'NCHW' data format
"""
# conv1
conv1_1 = self.conv_bn_layer(
'conv1_1', images, [3, 3, 3, 64], is_training, drop_rate=0.3)
conv1_2 = self.conv_bn_layer(
'conv1_2', conv1_1, [3, 3, 64, 64], is_training, drop_rate=0.0)
# pool1
pool1 = tf.nn.max_pool(
conv1_2,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME',
name='pool1')
# conv2
conv2_1 = self.conv_bn_layer(
'conv2_1', pool1, [3, 3, 64, 128], is_training, drop_rate=0.4)
conv2_2 = self.conv_bn_layer(
'conv2_2', conv2_1, [3, 3, 128, 128], is_training, drop_rate=0.0)
# pool2
pool2 = tf.nn.max_pool(
conv2_2,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME',
name='pool2')
# conv3
conv3_1 = self.conv_bn_layer(
'conv3_1', pool2, [3, 3, 128, 256], is_training, drop_rate=0.4)
conv3_2 = self.conv_bn_layer(
'conv3_2', conv3_1, [3, 3, 256, 256], is_training, drop_rate=0.4)
conv3_3 = self.conv_bn_layer(
'conv3_3', conv3_2, [3, 3, 256, 256], is_training, drop_rate=0.0)
# pool3
pool3 = tf.nn.max_pool(
conv3_3,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME',
name='pool3')
# conv4
conv4_1 = self.conv_bn_layer(
'conv4_1', pool3, [3, 3, 256, 512], is_training, drop_rate=0.4)
conv4_2 = self.conv_bn_layer(
'conv4_2', conv4_1, [3, 3, 512, 512], is_training, drop_rate=0.4)
conv4_3 = self.conv_bn_layer(
'conv4_3', conv4_2, [3, 3, 512, 512], is_training, drop_rate=0.0)
# pool4
pool4 = tf.nn.max_pool(
conv4_3,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME',
name='pool4')
# conv5
conv5_1 = self.conv_bn_layer(
'conv5_1', pool4, [3, 3, 512, 512], is_training, drop_rate=0.4)
conv5_2 = self.conv_bn_layer(
'conv5_2', conv5_1, [3, 3, 512, 512], is_training, drop_rate=0.4)
conv5_3 = self.conv_bn_layer(
'conv5_3', conv5_2, [3, 3, 512, 512], is_training, drop_rate=0.0)
# pool5
pool5 = tf.nn.max_pool(
conv5_3,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME',
name='pool4')
# flatten
shape = int(np.prod(pool5.get_shape()[1:]))
pool5_flat = tf.reshape(pool5, [-1, shape])
# fc1
drop = tf.layers.dropout(pool5_flat, rate=0.5, training=is_training)
fc1 = self.fc_layer('fc1', drop, [shape, 512])
# fc2
bn = self.batch_norm_relu(fc1, is_training)
drop = tf.layers.dropout(bn, rate=0.5, training=is_training)
fc2 = self.fc_layer('fc2', drop, [512, 512])
fc3 = self.fc_layer('fc3', fc2, [512, class_dim])
return fc3
def run_benchmark(cluster_spec, server):
"""Run benchmark on cifar10 or flowers."""
if args.data_set == "cifar10":
class_dim = 10
raw_shape = (3, 32, 32)
dat_shape = (None, 32, 32, 3) if args.data_format == 'NHWC' else (
None, 3, 32, 32)
else:
class_dim = 102
raw_shape = (3, 224, 224)
dat_shape = (None, 224, 224, 3) if args.data_format == 'NHWC' else (
None, 3, 224, 224)
device = tf.train.replica_device_setter(
worker_device="/job:worker/task:{}".format(args.task_index),
cluster=cluster_spec)
with tf.device(device):
images = tf.placeholder(tf.float32, shape=dat_shape)
labels = tf.placeholder(tf.int64, shape=(None, ))
is_training = tf.placeholder('bool')
onehot_labels = tf.one_hot(labels, depth=class_dim)
vgg16 = VGG16Model()
logits = vgg16.network(images, class_dim, is_training)
loss = tf.losses.softmax_cross_entropy(
onehot_labels=onehot_labels, logits=logits)
avg_loss = tf.reduce_mean(loss)
correct = tf.equal(tf.argmax(logits, 1), labels)
accuracy = tf.reduce_mean(tf.cast(correct, tf.float32))
optimizer = tf.train.AdamOptimizer(learning_rate=args.learning_rate)
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
global_step = tf.Variable(0, name='global_step', trainable=False)
with tf.control_dependencies(update_ops):
train_op = optimizer.minimize(avg_loss, global_step=global_step)
summary_op = tf.summary.merge_all()
init_op = tf.global_variables_initializer()
# data reader
train_reader = paddle.batch(
paddle.reader.shuffle(
paddle.dataset.cifar.train10()
if args.data_set == 'cifar10' else paddle.dataset.flowers.train(),
buf_size=5120),
batch_size=args.batch_size)
test_reader = paddle.batch(
paddle.reader.shuffle(
paddle.dataset.cifar.test10()
if args.data_set == 'cifar10' else paddle.dataset.flowers.test(),
buf_size=5120),
batch_size=args.batch_size)
# test
def test():
test_accs = []
for batch_id, data in enumerate(test_reader()):
test_images = np.array(
map(lambda x: np.transpose(x[0].reshape(raw_shape),
axes=[1, 2, 0]) if args.data_format == 'NHWC' else x[0], data)).astype("float32")
test_labels = np.array(map(lambda x: x[1], data)).astype('int64')
test_accs.append(
accuracy.eval(feed_dict={
images: test_images,
labels: test_labels,
is_training: False
}))
return np.mean(test_accs)
config = tf.ConfigProto(
intra_op_parallelism_threads=1,
inter_op_parallelism_threads=1,
log_device_placement=True)
config.gpu_options.allow_growth = True
hooks = [tf.train.StopAtStepHook(last_step=1000000)]
with tf.train.MonitoredTrainingSession(
master=server.target,
is_chief=(args.task_index == 0),
hooks=hooks,
config=config) as sess:
iters, num_samples, start_time = 0, 0, 0.0
for pass_id in range(args.num_passes):
# train
num_samples = 0
start_time = time.time()
for batch_id, data in enumerate(train_reader()):
train_images = np.array(
map(lambda x: np.transpose(x[0].reshape(raw_shape),
axes=[1, 2, 0]) if args.data_format == 'NHWC' else x[0], data)).astype("float32")
train_labels = np.array(map(lambda x: x[1], data)).astype(
'int64')
iter_begin_time = time.time()
_, loss, acc = sess.run([train_op, avg_loss, accuracy],
feed_dict={
images: train_images,
labels: train_labels,
is_training: True
})
iters += 1
print(
"Pass = %d, Iters = %d, Loss = %f, Accuracy = %f, Speed=%.2f imgs/sec"
% (pass_id, iters, loss, acc,
len(data) / (time.time() - iter_begin_time)))
num_samples += len(data)
train_elapsed = time.time() - start_time
# test
pass_test_acc = test()
print("Pass = %d, Train speed = %f imgs/s, Test accuracy = %f\n" %
(pass_id, num_samples / train_elapsed, pass_test_acc))
def print_arguments():
print('----------- Configuration Arguments -----------')
for arg, value in sorted(vars(args).iteritems()):
print('%s: %s' % (arg, value))
print('------------------------------------------------')
if __name__ == '__main__':
print_arguments()
ps_hosts = args.ps_hosts.split(",")
worker_hosts = args.worker_hosts.split(",")
# Create a cluster from the parameter server and worker hosts.
cluster_spec = tf.train.ClusterSpec({
"ps": ps_hosts,
"worker": worker_hosts
})
# Create and start a server for the local task.
server = tf.train.Server(
cluster_spec, job_name=args.job_name, task_index=args.task_index)
if args.job_name == "ps":
print("start pserver")
server.join()
elif args.job_name == "worker":
print("start worker")
run_benchmark(cluster_spec, server)
benchmark/cluster/vgg16/vgg16_v2.py 已删除 100644 → 0
浏览文件 @ 652332cd
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
#
#Licensed under the Apache License, Version 2.0 (the "License");
#you may not use this file except in compliance with the License.
#You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
#Unless required by applicable law or agreed to in writing, software
#distributed under the License is distributed on an "AS IS" BASIS,
#WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
#See the License for the specific language governing permissions and
#limitations under the License.
import gzip
import paddle.v2.dataset.cifar as cifar
import paddle.v2 as paddle
import time
import os
DATA_DIM = 3 * 32 * 32
CLASS_DIM = 10
BATCH_SIZE = os.getenv("BATCH_SIZE")
if BATCH_SIZE:
BATCH_SIZE = int(BATCH_SIZE)
else:
BATCH_SIZE = 128
print "batch_size", BATCH_SIZE
NODE_COUNT = int(os.getenv("TRAINERS"))
ts = 0
def vgg(input, nums, class_dim):
def conv_block(input, num_filter, groups, num_channels=None):
return paddle.networks.img_conv_group(
input=input,
num_channels=num_channels,
pool_size=2,
pool_stride=2,
conv_num_filter=[num_filter] * groups,
conv_filter_size=3,
conv_act=paddle.activation.Relu(),
pool_type=paddle.pooling.Max())
assert len(nums) == 5
# the channel of input feature is 3
conv1 = conv_block(input, 64, nums[0], 3)
conv2 = conv_block(conv1, 128, nums[1])
conv3 = conv_block(conv2, 256, nums[2])
conv4 = conv_block(conv3, 512, nums[3])
conv5 = conv_block(conv4, 512, nums[4])
fc_dim = 512
fc1 = paddle.layer.fc(input=conv5,
size=fc_dim,
act=paddle.activation.Relu(),
layer_attr=paddle.attr.Extra(drop_rate=0.5))
fc2 = paddle.layer.fc(input=fc1,
size=fc_dim,
act=paddle.activation.Relu(),
layer_attr=paddle.attr.Extra(drop_rate=0.5))
out = paddle.layer.fc(input=fc2,
size=class_dim,
act=paddle.activation.Softmax())
return out
def vgg13(input, class_dim):
nums = [2, 2, 2, 2, 2]
return vgg(input, nums, class_dim)
def vgg16(input, class_dim):
nums = [2, 2, 3, 3, 3]
return vgg(input, nums, class_dim)
def vgg19(input, class_dim):
nums = [2, 2, 4, 4, 4]
return vgg(input, nums, class_dim)
def main():
global ts
paddle.init(use_gpu=False)
image = paddle.layer.data(
name="image", type=paddle.data_type.dense_vector(DATA_DIM))
lbl = paddle.layer.data(
name="label", type=paddle.data_type.integer_value(CLASS_DIM))
extra_layers = None
# NOTE: for v2 distributed training need averaging updates.
learning_rate = 1e-3 / NODE_COUNT
out = vgg16(image, class_dim=CLASS_DIM)
cost = paddle.layer.classification_cost(input=out, label=lbl)
# Create parameters
parameters = paddle.parameters.create(cost)
# Create optimizer
optimizer = paddle.optimizer.Momentum(
momentum=0.9,
regularization=paddle.optimizer.L2Regularization(rate=0.0005 *
BATCH_SIZE),
learning_rate=learning_rate / BATCH_SIZE,
learning_rate_decay_a=0.1,
learning_rate_decay_b=128000 * 35,
learning_rate_schedule="discexp", )
train_reader = paddle.batch(
paddle.reader.shuffle(
cifar.train10(),
# To use other data, replace the above line with:
# reader.train_reader('train.list'),
buf_size=1000),
batch_size=BATCH_SIZE)
test_reader = paddle.batch(
cifar.test10(),
# To use other data, replace the above line with:
# reader.test_reader('val.list'),
batch_size=BATCH_SIZE)
# Create trainer
trainer = paddle.trainer.SGD(cost=cost,
parameters=parameters,
update_equation=optimizer,
extra_layers=extra_layers,
is_local=False)
# End batch and end pass event handler
def event_handler(event):
global ts, ts_pass
if isinstance(event, paddle.event.BeginPass):
ts_pass = time.time()
if isinstance(event, paddle.event.BeginIteration):
ts = time.time()
if isinstance(event, paddle.event.EndIteration):
if event.batch_id % 1 == 0:
print "\nPass %d, Batch %d, Cost %f, %s, spent: %f" % (
event.pass_id, event.batch_id, event.cost, event.metrics,
time.time() - ts)
if isinstance(event, paddle.event.EndPass):
print "Pass %d end, spent: %f" % (event.pass_id,
time.time() - ts_pass)
result = trainer.test(reader=test_reader)
print "\nTest with Pass %d, %s" % (event.pass_id, result.metrics)
trainer.train(
reader=train_reader, num_passes=200, event_handler=event_handler)
if __name__ == '__main__':
main()
benchmark/fluid/README.md 0 → 100644
浏览文件 @ fb3bfc3d
# Fluid Benchmark
This directory contains several models configurations and tools that used to run
Fluid benchmarks for local and distributed training.
## Run the Benchmark
To start, run the following command to get the full help message:
```bash
python fluid_benchmark.py --help
```
Currently supported `--model` argument include:
* mnist
* resnet
* you can chose to use different dataset using `--data_set cifar10` or
`--data_set flowers`.
* vgg
* stacked_dynamic_lstm
* machine_translation
* Run the following command to start a benchmark job locally:
```bash
python fluid_benchmark.py --model mnist --device GPU
```
You can choose to use GPU/CPU training. With GPU training, you can specify
`--gpus <gpu_num>` to run multi GPU training.
* Run distributed training with parameter servers:
* start parameter servers:
```bash
PADDLE_TRAINING_ROLE=PSERVER PADDLE_PSERVER_PORT=7164 PADDLE_PSERVER_IPS=127.0.0.1 PADDLE_TRAINERS=1 PADDLE_CURRENT_IP=127.0.0.1 PADDLE_TRAINER_ID=0 python fluid_benchmark.py --model mnist --device GPU --update_method pserver
```
* start trainers:
```bash
PADDLE_TRAINING_ROLE=TRAINER PADDLE_PSERVER_PORT=7164 PADDLE_PSERVER_IPS=127.0.0.1 PADDLE_TRAINERS=1 PADDLE_CURRENT_IP=127.0.0.1 PADDLE_TRAINER_ID=0 python fluid_benchmark.py --model mnist --device GPU --update_method pserver
```
* Run distributed training using NCCL2
```bash
PADDLE_PSERVER_PORT=7164 PADDLE_TRAINER_IPS=192.168.0.2,192.168.0.3 PADDLE_CURRENT_IP=127.0.0.1 PADDLE_TRAINER_ID=0 python fluid_benchmark.py --model mnist --device GPU --update_method nccl2
```
## Run Distributed Benchmark on Kubernetes Cluster
We provide a script `kube_gen_job.py` to generate Kubernetes yaml files to submit
distributed benchmark jobs to your cluster. To generate a job yaml, just run:
```bash
python kube_gen_job.py --jobname myjob --pscpu 4 --cpu 8 --gpu 8 --psmemory 20 --memory 40 --pservers 4 --trainers 4 --entry "python fluid_benchmark.py --model mnist --parallel 1 --device GPU --update_method pserver " --disttype pserver
```
Then the yaml files are generated under directory `myjob`, you can run:
```bash
kubectl create -f myjob/
```
The job shall start.
## Notes for Run Fluid Distributed with NCCL2 and RDMA
Before running NCCL2 distributed jobs, please check that whether your node has multiple network
interfaces, try to add the environment variable `export NCCL_SOCKET_IFNAME=eth0` to use your actual
network device.
To run high-performance distributed training, you must prepare your hardware environment to be
able to run RDMA enabled network communication, please check out [this](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/howto/cluster/nccl2_rdma_training.md)
note for details.
benchmark/fluid/fluid_benchmark.py 0 → 100644
浏览文件 @ fb3bfc3d
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import argparse
import cProfile
import time
import os
import numpy as np
import paddle.fluid as fluid
import paddle.fluid.core as core
import paddle.fluid.profiler as profiler
import paddle.fluid.transpiler.distribute_transpiler as distribute_transpiler
BENCHMARK_MODELS = [
"machine_translation", "resnet", "vgg", "mnist", "stacked_dynamic_lstm"
]
def parse_args():
parser = argparse.ArgumentParser('Fluid model benchmarks.')
parser.add_argument(
'--model',
type=str,
choices=BENCHMARK_MODELS,
default='resnet',
help='The model to run benchmark with.')
parser.add_argument(
'--batch_size', type=int, default=32, help='The minibatch size.')
parser.add_argument(
'--learning_rate',
type=float,
default=0.001,
help='The minibatch size.')
