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 1e8300fa 编写于 作者: T tensor-tang
浏览文件
操作

Merge remote-tracking branch 'upstream/develop' into merge

上级 3eb42bfd 8401039f
展开全部 隐藏空白更改
内联 并排
Showing with 4064 addition and 1629 deletion
CONTRIBUTING.md
浏览文件 @ 1e8300fa
./doc/howto/dev/contribute_to_paddle_en.md
# Contribute Code
We sincerely appreciate your contribution. This document explains our workflow and work style.
## Workflow
PaddlePaddle uses this [Git branching model](http://nvie.com/posts/a-successful-git-branching-model/). The following steps guide usual contributions.
1. Fork
Our development community has been growing fastly; it doesn't make sense for everyone to write into the official repo. So, please file Pull Requests from your fork. To make a fork, just head over to the GitHub page and click the ["Fork" button](https://help.github.com/articles/fork-a-repo/).
1. Clone
To make a copy of your fork to your local computers, please run
```bash
git clone https://github.com/your-github-account/paddle
cd paddle
```
1. Create the local feature branch
For daily works like adding a new feature or fixing a bug, please open your feature branch before coding:
```bash
git checkout -b my-cool-stuff
```
1. Commit
Before issuing your first `git commit` command, please install [`pre-commit`](http://pre-commit.com/) by running the following commands:
```bash
pip install pre-commit
pre-commit install
```
Our pre-commit configuration requires clang-format 3.8 for auto-formating C/C++ code and yapf for Python.
Once installed, `pre-commit` checks the style of code and documentation in every commit. We will see something like the following when you run `git commit`:
```
➜ git commit
CRLF end-lines remover...............................(no files to check)Skipped
yapf.................................................(no files to check)Skipped
Check for added large files..............................................Passed
Check for merge conflicts................................................Passed
Check for broken symlinks................................................Passed
Detect Private Key...................................(no files to check)Skipped
Fix End of Files.....................................(no files to check)Skipped
clang-formater.......................................(no files to check)Skipped
[my-cool-stuff c703c041] add test file
1 file changed, 0 insertions(+), 0 deletions(-)
create mode 100644 233
```
1. Build and test
Users can build PaddlePaddle natively on Linux and Mac OS X. But to unify the building environment and to make it easy for debugging, the recommended way is [using Docker](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/dev/build_en.md).
1. Keep pulling
An experienced Git user pulls from the official repo often -- daily or even hourly, so they notice conflicts with others work early, and it's easier to resolve smaller conflicts.
```bash
git remote add upstream https://github.com/PaddlePaddle/Paddle
git pull upstream develop
```
1. Push and file a pull request
You can "push" your local work into your forked repo:
```bash
git push origin my-cool-stuff
```
The push allows you to create a pull request, requesting owners of this [official repo](https://github.com/PaddlePaddle/Paddle) to pull your change into the official one.
To create a pull request, please follow [these steps](https://help.github.com/articles/creating-a-pull-request/).
If your change is for fixing an issue, please write ["Fixes <issue-URL>"](https://help.github.com/articles/closing-issues-using-keywords/) in the description section of your pull request. Github would close the issue when the owners merge your pull request.
Please remember to specify some reviewers for your pull request. If you don't know who are the right ones, please follow Github's recommendation.
1. Delete local and remote branches
To keep your local workspace and your fork clean, you might want to remove merged branches:
```bash
git push origin :my-cool-stuff
git checkout develop
git pull upstream develop
git branch -d my-cool-stuff
```
### Code Review
- Please feel free to ping your reviewers by sending them the URL of your pull request via IM or email. Please do this after your pull request passes the CI.
- Please answer reviewers' every comment. If you are to follow the comment, please write "Done"; please give a reason otherwise.
- If you don't want your reviewers to get overwhelmed by email notifications, you might reply their comments by [in a batch](https://help.github.com/articles/reviewing-proposed-changes-in-a-pull-request/).
- Reduce the unnecessary commits. Some developers commit often. It is recommended to append a sequence of small changes into one commit by running `git commit --amend` instead of `git commit`.
## Coding Standard
### Code Style
Our C/C++ code follows the [Google style guide](http://google.github.io/styleguide/cppguide.html).
Our Python code follows the [PEP8 style guide](https://www.python.org/dev/peps/pep-0008/).
Our build process helps to check the code style. In [`build.sh`](https://github.com/PaddlePaddle/Paddle/blob/b84e8226514b8bb4405c3c28e54aa5077193d179/paddle/scripts/docker/build.sh#L42), the entry point of our [builder Docker image](https://github.com/PaddlePaddle/Paddle/blob/b84e8226514b8bb4405c3c28e54aa5077193d179/Dockerfile#L88), the CMake argument `WITH_STYLE_CHECK` is set to `ON` by default. This flag is on
Please install pre-commit, which automatically reformat the changes to C/C++ and Python code whenever we run `git commit`. To check the whole codebase, we can run the command `pre-commit run -a`, as in the [`check_style.sh` file](https://github.com/PaddlePaddle/Paddle/blob/b84e8226514b8bb4405c3c28e54aa5077193d179/paddle/scripts/travis/check_style.sh#L30), which is invoked by [our Travis CI configuration](https://github.com/PaddlePaddle/Paddle/blob/b84e8226514b8bb4405c3c28e54aa5077193d179/.travis.yml#L43).
### Unit Tests
Please remember to add related unit tests.
- For C/C++ code, please follow [`google-test` Primer](https://github.com/google/googletest/blob/master/googletest/docs/Primer.md).
- For Python code, please use [Python's standard `unittest` package](http://pythontesting.net/framework/unittest/unittest-introduction/).
### Writing Logs
We use [glog](https://github.com/google/glog) for logging in our C/C++ code.
For general information, please use `LOG`. For debug information, please use [`VLOG`](http://htmlpreview.github.io/?https://github.com/google/glog/blob/master/doc/glog.html#verbose). The reason is at [here](https://groups.google.com/a/chromium.org/d/msg/chromium-dev/3NDNd1KzXeY/AZKMMx37fdQJ).
`VLOG` requires a *verbose level* parameter. For example:
```c++
VLOG(3) << "Operator FC is taking " << num_inputs << "inputs."
```
When we run a PaddlePaddle application or test, we can specify a verbose threshold. For example:
```bash
GLOG_vmodule=buddy_allocator=2 \
GLOG_v=10 \
python \
../python/paddle/v2/framework/tests/test_recurrent_op.py
```
This will enable VLOG messages generated by `buddy_allocator.{h,cc}` and in the verbose range of 0 to 3, so you will see above example VLOG message, which is in level 3. This suggests that we output overall messages in lower verbose levels, so they display with higher probability. When coding C++, please follow the verbose level convention as follows:
- verbose level 1: [framework](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/framework)
- verbose level 3: [operators](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators)
- verbose level 5: [memory](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/memory), [platform](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/platform)
- verbose level 7: [math](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/math)
doc/howto/cross_compiling/cross_compiling_for_ios_cn.md 0 → 100644
浏览文件 @ 1e8300fa
# 构建iOS平台上的PaddlePaddle库
交叉编译iOS平台上适用的PaddlePaddle库，需要在MacOS系统上进行。本文的将介绍在MacOS上，从源码交叉编译iOS平台上适用的PaddlePaddle库。
## 准备交叉编译环境
Apple官方为iOS开发提供了完整的交叉编译工具和集成开发环境，用户从App Store下载安装Xcode即可。也可自行前往官网下载，[Xcode](https://developer.apple.com/cn/xcode/)。安装完成之后，可在命令行执行`xcodebuild -version`，判断是否安装成功。
```bash
$ xcodebuild -version
Xcode 9.0
Build version 9A235
```
## 配置交叉编译参数
PaddlePaddle为交叉编译提供了工具链配置文档[cmake/cross_compiling/ios.cmake](https://github.com/PaddlePaddle/Paddle/blob/develop/cmake/cross_compiling/ios.cmake)，以提供一些默认的编译器和编译参数配置。
交叉编译iOS版本的PaddlePaddle库时，有一些必须配置的参数：
- `CMAKE_SYSTEM_NAME`，CMake编译的目标平台，必须设置为`iOS`。在设置`CMAKE_SYSTEM_NAME=iOS`后，PaddlePaddle的CMake系统会自动编译所有的第三方依赖库，并且强制设置一些PaddlePaddle参数的值（`WITH_C_API=ON`、`WITH_GPU=OFF`、`WITH_AVX=OFF`、`WITH_PYTHON=OFF`、`WITH_RDMA=OFF`）。
- `WITH_C_API`，是否编译C-API预测库，必须设置为ON。在iOS平台上只支持使用C-API来预测。
- `WITH_SWIG_PY`，必须设置为ON。在iOS平台上不支持通过swig调用来训练或者预测。
iOS平台可选配置参数：
- `IOS_PLATFORM`，可设置为`OS/SIMULATOR`，默认值为`OS`。
- `OS`，构建目标为`arm`架构的iPhone或者iPad等物理设备。
- `SIMULATOR`，构建目标为`x86`架构的模拟器平台。
- `IOS_ARCH`，目标架构。针对不同的`IOS_PLATFORM`，可设置的目标架构如下表所示：
| IOS_PLATFORM | IOS_ARCH |
|--------------|----------------------|
| OS | armv7, armv7s, arm64 (默认) |
| SIMULATOR | i386, x86_64 (默认) |
- `IOS_DEPLOYMENT_TARGET`，最小的iOS部署版本，默认值为`7.0`。
- `IOS_ENABLE_BITCODE`，是否使能[Bitcode](https://developer.apple.com/library/content/documentation/IDEs/Conceptual/AppDistributionGuide/AppThinning/AppThinning.html#//apple_ref/doc/uid/TP40012582-CH35-SW3)，可设置`ON/OFF`，默认值为`ON`。
- `IOS_USE_VECLIB_FOR_BLAS`，是否使用[vecLib](https://developer.apple.com/documentation/accelerate/veclib)框架进行BLAS矩阵计算，可设置`ON/OFF`，默认值为`OFF`。
- `IOS_DEVELOPMENT_ROOT`，`Developer`目录，可显式指定为`/path/to/platform/Developer`。若未显式指定，PaddlePaddle将会根据`IOS_PLATFORM`自动选择`Xcode`对应`platform`的`Developer`目录。
- `IOS_SDK_ROOT`，所使用`SDK`的根目录，可显式指定为`/path/to/platform/Developer/SDKs/SDK`。若未显式指定，PaddlePaddle将会自动选择`IOS_DEVELOPMENT_ROOT`目录下最新的`SDK`版本。
其他配置参数：
- `USE_EIGEN_FOR_BLAS`，是否使用Eigen库进行矩阵计算，在`IOS_USE_VECLIB_FOR_BLAS=OFF`时有效。可设置`ON/OFF`，默认值为`OFF`。
- `HOST_C/CXX_COMPILER`，宿主机的C/C++编译器。默认值为环境变量`CC/CXX`的值；若环境变量`CC/CXX`未设置，则使用`cc/c++`编译器。
常用的cmake配置如下：
```bash
cmake -DCMAKE_SYSTEM_NAME=iOS \
-DIOS_PLATFORM=OS \
-DIOS_ARCH="arm64" \
-DIOS_ENABLE_BITCODE=ON \
-DIOS_USE_VECLIB_FOR_BLAS=ON \
-DCMAKE_INSTALL_PREFIX=your/path/to/install \
-DWITH_C_API=ON \
-DWITH_TESTING=OFF \
-DWITH_SWIG_PY=OFF \
..
```
```bash
cmake -DCMAKE_SYSTEM_NAME=iOS \
-DIOS_PLATFORM=SIMULATOR \
-DIOS_ARCH="x86_64" \
-DIOS_USE_VECLIB_FOR_BLAS=ON \
-DCMAKE_INSTALL_PREFIX=your/path/to/install \
-DWITH_C_API=ON \
-DWITH_TESTING=OFF \
-DWITH_SWIG_PY=OFF \
..
```
用户还可根据自己的需求设置其他编译参数。比如希望最小化生成库的大小，可以设置`CMAKE_BUILD_TYPE`为`MinSizeRel`；若希望得到最快的执行速度，则可设置`CMAKE_BUILD_TYPE`为`Release`。亦可以通过手动设置`CMAKE_C/CXX_FLAGS`来影响PaddlePaddle的编译过程。
**性能TIPS**，为了达到最快的计算速度，在CMake参数配置上，有以下建议：
- 设置`CMAKE_BUILD_TYPE`为`Release`
- 设置`IOS_USE_VECLIB_FOR_BLAS=ON`，调用`vecLib`框架提供的BLAS函数进行矩阵计算。
## 编译和安装
CMake配置完成后，执行以下命令，PaddlePaddle将自动下载和编译所有第三方依赖库、编译和安装PaddlePaddle预测库。
```
$ make
$ make install
```
注意：如果你曾在源码目录下编译过其他平台的PaddlePaddle库，请先使用`rm -rf`命令删除`third_party`目录和`build`目录，以确保所有的第三方依赖库和PaddlePaddle代码都是针对新的CMake配置重新编译的。
执行完安装命令后，`your/path/to/install`目录中会包含以下内容：
- `include`目录，其中包含所有C-API的头文件
- `lib`目录，其中包含PaddlePaddle的C-API静态库
- `third_party`目录，其中包含所依赖的所有第三方库
注意，不同架构的PaddlePaddle库建议安装到不同的目录下，然后使用`lipo`工具将多个静态库合并成一个支持多个架构的fat库。
自此，PaddlePaddle库已经安装完成，用户可将合成的fat库用于深度学习相关的iOS App中，调用方法见C-API文档。
doc/howto/cross_compiling/cross_compiling_for_raspberry_cn.md
浏览文件 @ 1e8300fa
# 构建Raspberry Pi平台上的PaddlePaddle库
对于Rasspberry Pi系统，用户可通过ssh等方式登录到Raspberry Pi系统上，按照[源码编译PaddlePaddle](http://www.paddlepaddle.org/doc_cn/getstarted/build_and_install/cmake/build_from_source_cn.html)相关文档所述，直接编译Raspberry Pi平台上适用的PaddlePaddle库。
通常有两个方法来构建基于 Rasspberry Pi 的版本：
用户也可以在自己熟悉的开发平台上，通过交叉编译的方式来编译。这篇文档将以Linux x86-64平台为例，介绍交叉编译Raspberry Pi平台上适用的PaddlePaddle的方法和步骤。
1. 通过ssh等方式登录到Raspberry Pi系统上来构建。所需的开发工具和第三方库可以参考 [`/Dockerfile`](https://github.com/PaddlePaddle/Paddle/blob/develop/Dockerfile)。
## 准备交叉编译环境
1. 另一个方法是交叉编译。这篇文档介绍在 Linux/x64 上交叉编译Raspberry Pi平台上适用的PaddlePaddle的方法和步骤。
从源码交叉编译PaddlePaddle，用户需要提前准备好交叉编译环境。用户可自行前往[github](https://github.com/raspberrypi/tools)下载Raspberry Pi平台使用的C/C++交叉编译工具链，也可通过以下命令获取：
## 安装交叉编译器
克隆下面 Github repo
```bash
git clone https://github.com/raspberrypi/tools.git
```
该github仓库中包含若干个预编译好的、针对不同平台的编译工具。宿主机是Linux x86-64环境，则需选用`arm-bcm2708/gcc-linaro-arm-linux-gnueabihf-raspbian-x64`下的作为编译工具，所使用的编译器为arm-linux-gnueabihf-gcc 4.8.3。
注意，该编译工具链需要系统glibc支持2.14以上。
即可在 `./tools/tree/master/arm-bcm2708/gcc-linaro-arm-linux-gnueabihf-raspbian-x64` 目录里找到交叉编译器 arm-linux-gnueabihf-gcc 4.8.3。运行该编译工具链需要一台 Linux x64 机器上以及 2.14版本以上的 glibc。
## 配置交叉编译参数
CMake系统对交叉编译提供了支持[cmake-toolchains](https://cmake.org/cmake/help/v3.0/manual/cmake-toolchains.7.html#cross-compiling)。为了简化cmake配置，PaddlePaddle为交叉编译提供了工具链配置文档[cmake/cross_compiling/raspberry_pi.cmake](https://github.com/PaddlePaddle/Paddle/blob/develop/cmake/cross_compiling/raspberry_pi.cmake)，以提供一些默认的编译器和编译参数相关配置。
CMake[支持交叉编译](https://cmake.org/cmake/help/v3.0/manual/cmake-toolchains.7.html#cross-compiling)。PaddlePaddle for Raspberry Pi的配置信息在[cmake/cross_compiling/raspberry_pi.cmake](https://github.com/PaddlePaddle/Paddle/blob/develop/cmake/cross_compiling/raspberry_pi.cmake)。
交叉编译Raspberry Pi版本PaddlePaddle库时，有一些必须配置的参数：
- `CMAKE_SYSTEM_NAME`，CMake编译的目标平台，必须配置为`RPi`。在设置`CMAKE_SYSTEM_NAME=RPi`后，PaddlePaddle的CMake系统才认为在是在交叉编译Raspberry Pi系统的版本，并自动编译宿主机版protoc可执行文件、目标机版protobuf库、以及目标机版OpenBLAS库。
Raspberry Pi平台可选配置参数：
- `CMAKE_SYSTEM_NAME`：CMake编译的目标平台，必须配置为`RPi`。在设置`CMAKE_SYSTEM_NAME=RPi`后，PaddlePaddle的CMake系统才认为在是在交叉编译Raspberry Pi系统的版本，并自动编译宿主机版protoc可执行文件、目标机版protobuf库、以及目标机版OpenBLAS库。
- `RPI_TOOLCHAIN`，编译工具链所在的绝对路径，或者相对于构建目录的相对路径。PaddlePaddle的CMake系统将根据该值自动设置需要使用的交叉编译器；否则，用户需要在cmake时手动设置这些值。无默认值。
- `RPI_ARM_NEON`，是否使用NEON指令。目前必须设置成`ON`，默认值为`ON`。
- `RPI_TOOLCHAIN`：编译工具链所在的绝对路径，或者相对于构建目录的相对路径。PaddlePaddle的CMake系统将根据该值自动设置需要使用的交叉编译器；否则，用户需要在cmake时手动设置这些值。无默认值。
其他配置参数：
- `RPI_ARM_NEON`：是否使用NEON指令。目前必须设置成`ON`，默认值为`ON`。
- `HOST_C/CXX_COMPILER`，宿主机的C/C++编译器。在编译宿主机版protoc可执行文件和目标机版OpenBLAS库时需要用到。默认设置成环境变量`CC`的值；若环境变量`CC`没有设置，则设置成`cc`编译器。
cmake参数如下；
一个常用的CMake配置如下：
```
cmake -DCMAKE_SYSTEM_NAME=RPi \
......@@ -47,7 +44,9 @@ cmake -DCMAKE_SYSTEM_NAME=RPi \
..
