Skip to content

P
PaddleDetection

s920243400 / PaddleDetection
与 Fork 源项目一致

Fork自 PaddlePaddle / PaddleDetection

通知 2
Star 0
Fork 0
P
PaddleDetection
提交 fab63cc6 编写于 作者: T typhoonzero
浏览文件
操作

update design doc

上级 1ce2d613 9cfa5ce3
展开全部 隐藏空白更改
内联 并排
Showing with 37627 addition and 23609 deletion
.clang-format
浏览文件 @ fab63cc6
......@@ -25,4 +25,3 @@ AllowAllParametersOfDeclarationOnNextLine: true
BinPackParameters: false
BinPackArguments: false
...
.gitignore
浏览文件 @ fab63cc6
......@@ -28,3 +28,4 @@ cmake_install.cmake
paddle/.timestamp
python/paddlepaddle.egg-info/
paddle/pybind/pybind.h
python/paddle/version.py
.travis.yml
浏览文件 @ fab63cc6
......@@ -42,7 +42,7 @@ before_install:
script:
- |
timeout 2580 paddle/scripts/travis/${JOB}.sh # 43min timeout
RESULT=$?; if [ $RESULT -eq 0 ] || [ $RESULT -eq 142 ]; then true; else false; fi;
RESULT=$?; if [ $RESULT -eq 0 ] || [ $RESULT -eq 142 ]; then true ;else exit 1; fi;
- |
if [[ "$JOB" != "build_doc" ]]; then exit 0; fi;
if [[ "$TRAVIS_PULL_REQUEST" != "false" ]]; then exit 0; fi;
......
CMakeLists.txt
浏览文件 @ fab63cc6
......@@ -20,6 +20,8 @@ set(PADDLE_BINARY_DIR ${CMAKE_CURRENT_BINARY_DIR})
include(system)
project(paddle CXX C Go)
message(STATUS "CXX compiler: " ${CMAKE_CXX_COMPILER} ", version: " ${CMAKE_CXX_COMPILER_VERSION})
message(STATUS "C compiler: " ${CMAKE_C_COMPILER} ", version: " ${CMAKE_C_COMPILER_VERSION})
find_package(Sphinx)
if(NOT CMAKE_CROSSCOMPILING)
......@@ -54,7 +56,9 @@ option(WITH_C_API "Compile PaddlePaddle with C-API(Prediction)" OFF)
option(WITH_GOLANG "Compile PaddlePaddle with GOLANG" OFF)
option(GLIDE_INSTALL "Download and install go dependencies " ON)
option(USE_NNPACK "Compile PaddlePaddle with NNPACK library" OFF)
option(WITH_DISTRIBUTE "Compile with grpc distributed support" OFF)
option(USE_EIGEN_FOR_BLAS "Use matrix multiplication in Eigen" OFF)
option(WITH_ARM_FP16 "Use half precision support on armv8.2-a cpu" OFF)
# CMAKE_BUILD_TYPE
if(NOT CMAKE_BUILD_TYPE)
......@@ -67,9 +71,6 @@ if(ANDROID OR IOS)
if(ANDROID)
if(${CMAKE_SYSTEM_VERSION} VERSION_LESS "16")
message(FATAL_ERROR "Unsupport standalone toolchains with Android API level lower than 16")
elseif(${CMAKE_SYSTEM_VERSION} VERSION_LESS "21")
# TODO: support glog for Android api 16 ~ 19 in the future
message(WARNING "Using the unofficial git repository <https://github.com/Xreki/glog.git> instead")
endif()
endif()
......@@ -83,6 +84,8 @@ if(ANDROID OR IOS)
"Disable RDMA when cross-compiling for Android and iOS" FORCE)
set(WITH_MKL OFF CACHE STRING
"Disable MKL when cross-compiling for Android and iOS" FORCE)
set(WITH_GOLANG OFF CACHE STRING
"Disable golang when cross-compiling for Android and iOS" FORCE)
# Compile PaddlePaddle mobile inference library
if (NOT WITH_C_API)
......@@ -133,6 +136,8 @@ include(external/any) # download libn::any
include(external/eigen) # download eigen3
include(external/pybind11) # download pybind11
include(external/nccl)
include(external/cares)
include(external/grpc)
include(cudnn) # set cudnn libraries, must before configure
include(configure) # add paddle env configuration
......@@ -194,6 +199,10 @@ if(WITH_GOLANG)
endif(WITH_GOLANG)
set(PADDLE_PYTHON_BUILD_DIR "${CMAKE_CURRENT_BINARY_DIR}/python/build")
SET(CMAKE_CXX_FLAGS_RELWITHDEBINFO "-O3 -g -DNDEBUG")
SET(CMAKE_C_FLAGS_RELWITHDEBINFO "-O3 -g -DNDEBUG")
add_subdirectory(paddle)
if(WITH_PYTHON)
add_subdirectory(python)
......
Dockerfile
浏览文件 @ fab63cc6
......@@ -29,7 +29,7 @@ RUN apt-get update && \
automake locales clang-format swig doxygen cmake \
liblapack-dev liblapacke-dev libboost-dev \
clang-3.8 llvm-3.8 libclang-3.8-dev \
net-tools && \
net-tools libtool && \
apt-get clean -y
# Install Go and glide
......
README.md
浏览文件 @ fab63cc6
......@@ -2,8 +2,8 @@
[![Build Status](https://travis-ci.org/PaddlePaddle/Paddle.svg?branch=develop)](https://travis-ci.org/PaddlePaddle/Paddle)
[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://doc.paddlepaddle.org/develop/doc/)
[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://doc.paddlepaddle.org/develop/doc_cn/)
[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://www.paddlepaddle.org/docs/develop/documentation/en/getstarted/index_en.html)
[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://www.paddlepaddle.org/docs/develop/documentation/zh/getstarted/index_cn.html)
[![Coverage Status](https://coveralls.io/repos/github/PaddlePaddle/Paddle/badge.svg?branch=develop)](https://coveralls.io/github/PaddlePaddle/Paddle?branch=develop)
[![Release](https://img.shields.io/github/release/PaddlePaddle/Paddle.svg)](https://github.com/PaddlePaddle/Paddle/releases)
[![License](https://img.shields.io/badge/license-Apache%202-blue.svg)](LICENSE)
......@@ -36,7 +36,7 @@ Please refer to our [release announcement](https://github.com/PaddlePaddle/Paddl
examples:
- Optimized math operations through SSE/AVX intrinsics, BLAS libraries
(e.g. MKL, ATLAS, cuBLAS) or customized CPU/GPU kernels.
(e.g. MKL, OpenBLAS, cuBLAS) or customized CPU/GPU kernels.
- Highly optimized recurrent networks which can handle **variable-length**
sequence without padding.
- Optimized local and distributed training for models with high dimensional
......@@ -61,32 +61,32 @@ Please refer to our [release announcement](https://github.com/PaddlePaddle/Paddl
## Installation
It is recommended to check out the
[Docker installation guide](http://doc.paddlepaddle.org/develop/doc/getstarted/build_and_install/docker_install_en.html)
[Docker installation guide](http://www.paddlepaddle.org/docs/develop/documentation/en/getstarted/build_and_install/docker_install_en.html)
before looking into the
[build from source guide](http://doc.paddlepaddle.org/develop/doc/getstarted/build_and_install/build_from_source_en.html).
[build from source guide](http://www.paddlepaddle.org/docs/develop/documentation/en/getstarted/build_and_install/build_from_source_en.html).
## Documentation
We provide [English](http://doc.paddlepaddle.org/develop/doc/) and
[Chinese](http://doc.paddlepaddle.org/doc_cn/) documentation.
We provide [English](http://www.paddlepaddle.org/docs/develop/documentation/en/getstarted/index_en.html) and
[Chinese](http://www.paddlepaddle.org/docs/develop/documentation/zh/getstarted/index_cn.html) documentation.
- [Deep Learning 101](http://book.paddlepaddle.org/index.html)
- [Deep Learning 101](http://www.paddlepaddle.org/docs/develop/book/01.fit_a_line/index.html)
You might want to start from this online interactive book that can run in a Jupyter Notebook.
- [Distributed Training](http://doc.paddlepaddle.org/develop/doc/howto/usage/cluster/cluster_train_en.html)
- [Distributed Training](http://www.paddlepaddle.org/docs/develop/documentation/en/howto/usage/cluster/cluster_train_en.html)
You can run distributed training jobs on MPI clusters.
- [Distributed Training on Kubernetes](http://doc.paddlepaddle.org/develop/doc/howto/usage/k8s/k8s_en.html)
- [Distributed Training on Kubernetes](http://www.paddlepaddle.org/docs/develop/documentation/en/howto/usage/cluster/k8s_en.html)
You can also run distributed training jobs on Kubernetes clusters.
- [Python API](http://doc.paddlepaddle.org/develop/doc/api/index_en.html)
- [Python API](http://www.paddlepaddle.org/docs/develop/documentation/en/api/index_en.html)
Our new API enables much shorter programs.
- [How to Contribute](http://doc.paddlepaddle.org/develop/doc/howto/dev/contribute_to_paddle_en.html)
- [How to Contribute](http://www.paddlepaddle.org/docs/develop/documentation/en/howto/dev/contribute_to_paddle_en.html)
We appreciate your contributions!
......
RELEASE.cn.md
浏览文件 @ fab63cc6
# v0.11.0版本
## PaddlePaddle Fluid
- PaddlePaddle发布版本v0.11.0包含一个新的特性*PaddlePaddle Fluid*. Fluid 是设计用来让用户像Pytorch和Tensorflow Eager Execution一样执行程序。在这些系统中，不再有*模型*这个概念，应用也不再包含一个用于描述Operator图或者一系列层的符号描述，而是像通用程序那样描述训练或者预测的过程。而Fluid与PyTorch或Eager Execution的区别在于Fluid不依赖Python提供的控制流，例如 if-else-then或者for，而是提供了基于C++实现的控制流并暴露了对应的用with语法实现的Python接口。例如：
https://github.com/PaddlePaddle/Paddle/blob/3df78ed2a98d37f7ae6725894cc7514effd5664b/python/paddle/v2/fluid/tests/test_while_op.py#L36-L44
- 在v0.11.0版本中，我们提供了一个C++类`Executor`用于运行一个Fluid程序。Executor类似一个解释器。在未来的版本中，我们将提升和优化Executor成为一个调试器，就像GDB。并可能提供一些编译器，这个编译器会读取一个上文所描述的应用然后编译成一个等价的
源代码，这个源代码可以被nvcc编译成可以使用CUDA的二进制，或者被icc编译成可以充分利用Intel CPU的二进制。
## 新特点
* 发布 `PaddlePaddle Fluid`。
* 增加了用于模型预测的C-API。
* 用Fluid API实现了一个简单的GAN的例子。
* 增加了关于性能调优的文档。
* 为`paddle.v2.dataset`下载数据集提供了重试机制.
* C++中使用protobuf-lite替换protobuf减少了二进制的大小。
* 发布了新特性 [Elastic Deep Learning (EDL)](https://github.com/PaddlePaddle/cloud/tree/develop/doc/autoscale/experiment).
* 基于Bazel API利用cmake实现了一个的新的构建系统函数库。
* 当使用编译选项`WITH_MKL=ON`时自动下载和编译Intel® [MKLML](https://github.com/01org/mkl-dnn/releases/download/v0.11/mklml_lnx_2018.0.1.20171007.tgz) 函数库.
* [Intel® MKL-DNN on PaddlePaddle](https://github.com/PaddlePaddle/Paddle/tree/develop/doc/design/mkldnn):
- 完成了 11个 MKL-DNN 层: Convolution, Fully connectivity, Pooling, ReLU, Tanh, ELU, Softmax, BatchNorm, AddTo, Concat, LRN。
- 完成了 3个 MKL-DNN 网络: VGG-19, ResNet-50, GoogleNet
- 基于Intel Skylake 6148 CPU的[性能测试](https://github.com/PaddlePaddle/Paddle/blob/develop/benchmark/IntelOptimizedPaddle.md) : 相对于MKLML有2~3倍的训练加速。
* 增加 [softsign activation](http://www.paddlepaddle.org/docs/develop/documentation/zh/api/v2/config/activation.html#softsign)
* 增加 [dot product layer](http://www.paddlepaddle.org/docs/develop/documentation/zh/api/v2/config/layer.html#dot-prod)
* 增加 [L2 distance layer](http://www.paddlepaddle.org/docs/develop/documentation/zh/api/v2/config/layer.html#l2-distance)
* 增加 [sub-nested sequence layer](http://www.paddlepaddle.org/docs/develop/documentation/zh/api/v2/config/layer.html#sub-nested-seq)
* 增加 [kmax sequence score layer](http://www.paddlepaddle.org/docs/develop/documentation/zh/api/v2/config/layer.html#kmax-sequence-score)
* 增加 [sequence slice layer](http://www.paddlepaddle.org/docs/develop/documentation/zh/api/v2/config/layer.html#seq-slice)
* 增加 [row convolution layer](http://www.paddlepaddle.org/docs/develop/documentation/zh/api/v2/config/layer.html#row-conv)
* 增加移动端友好的网页
## 改进
* 使用一个Python`whl`包即可安装.
* [V2 API可以实现用户定制化评估](https://github.com/PaddlePaddle/models/tree/develop/ltr#训练过程中输出自定义评估指标)。
* 将 `PADDLE_ONLY_CPU` 改为 `PADDLE_WITH_GPU`, 因为我们会支持多种设备。
* 删除了有一些bug的BarrierStat。
* 清理和删除了paddle::Parameter中未使用的函数。
* 删除了ProtoDataProvider。
* Huber loss同时支持回归和分类。
* 为sequence pooling 层增加`stride`参数。
* v2 API自动使用cudnn batch normalization。
* 可以使用一个固定的参数名共享BN层的参数。
* 2D convolution operation支持variable-dimension input特性。
* 重构cmake中关于CUDA的部分并实现自动检测GPU架构的功能。
* 优化网页导航。
## 错误修复
* 修复ROI pooling的Bug. cc9a761
* 修复当label是dense vector是AUC变成0的问题. #5274
* 修复WarpCTC 层的Bug.
# v0.10.0版本
我们非常高兴发布了PaddlePaddle V0.10.0版，并开发了新的[Python API](http://research.baidu.com/paddlepaddles-new-api-simplifies-deep-learning-programs/)。
......
RELEASE.md
浏览文件 @ fab63cc6
# Release v0.11.0
## PaddlePaddle Fluid
- Release 0.11.0 includes a new feature *PaddlePaddle Fluid*. Fluid is
designed to allow users to program like PyTorch and TensorFlow Eager Execution.
In these systems, there is no longer the concept *model* and applications
do not include a symbolic description of a graph of operators nor a sequence
of layers. Instead, applications look exactly like a usual program that
describes a process of training or inference. The difference between
Fluid and PyTorch or Eager Execution is that Fluid doesn't rely on Python's
control-flow, `if-then-else` nor `for`. Instead, Fluid provides its
C++ implementations and their Python binding using the `with` statement. For an example
https://github.com/PaddlePaddle/Paddle/blob/3df78ed2a98d37f7ae6725894cc7514effd5664b/python/paddle/v2/fluid/tests/test_while_op.py#L36-L44
- In 0.11.0, we provides a C++ class `Executor` to run a Fluid program.
Executor works like an interpreter. In future version, we will improve
`Executor` into a debugger like GDB, and we might provide some compilers,
which, for example, takes an application like the above one, and outputs
an equivalent C++ source program, which can be compiled using
[`nvcc`](http://docs.nvidia.com/cuda/cuda-compiler-driver-nvcc/index.html)
to generate binaries that use CUDA, or using
[`icc`](https://software.intel.com/en-us/c-compilers) to generate binaries
that make full use of Intel CPUs.
## New Features
* Release `PaddlePaddle Fluid`.
* Add C-API for model inference
* Use fluid API to create a simple GAN demo.
* Add develop guide about performance tunning.
* Add retry when download `paddle.v2.dataset`.
* Linking protobuf-lite not protobuf in C++. Reduce the binary size.
* Feature [Elastic Deep Learning (EDL)](https://github.com/PaddlePaddle/cloud/tree/develop/doc/autoscale/experiment) released.
* A new style cmake functions for Paddle. It is based on Bazel API.
* Automatically download and compile with Intel® [MKLML](https://github.com/01org/mkl-dnn/releases/download/v0.11/mklml_lnx_2018.0.1.20171007.tgz) library as CBLAS when build `WITH_MKL=ON`.
* [Intel® MKL-DNN on PaddlePaddle](https://github.com/PaddlePaddle/Paddle/tree/develop/doc/design/mkldnn):
- Complete 11 MKL-DNN layers: Convolution, Fully connectivity, Pooling, ReLU, Tanh, ELU, Softmax, BatchNorm, AddTo, Concat, LRN.
- Complete 3 MKL-DNN networks: VGG-19, ResNet-50, GoogleNet
- [Benchmark](https://github.com/PaddlePaddle/Paddle/blob/develop/benchmark/IntelOptimizedPaddle.md) on Intel Skylake 6148 CPU: 2~3x training speedup compared with MKLML.
* Add the [`softsign` activation](http://www.paddlepaddle.org/docs/develop/documentation/zh/api/v2/config/activation.html#softsign).
* Add the [dot product layer](http://www.paddlepaddle.org/docs/develop/documentation/zh/api/v2/config/layer.html#dot-prod).
* Add the [L2 distance layer](http://www.paddlepaddle.org/docs/develop/documentation/zh/api/v2/config/layer.html#l2-distance).
* Add the [sub-nested sequence layer](http://www.paddlepaddle.org/docs/develop/documentation/zh/api/v2/config/layer.html#sub-nested-seq).
* Add the [kmax sequence score layer](http://www.paddlepaddle.org/docs/develop/documentation/zh/api/v2/config/layer.html#kmax-sequence-score).
* Add the [sequence slice layer](http://www.paddlepaddle.org/docs/develop/documentation/zh/api/v2/config/layer.html#seq-slice).
* Add the [row convolution layer](http://www.paddlepaddle.org/docs/develop/documentation/zh/api/v2/config/layer.html#row-conv)
* Add mobile friendly webpages.
## Improvements
* Build and install using a single `whl` package.
* [Custom evaluating in V2 API](https://github.com/PaddlePaddle/models/tree/develop/ltr#训练过程中输出自定义评估指标).
* Change `PADDLE_ONLY_CPU` to `PADDLE_WITH_GPU`, since we will support many kinds of devices.
* Remove buggy BarrierStat.
* Clean and remove unused functions in paddle::Parameter.
* Remove ProtoDataProvider.
* Huber loss supports both regression and classification.
* Add the `stride` parameter for sequence pooling layers.
* Enable v2 API use cudnn batch normalization automatically.
* The BN layer's parameter can be shared by a fixed the parameter name.
* Support variable-dimension input feature for 2D convolution operation.
* Refine cmake about CUDA to automatically detect GPU architecture.
* Improved website navigation.
## Bug Fixes
* Fix bug in ROI pooling. cc9a761
* Fix AUC is zero when label is dense vector. #5274
* Fix bug in WarpCTC layer.
# Release v0.10.0
We are glad to release version 0.10.0. In this version, we are happy to release the new
......
benchmark/IntelOptimizedPaddle.md
浏览文件 @ fab63cc6
......@@ -2,28 +2,27 @@
Machine:
- Server
- Intel(R) Xeon(R) Gold 6148 CPU @ 2.40GHz, 2 Sockets, 20 Cores per socket
- Laptop
- DELL XPS15-9560-R1745: i7-7700HQ 8G 256GSSD
- i5 MacBook Pro (Retina, 13-inch, Early 2015)
- Desktop
- i7-6700k
- Server: Intel(R) Xeon(R) Gold 6148 CPU @ 2.40GHz, 2 Sockets, 20 Cores per socket
- Laptop: TBD
System: CentOS release 6.3 (Final), Docker 1.12.1.
PaddlePaddle: paddlepaddle/paddle:latest (TODO: will rerun after 0.11.0)
- MKL-DNN tag v0.10
- MKLML 2018.0.20170720
- OpenBLAS v0.2.20
PaddlePaddle: (TODO: will rerun after 0.11.0)
- paddlepaddle/paddle:latest (for MKLML and MKL-DNN)
- MKL-DNN tag v0.11
- MKLML 2018.0.1.20171007
- paddlepaddle/paddle:latest-openblas (for OpenBLAS)
- OpenBLAS v0.2.20
On each machine, we will test and compare the performance of training on single node using MKL-DNN / MKLML / OpenBLAS respectively.
## Benchmark Model
### Server
#### Training
Test on batch size 64, 128, 256 on Intel(R) Xeon(R) Gold 6148 CPU @ 2.40GHz
Pay attetion that the speed below includes forward, backward and parameter update time. So we can not directly compare the data with the benchmark of caffe `time` [command](https://github.com/PaddlePaddle/Paddle/blob/develop/benchmark/caffe/image/run.sh#L9), which only contain forward and backward. The updating time of parameter would become very heavy when the weight size are large, especially on alexnet.
Input image size - 3 * 224 * 224, Time: images/second
......@@ -31,18 +30,78 @@ Input image size - 3 * 224 * 224, Time: images/second
| BatchSize | 64 | 128 | 256 |
|--------------|-------| -----| --------|
| OpenBLAS | 7.82 | 8.62 | 10.34 |
| MKLML | 11.02 | 12.86 | 15.33 |
| MKL-DNN | 27.69 | 28.8 | 29.27 |
| OpenBLAS | 7.80 | 9.00 | 10.80 |
| MKLML | 12.12 | 13.70 | 16.18 |
| MKL-DNN | 28.46 | 29.83 | 30.44 |
<img src="figs/vgg-cpu-train.png" width="500">
chart on batch size 128
TBD
- ResNet-50
| BatchSize | 64 | 128 | 256 |
|--------------|-------| ------| -------|
| OpenBLAS | 25.22 | 25.68 | 27.12 |
| MKLML | 32.52 | 31.89 | 33.12 |
| MKL-DNN | 81.69 | 82.35 | 84.08 |
<img src="figs/resnet-cpu-train.png" width="500">
- ResNet
- GoogLeNet
| BatchSize | 64 | 128 | 256 |
|--------------|-------| ------| -------|
| OpenBLAS | 89.52 | 96.97 | 108.25 |
| MKLML | 128.46| 137.89| 158.63 |
| MKL-DNN     | 250.46| 264.83| 269.50 |
<img src="figs/googlenet-cpu-train.png" width="500">
- Alexnet
| BatchSize | 64 | 128 | 256 |
|--------------|--------| ------ | -------|
| OpenBLAS | 2.13 | 2.45 | 2.68 |
| MKLML | 66.37 | 105.60 | 144.04 |
| MKL-DNN | 399.00 | 498.94 | 626.53 |
chart TBD
#### Inference
Test on batch size 1, 2, 4, 8, 16 on Intel(R) Xeon(R) Gold 6148 CPU @ 2.40GHz
- VGG-19
| BatchSize | 1 | 2 | 4 | 8 | 16 |
|-----------|-------|-------|-------|-------|-------|
| OpenBLAS | 1.07 | 1.08 | 1.06 | 0.88 | 0.65 |
| MKLML | 5.58 | 9.80 | 15.15 | 21.21 | 28.67 |
| MKL-DNN | 75.07 | 88.64 | 82.58 | 92.29 | 96.75 |
- ResNet-50
| BatchSize | 1 | 2 | 4 | 8 | 16 |
|-----------|-------|--------|--------|--------|--------|
| OpenBLAS | 3.35 | 3.19 | 3.09 | 2.55 | 1.96 |
| MKLML | 6.33 | 12.02 | 22.88 | 40.53 | 63.09 |
| MKL-DNN | 107.83| 148.84 | 177.78 | 189.35 | 217.69 |
- GoogLeNet
| BatchSize | 1 | 2 | 4 | 8 | 16 |
|-----------|--------|--------|--------|--------|--------|
| OpenBLAS | 12.04 | 11.31 | 10.00 | 9.07 | 4.34 |
| MKLML | 22.74 | 41.56 | 81.22 | 133.47 | 210.53 |
| MKL-DNN | 175.10 | 272.92 | 450.70 | 512.00 | 600.94 |
- Alexnet
| BatchSize | 1 | 2 | 4 | 8 | 16 |
|-----------|--------|--------|--------|--------|--------|
| OpenBLAS | | | | | |
| MKLML | 21.32 | 36.55 | 73.06 | 131.15 | 192.77 |
| MKL-DNN | 442.91 | 656.41 | 719.10 | 847.68 | 850.51 |
chart TBD
### Laptop
TBD
### Desktop
TBD
benchmark/paddle/image/alexnet.py
浏览文件 @ fab63cc6
......@@ -6,10 +6,24 @@ height = 227
width = 227
num_class = 1000
batch_size = get_config_arg('batch_size', int, 128)
gp = get_config_arg('layer_num', int, 1)
is_infer = get_config_arg("is_infer", bool, False)
num_samples = get_config_arg('num_samples', int, 2560)
args = {'height': height, 'width': width, 'color': True, 'num_class': num_class}
args = {
'height': height,
'width': width,
'color': True,
'num_class': num_class,
'is_infer': is_infer,
'num_samples': num_samples
}
define_py_data_sources2(
"train.list", None, module="provider", obj="process", args=args)
"train.list" if not is_infer else None,
"test.list" if is_infer else None,
module="provider",
obj="process",
args=args)
settings(
batch_size=batch_size,
......@@ -31,7 +45,7 @@ net = img_pool_layer(input=net, pool_size=3, stride=2)
# conv2
net = img_conv_layer(
input=net, filter_size=5, num_filters=256, stride=1, padding=2, groups=1)
input=net, filter_size=5, num_filters=256, stride=1, padding=2, groups=gp)
net = img_cmrnorm_layer(input=net, size=5, scale=0.0001, power=0.75)
net = img_pool_layer(input=net, pool_size=3, stride=2)
......@@ -40,11 +54,11 @@ net = img_conv_layer(
input=net, filter_size=3, num_filters=384, stride=1, padding=1)
# conv4
net = img_conv_layer(
input=net, filter_size=3, num_filters=384, stride=1, padding=1, groups=1)
input=net, filter_size=3, num_filters=384, stride=1, padding=1, groups=gp)
# conv5
net = img_conv_layer(
input=net, filter_size=3, num_filters=256, stride=1, padding=1, groups=1)
input=net, filter_size=3, num_filters=256, stride=1, padding=1, groups=gp)
net = img_pool_layer(input=net, pool_size=3, stride=2)
net = fc_layer(
......@@ -59,6 +73,9 @@ net = fc_layer(
layer_attr=ExtraAttr(drop_rate=0.5))
net = fc_layer(input=net, size=1000, act=SoftmaxActivation())
lab = data_layer('label', num_class)
loss = cross_entropy(input=net, label=lab)
outputs(loss)
if is_infer:
outputs(net)
else:
lab = data_layer('label', num_class)
loss = cross_entropy(input=net, label=lab)
outputs(loss)
benchmark/paddle/image/googlenet.py
浏览文件 @ fab63cc6
......@@ -5,10 +5,24 @@ height = 224
width = 224
num_class = 1000
batch_size = get_config_arg('batch_size', int, 128)
args = {'height': height, 'width': width, 'color': True, 'num_class': num_class}
use_gpu = get_config_arg('use_gpu', bool, True)
is_infer = get_config_arg("is_infer", bool, False)
num_samples = get_config_arg('num_samples', int, 2560)
args = {
'height': height,
'width': width,
'color': True,
'num_class': num_class,
'is_infer': is_infer,
'num_samples': num_samples
}
define_py_data_sources2(
"train.list", None, module="provider", obj="process", args=args)
"train.list" if not is_infer else None,
"test.list" if is_infer else None,
module="provider",
obj="process",
args=args)
settings(
batch_size=batch_size,
......@@ -16,6 +30,8 @@ settings(
learning_method=MomentumOptimizer(0.9),
regularization=L2Regularization(0.0005 * batch_size))
conv_projection = conv_projection if use_gpu else img_conv_layer
def inception2(name, input, channels, \
filter1,
filter3R, filter3,
......@@ -138,12 +154,11 @@ def inception(name, input, channels, \
cat = concat_layer(
name=name,
input=[cov1, cov3, cov5, covprj],
bias_attr=True,
bias_attr=True if use_gpu else False,
act=ReluActivation())
return cat
lab = data_layer(name="label", size=1000)
data = data_layer(name="input", size=3 * height * width)
# stage 1
......@@ -221,6 +236,10 @@ pool5 = img_pool_layer(
dropout = dropout_layer(name="dropout", input=pool5, dropout_rate=0.4)
out3 = fc_layer(
name="output3", input=dropout, size=1000, act=SoftmaxActivation())
loss3 = cross_entropy(name='loss3', input=out3, label=lab)
outputs(loss3)
if is_infer:
outputs(out3)
else:
lab = data_layer(name="label", size=num_class)
loss3 = cross_entropy(name='loss3', input=out3, label=lab)
outputs(loss3)
benchmark/paddle/image/provider.py
浏览文件 @ fab63cc6
......@@ -13,14 +13,21 @@ def initHook(settings, height, width, color, num_class, **kwargs):
settings.data_size = settings.height * settings.width * 3
else:
settings.data_size = settings.height * settings.width
settings.slots = [dense_vector(settings.data_size), integer_value(1)]
settings.is_infer = kwargs.get('is_infer', False)
settings.num_samples = kwargs.get('num_samples', 2560)
if settings.is_infer:
settings.slots = [dense_vector(settings.data_size)]
else:
settings.slots = [dense_vector(settings.data_size), integer_value(1)]
@provider(
init_hook=initHook, min_pool_size=-1, cache=CacheType.CACHE_PASS_IN_MEM)
def process(settings, file_list):
for i in xrange(1024):
for i in xrange(settings.num_samples):
img = np.random.rand(1, settings.data_size).reshape(-1, 1).flatten()
lab = random.randint(0, settings.num_class - 1)
yield img.astype('float32'), int(lab)
if settings.is_infer:
yield img.astype('float32')
else:
lab = random.randint(0, settings.num_class - 1)
yield img.astype('float32'), int(lab)
benchmark/paddle/image/resnet.py
浏览文件 @ fab63cc6
......@@ -6,11 +6,23 @@ width = 224
num_class = 1000
batch_size = get_config_arg('batch_size', int, 64)
layer_num = get_config_arg("layer_num", int, 50)
is_test = get_config_arg("is_test", bool, False)
args = {'height': height, 'width': width, 'color': True, 'num_class': num_class}
is_infer = get_config_arg("is_infer", bool, False)
num_samples = get_config_arg('num_samples', int, 2560)
args = {
'height': height,
'width': width,
'color': True,
'num_class': num_class,
'is_infer': is_infer,
'num_samples': num_samples
}
define_py_data_sources2(
"train.list", None, module="provider", obj="process", args=args)
"train.list" if not is_infer else None,
"test.list" if is_infer else None,
module="provider",
obj="process",
args=args)
settings(
batch_size=batch_size,
......@@ -45,7 +57,10 @@ def conv_bn_layer(name,
act=LinearActivation(),
bias_attr=False)
return batch_norm_layer(
name=name + "_bn", input=tmp, act=active_type, use_global_stats=is_test)
name=name + "_bn",
input=tmp,
act=active_type,
use_global_stats=is_infer)
def bottleneck_block(name, input, num_filters1, num_filters2):
......@@ -207,7 +222,9 @@ elif layer_num == 152:
else:
print("Wrong layer number.")