# TODO(wuyi): add "--use_fake_data" option back.
parser.add_argument(
'--skip_batch_num',
type=int,
default=5,
help='The first num of minibatch num to skip, for better performance test'
)
parser.add_argument(
'--iterations', type=int, default=80, help='The number of minibatches.')
parser.add_argument(
'--pass_num', type=int, default=100, help='The number of passes.')
parser.add_argument(
'--data_format',
type=str,
default='NCHW',
choices=['NCHW', 'NHWC'],
help='The data data_format, now only support NCHW.')
parser.add_argument(
'--device',
type=str,
default='GPU',
choices=['CPU', 'GPU'],
help='The device type.')
parser.add_argument(
'--gpus',
type=int,
default=1,
help='If gpus > 1, will use ParallelExecutor to run, else use Executor.')
parser.add_argument(
'--data_set',
type=str,
default='flowers',
choices=['cifar10', 'flowers'],
help='Optional dataset for benchmark.')
parser.add_argument(
'--infer_only', action='store_true', help='If set, run forward only.')
parser.add_argument(
'--use_cprof', action='store_true', help='If set, use cProfile.')
parser.add_argument(
'--use_nvprof',
action='store_true',
help='If set, use nvprof for CUDA.')
parser.add_argument(
'--no_test',
action='store_false',
help='If set, test the testset during training.')
parser.add_argument(
'--memory_optimize',
action='store_true',
help='If set, optimize runtime memory before start.')
parser.add_argument(
'--use_fake_data',
action='store_true',
help='If set ommit the actual read data operators.')
parser.add_argument(
'--profile', action='store_true', help='If set, profile a few steps.')
parser.add_argument(
'--update_method',
type=str,
default='local',
choices=['local', 'pserver', 'nccl2'],
help='Choose parameter update method, can be local, pserver, nccl2.')
args = parser.parse_args()
return args
def append_nccl2_prepare(trainer_id):
if trainer_id >= 0:
# append gen_nccl_id at the end of startup program
trainer_id = int(os.getenv("PADDLE_TRAINER_ID"))
port = os.getenv("PADDLE_PSERVER_PORT")
worker_ips = os.getenv("PADDLE_TRAINER_IPS")
worker_endpoints = []
for ip in worker_ips.split(","):
worker_endpoints.append(':'.join([ip, port]))
num_trainers = len(worker_endpoints)
current_endpoint = os.getenv("PADDLE_CURRENT_IP") + ":" + port
worker_endpoints.remove(current_endpoint)
nccl_id_var = fluid.default_startup_program().global_block().create_var(
name="NCCLID",
persistable=True,
type=fluid.core.VarDesc.VarType.RAW)
fluid.default_startup_program().global_block().append_op(
type="gen_nccl_id",
inputs={},
outputs={"NCCLID": nccl_id_var},
attrs={
"endpoint": current_endpoint,
"endpoint_list": worker_endpoints,
"trainer_id": trainer_id
})
return nccl_id_var, num_trainers, trainer_id
else:
raise Exception("must set positive PADDLE_TRAINER_ID env variables for "
"nccl-based dist train.")
def dist_transpile(trainer_id):
if trainer_id < 0:
return None, None
# the port of all pservers, needed by both trainer and pserver
port = os.getenv("PADDLE_PSERVER_PORT", "6174")
# comma separated ips of all pservers, needed by trainer and
# pserver
pserver_ips = os.getenv("PADDLE_PSERVER_IPS", "")
eplist = []
for ip in pserver_ips.split(","):
eplist.append(':'.join([ip, port]))
pserver_endpoints = ",".join(eplist)
# total number of workers/trainers in the job, needed by
# trainer and pserver
trainers = int(os.getenv("PADDLE_TRAINERS"))
# the IP of the local machine, needed by pserver only
current_endpoint = os.getenv("PADDLE_CURRENT_IP", "") + ":" + port
# the role, should be either PSERVER or TRAINER
training_role = os.getenv("PADDLE_TRAINING_ROLE")
t = distribute_transpiler.DistributeTranspiler()
t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers)
if training_role == "PSERVER":
pserver_program = t.get_pserver_program(current_endpoint)
pserver_startup_program = t.get_startup_program(current_endpoint,
pserver_program)
return pserver_program, pserver_startup_program
elif training_role == "TRAINER":
train_program = t.get_trainer_program()
return train_program, fluid.default_startup_program()
else:
raise ValueError(
'TRAINING_ROLE environment variable must be either TRAINER or PSERVER'
)
def test(exe, inference_program, test_reader, feeder, batch_acc):
accuracy_evaluator = fluid.metrics.Accuracy()
for batch_id, data in enumerate(test_reader()):
acc = exe.run(inference_program,
feed=feeder.feed(data),
fetch_list=[batch_acc])
accuracy_evaluator.update(value=np.array(acc), weight=len(data))
return accuracy_evaluator.eval()
# TODO(wuyi): replace train, train_parallel, test functions with new trainer
# API once it is ready.
def train(avg_loss, infer_prog, optimizer, train_reader, test_reader, batch_acc,
args, train_prog, startup_prog):
if os.getenv("PADDLE_TRAINING_ROLE") == "PSERVER":
place = core.CPUPlace()
exe = fluid.Executor(place)
exe.run(startup_prog)
exe.run(train_prog)
return
if args.use_fake_data:
raise Exception(
"fake data is not supported in single GPU test for now.")
place = core.CPUPlace() if args.device == 'CPU' else core.CUDAPlace(0)
exe = fluid.Executor(place)
exe.run(startup_prog)
feed_var_list = [
var for var in train_prog.global_block().vars.itervalues()
if var.is_data
]
feeder = fluid.DataFeeder(feed_var_list, place)
iters, num_samples, start_time = 0, 0, time.time()
for pass_id in range(args.pass_num):
train_losses = []
for batch_id, data in enumerate(train_reader()):
if iters == args.skip_batch_num:
start_time = time.time()
num_samples = 0
if iters == args.iterations:
break
loss = exe.run(train_prog,
feed=feeder.feed(data),
fetch_list=[avg_loss])
iters += 1
num_samples += len(data)
train_losses.append(loss)
print("Pass: %d, Iter: %d, Loss: %f\n" %
(pass_id, iters, np.mean(train_losses)))
train_elapsed = time.time() - start_time
examples_per_sec = num_samples / train_elapsed
print('\nTotal examples: %d, total time: %.5f, %.5f examples/sec\n' %
(num_samples, train_elapsed, examples_per_sec))
print("Pass: %d, Loss: %f" % (pass_id, np.mean(train_losses)))
# evaluation
if not args.no_test and batch_acc != None:
pass_test_acc = test(exe, infer_prog, test_reader, feeder,
batch_acc)
print(", Test Accuracy: %f" % pass_test_acc)
print("\n")
# TODO(wuyi): add warmup passes to get better perf data.
exit(0)
# TODO(wuyi): replace train, train_parallel, test functions with new trainer
# API once it is ready.
def train_parallel(avg_loss, infer_prog, optimizer, train_reader, test_reader,
batch_acc, args, train_prog, startup_prog, nccl_id_var,
num_trainers, trainer_id):
feed_var_list = [
var for var in train_prog.global_block().vars.itervalues()
if var.is_data
]
# generate fake:
if args.use_fake_data:
for var in feed_var_list:
v = startup_prog.global_block().clone_variable(var)
var.persistable = True
v.persistable = True
real_shape = list(var.shape)
real_shape[0] = args.batch_size / args.gpus
startup_prog.global_block().append_op(
outputs={"Out": v},
type="fill_constant",
attrs={"shape": real_shape,
"value": 1.0,
"dtype": var.dtype})
place = core.CPUPlace() if args.device == 'CPU' else core.CUDAPlace(0)
if nccl_id_var and trainer_id == 0:
#FIXME(wuyi): wait other trainer to start listening
time.sleep(30)
startup_exe = fluid.Executor(place)
startup_exe.run(startup_prog)
strategy = fluid.ExecutionStrategy()
strategy.num_threads = 1
strategy.allow_op_delay = False
exe = fluid.ParallelExecutor(
True,
avg_loss.name,
exec_strategy=strategy,
num_trainers=num_trainers,
trainer_id=trainer_id)
feeder = fluid.DataFeeder(feed_var_list, place)
for pass_id in range(args.pass_num):
num_samples = 0
iters = 0
start_time = time.time()
for batch_id, data in enumerate(train_reader()):
if args.profile and pass_id == 0 and batch_id == 5:
profiler.start_profiler("All")
elif args.profile and pass_id == 0 and batch_id == 10:
profiler.stop_profiler("total", "/tmp/profile_%d" % trainer_id)
if iters == args.skip_batch_num:
start_time = time.time()
num_samples = 0
if iters == args.iterations:
break
if args.use_fake_data:
loss, = exe.run([avg_loss.name])
else:
loss, = exe.run([avg_loss.name], feed=feeder.feed(data))
if args.update_method == "pserver":
exe.bcast_params()
num_samples += len(data)
iters += 1
if batch_id % 1 == 0:
print("Pass %d, batch %d, loss %s" %
(pass_id, batch_id, np.array(loss)))
train_elapsed = time.time() - start_time
examples_per_sec = num_samples / train_elapsed
print('\nTotal examples: %d, total time: %.5f, %.5f examples/sed\n' %
(num_samples, train_elapsed, examples_per_sec))
if not args.no_test and batch_acc != None:
test_acc = test(startup_exe, infer_prog, test_reader, feeder,
batch_acc)
print("Pass: %d, Test Accuracy: %f\n" % (pass_id, test_acc))
exit(0)
def print_arguments(args):
vars(args)['use_nvprof'] = (vars(args)['use_nvprof'] and
vars(args)['device'] == 'GPU')
print('----------- resnet Configuration Arguments -----------')
for arg, value in sorted(vars(args).iteritems()):
print('%s: %s' % (arg, value))
print('------------------------------------------------')
def main():
args = parse_args()
print_arguments(args)
# the unique trainer id, starting from 0, needed by trainer
# only
nccl_id_var, num_trainers, trainer_id = (
None, 1, int(os.getenv("PADDLE_TRAINER_ID", "-1")))
if args.use_cprof:
pr = cProfile.Profile()
pr.enable()
model_def = __import__("models.%s" % args.model, fromlist=["models"])
train_args = list(model_def.get_model(args))
train_args.append(args)
# Run optimizer.minimize(avg_loss)
train_args[2].minimize(train_args[0])
if args.memory_optimize:
fluid.memory_optimize(fluid.default_main_program())
if args.update_method == "pserver":
train_prog, startup_prog = dist_transpile(trainer_id)
if not train_prog:
raise Exception(
"Must configure correct environments to run dist train.")
train_args.extend([train_prog, startup_prog])
if args.gpus > 1 and os.getenv("PADDLE_TRAINING_ROLE") == "TRAINER":
train_args.extend([nccl_id_var, num_trainers, trainer_id])
train_parallel(*train_args)
train(*train_args)
exit(0)
# for other update methods, use default programs
train_args.append(fluid.default_main_program())
train_args.append(fluid.default_startup_program())
if args.update_method == "nccl2":
nccl_id_var, num_trainers, trainer_id = append_nccl2_prepare(trainer_id)
if args.gpus == 1:
# NOTE: parallel executor use profiler interanlly
if args.use_nvprof and args.device == 'GPU':
with profiler.cuda_profiler("cuda_profiler.txt", 'csv') as nvprof:
train(*train_args)
else:
train(*train_args)
else:
if args.device == "CPU":
raise Exception("Only support GPU perf with parallel exe")
train_args.extend([nccl_id_var, num_trainers, trainer_id])
train_parallel(*train_args)
if __name__ == "__main__":
main()
benchmark/fluid/kube_gen_job.py 0 → 100644
浏览文件 @ fb3bfc3d
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import yaml
import copy
import argparse
import random
import os
from kube_templates import pserver, trainer, envs
def parse_args():
parser = argparse.ArgumentParser(description='Generate dist job yamls.')
parser.add_argument(
'--jobname', default="paddlejob", help='unique job name')
parser.add_argument(
'--cpu', default=1, type=int, help='CPU cores per trainer node')
parser.add_argument(
'--pscpu', default=1, type=int, help='CPU cores per pserver node')
parser.add_argument(
'--gpu', default=0, type=int, help='num of GPUs per node')
parser.add_argument(
'--image',
default="bootstrapper:5000/fluid_benchmark:gpu",
help='num of GPUs per node')
parser.add_argument(
'--pservers', default=1, type=int, help='num of pservers')
parser.add_argument(
'--trainers', default=1, type=int, help='num of trainers')
parser.add_argument('--memory', default=1, type=int, help='trainer memory')
parser.add_argument(
'--psmemory', default=1, type=int, help='pserver memory')
parser.add_argument(
'--port', default=30236, type=int, help='num of trainers')
parser.add_argument(
'--entry', default="python train.py", help='command to run')
parser.add_argument(
'--fluid', default=1, type=int, help='whether is fluid job')
parser.add_argument(
'--rdma', action='store_ture', help='whether mount rdma libs')
parser.add_argument(
'--disttype',
default="pserver",
type=str,
choices=['pserver', 'nccl2', 'local'],
help='pserver or nccl2 or local')
args = parser.parse_args()
return args
def gen_job():
ps = pserver
tn = trainer
args = parse_args()
ps_container = ps["spec"]["template"]["spec"]["containers"][0]
tn_container = tn["spec"]["template"]["spec"]["containers"][0]
if args.fluid == 1:
ps_container["command"] = \
["paddle_k8s", "start_fluid"]
tn_container["command"] = \
["paddle_k8s", "start_fluid"]
ps["metadata"]["name"] = args.jobname + "-pserver"
ps["spec"]["template"]["metadata"]["labels"][
"paddle-job-pserver"] = args.jobname
tn["metadata"]["name"] = args.jobname + "-trainer"
tn["spec"]["template"]["metadata"]["labels"]["paddle-job"] = args.jobname
ps_container["image"] = args.image
tn_container["image"] = args.image
ps_container["resources"]["requests"]["cpu"] = str(args.pscpu)
ps_container["resources"]["requests"]["memory"] = str(args.psmemory) + "Gi"
ps_container["resources"]["limits"]["cpu"] = str(args.pscpu)
ps_container["resources"]["limits"]["memory"] = str(args.psmemory) + "Gi"
tn_container["resources"]["requests"]["cpu"] = str(args.cpu)
tn_container["resources"]["requests"]["memory"] = str(args.memory) + "Gi"
tn_container["resources"]["limits"]["cpu"] = str(args.cpu)
tn_container["resources"]["limits"]["memory"] = str(args.memory) + "Gi"
if args.gpu > 0:
tn_container["resources"]["requests"][
"alpha.kubernetes.io/nvidia-gpu"] = str(args.gpu)
tn_container["resources"]["limits"][
"alpha.kubernetes.io/nvidia-gpu"] = str(args.gpu)
ps["spec"]["replicas"] = int(args.pservers)
tn["spec"]["parallelism"] = int(args.trainers)
tn["spec"]["completions"] = int(args.trainers)
ps_container["ports"][0]["name"] = "jobport-" + str(args.port)
ps_container["ports"][0]["containerPort"] = args.port
spreadport = random.randint(40000, 60000)
tn_container["ports"][0]["name"] = "spr-" + str(spreadport)
tn_container["ports"][0]["containerPort"] = spreadport
envs.append({"name": "PADDLE_JOB_NAME", "value": args.jobname})
envs.append({"name": "TRAINERS", "value": str(args.trainers)})
envs.append({"name": "PSERVERS", "value": str(args.pservers)})
envs.append({"name": "ENTRY", "value": args.entry})
envs.append({"name": "PADDLE_INIT_PORT", "value": str(args.port)})
envs.append({"name": "PADDLE_PSERVER_PORT", "value": str(args.port)})