```
用户还可根据自己的需求设置其他编译参数。比如希望最小化生成的库的大小，可以设置`CMAKE_BUILD_TYPE`为`MinSizeRel`；若希望最快的执行速度，则可设置`CMAKE_BUILD_TYPE`为`Release`。亦可以通过手动设置`CMAKE_C/CXX_FLAGS_MINSIZEREL/RELEASE`来影响PaddlePaddle的编译过程。
其中`WITH_C_API=ON`表示需要构建推理库。
用户还可根据自己的需求设置其他编译参数。比如希望最小化生成的库的大小，可以设置`CMAKE_BUILD_TYPE`为`MinSizeRel`；若希望最快的执行速度，则可设置`CMAKE_BUILD_TYPE`为`Release`。
## 编译和安装
......@@ -60,6 +59,4 @@ make install
注意：如果你曾经在源码目录下编译过其他平台的PaddlePaddle库，请先使用`rm -rf`命令删除`third_party`目录和`build`目录，以确保所有的第三方依赖库和PaddlePaddle代码都是针对新的CMake配置重新编译的。
执行完安装命令后，由于上一步cmake配置中`WITH_C_API`设置为`ON`，`your/path/to/install`目录中会包含`include`和`lib`目录，其中`include`中包含C-API的头文件，`lib`中包含一个Raspberry Pi版本的库。
更多的编译配置见[源码编译PaddlePaddle](http://www.paddlepaddle.org/doc_cn/getstarted/build_and_install/cmake/build_from_source_cn.html)相关文档。
执行完安装命令后，`your/path/to/install`目录中会包含`include`和`lib`目录，其中`include`中包含C-API的头文件，`lib`中包含一个Raspberry Pi版本的库。
doc/howto/cross_compiling/cross_compiling_for_raspberry_en.md 0 → 100644
浏览文件 @ 1e8300fa
# Build PaddlePaddle for Raspberry Pi
You may use any of the following two approaches to build the inference library of PaddlePaddle for Raspberry Pi:
1. Build using SSH: Log in to a Raspberry Pi using SSH and build the library. The required development tools and third-party dependencies are listed in here: [`/Dockerfile`](https://github.com/PaddlePaddle/Paddle/blob/develop/Dockerfile).
1. Cross-compile: We talk about how to cross-compile PaddlePaddle for Raspberry Pi on a Linux/x64 machine, in more detail in this article.
## The Cross-Compiling Toolchain
Step 1. Clone the Github repo by running the following command.
```bash
git clone https://github.com/raspberrypi/tools.git
```
Step 2. Use the pre-built cross-compiler found in `./tools/tree/master/arm-bcm2708/gcc-linaro-arm-linux-gnueabihf-raspbian-x64`. To run it on a Linux computer, glibc version >= 2.14 is needed.
## CMake Arguments
CMake supports [cross-compiling](https://cmake.org/cmake/help/v3.0/manual/cmake-toolchains.7.html#cross-compiling). All CMake configuration arguments required for the cross-compilation for Raspberry Pi can be found in [`cmake/cross_compiling/raspberry_pi.cmake`](https://github.com/PaddlePaddle/Paddle/blob/develop/cmake/cross_compiling/raspberry_pi.cmake).
Some important arguments that need to be set:
- `CMAKE_SYSTEM_NAME`: The target platform. Must be `RPi`.
- `RPI_TOOLCHAIN`: The absolute path of the cross-compiling toolchain.
- `RPI_ARM_NEON`: Use ARM NEON Intrinsics. This is a required argument and set default to `ON`.
- `HOST_C/CXX_COMPILER`: The C/C++ compiler for the host. It is used to build building tools running on the host, for example, protoc.
A commonly-used CMake configuration is as follows:
```
cmake -DCMAKE_SYSTEM_NAME=RPi \
-DRPI_TOOLCHAIN=your/path/to/arm-bcm2708/gcc-linaro-arm-linux-gnueabihf-raspbian-x64 \
-DRPI_ARM_NEON=ON \
-DCMAKE_INSTALL_PREFIX=your/path/to/install \
-DWITH_GPU=OFF \
-DWITH_C_API=ON \
-DWITH_PYTHON=OFF \
-DWITH_SWIG_PY=OFF \
..
```
To build the inference library, please set the argument WITH\_C\_API to ON: `WITH_C_API=ON`.
You can add more arguments. For example, to minimize the size of the generated inference library, you may use `CMAKE_BUILD_TYPE=MinSizeRel`. For performance optimization, you may use `CMAKE_BUILD_TYPE=Release`.
## Build and Install
The following commands build the inference library of PaddlePaddle for Raspberry Pi and third-party dependencies.
```bash
make
make install
```
The intermediate files will be stored in `build`. Third-party libraries will be located in `build/third_party`. If you have already built it for other platforms like Android or iOS, you may want to clear these directories by running the command: `rm -rf build`.
The infernece library will be in `your/path/to/install/lib`, with related header files in `your/path/to/install/include`.
doc/howto/dev/contribute_to_paddle_en.md 已删除 100644 → 0
浏览文件 @ 3eb42bfd
# Contribute Code
We sincerely appreciate your contributions. You can use fork and pull request
workflow to merge your code.
## Code Requirements
- Your code comments must be fully documented by
[Doxygen](http://www.stack.nl/~dimitri/doxygen/) style.
- Make sure the compiler option `WITH_STYLE_CHECK` is on and the compiler
passes the code style check.
- All code must have unit test.
- Pass all unit tests.
The following tutorial guides you into submitting your contibution.
## [Creating a Fork](https://help.github.com/articles/fork-a-repo/)
Just head over to the GitHub page and click the "Fork" button.
It's just that simple.
## Clone
Clone remote repository.
```bash
➜ git clone https://github.com/USERNAME/Paddle
➜ cd Paddle
```
## Create a local branch
Paddle is currently using [Git-flow branching model](http://nvie.com/posts/a-successful-git-branching-model/).
All feature and bug fix development work should be done on a new branch, generally create new branch from `develop` branch .
```bash
➜ git checkout -b my-cool-stuff
```
Before the checkout, you need to keep the current branch directory clean, otherwise the untracked file will be brought to the new branch, which can be inspected by `git status`.
## Using `pre-commit` hook
Paddle developers use [pre-commit](http://pre-commit.com/) tool to manage git
pre-commit hooks. It can help us format source codes (cpp, python), check some
basic thing before commit (only one EOL for each file, do not add a huge file
in git). `pre-commit` tests is a part of unit tests in Travis-CI now, every
PR doesn't fit hook can not be merged into Paddle.
To use [pre-commit](http://pre-commit.com/), you should install it by
`pip install pre-commit`, and currently, Paddle uses `clang-format` to format
c/cpp sources. Please make sure clang-format 3.8+ installed.
Install and run it as follow:
```bash
➜ pip install pre-commit
➜ pre-commit install
```
When you commit your code, the pre-commit hook will check the local code if there is
anything not suitable to commit, and so on.
## Start to develop
In this tutorial, I delete a line in README.md and created a new file.
We can use `git status` to inspect the changes of current directory, `git diff` to see difference.
```bash
➜ git status
On branch test
Changes not staged for commit:
(use "git add <file>..." to update what will be committed)
(use "git checkout -- <file>..." to discard changes in working directory)
modified: README.md
Untracked files:
(use "git add <file>..." to include in what will be committed)
test
no changes added to commit (use "git add" and/or "git commit -a")
```
## Build and Test
We package PaddlePaddle's compile environment into a Docker image, called the develop image named `paddle:dev`, it contains all compiling tools that PaddlePaddle needs.
If you want to build the develop image, just run:
```bash
➜ docker build -t paddle:dev .
```
Then we can use the develop image to build PaddlePaddle source. For example:
```bash
➜ docker run -v $(pwd):/paddle -e "WITH_GPU=OFF" -e "WITH_AVX=ON" -e "WITH_TEST=ON" paddle:dev
```
The above command will compile PaddlePaddle and create a Dockerfile for building production image. All the generated files are in the build directory. "WITH_GPU" controls if the generated production image supports GPU. "WITH_AVX" controls if the generated production image supports AVX. "WITH_TEST" controls if the unit test will be generated.
Then we can generate the production image by copying the compiled PaddlePaddle program into the image by
```bash
➜ docker build -t paddle:prod -f build/Dockerfile .
```
Run unit test finally:
```bash
➜ docker run -it -v $(pwd):/paddle paddle:dev bash -c "cd /paddle/build && ctest"
```
For more details, you can read [this doc](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/getstarted/build_and_install/docker_install_en.rst).
## Commit
Next we cancel the changes to the README.md file and then commit our changes by following command lines:
```bash
➜ git checkout -- README.md
➜ git status
On branch test
Untracked files:
(use "git add <file>..." to include in what will be committed)
test
nothing added to commit but untracked files present (use "git add" to track)
➜ git add test
```
We should write a description of each commit by `git commit` to allow others to know
the changes in these files.
```bash
➜ git commit
CRLF end-lines remover...............................(no files to check)Skipped
yapf.................................................(no files to check)Skipped
Check for added large files..............................................Passed
Check for merge conflicts................................................Passed
Check for broken symlinks................................................Passed
Detect Private Key...................................(no files to check)Skipped
Fix End of Files.....................................(no files to check)Skipped
clang-formater.......................................(no files to check)Skipped
[my-cool-stuff c703c041] add test file
1 file changed, 0 insertions(+), 0 deletions(-)
create mode 100644 233
```
## Keeping Fork Up to Date
Before pull your request, you should sync your code from the latest PaddlePaddle.
To do this, you'll need to add a remote at first:
```bash
➜ git remote add upstream https://github.com/PaddlePaddle/Paddle
➜ git remote
origin
upstream
```
Update your fork with the latest upstream changes:
```bash
➜ git fetch upstream
➜ git pull upstream develop
```
Now, your local master branch is up-to-date with everything modified upstream.
## Push to GitHub
```bash
# push to your repository in Github
➜ git push origin my-cool-stuff
```
## Create an issue and a Pull Request
Create an Issue to describe the problem and record its number.
Go to the page for your fork on GitHub, select your development branch,
and click the `New pull request`.
<img width="295" alt="screen shot 2017-04-26 at 9 09 28 pm" src="https://cloud.githubusercontent.com/assets/11692045/25436054/a6d98c66-2ac4-11e7-9cb1-18dd13150230.png">
Then select the target branch:
<img width="750" alt="screen shot 2017-04-26 at 9 11 52 pm" src="https://cloud.githubusercontent.com/assets/11692045/25436139/f83b1e6c-2ac4-11e7-8c0e-add499023c46.png">
We can add `resolve #Issue number` in PR description to close the issue automatically after the PR is merge. More details in <https://help.github.com/articles/closing-issues-via-commit-messages/>.
Then wait for review, if there need to modify, refer to the above steps to update the corresponding origin branch.
## Delete origin branch
After the PR is merge into the main repository, we can delete the remote branch on the PR page.
<img width="775" alt="screen shot 2017-04-26 at 9 18 24 pm" src="https://cloud.githubusercontent.com/assets/11692045/25436457/e4cdd472-2ac5-11e7-9272-badc76c4a23e.png">
Or just run:
```bash
➜ git push origin :my-cool-stuff
```
## Delete local branch
Finally, we delete local branch:
```bash
➜ git checkout develop
# delete my-cool-stuff branch
➜ git branch -D my-cool-stuff
```
doc/howto/dev/contribute_to_paddle_en.md 0 → 120000
浏览文件 @ 1e8300fa
../../../CONTRIBUTING.md
\ No newline at end of file
doc/howto/index_cn.rst
浏览文件 @ 1e8300fa
......@@ -21,7 +21,6 @@
dev/build_cn.rst
dev/write_docs_cn.rst
dev/contribute_to_paddle_cn.md
模型配置
--------
......
doc/howto/usage/cluster/cluster_train_cn.md
浏览文件 @ 1e8300fa
......@@ -19,7 +19,7 @@
* [启动集群作业](#启动集群作业-1)
* [在Kubernetes集群中提交训练作业](#在kubernetes集群中提交训练作业)
# 概述
## 概述
本文将介绍如何使用PaddlePaddle在不同的集群框架下完成分布式训练。分布式训练架构如下图所示：
<img src="https://user-images.githubusercontent.com/13348433/31772175-5f419eca-b511-11e7-9db7-5231fe3d9ccb.png" width="500">
......@@ -32,7 +32,7 @@
在使用同步SGD训练神经网络时，PaddlePaddle使用同步屏障（barrier），使梯度的提交和参数的更新按照顺序方式执行。在异步SGD中，则并不会等待所有trainer提交梯度才更新参数，这样极大地提高了计算的并行性：参数服务器之间不相互依赖，并行地接收梯度和更新参数，参数服务器也不会等待计算节点全部都提交梯度之后才开始下一步，计算节点之间也不会相互依赖，并行地执行模型的训练。可以看出，虽然异步SGD方式会提高参数更新并行度, 但是并不能保证参数同步更新，在任意时间某一台参数服务器上保存的参数可能比另一台要更新，与同步SGD相比，梯度会有噪声。
# 环境准备
## 环境准备
1. 准备您的计算集群。计算集群通常由一组（几台到几千台规模）的Linux服务器组成。服务器之间可以通过局域网（LAN）联通，每台服务器具有集群中唯一的IP地址（或者可被DNS解析的主机名）。集群中的每台计算机通常被成为一个“节点”。
1. 我们需要在集群的所有节点上安装 PaddlePaddle。 如果要启用GPU，还需要在节点上安装对应的GPU驱动以及CUDA。PaddlePaddle的安装可以参考[build_and_install](https://github.com/PaddlePaddle/Paddle/tree/develop/doc/getstarted/build_and_install)的多种安装方式。我们推荐使用[Docker](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/getstarted/build_and_install/docker_install_cn.rst)安装方式来快速安装PaddlePaddle。
......@@ -51,8 +51,8 @@ PaddlePaddle 0.10.0, compiled with
下面以`doc/howto/usage/cluster/src/word2vec`中的代码作为实例，介绍使用PaddlePaddle v2 API完成分布式训练。
# 启动参数说明
## 启动参数服务器
## 启动参数说明
### 启动参数服务器
执行以下的命令启动一个参数服务器并等待和计算节点的数据交互
```bash
$ paddle pserver --port=7164 --ports_num=1 --ports_num_for_sparse=1 --num_gradient_servers=1
......@@ -70,7 +70,7 @@ $ stdbuf -oL /usr/bin/nohup paddle pserver --port=7164 --ports_num=1 --ports_num
| ports_num_for_sparse | 必选 | 1 | 用于稀疏类型参数通信的端口个数 |
| num_gradient_servers | 必选 | 1 | 当前训练任务pserver总数 |
## 启动计算节点
### 启动计算节点
执行以下命令启动使用python编写的trainer程序（文件名为任意文件名，如train.py）
```bash
$ python train.py
......@@ -117,7 +117,7 @@ paddle.init(
| pservers | 必选 | 127.0.0.1 | 当前训练任务启动的pserver的IP列表，多个IP使用“,”隔开 |
## 准备数据集
### 准备数据集
参考样例数据准备脚本[prepare.py](https://github.com/PaddlePaddle/Paddle/tree/develop/doc/howto/usage/cluster/src/word2vec/prepare.py)，准备训练数据和验证数据集，我们使用paddle.dataset.imikolov数据集，并根据分布式训练并发数（trainer节点个数），在`prepare.py`开头部分指定`SPLIT_COUNT`将数据切分成多份。
......@@ -149,7 +149,7 @@ test.txt-00002
对于不同的训练任务，训练数据格式和训练程序的`reader()`会大不相同，所以开发者需要根据自己训练任务的实际场景完成训练数据的分割和`reader()`的编写。
## 准备训练程序
### 准备训练程序
我们会对每个训练任务都会在每个节点上创建一个工作空间（workspace），其中包含了用户的训练程序、程序依赖、挂载或下载的训练数据分片。
......@@ -184,7 +184,7 @@ test.txt-00002
- `train_data_dir`：包含训练数据的目录，可以是从分布式存储挂载过来的，也可以是在任务启动前下载到本地的。
- `test_data_dir`：包含测试数据集的目录。
# 使用分布式计算平台或工具
## 使用分布式计算平台或工具
PaddlePaddle可以使用多种分布式计算平台构建分布式计算任务，包括：
- [Kubernetes](http://kubernetes.io) Google开源的容器集群的调度框架，支持大规模集群生产环境的完整集群方案。
......@@ -195,12 +195,12 @@ PaddlePaddle可以使用多种分布式计算平台构建分布式计算任务
在使用分布式计算平台进行训练时，任务被调度在集群中时，分布式计算平台通常会通过API或者环境变量提供任务运行需要的参数，比如节点的ID、IP和任务节点个数等。
## 使用Fabric启动集群作业
### 使用Fabric启动集群作业
### 准备一个Linux集群
#### 准备一个Linux集群
可以在`paddle/scripts/cluster_train_v2/fabric/docker_cluster`目录下，执行`kubectl -f ssh_servers.yaml`启动一个测试集群，并使用`kubectl get po -o wide`获得这些节点的IP地址。
### 启动集群作业
#### 启动集群作业
`paddle.py` 提供了自动化脚本来启动不同节点中的所有 PaddlePaddle 集群进程。默认情况下，所有命令行选项可以设置为 `paddle.py` 命令选项并且 `paddle.py` 将透明、自动地将这些选项应用到 PaddlePaddle 底层进程。
......@@ -216,10 +216,10 @@ sh run.sh
集群作业将会在几秒后启动。
### 终止集群作业
#### 终止集群作业
`paddle.py`能获取`Ctrl + C` SIGINT 信号来自动终止它启动的所有进程。只需中断 `paddle.py` 任务来终止集群作业。如果程序崩溃你也可以手动终止。
### 检查集群训练结果
#### 检查集群训练结果
详细信息请检查 $workspace/log 里的日志，每一个节点都有相同的日志结构。
`paddle_trainer.INFO`
......@@ -234,13 +234,13 @@ sh run.sh
`train.log`
提供训练过程的 stderr 和 stdout。训练失败时可以检查错误日志。
### 检查模型输出
#### 检查模型输出
运行完成后，模型文件将被写入节点 0 的 `output` 目录中。
工作空间中的 `nodefile` 表示当前集群作业的节点 ID。
## 在OpenMPI集群中提交训练作业
### 在OpenMPI集群中提交训练作业
### 准备OpenMPI集群
#### 准备OpenMPI集群
执行下面的命令以启动3个节点的OpenMPI集群和一个"head"节点：
......@@ -252,7 +252,7 @@ kubectl create -f mpi-nodes.yaml
然后可以从head节点ssh无密码登录到OpenMPI的每个节点上。
### 启动集群作业
#### 启动集群作业
您可以按照下面的步骤在OpenMPI集群中提交paddle训练任务：
......@@ -280,6 +280,6 @@ scp train.txt-00002 test.txt-00002 [node3IP]:/home/tutorial
mpirun -hostfile machines -n 3 /home/tutorial/start_mpi_train.sh
```
## 在Kubernetes集群中提交训练作业
### 在Kubernetes集群中提交训练作业
此部分的使用方法可以参考[here](../k8s/k8s_distributed_cn.md)。
doc/howto/usage/cluster/cluster_train_en.md
浏览文件 @ 1e8300fa
......@@ -19,7 +19,7 @@
* [Launching Cluster Job](#launching-cluster-job-1)
* [Cluster Training Using Kubernetes](#cluster-training-using-kubernetes)
# Introduction
## Introduction
In this article, we'll explain how to run distributed training jobs with PaddlePaddle on different types of clusters. The diagram below shows the main architecture of a distributed trainning job:
......@@ -33,7 +33,7 @@ PaddlePaddle can support both synchronize stochastic gradient descent (SGD) and
When training with synchronize SGD, PaddlePaddle uses an internal "synchronize barrier" which makes gradients update and parameter download in strict order. On the other hand, asynchronous SGD won't wait for all trainers to finish upload at a single step, this will increase the parallelism of distributed training: parameter servers do not depend on each other, they'll do parameter optimization concurrently. Parameter servers will not wait for trainers, so trainers will also do their work concurrently. But asynchronous SGD will introduce more randomness and noises in the gradient.