lbl = data_layer(name="label", size=num_class)
loss = cross_entropy(name='loss', input=resnet, label=lbl)
inputs(img, lbl)
outputs(loss)
if is_infer:
outputs(resnet)
else:
lbl = data_layer(name="label", size=num_class)
loss = cross_entropy(name='loss', input=resnet, label=lbl)
outputs(loss)
benchmark/paddle/image/run_mkl_infer.sh 0 → 100755
浏览文件 @ fab63cc6
set -e
function clock_to_seconds() {
hours=`echo $1 | awk -F ':' '{print $1}'`
mins=`echo $1 | awk -F ':' '{print $2}'`
secs=`echo $1 | awk -F ':' '{print $3}'`
echo `awk 'BEGIN{printf "%.2f",('$secs' + '$mins' * 60 + '$hours' * 3600)}'`
}
function infer() {
unset OMP_NUM_THREADS MKL_NUM_THREADS OMP_DYNAMIC KMP_AFFINITY
topology=$1
layer_num=$2
bs=$3
use_mkldnn=$4
if [ $4 == "True" ]; then
thread=1
log="logs/infer-${topology}-${layer_num}-mkldnn-${bs}.log"
elif [ $4 == "False" ]; then
thread=`nproc`
if [ $thread -gt $bs ]; then
thread=$bs
fi
log="logs/infer-${topology}-${layer_num}-${thread}mklml-${bs}.log"
else
echo "Wrong input $4, use True or False."
exit 0
fi
models_in="models/${topology}-${layer_num}/pass-00000/"
if [ ! -d $models_in ]; then
echo "Training model ${topology}_${layer_num}"
paddle train --job=train \
--config="${topology}.py" \
--use_mkldnn=True \
--use_gpu=False \
--trainer_count=1 \
--num_passes=1 \
--save_dir="models/${topology}-${layer_num}" \
--config_args="batch_size=128,layer_num=${layer_num},num_samples=256" \
> /dev/null 2>&1
echo "Done"
fi
log_period=$((256 / bs))
paddle train --job=test \
--config="${topology}.py" \
--use_mkldnn=$use_mkldnn \
--use_gpu=False \
--trainer_count=$thread \
--log_period=$log_period \
--config_args="batch_size=${bs},layer_num=${layer_num},is_infer=True" \
--init_model_path=$models_in \
2>&1 | tee ${log}
# calculate the last 5 logs period time of 1280 samples,
# the time before are burning time.
start=`tail ${log} -n 7 | head -n 1 | awk -F ' ' '{print $2}' | xargs`
end=`tail ${log} -n 2 | head -n 1 | awk -F ' ' '{print $2}' | xargs`
start_sec=`clock_to_seconds $start`
end_sec=`clock_to_seconds $end`
fps=`awk 'BEGIN{printf "%.2f",(1280 / ('$end_sec' - '$start_sec'))}'`
echo "Last 1280 samples start: ${start}(${start_sec} sec), end: ${end}(${end_sec} sec;" >> ${log}
echo "FPS: $fps images/sec" 2>&1 | tee -a ${log}
}
if [ ! -f "train.list" ]; then
echo " " > train.list
fi
if [ ! -f "test.list" ]; then
echo " " > test.list
fi
if [ ! -d "logs" ]; then
mkdir logs
fi
if [ ! -d "models" ]; then
mkdir -p models
fi
# inference benchmark
for use_mkldnn in True False; do
for batchsize in 1 2 4 8 16; do
infer vgg 19 $batchsize $use_mkldnn
infer resnet 50 $batchsize $use_mkldnn
infer googlenet v1 $batchsize $use_mkldnn
infer alexnet 2 $batchsize $use_mkldnn
done
done
benchmark/paddle/image/run_mkldnn.sh → benchmark/paddle/image/run_mkl_train.sh
浏览文件 @ fab63cc6
......@@ -8,13 +8,13 @@ function train() {
use_mkldnn=$4
if [ $4 == "True" ]; then
thread=1
log="logs/${topology}-${layer_num}-mkldnn-${bs}.log"
log="logs/train-${topology}-${layer_num}-mkldnn-${bs}.log"
elif [ $4 == "False" ]; then
thread=`nproc`
# each trainer_count use only 1 core to avoid conflict
log="logs/${topology}-${layer_num}-${thread}mklml-${bs}.log"
log="logs/train-${topology}-${layer_num}-${thread}mklml-${bs}.log"
else
echo "Wrong input $3, use True or False."
echo "Wrong input $4, use True or False."
exit 0
fi
args="batch_size=${bs},layer_num=${layer_num}"
......@@ -28,18 +28,25 @@ function train() {
--test_period=100 \
--config_args=$args \
2>&1 | tee ${log}
avg_time=`tail ${log} -n 1 | awk -F ' ' '{print $8}' | sed 's/avg=//'`
fps=`awk 'BEGIN{printf "%.2f",('$bs' / '$avg_time' * 1000)}'`
echo "FPS: $fps images/sec" 2>&1 | tee -a ${log}
}
if [ ! -d "train.list" ]; then
if [ ! -f "train.list" ]; then
echo " " > train.list
fi
if [ ! -d "logs" ]; then
mkdir logs
fi
# training benchmark
for use_mkldnn in True False; do
for batchsize in 64 128 256; do
train vgg 19 $batchsize $use_mkldnn
train resnet 50 $batchsize $use_mkldnn
train resnet 50 $batchsize $use_mkldnn
train googlenet v1 $batchsize $use_mkldnn
train alexnet 2 $batchsize $use_mkldnn
done
done
benchmark/paddle/image/run_openblas_infer.sh 0 → 100755
浏览文件 @ fab63cc6
set -e
function clock_to_seconds() {
hours=`echo $1 | awk -F ':' '{print $1}'`
mins=`echo $1 | awk -F ':' '{print $2}'`
secs=`echo $1 | awk -F ':' '{print $3}'`
echo `awk 'BEGIN{printf "%.2f",('$secs' + '$mins' * 60 + '$hours' * 3600)}'`
}
function infer() {
topology=$1
layer_num=$2
bs=$3
trainers=`nproc`
if [ $trainers -gt $bs ]; then
trainers=$bs
fi
log="logs/infer-${topology}-${layer_num}-${trainers}openblas-${bs}.log"
threads=$((`nproc` / trainers))
if [ $threads -eq 0 ]; then
threads=1
fi
export OPENBLAS_NUM_THREADS=$threads
models_in="models/${topology}-${layer_num}/pass-00000/"
if [ ! -d $models_in ]; then
echo "./run_mkl_infer.sh to save the model first"
exit 0
fi
log_period=$((32 / bs))
paddle train --job=test \
--config="${topology}.py" \
--use_mkldnn=False \
--use_gpu=False \
--trainer_count=$trainers \
--log_period=$log_period \
--config_args="batch_size=${bs},layer_num=${layer_num},is_infer=True,num_samples=256" \
--init_model_path=$models_in \
2>&1 | tee ${log}
# calculate the last 5 logs period time of 160(=32*5) samples,
# the time before are burning time.
start=`tail ${log} -n 7 | head -n 1 | awk -F ' ' '{print $2}' | xargs`
end=`tail ${log} -n 2 | head -n 1 | awk -F ' ' '{print $2}' | xargs`
start_sec=`clock_to_seconds $start`
end_sec=`clock_to_seconds $end`
fps=`awk 'BEGIN{printf "%.2f",(160 / ('$end_sec' - '$start_sec'))}'`
echo "Last 160 samples start: ${start}(${start_sec} sec), end: ${end}(${end_sec} sec;" >> ${log}
echo "FPS: $fps images/sec" 2>&1 | tee -a ${log}
}
if [ ! -f "train.list" ]; then
echo " " > train.list
fi
if [ ! -f "test.list" ]; then
echo " " > test.list
fi
if [ ! -d "logs" ]; then
mkdir logs
fi
# inference benchmark
for batchsize in 1 2 4 8 16; do
infer vgg 19 $batchsize
infer resnet 50 $batchsize
infer googlenet v1 $batchsize
infer alexnet 2 $batchsize
done
benchmark/paddle/image/run_openblas_train.sh 0 → 100755
浏览文件 @ fab63cc6
set -e
function train() {
export OPENBLAS_NUM_THREADS=1
topology=$1
layer_num=$2
bs=$3
thread=`nproc`
# each trainer_count use only 1 core to avoid conflict
log="logs/train-${topology}-${layer_num}-${thread}openblas-${bs}.log"
args="batch_size=${bs},layer_num=${layer_num}"
config="${topology}.py"
paddle train --job=time \
--config=$config \
--use_mkldnn=False \
--use_gpu=False \
--trainer_count=$thread \
--log_period=3 \
--test_period=30 \
--config_args=$args \
2>&1 | tee ${log}
avg_time=`tail ${log} -n 1 | awk -F ' ' '{print $8}' | sed 's/avg=//'`
fps=`awk 'BEGIN{printf "%.2f",('$bs' / '$avg_time' * 1000)}'`
echo "FPS: $fps images/sec" 2>&1 | tee -a ${log}
}
if [ ! -f "train.list" ]; then
echo " " > train.list
fi
if [ ! -d "logs" ]; then
mkdir logs
fi
# training benchmark
for batchsize in 64 128 256; do
train vgg 19 $batchsize
train resnet 50 $batchsize
train googlenet v1 $batchsize
train alexnet 2 $batchsize
done
benchmark/paddle/image/vgg.py
浏览文件 @ fab63cc6
......@@ -6,10 +6,23 @@ width = 224
num_class = 1000
batch_size = get_config_arg('batch_size', int, 64)
layer_num = get_config_arg('layer_num', int, 19)
is_infer = get_config_arg("is_infer", bool, False)
num_samples = get_config_arg('num_samples', int, 2560)
args = {'height': height, 'width': width, 'color': True, 'num_class': num_class}
args = {
'height': height,
'width': width,
'color': True,
'num_class': num_class,
'is_infer': is_infer,
'num_samples': num_samples
}
define_py_data_sources2(
"train.list", None, module="provider", obj="process", args=args)
"train.list" if not is_infer else None,
"test.list" if is_infer else None,
module="provider",
obj="process",
args=args)
settings(
batch_size=batch_size,
......@@ -98,6 +111,9 @@ elif layer_num == 19:
else:
print("Wrong layer number.")
lab = data_layer('label', num_class)
loss = cross_entropy(input=vgg, label=lab)
outputs(loss)
if is_infer:
outputs(vgg)
else:
lab = data_layer('label', num_class)
loss = cross_entropy(input=vgg, label=lab)
outputs(loss)
cmake/cblas.cmake
浏览文件 @ fab63cc6
......@@ -3,7 +3,7 @@
# It will search MKLML, atlas, OpenBlas, reference-cblas in order.
#
# If any cblas implementation found, the following variable will be set.
# CBLAS_PROVIDER # one of MKLML, ATLAS, OPENBLAS, REFERENCE
# CBLAS_PROVIDER # one of MKLML, OPENBLAS, REFERENCE
# CBLAS_INC_DIR # the include directory for cblas.
# CBLAS_LIBS # a list of libraries should be linked by paddle.
# # Each library should be full path to object file.
......@@ -17,7 +17,7 @@ if(WITH_MKLML AND MKLML_INC_DIR AND MKLML_LIB)
set(CBLAS_INC_DIR ${MKLML_INC_DIR})
set(CBLAS_LIBRARIES ${MKLML_LIB})
add_definitions(-DPADDLE_USE_MKLML)
add_definitions(-DPADDLE_WITH_MKLML)
add_definitions(-DLAPACK_FOUND)
message(STATUS "Found cblas and lapack in MKLML "
......@@ -25,42 +25,6 @@ if(WITH_MKLML AND MKLML_INC_DIR AND MKLML_LIB)
return()
endif()
## Then find atlas.
set(ATLAS_ROOT $ENV{ATLAS_ROOT} CACHE PATH "Folder contains Atlas")
set(ATLAS_INCLUDE_SEARCH_PATHS
${ATLAS_ROOT}/include
/usr/include
/usr/include/atlas)
set(ATLAS_LIB_SEARCH_PATHS
${ATLAS_ROOT}/lib
/usr/lib
/usr/lib/blas/atlas
/usr/lib/atlas
/usr/lib/atlas-base # special for ubuntu 14.04.
)
find_path(ATLAS_INC_DIR NAMES cblas.h
PATHS ${ATLAS_INCLUDE_SEARCH_PATHS})
find_path(ATLAS_CLAPACK_INC_DIR NAMES clapack.h
PATHS ${ATLAS_INCLUDE_SEARCH_PATHS})
find_library(ATLAS_CBLAS_LIB NAMES cblas libcblas.so.3
PATHS ${ATLAS_LIB_SEARCH_PATHS})
find_library(ATLAS_CLAPACK_LIB NAMES lapack_atlas liblapack_atlas.so.3
PATHS ${ATLAS_LIB_SEARCH_PATHS})
if(ATLAS_CLAPACK_INC_DIR AND ATLAS_INC_DIR AND ATLAS_CBLAS_LIB AND ATLAS_CLAPACK_LIB)
set(CBLAS_FOUND ON)
set(CBLAS_PROVIDER ATLAS)
set(CBLAS_INC_DIR ${ATLAS_INC_DIR} ${ATLAS_CLAPACK_INC_DIR})
set(CBLAS_LIBRARIES ${ATLAS_CLAPACK_LIB} ${ATLAS_CBLAS_LIB})
add_definitions(-DPADDLE_USE_ATLAS)
add_definitions(-DLAPACK_FOUND)
message(STATUS "Found ATLAS (include: ${ATLAS_INC_DIR}, library: ${CBLAS_LIBRARIES})")
message(STATUS "Found lapack in ATLAS (include: ${ATLAS_CLAPACK_INC_DIR})")
return()
endif()
## Then find openblas.
set(OPENBLAS_ROOT $ENV{OPENBLAS_ROOT} CACHE PATH "Folder contains Openblas")
set(OPENBLAS_INCLUDE_SEARCH_PATHS
......
cmake/configure.cmake
浏览文件 @ fab63cc6
......@@ -24,6 +24,11 @@ if(WITH_DOUBLE)
add_definitions(-DPADDLE_TYPE_DOUBLE)
endif(WITH_DOUBLE)
if(WITH_ARM_FP16)
add_definitions(-DPADDLE_ARM_FP16)
add_definitions("-march=armv8.2-a+fp16+simd")
endif(WITH_ARM_FP16)
if(WITH_TESTING)
add_definitions(-DPADDLE_WITH_TESTING)
endif(WITH_TESTING)
......
cmake/external/cares.cmake 0 → 100644
浏览文件 @ fab63cc6
# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
IF(MOBILE_INFERENCE OR NOT WITH_DISTRIBUTE)
return()
ENDIF()
include (ExternalProject)
# NOTE: c-ares is needed when linking with grpc.
SET(CARES_SOURCES_DIR ${THIRD_PARTY_PATH}/cares)
SET(CARES_INSTALL_DIR ${THIRD_PARTY_PATH}/install/cares)
SET(CARES_INCLUDE_DIR "${CARES_INSTALL_DIR}/include/" CACHE PATH "cares include directory." FORCE)
ExternalProject_Add(
extern_cares
GIT_REPOSITORY "https://github.com/c-ares/c-ares.git"
GIT_TAG "cares-1_13_0"
PREFIX ${CARES_SOURCES_DIR}
UPDATE_COMMAND ""
CONFIGURE_COMMAND ./buildconf && ./configure --disable-shared --prefix=${CARES_INSTALL_DIR}
BUILD_IN_SOURCE 1
BUILD_COMMAND make -j8
INSTALL_COMMAND make install
)
ADD_LIBRARY(cares STATIC IMPORTED GLOBAL)
SET_PROPERTY(TARGET cares PROPERTY IMPORTED_LOCATION
"${CARES_INSTALL_DIR}/lib/libcares.a")
include_directories(${CARES_INCLUDE_DIR})
ADD_DEPENDENCIES(cares extern_cares)
cmake/external/gflags.cmake
浏览文件 @ fab63cc6
......@@ -28,15 +28,8 @@ INCLUDE_DIRECTORIES(${GFLAGS_INCLUDE_DIR})
ExternalProject_Add(
extern_gflags
${EXTERNAL_PROJECT_LOG_ARGS}
# TODO(yiwang): The annoying warnings mentioned in
# https://github.com/PaddlePaddle/Paddle/issues/3277 are caused by
# gflags. I fired a PR https://github.com/gflags/gflags/pull/230
# to fix it. Before it gets accepted by the gflags team, we use
# my personal fork, which contains above fix, temporarily. Let's
# change this back to the official Github repo once my PR is
# merged.
GIT_REPOSITORY "https://github.com/wangkuiyi/gflags.git"
GIT_TAG 986964c07427ecb9cdb5bd73f73ebbd40e54dadb
GIT_REPOSITORY "https://github.com/gflags/gflags.git"
GIT_TAG 77592648e3f3be87d6c7123eb81cbad75f9aef5a
PREFIX ${GFLAGS_SOURCES_DIR}
UPDATE_COMMAND ""
CMAKE_ARGS -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
......
cmake/external/glog.cmake
浏览文件 @ fab63cc6
......@@ -26,12 +26,21 @@ ENDIF(WIN32)
INCLUDE_DIRECTORIES(${GLOG_INCLUDE_DIR})
IF(ANDROID AND ${CMAKE_SYSTEM_VERSION} VERSION_LESS "21")
# Using the unofficial glog for Android API < 21
SET(GLOG_REPOSITORY "https://github.com/Xreki/glog.git")
SET(GLOG_TAG "8a547150548b284382ccb6582408e9140ff2bea8")
ELSE()
SET(GLOG_REPOSITORY "https://github.com/google/glog.git")
SET(GLOG_TAG "v0.3.5")
ENDIF()
ExternalProject_Add(
extern_glog
${EXTERNAL_PROJECT_LOG_ARGS}
DEPENDS gflags
GIT_REPOSITORY "https://github.com/google/glog.git"
GIT_TAG v0.3.5
GIT_REPOSITORY ${GLOG_REPOSITORY}
GIT_TAG ${GLOG_TAG}
PREFIX ${GLOG_SOURCES_DIR}
UPDATE_COMMAND ""
CMAKE_ARGS -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
......
cmake/external/grpc.cmake 0 → 100644
浏览文件 @ fab63cc6
# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
IF(MOBILE_INFERENCE OR NOT WITH_DISTRIBUTE)
return()
ENDIF()
include (ExternalProject)
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)
IF(APPLE)
SET(BUILD_CMD make -n HAS_SYSTEM_PROTOBUF=false -s -j8 static grpc_cpp_plugin | sed "s/-Werror//g" | sh)
ELSE()
SET(BUILD_CMD make HAS_SYSTEM_PROTOBUF=false -s -j8 static grpc_cpp_plugin)
ENDIF()
ExternalProject_Add(
extern_grpc
DEPENDS protobuf zlib
GIT_REPOSITORY "https://github.com/grpc/grpc.git"
GIT_TAG "v1.7.x"
PREFIX ${GRPC_SOURCES_DIR}
UPDATE_COMMAND ""
CONFIGURE_COMMAND ""
BUILD_IN_SOURCE 1
# NOTE(yuyang18):
# Disable -Werror, otherwise the compile will fail in MacOS.
# It seems that we cannot configure that by make command.
# Just dry run make command and remove `-Werror`, then use a shell to run make commands
BUILD_COMMAND ${BUILD_CMD}
INSTALL_COMMAND make prefix=${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
"${GRPC_INSTALL_DIR}/lib/libgrpc++_unsecure.a")
ADD_LIBRARY(grpc++ STATIC IMPORTED GLOBAL)
SET_PROPERTY(TARGET grpc++ PROPERTY IMPORTED_LOCATION
"${GRPC_INSTALL_DIR}/lib/libgrpc++.a")
ADD_LIBRARY(gpr STATIC IMPORTED GLOBAL)
SET_PROPERTY(TARGET gpr PROPERTY IMPORTED_LOCATION
"${GRPC_INSTALL_DIR}/lib/libgpr.a")
ADD_LIBRARY(grpc_unsecure STATIC IMPORTED GLOBAL)
SET_PROPERTY(TARGET grpc_unsecure PROPERTY IMPORTED_LOCATION
"${GRPC_INSTALL_DIR}/lib/libgrpc_unsecure.a")
include_directories(${GRPC_INCLUDE_DIR})
ADD_DEPENDENCIES(grpc++_unsecure extern_grpc)
cmake/external/mkldnn.cmake
浏览文件 @ fab63cc6
......@@ -67,5 +67,5 @@ ADD_LIBRARY(mkldnn SHARED IMPORTED GLOBAL)
SET_PROPERTY(TARGET mkldnn PROPERTY IMPORTED_LOCATION ${MKLDNN_LIB})
ADD_DEPENDENCIES(mkldnn ${MKLDNN_PROJECT})
MESSAGE(STATUS "MKLDNN library: ${MKLDNN_LIB}")
add_definitions(-DPADDLE_USE_MKLDNN)
add_definitions(-DPADDLE_WITH_MKLDNN)
LIST(APPEND external_project_dependencies mkldnn)
cmake/external/openblas.cmake
浏览文件 @ fab63cc6
......@@ -114,11 +114,7 @@ INCLUDE_DIRECTORIES(${CBLAS_INC_DIR})
# linear algebra libraries for cc_library(xxx SRCS xxx.c DEPS cblas)
SET(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/cblas_dummy.c)
FILE(WRITE ${dummyfile} "const char * dummy = \"${dummyfile}\";")
IF("${CBLAS_PROVIDER}" STREQUAL "MKLML")
ADD_LIBRARY(cblas SHARED ${dummyfile})
ELSE()
ADD_LIBRARY(cblas STATIC ${dummyfile})
ENDIF()
ADD_LIBRARY(cblas STATIC ${dummyfile})
TARGET_LINK_LIBRARIES(cblas ${CBLAS_LIBRARIES})
IF(NOT ${CBLAS_FOUND})
......
cmake/external/protobuf.cmake
浏览文件 @ fab63cc6
......@@ -15,7 +15,18 @@
INCLUDE(ExternalProject)
# Always invoke `FIND_PACKAGE(Protobuf)` for importing function protobuf_generate_cpp
FIND_PACKAGE(Protobuf QUIET)
SET(PROTOBUF_FOUND "OFF")
macro(UNSET_VAR VAR_NAME)
UNSET(${VAR_NAME} CACHE)
UNSET(${VAR_NAME})
endmacro()
UNSET_VAR(PROTOBUF_INCLUDE_DIR)
UNSET_VAR(PROTOBUF_FOUND)
UNSET_VAR(PROTOBUF_PROTOC_EXECUTABLE)
UNSET_VAR(PROTOBUF_PROTOC_LIBRARY)
UNSET_VAR(PROTOBUF_LITE_LIBRARY)
UNSET_VAR(PROTOBUF_LIBRARY)
UNSET_VAR(PROTOBUF_INCLUDE_DIR)
UNSET_VAR(Protobuf_PROTOC_EXECUTABLE)
if(NOT COMMAND protobuf_generate_python) # before cmake 3.4, protobuf_genrerate_python is not defined.
function(protobuf_generate_python SRCS)
......@@ -110,7 +121,6 @@ macro(PROMPT_PROTOBUF_LIB)
# FIND_Protobuf.cmake uses `Protobuf_PROTOC_EXECUTABLE`.
# make `protobuf_generate_cpp` happy.
SET(Protobuf_PROTOC_EXECUTABLE ${PROTOBUF_PROTOC_EXECUTABLE})
FOREACH(dep ${protobuf_DEPS})
ADD_DEPENDENCIES(protobuf ${dep})
ADD_DEPENDENCIES(protobuf_lite ${dep})
......@@ -128,11 +138,11 @@ endmacro()
set(PROTOBUF_ROOT "" CACHE PATH "Folder contains protobuf")
if (NOT "${PROTOBUF_ROOT}" STREQUAL "")
find_path(PROTOBUF_INCLUDE_DIR google/protobuf/message.h PATHS ${PROTOBUF_ROOT}/include)
find_library(PROTOBUF_LIBRARY protobuf PATHS ${PROTOBUF_ROOT}/lib)
find_library(PROTOBUF_LITE_LIBRARY protobuf-lite PATHS ${PROTOBUF_ROOT}/lib)
find_library(PROTOBUF_PROTOC_LIBRARY protoc PATHS ${PROTOBUF_ROOT}/lib)
find_program(PROTOBUF_PROTOC_EXECUTABLE protoc PATHS ${PROTOBUF_ROOT}/bin)
find_path(PROTOBUF_INCLUDE_DIR google/protobuf/message.h PATHS ${PROTOBUF_ROOT}/include NO_DEFAULT_PATH)
find_library(PROTOBUF_LIBRARY protobuf PATHS ${PROTOBUF_ROOT}/lib NO_DEFAULT_PATH)
find_library(PROTOBUF_LITE_LIBRARY protobuf-lite PATHS ${PROTOBUF_ROOT}/lib NO_DEFAULT_PATH)
find_library(PROTOBUF_PROTOC_LIBRARY protoc PATHS ${PROTOBUF_ROOT}/lib NO_DEFAULT_PATH)
find_program(PROTOBUF_PROTOC_EXECUTABLE protoc PATHS ${PROTOBUF_ROOT}/bin NO_DEFAULT_PATH)
if (PROTOBUF_INCLUDE_DIR AND PROTOBUF_LIBRARY AND PROTOBUF_LITE_LIBRARY AND PROTOBUF_PROTOC_LIBRARY AND PROTOBUF_PROTOC_EXECUTABLE)
message(STATUS "Using custom protobuf library in ${PROTOBUF_ROOT}.")