# NOTE: these directories below are cluster specific, please modify
# this settings before you run on your own cluster.
envs.append({
"name": "LD_LIBRARY_PATH",
"value":
"/usr/local/lib:/usr/local/nvidia/lib64:/usr/local/rdma/lib64:/usr/lib64/mlnx_ofed/valgrind"
})
volumes = [{
"name": "nvidia-driver",
"hostPath": {
"path": "/usr/local/nvidia/lib64"
}
}]
volumeMounts = [{
"mountPath": "/usr/local/nvidia/lib64",
"name": "nvidia-driver"
}]
if args.rdma:
volumes.extend([{
"name": "ibetc",
"hostPath": {
"path": "/etc/libibverbs.d"
}
}, {
"name": "iblibs",
"hostPath": {
"path": "/usr/local/rdma"
}
}, {
"name": "valgrind",
"hostPath": {
"path": "/usr/lib64/mlnx_ofed/valgrind"
}
}])
volumeMounts.extend([{
"mountPath": "/etc/libibverbs.d",
"name": "ibetc"
}, {
"mountPath": "/usr/local/rdma",
"name": "iblibs"
}, {
"mountPath": "/usr/lib64/mlnx_ofed/valgrind",
"name": "valgrind"
}])
# append shm for NCCL2
volumes.append({"name": "dshm", "emptyDir": {"medium": "Memory"}})
volumeMounts.append({"mountPath": "/dev/shm", "name": "dshm"})
tn["spec"]["template"]["spec"]["volumes"] = volumes
tn_container["volumeMounts"] = volumeMounts
ps_container["env"] = envs
ps_container["env"].append({"name": "TRAINING_ROLE", "value": "PSERVER"})
tn_container["env"] = envs
if args.disttype == "pserver":
tn_container["env"].append({
"name": "TRAINING_ROLE",
"value": "TRAINER"
})
elif args.disttype == "nccl2" or args.disttype == "local":
# NCCL2 have no training role, set to plain WORKER
tn_container["env"].append({"name": "TRAINING_ROLE", "value": "WORKER"})
os.mkdir(args.jobname)
if args.disttype == "pserver":
with open("%s/pserver.yaml" % args.jobname, "w") as fn:
yaml.dump(ps, fn)
with open("%s/trainer.yaml" % args.jobname, "w") as fn:
yaml.dump(tn, fn)
if __name__ == "__main__":
gen_job()
benchmark/fluid/kube_templates/__init__.py 0 → 100644
浏览文件 @ fb3bfc3d
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from pserver import pserver
from trainer import trainer
__all__ = ["pserver", "trainer", "envs"]
envs = [
# envs that don't need to change
{
"name": "GLOG_v",
"value": "0"
},
{
"name": "GLOG_logtostderr",
"value": "1"
},
{
"name": "TOPOLOGY",
"value": ""
},
{
"name": "TRAINER_PACKAGE",
"value": "/workspace"
},
{
"name": "PADDLE_INIT_NICS",
"value": "eth2"
},
{
"name": "NAMESPACE",
"valueFrom": {
"fieldRef": {
"fieldPath": "metadata.namespace"
}
}
},
{
"name": "POD_IP",
"valueFrom": {
"fieldRef": {
"fieldPath": "status.podIP"
}
}
},
{
"name": "PADDLE_CURRENT_IP",
"valueFrom": {
"fieldRef": {
"fieldPath": "status.podIP"
}
}
}
]
benchmark/fluid/kube_templates/pserver.py 0 → 100644
浏览文件 @ fb3bfc3d
# Copyright (c) 2018 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.
pserver = {
"apiVersion": "extensions/v1beta1",
"kind": "ReplicaSet",
"metadata": {
"name": "jobname-pserver"
},
"spec": {
"replicas": 1,
"template": {
"metadata": {
"labels": {
"paddle-job-pserver": "jobname"
}
},
"spec": {
"hostNetwork": True,
"imagePullSecrets": [{
"name": "job-registry-secret"
}],
"containers": [{
"name": "pserver",
"image": "",
"imagePullPolicy": "Always",
"ports": [{
"name": "jobport-1",
"containerPort": 1
}],
"env": [],
"command": ["paddle_k8s", "start_pserver"],
"resources": {
"requests": {
"memory": "10Gi",
"cpu": "4"
},
"limits": {
"memory": "10Gi",
"cpu": "4"
}
}
}]
}
}
}
}
benchmark/fluid/kube_templates/trainer.py 0 → 100644
浏览文件 @ fb3bfc3d
# Copyright (c) 2018 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.
trainer = {
"apiVersion": "batch/v1",
"kind": "Job",
"metadata": {
"name": "jobname-pserver"
},
"spec": {
"parallelism": 4,
"completions": 4,
"template": {
"metadata": {
"labels": {
"paddle-job": "jobname"
}
},
"spec": {
"hostNetwork": True,
"imagePullSecrets": [{
"name": "job-registry-secret"
}],
"restartPolicy": "Never",
"containers": [{
"name": "trainer",
"image": "",
"imagePullPolicy": "Always",
# to let container set rlimit
"securityContext": {
"privileged": True
# TODO(wuyi): use below specific cap instead of privileged,
# using privileged will cause all GPU device are visible
# in the container.
# "capabilities": {
# "add": ["SYS_RESOURCE"]
# }
},
"ports": [{
"name": "jobport-1",
"containerPort": 1
}],
"env": [],
"command": ["paddle_k8s", "start_trainer", "v2"],
"resources": {
"requests": {
"memory": "10Gi",
"cpu": "4",
},
"limits": {
"memory": "10Gi",
"cpu": "4",
}
}
}]
}
}
}
}
benchmark/fluid/models/__init__.py 0 → 100644
浏览文件 @ fb3bfc3d
# Copyright (c) 2018 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.
__all__ = [
"machine_translation", "resnet", "vgg", "mnist", "stacked_dynamic_lstm"
]
benchmark/fluid/machine_translation.py → benchmark/fluid/models/machine_translation.py
浏览文件 @ fb3bfc3d
......@@ -27,74 +27,6 @@ import paddle.fluid.core as core
import paddle.fluid.framework as framework
from paddle.fluid.executor import Executor
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument(
"--embedding_dim",
type=int,
default=512,
help="The dimension of embedding table. (default: %(default)d)")
parser.add_argument(
"--encoder_size",
type=int,
default=512,
help="The size of encoder bi-rnn unit. (default: %(default)d)")
parser.add_argument(
"--decoder_size",
type=int,
default=512,
help="The size of decoder rnn unit. (default: %(default)d)")
parser.add_argument(
"--batch_size",
type=int,
default=16,
help="The sequence number of a mini-batch data. (default: %(default)d)")
parser.add_argument(
'--skip_batch_num',
type=int,
default=5,
help='The first num of minibatch num to skip, for better performance test')
parser.add_argument(
'--iterations', type=int, default=80, help='The number of minibatches.')
parser.add_argument(
"--dict_size",
type=int,
default=30000,
help="The dictionary capacity. Dictionaries of source sequence and "
"target dictionary have same capacity. (default: %(default)d)")
parser.add_argument(
"--pass_num",
type=int,
default=2,
help="The pass number to train. (default: %(default)d)")
parser.add_argument(
"--learning_rate",
type=float,
default=0.0002,
help="Learning rate used to train the model. (default: %(default)f)")
parser.add_argument(
"--infer_only", action='store_true', help="If set, run forward only.")
parser.add_argument(
"--beam_size",
type=int,
default=3,
help="The width for beam searching. (default: %(default)d)")
parser.add_argument(
'--device',
type=str,
default='GPU',
choices=['CPU', 'GPU'],
help="The device type.")
parser.add_argument(
"--max_length",
type=int,
default=250,
help="The maximum length of sequence when doing generation. "
"(default: %(default)d)")
parser.add_argument(
'--with_test',
action='store_true',
help='If set, test the testset during training.')
def lstm_step(x_t, hidden_t_prev, cell_t_prev, size):
def linear(inputs):
......@@ -264,116 +196,37 @@ def lodtensor_to_ndarray(lod_tensor):
return ndarray
def train():
def get_model(args):
embedding_dim = 512
encoder_size = 512
decoder_size = 512
dict_size = 30000
beam_size = 3
max_length = 250
avg_cost, feeding_list = seq_to_seq_net(
args.embedding_dim,
args.encoder_size,
args.decoder_size,
args.dict_size,
args.dict_size,
embedding_dim,
encoder_size,
decoder_size,
dict_size,
dict_size,
False,
beam_size=args.beam_size,
max_length=args.max_length)
beam_size=beam_size,
max_length=max_length)
# clone from default main program
inference_program = fluid.default_main_program().clone()
optimizer = fluid.optimizer.Adam(learning_rate=args.learning_rate)
optimizer.minimize(avg_cost)
fluid.memory_optimize(fluid.default_main_program())
train_batch_generator = paddle.batch(
paddle.reader.shuffle(
paddle.dataset.wmt14.train(args.dict_size), buf_size=1000),
paddle.dataset.wmt14.train(dict_size), buf_size=1000),
batch_size=args.batch_size)
test_batch_generator = paddle.batch(
paddle.reader.shuffle(
paddle.dataset.wmt14.test(args.dict_size), buf_size=1000),
paddle.dataset.wmt14.test(dict_size), buf_size=1000),
batch_size=args.batch_size)
place = core.CPUPlace() if args.device == 'CPU' else core.CUDAPlace(0)
exe = Executor(place)
exe.run(framework.default_startup_program())
def do_validation():
total_loss = 0.0
count = 0
for batch_id, data in enumerate(test_batch_generator()):
src_seq = to_lodtensor(map(lambda x: x[0], data), place)[0]
trg_seq = to_lodtensor(map(lambda x: x[1], data), place)[0]
lbl_seq = to_lodtensor(map(lambda x: x[2], data), place)[0]
fetch_outs = exe.run(inference_program,
feed={
feeding_list[0]: src_seq,
feeding_list[1]: trg_seq,
feeding_list[2]: lbl_seq
},
fetch_list=[avg_cost],
return_numpy=False)
total_loss += lodtensor_to_ndarray(fetch_outs[0])[0]
count += 1
return total_loss / count
iters, num_samples, start_time = 0, 0, time.time()
for pass_id in xrange(args.pass_num):
train_accs = []
train_losses = []
for batch_id, data in enumerate(train_batch_generator()):
if iters == args.skip_batch_num:
start_time = time.time()
num_samples = 0
if iters == args.iterations:
break
src_seq, word_num = to_lodtensor(map(lambda x: x[0], data), place)
num_samples += word_num
trg_seq, word_num = to_lodtensor(map(lambda x: x[1], data), place)
num_samples += word_num
lbl_seq, _ = to_lodtensor(map(lambda x: x[2], data), place)
fetch_outs = exe.run(framework.default_main_program(),
feed={
feeding_list[0]: src_seq,
feeding_list[1]: trg_seq,
feeding_list[2]: lbl_seq
},
fetch_list=[avg_cost])
iters += 1
loss = np.array(fetch_outs[0])
print(
"Pass = %d, Iter = %d, Loss = %f" % (pass_id, iters, loss)
) # The accuracy is the accumulation of batches, but not the current batch.
train_elapsed = time.time() - start_time
examples_per_sec = num_samples / train_elapsed
print('\nTotal examples: %d, total time: %.5f, %.5f examples/sed\n' %
(num_samples, train_elapsed, examples_per_sec))
# evaluation
if args.with_test:
test_loss = do_validation()
exit(0)
def infer():
pass
def print_arguments(args):
print('----------- seq2seq Configuration Arguments -----------')
for arg, value in sorted(vars(args).iteritems()):
print('%s: %s' % (arg, value))
print('------------------------------------------------')
if __name__ == '__main__':
args = parser.parse_args()
print_arguments(args)
if args.infer_only:
infer()
else:
train()
return avg_cost, inference_program, optimizer, train_batch_generator, \
test_batch_generator, None
benchmark/fluid/mnist.py → benchmark/fluid/models/mnist.py
浏览文件 @ fb3bfc3d
......@@ -19,6 +19,7 @@ from __future__ import print_function
import numpy as np
import argparse
import time
import cProfile
import paddle
import paddle.fluid as fluid
......@@ -31,42 +32,6 @@ DTYPE = "float32"
# fluid.default_startup_program().random_seed = SEED
def parse_args():
parser = argparse.ArgumentParser("mnist model benchmark.")
parser.add_argument(
'--batch_size', type=int, default=128, help='The minibatch size.')
parser.add_argument(
'--skip_batch_num',
type=int,
default=5,
help='The first num of minibatch num to skip, for better performance test'
)
parser.add_argument(
'--iterations', type=int, default=35, help='The number of minibatches.')
parser.add_argument(
'--pass_num', type=int, default=5, help='The number of passes.')
parser.add_argument(
'--device',
type=str,
default='GPU',
choices=['CPU', 'GPU'],
help='The device type.')
parser.add_argument(
'--infer_only', action='store_true', help='If set, run forward only.')
parser.add_argument(
'--use_cprof', action='store_true', help='If set, use cProfile.')
parser.add_argument(
'--use_nvprof',
action='store_true',
help='If set, use nvprof for CUDA.')
parser.add_argument(
'--with_test',
action='store_true',
help='If set, test the testset during training.')
args = parser.parse_args()
return args
def cnn_model(data):
conv_pool_1 = fluid.nets.simple_img_conv_pool(
input=data,
......@@ -99,36 +64,13 @@ def cnn_model(data):
return predict
def eval_test(exe, batch_acc, batch_size_tensor, inference_program):
test_reader = paddle.batch(
paddle.dataset.mnist.test(), batch_size=args.batch_size)
test_pass_acc = fluid.average.WeightedAverage()
for batch_id, data in enumerate(test_reader()):
img_data = np.array(map(lambda x: x[0].reshape([1, 28, 28]),
data)).astype(DTYPE)
y_data = np.array(map(lambda x: x[1], data)).astype("int64")
y_data = y_data.reshape([len(y_data), 1])
acc, weight = exe.run(inference_program,
feed={"pixel": img_data,
"label": y_data},
fetch_list=[batch_acc, batch_size_tensor])
test_pass_acc.add(value=acc, weight=weight)
pass_acc = test_pass_acc.eval()
return pass_acc
def run_benchmark(model, args):
if args.use_cprof:
pr = cProfile.Profile()
pr.enable()
start_time = time.time()
def get_model(args):
# Input data
images = fluid.layers.data(name='pixel', shape=[1, 28, 28], dtype=DTYPE)
label = fluid.layers.data(name='label', shape=[1], dtype='int64')
# Train program
predict = model(images)
predict = cnn_model(images)
cost = fluid.layers.cross_entropy(input=predict, label=label)
avg_cost = fluid.layers.mean(x=cost)
......@@ -143,82 +85,10 @@ def run_benchmark(model, args):
# Optimization
opt = fluid.optimizer.AdamOptimizer(
learning_rate=0.001, beta1=0.9, beta2=0.999)
opt.minimize(avg_cost)
fluid.memory_optimize(fluid.default_main_program())
# Initialize executor
place = fluid.CPUPlace() if args.device == 'CPU' else fluid.CUDAPlace(0)
exe = fluid.Executor(place)
# Parameter initialization
exe.run(fluid.default_startup_program())
# Reader
train_reader = paddle.batch(
paddle.dataset.mnist.train(), batch_size=args.batch_size)
accuracy = fluid.metrics.Accuracy()
iters, num_samples, start_time = 0, 0, time.time()
for pass_id in range(args.pass_num):
accuracy.reset()
train_accs = []
train_losses = []
for batch_id, data in enumerate(train_reader()):
if iters == args.skip_batch_num:
start_time = time.time()
num_samples = 0
if iters == args.iterations:
break
img_data = np.array(
map(lambda x: x[0].reshape([1, 28, 28]), data)).astype(DTYPE)
y_data = np.array(map(lambda x: x[1], data)).astype("int64")
y_data = y_data.reshape([len(y_data), 1])
outs = exe.run(
fluid.default_main_program(),
feed={"pixel": img_data,
"label": y_data},
fetch_list=[avg_cost, batch_acc, batch_size_tensor]
) # The accuracy is the accumulation of batches, but not the current batch.