# Preparations
## Preparations
1. Prepare your computer cluster. It's normally a bunch of Linux servers connected by LAN. Each server will be assigned a unique IP address. The computers in the cluster can be called "nodes".
2. Install PaddlePaddle on every node. If you are going to take advantage of GPU cards, you'll also need to install proper driver and CUDA libraries. To install PaddlePaddle please read [this build and install](https://github.com/PaddlePaddle/Paddle/tree/develop/doc/getstarted/build_and_install) document. We strongly recommend using [Docker installation](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/getstarted/build_and_install/docker_install_en.rst).
......@@ -52,9 +52,9 @@ PaddlePaddle 0.10.0rc, compiled with
We'll take `doc/howto/usage/cluster/src/word2vec` as an example to introduce distributed training using PaddlePaddle v2 API.
# Command-line arguments
## Command-line arguments
## Starting parameter server
### Starting parameter server
Type the below command to start a parameter server which will wait for trainers to connect:
......@@ -74,7 +74,7 @@ $ stdbuf -oL /usr/bin/nohup paddle pserver --port=7164 --ports_num=1 --ports_num
| ports_num_for_sparse | required | 1 | number of ports which serves sparse parameter update |
| num_gradient_servers | required | 1 | total number of gradient servers |
## Starting trainer
### Starting trainer
Type the command below to start the trainer(name the file whatever you want, like "train.py")
```bash
......@@ -122,7 +122,7 @@ paddle.init(
| trainer_id | required | 0 | ID for every trainer, start from 0 |
| pservers | required | 127.0.0.1 | list of IPs of parameter servers, separated by "," |
## Prepare Training Dataset
### Prepare Training Dataset
Here's some example code [prepare.py](https://github.com/PaddlePaddle/Paddle/tree/develop/doc/howto/usage/cluster/src/word2vec/prepare.py), it will download public `imikolov` dataset and split it into multiple files according to job parallelism(trainers count). Modify `SPLIT_COUNT` at the begining of `prepare.py` to change the count of output files.
......@@ -155,7 +155,7 @@ When job started, every trainer needs to get it's own part of data. In some dist
Different training jobs may have different data format and `reader()` function, developers may need to write different data prepare scripts and `reader()` functions for their job.
## Prepare Training program
### Prepare Training program
We'll create a *workspace* directory on each node, storing your training program, dependencies, mounted or downloaded dataset directory.
......@@ -191,7 +191,7 @@ Your workspace may looks like:
- `train_data_dir`: containing training data. Mount from storage service or copy trainning data to here.
- `test_data_dir`: containing testing data.
# Use cluster platforms or cluster management tools
## Use cluster platforms or cluster management tools
PaddlePaddle supports running jobs on several platforms including:
- [Kubernetes](http://kubernetes.io) open-source system for automating deployment, scaling, and management of containerized applications from Google.
......@@ -202,13 +202,13 @@ We'll introduce cluster job management on these platforms. The examples can be f
These cluster platforms provide API or environment variables for training processes, when the job is dispatched to different nodes. Like node ID, IP or total number of nodes etc.
## Cluster Training Using Fabric
### Cluster Training Using Fabric
### Prepare a Linux cluster
#### Prepare a Linux cluster
Run `kubectl -f ssh_servers.yaml` under the directory: `paddle/scripts/cluster_train_v2/fabric/docker_cluster` will launch a demo cluster. Run `kubectl get po -o wide` to get IP addresses of these nodes.
### Launching Cluster Job
#### Launching Cluster Job
`paddle.py` provides automatical scripts to start all PaddlePaddle cluster processes in different nodes. By default, all command line options can be set as `paddle.py` command options and `paddle.py` will transparently and automatically set these options to PaddlePaddle lower level processes.
`paddle.py`provides two distinguished command option for easy job launching.
......@@ -224,10 +224,10 @@ sh run.sh
The cluster Job will start in several seconds.
### Kill Cluster Job
#### Kill Cluster Job
`paddle.py` can capture `Ctrl + C` SIGINT signal to automatically kill all processes launched by it. So just stop `paddle.py` to kill cluster job. You should manually kill the job if the program crashed.
### Check Cluster Training Result
#### Check Cluster Training Result
Check log in $workspace/log for details, each node owns same log structure.
`paddle_trainer.INFO`
......@@ -242,13 +242,13 @@ It provides stderr and stdout of parameter server process. Check error log if tr
`train.log`
It provides stderr and stdout of trainer process. Check error log if training crashes.
### Check Model Output
#### Check Model Output
After one pass finished, model files will be written in `output` directory in node 0.
`nodefile` in workspace indicates the node id of current cluster job.
## Cluster Training Using OpenMPI
### Cluster Training Using OpenMPI
### Prepare an OpenMPI cluster
#### Prepare an OpenMPI cluster
Run the following command to start a 3-node MPI cluster and one "head" node.
......@@ -260,7 +260,7 @@ kubectl create -f mpi-nodes.yaml
Then you can log in to every OpenMPI node using ssh without input any passwords.
### Launching Cluster Job
#### Launching Cluster Job
Follow the steps to launch a PaddlePaddle training job in OpenMPI cluster:\
......@@ -288,6 +288,6 @@ scp train.txt-00002 test.txt-00002 [node3IP]:/home/tutorial
mpirun -hostfile machines -n 3 /home/tutorial/start_mpi_train.sh
```
## Cluster Training Using Kubernetes
### Cluster Training Using Kubernetes
The details can be found [here](../k8s/k8s_cn.md)
go/.gitignore
浏览文件 @ 1e8300fa
vendor/
.glide/
proto/*.go
go/glide.lock
浏览文件 @ 1e8300fa
hash: 51d9e2e46d7fd9173ff11ecada40f7b7728756be18d5e2f032535f66465e6e15
updated: 2017-10-24T15:04:09.987751592-07:00
hash: 107c058cf5c9163a75d40eef2273a793c36112683c25d72aa8288827fdde3a19
updated: 2017-10-30T03:46:19.137696069Z
imports:
- name: github.com/alecthomas/gometalinter
version: bae2f1293d092fd8167939d5108d1b025eaef9de
......
go/glide.yaml
浏览文件 @ 1e8300fa
......@@ -30,3 +30,4 @@ import:
version: v2.13
- package: github.com/go-stack/stack
version: v1.6.0
- package: github.com/golang/protobuf
go/proto/.gitignore 0 → 100644
浏览文件 @ 1e8300fa
# Ignore everything in this directory
*
# Except this file
!.gitignore
go/pserver/CMakeLists.txt
浏览文件 @ 1e8300fa
......@@ -13,5 +13,5 @@
# limitations under the License.
#
if(WITH_TESTING)
go_test(pserver_test DEPS paddle_go_optimizer)
go_test(pserver_test DEPS paddle_go_optimizer gen_proto_go)
endif()
go/pserver/service.go
浏览文件 @ 1e8300fa
......@@ -17,6 +17,7 @@ package pserver
import (
"bufio"
"bytes"
"encoding/binary"
"encoding/gob"
"encoding/json"
"errors"
......@@ -26,11 +27,15 @@ import (
"os"
"path"
"strconv"
"strings"
"sync"
"time"
"github.com/golang/protobuf/proto"
uuid "github.com/satori/go.uuid"
pb "github.com/PaddlePaddle/Paddle/go/proto"
log "github.com/inconshreveable/log15"
)
......@@ -65,6 +70,46 @@ type Parameter struct {
Content []byte
}
func float32ToString(b []byte) string {
f := make([]float32, len(b)/4)
buf := bytes.NewReader(b)
err := binary.Read(buf, binary.LittleEndian, &f)
if err != nil {
return ""
}
return fmt.Sprintf("%v", f)
}
func float32ByteToString(c []byte) string {
var a []byte
var b []byte
if len(c) <= 80 {
a = c
} else {
a = c[0:40]
b = c[len(c)-40:]
}
var s string
s = float32ToString(a)
if b == nil {
return s
}
s = strings.Replace(s, "]", "", -1) + "..." + strings.Replace(float32ToString(b), "[", "", -1)
return s
}
func (p Parameter) String() string {
if p.ElementType != Float32 {
return fmt.Sprintf("name:%v ElementType:%v",
p.Name, p.ElementType)
}
return float32ByteToString(p.Content)
}
// ParameterWithConfig contains the parameter and the configuration.
type ParameterWithConfig struct {
Param Parameter
......@@ -189,7 +234,9 @@ func (s *Service) InitParam(paramWithConfigs ParameterWithConfig, _ *int) error
default:
}
// TODO(helin): parse parameter config
c := &pb.OptimizerConfig{}
proto.Unmarshal(paramWithConfigs.Config, c)
log.Debug(fmt.Sprintf("OptimizerConfig:%v", c))
s.mu.Lock()
defer s.mu.Unlock()
......@@ -239,7 +286,8 @@ func (s *Service) SendGrad(g Gradient, _ *int) error {
select {
case <-s.initialized:
default:
log.Warn("received gradient before initialization.", "name", g.Name, "size", len(g.Content), "type", g.ElementType)
log.Warn("received gradient before initialization.",
"name", g.Name, "size", len(g.Content), "type", g.ElementType)
return errors.New(Uninitialized)
}
......@@ -248,10 +296,14 @@ func (s *Service) SendGrad(g Gradient, _ *int) error {
o, ok := s.optMap[g.Name]
if !ok {
log.Warn("received gradient but can't find name.",
"name", g.Name, "size", len(g.Content), "type", g.ElementType)
return fmt.Errorf("parameter: %s does not exist", g.Name)
}
log.Info("received gradient from trainer, updating gradient.", "name", g.Name, "size", len(g.Content), "type", g.ElementType)
log.Debug(Parameter(g).String())
log.Info("received gradient from trainer, updating gradient.",
"name", g.Name, "size", len(g.Content), "type", g.ElementType)
return o.UpdateParameter(g)
}
......@@ -277,7 +329,7 @@ func (s *Service) GetParam(name string, parameter *Parameter) error {
parameter.Name = name
parameter.ElementType = opt.elementType
parameter.Content = opt.GetWeights()
log.Debug(parameter.String())
log.Info("sending parameter to the trainer", "name", parameter.Name, "size", len(parameter.Content), "type", parameter.ElementType)
return nil
}
......
go/pserver/service_test.go
浏览文件 @ 1e8300fa
......@@ -15,6 +15,7 @@
package pserver_test
import (
"fmt"
"io/ioutil"
"reflect"
"sync"
......@@ -178,3 +179,33 @@ func TestBlockUntilInitialized(t *testing.T) {
wg.Wait()
}
func TestGradientString(t *testing.T) {
g := pserver.Parameter{}
g.ElementType = pserver.Float32
g.Content = []byte{0x18, 0x2d, 0x44, 0x54, 0xfb, 0x21, 0x09, 0x40, 0x18, 0x2d, 0x44, 0x54, 0xfb, 0x21, 0x09, 0x40}
if g.String() != "[3.3702806e+12 2.142699 3.3702806e+12 2.142699]" {
t.Fatal("get float data error!")
}
g.Content = []byte{0x18, 0x2d, 0x44, 0x54, 0xfb, 0x21, 0x09, 0x40,
0x18, 0x2d, 0x44, 0x54, 0xfb, 0x21, 0x09, 0x40,
0x18, 0x2d, 0x44, 0x54, 0xfb, 0x21, 0x09, 0x40,
0x18, 0x2d, 0x44, 0x54, 0xfb, 0x21, 0x09, 0x40,
0x18, 0x2d, 0x44, 0x54, 0xfb, 0x21, 0x09, 0x40,
0x18, 0x2d, 0x44, 0x54, 0xfb, 0x21, 0x09, 0x40,
0x18, 0x2d, 0x44, 0x54, 0xfb, 0x21, 0x09, 0x40,
0x18, 0x2d, 0x44, 0x54, 0xfb, 0x21, 0x09, 0x40,
0x18, 0x2d, 0x44, 0x54, 0xfb, 0x21, 0x09, 0x40,
0x18, 0x2d, 0x44, 0x54, 0xfb, 0x21, 0x09, 0x40,
0x18, 0x2d, 0x44, 0x54, 0xfb, 0x21, 0x09, 0x40,
0x18, 0x2d, 0x44, 0x54, 0xfb, 0x21, 0x09, 0x40,
0x18, 0x2d, 0x44, 0x54, 0xfb, 0x21, 0x09, 0x40,
0x18, 0x2d, 0x44, 0x54, 0xfb, 0x21, 0x09, 0x40,
0x18, 0x2d, 0x44, 0x54, 0xfb, 0x21, 0x09, 0x40,
0x18, 0x2d, 0x44, 0x54, 0xfb, 0x21, 0x09, 0x40}
if g.String() != "[3.3702806e+12 2.142699 3.3702806e+12 2.142699 3.3702806e+12 2.142699 3.3702806e+12 2.142699 3.3702806e+12 2.142699...3.3702806e+12 2.142699 3.3702806e+12 2.142699 3.3702806e+12 2.142699 3.3702806e+12 2.142699 3.3702806e+12 2.142699]" {
t.Fatal("get float data error!", g.String())
}
fmt.Println(g)
}
paddle/framework/attribute.cc
浏览文件 @ 1e8300fa
......@@ -19,7 +19,7 @@ limitations under the License. */
namespace paddle {
namespace framework {
Attribute GetAttrValue(const OpDesc::Attr& attr_desc, ProgramDesc* program) {
Attribute GetAttrValue(const OpDesc::Attr& attr_desc) {
switch (attr_desc.type()) {
case framework::AttrType::BOOLEAN: {
return attr_desc.b();
......@@ -61,13 +61,9 @@ Attribute GetAttrValue(const OpDesc::Attr& attr_desc, ProgramDesc* program) {
}
return val;
}
case framework::AttrType::BLOCK: {
PADDLE_ENFORCE(program != nullptr,
"Need to specify ProgramDesc when get a block attr");
return program->mutable_blocks(attr_desc.block_idx());
}
default:
PADDLE_THROW("Unsupport attr type %d", attr_desc.type());
}
PADDLE_ENFORCE(false, "Unknown OpDesc::AttrDesc::type !");
return boost::blank();
}
......
paddle/framework/attribute.h
浏览文件 @ 1e8300fa
......@@ -32,7 +32,7 @@ inline AttrType AttrTypeID() {
return static_cast<AttrType>(tmp.which() - 1);
}
Attribute GetAttrValue(const OpDesc::Attr& attr_desc, ProgramDesc* desc);
Attribute GetAttrValue(const OpDesc::Attr& attr_desc);
class AttrReader {
public:
......
paddle/framework/backward.cc
浏览文件 @ 1e8300fa
......@@ -18,6 +18,7 @@
#include <deque>
#include <list>
#include <memory>
#include <unordered_set>
#include "paddle/framework/block_desc.h"
#include "paddle/framework/op_registry.h"
......@@ -285,6 +286,15 @@ static bool AllGradInSet(const std::vector<std::string>& names,
return true;
}
static std::string FwdName(const std::string& grad_name) {
auto pos = grad_name.find("@GRAD");
if (pos == std::string::npos) {
return "";
} else {
return grad_name.substr(0, pos);
}
}
static void CreateGradVarInBlock(
size_t grad_op_start_index,
const std::unordered_map<std::string, std::string>& param_name_map,
......@@ -294,6 +304,7 @@ static void CreateGradVarInBlock(
for (size_t op_index = grad_op_start_index; op_index < ops.size();
++op_index) {
bool need_infer_shape = false;
std::unordered_set<std::string> new_vars;
ForEachVarName(ops[op_index]->Outputs(),
[&](const std::string& grad_var_name) {
if (block_desc->HasVar(grad_var_name)) {
......@@ -301,8 +312,7 @@ static void CreateGradVarInBlock(
}
need_infer_shape = true;
auto var = block_desc->Var(grad_var_name);
// FIXME(qiao) infer the datatype
var->SetDataType(framework::DataType::FP32);
new_vars.insert(var->Name());
auto it = param_name_map.find(grad_var_name);
if (it == param_name_map.end()) {
return false;
......@@ -316,6 +326,21 @@ static void CreateGradVarInBlock(
});
if (need_infer_shape) {
ops[op_index]->InferVarType(block_desc);
for (auto& arg : ops[op_index]->OutputArgumentNames()) {
if (new_vars.find(arg) == new_vars.end()) {
continue;
}
auto pname = FwdName(arg);
auto* param = block_desc->FindVar(pname);
auto* grad = block_desc->FindVar(arg);
if (param == nullptr) {
LOG(WARNING) << "Cannot find forward variable of " << arg
<< ". Set its gradient to FP32";
grad->SetDataType(DataType::FP32);
} else {
grad->SetDataType(param->GetDataType());
}
}
ops[op_index]->InferShape(*block_desc);
}
}
......@@ -368,7 +393,7 @@ std::vector<std::unique_ptr<OpDescBind>> MakeBlockBackward(
ProgramDescBind& program_desc, int block_idx,
std::unordered_set<std::string>* no_grad_vars,
std::unordered_map<std::string, std::string>* grad_to_var) {
BlockDescBind* cur_block = program_desc.Block(block_idx);
BlockDescBind* cur_block = program_desc.MutableBlock(block_idx);
std::vector<OpDescBind*> op_descs = cur_block->AllOps();
std::unordered_map<std::string, std::vector<size_t>> dup_out_ops;
size_t grad_desc_idx = 0;
......@@ -443,7 +468,7 @@ ParamGradInfoMap AppendBackward(
}
const int root_block_idx = 0;
auto root_block = program_desc.Block(root_block_idx);
auto root_block = program_desc.MutableBlock(root_block_idx);
// insert fill one op for target
// TODO(qiao) add some check to the target.