SET_PROTOBUF_VERSION()
......@@ -178,14 +188,26 @@ FUNCTION(build_protobuf TARGET_NAME BUILD_FOR_HOST)
SET(OPTIONAL_CACHE_ARGS "-DZLIB_ROOT:STRING=${ZLIB_ROOT}")
ENDIF()
SET(PROTOBUF_REPO "https://github.com/google/protobuf.git")
SET(PROTOBUF_TAG "9f75c5aa851cd877fb0d93ccc31b8567a6706546")
IF(MOBILE_INFERENCE)
# The reason why the official version is not used is described in
# https://github.com/PaddlePaddle/Paddle/issues/6114
SET(PROTOBUF_REPO "https://github.com/qingqing01/protobuf.git")
SET(PROTOBUF_TAG "v3.2.0")
IF(NOT BUILD_FOR_HOST)
SET(OPTIONAL_ARGS ${OPTIONAL_ARGS} "-Dprotobuf_BUILD_PROTOC_BINARIES=OFF")
ENDIF()
ENDIF()
ExternalProject_Add(
${TARGET_NAME}
${EXTERNAL_PROJECT_LOG_ARGS}
PREFIX ${PROTOBUF_SOURCES_DIR}
UPDATE_COMMAND ""
DEPENDS zlib
GIT_REPOSITORY "https://github.com/google/protobuf.git"
GIT_TAG "9f75c5aa851cd877fb0d93ccc31b8567a6706546"
GIT_REPOSITORY ${PROTOBUF_REPO}
GIT_TAG ${PROTOBUF_TAG}
CONFIGURE_COMMAND
${CMAKE_COMMAND} ${PROTOBUF_SOURCES_DIR}/src/${TARGET_NAME}/cmake
${OPTIONAL_ARGS}
......@@ -203,7 +225,11 @@ FUNCTION(build_protobuf TARGET_NAME BUILD_FOR_HOST)
)
ENDFUNCTION()
SET(PROTOBUF_VERSION 3.1)
IF(NOT MOBILE_INFERENCE)
SET(PROTOBUF_VERSION 3.1)
ELSE()
SET(PROTOBUF_VERSION 3.2)
ENDIF()
IF(CMAKE_CROSSCOMPILING)
build_protobuf(protobuf_host TRUE)
LIST(APPEND external_project_dependencies protobuf_host)
......@@ -227,9 +253,9 @@ IF(NOT PROTOBUF_FOUND)
IF(WITH_C_API)
INSTALL(DIRECTORY ${PROTOBUF_INCLUDE_DIR} DESTINATION third_party/protobuf)
IF(ANDROID)
INSTALL(FILES ${PROTOBUF_LIBRARY} DESTINATION third_party/protobuf/lib/${ANDROID_ABI})
INSTALL(FILES ${PROTOBUF_LITE_LIBRARY} DESTINATION third_party/protobuf/lib/${ANDROID_ABI})
ELSE()
INSTALL(FILES ${PROTOBUF_LIBRARY} DESTINATION third_party/protobuf/lib)
INSTALL(FILES ${PROTOBUF_LITE_LIBRARY} DESTINATION third_party/protobuf/lib)
ENDIF()
ENDIF()
......
cmake/external/zlib.cmake
浏览文件 @ fab63cc6
......@@ -50,6 +50,8 @@ ExternalProject_Add(
)
LIST(APPEND external_project_dependencies zlib)
ADD_LIBRARY(zlib_target STATIC IMPORTED GLOBAL)
SET_PROPERTY(TARGET zlib_target PROPERTY IMPORTED_LOCATION ${ZLIB_LIBRARIES})
IF(WITH_C_API)
INSTALL(DIRECTORY ${ZLIB_INCLUDE_DIR} DESTINATION third_party/zlib)
......
cmake/flags.cmake
浏览文件 @ fab63cc6
......@@ -111,6 +111,8 @@ set(COMMON_FLAGS
-Wno-error=sign-compare
-Wno-error=unused-local-typedefs
-Wno-error=parentheses-equality # Warnings in pybind11
-Wno-error=ignored-attributes # Warnings in Eigen, gcc 6.3
-Wno-error=terminate # Warning in PADDLE_ENFORCE
)
set(GPU_COMMON_FLAGS
......
cmake/generic.cmake
浏览文件 @ fab63cc6
......@@ -227,8 +227,8 @@ function(cc_test TARGET_NAME)
set(multiValueArgs SRCS DEPS)
cmake_parse_arguments(cc_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
add_executable(${TARGET_NAME} ${cc_test_SRCS})
target_link_libraries(${TARGET_NAME} ${cc_test_DEPS} gtest gtest_main)
add_dependencies(${TARGET_NAME} ${cc_test_DEPS} gtest gtest_main)
target_link_libraries(${TARGET_NAME} ${cc_test_DEPS} paddle_gtest_main paddle_memory gtest gflags)
add_dependencies(${TARGET_NAME} ${cc_test_DEPS} paddle_gtest_main paddle_memory gtest gflags)
add_test(NAME ${TARGET_NAME} COMMAND ${TARGET_NAME} WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR})
endif()
endfunction(cc_test)
......@@ -288,8 +288,8 @@ function(nv_test TARGET_NAME)
set(multiValueArgs SRCS DEPS)
cmake_parse_arguments(nv_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
cuda_add_executable(${TARGET_NAME} ${nv_test_SRCS})
target_link_libraries(${TARGET_NAME} ${nv_test_DEPS} gtest gtest_main)
add_dependencies(${TARGET_NAME} ${nv_test_DEPS} gtest gtest_main)
target_link_libraries(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main paddle_memory gtest gflags)
add_dependencies(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main paddle_memory gtest gflags)
add_test(${TARGET_NAME} ${TARGET_NAME})
endif()
endfunction(nv_test)
......@@ -459,11 +459,58 @@ function(py_test TARGET_NAME)
if(WITH_TESTING)
set(options STATIC static SHARED shared)
set(oneValueArgs "")
set(multiValueArgs SRCS DEPS)
cmake_parse_arguments(py_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
set(multiValueArgs SRCS DEPS ARGS)
cmake_parse_arguments(py_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
add_test(NAME ${TARGET_NAME}
COMMAND env PYTHONPATH=${PADDLE_PYTHON_BUILD_DIR}/lib-python
python2 ${py_test_SRCS}
${PYTHON_EXECUTABLE} -u ${py_test_SRCS} ${py_test_ARGS}
WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR})
endif()
endfunction()
# grpc_library generate grpc code using grpc_cpp_plugin and protoc
# then build the generated protobuf code and grpc code with your
# implementation source codes together. Use SRCS argument for your
# implementation source files and PROTO argument for your .proto
# files.
#
# Usage: grpc_library(my_target SRCS my_client.cc PROTO my_target.proto DEPS my_dep)
function(grpc_library TARGET_NAME)
set(oneValueArgs PROTO)
set(multiValueArgs SRCS DEPS)
set(options "")
cmake_parse_arguments(grpc_library "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
message(STATUS "generating grpc ${grpc_library_PROTO}")
get_filename_component(ABS_PROTO ${grpc_library_PROTO} ABSOLUTE)
get_filename_component(PROTO_WE ${grpc_library_PROTO} NAME_WE)
get_filename_component(PROTO_PATH ${ABS_PROTO} PATH)
protobuf_generate_cpp(grpc_proto_srcs grpc_proto_hdrs "${ABS_PROTO}")
set(grpc_grpc_srcs "${CMAKE_CURRENT_BINARY_DIR}/${PROTO_WE}.grpc.pb.cc")
set(grpc_grpc_hdrs "${CMAKE_CURRENT_BINARY_DIR}/${PROTO_WE}.grpc.pb.h")
cc_library("${TARGET_NAME}_proto" SRCS "${grpc_proto_srcs}")
add_custom_command(
OUTPUT "${grpc_grpc_srcs}" "${grpc_grpc_hdrs}"
COMMAND ${PROTOBUF_PROTOC_EXECUTABLE}
ARGS --grpc_out "${CMAKE_CURRENT_BINARY_DIR}" -I "${PROTO_PATH}"
--plugin=protoc-gen-grpc="${GRPC_CPP_PLUGIN}" "${ABS_PROTO}"
DEPENDS "${ABS_PROTO}" ${PROTOBUF_PROTOC_EXECUTABLE} extern_grpc)
# FIXME(typhoonzero): grpc generated code do not generate virtual-dtor, mark it
# as compiler warnings instead of error. Should try remove the warnings also.
set_source_files_properties(
${grpc_grpc_srcs}
PROPERTIES
COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
cc_library("${TARGET_NAME}_grpc" SRCS "${grpc_grpc_srcs}")
set_source_files_properties(
${grpc_library_SRCS}
PROPERTIES
COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
cc_library("${TARGET_NAME}" SRCS "${grpc_library_SRCS}" DEPS "${TARGET_NAME}_grpc" "${TARGET_NAME}_proto" "${grpc_library_DEPS}")
endfunction()
cmake/util.cmake
浏览文件 @ fab63cc6
......@@ -168,17 +168,3 @@ function(create_resources res_file output_file)
COMMAND python ARGS ${PADDLE_SOURCE_DIR}/cmake/make_resource.py ${res_file} ${output_file}
DEPENDS ${res_file} ${PADDLE_SOURCE_DIR}/cmake/make_resource.py)
endfunction()
# Create a python unittest using run_python_tests.sh,
# which takes care of making correct running environment
function(add_python_test TEST_NAME)
foreach(arg ${ARGN})
get_filename_component(py_fn ${arg} NAME_WE)
set(TRG_NAME ${TEST_NAME}_${py_fn})
add_test(NAME ${TRG_NAME}
COMMAND env PYTHONPATH=${PADDLE_PYTHON_PACKAGE_DIR}
python2 ${arg}
WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR})
endforeach()
endfunction()
doc/api/index_cn.rst
浏览文件 @ fab63cc6
......@@ -7,3 +7,4 @@ API
模型配置 <v2/model_configs.rst>
数据访问 <v2/data.rst>
训练与应用 <v2/run_logic.rst>
v2/fluid.rst
doc/api/index_en.rst
浏览文件 @ fab63cc6
......@@ -7,3 +7,4 @@ API
v2/model_configs.rst
v2/data.rst
v2/run_logic.rst
v2/fluid.rst
doc/api/v2/config/activation.rst
浏览文件 @ fab63cc6
......@@ -99,3 +99,10 @@ STanh
.. automodule:: paddle.v2.activation
:members: STanh
:noindex:
SoftSign
========
.. automodule:: paddle.v2.activation
:members: SoftSign
:noindex:
doc/api/v2/config/layer.rst
浏览文件 @ fab63cc6
......@@ -54,7 +54,7 @@ img_conv
.. _api_v2.layer_context_projection:
context_projection
context_projection
------------------
.. autoclass:: paddle.v2.layer.context_projection
:noindex:
......@@ -70,7 +70,7 @@ Image Pooling Layer
img_pool
--------
.. autoclass:: paddle.v2.layer.img_pool
:noindex:
:noindex:
spp
---
......@@ -104,7 +104,7 @@ sum_to_one_norm
---------------
.. autoclass:: paddle.v2.layer.sum_to_one_norm
:noindex:
cross_channel_norm
------------------
.. autoclass:: paddle.v2.layer.cross_channel_norm
......@@ -114,7 +114,7 @@ row_l2_norm
-----------
.. autoclass:: paddle.v2.layer.row_l2_norm
:noindex:
Recurrent Layers
================
......@@ -252,6 +252,11 @@ first_seq
.. autoclass:: paddle.v2.layer.first_seq
:noindex:
sub_seq
---------
.. autoclass:: paddle.v2.layer.sub_seq
:noindex:
concat
------
.. autoclass:: paddle.v2.layer.concat
......@@ -415,6 +420,13 @@ multiplex
.. autoclass:: paddle.v2.layer.multiplex
:noindex:
Factorization Machine Layer
============================
factorization_machine
---------------------
.. autoclass:: paddle.v2.layer.factorization_machine
:noindex:
Slicing and Joining Layers
==========================
......@@ -460,7 +472,7 @@ lambda_cost
:noindex:
square_error_cost
--------
-----------------
.. autoclass:: paddle.v2.layer.square_error_cost
:noindex:
......@@ -526,7 +538,7 @@ Miscs
=====
dropout
--------------
--------
.. autoclass:: paddle.v2.layer.dropout
:noindex:
......
doc/api/v2/fluid.rst 0 → 100644
浏览文件 @ fab63cc6
======================
Fluid
======================
.. toctree::
:maxdepth: 1
fluid/layers.rst
fluid/data_feeder.rst
fluid/executor.rst
fluid/initializer.rst
fluid/evaluator.rst
fluid/nets.rst
fluid/optimizer.rst
fluid/param_attr.rst
fluid/profiler.rst
fluid/regularizer.rst
doc/api/v2/fluid/data_feeder.rst 0 → 100644
浏览文件 @ fab63cc6
===========
DataFeeder
===========
DataFeeder
-----------
.. automodule:: paddle.v2.fluid.data_feeder
:members: DataFeeder
:noindex:
doc/api/v2/fluid/evaluator.rst 0 → 100644
浏览文件 @ fab63cc6
===========
Evaluator
===========
Evaluator
-----------
.. automodule:: paddle.v2.fluid.evaluator
:members: Evaluator
:noindex:
doc/api/v2/fluid/executor.rst 0 → 100644
浏览文件 @ fab63cc6
===========
Executor
===========
Executor
-----------
.. automodule:: paddle.v2.fluid.executor
:members: Executor
:noindex:
doc/api/v2/fluid/initializer.rst 0 → 100644
浏览文件 @ fab63cc6
===========
Initializer
===========
Initializer
-----------
.. automodule:: paddle.v2.fluid.initializer
:members: Initializer
:noindex:
ConstantInitializer
-------------------
.. automodule:: paddle.v2.fluid.initializer
:members: ConstantInitializer
:noindex:
UniformInitializer
------------------
.. automodule:: paddle.v2.fluid.initializer
:members: UniformInitializer
:noindex:
NormalInitializer
-----------------
.. automodule:: paddle.v2.fluid.initializer
:members: NormalInitializer
:noindex:
XavierInitializer
-----------------
.. automodule:: paddle.v2.fluid.initializer
:members: XavierInitializer
:noindex:
MSRAInitializer
---------------
.. automodule:: paddle.v2.fluid.initializer
:members: MSRAInitializer
:noindex:
doc/api/v2/fluid/layers.rst 0 → 100644
浏览文件 @ fab63cc6
==========
Layers
==========
fc
---
.. autofunction:: paddle.v2.fluid.layers.fc
:noindex:
embedding
---------
.. autofunction:: paddle.v2.fluid.layers.embedding
:noindex:
dynamic_lstm
------------
.. autofunction:: paddle.v2.fluid.layers.dynamic_lstm
:noindex:
data
----
.. autofunction:: paddle.v2.fluid.layers.data
:noindex:
mean
----
.. autofunction:: paddle.v2.fluid.layers.mean
:noindex:
mul
---
.. autofunction:: paddle.v2.fluid.layers.mul
:noindex:
elementwise_add
---------------
.. autofunction:: paddle.v2.fluid.layers.elementwise_add
:noindex:
elementwise_div
---------------
.. autofunction:: paddle.v2.fluid.layers.elementwise_div
:noindex:
dropout
-------
.. autofunction:: paddle.v2.fluid.layers.dropout
:noindex:
reshape
--------
.. autofunction:: paddle.v2.fluid.layers.reshape
:noindex:
sigmoid
---------
.. autofunction:: paddle.v2.fluid.layers.sigmoid
:noindex:
scale
---------
.. autofunction:: paddle.v2.fluid.layers.scale
:noindex:
transpose
---------
.. autofunction:: paddle.v2.fluid.layers.transpose
:noindex:
sigmoid_cross_entropy_with_logits
---------------------------------
.. autofunction:: paddle.v2.fluid.layers.esigmoid_cross_entropy_with_logits
:noindex:
cast
----
.. autofunction:: paddle.v2.fluid.layers.cast
:noindex:
concat
-------
.. autofunction:: paddle.v2.fluid.layers.concat
:noindex:
sums
----
.. autofunction:: paddle.v2.fluid.layers.sums
:noindex:
linear_chain_crf
----------------
.. autofunction:: paddle.v2.fluid.layers.linear_chain_crf
:noindex:
assign
-------
.. autofunction:: paddle.v2.fluid.layers.embedding
:noindex:
split_lod_tensor
----------------
.. autofunction:: paddle.v2.fluid.layers.split_lod_tensor
:noindex:
merge_lod_tensor
----------------
.. autofunction:: paddle.v2.fluid.layers.merge_lod_tensor
:noindex:
cos_sim
--------
.. autofunction:: paddle.v2.fluid.layers.cos_sim
:noindex:
cross_entropy
-------------
.. autofunction:: paddle.v2.fluid.layers.cross_entropy
:noindex:
square_error_cost
-----------------
.. autofunction:: paddle.v2.fluid.layers.square_error_cost
:noindex:
accuracy
---------
.. autofunction:: paddle.v2.fluid.layers.accuracy
:noindex:
sequence_conv
-------------
.. autofunction:: paddle.v2.fluid.layers.sequence_conv
:noindex:
conv2d
------
.. autofunction:: paddle.v2.fluid.layers.conv2d
:noindex:
sequence_pool
-------------
.. autofunction:: paddle.v2.fluid.layers.sequence_pool
:noindex:
sequence_first_step
-------------------
.. autofunction:: paddle.v2.fluid.layers.sequence_first_step
:noindex:
sequence_last_step
------------------
.. autofunction:: paddle.v2.fluid.layers.sequence_last_step
:noindex:
pool2d
------
.. autofunction:: paddle.v2.fluid.layers.pool2d
:noindex:
batch_norm
----------
.. autofunction:: paddle.v2.fluid.layers.batch_norm
:noindex:
beam_search_decode
------------------
.. autofunction:: paddle.v2.fluid.layers.beam_search_decode
:noindex:
lod_rank_table
--------------
.. autofunction:: paddle.v2.fluid.layers.lod_rank_table
:noindex:
max_sequence_len
----------------
.. autofunction:: paddle.v2.fluid.layers.max_sequence_len
:noindex:
topk
-----
.. autofunction:: paddle.v2.fluid.layers.topk
:noindex:
lod_tensor_to_array
-------------------
.. autofunction:: paddle.v2.fluid.layers.lod_tensor_to_array
:noindex:
array_to_lod_tensor
-------------------
.. autofunction:: paddle.v2.fluid.layers.array_to_lod_tensor
:noindex:
fill_constant
-------------
.. autofunction:: paddle.v2.fluid.layers.fill_constant
:noindex:
fill_constant_batch_size_like
-----------------------------
.. autofunction:: paddle.v2.fluid.layers.fill_constant_batch_size_like
:noindex:
ones
----
.. autofunction:: paddle.v2.fluid.layers.ones
:noindex:
zeros
-----
.. autofunction:: paddle.v2.fluid.layers.zeros
:noindex:
increment
---------
.. autofunction:: paddle.v2.fluid.layers.increment
:noindex:
array_write
-----------
.. autofunction:: paddle.v2.fluid.layers.array_write
:noindex:
create_array
------------
.. autofunction:: paddle.v2.fluid.layers.create_array
:noindex:
less_than
---------
.. autofunction:: paddle.v2.fluid.layers.less_than
:noindex:
array_read
----------
.. autofunction:: paddle.v2.fluid.layers.array_read
:noindex:
shrink_memory
--------------
.. autofunction:: paddle.v2.fluid.layers.shrink_memory
:noindex:
array_length
-------------
.. autofunction:: paddle.v2.fluid.layers.array_length
:noindex:
conv2d_transpose
----------------
.. autofunction:: paddle.v2.fluid.layers.conv2d_transpose
:noindex:
sequence_expand
---------------
.. autofunction:: paddle.v2.fluid.layers.sequence_expand
:noindex:
lstm_unit
---------
.. autofunction:: paddle.v2.fluid.layers.lstm_unit
:noindex:
sequence_softmax
----------------
.. autofunction:: paddle.v2.fluid.layers.sequence_softmax
:noindex:
reduce_sum
----------
.. autofunction:: paddle.v2.fluid.layers.reduce_sum
:noindex:
reduce_mean
-----------
.. autofunction:: paddle.v2.fluid.layers.reduce_mean
:noindex:
reduce_max
----------
.. autofunction:: paddle.v2.fluid.layers.reduce_max
:noindex:
reduce_min
----------
.. autofunction:: paddle.v2.fluid.layers.reduce_min
:noindex:
doc/api/v2/fluid/nets.rst 0 → 100644
浏览文件 @ fab63cc6
===========
Nets
===========
simple_img_conv_pool
--------------------
.. autofunction:: paddle.v2.fluid.nets.simple_img_conv_pool
:noindex:
img_conv_group
---------------
.. autofunction:: paddle.v2.fluid.nets.img_conv_group
:noindex:
sequence_conv_pool
------------------
.. autofunction:: paddle.v2.fluid.nets.sequence_conv_pool
:noindex:
doc/api/v2/fluid/optimizer.rst 0 → 100644
浏览文件 @ fab63cc6
===========
Optimizer
===========
Optimizer
-----------
.. automodule:: paddle.v2.fluid.optimizer
:members: Optimizer
:noindex:
SGDOptimizer
-----------
.. automodule:: paddle.v2.fluid.optimizer
:members: SGDOptimizer
:noindex:
MomentumOptimizer
-----------------
.. automodule:: paddle.v2.fluid.optimizer
:members: MomentumOptimizer
:noindex:
AdagradOptimizer
----------------
.. automodule:: paddle.v2.fluid.optimizer
:members: AdagradOptimizer
:noindex:
AdamOptimizer
-------------
.. automodule:: paddle.v2.fluid.optimizer
:members: AdamOptimizer
:noindex:
AdamaxOptimizer
-----------
.. automodule:: paddle.v2.fluid.optimizer
:members: AdamaxOptimizer
:noindex:
DecayedAdagradOptimizer
-----------------------
.. automodule:: paddle.v2.fluid.optimizer
:members: DecayedAdagradOptimizer
:noindex:
doc/api/v2/fluid/param_attr.rst 0 → 100644
浏览文件 @ fab63cc6
===========
ParamAttr
===========
ParamAttr
-----------
.. automodule:: paddle.v2.fluid.param_attr
:members: ParamAttr
:noindex:
doc/api/v2/fluid/profiler.rst 0 → 100644
浏览文件 @ fab63cc6
===========
Profiler
===========
Profiler
-----------
.. autofunction:: paddle.v2.fluid.profiler.cuda_profiler
:noindex:
doc/api/v2/fluid/regularizer.rst 0 → 100644
浏览文件 @ fab63cc6
===========
Regularizer
===========
WeightDecayRegularizer
----------------------
.. automodule:: paddle.v2.fluid.regularizer
:members: WeightDecayRegularizer
:noindex:
L2DecayRegularizer
------------------
.. automodule:: paddle.v2.fluid.regularizer
:members: L2DecayRegularizer
:noindex:
L1DecayRegularizer
-------------------
.. automodule:: paddle.v2.fluid.regularizer
:members: L1DecayRegularizer
doc/design/backward.md 0 → 100644
浏览文件 @ fab63cc6
# Backward Building
## Motivation
In Neural Network, most models are solved by the backpropagation algorithm(known as **BP**) at present. Technically, BP calculates the gradient of the loss function, then propagates it back through the networks following the chain rule. However, when configuring the model structure, users do not need to define the backward part. So a mechanism is required by the framework which can complete the model's backward part automatically according to the given forward part.
When implementing a specific `op`, the developer is also asked to implement its backward version, called `grad_op`. A `grad_op` takes gradients of its corresponding `op`'s outputs, and calculate gradients of the `op`'s inputs. During the building of a model's backward part, the framework creates each forward `op`'s `grad_op`, and then string them together in reverse order of forwarding part. In this way, gradients spread from the end to the beginning of the model, in another word, from the loss to parameters.
## Challenges
The motivation of backward building is apparent. However, implementation it correctly is not so easy. In the **Fluid** design, a deep learning model is described by `Program`, `Block`, `Op` and `Variable`. The `Block` itself can be nested. It means that the `op`s and `variable`s are scattered across different blocks rather than all be gathered in a single graph. Our backward building algorithm shall visit blocks in recursive order and be able to insert `grad_op`s and new created `variable`s into the right place.
## Usage
Although the whole algorithm is comprised of many functions, only one is exposed as API:
```python
def append_backward(loss, parameter_list=None, no_grad_set=None):
"""
Append backward part to main_program
Args:
loss(Variable): The variable generated by the cost function.
parameter_list(list): Parameters that need to be updated by optimizers.
If None, it means all parameters need to be updated.
no_grad_set(set): Variables that have no gradients in Block 0.
If None, the set will be generated inside the function and
contains all variables with `step_gradient=True` from all blocks.
Return:
(list[Variable]): list of (parameters, gradients) pair.
"""
```
By invoking this API, the framework appends backward part of the program where the `loss` is. It takes three arguments. `loss` means the final loss value. It must be a scalar and is usually the output of the loss layer. It is also where the gradient generated and backpropagation starts. `parameter_list` marks all parameters needs updating. If it's `None`, all parameter will be updated by optimizers. `no_grad_set` marks variables without gradient. if all outputs of some `grad_op` are in `no_grad_set`, the `grad_op` will not be run.
This API will be invoked automatically before optimizer building.
As a result, in most cases, users do not need to invoke the API by themselves to append backward part.
## Implementation
The implementation of backward building algorithm is in `backward.py` file. The whole algorithm can be divided into two independent parts: creating `grad_op`s and creating new variables.
### Creating `grad_op`s
The creating of `grad_op`s is implemented by:
```python
def _append_backward_ops_(target,
block,
target_block,
no_grad_dict,
grad_to_var):
"""
Create all grad ops, and insert them into given block
Args:
target(Variable): the target variable of forward pass
block(Block): the block where forward ops are
target_block(Block): the block which is going to hold new generated grad ops
no_grad_dict(dict):
key(int) block index
val(set) a set of varibale names. These varibales have no gradient
grad_to_var(dict)(output argument):
key(str): grad variable name
val(str): corresponding forward variable name
"""
```
Given a `block`, the function will traverses all `op`s in this block in reverse order, gets corresponding `grad_op` from the C++ core via `core.get_grad_op_desc()`, then append it to `target_block`.
However, some specific `op`(e.g. `while_op`, `if_else_op`) can hold its own sub-block. For these sub-blocks contains `op`s as well, the `grad_op` creating should be recursive.
During the reverse traversal, we check each `op` whether it has an attribute named `sub_block`. If so, it means there is a sub-block and we need to deal with it first. After creating a new block whose father is the one in `op`'s attribute, we invoke `_append_backward_ops_()` recursively, assigning the new block to parameter `target_block` and the one in `op`'s attribute to `block`. The *pseudo-code* shows this process:
```
******* pseudo-code ********
for op in reversed(block.ops):
if op has an attribute named 'sub_block':
Get the sub-block(`s_block`) from op's attribute.
Create a new block(`grad_s_block`), whose father is `s_block`.
Invoke _append_backward_ops_(), with `block=s_block` and `target_block=grad_s_block`
Invoke `core.get_grad_op_desc()` to get op's grad_op.
Insert name correspondings between variables and their gradients of the grad_op to grad_to_var
Assign grad_s_block to grad_op as it's 'sub_block' attribute.
Append grad_op to current target_block.
```
The first invoking of `_append_backward_ops_()` is initiated by `append_backward()`, in which parameters `block` and `target_block` are all assigned with root block(the block with index 0).
### Corner Cases of `grad_op` Creating
In the previous section, we show the regular process of `grad_op` creating. However, in some corner cases, the conventional algorithm is not enough to get the correct result and appending handling is required. These additional processes run after the algorithm mentioned above and do some special adjusts on its output `grad_op`s.
#### Shared Variables
If a variable is read by more than one `op` in the forward pass, its gradient is likely to be written by more than one `grad_op`s in the next backward pass. To make the gradient result being the sum of all `grad_op`s' outputs instead of the last running one, we assign each output with a temporary variable and then add a `sum_op` to add them up.
For the debug convenience, if the final gradient name is `w@GRAD`, it's corresponding temporary variables will be named as `w@GRAD@RENAME@0`, `w@GRAD@RENAME@1`...
See function `_addup_repetitive_outputs_` in `backward.py` for implementation details.
#### No Gradient Variables
In our framework, variables can be marked as *no_gradient*, it means that the gradient of this variable is unnecessary and can be considered as zero in model training. Apparently, when all the outputs of some `grad_op` are marked as *no_gradient*, the `grad_op` itself can be skipped in backward pass.
But these unnecessary gradients still need to be creating and initialized by something, otherwise following `grad_op`s who take these gradients as inputs take the risk of using uninitialized memory. In our code, we employ `fill_zeros_like_op` to initialize them as all zeros.
This features are implemented in function `_remove_no_grad_branch_`. It checks new created `grad_op`s one-by-one, removes whose outputs are all in `no_grad_set` or inserts `fill_zeros_like_op` when its necessary. We can get the `no_grad_set` from the `_append_backward_ops_` argument `no_grad_dict` or generate it on the fly by scanning all variables' `no_gradient` attribute(True or False).
### Creating Backward Variables
Up to now, we have completed all creating and adjusting jobs of `grad_op`s. However, backward variables have not been created. Now they are only represented by `grad_op`'s input and output arguments. The backward variable creating job will be done by:
```python
def _append_backward_vars_(block,
start_op_idx,
grad_to_var,
grad_info_map):
"""
Create new variables required by backward pass.
Args:
block(Block): the block where new variables will be created
start_op_idx(int): Only variables required by ops in block.ops[start_op_idx : ] will be created
grad_to_var(dict):
key(str): grad variable name
val(str): corresponding forward variable name
In most cases, this dict is generated by _append_backward_ops_()
grad_info_map(dict)(output argument):
key(str): forward variable name
val(tuple): a tuple of (str, int), str is the corresponding grad name, int is the block index
"""
```
Given a `block`, this function traverses all the `grad_op`s in it(The argument `start_op_idx` indicates where the grad_op sequence starts.) and creates all the uncreated outputs. The *pseudo-code* shows this process:
```
for op in block.ops[start_op_idx : ]:
if op has an attribute named 'sub_block':
Get the sub-block(`s_block`) from op's attribute.
Invoke _append_backward_vars_(), with `block=s_block`
for var_name in op.all_output_names():
if block.has_var_recursive(var_name) or var_name is the name of empty variable:
continue
create a new variable named 'var_name' in block
if grad_to_var.has_key(var_name):
set grad_info_map[grad_to_var[var_name]] as a tuple of (var_name. block)
do op's var type inference
do op's shape inference
```
doc/design/block.md
浏览文件 @ fab63cc6
......@@ -291,10 +291,10 @@ public:
}
void Run(const framework::Scope& scope,
const platform::DeviceContext& dev_ctx) const override {
const platform::Place& place) const override {
PADDLE_ENFORCE(symbols_ready_, "operators and variables should be created first.");
for (auto& op : runtime_table_.ops()) {
op->Run(scope, dev_ctx);
op->Run(scope, place);
}
}
......
doc/design/evaluator.md
浏览文件 @ fab63cc6
## Evaluator Design
### The Problem
### Problem Statement
During training or serving, we provide the evaluation function to measure the model performance, e.g., accuracy, precision. In the operator based framework design, the data go through the network pipeline batch by batch. As a result, inside the operator, we only can calculate one minibatch metrics. We need to provide a mechanism to calculate the metrics for each N pass/batch the user wanted.
During training or inference, we provide an evaluation function to measure the model performance, for example, accuracy, precision, etc. In the operator based framework design, the data passes through the network pipeline batch by batch. As a result, inside the operator, we only calculate the metrics for one minibatch. Thus, we need to provide a mechanism to calculate the metrics for each N pass/batch the user wants.
### Evaluator Design
Currently, every operation is expressed in the graph. we divide the evaluator process into three steps.
Currently, every operation is expressed in the graph. We divide the evaluator process into three steps.
1. Initialize the metric state and add it into the block.
2. Calculate the statistic of the metric state in every mini-batch. The single operator is only responsible for calculating necessary statistics for one mini-batch. For example, accuracy operator only calculate a minibatch data if run once.
2. Calculate the concerned metrics for every mini-batch. The single evaluator operator is only responsible for calculating the necessary statistics for one mini-batch. For example, the accuracy operator only calculates the accuracy for a minibatch data if run once.
3. Merge the mini-batch statistics to form the evaluation result for multiple mini-batches. When it comes to distributed training/Multi-GPU training, aggregate the value from different devices.
### Implementation
This design is shown in python API.
Each metric operator need to caculate the metric statistic and return the batch aware states, Python side responsible for accumulate the states for each pass.
This design is shown in the Python API.
Each metric operator needs to caculate the metric statistic and return the batch-aware states. Python side is responsible for accumulating the states for each pass.
```python
......
doc/design/executor.md
浏览文件 @ fab63cc6
# Executor Design Doc
## Motivation
In [fluid](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/fluid.md), we encourage the user to use deep learning programming paradigms to describe the training process. When the user-written Python program is executed, it will first create a protobuf message
[`ProgramDesc`](https://github.com/PaddlePaddle/Paddle/blob/a91efdde6910ce92a78e3aa7157412c4c88d9ee8/paddle/framework/framework.proto#L145) that describes the process and is conceptually like an [abstract syntax tree](https://en.wikipedia.org/wiki/Abstract_syntax_tree).
We use executor to do the runtime evaluation of a `ProgramDesc`.
The executor runs the `ProgramDesc` like an interpreter. `ProgramDesc` contains the intrinsics (operators in this case) and variables which will be used, executor explicitly executes the stored precompiled code.
## Overview
An executor takes a `ProgramDesc`, a `block_id` and a `Scope`. The `ProgramDesc` is a list of blocks and each block contains the protobuf definition of all the parameters and operators. The `block_id` specifies the entrance block. And the `Scope` is the container of all the variable instance, which is persistent throughout different runs.
An executor takes a `ProgramDesc`, a `block_id` and a `Scope`. The `ProgramDesc` is a list of blocks and each block contains the protobuf definition of all the parameters and operators in the block. The `block_id` specifies the entrance block. And the `Scope` is the container of all the variable instances, which is persistent throughout different runs.
### What does executor do?
## Executor
It evaluates all the operators in the `block_id`th block of a `ProgramDesc`.
The `Executor` explicitly executes all the intrinsics (operators here) in the `block_id`th block of a `ProgramDesc`. Essentially, it instantiates Variables and Operators, then runs all the operators in sequence one-by-one.
It is very similar to how a push stack frame works when entering a block, following which it cleans up all the temporary variables when a mini-batch is finished. It does not however, have the stack frame pop process.
### What does executor NOT do?
### The interface
```c++
Executor(places);
```
A executor does not own any computing resources, a user can only construct an executor using the specified places.