accuracy.update(value=outs[1], weight=outs[2])
iters += 1
num_samples += len(y_data)
loss = np.array(outs[0])
acc = np.array(outs[1])
train_losses.append(loss)
train_accs.append(acc)
print("Pass: %d, Iter: %d, Loss: %f, Accuracy: %f" %
(pass_id, iters, loss, acc))
print("Pass: %d, Loss: %f, Train Accuray: %f\n" %
(pass_id, np.mean(train_losses), np.mean(train_accs)))
train_elapsed = time.time() - start_time
examples_per_sec = num_samples / train_elapsed
print('\nTotal examples: %d, total time: %.5f, %.5f examples/sed\n' %
(num_samples, train_elapsed, examples_per_sec))
# evaluation
if args.with_test:
test_avg_acc = eval_test(exe, batch_acc, batch_size_tensor,
inference_program)
exit(0)
def print_arguments(args):
vars(args)['use_nvprof'] = (vars(args)['use_nvprof'] and
vars(args)['device'] == 'GPU')
print('----------- mnist Configuration Arguments -----------')
for arg, value in sorted(vars(args).iteritems()):
print('%s: %s' % (arg, value))
print('------------------------------------------------')
if __name__ == '__main__':
args = parse_args()
print_arguments(args)
if args.use_nvprof and args.device == 'GPU':
with profiler.cuda_profiler("cuda_profiler.txt", 'csv') as nvprof:
run_benchmark(cnn_model, args)
else:
run_benchmark(cnn_model, args)
test_reader = paddle.batch(
paddle.dataset.mnist.test(), batch_size=args.batch_size)
return avg_cost, inference_program, opt, train_reader, test_reader, batch_acc
benchmark/fluid/resnet.py → benchmark/fluid/models/resnet.py
浏览文件 @ fb3bfc3d
......@@ -16,7 +16,6 @@ from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import argparse
import functools
import numpy as np
import time
......@@ -29,64 +28,6 @@ import paddle.fluid.core as core
import paddle.fluid.profiler as profiler
def parse_args():
parser = argparse.ArgumentParser('Convolution model benchmark.')
parser.add_argument(
'--model',
type=str,
choices=['resnet_imagenet', 'resnet_cifar10'],
default='resnet_imagenet',
help='The model architecture.')
parser.add_argument(
'--batch_size', type=int, default=32, help='The minibatch size.')
parser.add_argument(
'--use_fake_data',
action='store_true',
help='use real data or fake data')
parser.add_argument(
'--skip_batch_num',
type=int,
default=5,
help='The first num of minibatch num to skip, for better performance test'
)
parser.add_argument(
'--iterations', type=int, default=80, help='The number of minibatches.')
parser.add_argument(
'--pass_num', type=int, default=100, help='The number of passes.')
parser.add_argument(
'--data_format',
type=str,
default='NCHW',
choices=['NCHW', 'NHWC'],
help='The data data_format, now only support NCHW.')
parser.add_argument(
'--device',
type=str,
default='GPU',
choices=['CPU', 'GPU'],
help='The device type.')
parser.add_argument(
'--data_set',
type=str,
default='flowers',
choices=['cifar10', 'flowers'],
help='Optional dataset for benchmark.')
parser.add_argument(
'--infer_only', action='store_true', help='If set, run forward only.')
parser.add_argument(
'--use_cprof', action='store_true', help='If set, use cProfile.')
parser.add_argument(
'--use_nvprof',
action='store_true',
help='If set, use nvprof for CUDA.')
parser.add_argument(
'--with_test',
action='store_true',
help='If set, test the testset during training.')
args = parser.parse_args()
return args
def conv_bn_layer(input, ch_out, filter_size, stride, padding, act='relu'):
conv1 = fluid.layers.conv2d(
input=input,
......@@ -100,7 +41,7 @@ def conv_bn_layer(input, ch_out, filter_size, stride, padding, act='relu'):
def shortcut(input, ch_out, stride):
ch_in = input.shape[1] if args.data_format == 'NCHW' else input.shape[-1]
ch_in = input.shape[1] # if args.data_format == 'NCHW' else input.shape[-1]
if ch_in != ch_out:
return conv_bn_layer(input, ch_out, 1, stride, 0, None)
else:
......@@ -172,23 +113,22 @@ def resnet_cifar10(input, class_dim, depth=32, data_format='NCHW'):
return out
def run_benchmark(model, args):
if args.use_cprof:
pr = cProfile.Profile()
pr.enable()
def get_model(args):
model = resnet_cifar10
if args.data_set == "cifar10":
class_dim = 10
if args.data_format == 'NCHW':
dshape = [3, 32, 32]
else:
dshape = [32, 32, 3]
model = resnet_cifar10
else:
class_dim = 102
if args.data_format == 'NCHW':
dshape = [3, 224, 224]
else:
dshape = [224, 224, 3]
model = resnet_imagenet
input = fluid.layers.data(name='data', shape=dshape, dtype='float32')
label = fluid.layers.data(name='label', shape=[1], dtype='int64')
......@@ -206,9 +146,6 @@ def run_benchmark(model, args):
target_vars=[batch_acc, batch_size_tensor])
optimizer = fluid.optimizer.Momentum(learning_rate=0.01, momentum=0.9)
opts = optimizer.minimize(avg_cost)
fluid.memory_optimize(fluid.default_main_program())
train_reader = paddle.batch(
paddle.reader.shuffle(
......@@ -221,93 +158,4 @@ def run_benchmark(model, args):
if args.data_set == 'cifar10' else paddle.dataset.flowers.test(),
batch_size=args.batch_size)
def test(exe):
test_accuracy = fluid.average.WeightedAverage()
for batch_id, data in enumerate(test_reader()):
img_data = np.array(map(lambda x: x[0].reshape(dshape),
data)).astype("float32")
y_data = np.array(map(lambda x: x[1], data)).astype("int64")
y_data = y_data.reshape([-1, 1])
acc, weight = exe.run(inference_program,
feed={"data": img_data,
"label": y_data},
fetch_list=[batch_acc, batch_size_tensor])
test_accuracy.add(value=acc, weight=weight)
return test_accuracy.eval()
place = core.CPUPlace() if args.device == 'CPU' else core.CUDAPlace(0)
exe = fluid.Executor(place)
exe.run(fluid.default_startup_program())
accuracy = fluid.average.WeightedAverage()
if args.use_fake_data:
data = train_reader().next()
image = np.array(map(lambda x: x[0].reshape(dshape), data)).astype(
'float32')
label = np.array(map(lambda x: x[1], data)).astype('int64')
label = label.reshape([-1, 1])
iters, num_samples, start_time = 0, 0, time.time()
for pass_id in range(args.pass_num):
accuracy.reset()
train_accs = []
train_losses = []
for batch_id, data in enumerate(train_reader()):
if iters == args.skip_batch_num:
start_time = time.time()
num_samples = 0
if iters == args.iterations:
break
if not args.use_fake_data:
image = np.array(map(lambda x: x[0].reshape(dshape),
data)).astype('float32')
label = np.array(map(lambda x: x[1], data)).astype('int64')
label = label.reshape([-1, 1])
loss, acc, weight = exe.run(
fluid.default_main_program(),
feed={'data': image,
'label': label},
fetch_list=[avg_cost, batch_acc, batch_size_tensor])
iters += 1
num_samples += len(label)
accuracy.add(value=acc, weight=weight)
train_losses.append(loss)
train_accs.append(acc)
print("Pass: %d, Iter: %d, Loss: %f, Accuracy: %f" %
(pass_id, iters, loss, acc))
print("Pass: %d, Loss: %f, Train Accuray: %f\n" %
(pass_id, np.mean(train_losses), np.mean(train_accs)))
train_elapsed = time.time() - start_time
examples_per_sec = num_samples / train_elapsed
print('\nTotal examples: %d, total time: %.5f, %.5f examples/sed\n' %
(num_samples, train_elapsed, examples_per_sec))
# evaluation
if args.with_test:
pass_test_acc = test(exe)
exit(0)
def print_arguments(args):
vars(args)['use_nvprof'] = (vars(args)['use_nvprof'] and
vars(args)['device'] == 'GPU')
print('----------- resnet Configuration Arguments -----------')
for arg, value in sorted(vars(args).iteritems()):
print('%s: %s' % (arg, value))
print('------------------------------------------------')
if __name__ == '__main__':
model_map = {
'resnet_imagenet': resnet_imagenet,
'resnet_cifar10': resnet_cifar10
}
args = parse_args()
print_arguments(args)
if args.data_format == 'NHWC':
raise ValueError('Only support NCHW data_format now.')
if args.use_nvprof and args.device == 'GPU':
with profiler.cuda_profiler("cuda_profiler.txt", 'csv') as nvprof:
run_benchmark(model_map[args.model], args)
else:
run_benchmark(model_map[args.model], args)
return avg_cost, inference_program, optimizer, train_reader, test_reader, batch_acc
benchmark/fluid/stacked_dynamic_lstm.py → benchmark/fluid/models/stacked_dynamic_lstm.py
浏览文件 @ fb3bfc3d
......@@ -29,57 +29,6 @@ import paddle.fluid as fluid
import paddle.batch as batch
import paddle.fluid.profiler as profiler
def parse_args():
parser = argparse.ArgumentParser("Understand Sentiment by Dynamic RNN.")
parser.add_argument(
'--batch_size',
type=int,
default=32,
help='The sequence number of a batch data. (default: %(default)d)')
parser.add_argument(
'--skip_batch_num',
type=int,
default=5,
help='The first num of minibatch num to skip, for better performance test'
)
parser.add_argument(
'--iterations', type=int, default=80, help='The number of minibatches.')
parser.add_argument(
'--emb_dim',
type=int,
default=512,
help='Dimension of embedding table. (default: %(default)d)')
parser.add_argument(
'--hidden_dim',
type=int,
default=512,
help='Hidden size of lstm unit. (default: %(default)d)')
parser.add_argument(
'--pass_num',
type=int,
default=100,
help='Epoch number to train. (default: %(default)d)')
parser.add_argument(
'--device',
type=str,
default='CPU',
choices=['CPU', 'GPU'],
help='The device type.')
parser.add_argument(
'--crop_size',
type=int,
default=int(os.environ.get('CROP_SIZE', '1500')),
help='The max sentence length of input. Since this model use plain RNN,'
' Gradient could be explored if sentence is too long')
parser.add_argument(
'--with_test',
action='store_true',
help='If set, test the testset during training.')
args = parser.parse_args()
return args
word_dict = imdb.word_dict()
......@@ -94,14 +43,15 @@ def crop_sentence(reader, crop_size):
return __impl__
def main():
args = parse_args()
lstm_size = args.hidden_dim
def get_model(args):
lstm_size = 512
emb_dim = 512
crop_size = 1500
data = fluid.layers.data(
name="words", shape=[1], lod_level=1, dtype='int64')
sentence = fluid.layers.embedding(
input=data, size=[len(word_dict), args.emb_dim])
input=data, size=[len(word_dict), emb_dim])
sentence = fluid.layers.fc(input=sentence, size=lstm_size, act='tanh')
......@@ -161,51 +111,17 @@ def main():
target_vars=[batch_acc, batch_size_tensor])
adam = fluid.optimizer.Adam()
adam.minimize(loss)
fluid.memory_optimize(fluid.default_main_program())
place = fluid.CPUPlace() if args.device == 'CPU' else fluid.CUDAPlace(0)
exe = fluid.Executor(place)
exe.run(fluid.default_startup_program())
train_reader = batch(
paddle.reader.shuffle(
crop_sentence(imdb.train(word_dict), args.crop_size),
buf_size=25000),
crop_sentence(imdb.train(word_dict), crop_size), buf_size=25000),
batch_size=args.batch_size)
test_reader = batch(
paddle.reader.shuffle(
crop_sentence(imdb.test(word_dict), crop_size), buf_size=25000),
batch_size=args.batch_size)
iters, num_samples, start_time = 0, 0, time.time()
for pass_id in range(args.pass_num):
train_accs = []
train_losses = []
for batch_id, data in enumerate(train_reader()):
if iters == args.skip_batch_num:
start_time = time.time()
num_samples = 0
if iters == args.iterations:
break
tensor_words = to_lodtensor([x[0] for x in data], place)
label = numpy.array([x[1] for x in data]).astype("int64")
label = label.reshape((-1, 1))
loss_np, acc, weight = exe.run(
fluid.default_main_program(),
feed={"words": tensor_words,
"label": label},
fetch_list=[loss, batch_acc, batch_size_tensor])
iters += 1
for x in data:
num_samples += len(x[0])
print(
"Pass = %d, Iter = %d, Loss = %f, Accuracy = %f" %
(pass_id, iters, loss_np, acc)
) # The accuracy is the accumulation of batches, but not the current batch.
train_elapsed = time.time() - start_time
examples_per_sec = num_samples / train_elapsed
print('\nTotal examples: %d, total time: %.5f, %.5f examples/sed\n' %
(num_samples, train_elapsed, examples_per_sec))
exit(0)
return loss, inference_program, adam, train_reader, test_reader, batch_acc
def to_lodtensor(data, place):
......@@ -221,16 +137,3 @@ def to_lodtensor(data, place):
res.set(flattened_data, place)
res.set_lod([lod])
return res
def print_arguments(args):
print('----------- lstm Configuration Arguments -----------')
for arg, value in sorted(vars(args).iteritems()):
print('%s: %s' % (arg, value))
print('------------------------------------------------')
if __name__ == '__main__':
args = parse_args()
print_arguments(args)
main()
benchmark/fluid/vgg.py → benchmark/fluid/models/vgg.py
浏览文件 @ fb3bfc3d
......@@ -23,46 +23,6 @@ import paddle.fluid.core as core
import argparse
import functools
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument(
'--batch_size', type=int, default=128, help="Batch size for training.")
parser.add_argument(
'--skip_batch_num',
type=int,
default=5,
help='The first num of minibatch num to skip, for better performance test')
parser.add_argument(
'--iterations', type=int, default=80, help='The number of minibatches.')
parser.add_argument(
'--learning_rate',
type=float,
default=1e-3,
help="Learning rate for training.")
parser.add_argument('--pass_num', type=int, default=50, help="No. of passes.")
parser.add_argument(
'--device',
type=str,
default='GPU',
choices=['CPU', 'GPU'],
help="The device type.")
parser.add_argument(
'--data_format',
type=str,
default='NCHW',
choices=['NCHW', 'NHWC'],
help='The data order, now only support NCHW.')
parser.add_argument(
'--data_set',
type=str,
default='cifar10',
choices=['cifar10', 'flowers'],
help='Optional dataset for benchmark.')
parser.add_argument(
'--with_test',
action='store_true',
help='If set, test the testset during training.')
args = parser.parse_args()
def vgg16_bn_drop(input):
def conv_block(input, num_filter, groups, dropouts):
......@@ -91,7 +51,7 @@ def vgg16_bn_drop(input):
return fc2
def main():
def get_model(args):
if args.data_set == "cifar10":
classdim = 10
if args.data_format == 'NCHW':
......@@ -128,16 +88,6 @@ def main():
# Optimization
optimizer = fluid.optimizer.Adam(learning_rate=args.learning_rate)
opts = optimizer.minimize(avg_cost)
fluid.memory_optimize(fluid.default_main_program())
# Initialize executor
place = core.CPUPlace() if args.device == 'CPU' else core.CUDAPlace(0)
exe = fluid.Executor(place)
# Parameter initialization
exe.run(fluid.default_startup_program())
# data reader
train_reader = paddle.batch(
......@@ -151,74 +101,4 @@ def main():
if args.data_set == 'cifar10' else paddle.dataset.flowers.test(),
batch_size=args.batch_size)
# test
def test(exe):
test_accuracy = fluid.average.WeightedAverage()
for batch_id, data in enumerate(test_reader()):
img_data = np.array(map(lambda x: x[0].reshape(data_shape),
data)).astype("float32")
y_data = np.array(map(lambda x: x[1], data)).astype("int64")
y_data = y_data.reshape([-1, 1])
acc, weight = exe.run(inference_program,
feed={"pixel": img_data,
"label": y_data},
fetch_list=[batch_acc, batch_size_tensor])
test_accuracy.add(value=acc, weight=weight)
return test_accuracy.eval()
iters, num_samples, start_time = 0, 0, time.time()
accuracy = fluid.average.WeightedAverage()
for pass_id in range(args.pass_num):
accuracy.reset()
train_accs = []
train_losses = []
for batch_id, data in enumerate(train_reader()):
if iters == args.skip_batch_num:
start_time = time.time()
num_samples = 0
if iters == args.iterations:
break
img_data = np.array(map(lambda x: x[0].reshape(data_shape),
data)).astype("float32")
y_data = np.array(map(lambda x: x[1], data)).astype("int64")
y_data = y_data.reshape([-1, 1])
loss, acc, weight = exe.run(
fluid.default_main_program(),
feed={"pixel": img_data,
"label": y_data},
fetch_list=[avg_cost, batch_acc, batch_size_tensor])
accuracy.add(value=acc, weight=weight)
iters += 1
num_samples += len(y_data)
print(
"Pass = %d, Iter = %d, Loss = %f, Accuracy = %f" %
(pass_id, iters, loss, acc)
) # The accuracy is the accumulation of batches, but not the current batch.