......@@ -492,7 +517,7 @@ ParamGradInfoMap AppendBackward(
CreateGradVarInBlock(forward_op_num, grad_to_var, root_block, &retv);
for (size_t block_index = forward_block_num;
block_index < program_desc.Size(); ++block_index) {
CreateGradVarInBlock(0, grad_to_var, program_desc.Block(block_index),
CreateGradVarInBlock(0, grad_to_var, program_desc.MutableBlock(block_index),
&retv);
}
return retv;
......
paddle/framework/backward_test.cc
浏览文件 @ 1e8300fa
......@@ -499,7 +499,7 @@ TEST(Backward, linear_net_intermediate_variable_has_no_grad) {
TEST(Backward, simple_single_op) {
f::ProgramDescBind program;
f::BlockDescBind *block = program.Block(0);
f::BlockDescBind *block = program.MutableBlock(0);
f::OpDescBind *op = block->AppendOp();
op->SetType("rowwise_add");
......@@ -535,7 +535,7 @@ TEST(Backward, simple_single_op) {
TEST(Backward, default_attribute) {
f::ProgramDescBind program;
f::BlockDescBind *block = program.Block(0);
f::BlockDescBind *block = program.MutableBlock(0);
f::OpDescBind *op = block->AppendOp();
op->SetType("mul");
op->SetInput("X", {"x"});
......@@ -561,7 +561,7 @@ TEST(Backward, default_attribute) {
TEST(Backward, simple_mult_op) {
f::ProgramDescBind program;
f::BlockDescBind *block = program.Block(0);
f::BlockDescBind *block = program.MutableBlock(0);
f::OpDescBind *op1 = block->AppendOp();
op1->SetType("rowwise_add");
op1->SetInput("X", {"x1"});
......@@ -644,7 +644,7 @@ TEST(Backward, simple_mult_op) {
TEST(Backward, intermedia_var_no_grad) {
f::ProgramDescBind program;
f::BlockDescBind *block = program.Block(0);
f::BlockDescBind *block = program.MutableBlock(0);
f::OpDescBind *op1 = block->AppendOp();
op1->SetType("rowwise_add");
op1->SetInput("X", {"x1"});
......@@ -714,7 +714,7 @@ TEST(Backward, intermedia_var_no_grad) {
TEST(Backward, var_no_grad) {
f::ProgramDescBind program;
f::BlockDescBind *block = program.Block(0);
f::BlockDescBind *block = program.MutableBlock(0);
f::OpDescBind *op1 = block->AppendOp();
op1->SetType("mult_in_out");
op1->SetInput("X", {"x1"});
......@@ -790,7 +790,7 @@ TEST(Backward, var_no_grad) {
TEST(Backward, shared_var) {
f::ProgramDescBind program;
f::BlockDescBind *block = program.Block(0);
f::BlockDescBind *block = program.MutableBlock(0);
f::OpDescBind *op1 = block->AppendOp();
op1->SetType("rowwise_add");
op1->SetInput("X", {"x1"});
......@@ -880,7 +880,7 @@ TEST(Backward, shared_var) {
TEST(Backward, half_backward) {
f::ProgramDescBind program;
f::BlockDescBind *block = program.Block(0);
f::BlockDescBind *block = program.MutableBlock(0);
auto *op1 = block->AppendOp();
op1->SetType("minus");
op1->SetInput("X", {"a"});
......
paddle/framework/block_desc.cc
浏览文件 @ 1e8300fa
......@@ -113,7 +113,7 @@ BlockDescBind *BlockDescBind::ParentBlock() const {
if (this->desc_->parent_idx() == kNoneBlockIndex) {
return nullptr;
}
return prog_->Block(static_cast<size_t>(this->desc_->parent_idx()));
return prog_->MutableBlock(static_cast<size_t>(this->desc_->parent_idx()));
}
BlockDesc *BlockDescBind::Proto() {
......
paddle/framework/executor.cc
浏览文件 @ 1e8300fa
......@@ -73,33 +73,32 @@ static void CreateTensor(Variable* var, VarDesc::VarType var_type) {
}
}
void Executor::Run(const ProgramDesc& pdesc, Scope* scope, int block_id) {
void Executor::Run(const ProgramDescBind& pdesc, Scope* scope, int block_id) {
// TODO(tonyyang-svail):
// - only runs on the first device (i.e. no interdevice communication)
// - will change to use multiple blocks for RNN op and Cond Op
PADDLE_ENFORCE_GT(pdesc.blocks_size(), block_id);
auto& block = pdesc.blocks(block_id);
PADDLE_ENFORCE_LT(block_id, pdesc.Size());
auto& block = pdesc.Block(block_id);
auto& device = device_contexts_[0];
Scope& local_scope = scope->NewScope();
for (auto& var : block.vars()) {
if (var.persistable()) {
auto* ptr = scope->Var(var.name());
CreateTensor(ptr, var.type());
VLOG(3) << "Create Variable " << var.name()
for (auto& var : block.AllVars()) {
if (var->Persistable()) {
auto* ptr = scope->Var(var->Name());
CreateTensor(ptr, var->GetType());
VLOG(3) << "Create Variable " << var->Name()
<< " global, which pointer is " << ptr;
} else {
auto* ptr = local_scope.Var(var.name());
CreateTensor(ptr, var.type());
VLOG(3) << "Create Variable " << var.name()
auto* ptr = local_scope.Var(var->Name());
CreateTensor(ptr, var->GetType());
VLOG(3) << "Create Variable " << var->Name()
<< " locally, which pointer is " << ptr;
}
}
for (auto& op_desc : block.ops()) {
auto op = paddle::framework::OpRegistry::CreateOp(
op_desc, const_cast<ProgramDesc*>(&pdesc));
for (auto& op_desc : block.AllOps()) {
auto op = paddle::framework::OpRegistry::CreateOp(*op_desc);
op->Run(local_scope, *device);
}
......
paddle/framework/executor.h
浏览文件 @ 1e8300fa
......@@ -14,8 +14,8 @@ limitations under the License. */
#pragma once
#include "paddle/framework/framework.pb.h"
#include "paddle/framework/op_info.h"
#include "paddle/framework/program_desc.h"
#include "paddle/framework/scope.h"
#include "paddle/framework/tensor.h"
......@@ -34,7 +34,7 @@ class Executor {
* ProgramDesc
* Scope
*/
void Run(const ProgramDesc&, Scope*, int);
void Run(const ProgramDescBind&, Scope*, int);
private:
std::vector<platform::DeviceContext*> device_contexts_;
......
paddle/framework/lod_tensor_test.cu
浏览文件 @ 1e8300fa
......@@ -36,8 +36,8 @@ TEST(LoDTensor, LoDInGPU) {
lod_tensor.mutable_data<float>(place);
lod_tensor.set_lod(src_lod);
CHECK_EQ(lod_tensor.lod_element(0, 2).first, 4UL);
CHECK_EQ(lod_tensor.lod_element(0, 4).first, 8UL);
EXPECT_EQ(lod_tensor.lod_element(0, 2).first, 4UL);
EXPECT_EQ(lod_tensor.lod_element(0, 4).first, 8UL);
auto lod = lod_tensor.lod();
......@@ -45,6 +45,6 @@ TEST(LoDTensor, LoDInGPU) {
cudaDeviceSynchronize();
for (size_t i = 0; i < src_lod[0].size(); ++i) {
CHECK_EQ(lod[0].data()[i], src_lod[0].data()[i] * 2);
EXPECT_EQ(lod[0].data()[i], src_lod[0].data()[i] * 2);
}
}
\ No newline at end of file
}
paddle/framework/op_desc.cc
浏览文件 @ 1e8300fa
......@@ -52,6 +52,22 @@ class CompileTimeInferShapeContext : public InferShapeContext {
const std::vector<std::string> &Outputs(
const std::string &name) const override;
void ShareLoD(const std::string &in, const std::string &out, size_t i = 0,
size_t j = 0) const override {
PADDLE_ENFORCE_LT(i, Inputs(in).size());
PADDLE_ENFORCE_LT(j, Outputs(out).size());
auto *in_var = block_.FindVarRecursive(Inputs(in)[i]);
auto *out_var = block_.FindVarRecursive(Outputs(out)[j]);
if (in_var->GetType() != VarDesc::LOD_TENSOR) {
VLOG(3) << "input " << in << "is not LodTensor";
return;
}
PADDLE_ENFORCE_EQ(in_var->GetType(), VarDesc::LOD_TENSOR,
"The %d-th output of Output(%s) must be LoDTensor.", j,
out);
in_var->SetLoDLevel(out_var->GetLodLevel());
}
private:
DDim GetDim(const std::string &name) const override;
......@@ -98,7 +114,12 @@ OpDescBind::OpDescBind(const OpDesc &desc, ProgramDescBind *prog)
// restore attrs_
for (const OpDesc::Attr &attr : desc_.attrs()) {
std::string attr_name = attr.name();
attrs_[attr_name] = GetAttrValue(attr, prog->Proto());
if (attr.type() != AttrType::BLOCK) {
attrs_[attr_name] = GetAttrValue(attr);
} else {
auto bid = attr.block_idx();
attrs_[attr_name] = prog->MutableBlock(bid);
}
}
}
......@@ -172,8 +193,7 @@ void OpDescBind::SetAttr(const std::string &name, const Attribute &v) {
}
void OpDescBind::SetBlockAttr(const std::string &name, BlockDescBind &block) {
BlockDesc *desc = block.Proto();
this->attrs_[name] = desc;
this->attrs_[name] = &block;
need_update_ = true;
}
......@@ -192,7 +212,7 @@ Attribute OpDescBind::GetAttr(const std::string &name) const {
int OpDescBind::GetBlockAttr(const std::string &name) const {
auto it = attrs_.find(name);
PADDLE_ENFORCE(it != attrs_.end(), "Attribute %s is not found", name);
return boost::get<BlockDesc *>(it->second)->idx();
return boost::get<BlockDescBind *>(it->second)->ID();
}
const std::unordered_map<std::string, Attribute> &OpDescBind::GetAttrMap()
......
paddle/framework/op_registry.cc
浏览文件 @ 1e8300fa
......@@ -43,13 +43,15 @@ static VariableNameMap ConvertOpDescVarsToVarNameMap(
return ret_val;
}
std::unique_ptr<OperatorBase> OpRegistry::CreateOp(const OpDesc& op_desc,
ProgramDesc* program) {
std::unique_ptr<OperatorBase> OpRegistry::CreateOp(const OpDesc& op_desc) {
VLOG(1) << "CreateOp directly from OpDesc is deprecated. It should only be"
"used in unit tests. Use CreateOp(const OpDescBind& op_desc) "
"instead.";
VariableNameMap inputs = ConvertOpDescVarsToVarNameMap(op_desc.inputs());
VariableNameMap outputs = ConvertOpDescVarsToVarNameMap(op_desc.outputs());
AttributeMap attrs;
for (auto& attr : op_desc.attrs()) {
attrs[attr.name()] = GetAttrValue(attr, program);
attrs[attr.name()] = GetAttrValue(attr);
}
return CreateOp(op_desc.type(), inputs, outputs, attrs);
......
paddle/framework/op_registry.h
浏览文件 @ 1e8300fa
......@@ -77,8 +77,7 @@ class OpRegistry {
const VariableNameMap& outputs,
AttributeMap attrs);
static std::unique_ptr<OperatorBase> CreateOp(const OpDesc& op_desc,
ProgramDesc* program);
static std::unique_ptr<OperatorBase> CreateOp(const OpDesc& op_desc);
static std::unique_ptr<OperatorBase> CreateOp(const OpDescBind& op_desc);
};
......
paddle/framework/op_registry_test.cc
浏览文件 @ 1e8300fa
......@@ -74,7 +74,7 @@ TEST(OpRegistry, CreateOp) {
attr->set_type(paddle::framework::AttrType::FLOAT);
attr->set_f(scale);
auto op = paddle::framework::OpRegistry::CreateOp(op_desc, nullptr);
auto op = paddle::framework::OpRegistry::CreateOp(op_desc);
paddle::framework::Scope scope;
paddle::platform::CPUDeviceContext dev_ctx;
op->Run(scope, dev_ctx);
......@@ -95,7 +95,7 @@ TEST(OpRegistry, IllegalAttr) {
bool caught = false;
try {
paddle::framework::OpRegistry::CreateOp(op_desc, nullptr);
paddle::framework::OpRegistry::CreateOp(op_desc);
} catch (paddle::platform::EnforceNotMet err) {
caught = true;
std::string msg = "larger_than check fail";
......@@ -115,7 +115,7 @@ TEST(OpRegistry, DefaultValue) {
ASSERT_TRUE(op_desc.IsInitialized());
auto op = paddle::framework::OpRegistry::CreateOp(op_desc, nullptr);
auto op = paddle::framework::OpRegistry::CreateOp(op_desc);
paddle::framework::Scope scope;
paddle::platform::CPUDeviceContext dev_ctx;
op->Run(scope, dev_ctx);
......@@ -131,7 +131,7 @@ TEST(OpRegistry, CustomChecker) {
// attr 'test_attr' is not set
bool caught = false;
try {
paddle::framework::OpRegistry::CreateOp(op_desc, nullptr);
paddle::framework::OpRegistry::CreateOp(op_desc);
} catch (paddle::platform::EnforceNotMet err) {
caught = true;
std::string msg = "Attribute 'test_attr' is required!";
......@@ -149,7 +149,7 @@ TEST(OpRegistry, CustomChecker) {
attr->set_i(3);
caught = false;
try {
paddle::framework::OpRegistry::CreateOp(op_desc, nullptr);
paddle::framework::OpRegistry::CreateOp(op_desc);
} catch (paddle::platform::EnforceNotMet err) {
caught = true;
std::string msg = "'test_attr' must be even!";
......@@ -166,7 +166,7 @@ TEST(OpRegistry, CustomChecker) {
attr->set_name("test_attr");
attr->set_type(paddle::framework::AttrType::INT);
attr->set_i(4);
auto op = paddle::framework::OpRegistry::CreateOp(op_desc, nullptr);
auto op = paddle::framework::OpRegistry::CreateOp(op_desc);
paddle::platform::CPUDeviceContext dev_ctx;
paddle::framework::Scope scope;
op->Run(scope, dev_ctx);
......
paddle/framework/operator.cc
浏览文件 @ 1e8300fa
......@@ -37,32 +37,32 @@ ExecutionContext::GetEigenDevice<platform::GPUPlace, Eigen::GpuDevice>() const {
std::string OperatorBase::Input(const std::string& name) const {
auto& ins = Inputs(name);
PADDLE_ENFORCE_LE(ins.size(), 1UL,
"Op %s input %s should contain only one variable", type_,
name);
"Operator %s's input %s should contain only one variable.",
type_, name);
return ins.empty() ? kEmptyVarName : ins[0];
}
const std::vector<std::string>& OperatorBase::Inputs(
const std::string& name) const {
auto it = inputs_.find(name);
PADDLE_ENFORCE(it != inputs_.end(), "Op %s do not have input %s", type_,
name);
PADDLE_ENFORCE(it != inputs_.end(), "Operator %s does not have the input %s.",
type_, name);
return it->second;
}
std::string OperatorBase::Output(const std::string& name) const {
auto& outs = Outputs(name);
PADDLE_ENFORCE_LE(outs.size(), 1UL,
"Op %s output %s should contain only one variable", type_,
name);
"Operator %s's output %s should contain only one variable.",
type_, name);
return outs.empty() ? kEmptyVarName : outs[0];
}
const std::vector<std::string>& OperatorBase::Outputs(
const std::string& name) const {
auto it = outputs_.find(name);
PADDLE_ENFORCE(it != outputs_.end(), "Op %s does not have output called %s",
type_, name);
PADDLE_ENFORCE(it != outputs_.end(),
"Operator %s does not have an output called %s.", type_, name);
return it->second;
}
......@@ -351,6 +351,20 @@ class RuntimeInferShapeContext : public InferShapeContext {
return op_.Outputs(name);
}
void ShareLoD(const std::string& in, const std::string& out, size_t i = 0,
size_t j = 0) const override {
PADDLE_ENFORCE_LT(i, Inputs(in).size());
PADDLE_ENFORCE_LT(j, Outputs(out).size());
Variable* in_var = scope_.FindVar(Inputs(in)[i]);
Variable* out_var = scope_.FindVar(Outputs(out)[j]);
if (!in_var->IsType<LoDTensor>()) return;
PADDLE_ENFORCE(out_var->IsType<LoDTensor>(),
"The %d-th output of Output(%s) must be LoDTensor.", j, out);
auto in_tensor = in_var->Get<LoDTensor>();
auto* out_tensor = out_var->GetMutable<LoDTensor>();
out_tensor->set_lod(in_tensor.lod());
}
private:
DDim GetDim(const std::string& name) const override {
Variable* var = scope_.FindVar(name);
......
paddle/framework/operator.h
浏览文件 @ 1e8300fa
......@@ -427,7 +427,8 @@ class OperatorWithKernel : public OperatorBase {
int tmp = static_cast<int>(ToDataType(t->type()));
VLOG(3) << "Input " << ipt_name << " with data_type " << tmp;
PADDLE_ENFORCE(tmp == data_type || data_type == -1,
"DataType of Paddle Op %s must be same.", Type());
"DataType of Paddle Op %s must be the same.",
Type());
data_type = tmp;
}
}
......
paddle/framework/operator_test.cc
浏览文件 @ 1e8300fa
......@@ -83,7 +83,7 @@ TEST(OperatorBase, all) {
paddle::platform::CPUDeviceContext device_context;
paddle::framework::Scope scope;
auto op = paddle::framework::OpRegistry::CreateOp(op_desc, nullptr);
auto op = paddle::framework::OpRegistry::CreateOp(op_desc);
scope.Var("OUT1");
ASSERT_EQ(paddle::framework::op_run_num, 0);
op->Run(scope, device_context);
......@@ -208,7 +208,7 @@ TEST(OpKernel, all) {
paddle::platform::CPUDeviceContext cpu_device_context;
paddle::framework::Scope scope;
auto op = paddle::framework::OpRegistry::CreateOp(op_desc, nullptr);
auto op = paddle::framework::OpRegistry::CreateOp(op_desc);
ASSERT_EQ(paddle::framework::cpu_kernel_run_num, 0);
op->Run(scope, cpu_device_context);
ASSERT_EQ(paddle::framework::cpu_kernel_run_num, 1);
......@@ -244,7 +244,7 @@ TEST(OpKernel, multi_inputs) {
scope.Var("y0")->GetMutable<LoDTensor>();
scope.Var("y1")->GetMutable<LoDTensor>();
auto op = paddle::framework::OpRegistry::CreateOp(op_desc, nullptr);
auto op = paddle::framework::OpRegistry::CreateOp(op_desc);
op->Run(scope, cpu_device_context);
}
......
paddle/framework/program_desc.h
浏览文件 @ 1e8300fa
......@@ -37,7 +37,9 @@ class ProgramDescBind {
BlockDescBind *AppendBlock(const BlockDescBind &parent);
BlockDescBind *Block(size_t idx) { return blocks_[idx].get(); }
BlockDescBind *MutableBlock(size_t idx) { return blocks_[idx].get(); }
const BlockDescBind &Block(size_t idx) const { return *blocks_[idx]; }
size_t Size() const { return blocks_.size(); }
......
paddle/framework/program_desc_test.cc
浏览文件 @ 1e8300fa
......@@ -20,7 +20,7 @@ namespace paddle {
namespace framework {
TEST(ProgramDesc, copy_ctor) {
ProgramDescBind program;
auto* global_block = program.Block(0);
auto* global_block = program.MutableBlock(0);
auto* x = global_block->Var("X");
x->SetType(VarDesc_VarType_LOD_TENSOR);
x->SetLoDLevel(0);
......@@ -44,7 +44,7 @@ TEST(ProgramDesc, copy_ctor) {
ProgramDescBind program_copy(program);
auto* global_block_copy = program_copy.Block(0);
auto* global_block_copy = program_copy.MutableBlock(0);
ASSERT_NE(global_block, global_block_copy);
auto assert_same_var = [&](const std::string& name, VarDescBind* var_before) {
......@@ -82,7 +82,7 @@ TEST(ProgramDesc, copy_ctor) {
TEST(ProgramDescBind, serialize_and_deserialize) {
ProgramDescBind program_origin;
auto* global_block = program_origin.Block(0);
auto* global_block = program_origin.MutableBlock(0);
auto* x = global_block->Var("X");
x->SetType(VarDesc_VarType_LOD_TENSOR);
x->SetLoDLevel(0);
......@@ -108,7 +108,7 @@ TEST(ProgramDescBind, serialize_and_deserialize) {
program_origin.Proto()->SerializeToString(&binary_str);
ProgramDescBind program_restored(binary_str);
auto* global_block_restored = program_restored.Block(0);
auto* global_block_restored = program_restored.MutableBlock(0);
ASSERT_NE(global_block, global_block_restored);
auto assert_same_var = [&](const std::string& name, VarDescBind* var_before) {
......