It does not do runtime optimization, meaning intelligently parse the dependency of each op a choose which one to be run and in which order they should be run.
### Running an Executor
It does not do graph partitioning, meaning dividing the `ProgramDesc` into several small pieces and executing them on different devices.
## Implementation
`Executor` evaluates a `ProgramDesc`. Essentially, it instantiates Variables and Operators, then run all the operators in sequence. [[code]](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/executor.cc)
```
void Run(ProgramDesc, Scope, block_id, create_local_scope);
```
An `Executor` only provides a unified way to execute `ProgramDesc`. `ProgramDesc` is the target that will be executed, the `Scope` specifies the variable container, the `block_id` indicates the entrance block and `create_local_scope` is a boolean that states whether it will destroy the temporary variables after the execution is finished.
doc/design/float16.md
浏览文件 @ fab63cc6
......@@ -28,6 +28,51 @@ The goal of float16 is to serve as a key for the executor to find and run the co
- [Eigen](https://github.com/RLovelett/eigen) >= 3.3 supports float16 calculation on both GPU and CPU using the `Eigen::half` class. It is mostly useful for Nvidia GPUs because of the overloaded arithmetic operators using cuda intrinsics. It falls back to using software emulation on CPU for calculation and there is no special treatment to ARM processors.
- [ARM compute library](https://github.com/ARM-software/ComputeLibrary) >= 17.02.01 supports NEON FP16 kernels (requires ARMv8.2-A CPU).
### CUDA version issue
There are currently three versions of CUDA that supports `__half` data type, namely, CUDA 7.5, 8.0, and 9.0.
CUDA 7.5 and 8.0 define `__half` as a simple struct that has a `uint16_t` data (see [`cuda_fp16.h`](https://github.com/ptillet/isaac/blob/9212ab5a3ddbe48f30ef373f9c1fb546804c7a8c/include/isaac/external/CUDA/cuda_fp16.h)) as follows:
```
typedef struct __align__(2) {
unsigned short x;
} __half;
typedef __half half;
```
This struct does not define any overloaded arithmetic operators. So you have to directly use `__hadd` instead of `+` to correctly add two half types:
```
__global__ void Add() {
half a, b, c;
c = __hadd(a, b); // correct
c = a + b; // compiler error: no operator "+" matches these operands
}
```
CUDA 9.0 provides a major update to the half data type. The related code can be found in the updated [`cuda_fp16.h`](https://github.com/ptillet/isaac/blob/master/include/isaac/external/CUDA/cuda_fp16.h) and the newly added [`cuda_fp16.hpp`](https://github.com/ptillet/isaac/blob/master/include/isaac/external/CUDA/cuda_fp16.hpp).
Essentially, CUDA 9.0 renames the original `__half` type in 7.5 and 8.0 as `__half_raw`, and defines a new `__half` class type that has constructors, conversion operators, and also provides overloaded arithmetic operators such as follows:
```
typedef struct __CUDA_ALIGN__(2) {
unsigned short x;
} __half_raw;
struct __CUDA_ALIGN__(2) __half {
protected:
unsigned short __x;
public:
// constructors and conversion operators from/to
// __half_raw and other built-in data types
}
typedef __half half;
__device__ __forceinline__
__half operator+(const __half &lh, const __half &rh) {
return __hadd(lh, rh);
}
// Other overloaded operators
```
This new design makes `c = a + b` work correctly for CUDA half data type.
## Implementation
The float16 class holds a 16-bit `uint16_t` data internally.
......
doc/design/fluid.md 0 → 100644
浏览文件 @ fab63cc6
# Design Doc: PaddlePaddle Fluid
## Why Fluid
When Baidu developed PaddlePaddle in 2013, the only well-known open source deep learning system at the time was Caffe. However, when PaddlePaddle was open-sourced in 2016, many other choices were available. There was a challenge -- what is the need for open sourcing yet another deep learning framework?
Fluid is the answer. Fluid is similar to PyTorch and TensorFlow Eager Execution, which describes the "process" of training or inference using the concept of a model. In fact in PyTorch, TensorFlow Eager Execution and Fluid, there is no concept of a model at all. The details are covered in the sections below. Fluid is currently more extreme in the above mentioned idea than PyTorch and Eager Execution, and we are trying to push Fluid towards the directions of a compiler and a new programming language for deep learning.
## The Evolution of Deep Learning Systems
Deep learning infrastructure is one of the fastest evolving technologies. Within four years, there have already been three generations of technologies invented.
| Existed since | model as sequence of layers | model as graph of operators | No model |
|--|--|--|--|
| 2013 | Caffe, Theano, Torch, PaddlePaddle | | |
| 2015 | | TensorFlow, MxNet, Caffe2, ONNX, n-graph | |
| 2016 | | | PyTorch, TensorFlow Eager Execution, PaddlePaddle Fluid |
From the above table, we see that the deep learning technology is evolving towards getting rid of the concept of a model. To understand the reasons behind this direction, a comparison of the *programming paradigms* or the ways to program deep learning applications using these systems, would be helpful. The following section goes over these.
## Deep Learning Programming Paradigms
With the systems listed as the first or second generation, e.g., Caffe or TensorFlow, an AI application training program looks like the following:
```python
x = layer.data("image")
l = layer.data("label")
f = layer.fc(x, W)
s = layer.softmax(f)
c = layer.mse(l, s)
for i in xrange(1000): # train for 1000 iterations
m = read_minibatch()
forward({input=x, data=m}, minimize=c)
backward(...)
print W # print the trained model parameters.
```
The above program includes two parts:
1. The first part describes the model, and
2. The second part describes the training process (or inference process) for the model.
This paradigm has a well-known problem that limits the productivity of programmers. If the programmer made a mistake in configuring the model, the error messages wouldn't show up until the second part is executed and `forward` and `backward` propagations are performed. This makes it difficult for the programmer to debug and locate a mistake that is located blocks away from the actual error prompt.
This problem of being hard to debug and re-iterate fast on a program is the primary reason that programmers, in general, prefer PyTorch over the older systems. Using PyTorch, we would write the above program as following:
```python
W = tensor(...)
for i in xrange(1000): # train for 1000 iterations
m = read_minibatch()
x = m["image"]
l = m["label"]
f = layer.fc(x, W)
s = layer.softmax(f)
c = layer.mse(l, s)
backward()
print W # print the trained model parameters.
```
We can see that the main difference is the moving the model configuration part (the first step) into the training loop. This change would allow the mistakes in model configuration to be reported where they actually appear in the programming block. This change also represents the model better, or its forward pass, by keeping the configuration process in the training loop.
## Describe Arbitrary Models for the Future
Describing the process instead of the model also brings Fluid, the flexibility to define different non-standard models that haven't been invented yet.
As we write out the program for the process, we can write an RNN as a loop, instead of an RNN as a layer or as an operator. A PyTorch example would look like the following:
```python
for i in xrange(1000):
m = read_minibatch()
x = m["sentence"]
for t in xrange x.len():
h[t] = the_step(x[t])
```
With Fluid, the training loop and the RNN in the above program are not really Python loops, but just a "loop structure" provided by Fluid and implemented in C++ as the following:
```python
train_loop = layers.While(cond)
with train_loop.block():
m = read_minibatch()
x = m["sentence"]
rnn = layers.While(...)
with rnn.block():
h[t] = the_step(input[t])
```
An actual Fluid example is described [here](https://github.com/PaddlePaddle/Paddle/blob/a91efdde6910ce92a78e3aa7157412c4c88d9ee8/python/paddle/v2/fluid/tests/test_while_op.py#L36-L44).
From the example, the Fluid programs look very similar to their PyTorch equivalent programs, except that Fluid's loop structure, wrapped with Python's `with` statement, could run much faster than just a Python loop.
We have more examples of the [`if-then-else`](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/if_else_op.md) structure of Fluid.
## Turing Completeness
In computability theory, a system of data-manipulation rules, such as a programming language, is said to be Turing complete if it can be used to simulate any Turing machine. For a programming language, if it provides if-then-else and loop, it is Turing complete. From the above examples, Fluid seems to be Turing complete; however, it is noteworthy to notice that there is a slight difference between the `if-then-else` of Fluid and that of a programming language. The difference being that the former runs both of its branches and splits the input mini-batch into two -- one for the True condition and another for the False condition. This hasn't been researched in depth if this is equivalent to the `if-then-else` in programming languages that makes them Turing-complete. Based on a conversation with [Yuang Yu](https://research.google.com/pubs/104812.html), it seems to be the case but this needs to be looked into in-depth.
## The Execution of a Fluid Program
There are two ways to execute a Fluid program. When a program is executed, it creates a protobuf message [`ProgramDesc`](https://github.com/PaddlePaddle/Paddle/blob/a91efdde6910ce92a78e3aa7157412c4c88d9ee8/paddle/framework/framework.proto#L145) that describes the process and is conceptually like an [abstract syntax tree](https://en.wikipedia.org/wiki/Abstract_syntax_tree).
There is a C++ class [`Executor`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/executor.h), which runs a `ProgramDesc`, similar to how an interpreter runs a Python program.
Fluid is moving towards the direction of a compiler, which is explain in more detail later in this article.
## Backward Compatibility of Fluid
Given all the advantages from the removal of the concept of a *model*, hardware manufacturers might still prefer the existence of the concept of a model, so it would be easier for them to support multiple frameworks all at once and could run a trained model during inference. For example, Nervana, a startup company acquired by Intel, has been working on an XPU that reads the models in the format known as [n-graph](https://github.com/NervanaSystems/ngraph). Similarly, [Movidius](https://www.movidius.com/) is producing a mobile deep learning chip that reads and runs graphs of operators. The well-known [ONNX](https://github.com/onnx/onnx) is also a file format of graphs of operators.
For Fluid, we can write a converter that extracts the parts in the `ProgramDesc` protobuf message, converts them into a graph of operators, and exports the graph into the ONNX or n-graph format.
## Towards a Deep Learning Language and the Compiler
We can change the `if-then-else` and loop structure a little bit in the above Fluid example programs, to make it into a new programming language, different than Python.
Even if we do not invent a new language, as long as we get the `ProgramDesc` message filled in, we can write a transpiler, which translates each invocation to an operator, into a C++ call to a kernel function of that operator. For example, a transpiler that weaves the CUDA kernels outputs an NVIDIA-friendly C++ program, which can be built using `nvcc`. Another transpiler could generate MKL-friendly code that should be built using `icc` from Intel. More interestingly, we can translate a Fluid program into its distributed version of two `ProgramDesc` messages, one for running on the trainer process, and the other one for the parameter server. For more details of the last example, the [concurrent programming design](concurrent_programming.md) document would be a good pointer. The following figure explains the proposed two-stage process:
![](fluid-compiler.png)
doc/design/kernel_hint_design.md 0 → 100644
浏览文件 @ fab63cc6
## Problem
In PaddlePaddle's [Design](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/switch_kernel.md), one Operator may have multiple kernels. Users may have some personal preference to choose a certain type of kernel for an operator, such as `force_cpu` to choose a CPU kernel, `use_cudnn` to choose a CUDNN kernel, we need to provide a way for users to do this.
In the current design, we use KernelType to describe one kernel.
```cpp
struct KernelType {
Place place_;
DataType data_type_;
LayoutType layout_;
};
```
`place_` `data_type_` and `layout_` can be got from the input tensors of the operator, `GetActualKernelType(inputs)` use inputs to infer the proper kernel key that fit the incoming data, but users can not directly configure it.
The [design](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/switch_kernel.md) also provides a virtual method `GetExpectedKernelType` that user can overload and use to choose the KernelType they want to use.
So we should send the information user defined in proto to `GetExpectedKernelType` for choosing a kernel.
The problem is, how should we define and send the information for `GetExpectedKernelType` to use?
## Solution
### Potential choice
1. Do nothing, let the user add the information they want to operator‘s attribute and get them inside `GetExpectedKernelType`, this can work properly. But there is a little problem that users may define many kinds of hints for the same purpose, such as `force_cpu`, `use_cpu`, `cpu_kernel` to choose CPU kernel, and `use_cudnn`, `force_cudnn`, `cudnn_kernel` to choose CUDNN kernel.
2. Pre-define all the needed option and use a single attr key such as `kernel_hint` for the user, this is not so flexible if the user wants to define some more kind of hint.
### Final choice
To provide enough flexibility while avoiding confusion definition, we can define some global constants for these attribute names, such as `force_cpu`, `use_cudnn`, `use_mkldnn` for a user to choose.
In C++
```cpp
const std::string kForceCPU = "force_cpu";
const std::string kUseCUDNN = "use_cudnn";
const std::string kUseMKLDNN = "use_mkldnn";
KernelType GetExpectedKernelType() {
if (Attr<bool>(kForceCPU)) {
return KernelType(CPUPlace, ...)
} else {
...
}
}
```
In Python code
```python
FORCE_CPU = core.kForceCPU()
def xx_layer(..., force_cpu=false):
layer_helper = LayerHelper(...)
layer_helper.append_op(
type="xx",
attr={FORCE_CPU: force_cpu})
```
doc/design/mkl/mkl_packed.md 0 → 100644
浏览文件 @ fab63cc6
# Intel® MKL Packed on PaddlePaddle: Design Doc
## Contents
- [Overview](#overview)
- [Key Points](#key-points)
- [Background](#background)
- [Solution](#solution)
- [Actions](#actions)
- [CMake](#cmake)
- [Layers](#layers)
- [Unit Tests](#unit-tests)
- [Python API](#python-api)
- [Benchmarking](#benchmarking)
## Overview
我们计划将 Intel® MKL 中引入的 GEMM Packed APIs\[[1](#references)\] 集成到 PaddlePaddle 中，充分发挥英特尔平台的优势，有效提升PaddlePaddle在英特尔架构上的性能。
现阶段的优化主要针对 Recurrent Neural Network（以下简称RNN）相关层（包括`RecurrentLayer`, `GatedRecurrentLayer`和`LstmLayer`）， 以及 PaddlePaddle V1 API。
## Key Points
### Background
目前PaddlePaddle采用了 Intel® MKL库的[cblas_?gemm](https://software.intel.com/en-us/mkl-developer-reference-c-cblas-gemm)函数，这个函数本身会在计算前将原数据转换为更适合英特尔平台的内部格式。
1. 转换耗时 \
这一数据格式的转换操作（Packing），在问题本身的计算量比较小的时候，显得相对来说较为耗时。例如在DeepSpeech2 \[[2](#references)\] 的Vanilla RNN部分中，矩阵大小是`batch_size * 2048`。
2. 转换冗余 \
由于在现有的某些情况下（例如RNN），多次调用 cblas_?gemm 会使用相同的原数据，因此，每次调用时对原数据的重复Packing便成为了冗余。
为了最大程度减少多次调用 cblas_?gemm 在Packing上的耗时，Intel® MKL 引入了以下四个API:
* [cblas_?gemm_alloc](https://software.intel.com/en-us/mkl-developer-reference-c-cblas-gemm-alloc)
* [cblas_?gemm_pack](https://software.intel.com/en-us/mkl-developer-reference-c-cblas-gemm-pack)
* [cblas_?gemm_compute](https://software.intel.com/en-us/mkl-developer-reference-c-cblas-gemm-compute)
* [cblas_?gemm_free](https://software.intel.com/en-us/mkl-developer-reference-c-cblas-gemm-free)
通过使用这些API，我们可以先完成对原数据的Packing操作，再把已转换为Packed格式的数据传递给那些复用同一数据的gemm_compute函数，从而避免了Packing冗余。
### Solution
在RNN的情况下，同一次前向、后向（forward/backward）过程中所有时间步（time step）共享同一个权重（weight）。当只做推断（inference）时，各次前向之间也都使用了相同的权重，没有必要在每次前向中每个时间步的计算时对权重进行重复的Packing操作。
我们通过使用新引入的GEMM Packed APIs，在层初始化的时候，先完成对权重的Packing操作，然后在前向，后向时复用已经转换过的权重，并在每次权重更新后，对新的权重进行转换用于下次迭代。
* 优化前，对于序列长度（sequence length）为`T`的网络模型（model）, `N`次迭代执行的转换次数为：
- `inference`： `N * T`
- `training`： `2 * N * T`
* 优化后，对于同样设置的网络模型，其转换次数减少至：
- `inference`： `1`
- `training`： `2 * N`
## Actions
添加的相关文件和目录结构如下：
```txt
PaddlePaddle/Paddle
├── ...
└── paddle/
├── ...
└── gserver/
├── ...
├── layers/
│ ├── ...
│ ├── MKLPackedRecurrentLayer.*
| ├── MKLPackedGatedRecurrentLayer.*
| ├── MKLPackedLstmLayer.*
| └── MKLPackedGemm.h
└── tests/
├── ...
└── test_MKLPacked.cpp
```
### CMake
在对应的`CMakeLists.txt`中根据`WITH_MKL`是否打开，来决定是否开启MKL Packed相关功能。
### Layers
所有的`MKLPacked*Layer`都继承于PaddlePaddle的基类`Layer`, 并添加头文件 `MKLPackedGemm.h`，该文件对相关GEMM Packed APIs做了封装。
### Unit Tests
我们会添加`test_MKLPacked.cpp`用于MKL Packed优化后layer的测试。
对于每一个新加的RNN layer，我们会对比如下2个方面：
1. 对比优化后layer自身，sequence mode（`rnn_use_batch=false`）与batch mode(`rnn_use_batch=true`)的结果。
2. 对比优化后layer与相对应的PaddlePaddle原有layer, 在batch mode下的结果。
### Python API
计划在`paddle/utils.Flags`中添加`use_mkl_packed`的flag，用于选择是否使用相关功能，并且当编译时`WITH_MKL=ON`的情况下，默认设置为`true`。
同时，在`python/paddle/trainer/config_parser.py`中对应的layer处，添加`use_mkl_packed`这个选择，方便用户在Python端选择是否启用这个功能。
具体实现方式比如：
```python
use_mkl_packed = bool(int(g_command_config_args.get("use_mkl_packed", 0)))
if use_mkl_packed:
self.layer_type = mkl_packed_*
```
所有相关的`layer_type`会以*mkl_packed_*开头，这些会在`MKLPacked*Layer`注册layer的时候保证，以示区分。
### Benchmarking
会添加相应的脚本用于测试和对比在使用MKL Packed recurrent layers 前后的网络性能。
## References
1. [Introducing the new Packed APIs for GEMM](https://software.intel.com/en-us/articles/introducing-the-new-packed-apis-for-gemm)
2. [DeepSpeech2 on PaddlePaddle](https://github.com/PaddlePaddle/DeepSpeech#deepspeech2-on-paddlepaddle)
doc/design/mkl/mkldnn.md 0 → 100644
浏览文件 @ fab63cc6
# Intel® MKL-DNN on PaddlePaddle: Design Doc
我们计划将英特尔深度神经网络数学库[Intel MKL-DNN](https://github.com/01org/mkl-dnn)
(Intel Math Kernel Library for Deep Neural Networks)集成到PaddlePaddle，
充分展现英特尔平台的优势，有效提升PaddlePaddle在英特尔架构上的性能。
<div align="center">
<img src="image/overview.png"><br/>
Figure 1. PaddlePaddle on IA
</div>
近期目标
- 完成常用Layer的MKL-DNN实现。
- 完成常见深度神经网络VGG，GoogLeNet 和 ResNet的MKL-DNN实现。
目前的优化，主要针对PaddlePaddle在重构之前的代码框架以及V1的API。
具体的完成状态可以参见[这里](https://github.com/PaddlePaddle/Paddle/projects/21)。
## Contents
- [Overview](#overview)
- [Actions](#actions)
- [CMake](#cmake)
- [Matrix](#matrix)
- [Layers](#layers)
- [Activations](#activations)
- [Parameters](#parameters)
- [Gradients](#gradients)
- [Unit Tests](#unit-tests)
- [Python API](#python-api)
- [Benchmarking](#benchmarking)
- [Others](#others)
- [Design Concerns](#design-concerns)
## Overview
我们会把MKL-DNN会作为第三方库集成进PaddlePaddle，与其他第三方库一样，会在编译PaddlePaddle的时候下载并编译MKL-DNN。
同时，为了进一步提升PaddlePaddle在基本数学运算的计算速度，我们也将MKLML即(MKL small library\[[1](#references)\])
作为另一个第三方库集成进PaddlePaddle，它只会包括生成好的动态库和头文件。
MKL，MKLML以及MKL-DNN三者关系如下表：
| Name | Open Source | License | Descriptions |
| :---------- | :--------------- | :---------- | :------------ |
| MKL | No | Proprietary | Accelerate math processing routines |
| MKLML | No | Proprietary | Small package of MKL, especially for Machine Learning |
| MKL-DNN | Yes | Apache 2.0 | Accelerate primitives processing routines especially for Deep Neural Networks |
MKLML可以与MKL-DNN共同使用，以此达到最好的性能。
<div align="center">
<img src="image/engine.png"><br/>
Figure 2. PaddlePaddle with MKL Engines
</div>
## Actions
添加的相关文件和目录结构如下：
```txt
PaddlePaddle/Paddle
├── ...
├── cmake/
│ ├── external/
│ │ ├── ...
│ │ ├── mkldnn.cmake
│ │ └── mklml.cmake
└── paddle/
├── ...
├── math/
│ ├── ...
│ └── MKLDNNMatrix.*
└── gserver/
├── ...
├── layers/
│ ├── ...
│ └── MKLDNN*Layer.*
├── activations/
│ ├── ...
│ └── MKLDNNActivations.*
└── tests/
├── ...
├── MKLDNNTester.*
└── test_MKLDNN.cpp
```
### CMake
在`CMakeLists.txt`中提供一个与MKL有关的总开关：`WITH_MKL`，它负责决定编译时是否使用MKLML和MKL-DNN
- `WITH_MKLML` 控制是否使用MKLML库。
当打开`WITH_MKL`时，会自动使用MKLML库作为PaddlePaddle的CBLAS和LAPACK库，同时会开启Intel OpenMP用于提高MKLML的性能。
编译时会把对应的头文件和库放在`build/third_party/install/mklml/*`目录下对应的地方。
MKLML的库目前都是动态库，主要包括`libiomp5.so`和`libmklml_intel.so`。
- `WITH_MKLDNN` 控制是否使用MKL-DNN。
当开启`WITH_MKL`时，会自动根据硬件配置[[2](#references)]选择是否编译MKL-DNN。
编译时会把对应的头文件和库放在`build/third_party/install/mkldnn/*`目录下对应的地方。
MKL-DNN的库目前只有动态库`libmkldnn.so`。
### Matrix
目前在PaddlePaddle中数据都是以`NCHW`的格式存储，但是在MKL-DNN中的排列方式不止这一种。
所以我们定义了一个`MKLDNNMatrix`用于管理MKL-DNN数据的不同格式以及相互之间的转换。
<div align="center">
<img src="image/matrix.png"><br/>
Figure 3. MKLDNNMatrix
</div>
### Layers
所有MKL-DNN的Layers都会继承于`MKLDNNLayer`，该类继承于PaddlePaddle的基类`Layer`。
在`MKLDNNLayer`中会提供一些必要的接口和函数，并且会写好`forward`和`backward`的基本逻辑，
子类只需要使用定义好的接口，实现具体的函数功能即可。
<div align="center">
<img src="image/layers.png"><br/>
Figure 4. MKLDNNLayer
</div>
每个MKLDNNLayer都包含用于内部存储和外部存储的一系列MKLDNNMatrix：
- 内部存储（internel memory）：`inVal_`,`inGrad_`,`outVal_`和`outGrad_`，分别代表输入数据，输入梯度，输出数据和输出梯度。
- 外部存储（external memory）：都是以ext开头，比如`extInVal_`和`extInGrad_`，它们主要是用于，
当数据格式与PaddlePaddle默认的`NCHW`格式不匹配时，转换内存的工作。
需要注意的是，PaddlePaddle的activation会直接使用`output_.value`和`output_.grad`，
所以`extOutVal_`和`extOutGrad_`必须分别与`output_.value`和`output_.grad`共享内存，
如果不需要外部存储用于转换，那么对应的内部存储也会与它们共享内存。
- 转换函数（resetXXX）： 包括`resetInValue`，`resetInGrad`，`resetOutValue`和`resetOutGrad`，
表示对输入数据，输入梯度，输出数据和输出梯度的转换。
这些函数会根据输入参数重新设置内部和外部存储，当然这两者也可以相等，即表示不需要转换。
注意：每个`MKLDNNlayer`的子类只需要使用内部存储就可以了，所有外部的转换工作都会在reset系列函数中都准备好。
### Activations
在重构前的PaddlePaddle中，激活函数是独立于`Layer`的概念，并且输入输出都是共用一块内存，
所以添加了对应的`MKLDNNActivation`来实现，方式类似于`MKLDNNLayer`。
### Parameters
对于有参数的层，我们会保证`MKLDNNLayer`使用的参数与PaddlePaddle申请的buffer共用一块内存。
如果存在数据排列格式不一样的情况时，我们会在网络训练之前把格式转换为MKL-DNN希望的格式，
在训练结束的时候再保存为PaddlePaddle的格式，但是整个训练过程中不需要任何转换。
这样既使得最终保存的参数格式与PaddlePaddle一致，又可以避免不必要的转换。
### Gradients
由于MKL-DNN的操作都是直接覆盖的形式，也就是说输出的结果不会在原来的数据上累加，
这样带来的好处就是不需要一直清空memory，节省了不必要的操作。
但是注意的是，当网络出现分支且在`backward`的时候，需要累加不同Layer传过来的梯度。
所以在`MKLDNNlayer`中实现了一个merge的方法，此时每个小分支的`Input Gradient`
会先临时保存在`MKLDNNMatrix`中，由分支处的Layer负责求和，并把结果放到当前层的`output_.grad`中。
所以整体上，在实现每个子类的时候就不需要关心分支的事情了。
<div align="center">
<img src="image/gradients.png"><br/>
Figure 5. Merge Gradients
</div>
### Unit Tests
我们会添加`test_MKLDNN.cpp`和`MKLDNNTester.*`用于MKL-DNN的测试。
测试分为每个Layer（或Activation）的单元测试和简单网络的整体测试。
每个测试会对比PaddlePaddle中CPU算出的结果与MKL-DNN的结果，小于某个比较小的阈值认为通过。
### Python API
目前只考虑**v1 API**。
计划在`python/paddle/trainer/config_parser.py`里面添加`use_mkldnn`这个选择，方便用户选择使用MKL-DNN的layers。
具体实现方式比如：
```python
use_mkldnn = bool(int(g_command_config_args.get("use_mkldnn", 0)))
if use_mkldnn
self.layer_type = mkldnn_*
```
所有MKL-DNN的`layer_type`会以*mkldnn_*开头，这些会在`MKLDNN*Layer`注册layer的时候保证，以示区分。
同时,会在`paddle/utils.Flags`中添加一个`use_mkldnn`的flag，用于选择是否使用MKL-DNN的相关功能。
### Benchmarking
会添加相应的脚本在[这里](https://github.com/PaddlePaddle/Paddle/tree/develop/benchmark/paddle/image)，用于测试和对比在使用MKL-DNN前后的CNN网络性能。
测试的性能对比结果会在[IntelOptimizedPaddle.md](https://github.com/PaddlePaddle/Paddle/blob/develop/benchmark/IntelOptimizedPaddle.md)
### Others
1. 如果在使用MKL-DNN的情况下，会把CPU的Buffer对齐为4096，具体可以参考MKL-DNN中的[memory](https://github.com/01org/mkl-dnn/blob/master/include/mkldnn.hpp#L673)。
2. 深入PaddlePaddle，寻找有没有其他可以优化的可能，进一步优化。比如可能会用OpenMP改进SGD的更新性能。
## Design Concerns
为了更好的符合PaddlePaddle的代码风格\[[3](#references)\]，同时又尽可能少的牺牲MKL-DNN的性能\[[4](#references)\]。
我们总结出一些特别需要注意的点：
1. 使用**deviceId_**。为了尽可能少的在父类Layer中添加变量或者函数，
我们决定使用已有的`deviceId_`变量来区分layer的属性，定义`-2`为`MKLDNNLayer`特有的设备ID。
2. 重写父类Layer的**init**函数，修改`deviceId_`为`-2`，代表这个layer是用于跑在MKL-DNN的环境下。
3. 创建`MKLDNNBase`，定义一些除了layer和memory相关的类和函数。
包括MKL-DNN会用到`MKLDNNStream`和`CPUEngine`，和未来可能还会用到`FPGAEngine`等。
4. 如果MKL-DNN layer的后面接有cpu device，那么就会使`output_.value`与`extOutVal_`共享内存，
同时数据格式就是`NCHW`，这样下一个cpu device就能拿到正确的数据。
在有普通的CPU layer时， `extOutVal_`和`extOutGrad_`的格式始终是`NCHW`或者`NC`。
## References
1. [MKL small library](https://github.com/01org/mkl-dnn#linking-your-application)是[Intel MKL](https://software.intel.com/en-us/mkl)的一个子集。
主要包括了深度学习相关的数学原语与操作，一般由MKL-DNN在发布[新版本](https://github.com/01org/mkl-dnn/releases)时一起更新。
2. [MKL-DNN System Requirements](https://github.com/01org/mkl-dnn#system-requirements)。
目前在PaddlePaddle中，仅会在支持AVX2指令集及以上的机器才使用MKL-DNN。
3. [原来的方案](https://github.com/PaddlePaddle/Paddle/pull/3096)会引入**nextLayer**的信息。
但是在PaddlePaddle中，无论是重构前的layer还是重构后的op，都不会想要知道next layer/op的信息。
4. MKL-DNN的高性能格式与PaddlePaddle原有的`NCHW`不同(PaddlePaddle中的cuDNN部分使用的也是`NCHW`，所以不存在这个问题)。
所以需要引入一个转换方法，并且只需要在必要的时候转换这种格式，才能更好的发挥MKL-DNN的性能。
doc/design/mkl/mkldnn_fluid.md 0 → 100644
浏览文件 @ fab63cc6
# Design Doc: Add MKLDNN Kernel in Fluid Operator
## Principles
First of all, we should follow some basical principles like:
1. [How to write a new operator](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/dev/new_op_en.md). We are trying to add a new kind of kernel into operators, so basically we should follow this doc.
2. [Supporting new Device/Library](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/support_new_device.md). Since MKLDNN is a new library to fluid, we should add `MKLDNNDeviceContext` and maybe `mkldnn_helper.h`, just like [cudnn_helper.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/platform/cudnn_helper.h).
3. [Switch Kernel](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/switch_kernel.md). Another important point is that we should ensure the data synchronization between different kernel types, which is this [topic](https://github.com/PaddlePaddle/Paddle/issues/6549). So basically we should override `GetExpectedKernelType` and `trans` functions to support switching kernels.