# pass_train_acc = accuracy.eval()
train_losses.append(loss)
train_accs.append(acc)
print("Pass: %d, Loss: %f, Train Accuray: %f\n" %
(pass_id, np.mean(train_losses), np.mean(train_accs)))
train_elapsed = time.time() - start_time
examples_per_sec = num_samples / train_elapsed
print('\nTotal examples: %d, total time: %.5f, %.5f examples/sed\n' %
(num_samples, train_elapsed, examples_per_sec))
# evaluation
if args.with_test:
pass_test_acc = test(exe)
exit(0)
def print_arguments():
print('----------- vgg Configuration Arguments -----------')
for arg, value in sorted(vars(args).iteritems()):
print('%s: %s' % (arg, value))
print('------------------------------------------------')
if __name__ == "__main__":
print_arguments()
main()
return avg_cost, inference_program, optimizer, train_reader, test_reader, batch_acc
cmake/configure.cmake
浏览文件 @ fb3bfc3d
......@@ -41,6 +41,10 @@ if(USE_EIGEN_FOR_BLAS)
add_definitions(-DPADDLE_USE_EIGEN_FOR_BLAS)
endif(USE_EIGEN_FOR_BLAS)
if(EIGEN_USE_THREADS)
add_definitions(-DEIGEN_USE_THREADS)
endif(EIGEN_USE_THREADS)
if(NOT WITH_PROFILER)
add_definitions(-DPADDLE_DISABLE_PROFILER)
endif(NOT WITH_PROFILER)
......
cmake/cpplint.cmake 已删除 100644 → 0
浏览文件 @ 652332cd
# util to check C++ file style
# * it basically use google cpplint.py.
# * It provide "add_style_check_target" for cmake.
# Usage see add_style_check_target's document
#
# TODO(yuyang18): Add python style check.
set(STYLE_FILTER)
# diable unwanted filters
# paddle do not indent public/potected/private in class
set(STYLE_FILTER "${STYLE_FILTER}-whitespace/indent,")
# paddle use mutable reference. BUT IT IS NOT RECOMMANDED
set(STYLE_FILTER "${STYLE_FILTER}-runtime/references,")
# paddle use relative path for include.
set(STYLE_FILTER "${STYLE_FILTER}-build/include,")
# paddle use <thread>, <mutex>, etc.
set(STYLE_FILTER "${STYLE_FILTER}-build/c++11,")
# paddle use c style casting. BUT IT IS NOT RECOMMANDED
set(STYLE_FILTER "${STYLE_FILTER}-readability/casting")
# IGNORE SOME FILES
set(IGNORE_PATTERN
.*ImportanceSampler.*
.*cblas\\.h.*
.*\\.pb\\.txt
.*MultiDataProvider.*
.*pb.*
.*pybind.h)
# add_style_check_target
#
# attach check code style step for target.
#
# first argument: target name to attach
# rest arguments: source list to check code style.
#
# NOTE: If WITH_STYLE_CHECK is OFF, then this macro just do nothing.
macro(add_style_check_target TARGET_NAME)
if(WITH_STYLE_CHECK)
set(SOURCES_LIST ${ARGN})
list(REMOVE_DUPLICATES SOURCES_LIST)
foreach(filename ${SOURCES_LIST})
foreach(pattern ${IGNORE_PATTERN})
if(filename MATCHES ${pattern})
list(REMOVE_ITEM SOURCES_LIST ${filename})
endif()
endforeach()
endforeach()
if(SOURCES_LIST)
add_custom_command(TARGET ${TARGET_NAME} POST_BUILD
COMMAND "${PYTHON_EXECUTABLE}" "${PADDLE_SOURCE_DIR}/paddle/scripts/cpplint.py"
"--filter=${STYLE_FILTER}"
${SOURCES_LIST}
COMMENT "cpplint: Checking source code style"
WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR})
endif()
endif()
endmacro()
cmake/cuda.cmake
浏览文件 @ fb3bfc3d
......@@ -172,6 +172,8 @@ set(CUDA_PROPAGATE_HOST_FLAGS OFF)
list(APPEND CUDA_NVCC_FLAGS "-std=c++11")
list(APPEND CUDA_NVCC_FLAGS "--use_fast_math")
list(APPEND CUDA_NVCC_FLAGS "-Xcompiler -fPIC")
# in cuda9, suppress cuda warning on eigen
list(APPEND CUDA_NVCC_FLAGS "-w")
# Set :expt-relaxed-constexpr to suppress Eigen warnings
list(APPEND CUDA_NVCC_FLAGS "--expt-relaxed-constexpr")
......
cmake/external/boost.cmake
浏览文件 @ fb3bfc3d
......@@ -23,8 +23,12 @@ set(BOOST_PROJECT "extern_boost")
# checked that the devtools package of CentOS 6 installs boost 1.41.0.
# So we use 1.41.0 here.
set(BOOST_VER "1.41.0")
set(BOOST_TAR "boost_1_41_0")
set(BOOST_URL "http://paddlepaddledeps.bj.bcebos.com/${BOOST_TAR}.tar.gz")
if((NOT DEFINED BOOST_TAR) OR (NOT DEFINED BOOST_URL))
message(STATUS "use pre defined download url")
set(BOOST_TAR "boost_1_41_0" CACHE STRING "" FORCE)
set(BOOST_URL "http://paddlepaddledeps.cdn.bcebos.com/${BOOST_TAR}.tar.gz" CACHE STRING "" FORCE)
endif()
MESSAGE(STATUS "BOOST_TAR: ${BOOST_TAR}, BOOST_URL: ${BOOST_URL}")
set(BOOST_SOURCES_DIR ${THIRD_PARTY_PATH}/boost)
set(BOOST_DOWNLOAD_DIR "${BOOST_SOURCES_DIR}/src/${BOOST_PROJECT}")
set(BOOST_INCLUDE_DIR "${BOOST_DOWNLOAD_DIR}/${BOOST_TAR}" CACHE PATH "boost include directory." FORCE)
......
cmake/external/grpc.cmake
浏览文件 @ fb3bfc3d
......@@ -23,17 +23,28 @@ SET(GRPC_SOURCES_DIR ${THIRD_PARTY_PATH}/grpc)
SET(GRPC_INSTALL_DIR ${THIRD_PARTY_PATH}/install/grpc)
SET(GRPC_INCLUDE_DIR "${GRPC_INSTALL_DIR}/include/" CACHE PATH "grpc include directory." FORCE)
SET(GRPC_CPP_PLUGIN "${GRPC_INSTALL_DIR}/bin/grpc_cpp_plugin" CACHE FILEPATH "GRPC_CPP_PLUGIN" FORCE)
include(ProcessorCount)
ProcessorCount(NUM_OF_PROCESSOR)
IF(APPLE)
SET(BUILD_CMD make -n HAS_SYSTEM_PROTOBUF=false -s -j static grpc_cpp_plugin | sed "s/-Werror//g" | sh)
SET(BUILD_CMD make -n HAS_SYSTEM_PROTOBUF=false -s -j ${NUM_OF_PROCESSOR} static grpc_cpp_plugin | sed "s/-Werror//g" | sh)
ELSE()
SET(BUILD_CMD make HAS_SYSTEM_PROTOBUF=false -s -j static grpc_cpp_plugin)
SET(BUILD_CMD make HAS_SYSTEM_PROTOBUF=false -s -j ${NUM_OF_PROCESSOR} static grpc_cpp_plugin)
ENDIF()
# FIXME(wuyi): do not build zlib cares protobuf twice, find a way to build grpc with them
ExternalProject_Add(
extern_grpc
DEPENDS protobuf zlib
GIT_REPOSITORY "https://github.com/grpc/grpc.git"
GIT_TAG "v1.10.x"
# NOTE(wuyi):
# this package is generated by following steps:
# 1. git clone -b v1.8.x https://github.com/grpc/grpc.git
# 2. submodule update --init
# 3. keep only zlib, cares, protobuf, boringssl under "third_party",
# checkout and clean other dirs under third_party
# 4. remove .git, and package the directory.
URL "http://paddlepaddledeps.bj.bcebos.com/grpc-v1.8.x.tar.gz"
PREFIX ${GRPC_SOURCES_DIR}
UPDATE_COMMAND ""
CONFIGURE_COMMAND ""
......@@ -46,7 +57,6 @@ ExternalProject_Add(
INSTALL_COMMAND make prefix=${GRPC_INSTALL_DIR} install
)
# FIXME(typhoonzero): hack to get static lib path, try a better way like merge them.
ADD_LIBRARY(grpc++_unsecure STATIC IMPORTED GLOBAL)
SET_PROPERTY(TARGET grpc++_unsecure PROPERTY IMPORTED_LOCATION
"${GRPC_INSTALL_DIR}/lib/libgrpc++_unsecure.a")
......
cmake/external/mkldnn.cmake
浏览文件 @ fb3bfc3d
......@@ -45,15 +45,15 @@ IF(${CBLAS_PROVIDER} STREQUAL "MKLML")
ELSE()
MESSAGE(FATAL_ERROR "Should enable MKLML when build MKLDNN")
ENDIF()
SET(MKLDNN_CFLAG "${CMAKE_C_FLAGS} -Wno-error=strict-overflow")
SET(MKLDNN_CXXFLAG "${CMAKE_CXX_FLAGS} -Wno-error=strict-overflow")
SET(MKLDNN_FLAG "-Wno-error=strict-overflow -Wno-error=unused-result -Wno-unused-result")
SET(MKLDNN_CFLAG "${CMAKE_C_FLAGS} ${MKLDNN_FLAG}")
SET(MKLDNN_CXXFLAG "${CMAKE_CXX_FLAGS} ${MKLDNN_FLAG}")
ExternalProject_Add(
${MKLDNN_PROJECT}
${EXTERNAL_PROJECT_LOG_ARGS}
DEPENDS ${MKLDNN_DEPENDS}
GIT_REPOSITORY "https://github.com/01org/mkl-dnn.git"
GIT_TAG "v0.11"
GIT_TAG "db3424ad44901513c03a1ea31ccaacdf633fbe9f"
PREFIX ${MKLDNN_SOURCES_DIR}
UPDATE_COMMAND ""
CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${MKLDNN_INSTALL_DIR}
......@@ -61,6 +61,7 @@ ExternalProject_Add(
CMAKE_ARGS -DMKLROOT=${MKLML_ROOT}
CMAKE_ARGS -DCMAKE_C_FLAGS=${MKLDNN_CFLAG}
CMAKE_ARGS -DCMAKE_CXX_FLAGS=${MKLDNN_CXXFLAG}
CMAKE_ARGS -DWITH_TEST=OFF -DWITH_EXAMPLE=OFF
CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${MKLDNN_INSTALL_DIR}
-DMKLROOT:PATH=${MKLML_ROOT}
)
......
cmake/external/mklml.cmake
浏览文件 @ fb3bfc3d
......@@ -27,8 +27,12 @@ ENDIF()
INCLUDE(ExternalProject)
SET(MKLML_PROJECT "extern_mklml")
SET(MKLML_VER "mklml_lnx_2018.0.1.20171007")
SET(MKLML_URL "http://paddlepaddledeps.bj.bcebos.com/${MKLML_VER}.tgz")
IF((NOT DEFINED MKLML_VER) OR (NOT DEFINED MKLML_URL))
MESSAGE(STATUS "use pre defined download url")
SET(MKLML_VER "mklml_lnx_2018.0.3.20180406" CACHE STRING "" FORCE)
SET(MKLML_URL "http://paddlepaddledeps.cdn.bcebos.com/${MKLML_VER}.tgz" CACHE STRING "" FORCE)
ENDIF()
MESSAGE(STATUS "MKLML_VER: ${MKLML_VER}, MKLML_URL: ${MKLML_URL}")
SET(MKLML_SOURCE_DIR "${THIRD_PARTY_PATH}/mklml")
SET(MKLML_DOWNLOAD_DIR "${MKLML_SOURCE_DIR}/src/${MKLML_PROJECT}")
SET(MKLML_DST_DIR "mklml")
......
cmake/external/protobuf.cmake
浏览文件 @ fb3bfc3d
......@@ -212,6 +212,7 @@ FUNCTION(build_protobuf TARGET_NAME BUILD_FOR_HOST)
${CMAKE_COMMAND} ${PROTOBUF_SOURCES_DIR}/src/${TARGET_NAME}/cmake
${OPTIONAL_ARGS}
-Dprotobuf_BUILD_TESTS=OFF
-DCMAKE_SKIP_RPATH=ON
-DCMAKE_POSITION_INDEPENDENT_CODE=ON
-DCMAKE_BUILD_TYPE=${THIRD_PARTY_BUILD_TYPE}
-DCMAKE_INSTALL_PREFIX=${PROTOBUF_INSTALL_DIR}
......
cmake/external/snappy.cmake
浏览文件 @ fb3bfc3d
......@@ -47,8 +47,6 @@ ExternalProject_Add(
-DCMAKE_INSTALL_LIBDIR:PATH=${SNAPPY_INSTALL_DIR}/lib
-DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON
-DCMAKE_BUILD_TYPE:STRING=${THIRD_PARTY_BUILD_TYPE}
BUILD_COMMAND make -j8
INSTALL_COMMAND make install
)
add_library(snappy STATIC IMPORTED GLOBAL)
......
cmake/external/snappystream.cmake
浏览文件 @ fb3bfc3d
......@@ -46,8 +46,6 @@ ExternalProject_Add(
-DCMAKE_INSTALL_PREFIX:PATH=${SNAPPYSTREAM_INSTALL_DIR}
-DCMAKE_INSTALL_LIBDIR:PATH=${SNAPPYSTREAM_INSTALL_DIR}/lib
-DCMAKE_BUILD_TYPE:STRING=${THIRD_PARTY_BUILD_TYPE}
BUILD_COMMAND make -j8
INSTALL_COMMAND make install
DEPENDS snappy
)
......
cmake/external/warpctc.cmake
浏览文件 @ fb3bfc3d
......@@ -38,8 +38,7 @@ ENDIF()
ExternalProject_Add(
extern_warpctc
${EXTERNAL_PROJECT_LOG_ARGS}
GIT_REPOSITORY "https://github.com/gangliao/warp-ctc.git"
GIT_TAG b63a0644654a3e0ed624c85a1767bc8193aead09
GIT_REPOSITORY "https://github.com/dzhwinter/warp-ctc.git"
PREFIX ${WARPCTC_SOURCES_DIR}
UPDATE_COMMAND ""
CMAKE_ARGS -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
......