paddle/framework/prune_test.cc
浏览文件 @ 1e8300fa
......@@ -52,7 +52,7 @@ void AddOp(const std::string &type, const f::VariableNameMap &inputs,
TEST(Prune, one_operator) {
f::ProgramDescBind program;
f::BlockDescBind *block = program.Block(0);
f::BlockDescBind *block = program.MutableBlock(0);
AddOp("one_one", {{"input", {"a"}}}, {{"output", {"b"}}}, {}, block);
......@@ -69,7 +69,7 @@ TEST(Prune, one_operator) {
TEST(Prune, forward) {
f::ProgramDescBind program;
f::BlockDescBind *block = program.Block(0);
f::BlockDescBind *block = program.MutableBlock(0);
AddOp("one_one", {{"input", {"a"}}}, {{"output", {"b"}}}, {}, block);
AddOp("one_one", {{"input", {"b"}}}, {{"output", {"c"}}}, {}, block);
......@@ -88,7 +88,7 @@ TEST(Prune, forward) {
TEST(Prune, multi_input_op) {
f::ProgramDescBind program;
f::BlockDescBind *block = program.Block(0);
f::BlockDescBind *block = program.MutableBlock(0);
AddOp("one_one", {{"input", {"a0"}}}, {{"output", {"b0"}}}, {}, block);
AddOp("one_one", {{"input", {"a1"}}}, {{"output", {"b1"}}}, {}, block);
......@@ -106,7 +106,7 @@ TEST(Prune, multi_input_op) {
TEST(Prune, multi_output_op) {
f::ProgramDescBind program;
f::BlockDescBind *block = program.Block(0);
f::BlockDescBind *block = program.MutableBlock(0);
AddOp("one_two", {{"input", {"a"}}}, {{"output", {"b", "c"}}}, {}, block);
AddOp("one_one", {{"input", {"b"}}}, {{"output", {"b1"}}}, {}, block);
......@@ -122,7 +122,7 @@ TEST(Prune, multi_output_op) {
TEST(Prune, multi_target) {
f::ProgramDescBind program;
f::BlockDescBind *block = program.Block(0);
f::BlockDescBind *block = program.MutableBlock(0);
AddOp("one_two", {{"input", {"a"}}}, {{"output", {"b", "c"}}}, {}, block);
AddOp("one_one", {{"input", {"b"}}}, {{"output", {"b1"}}}, {}, block);
......
paddle/framework/shape_inference.cc
浏览文件 @ 1e8300fa
......@@ -28,9 +28,6 @@ void InferShapeContext::SetOutputsDim(
SetDims(names, dims);
}
void InferShapeContext::ShareLoD(const std::string &in, const std::string &out,
size_t i, size_t j) const {}
std::vector<framework::DDim> InferShapeContext::GetDims(
const std::vector<std::string> &names) const {
std::vector<framework::DDim> ret;
......
paddle/framework/shape_inference.h
浏览文件 @ 1e8300fa
......@@ -43,9 +43,8 @@ class InferShapeContext {
virtual const std::vector<std::string> &Outputs(
const std::string &name) const = 0;
// TODO(qiao) implement this function
void ShareLoD(const std::string &in, const std::string &out, size_t i = 0,
size_t j = 0) const;
virtual void ShareLoD(const std::string &in, const std::string &out,
size_t i = 0, size_t j = 0) const = 0;
protected:
virtual framework::DDim GetDim(const std::string &name) const = 0;
......
paddle/framework/tensor.h
浏览文件 @ 1e8300fa
......@@ -118,10 +118,12 @@ class Tensor {
const platform::DeviceContext& ctx);
/**
* @brief Return the slice of the tensor.
* @brief Return a sub-tensor of the given tensor.
*
* @param[in] begin_idx The begin index of the slice.
* @param[in] end_idx The end index of the slice.
* @param[in] begin_idx The index of the start row(inclusive) to slice.
* The index number begins from 0.
* @param[in] end_idx The index of the end row(exclusive) to slice.
* The index number begins from 0.
*/
inline Tensor Slice(const int& begin_idx, const int& end_idx) const;
......
paddle/framework/tensor_impl.h
浏览文件 @ 1e8300fa
......@@ -112,9 +112,10 @@ inline void* Tensor::mutable_data(platform::Place place, std::type_index type) {
if (holder_ != nullptr) {
holder_->set_type(type);
}
PADDLE_ENFORCE_GT(numel(), 0,
"Tensor's numel must be larger than zero to call "
"Tensor::mutable_data. Call Tensor::set_dim first.");
PADDLE_ENFORCE_GT(
numel(), 0,
"When calling this method, the Tensor's numel must be larger than zero. "
"Please check Tensor::Resize has been called first.");
int64_t size = numel() * SizeOfType(type);
/* some versions of boost::variant don't have operator!= */
if (holder_ == nullptr || !(holder_->place() == place) ||
......@@ -229,10 +230,12 @@ inline void Tensor::CopyFromVector(const std::vector<T>& src,
inline Tensor Tensor::Slice(const int& begin_idx, const int& end_idx) const {
check_memory_size();
PADDLE_ENFORCE_GE(begin_idx, 0, "Slice begin index is less than zero.");
PADDLE_ENFORCE_LE(end_idx, dims_[0], "Slice end index is out of bound.");
PADDLE_ENFORCE_LT(begin_idx, end_idx,
"Begin index must be less than end index.");
PADDLE_ENFORCE_GE(begin_idx, 0,
"The start row index must be greater than 0.");
PADDLE_ENFORCE_LE(end_idx, dims_[0], "The end row index is out of bound.");
PADDLE_ENFORCE_LT(
begin_idx, end_idx,
"The start row index must be lesser than the end row index.");
if (dims_[0] == 1) {
return *this;
......
paddle/framework/type_defs.h
浏览文件 @ 1e8300fa
......@@ -36,7 +36,7 @@ using VariableNameMap = std::map<std::string, std::vector<std::string>>;
using Attribute =
boost::variant<boost::blank, int, float, std::string, std::vector<int>,
std::vector<float>, std::vector<std::string>, bool,
std::vector<bool>, BlockDesc*>;
std::vector<bool>, BlockDescBind*>;
using AttributeMap = std::unordered_map<std::string, Attribute>;
......
paddle/framework/var_type_inference_test.cc
浏览文件 @ 1e8300fa
......@@ -63,41 +63,43 @@ namespace framework {
TEST(InferVarType, sum_op) {
ProgramDescBind prog;
auto *op = prog.Block(0)->AppendOp();
auto *op = prog.MutableBlock(0)->AppendOp();
op->SetType("sum");
op->SetInput("X", {"test_a", "test_b", "test_c"});
op->SetOutput("Out", {"test_out"});
prog.Block(0)->Var("test_a")->SetType(VarDesc::SELECTED_ROWS);
prog.Block(0)->Var("test_b")->SetType(VarDesc::SELECTED_ROWS);
prog.Block(0)->Var("test_c")->SetType(VarDesc::SELECTED_ROWS);
prog.Block(0)->Var("test_out");
prog.MutableBlock(0)->Var("test_a")->SetType(VarDesc::SELECTED_ROWS);
prog.MutableBlock(0)->Var("test_b")->SetType(VarDesc::SELECTED_ROWS);
prog.MutableBlock(0)->Var("test_c")->SetType(VarDesc::SELECTED_ROWS);
prog.MutableBlock(0)->Var("test_out");
op->InferVarType(prog.Block(0));
op->InferVarType(prog.MutableBlock(0));
ASSERT_EQ(VarDesc::SELECTED_ROWS, prog.Block(0)->Var("test_out")->GetType());
ASSERT_EQ(VarDesc::SELECTED_ROWS,
prog.MutableBlock(0)->Var("test_out")->GetType());
prog.Block(0)->Var("test_b")->SetType(VarDesc::LOD_TENSOR);
op->InferVarType(prog.Block(0));
ASSERT_EQ(VarDesc::LOD_TENSOR, prog.Block(0)->Var("test_out")->GetType());
prog.MutableBlock(0)->Var("test_b")->SetType(VarDesc::LOD_TENSOR);
op->InferVarType(prog.MutableBlock(0));
ASSERT_EQ(VarDesc::LOD_TENSOR,
prog.MutableBlock(0)->Var("test_out")->GetType());
}
TEST(InferVarType, sum_op_without_infer_var_type) {
ProgramDescBind prog;
auto *op = prog.Block(0)->AppendOp();
auto *op = prog.MutableBlock(0)->AppendOp();
op->SetType("sum_without_infer_var_type");
op->SetInput("X", {"test2_a", "test2_b", "test2_c"});
op->SetOutput("Out", {"test2_out"});
prog.Block(0)->Var("test2_a")->SetType(VarDesc::SELECTED_ROWS);
prog.Block(0)->Var("test2_b")->SetType(VarDesc::SELECTED_ROWS);
prog.Block(0)->Var("test2_c")->SetType(VarDesc::SELECTED_ROWS);
prog.Block(0)->Var("test2_out");
prog.MutableBlock(0)->Var("test2_a")->SetType(VarDesc::SELECTED_ROWS);
prog.MutableBlock(0)->Var("test2_b")->SetType(VarDesc::SELECTED_ROWS);
prog.MutableBlock(0)->Var("test2_c")->SetType(VarDesc::SELECTED_ROWS);
prog.MutableBlock(0)->Var("test2_out");
op->InferVarType(prog.Block(0));
op->InferVarType(prog.MutableBlock(0));
ASSERT_EQ(VarDesc_VarType_LOD_TENSOR,
prog.Block(0)->Var("test2_out")->GetType());
prog.MutableBlock(0)->Var("test2_out")->GetType());
}
} // namespace framework
......
paddle/gserver/layers/CRFLayer.cpp
浏览文件 @ 1e8300fa
......@@ -101,8 +101,10 @@ void CRFLayer::backward(const UpdateCallback& callback) {
: real(1.0f);
instanceWeight *= coeff_;
MatrixPtr grad = output.grad->subRowMatrix(starts[i], starts[i + 1]);
grad->add(*crfs_[i].getXGrad(), real(1.0f), instanceWeight);
if (output.grad) {
MatrixPtr grad = output.grad->subRowMatrix(starts[i], starts[i + 1]);
grad->add(*crfs_[i].getXGrad(), real(1.0f), instanceWeight);
}
if (needWGrad) {
weight_->getWGrad()->add(
*crfs_[i].getWGrad(), real(1.0f), instanceWeight);
......
paddle/gserver/layers/LinearChainCRF.cpp
浏览文件 @ 1e8300fa
......@@ -102,7 +102,6 @@ real LinearChainCRF::forward(real* x, int* s, int length) {
}
void LinearChainCRF::backward(real* x, int* s, int length, bool needWGrad) {
MatrixPtr matX = Matrix::create(x, length, numClasses_);
Matrix::resizeOrCreate(matGrad_, length, numClasses_);
Matrix::resizeOrCreate(beta_, length, numClasses_);
real* b = b_->getData();
......
paddle/gserver/layers/SequenceReshapeLayer.cpp
浏览文件 @ 1e8300fa
......@@ -70,11 +70,23 @@ void SequenceReshapeLayer::forward(PassType passType) {
size_t outDim = getSize();
size_t numSequences = input.getNumSequences();
auto startPositions = input.sequenceStartPositions->getVector(false);
const int* starts = startPositions->getData();
CHECK_EQ(starts[numSequences], input.getBatchSize());
CHECK_EQ(numSequences, startPositions->getSize() - 1);
// by default, we assume each instance as a sequence
IVectorPtr seqStarts;
IVector::resizeOrCreate(seqStarts, input.getBatchSize() + 1, false);
int* startsData = seqStarts->getData();
for (int i = 0; i < input.getBatchSize() + 1; i++) {
startsData[i] = i;
}
const int* starts = startsData;
// if there is sequence, then use start positions
if (input.sequenceStartPositions) {
auto startPositions = input.sequenceStartPositions->getVector(false);
starts = startPositions->getData();
CHECK_EQ(starts[numSequences], input.getBatchSize());
CHECK_EQ(numSequences, startPositions->getSize() - 1);
}
for (size_t seqID = 0; seqID < numSequences; seqID++) {
size_t inNumIns = starts[seqID + 1] - starts[seqID];
......
paddle/gserver/tests/CMakeLists.txt
浏览文件 @ 1e8300fa
# gserver pacakge unittests
if(NOT MOBILE_INFERENCE)
################### test_ProtoDataProvider ############
add_unittest_without_exec(test_ProtoDataProvider
test_ProtoDataProvider.cpp)
# test_ProtoDataProvider will mkdir as same name,
# so if WORKING_DIRECTORY is default directory, then
# mkdir will get error.
add_test(NAME test_ProtoDataProvider
COMMAND ${CMAKE_CURRENT_BINARY_DIR}/test_ProtoDataProvider
WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle)
endif()
add_simple_unittest(test_LinearChainCRF)
add_simple_unittest(test_MultinomialSampler)
add_simple_unittest(test_RecurrentLayer)
################# test_LayerGrad #######################
add_unittest_without_exec(test_LayerGrad
test_LayerGrad.cpp
LayerGradUtil.cpp)
add_test(NAME test_LayerGrad
COMMAND test_LayerGrad)
function(gserver_test TARGET)
add_unittest_without_exec(${TARGET}
${TARGET}.cpp
LayerGradUtil.cpp)
add_test(NAME ${TARGET}
COMMAND ${TARGET})
endfunction()
gserver_test(test_LayerGrad)
gserver_test(test_CRFLayerGrad)
gserver_test(test_CrossEntropyOverBeamGrad)
gserver_test(test_SeqSliceLayerGrad)
gserver_test(test_ActivationGrad)
gserver_test(test_ConvTrans)
gserver_test(test_PriorBox)
gserver_test(test_DetectionOutput)
gserver_test(test_ConvUnify)
gserver_test(test_BatchNorm)
gserver_test(test_KmaxSeqScore)
gserver_test(test_Expand)
########## test_Mkldnn layers and activations ##########
if(WITH_MKLDNN)
......@@ -32,89 +37,6 @@ if(WITH_MKLDNN)
WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle)
endif()
################ test_CRFLayerGrad ####################
add_unittest_without_exec(test_CRFLayerGrad
test_CRFLayerGrad.cpp
LayerGradUtil.cpp)
add_test(NAME test_CRFLayerGrad
COMMAND test_CRFLayerGrad)
################ test_CrossEntropyOverBeam ####################
add_unittest_without_exec(test_CrossEntropyOverBeam
test_CrossEntropyOverBeamGrad.cpp
LayerGradUtil.cpp)
add_test(NAME test_CrossEntropyOverBeam
COMMAND test_CrossEntropyOverBeam)
################ test_SeqSliceLayerGrad ####################
add_unittest_without_exec(test_SeqSliceLayerGrad
test_SeqSliceLayerGrad.cpp
LayerGradUtil.cpp)
add_test(NAME test_SeqSliceLayerGrad
COMMAND test_SeqSliceLayerGrad)
add_unittest_without_exec(test_ActivationGrad
test_ActivationGrad.cpp
LayerGradUtil.cpp)
add_test(NAME test_ActivationGrad
COMMAND test_ActivationGrad)
################# test_ConvTrans #######################
add_unittest_without_exec(test_ConvTrans
test_ConvTrans.cpp
LayerGradUtil.cpp)
add_test(NAME test_ConvTrans
COMMAND test_ConvTrans)
################# test_PriorBox #######################
add_unittest_without_exec(test_PriorBox
test_PriorBox.cpp
LayerGradUtil.cpp)
add_test(NAME test_PriorBox
COMMAND test_PriorBox)
################# test_DetectionOutput #######################
add_unittest_without_exec(test_DetectionOutput
test_DetectionOutput.cpp
LayerGradUtil.cpp)
add_test(NAME test_DetectionOutput
COMMAND test_DetectionOutput)
################# test_ConvUnify #######################
add_unittest_without_exec(test_ConvUnify
test_ConvUnify.cpp
LayerGradUtil.cpp)
add_test(NAME test_ConvUnify
COMMAND test_ConvUnify)
################# test_BatchNorm #######################
add_unittest_without_exec(test_BatchNorm
test_BatchNorm.cpp
LayerGradUtil.cpp)
add_test(NAME test_BatchNorm
COMMAND test_BatchNorm)
################# test_KmaxSeqScore #######################
add_unittest_without_exec(test_KmaxSeqScore
test_KmaxSeqScore.cpp
LayerGradUtil.cpp)
add_test(NAME test_KmaxSeqScore
COMMAND test_KmaxSeqScore)
if(NOT MOBILE_INFERENCE)
################## test_Evaluator #######################
add_unittest(test_Evaluator
test_Evaluator.cpp)
endif()
################ test_LinearChainCRF ####################
add_simple_unittest(test_LinearChainCRF)
############## test_MultinomialSampler ###################
add_simple_unittest(test_MultinomialSampler)
############## test_PyDataProvider ########################
if(WITH_PYTHON)
add_unittest_without_exec(test_PyDataProvider
......@@ -125,9 +47,6 @@ if(WITH_PYTHON)
WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle)
endif()
############### test_RecurrentLayer #######################
add_simple_unittest(test_RecurrentLayer)
############### test_WarpCTCLayer #######################
if(NOT WITH_DOUBLE)
add_unittest_without_exec(test_WarpCTCLayer
......@@ -139,19 +58,33 @@ if(NOT WITH_DOUBLE)
endif()
if(NOT MOBILE_INFERENCE)
############### test_RecurrentGradientMachine ###############
# TODO(yuyang18): There is some bug in test_RecurrentGradientMachine
# I will fix it.
add_unittest_without_exec(test_RecurrentGradientMachine
test_RecurrentGradientMachine.cpp)
add_test(NAME test_RecurrentGradientMachine
COMMAND .set_python_path.sh -d
${PADDLE_SOURCE_DIR}/python:${PADDLE_SOURCE_DIR}/paddle/gserver/tests
${CMAKE_CURRENT_BINARY_DIR}/test_RecurrentGradientMachine
WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle)
endif()
################### test_ProtoDataProvider ############
add_unittest_without_exec(test_ProtoDataProvider
test_ProtoDataProvider.cpp)
if(NOT MOBILE_INFERENCE)
# test_ProtoDataProvider will mkdir as same name,
# so if WORKING_DIRECTORY is default directory, then
# mkdir will get error.
add_test(NAME test_ProtoDataProvider
COMMAND ${CMAKE_CURRENT_BINARY_DIR}/test_ProtoDataProvider
WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle)
################## test_Evaluator #######################
add_unittest(test_Evaluator
test_Evaluator.cpp)
############### test_RecurrentGradientMachine ###############
# TODO(yuyang18): There is some bug in test_RecurrentGradientMachine
# I will fix it.
add_unittest_without_exec(test_RecurrentGradientMachine
test_RecurrentGradientMachine.cpp)
add_test(NAME test_RecurrentGradientMachine
COMMAND .set_python_path.sh -d
${PADDLE_SOURCE_DIR}/python:${PADDLE_SOURCE_DIR}/paddle/gserver/tests
${CMAKE_CURRENT_BINARY_DIR}/test_RecurrentGradientMachine
WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle)
############### test_NetworkCompare ###############
add_unittest_without_exec(test_NetworkCompare
test_NetworkCompare.cpp)
if(WITH_GPU)
......
paddle/gserver/tests/test_Expand.cpp 0 → 100644
浏览文件 @ 1e8300fa
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include <gtest/gtest.h>
#include <string>
#include <vector>
#include "LayerGradUtil.h"
#include "paddle/testing/TestUtil.h"
using namespace paddle; // NOLINT
using namespace std; // NOLINT
// Do one forward pass of expand layer and check to see if its output
// matches the given result.(Test onlyCPU currently.)
void doOneExpandTest(string trans_type,
bool hasSubseq,
bool useGpu,
Argument& input1,
Argument& input2,
Argument& result) {
FLAGS_use_gpu = false;
// Setting up the expand layer
TestConfig config;
config.layerConfig.set_type("expand");
auto inputType1 =
trans_type == "non-seq" ? INPUT_DENSE_DIM_DATA : INPUT_SEQUENCE_DATA;
config.inputDefs.push_back({inputType1, "layer0", 1, 0});
auto inputType2 =
hasSubseq ? INPUT_HASSUB_SEQUENCE_DATA : INPUT_SEQUENCE_DATA;
config.inputDefs.push_back({inputType2, "layer1", 1, 0});
config.layerConfig.add_inputs();
config.layerConfig.add_inputs();
config.layerConfig.set_trans_type(trans_type);
// data layer initialize
std::vector<DataLayerPtr> dataLayers;
LayerMap layerMap;
vector<Argument> datas;
initDataLayer(
config, &dataLayers, &datas, &layerMap, "expand", 1, false, useGpu);
dataLayers[0]->getOutput() = input1;
dataLayers[1]->getOutput() = input2;
// test layer initialize
std::vector<ParameterPtr> parameters;
LayerPtr expandLayer;
initTestLayer(config, &layerMap, &parameters, &expandLayer);
expandLayer->forward(PASS_GC);
checkMatrixEqual(expandLayer->getOutputValue(), result.value);
}
TEST(Layer, ExpandLayerFwd) {
bool useGpu = false;
// Assume batch_size =3 in all cases.