4. [The Keys of Operator Kernel Type](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/operator_kernel_type.md). Kernel Type is a pivotal conception which can record the `Place`, `Library`, `DataType` and `Layout`.
## Sulution
In general, there are four parts we should follow to run a MKL-DNN primitive.
- Create a primitive descriptor that describe this operator
- Create a primitive itself by primitive descriptor and the engine
- Create all memory buffers that primitive needed
- Launch a stream to execute the primitive created
More details can refer to [here](http://01org.github.io/mkl-dnn).
It's better to avoid reinitialization of primitives and memory handles in the first three stages in every iteration. \
So we plan to create a map to record all the `primitive` and `memory`, which should not take too much memories as discussed [here](https://github.com/PaddlePaddle/Paddle/issues/6822).
It's assumed that following three conditions should be satisfied.
1. there is a unique key for each operator instance. May be the actual name of `Output Tensor`.
2. the `Input Tensor` inside `Compute` function is the one after converted.
3. we can get the phase(eg. `is_test`) inside `Compute` function, otherwise we need to expose this attribue to user.
### Compute
The algorithm of `Compute` would be described as follow, let's take conv like an example.
```c++
PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()), "It must use CPUPlace.");
PADDLE_ENFORCE(platform::is_mkldnn_library(ctx.GetLibrary()), "It must use MKLDNN Library.");
auto& dev_ctx = ctx.template device_context<platform::MKLDNNDeviceContext>();
// find primitive by unique key from mkldnn context
// the op_key should be a unique name of this op instance
auto& p = dev_ctx.findPrimitive(op_key + "_fwd");
// assuming the input tensor inside this compute function is the one after converted
// this point should be guarantee by another mechanism
auto& i = dev_ctx.findMemory(op_key + "_input");
if (p == nullptr || i == nullptr || inputSizeChanged(p, i)) {
auto fwd_primitive_desc = createPrimitiveDesc(ctx);
auto* input = ctx.Input<Tensor>("Input");
auto* filter = ctx.Input<Tensor>("Filter");
auto* output = ctx.Output<Tensor>("Output");
shared_ptr<mkldnn::memory> in(new mkldnn::memory(fwd_primitive_desc->src_primitive_desc(), input->data<T>()));
shared_ptr<mkldnn::memory> wgt(new mkldnn::memory(fwd_primitive_desc->weights_primitive_desc(), filter->data<T>()));
shared_ptr<mkldnn::memory> out(new mkldnn::memory(fwd_primitive_desc->dst_primitive_desc(), output->mutable_data<T>(ctx.GetPlace())));
shared_ptr<mkldnn::conv_fwd> fwd_primitive(new mkldnn::conv_fwd(*fwd_primitive_desc, *in, *wgt, *out));
dev_ctx.addMemory(op_key+"_input", in);
dev_ctx.addMemory(op_key+"_output", out);
dev_ctx.addMemory(op_key+"_filer", wgt);
dev_ctx.addPrimitive(op_key+"_fwd", fwd_primitive);
dev_ctx.addPrimitiveDesc(op_key+"_fwd_PD", fwd_primitive_desc);
}
p = dev_ctx.findPrimitive(op_key + "_fwd");
PADDLE_ENFORCE(p, "Should have forward Primitive");
PADDLE_ENFORCE(dev_ctx.findMemory(op_unique_key+"_input"), "Should have input memory");
PADDLE_ENFORCE(dev_ctx.findMemory(op_unique_key+"_output"), "Should have output memory");
PADDLE_ENFORCE(dev_ctx.findMemory(op_unique_key+"_filter"), "Should have filter memory");
PADDLE_ENFORCE(dev_ctx.findPrimitiveDesc(op_unique_key+"_fwd_PD"), "Should have forward PrimitiveDesc");
dev_ctx.submit(p);
dev_ctx.execute(); // the convert primitive should have already contained.
```
The `createPrimitiveDesc` returns the primitive descripotor of this operator, would be like this:
```c++
auto* input = ctx.Input<Tensor>("Input");
auto* filter = ctx.Input<Tensor>("Filter");
auto* output = ctx.Output<Tensor>("Output");
std::vector<int> strides = ctx.Attr<std::vector<int>>("strides");
std::vector<int> paddings = ctx.Attr<std::vector<int>>("paddings");
std::vector<int> dilations = ctx.Attr<std::vector<int>>("dilations");
int groups = ctx.Attr<int>("groups");
algorithm algo = static_cast<algorithm>(ctx.Attr<int>("convolution_algorithm_option"));
prop_kind pk = ctx.Attr<bool>("is_test") ? prop_kind::forward_inference : prop_kind::forward_training;
auto fwd_desc = mkldnn::conv_fwd::desc(/* all the setting above*/);
shared_ptr<mkldnn::conv_fwd::primitive_desc> fwd_primitive_desc(new mkldnn::conv_fwd::primitive_desc(fwd_desc, ctx.getEngine()));
return fwd_primitive_desc;
}
```
### MKLDNNDeviceContext
`MKLDNNDeviceContext`, which is very straightforward, should contain some base information like: `stream`, `engine` and the map needed.
### mkldnn_helper
Some functions would be put in `paddle/platform/mkldnn_helper.h`.
- create MKLDNN memories
- create MKLDNN primitives
- error check function
- etc
### Kernel Switch
We should `reorder` the different Layout from other device or to other device. `GetExpectedKernelType` and `trans` functions can help us to implement it.
`GetExpectedKernelType` should get the context, and this operator can return the best `KernelType`.
`trans` would be like this:
```c++
void trans(inputs, ctx) override {
if (NoNeedTrans()) {
return;
}
// find reorder primitive by op_key from context
auto& dev_ctx = ctx.template device_context<platform::MKLDNNDeviceContext>();
auto& p = dev_ctx.findPrimitive(op_key + "_reorder_input");
auto& i = dev_ctx.findMemory(op_key + "_src_input");
if (p == nullptr || i == nullptr || changeSized(i, input)) {
auto prim = createPrimitiveDesc(ctx);
auto src = createMemory(memoryDesc(input->dims(), actual_layout), input->data);
auto newbuffer = paddle::memory::Alloc(ctx.GetPlace(), input->size_in_bytes());
auto dst = createMemory(p->expected_desc(), newbuffer->data);
auto reorder_primitive(new mkldnn::reorder(src, dst));
dev_ctx.addMemory(op_key+"_src_input", src);
dev_ctx.addMemory(op_key+"_input", dst);
dev_ctx.addPrimitive(op_key+"_reorder_input", reorder_primitive);
}
p = dev_ctx.findPrimitive(op_key + "_reorder_input");
PADDLE_ENFORCE(p, "Should have Reorder Primitive");
dev_ctx.submit(p);
if (! this->isMKLDNNKernel()) {
// execute immediately only if this is not mkldnn kernel function.
// otherwise, it can be executed with the operator primitive in Compute
dev_ctx.stream();
}
// after submit, the input tensor in ExecutionContext should be changed as the converted one
// there should be another mechanism to ensure this
}
```
### Unit Test
All the functions should be tested corresponding.
TBD
doc/design/mkldnn/README.MD 已删除 100644 → 0
浏览文件 @ 1ce2d613
# Intel® MKL-DNN on PaddlePaddle: Design Doc
我们计划将Intel深度神经网络数学库(**MKL-DNN**\[[1](#references)\])集成到PaddlePaddle，充分展现英特尔平台的优势，有效提升PaddlePaddle在英特尔架构上的性能。
我们短期内的基本目标是：
- 完成常用layer的MKL-DNN实现。
- 完成常见深度神经网络VGG，GoogLeNet 和 ResNet的MKL-DNN实现。
## Contents
- [Overview](#overview)
- [Actions](#actions)
- [CMake](#cmake)
- [Layers](#layers)
- [Activations](#activations)
- [Weights](#weights)
- [Unit Tests](#unit-tests)
- [Protobuf Messages](#protobuf-messages)
- [Python API](#python-api)
- [Demos](#demos)
- [Benchmarking](#benchmarking)
- [Others](#others)
- [Design Concerns](#design-concerns)
## Overview
我们会把MKL-DNN作为第三方库集成进PaddlePaddle，整体框架图
<div align="center">
<img src="image/overview.png" width=350><br/>
Figure 1. PaddlePaddle on IA.
</div>
## Actions
我们把集成方案大致分为了如下几个方面。
### CMake
我们会在`CMakeLists.txt`中会给用户添加一个`WITH_MKL`的开关，他是负责`WITH_MKLML`和`WITH_MKLDNN`的总开关。
当打开`WITH_MKL`时，会开启MKLML的功能，作为PaddlePaddle的CBLAS和LAPACK库，同时会开启Intel OpenMP用于提高MKLML的性能。 如果系统支持AVX2指令集及以上，同时会开启MKL-DNN功能。
当关闭`WITH_MKL`时，MKLML和MKL-DNN功能会同时关闭。
所以，我们会在`cmake/external`目录新建`mkldnn.cmake`和`mklml.cmake`文件，它们会在编译PaddlePaddle的时候下载对应的软件包，并放到PaddlePaddle的third party目录中。
### Layers
所有MKL-DNN相关的C++ layers，都会按照PaddlePaddle的目录结构存放在
`paddle/gserver/layers`中，并且文件名都会一以*MKLDNN*开头。
所有MKL-DNN的layers都会继承于一个叫做`MKLDNNLayer`的父类，该父类继承于PaddlePaddle的基类`Layer`。
在`MKLDNNLayer`中会提供一些必要的接口和函数，并且会写好`forward`和`backward`的基本逻辑。部分函数定义为纯虚函数，子类只需要实现这些函数即可。
### Activations
由于在PaddlePaddle中，激活函数是独立于layer概念的，所以会在`paddle/gserver/activations`目录下添加`MKLDNNActivation.h`和`MKLDNNActivation.cpp`文件用于定义和使用MKL-DNN的接口。
### Weights
由于有些layer是含有参数的，我们会尽量让MKL-DNN的参数与PaddlePaddle中`parameter`共享一块内存。
同时，由于MKL-DNN在训练时使用的参数layout可能与PaddlePaddle默认的`nchw`不一致，我们会在网络训练的开始和结束时分别转换这个layout，使得最终保存的参数格式与PaddlePaddle一致。
### Unit Tests
会在`paddle/gserver/test`目录下添加`test_MKLDNN.cpp`和`MKLDNNTester.*`用于MKL-DNN的测试。
测试分为每个layer(或activation)的单元测试和简单网络的整体测试。
每个测试会对比PaddlePaddle中CPU算出的结果与MKL-DNN的结果，小于某个比较小的阈值认为通过。
### Protobuf Messages
根据具体layer的需求可能会在`proto/ModelConfig.proto`里面添加必要的选项。
### Python API
目前只考虑**v1 API**。
计划在`python/paddle/trainer/config_parser.py`里面添加`use_mkldnn`这个选择，方便用户选择使用MKL-DNN的layers。
具体实现方式比如：
```python
use_mkldnn = bool(int(g_command_config_args.get("use_mkldnn", 0)))
if use_mkldnn
self.layer_type = mkldnn_*
```
所有MKL-DNN的layer type会以*mkldnn_*开头，以示区分。
并且可能在`python/paddle/trainer_config_helper`目录下的`activations.py `和`layers.py`里面添加必要的MKL-DNN的接口。
### Demos
会在`v1_api_demo`目录下添加一个`mkldnn`的文件夹，里面放入一些用于MKL-DNN测试的demo脚本。
### Benchmarking
会添加`benchmark/paddle/image/run_mkldnn.sh`，用于测试使用MKL-DNN之后的性能。
### Others
1. 如果在使用MKL-DNN的情况下，会把CPU的Buffer对齐为64。
2. 深入PaddlePaddle，寻找有没有其他可以优化的可能，进一步优化。比如可能会用OpenMP改进SGD的更新性能。
## Design Concerns
为了更好的符合PaddlePaddle的代码风格\[[2](#references)\]，同时又尽可能少的牺牲MKL-DNN的性能\[[3](#references)\]。
我们总结出一些特别需要注意的点：
1. 使用**deviceId_**。为了尽可能少的在父类Layer中添加变量或者函数，我们决定使用已有的`deviceId_`变量来区分layer的属性，定义`-2`为`MKLDNNLayer`特有的设备ID。
2. 重写父类Layer的**init**函数，修改`deviceId_`为`-2`，代表这个layer是用于跑在MKL-DNN的环境下。
3. 创建`MKLDNNMatrix`，同时继承`CpuMatrix`和`mkldnn::memory`。用于管理MKL-DNN会用到的相关memory函数、接口以及会用的到格式信息。
4. 创建`MKLDNNBase`，定义一些除了layer和memory相关的类和函数。包括MKL-DNN会用到`MKLDNNStream`和`CPUEngine`，和未来可能还会用到`FPGAEngine`等。
5. 每个`MKLDNNlayer`都会有`inVal_`,`inGrad_`,`outVal_`和`outGrad_`，分别代表input value， input gradient，output value和output gradient。他们会存放MKL-DNN用到的internal memory。同时还会定义以*ext*开头的`MKLDNNMatrix`(表示external的memory)，主要是在格式与PaddlePaddle默认的`nchw`格式不匹配时，用于转换内存的工作。必要的转换函数也会在`MKLDNNLayer`中提前定义好，每个子类只需要调用定义好的reset buffer函数即可。
6. 每个`MKLDNNlayer`的resetbuffer相关的函数（包括reset input、output的Value和grad），他们会根据输入参数reset internal和external的memory，当然这两者也可以相等，即表示不需要转换。只需要把握一个原则，每个`MKLDNNlayer`的子类，只需要使用internal的memory就可以了，所有external的转换工作在父类的reset函数中都提前准备好了。
7. 一般来说，external的memory会尽量与PaddlePaddle中的`value`和`grad`共享内存。同时每个`MKLDNNLayer`中的external output value和gradient(也就是`extOutVal_`和`extOutGrad_`)必须分别与`output_.value`和`output_.grad`共享内存，因为PaddlePaddle的activation会直接使用`output_.value`和`output_.grad`。如果不需要external的buffer用于转换，那么internal的buffer也会与他们共享内存。
8. 如果MKL-DNN layer的后面接有cpu device，那么就会使`output_.value`与`extOutVal_`共享内存，同时数据格式就是`nchw`，这样下一个cpu device就能拿到正确的数据。在有cpu device的时候，external的memory的格式始终是`nchw`或者`nc`。
9. 由于MKL-DNN的输出操作都是覆盖data的，不是在原来的数据上累加，所以当网络出现分支时，在`backward`时会需要merge不同layer的梯度。`MKLDNNlayer`中会实现merge的方法，此时每个小分支的input gradient会先临时保存在一个`MKLDNNMatrix`中，由分支处的layer负责求和，并把结果放到这个layer的`output_.grad`中。所以整体上，每个子类并不会需要关心分支的事情，也是在父类都实现好了。
10. 在原来的`FLAGS`中添加一个`use_mkldnn`的flag，用于选择是否使用MKL-DNN的相关功能。
## References
1. [Intel Math Kernel Library for Deep Neural Networks (Intel MKL-DNN)](https://github.com/01org/mkl-dnn "Intel MKL-DNN")
2. [原来的方案](https://github.com/PaddlePaddle/Paddle/pull/3096)会引入**nextLayer**的信息。但是在PaddlePaddle中，无论是重构前的layer还是重构后的op，都不会想要知道next layer/op的信息。
3. MKL-DNN的高性能格式与PaddlePaddle原有的`NCHW`不同(PaddlePaddle中的CUDNN部分使用的也是`NCHW`，所以不存在这个问题)，所以需要引入一个转换方法，并且只需要在必要的时候转换这种格式，才能更好的发挥MKL-DNN的性能。
doc/design/operator_kernel_type.md 0 → 100644
浏览文件 @ fab63cc6
# Design Doc: The Keys of Operator Kernel Type
## Problem
An operator can have different kernel implementations, and each operator will have a map to store the related kernels. Fluid uses `OpKernelType` as a key to identify a unique Kernel. Before an operator runs, an certain kernel must be chosen by a key of `OpKernelType`. Currently, `OpKernelType` is defined as follows:
```cpp
struct OpKernelType {
platform::Place place_;
proto::DataType data_type_;
};
```
For more details, please refer to [codes](https://github.com/PaddlePaddle/Paddle/blob/2d5ec16bc8a09fb8e0f62c89b116b0cd1d333907/paddle/framework/operator.h#L348-L374) in github.
It contains two keys, `Place` and `DataType`. And these two keys will be hashed to a unique key to represent a certain type of kernel. However, these two keys are not enough. We need a more complete representation of `OpKernelType`.
We often implement a kernel of an operator with some computing library in certain device(place). Please remind that computing library and device are not one-to-one corresponding. A device can have a lot of computing libraries and a computing library can also support several devices.
For example, Eigen library can support Nvidia GPU/AMD GPU/CPU. And MKLDNN library can support Intel CPU/Intel FPGA. Both `Place` and `Library` should be a key of `OpKernelType`.
It's obvious that different DataTypes, like fp64/fp32/int8 will have different kernels. But the data layout of a Tensor will also lead to different implementation. Please refer to the batch norm operator [kernels](https://github.com/PaddlePaddle/Paddle/blob/a948fac4d0ad7e0412d373b8aabeb711c2899563/paddle/operators/batch_norm_op.cc#L180-L209). Data Layout should also be taken into consideration.
## Solution
There are four keys to determine a kernel type of an operator: `Place`/`Library`/`DataType`/`Layout`.
```cpp
struct OpKernelType {
platform::Place place_;
platform::Library library_;
proto::DataType data_type_;
framework::Layout layout_;
};
```
Following is the details:
### Place
`Place` is defined as follows:
```cpp
typedef boost::variant<CUDAPlace, ROCmPlace, FPGAPlace, CPUPlace> Place;
```
`Place` is to represent the device memory where data is locating.
### Library
One operator kernel is usually implemented based on one library. `Library` is defined as a enum variable:
```cpp
enum Library { Plain, MKLDNN, CUDNN };
```
We use `Plain` enumerator to represent default library. Since most operators in Fluid are implemented based on `Eigen` library, we take `Eigen` library as the `Plain` enumerator.
A library usually has a corresponding `DeviceContext` which contains some handles needed by computation. Fluid now have two default DeviceContexts in CPU and CUDA, `CPUDeviceContext` and `CUDADeviceContext`. `CPUDeviceContext` contains a Eigen library handle and `CDUADeviceContext` contains a Eigen library handle and cuBLAS handle.
If we want to support new Library, a new enumerator need to be added to `Library` and a new corresponding `LibraryDeviceContext` will be created.
### DataType
`DataType` is defined in [framework.proto](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/framework.proto). Currently, int32/int64/fp32/fp64 are supported.
### Layout
Actually, a Tensor is a view of a block of memory. Besides a pointer to the memory, we also have to get some other descriptions of this block of memory, such as shape(ddim), stride, and layout.
Different layout leads to different implementation of operator kernel. There are mainly 4 principles we have to follow to support layout in our fluid framework.
- We take layout as a data member of Tensor. Layout is actually a enum variable. If fluid is built with MKLDNN, then, the memory format in MKLDNN will be added into this enum variable too.
- Users have to set layout for input data. And some operators like fill_constant/random, also have to set layout of generating data. Of course, we can have some default layout, like NCHW.
- The inference of Layout is at run-time, not compile-time.
- Every operator have to implement different kernels for different layouts. Let's take MKLDNN as an example, if we want to implement a MKLDNN convolution operator, we have to realize all the kernels for different layout, list at [here](http://01org.github.io/mkl-dnn/structmkldnn_1_1memory.html). And we will have a special macro to do registering kernels for MKLDNN operators.
`Layout` is also defined as a enum variable:
```cpp
enum Layout {
kNCHW,
kNHWC,
#ifdef PADDLE_WITH_MKLDNN
knChw8c
...
#endif
};
```
doc/design/optimizer.md
浏览文件 @ fab63cc6
......@@ -79,7 +79,7 @@ class Optimizer(object):
def minimize(self, loss, parameter_list):
"""Add operations to minimize `loss` by updating `parameter_list`.
This method combines interface `append_backward_ops()` and
This method combines interface `append_backward()` and
`create_optimization_pass()` into one.
"""
params_grads = self.create_backward_pass(loss, parameter_list)
......
doc/design/paddle_nccl.md 0 → 100644
浏览文件 @ fab63cc6
# Design Doc: NCCL support in Paddle Fluid
## Abstract
This Design Doc refers to the NCCL feature in paddle. We propose an approach to support NCCL library both on a single machine and multiple machines. We wrapper the NCCL primitives `Broadcast`, `Allreduce`, `Reduce` as operators to utilize Multi-GPU powers in one script.
## Motivation
[NCCL](https://developer.nvidia.com/nccl) is a NVIDIA library support Multi-GPU communicating and optimized for NVIDIA GPUs, it provides routines such as all-gather, all-reduce, broadcast, reduce, reduce-scatter, that can achieve high bandwidth over PCIe and NVLink high-speed interconnect. With NCCL library, we can easily accelerate the training in parallel.
- Pros
1. easily plug-in with [NCCL2](https://developer.nvidia.com/nccl) library.
1. high performance in NVIDIA GPUs.
1. MPI like primitives, which have low learning cost for users.
- Cons
1. Only design for NVIDIA GPUs, not a general multi-device solution.
1. Although NCCL1 is opensourced under BSD license, but NCCL2 is not opensourced anymore.
At the beginning of training, the framework needs to distribute the same parameters to every GPU, and merge the gradients at any time user interests.
As a result, during training, we need the operations of peer to peer copy between different GPUs, aggregating gradients/parameters from GPUs, and broadcasting parameters to GPUs. Every GPU only need to run the operator with correct place information.
Besides, it needs interfaces to synchronize model update with each different GPU Cards.
## Implementation
As mentioned above, we wrap the NCCL routines as several kinds of operators. Need to note that NCCL need to create Communicator between gpu at the beginning, so there is a NCCLInit operator created.
### Transpiler
To be compatible with [parameter server design doc](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/ops/dist_train.md), the transpiler compiles the user defined operation graph into sub-graphs to be executed on different devices.
1. The user-defined model will be a single device program
2. Broadcast/Reduce operators between GPUs will be inserted into the program, even for the multi-node, may insert the `Send`, `Recv` operator.
*Broadcast, AllReduce in a single machine. And Broadcast, AllReduce, [Send, Recv](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/ops/dist_train.md#graph-converter) in multiple machines*
<img src="images/multigpu_before_convert.png" width="300"/>
After compiling, the graph as shows
<img src="images/multigpu_allreduce.png" width="1000"/>
Operators are added to the sub-graphs. Every GPU assigned a role of `rank0`, `rank1` etc.
- **Broadcast**. Broadcast operator distribute initialized parameter to all the GPUs from the GPU who owns it. e.g. from`rank0` GPU.
- **AllReduce**. AllReduce operator synchronizes parameters/gradients between GPUs. AllReduce implemented in the Ring-Based communicating method, avoid of the bottle neck in a single GPU.
Need to notice that AllReduce operator force GPUs synchronized at that point. The whole training process in asynchronous or synchronous mode depends on the AllReduce point in the graph.
As it shown in the picture, when each GPU compute the gradient of `W`, followed with a `AllReduce` operator, accumulate the `dW` to full batch of data, then run the optimize process individually and apply the gradient to its `W`.
- **AllReduce**
Need to note that our AllReduce operator is a ring-base AllReduce implementation. If we use the NCCL2 AllReduce primitive, every GPU optimized full batch of data, wasted (n-1) GPU compute resources. In addition, NCCL2 built-in AllReduce will only utilize the communicating resource during synchronization, then update the gradient will be a subsequent phase. In fact, we can amortize the update gradient time cost into the communicating phase. The process is
1. Every parameter has its root card. That card will responsible for aggregating the gradients from GPUs.
2. The whole model's parameter will be hashed to different root card, ensure the load balance between GPUs.
3. Logically neighberhood card will start send parameter to the next one. After one round, the parameter main card will aggregate the full gradients.
4. Then the root card will optimize the parameter.
5. This parameter card will send its optimized result to its neighberhood, then the neighberhood will send parameter to its next one.
6. Finish the sychronization round.
The total time cost will be 2 * (n-1) * per-parameter-send-time, we reach the goal of amortize the upgrade time into communicating phase.
doc/design/reader/README.md
浏览文件 @ fab63cc6
# Python Data Reader Design Doc
At training and testing time, PaddlePaddle programs need to read data. To ease the users' work to write data reading code, we define that
During the training and testing phases, PaddlePaddle programs need to read data. To help the users write code that performs reading input data, we define the following:
- A *reader* is a function that reads data (from file, network, random number generator, etc) and yields data items.
- A *reader creator* is a function that returns a reader function.
- A *reader decorator* is a function, which accepts one or more readers, and returns a reader.
- A *batch reader* is a function that reads data (from *reader*, file, network, random number generator, etc) and yields a batch of data items.
- A *reader*: A function that reads data (from file, network, random number generator, etc) and yields the data items.
- A *reader creator*: A function that returns a reader function.
- A *reader decorator*: A function, which takes in one or more readers, and returns a reader.
- A *batch reader*: A function that reads data (from *reader*, file, network, random number generator, etc) and yields a batch of data items.
and provide function which converts reader to batch reader, frequently used reader creators and reader decorators.
and also provide a function which can convert a reader to a batch reader, frequently used reader creators and reader decorators.
## Data Reader Interface
Indeed, *data reader* doesn't have to be a function that reads and yields data items. It can be any function with no parameter that creates a iterable (anything can be used in `for x in iterable`):
*Data reader* doesn't have to be a function that reads and yields data items. It can just be any function without any parameters that creates an iterable (anything can be used in `for x in iterable`) as follows:
```
iterable = data_reader()
```
Element produced from the iterable should be a **single** entry of data, **not** a mini batch. That entry of data could be a single item, or a tuple of items. Item should be of [supported type](http://www.paddlepaddle.org/doc/ui/data_provider/pydataprovider2.html?highlight=dense_vector#input-types) (e.g., numpy 1d array of float32, int, list of int)
The item produced from the iterable should be a **single** entry of data and **not** a mini batch. The entry of data could be a single item or a tuple of items. Item should be of one of the [supported types](http://www.paddlepaddle.org/doc/ui/data_provider/pydataprovider2.html?highlight=dense_vector#input-types) (e.g., numpy 1d array of float32, int, list of int etc.)
An example implementation for single item data reader creator:
An example implementation for single item data reader creator is as follows:
```python
def reader_creator_random_image(width, height):
......@@ -29,7 +29,7 @@ def reader_creator_random_image(width, height):
return reader
```
An example implementation for multiple item data reader creator:
An example implementation for multiple item data reader creator is as follows:
```python
def reader_creator_random_image_and_label(width, height, label):
def reader():
......@@ -40,9 +40,10 @@ def reader_creator_random_image_and_label(width, height, label):
## Batch Reader Interface
*batch reader* can be any function with no parameter that creates a iterable (anything can be used in `for x in iterable`). The output of the iterable should be a batch (list) of data items. Each item inside the list must be a tuple.
*Batch reader* can be any function without any parameters that creates an iterable (anything can be used in `for x in iterable`). The output of the iterable should be a batch (list) of data items. Each item inside the list should be a tuple.
Here are some valid outputs:
Here are valid outputs:
```python
# a mini batch of three data items. Each data item consist three columns of data, each of which is 1.
[(1, 1, 1),
......@@ -58,20 +59,22 @@ Here are valid outputs:
Please note that each item inside the list must be a tuple, below is an invalid output:
```python
# wrong, [1,1,1] needs to be inside a tuple: ([1,1,1],).
# Otherwise it's ambiguous whether [1,1,1] means a single column of data [1, 1, 1],
# or three column of datas, each of which is 1.
# Otherwise it is ambiguous whether [1,1,1] means a single column of data [1, 1, 1],
# or three columns of data, each of which is 1.
[[1,1,1],
[2,2,2],
[3,3,3]]
```
It's easy to convert from reader to batch reader:
It is easy to convert from a reader to a batch reader:
```python
mnist_train = paddle.dataset.mnist.train()
mnist_train_batch_reader = paddle.batch(mnist_train, 128)
```
Also easy to create custom batch reader:
It is also straight forward to create a custom batch reader:
```python
def custom_batch_reader():
while True:
......@@ -85,7 +88,8 @@ mnist_random_image_batch_reader = custom_batch_reader
## Usage
batch reader, mapping from item(s) read to data layer, batch size and number of total pass will be passed into `paddle.train`:
Following is how we can use the reader with PaddlePaddle:
The batch reader, a mapping from item(s) to data layer, the batch size and the number of total passes will be passed into `paddle.train` as follows:
```python
# two data layer is created:
......@@ -99,13 +103,13 @@ paddle.train(batch_reader, {"image":0, "label":1}, 128, 10, ...)
## Data Reader Decorator
*Data reader decorator* takes a single or multiple data reader, returns a new data reader. It is similar to a [python decorator](https://wiki.python.org/moin/PythonDecorators), but it does not use `@` syntax.
The *Data reader decorator* takes in a single reader or multiple data readers and returns a new data reader. It is similar to a [python decorator](https://wiki.python.org/moin/PythonDecorators), but it does not use `@` in the syntax.
Since we have a strict interface for data readers (no parameter, return a single data item). Data reader can be used flexiable via data reader decorators. Following are a few examples:
Since we have a strict interface for data readers (no parameters and return a single data item), a data reader can be used in a flexible way using data reader decorators. Following are a few examples:
### Prefetch Data
Since reading data may take time and training can not proceed without data. It is generally a good idea to prefetch data.
Since reading data may take some time and training can not proceed without data, it is generally a good idea to prefetch the data.
Use `paddle.reader.buffered` to prefetch data:
......@@ -117,9 +121,9 @@ buffered_reader = paddle.reader.buffered(paddle.dataset.mnist.train(), 100)
### Compose Multiple Data Readers
For example, we want to use a source of real images (reusing mnist dataset), and a source of random images as input for [Generative Adversarial Networks](https://arxiv.org/abs/1406.2661).
For example, if we want to use a source of real images (say reusing mnist dataset), and a source of random images as input for [Generative Adversarial Networks](https://arxiv.org/abs/1406.2661).
We can do:
We can do the following :
```python
def reader_creator_random_image(width, height):
......@@ -139,13 +143,13 @@ false_reader = reader_creator_bool(False)
reader = paddle.reader.compose(paddle.dataset.mnist.train(), data_reader_creator_random_image(20, 20), true_reader, false_reader)
# Skipped 1 because paddle.dataset.mnist.train() produces two items per data entry.
# And we don't care second item at this time.
# And we don't care about the second item at this time.
paddle.train(paddle.batch(reader, 128), {"true_image":0, "fake_image": 2, "true_label": 3, "false_label": 4}, ...)