cmake/generic.cmake
浏览文件 @ fb3bfc3d
......@@ -206,8 +206,6 @@ function(cc_library TARGET_NAME)
list(APPEND cc_library_HEADERS ${CMAKE_CURRENT_SOURCE_DIR}/${source}.h)
endif()
endforeach()
add_style_check_target(${TARGET_NAME} ${cc_library_SRCS} ${cc_library_HEADERS})
else(cc_library_SRCS)
if(cc_library_DEPS)
merge_static_libs(${TARGET_NAME} ${cc_library_DEPS})
......@@ -231,7 +229,7 @@ endfunction(cc_binary)
function(cc_test TARGET_NAME)
if(WITH_TESTING)
set(options "")
set(options SERIAL)
set(oneValueArgs "")
set(multiValueArgs SRCS DEPS ARGS)
cmake_parse_arguments(cc_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
......@@ -241,6 +239,9 @@ function(cc_test TARGET_NAME)
add_test(NAME ${TARGET_NAME}
COMMAND ${TARGET_NAME} ${cc_test_ARGS}
WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
if (${cc_test_SERIAL})
set_property(TEST ${TARGET_NAME} PROPERTY SERIAL 1)
endif()
endif()
endfunction(cc_test)
......@@ -268,7 +269,6 @@ function(nv_library TARGET_NAME)
list(APPEND nv_library_HEADERS ${CMAKE_CURRENT_SOURCE_DIR}/${source}.h)
endif()
endforeach()
add_style_check_target(${TARGET_NAME} ${nv_library_SRCS} ${nv_library_HEADERS})
else(nv_library_SRCS)
if (nv_library_DEPS)
merge_static_libs(${TARGET_NAME} ${nv_library_DEPS})
......@@ -295,7 +295,7 @@ endfunction(nv_binary)
function(nv_test TARGET_NAME)
if (WITH_GPU AND WITH_TESTING)
set(options "")
set(options SERIAL)
set(oneValueArgs "")
set(multiValueArgs SRCS DEPS)
cmake_parse_arguments(nv_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
......@@ -303,6 +303,9 @@ function(nv_test TARGET_NAME)
target_link_libraries(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main memory gtest gflags glog)
add_dependencies(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main memory gtest gflags glog)
add_test(${TARGET_NAME} ${TARGET_NAME})
if (nv_test_SERIAL)
set_property(TEST ${TARGET_NAME} PROPERTY SERIAL 1)
endif()
endif()
endfunction(nv_test)
......@@ -338,7 +341,6 @@ function(hip_library TARGET_NAME)
list(APPEND hip_library_HEADERS ${CMAKE_CURRENT_SOURCE_DIR}/${source}.h)
endif()
endforeach()
add_style_check_target(${TARGET_NAME} ${hip_library_SRCS} ${hip_library_HEADERS})
else(hip_library_SRCS)
if (hip_library_DEPS)
merge_static_libs(${TARGET_NAME} ${hip_library_DEPS})
......
cmake/inference_lib.cmake
浏览文件 @ fb3bfc3d
......@@ -52,66 +52,91 @@ function(copy TARGET)
endfunction()
# third party
set(dst_dir "${CMAKE_INSTALL_PREFIX}/third_party/eigen3")
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/eigen3")
copy(eigen3_lib
SRCS ${EIGEN_INCLUDE_DIR}/Eigen/Core ${EIGEN_INCLUDE_DIR}/Eigen/src ${EIGEN_INCLUDE_DIR}/unsupported/Eigen
DSTS ${dst_dir}/Eigen ${dst_dir}/Eigen ${dst_dir}/unsupported
DEPS eigen3
)
set(dst_dir "${CMAKE_INSTALL_PREFIX}/third_party/install/gflags")
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/gflags")
copy(gflags_lib
SRCS ${GFLAGS_INCLUDE_DIR} ${GFLAGS_LIBRARIES}
DSTS ${dst_dir} ${dst_dir}/lib
DEPS gflags
)
set(dst_dir "${CMAKE_INSTALL_PREFIX}/third_party/install/glog")
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/glog")
copy(glog_lib
SRCS ${GLOG_INCLUDE_DIR} ${GLOG_LIBRARIES}
DSTS ${dst_dir} ${dst_dir}/lib
DEPS glog
)
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/boost/")
copy(boost_lib
SRCS ${BOOST_INCLUDE_DIR}/boost
DSTS ${dst_dir}
DEPS boost
)
if(NOT PROTOBUF_FOUND)
set(dst_dir "${CMAKE_INSTALL_PREFIX}/third_party/install/protobuf")
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/protobuf")
copy(protobuf_lib
SRCS ${PROTOBUF_INCLUDE_DIR} ${PROTOBUF_LIBRARY}
DSTS ${dst_dir} ${dst_dir}/lib
DEPS extern_protobuf
)
endif()
if(NOT CBLAS_FOUND)
set(dst_dir "${CMAKE_INSTALL_PREFIX}/third_party/install/openblas")
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/openblas")
copy(openblas_lib
SRCS ${CBLAS_INSTALL_DIR}/lib ${CBLAS_INSTALL_DIR}/include
DSTS ${dst_dir} ${dst_dir}
DEPS extern_openblas
)
elseif (WITH_MKLML)
set(dst_dir "${CMAKE_INSTALL_PREFIX}/third_party/install/mklml")
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/mklml")
copy(mklml_lib
SRCS ${MKLML_LIB} ${MKLML_IOMP_LIB} ${MKLML_INC_DIR}
DSTS ${dst_dir}/lib ${dst_dir}/lib ${dst_dir}
DEPS mklml
)
endif()
if(WITH_MKLDNN)
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/mkldnn")
copy(mkldnn_lib
SRCS ${MKLDNN_INC_DIR} ${MKLDNN_SHARED_LIB}
DSTS ${dst_dir} ${dst_dir}/lib
DEPS mkldnn
)
endif()
if(NOT MOBILE_INFERENCE AND NOT RPI)
set(dst_dir "${CMAKE_INSTALL_PREFIX}/third_party/install/snappy")
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/snappy")
copy(snappy_lib
SRCS ${SNAPPY_INCLUDE_DIR} ${SNAPPY_LIBRARIES}
DSTS ${dst_dir} ${dst_dir}/lib)
DSTS ${dst_dir} ${dst_dir}/lib
DEPS snappy)
set(dst_dir "${CMAKE_INSTALL_PREFIX}/third_party/install/snappystream")
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/snappystream")
copy(snappystream_lib
SRCS ${SNAPPYSTREAM_INCLUDE_DIR} ${SNAPPYSTREAM_LIBRARIES}
DSTS ${dst_dir} ${dst_dir}/lib)
DSTS ${dst_dir} ${dst_dir}/lib
DEPS snappystream)
set(dst_dir "${CMAKE_INSTALL_PREFIX}/third_party/install/zlib")
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/zlib")
copy(zlib_lib
SRCS ${ZLIB_INCLUDE_DIR} ${ZLIB_LIBRARIES}
DSTS ${dst_dir} ${dst_dir}/lib)
DSTS ${dst_dir} ${dst_dir}/lib
DEPS zlib)
endif()
# paddle fluid module
set(src_dir "${PADDLE_SOURCE_DIR}/paddle/fluid")
set(dst_dir "${CMAKE_INSTALL_PREFIX}/paddle/fluid")
set(dst_dir "${FLUID_INSTALL_DIR}/paddle/fluid")
set(module "framework")
copy(framework_lib DEPS framework_py_proto
SRCS ${src_dir}/${module}/*.h ${src_dir}/${module}/details/*.h ${PADDLE_BINARY_DIR}/paddle/fluid/framework/framework.pb.h
......@@ -142,4 +167,31 @@ copy(string_lib
DSTS ${dst_dir}/${module} ${dst_dir}/${module}/tinyformat
)
set(module "pybind")
copy(pybind_lib
SRCS ${CMAKE_CURRENT_BINARY_DIR}/paddle/fluid/${module}/pybind.h
DSTS ${dst_dir}/${module}
)
# CMakeCache Info
copy(cmake_cache
SRCS ${CMAKE_CURRENT_BINARY_DIR}/CMakeCache.txt
DSTS ${FLUID_INSTALL_DIR})
add_custom_target(inference_lib_dist DEPENDS ${inference_lib_dist_dep})
# paddle fluid version
execute_process(
COMMAND ${GIT_EXECUTABLE} log --pretty=format:%H -1
WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}
OUTPUT_VARIABLE PADDLE_GIT_COMMIT)
set(version_file ${FLUID_INSTALL_DIR}/version.txt)
file(WRITE ${version_file}
"GIT COMMIT ID: ${PADDLE_GIT_COMMIT}\n"
"WITH_MKL: ${WITH_MKL}\n"
"WITH_GPU: ${WITH_GPU}\n")
if(WITH_GPU)
file(APPEND ${version_file}
"CUDA version: ${CUDA_VERSION}\n"
"CUDNN version: v${CUDNN_MAJOR_VERSION}\n")
endif()
doc/fluid/CMakeLists.txt
浏览文件 @ fb3bfc3d
......@@ -15,6 +15,9 @@ set(SPHINX_CACHE_DIR_EN "${CMAKE_CURRENT_BINARY_DIR}/en/_doctrees")
# HTML output director
set(SPHINX_HTML_DIR_EN "${CMAKE_CURRENT_BINARY_DIR}/en/html")
set(IMPORT_PADDLE_STRING "")
set(IMPORT_PADDLEV2_STRING "")
configure_file(
"${CMAKE_CURRENT_SOURCE_DIR}/../templates/conf.py.en.in"
"${BINARY_BUILD_DIR_EN}/conf.py"
......@@ -27,8 +30,6 @@ sphinx_add_target(paddle_fluid_docs
${CMAKE_CURRENT_SOURCE_DIR}
${SPHINX_HTML_DIR_EN})
add_dependencies(paddle_fluid_docs gen_proto_py paddle_python)
# configured documentation tools and intermediate build results
set(BINARY_BUILD_DIR_CN "${CMAKE_CURRENT_BINARY_DIR}/cn/_build")
......@@ -50,6 +51,4 @@ sphinx_add_target(paddle_fluid_docs_cn
${CMAKE_CURRENT_SOURCE_DIR}
${SPHINX_HTML_DIR_CN})
add_dependencies(paddle_fluid_docs_cn gen_proto_py paddle_python)
add_subdirectory(api)
doc/fluid/api/CMakeLists.txt
浏览文件 @ fb3bfc3d
......@@ -7,6 +7,9 @@ set(SPHINX_CACHE_DIR_EN "${CMAKE_CURRENT_BINARY_DIR}/en/_doctrees")
# HTML output director
set(SPHINX_HTML_DIR_EN "${CMAKE_CURRENT_BINARY_DIR}/en/html")
set(IMPORT_PADDLE_STRING "import paddle")
set(IMPORT_PADDLEV2_STRING "import paddle.v2")
configure_file(
"${CMAKE_CURRENT_SOURCE_DIR}/../../templates/conf.py.en.in"
"${BINARY_BUILD_DIR_EN}/conf.py"
......
doc/fluid/api/clip.rst 0 → 100644
浏览文件 @ fb3bfc3d
.. THIS FILE IS GENERATED BY `gen_doc.{py|sh}`
!DO NOT EDIT THIS FILE MANUALLY!
====
clip
====
ErrorClipByValue
----------------
.. autoclass:: paddle.fluid.clip.ErrorClipByValue
:members:
:noindex:
GradientClipByValue
-------------------
.. autoclass:: paddle.fluid.clip.GradientClipByValue
:members:
:noindex:
GradientClipByNorm
------------------
.. autoclass:: paddle.fluid.clip.GradientClipByNorm
:members:
:noindex:
GradientClipByGlobalNorm
------------------------
.. autoclass:: paddle.fluid.clip.GradientClipByGlobalNorm
:members:
:noindex:
append_gradient_clip_ops
------------------------
.. autofunction:: paddle.fluid.clip.append_gradient_clip_ops
:noindex:
error_clip_callback
-------------------
.. autofunction:: paddle.fluid.clip.error_clip_callback
:noindex:
doc/fluid/api/data/data_reader.rst
浏览文件 @ fb3bfc3d
......@@ -56,11 +56,11 @@ DataFeeder
Reader
======
.. automodule:: paddle.v2.reader
.. automodule:: paddle.reader
:members:
:noindex:
.. automodule:: paddle.v2.reader.creator
.. automodule:: paddle.reader.creator
:members:
:noindex:
......
doc/fluid/api/evaluator.rst
浏览文件 @ fb3bfc3d
......@@ -5,24 +5,3 @@
evaluator
=========
ChunkEvaluator
--------------
.. autoclass:: paddle.fluid.evaluator.ChunkEvaluator
:members:
:noindex:
EditDistance
--------------
.. autoclass:: paddle.fluid.evaluator.EditDistance
:members:
:noindex:
DetectionMAP
--------------
.. autoclass:: paddle.fluid.evaluator.DetectionMAP
:members:
:noindex:
doc/fluid/api/executor.rst
浏览文件 @ fb3bfc3d
......@@ -30,3 +30,9 @@ switch_scope
.. autofunction:: paddle.fluid.executor.switch_scope
:noindex:
fetch_var
---------
.. autofunction:: paddle.fluid.executor.fetch_var
:noindex:
doc/fluid/api/gen_doc.sh
浏览文件 @ fb3bfc3d
#!/bin/bash
python gen_doc.py layers --submodules control_flow device io nn ops tensor > layers.rst
for module in io data_feeder evaluator executor initializer io nets optimizer param_attr profiler regularizer
for module in data_feeder clip metrics executor initializer io nets optimizer param_attr profiler regularizer
do
python gen_doc.py ${module} > ${module}.rst
done
doc/fluid/api/index_en.rst
浏览文件 @ fb3bfc3d
......@@ -9,8 +9,9 @@ Fluid
data_feeder.rst
executor.rst
initializer.rst
evaluator.rst
metrics.rst
nets.rst
clip.rst
optimizer.rst
param_attr.rst
profiler.rst
......