// CPU case 1. non-seq expand to seq
// input1 = 1,2,3
// input2 = [4,5],[6],[7,8,9]
// result = [1,1],[2],[3,3,3]
Argument input1, input2, result;
input1.value = Matrix::create(3, 1, false, useGpu);
real input1Data[] = {1, 2, 3};
input1.value->setData(input1Data);
input2.value = Matrix::create(6, 1, false, useGpu);
real input2Data[] = {4, 5, 6, 7, 8, 9};
input2.value->setData(input2Data);
input2.sequenceStartPositions = ICpuGpuVector::create(4, useGpu);
int input2Seq[] = {0, 2, 3, 6};
input2.sequenceStartPositions->copyFrom(input2Seq, 4, useGpu);
result.value = Matrix::create(6, 1, false, useGpu);
real resultData[] = {1, 1, 2, 3, 3, 3};
result.value->setData(resultData);
doOneExpandTest("non-seq", false, useGpu, input1, input2, result);
// CPU case 2. non-seq expand to sub-seq
// NOTE: input1.batch_size == input2.sequencelength in this case.
// i.e, input1 expands by input2.sequence
// input1 = 1,2,3
// input2 = [[4,5]],[[6]],[[7],[8,9]]
// result = [[1,1]],[[2]],[[3],[3,3]]
input2.subSequenceStartPositions = ICpuGpuVector::create(5, useGpu);
int input2SubSeq[] = {0, 2, 3, 4, 6};
input2.subSequenceStartPositions->copyFrom(input2SubSeq, 5, useGpu);
doOneExpandTest("non-seq", true, useGpu, input1, input2, result);
// CPU case 3. seq expand to sub-seq
// input1 = [1,2],[3],[4]
// input2 = [[4,5]],[[6]],[[7],[8,9]]
// result = [[1,1]],[[2]],[[3],[4,4]]
Matrix::resizeOrCreate(input1.value, 4, 1, false, useGpu);
real input1Data_case3[] = {1, 2, 3, 4};
input1.value->setData(input1Data_case3);
input1.sequenceStartPositions = ICpuGpuVector::create(4, useGpu);
int input1Seq[] = {0, 2, 3, 4};
input1.sequenceStartPositions->copyFrom(input1Seq, 4, useGpu);
real resultData_case3[] = {1, 1, 2, 3, 4, 4};
result.value->setData(resultData_case3);
doOneExpandTest("seq", true, useGpu, input1, input2, result);
}
int main(int argc, char** argv) {
testing::InitGoogleTest(&argc, argv);
initMain(argc, argv);
return RUN_ALL_TESTS();
}
paddle/memory/detail/buddy_allocator.cc
浏览文件 @ 1e8300fa
......@@ -27,11 +27,11 @@ BuddyAllocator::BuddyAllocator(SystemAllocator* system_allocator,
system_allocator_(std::move(system_allocator)) {}
BuddyAllocator::~BuddyAllocator() {
VLOG(3) << "BuddyAllocator Disconstructor makes sure that all of these "
"have actually been freed";
VLOG(10) << "BuddyAllocator Disconstructor makes sure that all of these "
"have actually been freed";
while (!pool_.empty()) {
auto block = static_cast<MemoryBlock*>(std::get<2>(*pool_.begin()));
VLOG(3) << "Free from block (" << block << ", " << max_chunk_size_ << ")";
VLOG(10) << "Free from block (" << block << ", " << max_chunk_size_ << ")";
system_allocator_->Free(block, max_chunk_size_, block->index(cache_));
cache_.invalidate(block);
......@@ -51,11 +51,12 @@ void* BuddyAllocator::Alloc(size_t unaligned_size) {
// acquire the allocator lock
std::lock_guard<std::mutex> lock(mutex_);
VLOG(3) << "Allocate " << unaligned_size << " bytes from chunk size " << size;
VLOG(10) << "Allocate " << unaligned_size << " bytes from chunk size "
<< size;
// if the allocation is huge, send directly to the system allocator
if (size > max_chunk_size_) {
VLOG(3) << "Allocate from system allocator.";
VLOG(10) << "Allocate from system allocator.";
return SystemAlloc(size);
}
......@@ -70,9 +71,9 @@ void* BuddyAllocator::Alloc(size_t unaligned_size) {
return nullptr;
}
} else {
VLOG(3) << "Allocation from existing memory block " << std::get<2>(*it)
<< " at address "
<< reinterpret_cast<MemoryBlock*>(std::get<2>(*it))->data();
VLOG(10) << "Allocation from existing memory block " << std::get<2>(*it)
<< " at address "
<< reinterpret_cast<MemoryBlock*>(std::get<2>(*it))->data();
}
total_used_ += size;
......@@ -89,10 +90,10 @@ void BuddyAllocator::Free(void* p) {
// Acquire the allocator lock
std::lock_guard<std::mutex> lock(mutex_);
VLOG(3) << "Free from address " << block;
VLOG(10) << "Free from address " << block;
if (block->type(cache_) == MemoryBlock::HUGE_CHUNK) {
VLOG(3) << "Free directly from system allocator";
VLOG(10) << "Free directly from system allocator";
system_allocator_->Free(block, block->total_size(cache_),
block->index(cache_));
......@@ -109,8 +110,8 @@ void BuddyAllocator::Free(void* p) {
// Trying to merge the right buddy
if (block->has_right_buddy(cache_)) {
VLOG(3) << "Merging this block " << block << " with its right buddy "
<< block->right_buddy(cache_);
VLOG(10) << "Merging this block " << block << " with its right buddy "
<< block->right_buddy(cache_);
auto right_buddy = block->right_buddy(cache_);
......@@ -127,8 +128,8 @@ void BuddyAllocator::Free(void* p) {
// Trying to merge the left buddy
if (block->has_left_buddy(cache_)) {
VLOG(3) << "Merging this block " << block << " with its left buddy "
<< block->left_buddy(cache_);
VLOG(10) << "Merging this block " << block << " with its left buddy "
<< block->left_buddy(cache_);
auto left_buddy = block->left_buddy(cache_);
......@@ -144,8 +145,8 @@ void BuddyAllocator::Free(void* p) {
}
// Dumping this block into pool
VLOG(3) << "Inserting free block (" << block << ", "
<< block->total_size(cache_) << ")";
VLOG(10) << "Inserting free block (" << block << ", "
<< block->total_size(cache_) << ")";
pool_.insert(
IndexSizeAddress(block->index(cache_), block->total_size(cache_), block));
......@@ -164,7 +165,7 @@ void* BuddyAllocator::SystemAlloc(size_t size) {
size_t index = 0;
void* p = system_allocator_->Alloc(index, size);
VLOG(3) << "Allocated " << p << " from system allocator.";
VLOG(10) << "Allocated " << p << " from system allocator.";
if (p == nullptr) return nullptr;
......@@ -190,8 +191,8 @@ BuddyAllocator::PoolSet::iterator BuddyAllocator::RefillPool() {
if (p == nullptr) return pool_.end();
VLOG(3) << "Creating and inserting new block " << p
<< " from system allocator";
VLOG(10) << "Creating and inserting new block " << p
<< " from system allocator";
static_cast<MemoryBlock*>(p)->init(cache_, MemoryBlock::FREE_CHUNK, index,
max_chunk_size_, nullptr, nullptr);
......@@ -235,19 +236,19 @@ void* BuddyAllocator::SplitToAlloc(BuddyAllocator::PoolSet::iterator it,
auto block = static_cast<MemoryBlock*>(std::get<2>(*it));
pool_.erase(it);
VLOG(3) << "Split block (" << block << ", " << block->total_size(cache_)
<< ") into";
VLOG(10) << "Split block (" << block << ", " << block->total_size(cache_)
<< ") into";
block->split(cache_, size);
VLOG(3) << "Left block (" << block << ", " << block->total_size(cache_)
<< ")";
VLOG(10) << "Left block (" << block << ", " << block->total_size(cache_)
<< ")";
block->set_type(cache_, MemoryBlock::ARENA_CHUNK);
// the rest of memory if exist
if (block->has_right_buddy(cache_)) {
if (block->right_buddy(cache_)->type(cache_) == MemoryBlock::FREE_CHUNK) {
VLOG(3) << "Insert right block (" << block->right_buddy(cache_) << ", "
<< block->right_buddy(cache_)->total_size(cache_) << ")";
VLOG(10) << "Insert right block (" << block->right_buddy(cache_) << ", "
<< block->right_buddy(cache_)->total_size(cache_) << ")";
pool_.insert(
IndexSizeAddress(block->right_buddy(cache_)->index(cache_),
......@@ -274,7 +275,7 @@ void BuddyAllocator::CleanIdleFallBackAlloc() {
return;
}
VLOG(3) << "Return block " << block << " to fallback allocator.";
VLOG(10) << "Return block " << block << " to fallback allocator.";
system_allocator_->Free(block, max_chunk_size_, block->index(cache_));
cache_.invalidate(block);
......@@ -310,7 +311,7 @@ void BuddyAllocator::CleanIdleNormalAlloc() {
MemoryBlock* block = static_cast<MemoryBlock*>(std::get<2>(*pool));
VLOG(3) << "Return block " << block << " to base allocator.";
VLOG(10) << "Return block " << block << " to base allocator.";
system_allocator_->Free(block, max_chunk_size_, block->index(cache_));
cache_.invalidate(block);
......
paddle/memory/detail/meta_cache.cc
浏览文件 @ 1e8300fa
......@@ -30,7 +30,7 @@ Metadata MetadataCache::load(const MemoryBlock* block) {
return existing_metadata->second;
} else {
auto* meta = reinterpret_cast<const Metadata*>(block);
VLOG(3) << "Load MetaData type=" << meta->type;
VLOG(10) << "Load MetaData type=" << meta->type;
PADDLE_ASSERT(meta->check_guards());
return *reinterpret_cast<const Metadata*>(block);
}
......
paddle/memory/detail/system_allocator.cc
浏览文件 @ 1e8300fa
......@@ -41,7 +41,16 @@ void* CPUAllocator::Alloc(size_t& index, size_t size) {
index = 0; // unlock memory
void* p = malloc(size);
void* p;
#ifdef PADDLE_USE_MKLDNN
// refer to https://github.com/01org/mkl-dnn/blob/master/include/mkldnn.hpp
// memory alignment
PADDLE_ENFORCE_EQ(posix_memalign(&p, 4096ul, size), 0);
#else
PADDLE_ENFORCE_EQ(posix_memalign(&p, 32ul, size), 0);
#endif
PADDLE_ENFORCE(p, "Fail to allocate CPU memory: size = %d .", size);
if (p != nullptr) {
if (FLAGS_use_pinned_memory) {
......
paddle/memory/memory.cc
浏览文件 @ 1e8300fa
......@@ -39,15 +39,15 @@ BuddyAllocator* GetCPUBuddyAllocator() {
template <>
void* Alloc<platform::CPUPlace>(platform::CPUPlace place, size_t size) {
VLOG(3) << "Allocate " << size << " bytes on " << platform::Place(place);
VLOG(10) << "Allocate " << size << " bytes on " << platform::Place(place);
void* p = GetCPUBuddyAllocator()->Alloc(size);
VLOG(3) << " pointer=" << p;
VLOG(10) << " pointer=" << p;
return p;
}
template <>
void Free<platform::CPUPlace>(platform::CPUPlace place, void* p) {
VLOG(3) << "Free pointer=" << p << " on " << platform::Place(place);
VLOG(10) << "Free pointer=" << p << " on " << platform::Place(place);
GetCPUBuddyAllocator()->Free(p);
}
......@@ -69,11 +69,12 @@ BuddyAllocator* GetGPUBuddyAllocator(int gpu_id) {
platform::GpuMinChunkSize(),
platform::GpuMaxChunkSize());
}
VLOG(3) << "\n\nNOTE: each GPU device use "
<< FLAGS_fraction_of_gpu_memory_to_use * 100 << "% of GPU memory.\n"
<< "You can set environment variable '"
<< platform::kEnvFractionGpuMemoryToUse
<< "' to change the fraction of GPU usage.\n\n";
VLOG(10) << "\n\nNOTE: each GPU device use "
<< FLAGS_fraction_of_gpu_memory_to_use * 100
<< "% of GPU memory.\n"
<< "You can set environment variable '"
<< platform::kEnvFractionGpuMemoryToUse
<< "' to change the fraction of GPU usage.\n\n";
}
platform::SetDeviceId(gpu_id);
return as[gpu_id];
......
paddle/operators/accuracy_op.cc
浏览文件 @ 1e8300fa
......@@ -22,22 +22,35 @@ class AccuracyOp : public framework::OperatorWithKernel {
using framework::OperatorWithKernel::OperatorWithKernel;
void InferShape(framework::InferShapeContext *ctx) const override {
PADDLE_ENFORCE(ctx->HasInput("Inference"),
"Input(Inference) of AccuracyOp should not be null.");
PADDLE_ENFORCE(ctx->HasInput("Out"),
"Input (Out) of accuracy op should not be null.");
PADDLE_ENFORCE(ctx->HasInput("Indices"),
"Input (Indices) of accuracy op should not be null.");
PADDLE_ENFORCE(ctx->HasInput("Label"),
"Input(Label) of AccuracyOp should not be null.");
"Input (Label) of accuracy op should not be null.");
PADDLE_ENFORCE(ctx->HasOutput("Accuracy"),
"Output(Accuracy) of AccuracyOp should not be null.");
"Output (Accuracy) of AccuracyOp should not be null.");
auto inference_dim = ctx->GetInputDim("Inference");
auto inference_dim = ctx->GetInputDim("Out");
auto label_dim = ctx->GetInputDim("Label");
// Assume indices has same shape with infernece, because
// it's the output of topk.
PADDLE_ENFORCE_EQ(label_dim.size(), 1, "label must be a vector");
PADDLE_ENFORCE_EQ(label_dim.size(), 2, "label's rank must be 2.");
PADDLE_ENFORCE_EQ(label_dim[1], 1, "label's second dimension must be 1");
PADDLE_ENFORCE_EQ(inference_dim[0], label_dim[0],
"inference size must be the same as label size");
"the inference tensor's num_rows must be"
" the same as label.");
ctx->SetOutputDim("Accuracy", {1});
ctx->ShareLoD("Inference", /*->*/ "Accuracy");
ctx->ShareLoD("Out", /*->*/ "Accuracy");
}
protected:
// IndicateDataType
framework::DataType IndicateDataType(
const framework::ExecutionContext &ctx) const override {
return framework::ToDataType(ctx.Input<Tensor>("Out")->type());
}
};
......@@ -47,7 +60,8 @@ class AccuracyOpMaker : public framework::OpProtoAndCheckerMaker {
framework::OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
// TODO(typhoonzero): support both inference value and indices.
AddInput("Inference", "topk(indices) the network output");
AddInput("Out", "topk (inferences) the network output");
AddInput("Indices", "topk (indices) the network output");
AddInput("Label", "Label of the training data");
// TODO(typhoonzero): AddInput("Weight", ...
AddOutput("Accuracy", "The accuracy of current batch");
......@@ -58,7 +72,7 @@ The accuracy is:
.. math::
accuracy = \\frac{NumOfCorrectPredicts}{NumOfAllSamples})
Both the input `Inference` and `Label` can carry the LoD (Level of Details)
Both the input `Out` and `Label` can carry the LoD (Level of Details)
information, or not. But the output only shares the LoD with input `Inference`.
)DOC");
}
......@@ -68,7 +82,10 @@ information, or not. But the output only shares the LoD with input `Inference`.
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP_WITHOUT_GRADIENT(accuracy, ops::AccuracyOp, ops::AccuracyOpMaker);
REGISTER_OP_CPU_KERNEL(
accuracy, ops::AccuracyKernel<paddle::platform::CPUPlace, int>,
ops::AccuracyKernel<paddle::platform::CPUPlace, int64_t>);
REGISTER_OPERATOR(accuracy, ops::AccuracyOp, ops::AccuracyOpMaker,
paddle::framework::EmptyGradOpMaker);
// FIXME(typhoonzero): types of T is for infernece data.
// label data is always int.
REGISTER_OP_CPU_KERNEL(accuracy,
ops::AccuracyKernel<paddle::platform::CPUPlace, float>,
ops::AccuracyKernel<paddle::platform::CPUPlace, double>);
paddle/operators/accuracy_op.cu
浏览文件 @ 1e8300fa
......@@ -21,9 +21,10 @@ namespace paddle {
namespace operators {
using platform::PADDLE_CUDA_NUM_THREADS;
template <typename T, int BlockSize>
__global__ void AccuracyCudaKernel(const int N, const int D, const T* Xdata,
const T* labeldata, float* accuracy) {
template <int BlockSize>
__global__ void AccuracyCudaKernel(const int N, const int D,
const int64_t* Xdata,
const int64_t* labeldata, float* accuracy) {
int count = 0;
__shared__ int total[BlockSize];
......@@ -52,13 +53,14 @@ class AccuracyOpCUDAKernel : public framework::OpKernel<T> {
void Compute(const framework::ExecutionContext& ctx) const override {
PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
"It must use GPUPlace.");
auto* inference = ctx.Input<Tensor>("Inference");
auto* inference = ctx.Input<Tensor>("Out");
auto* indices = ctx.Input<Tensor>("Indices");
auto* label = ctx.Input<Tensor>("Label");
auto* accuracy = ctx.Output<Tensor>("Accuracy");
// FIXME(typhoonzero): only support indices currently
// if add support for output values, how to detect the data type?
const T* inference_data = inference->data<T>();
const T* label_data = label->data<T>();
const int64_t* indices_data = indices->data<int64_t>();
const int64_t* label_data = label->data<int64_t>();
float* accuracy_data = accuracy->mutable_data<float>(ctx.GetPlace());
size_t num_samples = inference->dims()[0];
......@@ -69,11 +71,11 @@ class AccuracyOpCUDAKernel : public framework::OpKernel<T> {
return;
}
AccuracyCudaKernel<T, PADDLE_CUDA_NUM_THREADS><<<
AccuracyCudaKernel<PADDLE_CUDA_NUM_THREADS><<<
1, PADDLE_CUDA_NUM_THREADS, 0,
reinterpret_cast<const platform::CUDADeviceContext&>(
ctx.device_context())
.stream()>>>(num_samples, infer_width, inference_data, label_data,
.stream()>>>(num_samples, infer_width, indices_data, label_data,
accuracy_data);
}
};
......@@ -81,5 +83,7 @@ class AccuracyOpCUDAKernel : public framework::OpKernel<T> {
} // namespace operators
} // namespace paddle
REGISTER_OP_GPU_KERNEL(accuracy, paddle::operators::AccuracyOpCUDAKernel<int>,
paddle::operators::AccuracyOpCUDAKernel<int64_t>);
// FIXME(typhoonzero): types of T is for infernece data.