```
### Shuffle
Given shuffle buffer size `n`, `paddle.reader.shuffle` will return a data reader that buffers `n` data entries and shuffle them before a data entry is read.
Given the shuffle buffer size `n`, `paddle.reader.shuffle` returns a data reader that buffers `n` data entries and shuffles them before a data entry is read.
Example:
```python
......@@ -154,21 +158,21 @@ reader = paddle.reader.shuffle(paddle.dataset.mnist.train(), 512)
## Q & A
### Why reader return only a single entry, but not a mini batch?
### Why does a reader return only a single entry, and not a mini batch?
Always returning a single entry make reusing existing data readers much easier (e.g., if existing reader return not a single entry but 3 entries, training code will be more complex because it need to handle cases like batch size 2).
Returning a single entry makes reusing existing data readers much easier (for example, if an existing reader returns 3 entries instead if a single entry, the training code will be more complicated because it need to handle cases like a batch size 2).
We provide function `paddle.batch` to turn (single entry) reader into batch reader.
We provide a function: `paddle.batch` to turn (a single entry) reader into a batch reader.
### Why do we need batch reader, isn't train take reader and batch_size as arguments sufficient?
### Why do we need a batch reader, isn't is sufficient to give the reader and batch_size as arguments during training ?
In most of the case, train taking reader and batch_size as arguments would be sufficent. However sometimes user want to customize order of data entries inside a mini batch. Or even change batch size dynamically.
In most of the cases, it would be sufficient to give the reader and batch_size as arguments to the train method. However sometimes the user wants to customize the order of data entries inside a mini batch, or even change the batch size dynamically. For these cases using a batch reader is very efficient and helpful.
### Why use a dictionary but not a list to provide mapping?
### Why use a dictionary instead of a list to provide mapping?
We decided to use dictionary (`{"image":0, "label":1}`) instead of list (`["image", "label"]`) is because that user can easily resue item (e.g., using `{"image_a":0, "image_b":0, "label":1}`) or skip item (e.g., using `{"image_a":0, "label":2}`).
Using a dictionary (`{"image":0, "label":1}`) instead of a list (`["image", "label"]`) gives the advantage that the user can easily reuse the items (e.g., using `{"image_a":0, "image_b":0, "label":1}`) or even skip an item (e.g., using `{"image_a":0, "label":2}`).
### How to create custom data reader creator
### How to create a custom data reader creator ?
```python
def image_reader_creator(image_path, label_path, n):
......@@ -192,7 +196,7 @@ paddle.train(paddle.batch(reader, 128), {"image":0, "label":1}, ...)
### How is `paddle.train` implemented
An example implementation of paddle.train could be:
An example implementation of paddle.train is:
```python
def train(batch_reader, mapping, batch_size, total_pass):
......
doc/design/refactor/distributed_architecture.md
浏览文件 @ fab63cc6
......@@ -2,148 +2,131 @@
## Abstract
PaddlePaddle v0.10.0 uses the "trainer-parameter server"
architecture. We run multiple replicated instances of trainers (runs
the same code written by the user) and parameter servers for
distributed training. This architecture served us well, but has some
limitations:
PaddlePaddle version 0.10.0 uses the "trainer-parameter server" architecture. We run multiple instances of trainers (where each trainer runs the same model) and parameter servers for distributed training. This architecture serves well, but has few limitations:
1. Need to write special code to handle tasks which should only be run
by a single trainer. E.g., initializing model and saving model.
1. There is a need to write special code that handles tasks which should only be run on a single trainer. E.g., initializing the model, saving the model etc.
2. Model parallelism is hard: need to write if-else branches conditioned
on the trainer ID to partition model onto each trainer, and manually
write the inter-model-shard communication code.
2. Model parallelism is hard: It would need all the if-else branches conditioned on the trainer ID to partition the model onto the trainers, and eventually manually writing out the inter-model-shard communication code to communicate between different trainers.
3. The user can not directly specify the parameter update rule: need
to modify the parameter server C++ code and compile a new binary.
This adds complication for researchers: A lot of extra effort is
required. Besides, the training job submission program
may not allow running arbitrary binaries.
3. The user can not directly specify the parameter update rule: This would need to modify the parameter server code and compile a new binary. This makes things more complicated for researchers: A lot of extra effort is required to make this work. Besides, the training job submission program may not allow running arbitrary binaries.
This design doc discusses PaddlePaddle's new distributed training
architecture that addresses the above limitations.
This design doc discusses PaddlePaddle's new distributed training architecture that addresses the above mentioned limitations.
## Analysis
We will assume the user writes the trainer program by Python, the same
analysis holds if the trainer program is written in C++.
The assumption is that the user writes the trainer program in either Python or C++.
### Limitation 1
If we look at the Python code that the user writes, there are two
kinds of functionalities:
There are two basic functionalities in the trainer program:
- The training logic such as load / save model and print log.
- The neural network definition such as the definition of the data
layer, the fully connected layer, the cost function and the
1. The training logic such as loading / saving the model and printing out the logs.
2. The neural network definition such as the definition of the data layer, the fully connected layer, the cost function and the
optimizer.
When we training with PaddlePaddle v0.10.0 distributedly, multiple
replicated Python instances are running on different nodes: both the
training logic and the neural network computation is replicated.
When we train using PaddlePaddle v0.10.0 in a distributed fashion, multiple instances of the same Python code are run on different nodes, hence both: the
training logic as well as the neural network computation logic, is replicated.
The tasks that should only run once all belong to the training logic,
if we only replicate the neural network computation but do **not**
replicate the training logic, the limitation could be solved.
The tasks that only need to be run once belong to the training logic. Hence if we only replicate the neural network computation part, and do **not**
replicate the training logic, the limitation mentioned above can be avoided.
### Limitation 2
Model parallelism means running a single model on multiple nodes by
partitioning the model onto different nodes and managing the
inter-model-shard communications.
Model parallelism means that a single model is partitioned into different components and each node runs one of the component separately. This comes at the extra cost of managing the
inter-model-shard communication between nodes.
PaddlePaddle should be able to modify the neural network computation
definition to support model parallelism automatically. However, the
computation is only specified in Python code, and PaddlePaddle cannot
modify Python code.
PaddlePaddle should ideally be able to modify the neural network computation and figure out the support for model parallelism automatically. However, the
computation is only specified in Python code which sits outside of PaddlePaddle, hence PaddlePaddle can not support the feature in this setup.
Just like compiler uses an intermediate representation (IR) so that
the programmer does not need to manually optimize their code in most of
the cases - the compiler will optimize the IR:
Similar to how a compiler uses an intermediate representation (IR) so that the programmer does not need to manually optimize their code for most of the cases, we can have an intermediate representation in PaddlePaddle as well. The compiler optimizes the IR as follows:
<img src="src/compiler.png"/>
We can have our own IR which is called [Program](../program.md).
PaddlePaddle can support model parallel by
converting the IR so the user no longer need to manually do it in
Python:
PaddlePaddle can support model parallelism by converting the IR so that the user no longer needs to manually perform the computation and operations in the Python component:
<img src="src/paddle-compile.png"/>
The IR for PaddlePaddle after refactoring is called a `Block`, it specifies the computation dependency graph and the variables used in the computation.
### Limitation 3
The user can not directly specify the parameter update rule for the
parameter server because the previous implementation hard coded that
parameter server only do vector's optimization algorithm by
configuration. The user can not specify the parameter server's
computation layer by layer.
The user can not directly specify the parameter update rule for the parameter server in the Python module, since the parameter server does not use the same computation definition as the trainer. Instead, the update rule is baked inside the parameter server. The user can not specify the update rule explicitly.
This could be fixed by making the parameter server run a separated
IR according to the trainer's variable (tensors, selectedrows)
definition.
the same
computation definition of the trainer. For a detailed explanation,
please
see
[Design Doc: Operation Graph-Based Parameter Server](./parameter_server.md)
This could be fixed by making the parameter server run the same computation definition as the trainer (the user's Python module). For a detailed explanation, refer to this document -
[Design Doc: Operation Graph Based Parameter Server](./parameter_server.md)
## Distributed Training Architecture
The new distributed training architecture can address the above
limitations. Below is the illustration:
The revamped distributed training architecture can address the above discussed limitations. Below is the illustration of how it does so:
<img src="src/distributed_architecture.png"/>
The architecture includes major components: *PaddlePaddle Python*,
*PaddlePaddle converter* and *PaddlePaddle runtime*:
The major components are: *Python API*, *Distribute Transpiler* and *Remote Executor*.
### PaddlePaddle Python
### Python API
PaddlePaddle Python is the Python library that user's Python trainer
invoke to build the neural network topology, start training, etc.
Python API is the Python library that user's Python code invokes, to read the data, build the neural network topology, and start training, etc.
```Python
paddle.init()
input = paddle.op.recordIO("/home/data/mnist.recordio") # file stored on the cluster
img, label = input[0], input[1]
hidden = paddle.layer.fc(input=img, size=200, act=paddle.activation.Tanh())
prediction = paddle.layer.fc(input=img, size=10, act=paddle.activation.Softmax())
cost = paddle.layer.classification_cost(input=prediction, label=label)
optimizer = paddle.optimizer.SGD(learning_rate=0.01)
opts = optimizer.minimize(cost)
exe = RemoteExecutor(num_trainer=3, num_ps=2, GPU_per_trainer=2, sync_batches=1)
# this will init variable data on both server and trainer
exe.run(framework.default_startup_program())
exe.sync()
for i in range(1000):
# feed data
...
cost, acc = exe.run(framework.default_main_program(),
fetch_list=[avg_cost, acc_out])
print cost, acc
images = fluid.layers.data(name='pixel', shape=[1, 28, 28], dtype='float32')
label = fluid.layers.data(name='label', shape=[1], dtype='int64')
...
predict = fluid.layers.fc(input=conv_pool_2, size=10, act="softmax")
cost = fluid.layers.cross_entropy(input=predict, label=label)
avg_cost = fluid.layers.mean(x=cost)
optimizer = fluid.optimizer.Adam(learning_rate=0.01)
optimizer.minimize(avg_cost)
train_reader = paddle.batch(
paddle.reader.shuffle(
paddle.dataset.mnist.train(), buf_size=500),
batch_size=BATCH_SIZE)
place = fluid.CPUPlace()
exe = fluid.Executor(place)
for pass_id in range(10):
for data in train_reader():
loss, acc = exe.run(trainer_prog,
feed=feeder.feed(data),
fetch_list=[avg_cost])
```
The code above is a typical Python trainer code, the neural network
topology is built using helper functions such as
`paddle.layer.fc`. The training is done by calling `Executor.run`
The code above is a typical local training program, the "Training Program" is built using helper functions such as
`fluid.layer.fc`. The training is done by calling `Executor.run`
iteratively.
#### RemoteExecutor
For more details, the implementation of IR is [Program](../program.md), and `ProgramDesc` is the protobuf type.
[Executor](../executor.md) simply runs the `ProgramDesc`. For local training you generally use
`Executor` to run the program locally. For any kind of distributed training, you can use
`RemoteExecutor` to specify desired distributed training method with some optional arguments.
### Distributed Transpiler
The Distributed Transpiler automatically converts the IR (in protobuf format) to partitioned IRs. Then
the Remote Executor dispatches the new IRs to Remote Executors across the cluster.
Below are the steps that are followed :
1. User only need to change `Executor` to `RemoteExecutor` to change local program to distributed program.
1. `RemoteExecutor` calls `Distributed Transpiler` to "transpile" user's program to several IRs representing a
distributed training program:
1. Parse configurations from `RemoteExecutor`.
1. Determine the type of distributed program, can be DataParallelism, ModelParallelism or Streaming.
1. Partition the `ProgramDesc` according to type and add `send` / `recv` OP pair on the boundaries. For
DataParallelism type for example, it removes the optimization operators and add a `send` OP to the
"trainer" role, then add the optimization operators to the parameter server role within the `recv` OP.
1. Dispatch the partitioned graph to different `RemoteExecutor` in the cluster.
1. `RemoteExecutor` on each node run the received `ProgramDesc` utill the end.
As shown in the graph, `RemoteExecutor.run` sends the IR to the
PaddlePaddle cluster for Execution. You can also use parameter
`fetch_list` to interactively fetch variable back to local for
### RemoteExecutor
As shown in the graph, `RemoteExecutor.run` sends the IR to the cluster for Execution.
You can also use parameter `fetch_list` to interactively fetch variable back to local for
log printing.
The Python `RemoteExecutor` is derived from `Executor` class.
For more information about `RemoteExecutor`, please
see [Design Doc: RemoteExecutor](./remote_executor.md).
The `RemoteExecutor.run` interface defination is:
```python
run(self,
......@@ -167,87 +150,24 @@ run(self,
`JobDesc` object describe the distributed job resource specification to run on
Cluster environment.
By default, `Executor.run` starts a PaddlePaddle Cloud
[TrainingJob](https://github.com/PaddlePaddle/cloud/blob/develop/doc/autoscale/README.md#training-job-resource),
or you can run each component in the
executor by your own method:
- Data Parallelism
```python
if os.getenv('PLACE_PSERVER'):
exe.run_pserver()
elif os.getenv('PLACE_TRAINER'):
exe.run_trainer()
```
- Model Parrallelism
```python
for part in exe.get_parralle_parts():
exe.run_part(part)
```
#### Program and Executor
As mentioned above, the implementation of IR is [Program](../program.md).
[Executor](../executor.md) converts and parses the IR to a preferred
graph for final execution. For local training you generally use
`Executor` to run the graph locally. For any kind of distributed
training, you can use `RemoteExecutor` to specify desired distributed
training method with some optional arguments.
### PaddlePaddle Converter
PaddlePaddle converter automatically converts the IR in the request
(IR and evaluation inputs/targets) from PaddlePaddle Python to new
partitioned IRs and dispatch the new IRs and evaluation inputs/targets
to different PaddlePaddle runtimes. Below are the steps:
1. Add `feed` OP that feeds the eval inputs, and `fetch` OP that
fetches the eval targets to the IR.
1. Extract a new computation (sub)graph with `feed` and `fetch` OP as
the boundary. The runtime does not need to run the OP that is not
dependent on the `fetch` OP.
1. Optimizes the computation graph.
1. Place the OPs in the graph onto different devices on different
PaddlePaddle runtime according to a placement algorithm and device
constraint specified by the user.
1. Partition the graph according to runtime boundaries and add `send` /
`recv` OP pair on the runtime boundaries.
1. Dispatch the partitioned graph to different PaddlePaddle runtimes.
<img src="src/remote_executor.png"/>
1. PaddlePaddle runtimes with the `fetch` OP reports evaluation
results back to the converter, the convert reports the evaluation
results back to the PaddlePaddle Python.
The output IRs will be cached to optimize the conversion latency.
`RemoteExecutor.run` sends the `ProgramDesc` and
[TrainingJob](https://github.com/PaddlePaddle/cloud/blob/develop/doc/autoscale/README.md#training-job-resource)
to a server in the cluster which executes `RemoteExecutor.listen`. This server is responsible
to start the final Kubernetes Jobs to run the different role of `ProgramDesc`.
#### Placement Algorithm
### Placement Algorithm
Our first implementation will only support "trainer-parameter server"
placement: the parameters, initializers, and optimizers are placed on
the PaddlePaddle runtimes with the parameter server role. And
everything else will be placed on the PaddlePaddle runtimes with the
trainer role. This has the same functionality of our
"trainer-parameter server" architecture of PaddlePaddle v0.10.0, but
is more general and flexible.
Our first implementation will only support "trainer-parameter server" placement: the parameters, initializers, and optimizers are all placed on the PaddlePaddle runtimes with the parameter server role. Everything else will be placed on the PaddlePaddle runtimes with the trainer role. This has the same functionality as the "trainer-parameter server" architecture of PaddlePaddle v0.10.0, but is more generic and flexible.
In the future, we will implement the general placement algorithm,
which makes placements according to the input IR, and a model of
device computation time and device communication time. Model
parallelism requires the general placement algorithm.
In the future, a more general placement algorithm should be implemented, which makes placements according to the input IR, and a model of device computation time and device communication time. Model parallelism requires the generic placement algorithm.
### Local Training Architecture
The local training architecture will be the same as the distributed
training architecture, the differences are everything runs locally,
and there is just one PaddlePaddle runtime:
The local training architecture will be the same as the distributed training architecture, the difference is that everything runs locally, and there is just one PaddlePaddle runtime:
<img src="src/local_architecture.png"/>
......
doc/design/refactor/multi_cpu.md 0 → 100644
浏览文件 @ fab63cc6
# Design Doc: Execute the Program with Multi CPU
## Abstract
This Design Doc propose an approach to make the user-defined Op graph
running with multi-CPU, we will use an auto transpiler to convert the user-defined
Op graph to a multi-CPU Op graph, and run `ParallelDo` Op to run the graph.
## Transpiler
<img src="src/multi-threads/single-thread@3x.png" width="300">
After converted:
<img src="src/multi-threads/multi-threads@3x.png" width="1000">
## Implement
- `Multi-CPU Transpiler` will convert the graph to a multi-CPU graph
which would be executed with multi-threads.
- `BlockingCounter` will `Init/Decrement` an atomic counter, and Blocking `Wait`
for the atomic counter become `0`:
```cpp
BlockingCounter bc(thread_count);
for (int i = 0; i < thread_count; ++i) {
thread_pool->Start([&bc] {bc.DecrementCount(); })
}
bc.Wait();
```
- `ParallelDo` Operator
- Initialize a thread pool which is a Singleton.
- Use a block id as the input, and create run the specify Block on independent scope
with multi-threads.
- Initialize a `BlockingCounter` instance and wait until all threads are done.
- `Split` Operator will split the Input Tensor into a TensorArray.
- `Merge` merge all the gradients which calculated in different threads
with `mean/sum/max/min...` method, and then run the Optimizer Op to optimize `W`.
## TODO
- Improve the optimizer stage with multi-threads, since we could
assign the parameters to the different threads and execute
optimizer with multi-threads.
doc/design/releasing_process.md
浏览文件 @ fab63cc6
......@@ -5,8 +5,9 @@ PaddlePaddle使用git-flow branching model做分支管理，使用[Semantic Vers
PaddlePaddle每次发新的版本，遵循以下流程:
1. 从`develop`分支派生出新的分支，分支名为`release/版本号`。例如，`release/0.10.0`
2. 将新分支的版本打上tag，tag为`版本号rc.Patch号`。第一个tag为`0.10.0rc1`，第二个为`0.10.0rc2`，依次类推。
3. 对这个版本的提交，做如下几个操作:
1. 将新分支的版本打上tag，tag为`版本号rc.Patch号`。第一个tag为`0.10.0rc1`，第二个为`0.10.0rc2`，依次类推。
1. 对这个版本的提交，做如下几个操作:
* 修改`python/setup.py.in`中的版本信息,并将`istaged`字段设为`True`。
* 编译这个版本的Docker发行镜像，发布到dockerhub。如果失败，修复Docker编译镜像问题，Patch号加一，返回第二步
* 编译这个版本的Ubuntu Deb包。如果失败，修复Ubuntu Deb包编译问题，Patch号加一，返回第二步。
* 使用Regression Test List作为检查列表，测试Docker镜像/ubuntu安装包的功能正确性
......@@ -20,9 +21,9 @@ PaddlePaddle每次发新的版本，遵循以下流程:
pip install twine
twine upload dist/[package to upload]
```
4. 第三步完成后，将`release/版本号`分支合入master分支，并删除`release/版本号`分支。将master分支的合入commit打上tag，tag为`版本号`。同时再将`master`分支合入`develop`分支。最后删除`release/版本号`分支。
5. 编译master分支的Docker发行镜像，发布到dockerhub。编译ubuntu的deb包，发布到github release页面
6. 协同完成Release Note的书写
1. 第三步完成后，将`release/版本号`分支合入master分支，并删除`release/版本号`分支。将master分支的合入commit打上tag，tag为`版本号`。同时再将`master`分支合入`develop`分支。最后删除`release/版本号`分支。
1. 编译master分支的Docker发行镜像，发布到dockerhub。编译ubuntu的deb包，发布到github release页面
1. 协同完成Release Note的书写
需要注意的是:
......@@ -30,7 +31,7 @@ PaddlePaddle每次发新的版本，遵循以下流程:
* `release/版本号`分支一旦建立，一般不允许再从`develop`分支合入`release/版本号`。这样保证`release/版本号`分支功能的封闭，方便测试人员测试PaddlePaddle的行为。
* 在`release/版本号`分支存在的时候，如果有bugfix的行为，需要将bugfix的分支同时merge到`master`, `develop`和`release/版本号`这三个分支。
# PaddlePaddle 分支规范
## PaddlePaddle 分支规范
PaddlePaddle开发过程使用[git-flow](http://nvie.com/posts/a-successful-git-branching-model/)分支规范，并适应github的特性做了一些区别。
......@@ -47,11 +48,11 @@ PaddlePaddle开发过程使用[git-flow](http://nvie.com/posts/a-successful-git-
* BugFix分支也是在开发者自己的fork版本库维护，与功能分支不同的是，BugFix分支需要分别给主版本库的`master`、`develop`与可能有的`release/版本号`分支，同时提起`Pull Request`。
# PaddlePaddle回归测试列表
## PaddlePaddle回归测试列表
本列表说明PaddlePaddle发版之前需要测试的功能点。
## PaddlePaddle Book中所有章节
### PaddlePaddle Book中所有章节
PaddlePaddle每次发版本首先要保证PaddlePaddle Book中所有章节功能的正确性。功能的正确性包括验证PaddlePaddle目前的`paddle_trainer`训练和纯使用`Python`训练模型正确性。
......
doc/design/support_new_device.md 0 → 100644
浏览文件 @ fab63cc6
# Design Doc: Supporting new Device/Library
## Background
Deep learning has a high demand for computing resources. New high-performance devices and computing libraries are appearing very frequently. Deep learning frameworks have to integrate these high-performance devices and computing libraries flexibly and efficiently.
On one hand, hardware and computing libraries usually do not have a one-to-one correspondence. For example,Intel CPUs support Eigen and MKL computing libraries while Nvidia GPUs support Eigen and cuDNN computing libraries. We have to implement operator specific kernels for each computing library.
On the other hand, users usually do not want to care about the low-level hardware and computing libraries when writing a neural network configuration. In Fluid, `Layer` is exposed in `Python`, and `Operator` is exposed in `C++`. Both `Layer` and `Operator` are hardware independent.
So, how to support a new Device/Library in Fluid becomes a challenge.
## Basic: Integrate A New Device/Library
For a general overview of fluid, please refer to the [overview doc](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/read_source.md).
There are mainly three parts that we have to consider while integrating a new device/library:
- Place and DeviceContext: indicates the device id and manages hardware resources
- Memory and Tensor: malloc/free data on certain device
- Math Functor and OpKernel: implement computing unit on certain devices/libraries
### Place and DeviceContext
Please remind that device and computing library are not one-to-one corresponding. A device can have a lot of computing libraries and a computing library can also support several devices.
#### Place
Fluid uses class [Place](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/platform/place.h#L55) to represent the device memory where data is located. If we add another device, we have to add corresponding `DevicePlace`.
```
| CPUPlace
Place --| CUDAPlace
| FPGAPlace
```
And `Place` is defined as follows:
```
typedef boost::variant<CUDAPlace, CPUPlace, FPGAPlace> Place;
```
#### DeviceContext
Fluid uses class [DeviceContext](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/platform/device_context.h#L30) to manage the resources in different libraries, such as CUDA stream in `CDUADeviceContext`. There are also inheritance relationships between different kinds of `DeviceContext`.
```
/-> CPUDeviceContext --> MKLDeviceContext
DeviceContext ----> CUDADeviceContext --> CUDNNDeviceContext
\-> FPGADeviceContext
```
An example of Nvidia GPU is as follows:
- DeviceContext
```
class DeviceContext {
virtual Place GetPlace() const = 0;
};
```
- CUDADeviceContext
```
class CUDADeviceContext : public DeviceContext {
Place GetPlace() const override { return place_; }
private:
CUDAPlace place_;
cudaStream_t stream_;
cublasHandle_t cublas_handle_;
std::unique_ptr<Eigen::GpuDevice> eigen_device_; // binds with stream_
};
```
- CUDNNDeviceContext
```
class CUDNNDeviceContext : public CUDADeviceContext {
private:
cudnnHandle_t cudnn_handle_;
};
```
### Memory and Tensor
#### memory module
Fluid provides the following [memory interfaces](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/memory/memory.h#L36):
```
template <typename Place>
void* Alloc(Place place, size_t size);
template <typename Place>
void Free(Place place, void* ptr);
template <typename Place>
size_t Used(Place place);
```
To implement these interfaces, we have to implement MemoryAllocator for different Devices.
#### Tensor
[Tensor](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/tensor.h#L36) holds data with some shape in a specific Place.
```cpp
class Tensor {
public:
/*! Return a pointer to mutable memory block. */
template <typename T>
inline T* data();
/**
* @brief Return a pointer to mutable memory block.
* @note If not exist, then allocation.
*/
template <typename T>
inline T* mutable_data(platform::Place place);
/**
* @brief Return a pointer to mutable memory block.
*
* @param[in] dims The dimensions of the memory block.
* @param[in] place The place of the memory block.
*
* @note If not exist, then allocation.
*/
template <typename T>
inline T* mutable_data(DDim dims, platform::Place place);
/*! Resize the dimensions of the memory block. */
inline Tensor& Resize(const DDim& dims);
/*! Return the dimensions of the memory block. */
inline const DDim& dims() const;
private:
/*! holds the memory block if allocated. */
std::shared_ptr<Placeholder> holder_;
/*! points to dimensions of memory block. */
DDim dim_;
};
```
`Placeholder` is used to delay memory allocation; that is, we can first define a tensor, using `Resize` to configure its shape, and then call `mutuable_data` to allocate the actual memory.
```cpp
paddle::framework::Tensor t;
paddle::platform::CPUPlace place;
// set size first
t.Resize({2, 3});
// allocate memory on CPU later
t.mutable_data(place);
```
### Math Functor and OpKernel
Fluid implements computing units based on different DeviceContexts. Some computing units are shared between operators. This common part will be put in operators/math directory as basic Functors.
Let's take [MaxOutFunctor](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/math/maxouting.h#L27) as an example:
The interface is defined in header file.
```
template <typename DeviceContext, typename T>
class MaxOutFunctor {
public:
void operator()(const DeviceContext& context, const framework::Tensor& input,
framework::Tensor* output, int groups);
};
```
CPU implemention is in .cc file
```
template <typename T>
class MaxOutFunctor<platform::CPUDeviceContext, T> {
public:
void operator()(const platform::CPUDeviceContext& context,
const framework::Tensor& input, framework::Tensor* output,
int groups) {
...
}
};
```
CUDA implemention is in .cu file
```
template <typename T>
class MaxOutFunctor<platform::CUDADeviceContext, T> {
public:
void operator()(const platform::CUDADeviceContext& context,
const framework::Tensor& input, framework::Tensor* output,
int groups) {
...
}
};
```
We get computing handle from a concrete DeviceContext, and make compution on tensors.
The implemention of `OpKernel` is similar to math functors, the extra thing we need to do is to register the OpKernel in a global map.
Fluid provides different register interfaces in op_registry.h
Let's take [Crop](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/crop_op.cc#L134) operator as an example:
In .cc file:
```
REGISTER_OP_CPU_KERNEL(crop, ops::CropKernel<float>);
REGISTER_OP_CPU_KERNEL(
crop_grad, ops::CropGradKernel<paddle::platform::CPUDeviceContext, float>);
```
In .cu file:
```
REGISTER_OP_CUDA_KERNEL(crop, ops::CropKernel<float>);
REGISTER_OP_CUDA_KERNEL(
crop_grad, ops::CropGradKernel<paddle::platform::CUDADeviceContext, float>);
```
## Advanced topics: How to switch between different Device/Library
Generally, we will impelement OpKernel for all Device/Library of an Operator. We can easily train a Convolutional Neural Network in GPU. However, some OpKernel is not sutibale on a specific Device. For example, crf operator can only run on CPU, whereas most other operators can run at GPU. To achieve high performance in such circumstance, we have to switch between different Device/Library.
For more details, please refer to following docs:
- operator kernel type [doc](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/operator_kernel_type.md)
- switch kernel [doc](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/switch_kernel.md)
doc/design/switch_kernel.md 0 → 100644
浏览文件 @ fab63cc6
## Background
Every operator has many kernels because there are multiple data types, places, data layout that Fluid supports. We use the `KernelType` to describe kernel types that operators can hold.
The `KernelType` is as follows.
```
struct KernelType {
Place place_;
DataType data_type_;
LayoutType layout_;
};
```
The `place_` is a descriptor of the device and the computational library, e.g., `MKLDNNPlace`, `CUDAPlace`.
The `data_type_` is the data type that this kernel performs on, e.g., `FP32`, `INT64`. Note that one kernel may have inputs with different data types. However, it will be a major `data_type`. For example, the `cross_entropy` takes `int64` as it label, and `double`/`float` as its input logit and output cost. The major `data_type` of `cross_entropy` is `float`/`double`.
The `layout` is useful for some computational library. One example is that MKLDNN uses many kinds of layout, such as `nChw8c`. Each kind of layout will invoke the different kernel.
## Problem
We register a kernel for every operator and every kernel type ideally. However, it is impracticable for the following situations.
1. Some operators, like CRF, are complicated and inefficient to be implemented on GPU. The CRF operator will only have a CPU kernel.
2. Some operators will take too many memory. It is better to force them into CPU. However, the rest of operators in this neural network will be performed on GPU, i.e., model parallel problem.
3. Some layout and place are particular. One example is that MKLDNN uses `nChw8` and there is no other library uses `nChw8c`.
Problems under these situations are similar. We can formalise this problem as follow.
We register kernels with types $KT = \{kt_1, kt_2, kt_3, ...\}$ for one operator. The inputs of this operator should be run on kernel type $kt_{?}$, which the $kt_{?} \notin KT$. How to cast the input of this operator from $kt_{?}$ to any of kernel type in $KT$.
## Solution
It is clearly that transforming inputs of an operator toadapt another kernel type is not related to the particular operator. So we should register these transformation methods as global methods.
We can infer a kernel type from the inputs of an operators. We let this kernel type as `actual kernel type`, which means this kernel type is the actually kernel type that operator should be performed.
We can get a kernel type by 1) The configuration of operator description. (Users may want to force use `MKL` for `conv` operator). 2) The place of the current executor. (Executor is running on GPU). This kernel type is what we expect the operator will be performed on. We let this kernel type as `expect kernel type`.