doc/fluid/api/initializer.rst
浏览文件 @ fb3bfc3d
......@@ -33,11 +33,16 @@ Xavier
:members:
:noindex:
MSRA
------
force_init_on_cpu
-----------------
.. autoclass:: paddle.fluid.initializer.MSRA
:members:
.. autofunction:: paddle.fluid.initializer.force_init_on_cpu
:noindex:
init_on_cpu
-----------
.. autofunction:: paddle.fluid.initializer.init_on_cpu
:noindex:
ConstantInitializer
......@@ -68,9 +73,3 @@ XavierInitializer
:members:
:noindex:
MSRAInitializer
-----------------
.. autoclass:: paddle.fluid.initializer.MSRAInitializer
:members:
:noindex:
doc/fluid/api/layers.rst
浏览文件 @ fb3bfc3d
......@@ -55,6 +55,13 @@ While
:members:
:noindex:
Switch
------
.. autoclass:: paddle.fluid.layers.Switch
:members:
:noindex:
lod_rank_table
--------------
......@@ -67,12 +74,6 @@ max_sequence_len
.. autofunction:: paddle.fluid.layers.max_sequence_len
:noindex:
topk
----
.. autofunction:: paddle.fluid.layers.topk
:noindex:
lod_tensor_to_array
-------------------
......@@ -109,6 +110,12 @@ less_than
.. autofunction:: paddle.fluid.layers.less_than
:noindex:
equal
-----
.. autofunction:: paddle.fluid.layers.equal
:noindex:
array_read
----------
......@@ -212,6 +219,42 @@ Send
.. autofunction:: paddle.fluid.layers.Send
:noindex:
open_recordio_file
------------------
.. autofunction:: paddle.fluid.layers.open_recordio_file
:noindex:
open_files
----------
.. autofunction:: paddle.fluid.layers.open_files
:noindex:
read_file
---------
.. autofunction:: paddle.fluid.layers.read_file
:noindex:
shuffle
-------
.. autofunction:: paddle.fluid.layers.shuffle
:noindex:
batch
-----
.. autofunction:: paddle.fluid.layers.batch
:noindex:
double_buffer
-------------
.. autofunction:: paddle.fluid.layers.double_buffer
:noindex:
nn
==
......@@ -281,12 +324,6 @@ square_error_cost
.. autofunction:: paddle.fluid.layers.square_error_cost
:noindex:
accuracy
--------
.. autofunction:: paddle.fluid.layers.accuracy
:noindex:
chunk_eval
----------
......@@ -311,6 +348,18 @@ sequence_pool
.. autofunction:: paddle.fluid.layers.sequence_pool
:noindex:
sequence_softmax
----------------
.. autofunction:: paddle.fluid.layers.sequence_softmax
:noindex:
softmax
-------
.. autofunction:: paddle.fluid.layers.softmax
:noindex:
pool2d
------
......@@ -323,12 +372,6 @@ batch_norm
.. autofunction:: paddle.fluid.layers.batch_norm
:noindex:
layer_norm
----------
.. autofunction:: paddle.fluid.layers.layer_norm
:noindex:
beam_search_decode
------------------
......@@ -377,6 +420,12 @@ reduce_min
.. autofunction:: paddle.fluid.layers.reduce_min
:noindex:
reduce_prod
-----------
.. autofunction:: paddle.fluid.layers.reduce_prod
:noindex:
sequence_first_step
-------------------
......@@ -425,6 +474,12 @@ matmul
.. autofunction:: paddle.fluid.layers.matmul
:noindex:
topk
----
.. autofunction:: paddle.fluid.layers.topk
:noindex:
warpctc
-------
......@@ -473,25 +528,34 @@ multiplex
.. autofunction:: paddle.fluid.layers.multiplex
:noindex:
label_smooth
------------
layer_norm
----------
.. autofunction:: paddle.fluid.layers.label_smooth
.. autofunction:: paddle.fluid.layers.layer_norm
:noindex:
ops
===
softmax_with_cross_entropy
--------------------------
mean
----
.. autofunction:: paddle.fluid.layers.softmax_with_cross_entropy
:noindex:
.. autofunction:: paddle.fluid.layers.mean
smooth_l1
---------
.. autofunction:: paddle.fluid.layers.smooth_l1
:noindex:
mul
---
one_hot
-------
.. autofunction:: paddle.fluid.layers.mul
.. autofunction:: paddle.fluid.layers.one_hot
:noindex:
autoincreased_step_counter
--------------------------
.. autofunction:: paddle.fluid.layers.autoincreased_step_counter
:noindex:
reshape
......@@ -500,12 +564,52 @@ reshape
.. autofunction:: paddle.fluid.layers.reshape
:noindex:
lod_reset
---------
.. autofunction:: paddle.fluid.layers.lod_reset
:noindex:
lrn
---
.. autofunction:: paddle.fluid.layers.lrn
:noindex:
pad
---
.. autofunction:: paddle.fluid.layers.pad
:noindex:
label_smooth
------------
.. autofunction:: paddle.fluid.layers.label_smooth
:noindex:
roi_pool
--------
.. autofunction:: paddle.fluid.layers.roi_pool
:noindex:
ops
===
mean
----
.. autofunction:: paddle.fluid.layers.mean
:noindex:
mul
---
.. autofunction:: paddle.fluid.layers.mul
:noindex:
scale
-----
......@@ -572,10 +676,70 @@ clip_by_norm
.. autofunction:: paddle.fluid.layers.clip_by_norm
:noindex:
sequence_softmax
----------------
logical_and
-----------
.. autofunction:: paddle.fluid.layers.sequence_softmax
.. autofunction:: paddle.fluid.layers.logical_and
:noindex:
logical_or
----------
.. autofunction:: paddle.fluid.layers.logical_or
:noindex:
logical_xor
-----------
.. autofunction:: paddle.fluid.layers.logical_xor
:noindex:
logical_not
-----------
.. autofunction:: paddle.fluid.layers.logical_not
:noindex:
uniform_random
--------------
.. autofunction:: paddle.fluid.layers.uniform_random
:noindex:
uniform_random_batch_size_like
------------------------------
.. autofunction:: paddle.fluid.layers.uniform_random_batch_size_like
:noindex:
gaussian_random
---------------
.. autofunction:: paddle.fluid.layers.gaussian_random
:noindex:
gaussian_random_batch_size_like
-------------------------------
.. autofunction:: paddle.fluid.layers.gaussian_random_batch_size_like
:noindex:
cumsum
------
.. autofunction:: paddle.fluid.layers.cumsum
:noindex:
scatter
-------
.. autofunction:: paddle.fluid.layers.scatter
:noindex:
sum
---
.. autofunction:: paddle.fluid.layers.sum
:noindex:
sigmoid
......@@ -644,6 +808,18 @@ floor
.. autofunction:: paddle.fluid.layers.floor
:noindex:
cos
---
.. autofunction:: paddle.fluid.layers.cos
:noindex:
sin
---
.. autofunction:: paddle.fluid.layers.sin
:noindex:
round
-----
......@@ -820,3 +996,16 @@ topk
.. autofunction:: paddle.fluid.layers.topk
:noindex:
dice_loss
----
.. autofunction:: paddle.fluid.layers.dice_loss
:noindex:
upsampling_bilinear2d
____
.. autofunction:: paddle.fluid.layers.upsampling_bilinear2d
:noindex:
doc/fluid/api/metrics.rst 0 → 100644
浏览文件 @ fb3bfc3d
.. THIS FILE IS GENERATED BY `gen_doc.{py|sh}`
!DO NOT EDIT THIS FILE MANUALLY!
=======
metrics
=======
MetricBase
----------
.. autoclass:: paddle.fluid.metrics.MetricBase
:members:
:noindex:
CompositeMetric
---------------
.. autoclass:: paddle.fluid.metrics.CompositeMetric
:members:
:noindex:
Accuracy
--------
.. autoclass:: paddle.fluid.metrics.Accuracy
:members:
:noindex:
ChunkEvaluator
--------------
.. autoclass:: paddle.fluid.metrics.ChunkEvaluator
:members:
:noindex:
EditDistance
------------
.. autoclass:: paddle.fluid.metrics.EditDistance
:members:
:noindex:
DetectionMAP
------------
.. autoclass:: paddle.fluid.metrics.DetectionMAP
:members:
:noindex:
Auc
---
.. autoclass:: paddle.fluid.metrics.Auc
:members:
:noindex:
doc/fluid/api/optimizer.rst
浏览文件 @ fb3bfc3d
......@@ -47,6 +47,28 @@ DecayedAdagrad
:members:
:noindex:
Adadelta
-----------------
.. autoclass:: paddle.fluid.optimizer.Adadelta
:members:
:noindex:
RMSProp
-----------------
.. autoclass:: paddle.fluid.optimizer.RMSProp
:members:
:noindex:
ModelAverage
-----------------
.. autoclass:: paddle.fluid.optimizer.ModelAverage
:members:
:noindex:
SGDOptimizer
------------
......@@ -89,9 +111,25 @@ DecayedAdagradOptimizer
:members:
:noindex:
Adadelta
--------------
AdadeltaOptimizer
-----------------
.. autoclass:: paddle.fluid.optimizer.AdadeltaOptimizer
:members:
:noindex:
RMSPropOptimizer
-----------------
.. autoclass:: paddle.fluid.optimizer.RMSPropOptimizer
:members:
:noindex:
Optimizer
---------
.. autoclass:: paddle.fluid.optimizer.Optimizer
:members:
:noindex:
doc/fluid/api/regularizer.rst
浏览文件 @ fb3bfc3d
......@@ -11,6 +11,13 @@ append_regularization_ops
.. autofunction:: paddle.fluid.regularizer.append_regularization_ops
:noindex:
WeightDecayRegularizer
----------------------
.. autoclass:: paddle.fluid.regularizer.WeightDecayRegularizer
:members:
:noindex:
L1Decay
-------
......@@ -26,15 +33,16 @@ L2Decay
:noindex:
L1DecayRegularizer
---------------------
------------------
.. autoclass:: paddle.fluid.regularizer.L1DecayRegularizer
:members:
:noindex:
L2DecayRegularizer
---------------------
------------------
.. autoclass:: paddle.fluid.regularizer.L2DecayRegularizer
:members:
:noindex:
doc/fluid/build_and_install/paddleci.png 0 → 120000
浏览文件 @ fb3bfc3d
../../v2/build_and_install/paddleci.png
\ No newline at end of file
doc/fluid/design/algorithm/parameter_average.md
浏览文件 @ fb3bfc3d
# 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.
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 to obtain the optimal values of parameters by going through the data in as few passes as possible.
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.
......@@ -16,16 +16,16 @@ We propose averaging for any optimizer similar to how ASGD performs it, as menti
### 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
1. It will take in an instance of an 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.
1. In theory, the values of this copy are the average of the values of the parameters in the most recent N batches.
2. However, saving all 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:
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.
......
doc/fluid/design/concepts/functions_operators_layers.md
浏览文件 @ fb3bfc3d
......@@ -40,7 +40,7 @@ template <typename T>
class FCOp : public OperatorBase {
public:
void Run(...) {
add(mul(Input<T>("X"), Input<T>("W")), Input<T>("b");
add(mul(Input<T>("X"), Input<T>("W")), Input<T>("b"));
}
};
REGISTER_OP(FCOp, "fc");
......
doc/fluid/design/concepts/lod_tensor.md
浏览文件 @ fb3bfc3d
......@@ -155,7 +155,7 @@ into offsets
3 2+3 4+5 1+9 2+10 3+12
```
so we know that the first sentence is from word 0 to word 3, and the second sentence from work 3 to word 5.
so we know that the first sentence is from word 0 to word 3, and the second sentence from word 3 to word 5.
Similarly, the lengths in the top level LoD
......
doc/fluid/design/data_type/float16.md
浏览文件 @ fb3bfc3d
......@@ -3,7 +3,7 @@
## 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
When high precision computation is not required (which is usually the case at least in the deep learning inference stage), 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;
......@@ -12,7 +12,7 @@ When high precision computation is not required, using float16 data type could p
## 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.
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 kernels. 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.
......@@ -95,11 +95,89 @@ 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:
## float16 inference
In Fluid, a neural network is represented as a protobuf message called [ProgramDesc](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/program.md), whose Python wrapper is a [Program](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/python_api.md#program). The basic structure of a program is some nested [blocks](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/python_api.md#block), where each block consists of some [variable](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/python_api.md#variable) definitions and a sequence of [operators](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/python_api.md#operator). An [executor](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/executor.md) will run a given program desc by executing the sequence of operators in the entrance block of the program one by one.
- Update pybind/tensor_py.h to bind c++ float16 with numpy float16.
### Operator level requirement
Each operator has many kernels for different data types, devices, and library types. The operator will select the appropriate kernel to run based on, among other things, the data type of the input variables. By default, every Fluid operator has a float data type kernel that takes float variables as input and generates float output.
- Modify `GetKernelType()` method in `framework/operator.h` to make it compatible with float16.
This means that if we provide float input to the first operator in a program, then each opeartor will use float kernel to compute float output and send it as input to the next operator to trigger the float kernel. Overall, the program will run in float mode and give us a final output of float data type.
- Create a type-casting operator that can convert the data type in tensor between float16 and other types.
The same principle applies if we want a program to run in float16 mode. We provide input variable of float16 data type to the first operator, and then one by one, each operator in the program will run the float16 kernel (provided that each operator in this program has float16 kernels registered) until we finally obtain a float16 output variable.
So the preliminary requirement for float16 inference is to add float16 kernel to operators that are needed in a specific kind of program. For example, float16 inference on an image classification neural network like Vgg or Resnet, typically requires the following operators to have float16 kernels: convolution, pooling, multiplication, addition, batch norm, dropout, relu, and softmax. Please refer to [new_op_en](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/dev/new_op_en.md) for details of how to add new kernels to an operator.
### Variable level requirement
Operators including convolution and multiplication (used in fully-connected layers) takes as input not only the variables generated by the preceding operators but also [parameter](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/python_api.md#parameter) variables, which contains the trained weights to apply to the input data. These weights are obtained in the Fluid training process and are by default of float data type.
When these operators are running in float16 mode, the float16 kernel requires those parameter variables to contain weights of Fluid float16 data type. Thus, we need a convenient way to convert the original float weights to float16 weights.
In Fluid, we use tensor to hold actual data for a variable on the c++ end. [Pybind](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/pybind/tensor_py.h) is used to bind c++ tensors of certain data type with numpy array of the correponding numpy data type on the Python end. Each common c++ built-in data type has a corresponding numpy data type of the same name. However, since there is no built-in float16 type in c++, we cannot directly bind numpy float16 data type with the Fluid float16 class. Since both Fluid float16 and numpy float16 use uint16 as the internal data storage type, we use c++ built-in type `uint16_t` and the corresponding numpy uint16 data type to bridge the gap via [Pybind](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/pybind/tensor_py.h).
The following code demonstrates how to do the tensor conversion.
```Python
# var is the variable of float weights
# tensor is a numpy array of data copied from the tensor data in var
# fp16_var is the variable that will contain float16 weights converted from var
tensor = numpy.array(var.get_tensor())
fp16_tensor = fp16_var.get_tensor()
# After the original tensor data is converted to numpy float16 data type,
# view(numpy.uint16) is used so that the internal memory of the numpy array
# will be reinterpreted to be of uint16 data type, which is binded to
# Fluid float16 class via pybind with the help of uint16_t built-in c++ type
fp16_tensor.set(tensor.astype(numpy.float16).view(numpy.uint16), GPUPlace)
```
### Consistent API requirement
The basic inference in float16 mode requires users to feed input and obtain output both of float16 data type. However, in this way, the inference APIs are not consistent between float16 mode and float mode, and users may find it confusing and diffcult to use float16 inference since they need to do extra steps to provide float16 input data and convert float16 output data back to float. To have consistent API for different inference modes, we need to transpile the program desc in some way so that we can run float16 inference by feeding and fetching variables of float data type.
This problem can be solved by introducing a type-casting operator which takes an input variable of certain data type, cast it to another specified data type, and put the casted data into the output variable. Insert cast operator where needed can make a program internally run in float16 mode.
### float16 transpiler
Put all the above requirements in mind, we designed a float16 inference transpiler that can tranpile a float32 mode inference program desc to a float16 mode one.
Given a float inference program and the corresponding variables of float32 weights in the [scope](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/scope.md),
this transpiler mainly does the following modifications:
1. Insert cast operators at the beginning of the program so that the input float data will be converted to float16 data type before feeding to subsequent operators to invoke the float16 kernel.
2. Insert cast operators at the end of the program so that the output float16 data will be converted back to float data type before users obtain the result.
3. For each parameter variable of float weights, create in the scope a corresponding variable of float16 weights which are converted from the corresponding float weights and add this new float16 variable to the program.
4. Update the operator information in the program so that each relevant operator use the newly created float16 variable instead of its float counterpart.
Below is an example of usage:
```Python
# Get the float inference program
[float_inference_program, feed_target_names,
fetch_targets] = fluid.io.load_inference_model(save_dirname, exe)
# Prepare the float input data
tensor_img = numpy.random.rand(1, 3, 32, 32).astype(numpy.float32)
# Running inference_program in float mode
float_results = exe.run(float_inference_program,
feed={feed_target_names[0]: tensor_img},
fetch_list=fetch_targets)
# Use float16 transpiler to speedup
float16_inference_program = float_inference_program.clone()
t = fluid.InferenceTranspiler()
t.float16_transpile(float16_inference_program, GPUPlace)
# Running
float16_results = exe.run(float16_inference_program,
feed={feed_target_names[0]: tensor_img},
fetch_list=fetch_targets)
```
As we can see from the example above, users can simply use the `float16_transpile` method provided by the infernece transpiler class on an existing float inference program to run inference in float16 mode.
### Speedup on GPU
Currently, Fluid inference in float16 mode is only supported on Nvidia GPU device. There is no motivation to support float16 inference on non-ARM CPUs because float16 is not natively supported there and float16 calculation will only be slower than its float counterpart.
Nvidia started to support its native float16 data type (which has the same internal memory representation as Fluid float16 class) on CUDA 7.5. Moreover, float16 speedups on common computational intensive tasks including GEMM (general matrix-matrix multiplication) and convolution are supported since cublas 7.5 and cuDNN 5.0.
Recently, the introduction of [tensor core](https://devblogs.nvidia.com/programming-tensor-cores-cuda-9/) in volta architecture GPUs and the support of tensor core calculation in CUDA 9.0 and cuDNN 7.0 make float16 truly superior to float in certain deep learning applications. Please refer to this [benchmark report](https://github.com/kexinzhao/Paddle_benchmark/blob/master/float16_benchmark.md) for more details.
doc/fluid/design/motivation/api.md
浏览文件 @ fb3bfc3d
......@@ -77,8 +77,7 @@ print "The sematic-vector of testA: ", paddle.infer(fA, parameters, testA)
### Example 2. Sharing Parameters between "Models"
We use [GAN](https://github.com/PaddlePaddle/book/tree/develop/gan) in
this example. In the following example program, `d0` and `d1`
We use GAN in this example. In the following example program, `d0` and `d1`
correspond to the two networks in the following figure:
<img src="https://github.com/wangyang59/book/raw/00036f4b0da5225041a6824587c1a01cf20159b1/gan/image/gan_ig.png" width=400 />
......
doc/fluid/design/motivation/refactorization.md
浏览文件 @ fb3bfc3d
......@@ -125,12 +125,12 @@ Compile Time -> IR -> Runtime
## Operator/OpWithKernel/OpKernel
![class_diagram](http://api.paddlepaddle.org/graphviz?dot=https://gist.githubusercontent.com/reyoung/53df507f6749762675dff3e7ce53372f/raw/49caf1fb70820fb4a6c217634317c9306f361f36/op_op_with_kern_class_diagram.dot)
![class_diagram](https://raw.githubusercontent.com/PaddlePaddle/Paddle/develop/doc/fluid/images/op_op_with_kern_class_diagram.dot)
---
## Operator
![class_diagram](http://api.paddlepaddle.org/graphviz?dot=https://gist.githubusercontent.com/reyoung/53df507f6749762675dff3e7ce53372f/raw/dd598e8f1976f5759f58af5e5ef94738a6b2e661/op.dot)
![class_diagram](https://raw.githubusercontent.com/PaddlePaddle/Paddle/develop/doc/fluid/images/op.dot)
* `Operator` is the fundamental building block of the user interface.