// label data is always int
REGISTER_OP_GPU_KERNEL(accuracy, paddle::operators::AccuracyOpCUDAKernel<float>,
paddle::operators::AccuracyOpCUDAKernel<double>);
paddle/operators/accuracy_op.h
浏览文件 @ 1e8300fa
......@@ -38,14 +38,15 @@ template <typename Place, typename T>
class AccuracyKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
auto* inference = ctx.Input<Tensor>("Inference");
auto* inference = ctx.Input<Tensor>("Out");
auto* indices = ctx.Input<Tensor>("Indices");
auto* label = ctx.Input<Tensor>("Label");
auto* accuracy = ctx.Output<Tensor>("Accuracy");
float* accuracy_data = accuracy->mutable_data<float>(ctx.GetPlace());
const T* inference_data = inference->data<T>();
const T* label_data = label->data<T>();
const int64_t* indices_data = indices->data<int64_t>();
const int64_t* label_data = label->data<int64_t>();
size_t num_samples = inference->dims()[0];
size_t class_dim = inference->dims()[1];
......@@ -60,7 +61,7 @@ class AccuracyKernel : public framework::OpKernel<T> {
for (size_t i = 0; i < num_samples; ++i) {
PADDLE_ENFORCE_GE(label_data[i], 0, "label must >= 0");
for (size_t j = 0; j < class_dim; ++j) {
if (inference_data[i * class_dim + j] == label_data[i]) {
if (indices_data[i * class_dim + j] == label_data[i]) {
++num_correct;
break;
}
......
paddle/operators/activation_op.h
浏览文件 @ 1e8300fa
......@@ -547,6 +547,7 @@ struct ELUGradFunctor : public BaseActivationFunctor<T> {
}
};
// FIXME(qijun) https://github.com/PaddlePaddle/Paddle/issues/5198
template <typename T>
struct PowFunctor : public BaseActivationFunctor<T> {
float factor;
......
paddle/operators/auc_op.cc
浏览文件 @ 1e8300fa
......@@ -23,18 +23,26 @@ class AucOp : public framework::OperatorWithKernel {
protected:
void InferShape(framework::InferShapeContext *ctx) const override {
PADDLE_ENFORCE(ctx->HasInput("Inference"),
"Input of Inference must be initialized.");
PADDLE_ENFORCE(ctx->HasInput("Out"), "Input of Out must be initialized.");
PADDLE_ENFORCE(ctx->HasInput("Indices"),
"Input of Indices must be initialized.");
PADDLE_ENFORCE(ctx->HasInput("Label"),
"Input of Label must be initialized.");
auto inference_dim = ctx->GetInputDim("Inference");
auto label_dim = ctx->GetInputDim("Label");
auto inference_height = ctx->GetInputDim("Out")[0];
auto label_height = ctx->GetInputDim("Label")[0];
PADDLE_ENFORCE_EQ(inference_dim, label_dim,
"inference and label should have same shape");
PADDLE_ENFORCE_EQ(inference_height, label_height,
"Out and Label should have same height.");
ctx->SetOutputDim("AUC", {1});
ctx->ShareLoD("Inference", /*->*/ "AUC");
ctx->ShareLoD("Out", /*->*/ "AUC");
}
protected:
// IndicateDataType
framework::DataType IndicateDataType(
const framework::ExecutionContext &ctx) const override {
return framework::ToDataType(ctx.Input<Tensor>("Out")->type());
}
};
......@@ -42,12 +50,18 @@ class AucOpMaker : public framework::OpProtoAndCheckerMaker {
public:
AucOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("Inference",
"A floating point tensor of arbitrary shape and whose values"
"are in the range [0, 1].");
AddInput("Out",
"A floating point 2D tensor, values are in the range [0, 1]."
"Each row is descend sorted. This input should be the"
"output of topk."
"Typically, this tensor indicates the probability of each label");
AddInput("Indices",
"An int 2D tensor, indicating the indices of original"
"tensor before sort. Typically, this tensor indicates which label"
"the probability stands for.");
AddInput("Label",
"A tensor whose shape matches "
"Inference. Will be cast to bool.");
"A 2D int tensor indicating the label of the training data."
"The height is batch size and width is always 1.");
// TODO(typhoonzero): support weight input
AddOutput("AUC",
"A scalar representing the "
......
paddle/operators/auc_op.h
浏览文件 @ 1e8300fa
......@@ -29,7 +29,7 @@ template <typename Place, typename T>
class AucKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
auto* inference = ctx.Input<Tensor>("Inference");
auto* inference = ctx.Input<Tensor>("Out");
auto* label = ctx.Input<Tensor>("Label");
auto* auc = ctx.Output<Tensor>("AUC");
......@@ -46,18 +46,11 @@ class AucKernel : public framework::OpKernel<T> {
thresholds_list[0] = 0.0f - kEpsilon;
thresholds_list[num_thresholds - 1] = 1.0f + kEpsilon;
size_t num_samples = inference->numel();
size_t batch_size = inference->dims()[0];
size_t inference_width = inference->dims()[1];
const T* inference_data = inference->data<T>();
Tensor label_casted;
label_casted.Resize(label->dims());
bool* label_casted_data = label_casted.mutable_data<bool>(ctx.GetPlace());
const int* label_data = label->data<int>();
// cast label_data to bool
for (size_t i = 0; i < num_samples; i++) {
label_casted_data[i] = static_cast<bool>(label_data[i]);
}
const int64_t* label_data = label->data<int64_t>();
// Create local tensor for storing the curve: TP, FN, TN, FP
// TODO(typhoonzero): use eigen op to caculate these values.
......@@ -68,23 +61,27 @@ class AucKernel : public framework::OpKernel<T> {
true_negative.Resize({num_thresholds});
false_positive.Resize({num_thresholds});
int* tp_data = true_positive.mutable_data<int>(ctx.GetPlace());
int* fn_data = false_negative.mutable_data<int>(ctx.GetPlace());
int* tn_data = true_negative.mutable_data<int>(ctx.GetPlace());
int* fp_data = false_positive.mutable_data<int>(ctx.GetPlace());
int64_t* tp_data = true_positive.mutable_data<int64_t>(ctx.GetPlace());
int64_t* fn_data = false_negative.mutable_data<int64_t>(ctx.GetPlace());
int64_t* tn_data = true_negative.mutable_data<int64_t>(ctx.GetPlace());
int64_t* fp_data = false_positive.mutable_data<int64_t>(ctx.GetPlace());
for (int idx_thresh = 0; idx_thresh < num_thresholds; idx_thresh++) {
// caculate TP, FN, TN, FP for current thresh
int tp = 0, fn = 0, tn = 0, fp = 0;
for (size_t i = 0; i < num_samples; i++) {
if (label_casted_data[i]) {
if (inference_data[i] >= (thresholds_list[idx_thresh])) {
int64_t tp = 0, fn = 0, tn = 0, fp = 0;
for (size_t i = 0; i < batch_size; i++) {
// NOTE: label_data used as bool, labels >0 will be treated as true.
if (label_data[i]) {
// use first(max) data in each row
if (inference_data[i * inference_width] >=
(thresholds_list[idx_thresh])) {
tp++;
} else {
fn++;
}
} else {
if (inference_data[i] >= (thresholds_list[idx_thresh])) {
if (inference_data[i * inference_width] >=
(thresholds_list[idx_thresh])) {
fp++;
} else {
tn++;
......
paddle/operators/cross_entropy_op.cc
浏览文件 @ 1e8300fa
......@@ -28,8 +28,9 @@ class CrossEntropyOp : public framework::OperatorWithKernel {
auto x_dims = ctx->GetInputDim("X");
auto label_dims = ctx->GetInputDim("Label");
PADDLE_ENFORCE_EQ(x_dims.size(), 2, "Input(X)'s rank should be 2.");
PADDLE_ENFORCE_EQ(label_dims.size(), 2, "Input(Label)'s rank should be 2.");
PADDLE_ENFORCE_EQ(x_dims.size(), 2UL, "Input(X)'s rank should be 2.");
PADDLE_ENFORCE_EQ(label_dims.size(), 2UL,
"Input(Label)'s rank should be 2.");
PADDLE_ENFORCE_EQ(x_dims[0], label_dims[0],
"The 1st dimension of Input(X) and Input(Label) should "
"be equal.");
......@@ -38,8 +39,8 @@ class CrossEntropyOp : public framework::OperatorWithKernel {
"If Attr(soft_label) == true, the 2nd dimension of "
"Input(X) and Input(Label) should be equal.");
} else {
PADDLE_ENFORCE_EQ(label_dims[1], 1,
"If Attr(soft_label) == false, the 2nd dimension of "
PADDLE_ENFORCE_EQ(label_dims[1], 1UL,
"If Attr(softLabel) == false, the 2nd dimension of "
"Input(Label) should be 1.");
}
......@@ -48,7 +49,8 @@ class CrossEntropyOp : public framework::OperatorWithKernel {
}
protected:
// CrossEntropy's data type just determined by "X"
// Explicitly set that data type of the output of the cross_entropy operator
// is determined by its input "X".
framework::DataType IndicateDataType(
const framework::ExecutionContext& ctx) const override {
return framework::ToDataType(ctx.Input<Tensor>("X")->type());
......
paddle/operators/dynamic_recurrent_op_test.cc
浏览文件 @ 1e8300fa
......@@ -51,7 +51,7 @@ class RNNAlgorithmTestHelper : public ::testing::Test {
CreateGlobalVariables();
auto op_desc = CreateOpDesc();
op = paddle::framework::OpRegistry::CreateOp(op_desc, nullptr);
op = paddle::framework::OpRegistry::CreateOp(op_desc);
dop = &(dynamic_cast<DynamicRecurrentOp*>(op.get())->rnn);
InitCacheManually();
InitStepNet();
......
paddle/operators/gaussian_random_op.cc
浏览文件 @ 1e8300fa
......@@ -45,14 +45,14 @@ class GaussianRandomOp : public framework::OperatorWithKernel {
void InferShape(framework::InferShapeContext* ctx) const override {
PADDLE_ENFORCE(ctx->HasOutput("Out"),
"Output(Out) of GaussianRandomOp should not be null.");
auto dims = ctx->Attrs().Get<std::vector<int>>("dims");
auto shape = ctx->Attrs().Get<std::vector<int>>("shape");
std::vector<int64_t> temp;
temp.reserve(dims.size());
for (auto dim : dims) {
temp.reserve(shape.size());
for (auto dim : shape) {
temp.push_back(static_cast<int64_t>(dim));
}
PADDLE_ENFORCE(dims.size() > 0UL,
"dims can be one int or array. dims must be set.");
PADDLE_ENFORCE(shape.size() > 0UL,
"shape can be one int or array. shape must be set.");
ctx->SetOutputDim("Out", framework::make_ddim(temp));
}
......@@ -74,7 +74,7 @@ GaussianRandom operator.
Use to initialize tensor with gaussian random generator.
)DOC");
AddAttr<std::vector<int>>("dims", "The dimension of random tensor.");
AddAttr<std::vector<int>>("shape", "The dimension of random tensor.");
AddAttr<float>("mean", "mean of random tensor.").SetDefault(.0f);
AddAttr<float>("std", "std of random tensor.").SetDefault(1.0f);
AddAttr<int>("seed",
......
paddle/operators/linear_chain_crf_op.cc 0 → 100644
浏览文件 @ 1e8300fa
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/operators/linear_chain_crf_op.h"
namespace paddle {
namespace operators {
class LinearChainCRFOpMaker : public framework::OpProtoAndCheckerMaker {
public:
LinearChainCRFOpMaker(framework::OpProto* proto,
framework::OpAttrChecker* op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput(
"Emission",
"(LoDTensor, default: LoDTensor<float>). "
"The unscaled emission weight matrix for the linear chain CRF. "
"This input is a LoDTensor with shape [N x D] where N is the size of "
"the mini-batch and D is the total tag number.");
AddInput(
"Transition",
"(Tensor, default: Tensor<float>). A Tensor with shape [(D + 2) x D]. "
"The learnable parameter for the linear_chain_crf operator. "
"See more details in the operator's comments.");
AddInput(
"Label",
"(LoDTensor, default: LoDTensor<int>). The ground truth which is a 2-D "
"LoDTensor with shape [N x 1], where N is the total element number in "
"a mini-batch.");
AddOutput(
"Alpha",
"Tensor, default: Tensor<float>. The forward vectors for the entire "
"batch. A two dimensional tensor with shape [N x D], "
"denoted as \f$\alpha\f$. \f$\alpha$\f is a memo table used to "
"calculate the normalization factor in CRF. \f$\alpha[k, v]$\f stores "
"the unnormalized probabilites of all possible unfinished sequences of "
"tags that end at position \f$k$\f with tag \f$v$\f. For each \f$k$\f, "
"\f$\alpha[k, v]$\f is a vector of length \f$D$\f with a component for "
"each tag value \f$v$\f. This vector is called a forward vecotr and "
"will also be used in backward computations.")
.AsIntermediate();
AddOutput("EmissionExps",
"The exponentials of Input(Emission). This is an intermediate "
"computational result in forward computation, and will be reused "
"in backward computation.")
.AsIntermediate();
AddOutput("TransitionExps",
"The exponentials of Input(Transition). This is an intermediate "
"computational result in forward computation, and will be reused "
"in backward computation.")
.AsIntermediate();
AddOutput(
"LogLikelihood",
"(Tensor, default: Tensor<float>). The logarithm of the conditional "
"likelihood of each training sample in a mini-batch. This is a 2-D "
"tensor with shape [S x 1], where S is the sequence number in a "
"mini-batch. Note: S is equal to the sequence number in a mini-batch. "
"The output is no longer a LoDTensor.");
AddComment(R"DOC(
Conditional Random Field defines an undirected probabilistic graph with nodes
denoting random variables and edges denoting dependencies between these
variables. CRF learns the conditional probability \f$P(Y|X)\f$, where
\f$X = (x_1, x_2, ... , x_n)\f$ are structured inputs and
\f$Y = (y_1, y_2, ... , y_n)\f$ are labels for the inputs.
Linear chain CRF is a special case of CRF that is useful for sequence labeling
task. Sequence labeling tasks do not assume a lot of conditional
independences among inputs. The only constraint they impose is that the input
and output must be linear sequences. Thus, the graph of such a CRF is a simple
chain or a line, which results in the linear chain CRF.
This operator implements the Forward-Backward algorithm for the linear chain
CRF. Please see http://www.cs.columbia.edu/~mcollins/fb.pdf and
http://cseweb.ucsd.edu/~elkan/250Bwinter2012/loglinearCRFs.pdf for reference.
Equation:
- Denote Input(Emission) to this operator as \f$x\f$ here.
- The first D values of Input(Transition) to this operator are for starting
weights, denoted as \f$a\f$ here.
- The next D values of Input(Transition) of this operator are for ending
weights, denoted as \f$b\f$ here.
- The remaning values of Input(Transition) are for transition weights,
denoted as \f$w\f$ here.
- Denote Input(Label) as \f$s\f$ here.
The probability of a sequence \f$s\f$ of length \f$L\f$ is defined as:
\f$P(s) = (1/Z) exp(a_{s_1} + b_{s_L}
+ \sum_{l=1}^L x_{s_l}
+ \sum_{l=2}^L w_{s_{l-1},s_l})\f$
where \f$Z\f$ is a normalization value so that the sum of \f$P(s)\f$ over
all possible sequences is \f$1\f$, and \f$x\f$ is the emission feature weight
to the linear chain CRF.
Finaly, the linear chain CRF operator outputs the logarithm of the conditional
likelihood of each training sample in a mini-batch.
NOTE:
1. The feature function for a CRF is made up of the emission features and the
transition features. The emission feature weights are NOT computed in
this operator. They MUST be computed first before this operator is called.
2. Because this operator performs global normalization over all possible
sequences internally, it expects UNSCALED emission feature weights.
Please do not call this op with the emission feature being output of any
nonlinear activation.
3. The 2nd dimension of Input(Emission) MUST be equal to the tag number.
)DOC");
}
};
class LinearChainCRFOp : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
void InferShape(framework::InferShapeContext* ctx) const override {
PADDLE_ENFORCE(ctx->HasInput("Emission"),
"Input(Emission) should be not null.");
PADDLE_ENFORCE(ctx->HasInput("Transition"),
"Input(Transition) should be not null.");
PADDLE_ENFORCE(ctx->HasInput("Label"), "Input(Label) should be not null.");
PADDLE_ENFORCE(ctx->HasOutput("Alpha"),
"Output(Alpha) should be not null.");
PADDLE_ENFORCE(ctx->HasOutput("EmissionExps"),
"Output(EmissionExps) should be not null.");
PADDLE_ENFORCE(ctx->HasOutput("TransitionExps"),
"Output(TransitionExps) should be not null.");
PADDLE_ENFORCE(ctx->HasOutput("LogLikelihood"),
"Output(LogLikelihood) should be not null.");
auto emission_dims = ctx->GetInputDim("Emission");
PADDLE_ENFORCE_EQ(emission_dims.size(), 2UL,
"The Input(Emission) should be a 2-D tensor.");
PADDLE_ENFORCE(emission_dims[0], "An empty mini-batch is not allowed.");
auto transition_dims = ctx->GetInputDim("Transition");
PADDLE_ENFORCE_EQ(transition_dims.size(), 2UL,
"The Input(Transition) should be a 2-D tensor.");
PADDLE_ENFORCE_EQ(
transition_dims[0] - 2, transition_dims[1],
"An invalid dimension for the Input(Transition), which should "
"be a 2-D tensor with shape [(D + 2) x D].");
PADDLE_ENFORCE_EQ(
emission_dims[1], transition_dims[1],
"The 2nd dimension of the Input(Emission) and the Input(Transition) "
"should be equal to the tag number.");
auto label_dims = ctx->GetInputDim("Label");
PADDLE_ENFORCE(label_dims.size() == 2UL && label_dims[1] == 1UL,
"The Input(Label) should be a 2-D tensor with the 2nd "
"dimensions fixed to 1.");
PADDLE_ENFORCE_EQ(
emission_dims[0], label_dims[0],
"The height of Input(Emission) and the height of Input(Label) "
"should be the same.");
ctx->SetOutputDim("Alpha", emission_dims);
ctx->SetOutputDim("EmissionExps", emission_dims);
ctx->SetOutputDim("TransitionExps", transition_dims);
// TODO(caoying) This is tricky. The 1st dimension of Output(LogLikelihood)
// is the sequence number in a mini-batch. The dimension set here should be
// resized to its correct size in the function Compute. Fix this once we can
// get LoD information in the InferShape interface.
ctx->SetOutputDim("LogLikelihood", {emission_dims[0], 1});
}
protected:
// Explicitly set that the data type of output of the linear_chain_crf
// operator is determined by its input "Emission".