We transform the input data from `actual` to `expect` if the expect kernel type is not as same as actual kernel type.
The algorithm is described as follow
```cpp
using DataTransformationFN = std::function<void(const Tensor& in, Tensor* out)>;
using KernelTypePair = std::pair<KernelType, KernelType>;
map<KernelTypePair, DataTransformationFN> g_data_transformation_;
void OpWithKernel::Run() {
vec<Tensor> inputs = ...
auto actual_kernel_type = GetActualKernelType(inputs);
// The expected kernel type is related to actual kernel type.
// For the most operators, the expected kernel type is as same as
// actual kernel type.
//
// So we pass `actual_kernel_type` as a parameter of
// GetExpectedKernelType
auto expect_kernel_type = GetExpectedKernelType(actual_kernel_type);
auto trans = g_data_transformation_[{actual_kernel_type, expect_kernel_type}];
kernel.run(trans(inputs));
}
```
doc/faq/build_and_install/index_cn.rst
浏览文件 @ fab63cc6
......@@ -14,7 +14,7 @@
$ export CUDA_SO="$(\ls usr/lib64/libcuda* | xargs -I{} echo '-v {}:{}') $(\ls /usr/lib64/libnvidia* | xargs -I{} echo '-v {}:{}')"
$ export DEVICES=$(\ls /dev/nvidia* | xargs -I{} echo '--device {}:{}')
$ docker run ${CUDA_SO} ${DEVICES} -it paddledev/paddlepaddle:latest-gpu
$ docker run ${CUDA_SO} ${DEVICES} -it paddlepaddle/paddle:latest-gpu
更多关于Docker的安装与使用, 请参考 `PaddlePaddle Docker 文档 <http://www.paddlepaddle.org/doc_cn/build_and_install/install/docker_install.html>`_ 。
......@@ -109,3 +109,31 @@ PaddlePaddle使用avx SIMD指令提高cpu执行效率，因此错误的使用二
解决办法是：
* 卸载PaddlePaddle包 :code:`pip uninstall paddle`, 清理掉老旧的PaddlePaddle安装包，使得单元测试有一个干净的环境。如果PaddlePaddle包已经在python的site-packages里面，单元测试会引用site-packages里面的python包，而不是源码目录里 :code:`/python` 目录下的python包。同时，即便设置 :code:`PYTHONPATH` 到 :code:`/python` 也没用，因为python的搜索路径是优先已经安装的python包。
8. 下载MKLML库失败
------------------
.. code-block:: bash
make[2]: *** [third_party/mklml/src/extern_mklml-stamp/extern_mklml-download] 错误 4
make[1]: *** [CMakeFiles/extern_mklml.dir/all] 错误 2
make[1]: *** 正在等待未完成的任务....
原因：网速或SSL链接原因，导致MKLML库下载不成功。
解决办法是：手动下载并安装，具体步骤如下。
.. code-block:: bash
// 1. 进入对应的目录
cd build/third_party/mklml/src/extern_mklml
// 2. 查看包的大小， 正常情况下是75M，如果小于75M，即下载失败：
du -sh mklml_lnx_2018.0.1.20171007.tgz
// 3. 手动下载且解压缩，并手动生成download成功标签：
wget --no-check-certificate https://github.com/01org/mkl-dnn/releases/download/v0.11/mklml_lnx_2018.0.1.20171007.tgz -c -O mklml_lnx_2018.0.1.20171007.tgz
tar zxf mklml_lnx_2018.0.1.20171007.tgz
touch ../extern_mklml-stamp/extern_mklml-download
// 4. 接着编译即可
doc/getstarted/basic_usage/index_cn.rst 已删除 100644 → 0
浏览文件 @ 1ce2d613
经典的线性回归任务
==================
PaddlePaddle是源于百度的一个深度学习平台。这份简短的介绍将向你展示如何利用PaddlePaddle来解决一个经典的线性回归问题。
任务简介
--------
我们展示如何用PaddlePaddle解决 `单变量的线性回归 <https://www.baidu.com/s?wd=单变量线性回归>`_ 问题。线性回归的输入是一批点 `(x, y)` ，其中 `y = wx + b + ε`， 而 ε 是一个符合高斯分布的随机变量。线性回归的输出是从这批点估计出来的参数 `w` 和 `b` 。
一个例子是房产估值。我们假设房产的价格（y）是其大小（x）的一个线性函数，那么我们可以通过收集市场上房子的大小和价格，用来估计线性函数的参数w 和 b。
准备数据
-----------
假设变量 `x` 和 `y` 的真实关系为： `y = 2x + 0.3 + ε`，这里展示如何使用观测数据来拟合这一线性关系。首先，Python代码将随机产生2000个观测点，作为线性回归的输入。下面脚本符合PaddlePaddle期待的读取数据的Python程序的模式。
.. code-block:: python
# dataprovider.py
from paddle.trainer.PyDataProvider2 import *
import random
# 定义输入数据的类型: 2个浮点数
@provider(input_types=[dense_vector(1), dense_vector(1)],use_seq=False)
def process(settings, input_file):
for i in xrange(2000):
x = random.random()
yield [x], [2*x+0.3]
训练模型
-----------
为了还原 `y = 2x + 0.3`，我们先从一条随机的直线 `y' = wx + b` 开始，然后利用观测数据调整 `w` 和 `b` 使得 `y'` 和 `y` 的差距不断减小，最终趋于接近。这个过程就是模型的训练过程，而 `w` 和 `b` 就是模型的参数，即我们的训练目标。
在PaddlePaddle里，该模型的网络配置如下。
.. code-block:: python
# trainer_config.py
from paddle.trainer_config_helpers import *
# 1. 定义数据来源，调用上面的process函数获得观测数据
data_file = 'empty.list'
with open(data_file, 'w') as f: f.writelines(' ')
define_py_data_sources2(train_list=data_file, test_list=None,
module='dataprovider', obj='process',args={})
# 2. 学习算法。控制如何改变模型参数 w 和 b
settings(batch_size=12, learning_rate=1e-3, learning_method=MomentumOptimizer())
# 3. 神经网络配置
x = data_layer(name='x', size=1)
y = data_layer(name='y', size=1)
# 线性计算网络层: ȳ = wx + b
ȳ = fc_layer(input=x, param_attr=ParamAttr(name='w'), size=1, act=LinearActivation(), bias_attr=ParamAttr(name='b'))
# 计算误差函数，即 ȳ 和真实 y 之间的距离
cost = square_error_cost(input= ȳ, label=y)
outputs(cost)
这段简短的配置展示了PaddlePaddle的基本用法：
- 第一部分定义了数据输入。一般情况下，PaddlePaddle先从一个文件列表里获得数据文件地址，然后交给用户自定义的函数（例如上面的 `process`函数）进行读入和预处理从而得到真实输入。本文中由于输入数据是随机生成的不需要读输入文件，所以放一个空列表（`empty.list`）即可。
- 第二部分主要是选择学习算法，它定义了模型参数改变的规则。PaddlePaddle提供了很多优秀的学习算法，这里使用一个基于momentum的随机梯度下降(SGD)算法，该算法每批量(batch)读取12个采样数据进行随机梯度计算来更新更新。
- 最后一部分是神经网络的配置。由于PaddlePaddle已经实现了丰富的网络层，所以很多时候你需要做的只是定义正确的网络层并把它们连接起来。这里使用了三种网络单元：
- **数据层**：数据层 `data_layer` 是神经网络的入口，它读入数据并将它们传输到接下来的网络层。这里数据层有两个，分别对应于变量 `x` 和 `y`。
- **全连接层**：全连接层 `fc_layer` 是基础的计算单元，这里利用它建模变量之间的线性关系。计算单元是神经网络的核心，PaddlePaddle支持大量的计算单元和任意深度的网络连接，从而可以拟合任意的函数来学习复杂的数据关系。
- **回归误差代价层**：回归误差代价层 `square_error_cost` 是众多误差代价函数层的一种，它们在训练过程作为网络的出口，用来计算模型的误差，是模型参数优化的目标函数。
定义了网络结构并保存为 `trainer_config.py` 之后，运行以下训练命令：
.. code-block:: bash
paddle train --config=trainer_config.py --save_dir=./output --num_passes=30
PaddlePaddle将在观测数据集上迭代训练30轮，并将每轮的模型结果存放在 `./output` 路径下。从输出日志可以看到，随着轮数增加误差代价函数的输出在不断的减小，这意味着模型在训练数据上不断的改进，直到逼近真实解：` y = 2x + 0.3 `
模型检验
-----------
训练完成后，我们希望能够检验模型的好坏。一种常用的做法是用学习的模型对另外一组测试数据进行预测，评价预测的效果。在这个例子中，由于已经知道了真实答案，我们可以直接观察模型的参数是否符合预期来进行检验。
PaddlePaddle将每个模型参数作为一个numpy数组单独存为一个文件，所以可以利用如下方法读取模型的参数。
.. code-block:: python
import numpy as np
import os
def load(file_name):
with open(file_name, 'rb') as f:
f.read(16) # skip header for float type.
return np.fromfile(f, dtype=np.float32)
print 'w=%.6f, b=%.6f' % (load('output/pass-00029/w'), load('output/pass-00029/b'))
# w=1.999743, b=0.300137
.. image:: ./parameters.png
:align: center
:scale: 80 %
从图中可以看到，虽然 `w` 和 `b` 都使用随机值初始化，但在起初的几轮训练中它们都在快速逼近真实值，并且后续仍在不断改进，使得最终得到的模型几乎与真实模型一致。
这样，我们用PaddlePaddle解决了单变量线性回归问题， 包括数据输入、模型训练和最后的结果验证。
doc/getstarted/basic_usage/index_en.rst 已删除 100644 → 0
浏览文件 @ 1ce2d613
Simple Linear Regression
========================
PaddlePaddle is a deep learning platform open-sourced by Baidu. With PaddlePaddle, you can easily train a classic neural network within a couple lines of configuration, or you can build sophisticated models that provide state-of-the-art performance on difficult learning tasks like sentiment analysis, machine translation, image caption and so on.
Problem Background
------------------
Now, to give you a hint of what using PaddlePaddle looks like, let's start with a fundamental learning problem - `simple linear regression <https://en.wikipedia.org/wiki/Simple_linear_regression>`_: you have observed a set of two-dimensional data points of ``X`` and ``Y``, where ``X`` is an explanatory variable and ``Y`` is corresponding dependent variable, and you want to recover the underlying correlation between ``X`` and ``Y``. Linear regression can be used in many practical scenarios. For example, ``X`` can be a variable about house size, and ``Y`` a variable about house price. You can build a model that captures relationship between them by observing real estate markets.
Prepare the Data
-----------------
Suppose the true relationship can be characterized as ``Y = 2X + 0.3``, let's see how to recover this pattern only from observed data. Here is a piece of python code that feeds synthetic data to PaddlePaddle. The code is pretty self-explanatory, the only extra thing you need to add for PaddlePaddle is a definition of input data types.
.. code-block:: python
# dataprovider.py
from paddle.trainer.PyDataProvider2 import *
import random
# define data types of input: 2 real numbers
@provider(input_types=[dense_vector(1), dense_vector(1)],use_seq=False)
def process(settings, input_file):
for i in xrange(2000):
x = random.random()
yield [x], [2*x+0.3]
Train a NeuralNetwork
----------------------
To recover this relationship between ``X`` and ``Y``, we use a neural network with one layer of linear activation units and a square error cost layer. Don't worry if you are not familiar with these terminologies, it's just saying that we are starting from a random line ``Y' = wX + b`` , then we gradually adapt ``w`` and ``b`` to minimize the difference between ``Y'`` and ``Y``. Here is what it looks like in PaddlePaddle:
.. code-block:: python
# trainer_config.py
from paddle.trainer_config_helpers import *
# 1. read data. Suppose you saved above python code as dataprovider.py
data_file = 'empty.list'
with open(data_file, 'w') as f: f.writelines(' ')
define_py_data_sources2(train_list=data_file, test_list=None,
module='dataprovider', obj='process',args={})
# 2. learning algorithm
settings(batch_size=12, learning_rate=1e-3, learning_method=MomentumOptimizer())
# 3. Network configuration
x = data_layer(name='x', size=1)
y = data_layer(name='y', size=1)
y_predict = fc_layer(input=x, param_attr=ParamAttr(name='w'), size=1, act=LinearActivation(), bias_attr=ParamAttr(name='b'))
cost = square_error_cost(input=y_predict, label=y)
outputs(cost)
Some of the most fundamental usages of PaddlePaddle are demonstrated:
- The first part shows how to feed data into PaddlePaddle. In general cases, PaddlePaddle reads raw data from a list of files, and then do some user-defined process to get real input. In this case, we only need to create a placeholder file since we are generating synthetic data on the fly.
- The second part describes learning algorithm. It defines in what ways adjustments are made to model parameters. PaddlePaddle provides a rich set of optimizers, but a simple momentum based optimizer will suffice here, and it processes 12 data points each time.
- Finally, the network configuration. It usually is as simple as "stacking" layers. Three kinds of layers are used in this configuration:
- **Data Layer**: a network always starts with one or more data layers. They provide input data to the rest of the network. In this problem, two data layers are used respectively for ``X`` and ``Y``.
- **FC Layer**: FC layer is short for Fully Connected Layer, which connects all the input units to current layer and does the actual computation specified as activation function. Computation layers like this are the fundamental building blocks of a deeper model.
- **Cost Layer**: in training phase, cost layers are usually the last layers of the network. They measure the performance of current model, and provide guidence to adjust parameters.
Now that everything is ready, you can train the network with a simple command line call:
.. code-block:: bash
paddle train --config=trainer_config.py --save_dir=./output --num_passes=30
This means that PaddlePaddle will train this network on the synthectic dataset for 30 passes, and save all the models under path ``./output``. You will see from the messages printed out during training phase that the model cost is decreasing as time goes by, which indicates we are getting a closer guess.
Evaluate the Model
-------------------
Usually, a different dataset that left out during training phase should be used to evalute the models. However, we are lucky enough to know the real answer: ``w=2, b=0.3``, thus a better option is to check out model parameters directly.
In PaddlePaddle, training is just to get a collection of model parameters, which are ``w`` and ``b`` in this case. Each parameter is saved in an individual file in the popular ``numpy`` array format. Here is the code that reads parameters from last pass.
.. code-block:: python
import numpy as np
import os
def load(file_name):
with open(file_name, 'rb') as f:
f.read(16) # skip header for float type.
return np.fromfile(f, dtype=np.float32)
print 'w=%.6f, b=%.6f' % (load('output/pass-00029/w'), load('output/pass-00029/b'))
# w=1.999743, b=0.300137
.. image:: parameters.png
:align: center
Although starts from a random guess, you can see that value of ``w`` changes quickly towards 2 and ``b`` changes quickly towards 0.3. In the end, the predicted line is almost identical with real answer.
There, you have recovered the underlying pattern between ``X`` and ``Y`` only from observed data.
doc/getstarted/build_and_install/build_from_source_cn.rst 0 → 100644
浏览文件 @ fab63cc6
从源码编译
======================
.. _build_step:
编译方法
----------------
PaddlePaddle主要使用 `CMake <https://cmake.org>`_ 以及GCC, G++作为编译工具。
我们推荐您使用PaddlePaddle Docker编译环境镜像完成编译，这样可以免去单独安装编译依赖的步骤，可选的不同编译环境Docker镜像
可以在 `这里 <https://hub.docker.com/r/paddlepaddle/paddle_manylinux_devel/tags/>`_ 找到。
如果您选择不使用Docker镜像，则需要在本机安装下面章节列出的 `编译依赖`_ 之后才能开始编译的步骤。
编译PaddlePaddle，需要执行：
.. code-block:: bash
git clone https://github.com/PaddlePaddle/Paddle.git
cd Paddle
# 如果使用Docker编译环境，执行下面的命令编译CPU-Only的二进制
docker run -it -v $PWD:/paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=OFF" paddlepaddle/paddle_manylinux_devel:cuda8.0_cudnn5 bash -x /paddle/paddle/scripts/docker/build.sh
# 如果不使用Docker编译环境，执行下面的命令
mkdir build
cd build
cmake -DWITH_GPU=OFF -DWITH_TESTING=OFF ..
make
编译完成后会在build/python/dist目录下生成输出的whl包，可以选在在当前机器安装也可以拷贝到目标机器安装：
.. code-block:: bash
pip install build/python/dist/*.whl
.. _run_test:
执行单元测试
----------------
如果您期望在编译完成后立即执行所有的单元测试，可以按照下面的方法：
使用Docker的情况下，设置 :code:`RUN_TEST=ON` 和 :code:`WITH_TESTING=ON` 就会在完成编译之后，立即执行单元测试。
开启 :code:`WITH_GPU=ON` 可以指定同时执行GPU上的单元测试。
.. code-block:: bash
docker run -it -v $PWD:/paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=ON" -e "RUN_TEST=ON" paddlepaddle/paddle_manylinux_devel:cuda8.0_cudnn5 bash -x /paddle/paddle/scripts/docker/build.sh
如果不使用Docker，可以执行ctest命令即可：
.. code-block:: bash
mkdir build
cd build
cmake -DWITH_GPU=OFF -DWITH_TESTING=OFF ..
make
ctest
# 指定执行其中一个单元测试 test_mul_op
ctest -R test_mul_op
.. _compile_deps:
编译依赖
----------------
PaddlePaddle编译需要使用到下面的依赖（包含但不限于），其他的依赖软件，会自动在编译时下载。
.. csv-table:: PaddlePaddle编译依赖
:header: "依赖", "版本", "说明"
:widths: 10, 15, 30
"CMake", ">=3.2", ""
"GCC", "4.8.2", "推荐使用CentOS的devtools2"
"Python", "2.7.x", "依赖libpython2.7.so"
"pip", ">=9.0", ""
"numpy", "", ""
"SWIG", ">=2.0", ""
"Go", ">=1.8", "可选"
.. _build_options:
编译选项
----------------
PaddlePaddle的编译选项，包括生成CPU/GPU二进制文件、链接何种BLAS库等。
用户可在调用cmake的时候设置它们，详细的cmake使用方法可以参考
`官方文档 <https://cmake.org/cmake-tutorial>`_ 。
在cmake的命令行中，通过使用 ``-D`` 命令设置该类编译选项，例如：
.. code-block:: bash
cmake .. -DWITH_GPU=OFF
.. csv-table:: 编译选项说明
:header: "选项", "说明", "默认值"
:widths: 1, 7, 2
"WITH_GPU", "是否支持GPU", "ON"
"WITH_C_API", "是否仅编译CAPI", "OFF"
"WITH_DOUBLE", "是否使用双精度浮点数", "OFF"
"WITH_DSO", "是否运行时动态加载CUDA动态库，而非静态加载CUDA动态库。", "ON"
"WITH_AVX", "是否编译含有AVX指令集的PaddlePaddle二进制文件", "ON"
"WITH_PYTHON", "是否内嵌PYTHON解释器", "ON"
"WITH_STYLE_CHECK", "是否编译时进行代码风格检查", "ON"
"WITH_TESTING", "是否开启单元测试", "ON"
"WITH_DOC", "是否编译中英文文档", "OFF"
"WITH_SWIG_PY", "是否编译PYTHON的SWIG接口，该接口可用于预测和定制化训练", "Auto"
"WITH_GOLANG", "是否编译go语言的可容错parameter server", "ON"
"WITH_MKL", "是否使用MKL数学库，如果为否则是用OpenBLAS", "ON"
BLAS
+++++
PaddlePaddle支持 `MKL <https://software.intel.com/en-us/intel-mkl>`_ 和
`OpenBlAS <http://www.openblas.net/>`_ 两种BLAS库。默认使用MKL。如果使用MKL并且机器含有AVX2指令集，
还会下载MKL-DNN数学库，详细参考 `这里 <https://github.com/PaddlePaddle/Paddle/tree/develop/doc/design/mkldnn#cmake>`_ 。
如果关闭MKL，则会使用OpenBLAS作为BLAS库。
CUDA/cuDNN
+++++++++++
PaddlePaddle在编译时/运行时会自动找到系统中安装的CUDA和cuDNN库进行编译和执行。
使用参数 :code:`-DCUDA_ARCH_NAME=Auto` 可以指定开启自动检测SM架构，加速编译。
PaddlePaddle可以使用cuDNN v5.1之后的任何一个版本来编译运行，但尽量请保持编译和运行使用的cuDNN是同一个版本。
我们推荐使用最新版本的cuDNN。
编译选项的设置
++++++++++++++
PaddePaddle通过编译时指定路径来实现引用各种BLAS/CUDA/cuDNN库。cmake编译时，首先在系统路径（ :code:`/usr/lib:/usr/local/lib` ）中搜索这几个库，同时也会读取相关路径变量来进行搜索。 通过使用 ``-D`` 命令可以设置，例如
.. code-block:: bash
cmake .. -DWITH_GPU=ON -DWITH_TESTING=OFF -DCUDNN_ROOT=/opt/cudnnv5
**注意：这几个编译选项的设置，只在第一次cmake的时候有效。如果之后想要重新设置，推荐清理整个编译目录（** :code:`rm -rf` ）**后，再指定。**
doc/getstarted/build_and_install/build_from_source_en.md 已删除 100644 → 0
浏览文件 @ 1ce2d613
Installing from Sources
==========================
* [1. Download and Setup](#download)
* [2. Requirements](#requirements)
* [3. Build on Ubuntu](#ubuntu)
* [4. Build on Centos](#centos)
## <span id="download">Download and Setup</span>
You can download PaddlePaddle from the [github source](https://github.com/PaddlePaddle/Paddle).
```bash
git clone https://github.com/PaddlePaddle/Paddle paddle
cd paddle
```
## <span id="requirements">Requirements</span>
To compile the source code, your computer must be equipped with the following dependencies.
- **Compiler**: GCC >= 4.8 or Clang >= 3.3 (AppleClang >= 5.1) and gfortran compiler
- **CMake**: CMake >= 3.0 (at least CMake 3.4 on Mac OS X)
- **BLAS**: MKL, OpenBlas or ATLAS
- **Python**: only support Python 2.7
- **Go**
**Note:** For CUDA 7.0 and CUDA 7.5, GCC 5.0 and up are not supported!
For CUDA 8.0, GCC versions later than 5.3 are not supported!
### Options
PaddlePaddle supports some build options.
<html>
<table>
<thead>
<tr>
<th scope="col" class="left">Optional</th>
<th scope="col" class="left">Description</th>
</tr>
</thead>
<tbody>
<tr><td class="left">WITH_GPU</td><td class="left">Compile PaddlePaddle with NVIDIA GPU</td></tr>
<tr><td class="left">WITH_AVX</td><td class="left">Compile PaddlePaddle with AVX intrinsics</td></tr>
<tr><td class="left">WITH_DSO</td><td class="left">Compile PaddlePaddle with dynamic linked CUDA</td></tr>
<tr><td class="left">WITH_TESTING</td><td class="left">Compile PaddlePaddle with unit testing</td></tr>
<tr><td class="left">WITH_SWIG_PY</td><td class="left">Compile PaddlePaddle with inference api</td></tr>
<tr><td class="left">WITH_STYLE_CHECK</td><td class="left">Compile PaddlePaddle with style check</td></tr>
<tr><td class="left">WITH_PYTHON</td><td class="left">Compile PaddlePaddle with python interpreter</td></tr>
<tr><td class="left">WITH_DOUBLE</td><td class="left">Compile PaddlePaddle with double precision</td></tr>
<tr><td class="left">WITH_RDMA</td><td class="left">Compile PaddlePaddle with RDMA support</td></tr>
<tr><td class="left">WITH_TIMER</td><td class="left">Compile PaddlePaddle with stats timer</td></tr>
<tr><td class="left">WITH_PROFILER</td><td class="left">Compile PaddlePaddle with GPU profiler</td></tr>
<tr><td class="left">WITH_DOC</td><td class="left">Compile PaddlePaddle with documentation</td></tr>
<tr><td class="left">WITH_COVERAGE</td><td class="left">Compile PaddlePaddle with code coverage</td></tr>
<tr><td class="left">COVERALLS_UPLOAD</td><td class="left">Package code coverage data to coveralls</td></tr>
<tr><td class="left">ON_TRAVIS</td><td class="left">Exclude special unit test on Travis CI</td></tr>
</tbody>
</table>
</html>
**Note:**
- The GPU version works best with Cuda Toolkit 8.0 and cuDNN v5.
- Other versions like Cuda Toolkit 7.0, 7.5 and cuDNN v3, v4 are also supported.
- **To utilize cuDNN v5, Cuda Toolkit 7.5 is prerequisite and vice versa.**
As a simple example, consider the following:
1. **BLAS Dependencies(optional)**
CMake will search BLAS libraries from the system. If not found, OpenBLAS will be downloaded, built and installed automatically.
To utilize preinstalled BLAS， you can simply specify MKL, OpenBLAS or ATLAS via `MKL_ROOT`, `OPENBLAS_ROOT` or `ATLAS_ROOT`.
```bash
# specify MKL
cmake .. -DMKL_ROOT=<mkl_path>
# or specify OpenBLAS
cmake .. -DOPENBLAS_ROOT=<openblas_path>
```
2. **Doc Dependencies(optional)**
To generate PaddlePaddle's documentation, install dependencies and set `-DWITH_DOC=ON` as follows:
```bash
pip install 'sphinx>=1.4.0'
pip install sphinx_rtd_theme recommonmark
# install doxygen on Ubuntu
sudo apt-get install doxygen
# install doxygen on Mac OS X
brew install doxygen
# active docs in cmake
cmake .. -DWITH_DOC=ON`
```
## <span id="ubuntu">Build on Ubuntu 14.04</span>
### Install Dependencies
- **Paddle Dependencies**
```bash
# necessary
sudo apt-get update
sudo apt-get install -y git curl gcc g++ gfortran make build-essential automake
sudo apt-get install -y python python-pip python-numpy libpython-dev bison
sudo pip install 'protobuf==3.1.0.post1'
# Install Go
# You can follow https://golang.org/doc/install for a detailed explanation.
wget -O go.tgz https://storage.googleapis.com/golang/go1.8.1.linux-amd64.tar.gz && \
tar -C $HOME -xzf go.tgz && \
mkdir $HOME/gopath && \
rm go.tgz
# Setup environment variables
export GOROOT=$HOME/go
export GOPATH=$HOME/gopath
export PATH=$PATH:$GOROOT/bin
# install cmake 3.4
curl -sSL https://cmake.org/files/v3.4/cmake-3.4.1.tar.gz | tar -xz && \
cd cmake-3.4.1 && ./bootstrap && make -j4 && sudo make install && \
cd .. && rm -rf cmake-3.4.1
```
- **GPU Dependencies (optional)**
To build GPU version, you will need the following installed:
1. a CUDA-capable GPU
2. A supported version of Linux with a GCC compiler and toolchain
3. NVIDIA CUDA Toolkit (available at http://developer.nvidia.com/cuda-downloads)
4. NVIDIA cuDNN Library (available at https://developer.nvidia.com/cudnn)
The CUDA development environment relies on tight integration with the host development environment,
including the host compiler and C runtime libraries, and is therefore only supported on
distribution versions that have been qualified for this CUDA Toolkit release.
After downloading cuDNN library, issue the following commands:
```bash
sudo tar -xzf cudnn-7.5-linux-x64-v5.1.tgz -C /usr/local
sudo chmod a+r /usr/local/cuda/include/cudnn.h /usr/local/cuda/lib64/libcudnn*
```
Then you need to set LD\_LIBRARY\_PATH, PATH environment variables in ~/.bashrc.
```bash
export LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH
export PATH=/usr/local/cuda/bin:$PATH
```
### Build and Install
As usual, the best option is to create build folder under paddle project directory.
```bash
mkdir build && cd build
```
Finally, you can build and install PaddlePaddle:
```bash
# you can add build option here, such as:
cmake .. -DCMAKE_INSTALL_PREFIX=<path to install>
# please use sudo make install, if you want to install PaddlePaddle into the system
make -j `nproc` && make install
# set PaddlePaddle installation path in ~/.bashrc
export PATH=<path to install>/bin:$PATH
# install PaddlePaddle Python modules.
sudo pip install <path to install>/opt/paddle/share/wheels/*.whl
```
## <span id="centos">Build on Centos 7</span>
### Install Dependencies
- **CPU Dependencies**
```bash
# necessary
sudo yum update
sudo yum install -y epel-release
sudo yum install -y make cmake3 python-devel python-pip gcc-gfortran swig git
sudo pip install wheel numpy
sudo pip install 'protobuf>=3.0.0'
```
- **GPU Dependencies (optional)**
To build GPU version, you will need the following installed:
1. a CUDA-capable GPU
2. A supported version of Linux with a GCC compiler and toolchain
3. NVIDIA CUDA Toolkit (available at http://developer.nvidia.com/cuda-downloads)
4. NVIDIA cuDNN Library (available at https://developer.nvidia.com/cudnn)
The CUDA development environment relies on tight integration with the host development environment,
including the host compiler and C runtime libraries, and is therefore only supported on
distribution versions that have been qualified for this CUDA Toolkit release.
After downloading cuDNN library, issue the following commands:
```bash
sudo tar -xzf cudnn-7.5-linux-x64-v5.1.tgz -C /usr/local
sudo chmod a+r /usr/local/cuda/include/cudnn.h /usr/local/cuda/lib64/libcudnn*
```
Then you need to set LD\_LIBRARY\_PATH, PATH environment variables in ~/.bashrc.
```bash
export LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH
export PATH=/usr/local/cuda/bin:$PATH
```
### Build and Install
As usual, the best option is to create build folder under paddle project directory.
```bash
mkdir build && cd build
```
Finally, you can build and install PaddlePaddle:
```bash
# you can add build option here, such as:
cmake3 .. -DCMAKE_INSTALL_PREFIX=<path to install>
# please use sudo make install, if you want to install PaddlePaddle into the system
make -j `nproc` && make install
# set PaddlePaddle installation path in ~/.bashrc
export PATH=<path to install>/bin:$PATH
# install PaddlePaddle Python modules.
sudo pip install <path to install>/opt/paddle/share/wheels/*.whl
```
doc/getstarted/build_and_install/build_from_source_en.rst 0 → 100644
浏览文件 @ fab63cc6
Build from Sources
==========================
.. _build_step:
How To Build
----------------
PaddlePaddle mainly uses `CMake <https://cmake.org>`_ and GCC, G++ as compile
tools. We recommend you to use our pre-built Docker image to run the build
to avoid installing dependencies by yourself. We have several build environment
Docker images `here <https://hub.docker.com/r/paddlepaddle/paddle_manylinux_devel/tags/>`_ .
If you choose not to use Docker image for your build, you need to install the
below `Compile Dependencies`_ before run the build.