* Operator stores input/output variable names and attributes.
......@@ -141,7 +141,7 @@ Compile Time -> IR -> Runtime
## OpWithKernel/Kernel
![class_diagram](http://api.paddlepaddle.org/graphviz?dot=https://gist.githubusercontent.com/reyoung/53df507f6749762675dff3e7ce53372f/raw/9d7f4eba185cf41c8e2fbfb40ae21890dbddcd39/op_with_kernel.dot)
![class_diagram](https://raw.githubusercontent.com/PaddlePaddle/Paddle/develop/doc/fluid/images/op_with_kernel.dot)
* `OpWithKernel` inherits `Operator`.
* `OpWithKernel` contains a Kernel map.
......
doc/fluid/getstarted/index_cn.rst
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/fluid/getstarted/index_en.rst
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/fluid/howto/cluster/nccl2_rdma_training.md 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/fluid/howto/index_cn.rst
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/fluid/howto/index_en.rst
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/fluid/howto/inference/inference_support_in_fluid.md 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/fluid/images/1.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/fluid/images/2.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/fluid/images/3.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/fluid/images/4.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/fluid/images/LoDTensor.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/fluid/images/compile_run_time.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/fluid/images/executor.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/fluid/images/fluid_examples.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/fluid/images/fluid_module_1.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/fluid/images/fluid_module_2.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/fluid/images/layer.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/fluid/images/op.dot 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/fluid/images/op_op_with_kern_class_diagram.dot 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/fluid/images/op_with_kernel.dot 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/fluid/images/operator1.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/fluid/images/operator2.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/fluid/images/place.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/fluid/images/print_fluid_program.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/fluid/images/program_desc1.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/fluid/images/program_desc2.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/fluid/images/raw_input.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/fluid/images/scope_variable_tensor.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/fluid/images/sorted_input.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/fluid/images/transpiler.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/fluid/images/user_interface.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/mobile/CMakeLists.txt
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/mobile/index_cn.rst
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/templates/conf.py.cn.in
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/templates/conf.py.en.in
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/CMakeLists.txt
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/api/CMakeLists.txt
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/api/config/layer.rst
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/api/data/data_reader.rst
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/design/mkl/mkldnn.md
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/FullyConnected.jpg 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/add_security_group.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/bi_lstm.jpg 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/checkpointing.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/create_efs.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/csr.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/data_dispatch.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/dataset.graffle 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/dataset.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/doc_en.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/efs_mount.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/encoder-decoder-attention-model.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/engine.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/file_storage.graffle 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/file_storage.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/glossary_rnn.dot 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/glossary_rnn_with_memory.dot 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/gradients.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/init_lock.graffle 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/init_lock.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/k8s-paddle-arch.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/layers.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/managed_policy.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/matrix.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/nvvp1.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/nvvp2.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/nvvp3.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/nvvp4.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/overview.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/paddle-cloud-in-data-center.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/paddle-etcd.graffle 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/paddle-etcd.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/paddle-model-sharding.graffle 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/paddle-model-sharding.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/paddle-ps-0.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/paddle-ps-1.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/paddle-ps.graffle 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/paddle-task-queues.graffle 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/paddle-task-queues.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/paddle-task-states.graffle 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/paddle-task-states.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/ps_cn.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/ps_en.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/pserver_and_trainer.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/pserver_init.graffle 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/pserver_init.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/route53_create_recordset.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/route53_create_zone.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/sequence_data.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/simple_full_hierarchical_recurrent.dot 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/simple_full_recurrent.dot 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/submit-job.graffle 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/submit-job.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/trainer.graffle 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/trainer.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/trainer_cn.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/worker_security_group.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
doc/v2/images/workflow_of_CAPI.png 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/.gitignore
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/CMakeLists.txt
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/api/GradientMachine.cpp
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/api/PaddleAPI.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/api/SequenceGenerator.cpp
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/capi/CMakeLists.txt
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/capi/Matrix.cpp
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/capi/gradient_machine.cpp
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/contrib/CMakeLists.txt 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/contrib/float16/.gitignore 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/contrib/float16/README.md 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/contrib/float16/float16_benchmark.md 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/contrib/float16/float16_inference_demo.py 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/contrib/float16/float16_transpiler.py 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/contrib/float16/run_float16_demo.sh 0 → 100755
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/contrib/inference/CMakeLists.txt 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/contrib/inference/README.md 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/contrib/inference/paddle_inference_api.cc 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/contrib/inference/paddle_inference_api.h 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/contrib/inference/paddle_inference_api_impl.cc 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/contrib/inference/paddle_inference_api_impl.h 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/contrib/inference/test_paddle_inference_api.cc 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/cuda/CMakeLists.txt
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/cuda/include/hl_base.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/cuda/include/hl_tensor_ops.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/cuda/src/hl_cuda_lstm.cu
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/cuda/src/hl_top_k.cu
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/framework/block_desc.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/framework/data_type.cc 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/framework/data_type.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/framework/details/build_strategy.h 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/framework/details/execution_strategy.h 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/framework/executor.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/framework/executor.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/framework/init.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/framework/init.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/framework/op_desc.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/framework/op_desc.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/framework/operator.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/framework/operator.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/framework/reader.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/inference/analysis/CMakeLists.txt 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/inference/analysis/data_flow_graph.cc 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/inference/analysis/data_flow_graph.h 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/inference/analysis/device.h 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/inference/analysis/dot.cc 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/inference/analysis/dot.h 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/inference/analysis/dot_tester.cc 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/inference/analysis/graph_traits.cc 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/inference/analysis/graph_traits.h 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/inference/analysis/helper.h 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/inference/analysis/node.cc 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/inference/analysis/node.h 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/inference/analysis/node_tester.cc 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/inference/analysis/pass.cc 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/inference/analysis/pass.h 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/inference/analysis/subgraph_splitter.cc 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/inference/analysis/subgraph_splitter.h 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/inference/analysis/ut_helper.h 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/inference/engine.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/inference/io.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/inference/io.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/inference/tensorrt/convert/CMakeLists.txt 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/inference/tensorrt/convert/conv2d_op.cc 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/inference/tensorrt/convert/io_converter.cc 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/inference/tensorrt/convert/io_converter.h 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/inference/tensorrt/convert/mul_op.cc 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/inference/tensorrt/convert/op_converter.h 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/inference/tensorrt/convert/test_mul_op.cc 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/inference/tensorrt/convert/ut_helper.h 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/inference/utils/singleton.h 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/adam_op.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/adam_op.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/adamax_op.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/assign_op.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/auc_op.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/batch_norm_mkldnn_op.cc 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/cast_op.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/cast_op.cu
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/clip_op.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/concat_op.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/concat_op.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/conv_op.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/conv_op.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/crop_op.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/cumsum_op.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/detection/CMakeLists.txt 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/expand_op.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/fake_dequantize_op.cc 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/fake_dequantize_op.h 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/fc_op.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/fc_op.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/feed_op.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/fetch_barrier_op.cc 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/fetch_op.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/fill_op.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/ftrl_op.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/gather_op.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/gen_nccl_id_op.cc 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/go_op.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/gru_op.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/gru_op.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/is_empty_op.h 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/load_op.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/lookup_sparse_table_op.cc 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/lrn_op.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/lstm_op.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/lstm_op.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/lstmp_op.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/lstmp_op.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/math/blas.cc 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/math/blas.h 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/math/blas_impl.cu.h 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/math/blas_impl.h 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/math/matmul.h 已删除 100644 → 0
浏览文件 @ 652332cd
此差异已折叠。
paddle/fluid/operators/matmul_op.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/matmul_op.h 已删除 100644 → 0
浏览文件 @ 652332cd
此差异已折叠。
paddle/fluid/operators/maxout_op.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/mean_op.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/minus_op.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/mul_mkldnn_op.cc 已删除 100644 → 0
浏览文件 @ 652332cd
此差异已折叠。
paddle/fluid/operators/mul_op.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/mul_op.cu.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/mul_op.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/nccl_op.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/nce_op.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/norm_op.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/pad_op.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/pool_op.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/pool_op.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/prelu_op.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/print_op.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/random_crop_op.cc 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/random_crop_op.cu 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/random_crop_op.h 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/read_op.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/recv_op.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/reduce_op.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/reduce_op.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/reshape_op.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/save_op.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/scale_op.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/select_op.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/send_op.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/sgd_op.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/sgd_op.cu
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/sgd_op.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/sign_op.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/split_op.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/spp_op.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/sum_op.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/tensorrt_engine_op.cc 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/tensorrt_engine_op.h 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/test_send_nccl_id.cc 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/top_k_op.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/top_k_op.cu
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/top_k_op.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/unpool_op.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/warpctc_op.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/operators/while_op.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/platform/cuda_device_function.h 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/platform/nccl_helper.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/platform/profiler.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/platform/profiler.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/pybind/const_value.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/pybind/protobuf.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/pybind/pybind.cc
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/pybind/tensor_py.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/train/demo/CMakeLists.txt 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/train/demo/README.md 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/fluid/train/demo/demo_trainer.cc 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/function/BlockExpandOp.cpp
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/function/BufferArg.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/function/CMakeLists.txt
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/function/ConvOp.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/function/CosSimOp.cpp
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/function/CropOp.cpp
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/function/DepthwiseConvOp.cpp
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/function/DepthwiseConvOp.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/function/EigenGemm.cpp
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/function/EigenThreadDevice.h 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/function/Function.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/function/FunctionTest.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/function/GemmConvOp.cpp
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/function/Im2Col.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/function/Im2ColOp.cpp
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/function/Im2ColOpGpu.cu
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/function/MulOp.cpp
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/function/NaiveConvOp.cpp
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/function/PadOp.cpp
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/function/RowConvOp.cpp
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/function/RowConvOpGpu.cu
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/function/SwitchOp.cpp
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/function/TensorShape.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/gserver/CMakeLists.txt
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/gserver/layers/AddtoLayer.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/gserver/layers/AgentLayer.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/gserver/layers/CRFLayer.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/gserver/layers/CTCLayer.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/gserver/layers/ClipLayer.cpp
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/gserver/layers/Conv3DLayer.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/gserver/layers/CosSimLayer.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/gserver/layers/CostLayer.cpp
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/gserver/layers/CostLayer.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/gserver/layers/CropLayer.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/gserver/layers/DataLayer.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/gserver/layers/ExpandLayer.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/gserver/layers/GruCompute.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/gserver/layers/Layer.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/gserver/layers/LstmCompute.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/gserver/layers/LstmLayer.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/gserver/layers/MKLDNNBase.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/gserver/layers/MKLDNNLayer.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/gserver/layers/MaxLayer.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/gserver/layers/MaxOutLayer.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/gserver/layers/MixedLayer.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/gserver/layers/NCELayer.cpp
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/gserver/layers/NormLayer.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/gserver/layers/Operator.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/gserver/layers/PadLayer.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/gserver/layers/Pool3DLayer.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/gserver/layers/PoolLayer.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/gserver/layers/PriorBox.cpp
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/gserver/layers/Projection.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/gserver/layers/RotateLayer.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/gserver/layers/TensorLayer.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/gserver/layers/TransLayer.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/gserver/tests/MKLDNNTester.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/math/Allocator.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/math/BaseMatrix.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/math/CMakeLists.txt
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/math/CpuSparseMatrix.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/math/ExecViaCpu.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/math/MKLDNNMatrix.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/math/MathFunctions.cpp
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/math/Matrix.cpp
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/math/Matrix.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/math/MatrixBitCode.cpp
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/math/MemoryHandle.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/math/PoolAllocator.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/math/RowBuffer.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/math/SparseMatrix.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/math/SparseRowMatrix.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/math/Storage.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/math/TensorApply.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/math/TensorAssign.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/math/TensorExpression.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/math/Vector.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/math/tests/TensorCheck.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/math/tests/TestUtils.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/optimizer/CMakeLists.txt
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/optimizer/adam_optimizer.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/optimizer/lr_policy.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/optimizer/sgd_optimizer.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/optimizer/tensor.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/parameter/AverageOptimizer.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/parameter/CMakeLists.txt
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/parameter/Parameter.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/parameter/Regularizer.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/parameter/Weight.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/pserver/BaseClient.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/pserver/CMakeLists.txt
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/pserver/LightNetwork.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/pserver/ParameterClient2.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/pserver/ParameterServer2.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/pserver/ProtoServer.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/pserver/SocketChannel.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/pserver/test/SocketTest.cpp
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/scripts/docker/build.sh 已删除 100755 → 0
浏览文件 @ 652332cd
此差异已折叠。
paddle/scripts/docker/build_android.sh 已删除 100644 → 0
浏览文件 @ 652332cd
此差异已折叠。
paddle/scripts/docker/entrypoint 已删除 100755 → 0
浏览文件 @ 652332cd
此差异已折叠。
paddle/scripts/docker/test.sh 已删除 100755 → 0
浏览文件 @ 652332cd
此差异已折叠。
paddle/scripts/paddle_build.sh
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/scripts/travis/build_doc.sh 已删除 100755 → 0
浏览文件 @ 652332cd
此差异已折叠。
paddle/scripts/travis/build_ios.sh 已删除 100755 → 0
浏览文件 @ 652332cd
此差异已折叠。
paddle/scripts/travis/check_style.sh 已删除 100755 → 0
浏览文件 @ 652332cd
此差异已折叠。
paddle/scripts/travis/deploy_key.enc 已删除 100644 → 0
浏览文件 @ 652332cd
此差异已折叠。
paddle/trainer/CMakeLists.txt
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/trainer/ParamUtil.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/trainer/ParameterUpdater.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/trainer/Tester.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/trainer/Trainer.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/trainer/TrainerInternal.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/trainer/tests/picojson.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/utils/CMakeLists.txt
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/utils/ClassRegistrar.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/utils/CpuId.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/utils/CustomStackTrace.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/utils/Error.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/utils/GlobalConstants.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/utils/Locks.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/utils/PythonUtil.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/utils/Queue.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/utils/Stat.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/utils/Thread.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/utils/ThreadLocal.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/utils/Util.h
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/utils/arch/linux/Locks.cpp
浏览文件 @ fb3bfc3d
此差异已折叠。
paddle/utils/arch/osx/Locks.cpp
浏览文件 @ fb3bfc3d
此差异已折叠。
python/paddle/dataset/wmt16.py
浏览文件 @ fb3bfc3d
此差异已折叠。
python/paddle/fluid/__init__.py
浏览文件 @ fb3bfc3d
此差异已折叠。
python/paddle/fluid/backward.py
浏览文件 @ fb3bfc3d
此差异已折叠。
python/paddle/fluid/clip.py
浏览文件 @ fb3bfc3d
此差异已折叠。
python/paddle/fluid/data_feeder.py
浏览文件 @ fb3bfc3d
此差异已折叠。
python/paddle/fluid/evaluator.py
浏览文件 @ fb3bfc3d
此差异已折叠。
python/paddle/fluid/executor.py
浏览文件 @ fb3bfc3d
此差异已折叠。
python/paddle/fluid/framework.py
浏览文件 @ fb3bfc3d
此差异已折叠。
python/paddle/fluid/inferencer.py 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
python/paddle/fluid/io.py
浏览文件 @ fb3bfc3d
此差异已折叠。
python/paddle/fluid/layer_helper.py
浏览文件 @ fb3bfc3d
此差异已折叠。
python/paddle/fluid/layers/io.py
浏览文件 @ fb3bfc3d
此差异已折叠。
python/paddle/fluid/layers/nn.py
浏览文件 @ fb3bfc3d
此差异已折叠。
python/paddle/fluid/lod_tensor.py 0 → 100644
浏览文件 @ fb3bfc3d
此差异已折叠。
python/paddle/fluid/metrics.py
浏览文件 @ fb3bfc3d
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