Then run:
.. code-block:: bash
git clone https://github.com/PaddlePaddle/Paddle.git
cd Paddle
# run the following command to build a CPU-Only binaries if you are using docker
docker run -it -v $PWD:/paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=OFF" paddlepaddle/paddle_manylinux_devel:cuda8.0_cudnn5 bash -x /paddle/paddle/scripts/docker/build.sh
# else run these commands
mkdir build
cd build
cmake -DWITH_GPU=OFF -DWITH_TESTING=OFF ..
make
When the compile finishes, you can get the output whl package under
build/python/dist, then you can choose to install the whl on local
machine or copy it to the target machine.
.. code-block:: bash
pip install build/python/dist/*.whl
.. _run_test:
Run Tests
----------------
If you wish to run the tests, you may follow the below steps:
When using Docker, set :code:`RUN_TEST=ON` and :code:`WITH_TESTING=ON` will run test immediately after the build.
Set :code:`WITH_GPU=ON` Can also run tests on GPU.
.. code-block:: bash
docker run -it -v $PWD:/paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=ON" -e "RUN_TEST=ON" paddlepaddle/paddle_manylinux_devel:cuda8.0_cudnn5 bash -x paddle/paddle/scripts/docker/build.sh
If you don't use Docker, just run ctest will start the tests:
.. code-block:: bash
mkdir build
cd build
cmake -DWITH_GPU=OFF -DWITH_TESTING=ON ..
make
ctest
# run a single test like test_mul_op
ctest -R test_mul_op
.. _compile_deps:
Compile Dependencies
----------------
PaddlePaddle need the following dependencies when compiling, other dependencies
will be downloaded automatically.
.. csv-table:: PaddlePaddle Compile Dependencies
:header: "Dependency", "Version", "Description"
:widths: 10, 15, 30
"CMake", ">=3.2", ""
"GCC", "4.8.2", "Recommend devtools2 for CentOS"
"Python", "2.7.x", "Need libpython2.7.so"
"pip", ">=9.0", ""
"numpy", "", ""
"SWIG", ">=2.0", ""
"Go", ">=1.8", "Optional"
.. _build_options:
Build Options
----------------
Build options include whether build binaries for CPU or GPU, which BLAS
library to use etc. You may pass these settings when running cmake.
For detailed cmake tutorial please refer to `here <https://cmake.org/cmake-tutorial>`_ 。
.. _build_options_bool:
Bool Type Options
----------------
You can add :code:`-D` argument to pass such options, like:
.. code-block:: bash
cmake .. -DWITH_GPU=OFF
.. csv-table:: Bool Type Options
:header: "Option", "Description", "Default"
:widths: 1, 7, 2
"WITH_GPU", "Build with GPU support", "ON"
"WITH_C_API", "Build only CAPI", "OFF"
"WITH_DOUBLE", "Build with double precision", "OFF"
"WITH_DSO", "Dynamically load CUDA libraries", "ON"
"WITH_AVX", "Build with AVX support", "ON"
"WITH_PYTHON", "Build with integrated Python interpreter", "ON"
"WITH_STYLE_CHECK", "Check code style when building", "ON"
"WITH_TESTING", "Build unit tests", "ON"
"WITH_DOC", "Build documentations", "OFF"
"WITH_SWIG_PY", "Build Python SWIG interface for V2 API", "Auto"
"WITH_GOLANG", "Build fault-tolerant parameter server written in go", "ON"
"WITH_MKL", "Use MKL as BLAS library, else use OpenBLAS", "ON"
BLAS
+++++
PaddlePaddle supports `MKL <https://software.intel.com/en-us/intel-mkl>`_ and
`OpenBlAS <http://www.openblas.net/>`_ as BLAS library。By default it uses MKL.
If you are using MKL and your machine supports AVX2, MKL-DNN will also be downloaded
and used, for more `details <https://github.com/PaddlePaddle/Paddle/tree/develop/doc/design/mkldnn#cmake>`_ .
If you choose not to use MKL, then OpenBlAS will be used.
CUDA/cuDNN
+++++++++++
PaddlePaddle will automatically find CUDA and cuDNN when compiling and running.
parameter :code:`-DCUDA_ARCH_NAME=Auto` can be used to detect SM architecture
automatically in order to speed up the build.
PaddlePaddle can build with any version later than cuDNN v5.1, and we intend to
keep on with latest cuDNN versions. Be sure to run with the same version of cuDNN
you built.
Pass Compile Options
++++++++++++++
You can pass compile options to use intended BLAS/CUDA/Cudnn libraries.
When running cmake command, it will search system paths like
:code:`/usr/lib:/usr/local/lib` and then search paths that you
passed to cmake, i.e.
.. code-block:: bash
cmake .. -DWITH_GPU=ON -DWITH_TESTING=OFF -DCUDNN_ROOT=/opt/cudnnv5
**NOTE: These options only take effect when running cmake for the first time, you need to clean the cmake cache or clean the build directory (** :code:`rm -rf` **) if you want to change it.**
doc/getstarted/build_and_install/cmake/build_from_source_cn.rst 已删除 100644 → 0
浏览文件 @ 1ce2d613
PaddlePaddle的编译选项
======================
PaddlePaddle的编译选项，包括生成CPU/GPU二进制文件、链接何种BLAS库等。用户可在调用cmake的时候设置它们，详细的cmake使用方法可以参考 `官方文档 <https://cmake.org/cmake-tutorial>`_ 。
Bool型的编译选项
----------------
用户可在cmake的命令行中，通过使用 ``-D`` 命令设置该类编译选项，例如
.. code-block:: bash
cmake .. -DWITH_GPU=OFF
.. csv-table:: Bool型的编译选项
:widths: 1, 7, 2
:file: compile_options.csv
BLAS/CUDA/Cudnn的编译选项
--------------------------
BLAS
+++++
PaddlePaddle支持以下任意一种BLAS库：`MKL <https://software.intel.com/en-us/intel-mkl>`_ ，`ATLAS <http://math-atlas.sourceforge.net/>`_ ，`OpenBlAS <http://www.openblas.net/>`_ 和 `REFERENCE BLAS <http://www.netlib.org/blas/>`_ 。
.. csv-table:: BLAS路径相关的编译选项
:widths: 1, 2, 7
:file: cblas_settings.csv
CUDA/Cudnn
+++++++++++
PaddlePaddle可以使用cudnn v2之后的任何一个版本来编译运行，但尽量请保持编译和运行使用的cudnn是同一个版本。 我们推荐使用最新版本的cudnn v5.1。
编译选项的设置
++++++++++++++
PaddePaddle通过编译时指定路径来实现引用各种BLAS/CUDA/Cudnn库。cmake编译时，首先在系统路径(/usr/lib\:/usr/local/lib)中搜索这几个库，同时也会读取相关路径变量来进行搜索。 通过使用 ``-D`` 命令可以设置，例如
.. code-block:: bash
cmake .. -DMKL_ROOT=/opt/mkl/ -DCUDNN_ROOT=/opt/cudnnv5
注意：这几个编译选项的设置，只在第一次cmake的时候有效。如果之后想要重新设置，推荐清理整个编译目录（``rm -rf``）后，再指定。
doc/getstarted/build_and_install/cmake/cblas_settings.csv 已删除 100644 → 0
浏览文件 @ 1ce2d613
编译选项,描述,注意
MKL_ROOT,MKL的路径,${MKL_ROOT}/include下需要包含mkl.h，${MKL_ROOT}/lib目录下需要包含mkl_core，mkl_sequential和mkl_intel_lp64三个库。
ATLAS_ROOT,ATLAS的路径,${ATLAS_ROOT}/include下需要包含cblas.h，${ATLAS_ROOT}/lib下需要包含cblas和atlas两个库。
OPENBLAS_ROOT,OpenBLAS的路径,${OPENBLAS_ROOT}/include下需要包含cblas.h，${OPENBLAS_ROOT}/lib下需要包含openblas库。
REFERENCE_CBLAS_ROOT,REFERENCE BLAS的路径,${REFERENCE_CBLAS_ROOT}/include下需要包含cblas.h，${REFERENCE_CBLAS_ROOT}/lib下需要包含cblas库。
\ No newline at end of file
doc/getstarted/build_and_install/cmake/compile_options.csv 已删除 100644 → 0
浏览文件 @ 1ce2d613
选项,说明,默认值
WITH_GPU,是否支持GPU。,取决于是否寻找到CUDA工具链
WITH_DOUBLE,是否使用双精度浮点数。,否
WITH_DSO,是否运行时动态加载CUDA动态库，而非静态加载CUDA动态库。,是
WITH_AVX,是否编译含有AVX指令集的PaddlePaddle二进制文件,是
WITH_PYTHON,是否内嵌PYTHON解释器。方便今后的嵌入式移植工作。,是
WITH_STYLE_CHECK,是否编译时进行代码风格检查,是
WITH_RDMA,是否开启RDMA,否
WITH_TIMER,是否开启计时功能。如果开启会导致运行略慢，打印的日志变多，但是方便调试和测Benchmark,否
WITH_TESTING,是否开启单元测试,取决于是否寻找到GTEST
WITH_DOC,是否编译中英文文档,否
WITH_SWIG_PY,是否编译PYTHON的SWIG接口，该接口可用于预测和定制化训练,取决于是否寻找到SWIG
\ No newline at end of file
doc/getstarted/build_and_install/paddleci.png 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
doc/getstarted/build_and_install/pip_install_cn.rst 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
doc/getstarted/build_and_install/pip_install_en.rst 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
doc/getstarted/concepts/src/infer.py 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
doc/getstarted/index_cn.rst
浏览文件 @ fab63cc6
此差异已折叠。
doc/getstarted/index_en.rst
浏览文件 @ fab63cc6
此差异已折叠。
doc/howto/dev/build_cn.md
浏览文件 @ fab63cc6
此差异已折叠。
doc/howto/dev/build_en.md
浏览文件 @ fab63cc6
此差异已折叠。
doc/howto/dev/new_op_cn.md
浏览文件 @ fab63cc6
此差异已折叠。
doc/howto/dev/new_op_en.md
浏览文件 @ fab63cc6
此差异已折叠。
doc/howto/dev/write_docs_cn.rst
浏览文件 @ fab63cc6
此差异已折叠。
doc/howto/dev/write_docs_en.rst 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
doc/howto/index_cn.rst
浏览文件 @ fab63cc6
此差异已折叠。
doc/howto/index_en.rst
浏览文件 @ fab63cc6
此差异已折叠。
doc/howto/optimization/cpu_profiling.md 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
doc/howto/optimization/cpu_profiling_cn.md 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
doc/howto/optimization/pprof_1.png 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
doc/howto/optimization/pprof_2.png 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
doc/howto/read_source.md 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
doc/howto/usage/cluster/fabric_cn.md 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
doc/howto/usage/cluster/fabric_en.md 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
doc/howto/usage/cluster/k8s_aws_cn.md 0 → 120000
浏览文件 @ fab63cc6
此差异已折叠。
doc/howto/usage/cluster/openmpi_cn.md 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
doc/howto/usage/cluster/openmpi_en.md 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
doc/howto/usage/cluster/src/k8s-paddle-arch.png 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
doc/howto/usage/k8s/k8s_basis_cn.md 已删除 100644 → 0
浏览文件 @ 1ce2d613
此差异已折叠。
doc/howto/usage/k8s/src/k8s-paddle-arch.png 已删除 100644 → 0
浏览文件 @ 1ce2d613
此差异已折叠。
doc/mobile/cross_compiling_for_ios_en.md 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
doc/mobile/index_en.rst
浏览文件 @ fab63cc6
此差异已折叠。
paddle/api/CMakeLists.txt
浏览文件 @ fab63cc6
此差异已折叠。
paddle/capi/CMakeLists.txt
浏览文件 @ fab63cc6
此差异已折叠。
paddle/capi/Main.cpp
浏览文件 @ fab63cc6
此差异已折叠。
paddle/capi/Matrix.cpp
浏览文件 @ fab63cc6
此差异已折叠。
paddle/capi/config.h.in
浏览文件 @ fab63cc6
此差异已折叠。
paddle/capi/error.cpp 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/capi/error.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/capi/main.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/capi/matrix.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/cuda/include/hl_cnn.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/cuda/src/hl_cuda_cnn.cu
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/CMakeLists.txt
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/attribute.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/attribute.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/backward.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/backward.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/backward.md 已删除 100644 → 0
浏览文件 @ 1ce2d613
此差异已折叠。
paddle/framework/backward_test.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/block_desc.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/block_desc.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/data_layout.h 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/data_transform.cc 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/data_transform.h 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/data_transform_test.cc 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/data_type.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/ddim_test.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/executor.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/executor.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/feed_fetch_type.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/framework.proto
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/init.cc 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/init.h 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/init_test.cc 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/library_type.h 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/lod_rank_table.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/lod_rank_table.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/lod_tensor.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/lod_tensor.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/lod_tensor_array.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/lod_tensor_test.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/lod_tensor_test.cu
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/op_desc.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/op_desc.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/op_info.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/op_info.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/op_kernel_type.h 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/op_kernel_type_test.cc 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/op_proto_maker.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/op_registry.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/op_registry.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/operator.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/operator.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/operator_test.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/program_desc.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/program_desc.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/prune.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/prune.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/prune_test.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/scope.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/scope.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/scope_test.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/selected_rows.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/selected_rows.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/shape_inference.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/shape_inference.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/tensor.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/tensor.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/tensor.md
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/tensor_array.cc 已删除 100644 → 0
浏览文件 @ 1ce2d613
此差异已折叠。
paddle/framework/tensor_array.h 已删除 100644 → 0
浏览文件 @ 1ce2d613
此差异已折叠。
paddle/framework/tensor_array_test.cc 已删除 100644 → 0
浏览文件 @ 1ce2d613
此差异已折叠。
paddle/framework/tensor_impl.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/tensor_test.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/tensor_util.cc 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/tensor_util.cu 0 → 120000
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/tensor_util.h 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/tensor_util_test.cc 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/threadpool.cc 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/threadpool.h 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/threadpool_test.cc 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/type_defs.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/var_desc.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/var_desc.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/framework/var_type.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/function/EigenGemm.cpp
浏览文件 @ fab63cc6
此差异已折叠。
paddle/function/GemmConvOp.cpp
浏览文件 @ fab63cc6
此差异已折叠。
paddle/function/Im2Col.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/function/Im2ColTest.cpp
浏览文件 @ fab63cc6
此差异已折叠。
paddle/gserver/layers/FactorizationMachineLayer.cpp 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/gserver/layers/FactorizationMachineLayer.h 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/gserver/layers/MKLDNNLRNLayer.cpp 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/gserver/layers/MKLDNNLRNLayer.h 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/gserver/layers/MKLDNNLayer.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/gserver/layers/PoolLayer.cpp
浏览文件 @ fab63cc6
此差异已折叠。
paddle/gserver/layers/PoolLayer.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/gserver/tests/CMakeLists.txt
浏览文件 @ fab63cc6
此差异已折叠。
paddle/math/Allocator.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/math/CMakeLists.txt
浏览文件 @ fab63cc6
此差异已折叠。
paddle/math/CpuSparseMatrix.cpp
浏览文件 @ fab63cc6
此差异已折叠。
paddle/math/CpuSparseMatrix.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/math/MathFunctions.cpp
浏览文件 @ fab63cc6
此差异已折叠。
paddle/math/MathFunctions.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/math/Matrix.cpp
浏览文件 @ fab63cc6
此差异已折叠。
paddle/math/Matrix.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/math/NEONFunctions.cpp
浏览文件 @ fab63cc6
此差异已折叠。
paddle/math/NEONFunctions.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/math/SIMDFunctions.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/math/float16.h 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/math/tests/CMakeLists.txt
浏览文件 @ fab63cc6
此差异已折叠。
paddle/math/tests/test_float16.cpp 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/math/tests/test_float16.cu 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/memory/CMakeLists.txt
浏览文件 @ fab63cc6
此差异已折叠。
paddle/memory/README.md
浏览文件 @ fab63cc6
此差异已折叠。
paddle/memory/detail/memory_block.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/memory/detail/meta_cache.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/memory/detail/meta_cache.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/memory/detail/meta_data.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/memory/detail/meta_data.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/memory/memcpy.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/memory/memory.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/memory/memory_test.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/CMakeLists.txt
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/accuracy_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/accuracy_op.cu
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/accuracy_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/activation_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/activation_op.cu
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/activation_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/adadelta_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/adadelta_op.cu
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/adadelta_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/adagrad_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/adagrad_op.cu
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/adagrad_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/adam_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/adam_op.cu
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/adam_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/adamax_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/adamax_op.cu
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/adamax_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/array_operator.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/assign_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/auc_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/auc_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/batch_norm_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/batch_norm_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/beam_search_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/beam_search_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/cast_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/cast_op.cu
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/cast_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/chunk_eval_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/chunk_eval_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/clip_by_norm_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/clip_by_norm_op.cu
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/clip_by_norm_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/clip_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/clip_op.cu
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/clip_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/compare_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/compare_op.cu
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/compare_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/concat_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/concat_op.cu.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/concat_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/cond_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/cond_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/conv_cudnn_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/conv_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/conv_op.cu.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/conv_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/conv_shift_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/conv_shift_op.cu
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/conv_shift_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/cos_sim_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/cos_sim_op.cu
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/cos_sim_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/crf_decoding_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/crf_decoding_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/crop_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/crop_op.cu
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/crop_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/cross_entropy_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/detail/CMakeLists.txt 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/detail/recv_impl.cc 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/detail/safe_ref.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/detail/send_impl.cc 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/detail/send_recv.proto 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/detail/send_recv_impl.h 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/detail/simple_block_queue.h 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/detection_output_op.cc 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/detection_output_op.cu.cc 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/detection_output_op.h 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/dropout_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/dropout_op.cu
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/dropout_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/dynamic_recurrent_op.cc 已删除 100644 → 0
浏览文件 @ 1ce2d613
此差异已折叠。
paddle/operators/dynamic_recurrent_op.h 已删除 100644 → 0
浏览文件 @ 1ce2d613
此差异已折叠。
paddle/operators/dynamic_recurrent_op_test.cc 已删除 100644 → 0
浏览文件 @ 1ce2d613
此差异已折叠。
paddle/operators/elementwise_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/expand_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/expand_op.cu
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/expand_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/feed_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/fetch_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/fill_op.cc 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/ftrl_op.cc 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/ftrl_op.cu 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/ftrl_op.h 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/gather.cu.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/gather_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/gather_op.cu
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/gather_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/gru_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/gru_op.cu.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/gru_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/gru_unit_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/gru_unit_op.cu
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/gru_unit_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/hinge_loss_op.cc 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/hinge_loss_op.cu 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/hinge_loss_op.h 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/huber_loss_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/huber_loss_op.cu
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/huber_loss_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/increment_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/is_empty_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/l1_norm_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/l1_norm_op.cu
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/l1_norm_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/load_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/lod_reset_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/lod_reset_op.cu
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/lod_reset_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/log_loss_op.cc 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/log_loss_op.cu 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/log_loss_op.h 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/logical_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/logical_op.cu
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/logical_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/lookup_table_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/lookup_table_op.cu
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/lookup_table_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/lrn_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/lrn_op.cu
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/lrn_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/lstm_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/lstm_op.cu.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/lstm_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/lstm_unit_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/lstm_unit_op.cu
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/lstm_unit_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/math/cos_sim_functor.cc 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/math/cos_sim_functor.cu 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/math/cos_sim_functor.h 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/math/detection_util.h 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/math/gru_compute.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/math/im2col.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/math/im2col.cu
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/math/im2col.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/math/matmul.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/math/maxouting.cc 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/math/maxouting.cu 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/math/maxouting.h 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/math/pooling.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/math/pooling.cu
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/math/pooling.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/math/softmax.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/math/softmax.cu
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/math/softmax.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/math/unpooling.cc 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/math/unpooling.cu 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/math/unpooling.h 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/math/vol2col.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/math/vol2col.cu
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/math/vol2col.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/matmul_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/matmul_op.cu.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/matmul_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/max_sequence_len_op.cc 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/maxout_op.cc 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/maxout_op.cu.cc 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/maxout_op.h 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/mean_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/mean_op.cu
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/mean_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/minus_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/minus_op.cu
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/minus_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/momentum_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/momentum_op.cu
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/momentum_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/mul_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/mul_op.cu.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/mul_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/multiplex_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/multiplex_op.cu
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/multiplex_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/nccl_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/nccl_op.cu.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/nccl_op_test.cu.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/nce_op.cc 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/nce_op.h 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/net_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/net_op_test.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/op_documentation/op_markdown_format.md 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/pad_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/pad_op.cu
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/pad_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/pool_cudnn_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/pool_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/pool_op.cu.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/pool_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/prelu_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/prelu_op.cu
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/prelu_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/proximal_gd_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/proximal_gd_op.cu
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/proximal_gd_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/rank_loss_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/rank_loss_op.cu
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/rank_loss_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/recurrent_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/recv_op.cc 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/reduce_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/reduce_op.cu
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/reduce_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/reorder_lod_tensor_by_rank_op.cc 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/reshape_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/reshape_op.cu 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/reshape_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/rmsprop_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/rmsprop_op.cu
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/rmsprop_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/rnn/recurrent_op_utils.cc 已删除 100644 → 0
浏览文件 @ 1ce2d613
此差异已折叠。
paddle/operators/rnn/recurrent_op_utils.h 已删除 100644 → 0
浏览文件 @ 1ce2d613
此差异已折叠。
paddle/operators/roi_pool_op.cc 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/roi_pool_op.cu 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/roi_pool_op.h 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/row_conv_op.cc 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/row_conv_op.cu 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/row_conv_op.h 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/save_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/scale_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/scale_op.cu
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/scale_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/scatter.cu.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/scatter_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/scatter_op.cu
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/send_op.cc 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/send_recv_op_test.cc 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/seq_expand_op.cu 已删除 100644 → 0
浏览文件 @ 1ce2d613
此差异已折叠。
paddle/operators/sequence_conv_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/sequence_expand_op.cu 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/sequence_pool_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/sequence_slice_op.cc 100755 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/sequence_slice_op.h 100755 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/sgd_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/sgd_op.cu
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/sgd_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/sign_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/sign_op.cu
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/sign_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/softmax_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/softmax_op.cu.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/softmax_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/split_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/split_op.cu.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/split_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/spp_op.cc 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/spp_op.cu.cc 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/spp_op.h 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/strided_memcpy.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/sum_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/sum_op.cu
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/sum_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/tensor.save 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/top_k_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/top_k_op.cu
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/top_k_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/transpose_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/transpose_op.cu.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/transpose_op.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/unpool_op.cc 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/unpool_op.cu.cc 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/unpool_op.h 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/operators/while_op.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/optimizer/adam_optimizer.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/optimizer/adam_optimizer.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/optimizer/optimizer.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/optimizer/optimizer.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/optimizer/sgd_optimizer.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/optimizer/sgd_optimizer.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/platform/CMakeLists.txt
浏览文件 @ fab63cc6
此差异已折叠。
paddle/platform/call_once.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/platform/cuda_helper.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/platform/cuda_profiler.h 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/platform/cudnn_helper.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/platform/device_context.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/platform/device_context.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/platform/dynload/cudnn.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/platform/dynload/cudnn.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/platform/dynload/nccl.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/platform/dynload/nccl.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/platform/enforce.cc 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/platform/enforce.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/platform/for_range.h 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/platform/gpu_info.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/platform/gpu_info.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/platform/nccl_test.cu
浏览文件 @ fab63cc6
此差异已折叠。
paddle/platform/place.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/platform/place.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/platform/place_test.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/platform/transform.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/platform/transform_test.cu
浏览文件 @ fab63cc6
此差异已折叠。
paddle/platform/variant.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/pserver/CMakeLists.txt
浏览文件 @ fab63cc6
此差异已折叠。
paddle/pybind/CMakeLists.txt
浏览文件 @ fab63cc6
此差异已折叠。
paddle/pybind/const_value.cc 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/pybind/const_value.h 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/pybind/exception.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/pybind/exception.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/pybind/protobuf.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/pybind/pybind.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/pybind/tensor_py.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/scripts/CMakeLists.txt
浏览文件 @ fab63cc6
此差异已折叠。
paddle/scripts/check_env.sh 0 → 100755
浏览文件 @ fab63cc6
此差异已折叠。
paddle/scripts/docker/README.md
浏览文件 @ fab63cc6
此差异已折叠。
paddle/scripts/docker/build.sh
浏览文件 @ fab63cc6
此差异已折叠。
paddle/scripts/submit_local.sh.in
浏览文件 @ fab63cc6
此差异已折叠。
paddle/scripts/tools/usage_stat/usage.sh 已删除 100755 → 0
浏览文件 @ 1ce2d613
此差异已折叠。
paddle/scripts/travis/build_doc.sh
浏览文件 @ fab63cc6
此差异已折叠。
paddle/string/to_string.h
浏览文件 @ fab63cc6
此差异已折叠。
paddle/string/to_string_test.cc
浏览文件 @ fab63cc6
此差异已折叠。
paddle/testing/CMakeLists.txt
浏览文件 @ fab63cc6
此差异已折叠。
paddle/testing/paddle_gtest_main.cc 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
paddle/trainer/CMakeLists.txt
浏览文件 @ fab63cc6
此差异已折叠。
paddle/trainer/Trainer.cpp
浏览文件 @ fab63cc6
此差异已折叠。
paddle/trainer/tests/CMakeLists.txt
浏览文件 @ fab63cc6
此差异已折叠。
paddle/trainer/tests/chunking.conf 已删除 100644 → 0
浏览文件 @ 1ce2d613
此差异已折叠。
paddle/trainer/tests/compare_sparse_data 已删除 100644 → 0
浏览文件 @ 1ce2d613
此差异已折叠。
paddle/trainer/tests/data_bin_part 已删除 100644 → 0
浏览文件 @ 1ce2d613
此差异已折叠。
paddle/trainer/tests/gen_proto_data.py 已删除 100644 → 0
浏览文件 @ 1ce2d613
此差异已折叠。
paddle/trainer/tests/test.txt 已删除 100644 → 0
浏览文件 @ 1ce2d613
此差异已折叠。
paddle/trainer/tests/test_files.txt 已删除 100644 → 0
浏览文件 @ 1ce2d613
此差异已折叠。
paddle/trainer/tests/train.list 已删除 100644 → 0
浏览文件 @ 1ce2d613
此差异已折叠。
paddle/trainer/tests/train.txt 已删除 100644 → 0
浏览文件 @ 1ce2d613
此差异已折叠。
paddle/trainer/tests/train_files.txt 已删除 100644 → 0
浏览文件 @ 1ce2d613
此差异已折叠。
paddle/trainer/tests/train_sparse.list 已删除 100644 → 0
浏览文件 @ 1ce2d613
此差异已折叠。
paddle/utils/Flags.cpp
浏览文件 @ fab63cc6
此差异已折叠。
proto/ModelConfig.proto
浏览文件 @ fab63cc6
此差异已折叠。
python/.gitignore
浏览文件 @ fab63cc6
此差异已折叠。
python/CMakeLists.txt
浏览文件 @ fab63cc6
此差异已折叠。
python/paddle/__init__.py
浏览文件 @ fab63cc6
此差异已折叠。
python/paddle/v2/__init__.py
浏览文件 @ fab63cc6
此差异已折叠。
python/paddle/v2/dataset/common.py
浏览文件 @ fab63cc6
此差异已折叠。
python/paddle/v2/dataset/flowers.py
浏览文件 @ fab63cc6
此差异已折叠。
python/paddle/v2/dataset/imdb.py
浏览文件 @ fab63cc6
此差异已折叠。
python/paddle/v2/fluid/__init__.py
浏览文件 @ fab63cc6
此差异已折叠。
python/paddle/v2/fluid/backward.py
浏览文件 @ fab63cc6
此差异已折叠。
python/paddle/v2/fluid/clip.py 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
python/paddle/v2/fluid/data_feeder.py 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
python/paddle/v2/fluid/distribute_transpiler.py 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
python/paddle/v2/fluid/evaluator.py
浏览文件 @ fab63cc6
此差异已折叠。
python/paddle/v2/fluid/executor.py
浏览文件 @ fab63cc6
此差异已折叠。
python/paddle/v2/fluid/framework.py
浏览文件 @ fab63cc6
此差异已折叠。
python/paddle/v2/fluid/io.py
浏览文件 @ fab63cc6
此差异已折叠。
python/paddle/v2/fluid/layers/__init__.py 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
python/paddle/v2/fluid/layers/io.py 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
python/paddle/v2/fluid/layers/nn.py 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
python/paddle/v2/fluid/layers/ops.py 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
python/paddle/v2/fluid/layers/tensor.py 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
python/paddle/v2/fluid/nets.py
浏览文件 @ fab63cc6
此差异已折叠。
python/paddle/v2/fluid/optimizer.py
浏览文件 @ fab63cc6
此差异已折叠。
python/paddle/v2/fluid/param_attr.py 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
python/paddle/v2/fluid/profiler.py 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
python/paddle/v2/fluid/registry.py 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
python/paddle/v2/fluid/tests/decorators.py 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
python/paddle/v2/fluid/tests/demo/fc_gan.py 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
python/paddle/v2/fluid/tests/test_const_value.py 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
python/paddle/v2/fluid/tests/test_data_feeder.py 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
python/paddle/v2/fluid/tests/test_dyn_rnn.py 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
python/paddle/v2/fluid/tests/test_fill_op.py 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
python/paddle/v2/fluid/tests/test_ftrl_op.py 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
python/paddle/v2/fluid/tests/test_hinge_loss_op.py 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
python/paddle/v2/fluid/tests/test_log_loss_op.py 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
python/paddle/v2/fluid/tests/test_maxout_op.py 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
python/paddle/v2/fluid/tests/test_mnist_if_else_op.py 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
python/paddle/v2/fluid/tests/test_nce.py 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
python/paddle/v2/fluid/tests/test_profiler.py 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
python/paddle/v2/fluid/tests/test_registry.py 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
python/paddle/v2/fluid/tests/test_reorder_lod_tensor.py 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
python/paddle/v2/fluid/tests/test_roi_pool_op.py 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
python/paddle/v2/fluid/tests/test_row_conv_op.py 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
python/paddle/v2/fluid/tests/test_sequence_slice_op.py 100755 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
python/paddle/v2/fluid/tests/test_spp_op.py 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
python/paddle/v2/fluid/tests/test_unpool_op.py 0 → 100644
浏览文件 @ fab63cc6
此差异已折叠。
python/paddle/v2/parameters.py
浏览文件 @ fab63cc6
此差异已折叠。
python/setup.py.in
浏览文件 @ fab63cc6
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