framework::DataType IndicateDataType(
const framework::ExecutionContext& ctx) const override {
return framework::ToDataType(ctx.Input<LoDTensor>("Emission")->type());
}
};
class LinearChainCRFGradOp : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
void InferShape(framework::InferShapeContext* ctx) const override {
PADDLE_ENFORCE(ctx->HasInput("EmissionExps"),
"Input(EmissionExps) should be not null.");
PADDLE_ENFORCE(ctx->HasInput("TransitionExps"),
"Input(TransitionExps) should be not null.");
PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("LogLikelihood")),
"Input(LogLikelihood@GRAD) shoudl be not null.");
auto emission_exps_dims = ctx->GetInputDim("EmissionExps");
PADDLE_ENFORCE_EQ(emission_exps_dims.size(), 2UL,
"The Input(EmissionExps) should be a 2-D tensor.");
PADDLE_ENFORCE(emission_exps_dims[0],
"An empty mini-batch is not allowed.");
auto transition_exps_dims = ctx->GetInputDim("TransitionExps");
PADDLE_ENFORCE_EQ(transition_exps_dims.size(), 2UL,
"The Input(TransitionExps) should be a 2-D tensor.");
PADDLE_ENFORCE_EQ(
transition_exps_dims[0] - 2, transition_exps_dims[1],
"An invalid dimension for the Input(TransitionExps), which should "
"be a 2-D tensor with shape [(D + 2) x D].");
PADDLE_ENFORCE_EQ(
emission_exps_dims[1], transition_exps_dims[1],
"The 2nd dimension of the Input(EmissionExps) and the "
"Input(TransitionExps) should be equal to the tag number.");
auto label_dims = ctx->GetInputDim("Label");
PADDLE_ENFORCE(label_dims.size() == 2UL && label_dims[1] == 1UL,
"The Input(Label) should be a 2-D tensor with the 2nd "
"dimensions fixed to 1.");
PADDLE_ENFORCE_EQ(
emission_exps_dims[0], label_dims[0],
"The height of Input(EmissionExps) and the height of Input(Label) "
"should be the same.");
if (ctx->HasOutput(framework::GradVarName("Emission"))) {
ctx->SetOutputDim(framework::GradVarName("Emission"), emission_exps_dims);
}
if (ctx->HasOutput(framework::GradVarName("Transition"))) {
ctx->SetOutputDim(framework::GradVarName("Transition"),
transition_exps_dims);
}
}
protected:
// Explicitly set that the data type of output of the linear_chain_crf_grad
// operator is determined by its input: gradients of LogLikelihood.
framework::DataType IndicateDataType(
const framework::ExecutionContext& ctx) const override {
return framework::ToDataType(
ctx.Input<LoDTensor>(framework::GradVarName("LogLikelihood"))->type());
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP(linear_chain_crf, ops::LinearChainCRFOp, ops::LinearChainCRFOpMaker,
linear_chain_crf_grad, ops::LinearChainCRFGradOp);
REGISTER_OP_CPU_KERNEL(
linear_chain_crf,
ops::LinearChainCRFOpKernel<paddle::platform::CPUPlace, float>,
ops::LinearChainCRFOpKernel<paddle::platform::CPUPlace, double>);
REGISTER_OP_CPU_KERNEL(
linear_chain_crf_grad,
ops::LinearChainCRFGradOpKernel<paddle::platform::CPUPlace, float>,
ops::LinearChainCRFGradOpKernel<paddle::platform::CPUPlace, double>);
paddle/operators/linear_chain_crf_op.cu 0 → 100644
浏览文件 @ 1e8300fa
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/operators/linear_chain_crf_op.h"
namespace ops = paddle::operators;
REGISTER_OP_GPU_KERNEL(
linear_chain_crf,
ops::LinearChainCRFOpKernel<paddle::platform::GPUPlace, float>,
ops::LinearChainCRFOpKernel<paddle::platform::GPUPlace, double>);
REGISTER_OP_GPU_KERNEL(
linear_chain_crf_grad,
ops::LinearChainCRFGradOpKernel<paddle::platform::GPUPlace, float>,
ops::LinearChainCRFGradOpKernel<paddle::platform::GPUPlace, double>);
paddle/operators/linear_chain_crf_op.h 0 → 100644
浏览文件 @ 1e8300fa
此差异已折叠。
paddle/operators/lookup_table_op.cc
浏览文件 @ 1e8300fa
......@@ -43,7 +43,7 @@ class LookupTableOp : public framework::OperatorWithKernel {
protected:
framework::DataType IndicateDataType(
const framework::ExecutionContext& ctx) const override {
return framework::ToDataType(ctx.Input<Tensor>("W")->type());
return framework::ToDataType(ctx.Input<LoDTensor>("W")->type());
}
};
......@@ -93,7 +93,7 @@ class LookupTableOpGrad : public framework::OperatorWithKernel {
protected:
framework::DataType IndicateDataType(
const framework::ExecutionContext& ctx) const override {
return framework::ToDataType(ctx.Input<Tensor>("W")->type());
return framework::ToDataType(ctx.Input<LoDTensor>("W")->type());
}
};
......
paddle/operators/lookup_table_op.cu
浏览文件 @ 1e8300fa
......@@ -61,16 +61,16 @@ template <typename T>
class LookupTableCUDAKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& context) const override {
auto table_t = context.Input<Tensor>("W");
auto ids_t = context.Input<Tensor>("Ids");
auto output_t = context.Output<Tensor>("Out");
auto* table_t = context.Input<LoDTensor>("W");
auto* ids_t = context.Input<LoDTensor>("Ids");
auto* output_t = context.Output<LoDTensor>("Out");
size_t N = table_t->dims()[0];
size_t D = table_t->dims()[1];
size_t K = ids_t->numel();
auto ids = ids_t->data<int64_t>();
auto table = table_t->data<T>();
auto output = output_t->mutable_data<T>(context.GetPlace());
auto* ids = ids_t->data<int64_t>();
auto* table = table_t->data<T>();
auto* output = output_t->mutable_data<T>(context.GetPlace());
dim3 threads(128, 8);
dim3 grids(8, 1);
......@@ -87,9 +87,9 @@ class LookupTableGradCUDAKernel : public framework::OpKernel<T> {
void Compute(const framework::ExecutionContext& context) const override {
bool is_sparse = context.Attr<bool>("is_sparse");
if (is_sparse) {
auto* ids = context.Input<Tensor>("Ids");
auto* table = context.Input<Tensor>("W");
auto* d_output = context.Input<Tensor>(framework::GradVarName("Out"));
auto* ids = context.Input<LoDTensor>("Ids");
auto* table = context.Input<LoDTensor>("W");
auto* d_output = context.Input<LoDTensor>(framework::GradVarName("Out"));
auto* d_table = context.Output<SelectedRows>(framework::GradVarName("W"));
auto* ids_data = ids->data<int64_t>();
......@@ -116,12 +116,12 @@ class LookupTableGradCUDAKernel : public framework::OpKernel<T> {
auto* d_output_data = d_output->data<T>();
PADDLE_ENFORCE_EQ(d_table_value->dims(), d_output->dims());
memory::Copy(gpu_place, d_table_data, gpu_place, d_output_data,
d_output->numel(), stream);
d_output->numel() * sizeof(T), stream);
} else {
auto ids_t = context.Input<Tensor>("Ids");
auto d_output_t = context.Input<Tensor>(framework::GradVarName("Out"));
auto d_table_t = context.Output<Tensor>(framework::GradVarName("W"));
auto ids_t = context.Input<LoDTensor>("Ids");
auto d_output_t = context.Input<LoDTensor>(framework::GradVarName("Out"));
auto d_table_t = context.Output<LoDTensor>(framework::GradVarName("W"));
int N = d_table_t->dims()[0];
int D = d_table_t->dims()[1];
......
paddle/operators/lookup_table_op.h
浏览文件 @ 1e8300fa
......@@ -19,22 +19,22 @@
namespace paddle {
namespace operators {
using Tensor = framework::Tensor;
using LoDTensor = framework::LoDTensor;
using SelectedRows = framework::SelectedRows;
template <typename T>
class LookupTableKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& context) const override {
auto table_t = context.Input<Tensor>("W"); // float tensor
auto ids_t = context.Input<Tensor>("Ids"); // int tensor
auto output_t = context.Output<Tensor>("Out"); // float tensor
auto* table_t = context.Input<LoDTensor>("W"); // float tensor
auto* ids_t = context.Input<LoDTensor>("Ids"); // int tensor
auto* output_t = context.Output<LoDTensor>("Out"); // float tensor
int N = table_t->dims()[0];
int D = table_t->dims()[1];
auto ids = ids_t->data<int64_t>();
auto table = table_t->data<T>();
auto output = output_t->mutable_data<T>(context.GetPlace());
auto* ids = ids_t->data<int64_t>();
auto* table = table_t->data<T>();
auto* output = output_t->mutable_data<T>(context.GetPlace());
for (int64_t i = 0; i < ids_t->numel(); ++i) {
PADDLE_ENFORCE_LT(ids[i], N);
PADDLE_ENFORCE_GE(ids[i], 0);
......@@ -49,9 +49,9 @@ class LookupTableGradKernel : public framework::OpKernel<T> {
void Compute(const framework::ExecutionContext& context) const override {
bool is_sparse = context.Attr<bool>("is_sparse");
if (is_sparse) {
auto* ids = context.Input<Tensor>("Ids");
auto* table = context.Input<Tensor>("W");
auto* d_output = context.Input<Tensor>(framework::GradVarName("Out"));
auto* ids = context.Input<LoDTensor>("Ids");
auto* table = context.Input<LoDTensor>("W");
auto* d_output = context.Input<LoDTensor>(framework::GradVarName("Out"));
auto* d_table = context.Output<SelectedRows>(framework::GradVarName("W"));
auto* ids_data = ids->data<int64_t>();
......@@ -76,10 +76,10 @@ class LookupTableGradKernel : public framework::OpKernel<T> {
PADDLE_ENFORCE_EQ(d_table_value->dims(), d_output->dims());
memcpy(d_table_data, d_output_data, sizeof(T) * d_output->numel());
} else {
auto* ids = context.Input<Tensor>("Ids");
auto* d_output = context.Input<Tensor>(framework::GradVarName("Out"));
auto* d_table = context.Output<Tensor>(framework::GradVarName("W"));
auto* table = context.Input<Tensor>("W");
auto* ids = context.Input<LoDTensor>("Ids");
auto* d_output = context.Input<LoDTensor>(framework::GradVarName("Out"));
auto* d_table = context.Output<LoDTensor>(framework::GradVarName("W"));
auto* table = context.Input<LoDTensor>("W");
auto* ids_data = ids->data<int64_t>();
auto ids_dim = ids->dims();
......
paddle/operators/lstm_op.cc
浏览文件 @ 1e8300fa
......@@ -21,7 +21,6 @@ class LSTMOp : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
protected:
void InferShape(framework::InferShapeContext* ctx) const override {
PADDLE_ENFORCE(ctx->HasInput("Input"),
"Input(Input) of LSTM should not be null.");
......@@ -29,9 +28,13 @@ class LSTMOp : public framework::OperatorWithKernel {
"Output(Hidden) of LSTM should not be null.");
PADDLE_ENFORCE(ctx->HasOutput("Cell"),
"Output(Cell) of LSTM should not be null.");
PADDLE_ENFORCE(ctx->HasOutput("BatchGate"),
"Output(BatchGate) of LSTM should not be null.");
PADDLE_ENFORCE(ctx->HasOutput("BatchCellPreAct"),
"Output(BatchGate) of LSTM should not be null.");
auto x_dims = ctx->GetInputDim("Input");
PADDLE_ENFORCE_EQ(x_dims.size(), 2, "Input(X)'s rank must be 2.");
auto in_dims = ctx->GetInputDim("Input");
PADDLE_ENFORCE_EQ(in_dims.size(), 2, "Input(X)'s rank must be 2.");
if (ctx->HasInput("H0")) {
PADDLE_ENFORCE(ctx->HasInput("C0"),
......@@ -44,7 +47,7 @@ class LSTMOp : public framework::OperatorWithKernel {
"should be the same.");
}
int frame_size = x_dims[1] / 4;
int frame_size = in_dims[1] / 4;
auto w_dims = ctx->GetInputDim("Weight");
PADDLE_ENFORCE_EQ(w_dims.size(), 2,
"The rank of Input(Weight) should be 2.");
......@@ -71,12 +74,21 @@ class LSTMOp : public framework::OperatorWithKernel {
"4 * %d if disable peepholes connection",
frame_size);
}
ctx->SetOutputDim("Hidden", {x_dims[0], frame_size});
ctx->SetOutputDim("Cell", {x_dims[0], frame_size});
ctx->SetOutputDim("BatchGate", x_dims);
framework::DDim out_dims({in_dims[0], frame_size});
ctx->SetOutputDim("Hidden", out_dims);
ctx->SetOutputDim("Cell", out_dims);
ctx->SetOutputDim("BatchGate", in_dims);
ctx->SetOutputDim("BatchCellPreAct", out_dims);
ctx->ShareLoD("Input", "Hidden");
ctx->ShareLoD("Input", "Cell");
}
protected:
framework::DataType IndicateDataType(
const framework::ExecutionContext& ctx) const override {
return framework::ToDataType(
ctx.Input<framework::LoDTensor>("Input")->type());
}
};
class LSTMOpMaker : public framework::OpProtoAndCheckerMaker {
......@@ -86,16 +98,18 @@ class LSTMOpMaker : public framework::OpProtoAndCheckerMaker {
AddInput("Input",
"(LoDTensor) the first input is a LodTensor, which support "
"variable-time length input sequence. The underlying tensor in "
"this LoDTensor is a matrix with shape (T X 4D), where, T is the "
"this LoDTensor is a matrix with shape (T X 4D), where T is the "
"total time steps in this mini-batch, D is the hidden size.");
AddInput("H0",
"(Tensor, optional) the initial hidden state is an optional "
"input. This is a tensor with shape (N x D), where N is the "
"batch size, D is the hidden size.");
"batch size, D is the hidden size.")
.AsDispensable();
AddInput("C0",
"(Tensor, optional) the initial cell state is an optional "
"input. This is a tensor with shape (N x D), where N is the "
"batch size. `H0` and `C0` can be NULL but only at the same time");
"batch size. `H0` and `C0` can be NULL but only at the same time")
.AsDispensable();
AddInput("Weight",
"(Tensor) the learnable hidden-hidden weights."
" - The shape is (D x 4D), where D is the hidden size. "
......@@ -109,22 +123,27 @@ class LSTMOpMaker : public framework::OpProtoAndCheckerMaker {
" - Bias = {b_c, b_i, b_f, b_o}."
"2. `usePeepholes = True` "
" - The shape is (1 x 7D). "
" - Bias = {b_c, b_i, b_f, b_o, W_ic, W_fc, W_oc}.");
" - Bias = {b_c, b_i, b_f, b_o, W_ic, W_fc, W_oc}.")
.AsDispensable();
AddOutput("Hidden",
"(LoDTensor) the hidden state of LSTM operator. "
"The shape is (T x D), and lod is the same with the `Input`.");
AddOutput("Cell",
"(LoDTensor) the cell state of LSTM operator. "
"The shape is (T x D), and lod is the same with the `Input`.");
AddOutput("BatchGate",
"(LoDTensor) This LoDTensor contains input gate, forget gate "
"and output gate after the nonlinear computation. This "
"LoDTensor has the same shape with the reorganized input, which "
"was also be called batch input. The LoD size is 2. The first "
"is also be called batch input. The LoD size is 2. The first "
"LoD is the batch offsets and the second LoD contains the "
"indexes, which denote the position of reorganized sequence "
"in the raw input.")
.AsIntermediate();
AddOutput("Hidden",
"(LoDTensor) the hidden state lod tensor of LSTM operator. "
"The shape and lod is the same with the `Input`.");
AddOutput("Cell",
"(LoDTensor) the cell state lod tensor of LSTM operator. "
"The shape and lod is the same with the `Input`.");
AddOutput("BatchCellPreAct",
"(LoDTensor) This LoDTensor is got in the forward and used "
"in the backward.")
.AsIntermediate();
AddAttr<bool>("usePeepholes",
"(bool, defalut: True) "
"whether to enable diagonal/peephole connections.")
......@@ -202,15 +221,37 @@ class LSTMGradOp : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
protected:
void InferShape(framework::InferShapeContext* ctx) const override {
PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Hidden")),
"Input(Hidden@GRAD) should not be null");
PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Cell")),
"Input(Cell@GRAD) should not be null");
ctx->SetOutputDim(framework::GradVarName("Weight"),
ctx->GetInputDim("Weight"));
ctx->SetOutputDim(framework::GradVarName("Bias"), ctx->GetInputDim("Bias"));
PADDLE_ENFORCE(ctx->HasInput("Input"),
"Input(Input) of LSTM should not be null.");
PADDLE_ENFORCE(ctx->HasInput("Hidden"),
"Input(Hidden) of LSTM should not be null.");
PADDLE_ENFORCE(ctx->HasInput("Cell"),
"Input(Cell) of LSTM should not be null.");
PADDLE_ENFORCE(ctx->HasInput("BatchGate"),
"Input(BatchGate) of LSTM should not be null.");
PADDLE_ENFORCE(ctx->HasInput("BatchCellPreAct"),
"Input(BatchGate) of LSTM should not be null.");
auto in_g_name = framework::GradVarName("Input");
if (ctx->HasOutput(in_g_name))
ctx->SetOutputDim(in_g_name, ctx->GetInputDim("Input"));
auto w_g_name = framework::GradVarName("Weight");
if (ctx->HasOutput(w_g_name))
ctx->SetOutputDim(w_g_name, ctx->GetInputDim("Weight"));
auto b_g_name = framework::GradVarName("Bias");
if (ctx->HasOutput(b_g_name))
ctx->SetOutputDim(b_g_name, ctx->GetInputDim("Bias"));
}
protected:
framework::DataType IndicateDataType(
const framework::ExecutionContext& ctx) const override {
return framework::ToDataType(
ctx.Input<framework::LoDTensor>("Input")->type());
}
};
......
paddle/operators/lstm_op.h
浏览文件 @ 1e8300fa
......@@ -21,8 +21,9 @@ limitations under the License. */
namespace paddle {
namespace operators {
using framework::LoDTensor;
using framework::Tensor;
using LoDTensor = framework::LoDTensor;
using Tensor = framework::Tensor;
template <typename T, int MajorType = Eigen::RowMajor,
typename IndexType = Eigen::DenseIndex>
using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
......@@ -31,15 +32,15 @@ template <typename Place, typename T>
class LSTMKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
auto* input = ctx.Input<framework::LoDTensor>("Input");
auto* weight = ctx.Input<framework::Tensor>("Weight");
auto* bias = ctx.Input<framework::Tensor>("Bias");
auto* input = ctx.Input<LoDTensor>("Input");
auto* weight = ctx.Input<Tensor>("Weight");
auto* bias = ctx.Input<Tensor>("Bias");
auto* batch_gate = ctx.Output<framework::LoDTensor>("BatchGate");
auto* batch_gate = ctx.Output<LoDTensor>("BatchGate");
batch_gate->mutable_data<T>(ctx.GetPlace());
auto* hidden_out = ctx.Output<framework::LoDTensor>("Hidden");
auto* hidden_out = ctx.Output<LoDTensor>("Hidden");
hidden_out->mutable_data<T>(ctx.GetPlace());
auto* cell_out = ctx.Output<framework::LoDTensor>("Cell");
auto* cell_out = ctx.Output<LoDTensor>("Cell");
cell_out->mutable_data<T>(ctx.GetPlace());
// Now the function ShareLoD in InferShape is not implemented.
......@@ -49,7 +50,8 @@ class LSTMKernel : public framework::OpKernel<T> {
bool is_reverse = ctx.Attr<bool>("isReverse");
math::LoDTensor2BatchFunctor<Place, T> to_batch;
to_batch(ctx.device_context(), *input, *batch_gate, is_reverse);
auto& device_ctx = ctx.device_context();
to_batch(device_ctx, *input, *batch_gate, true, is_reverse);
auto in_dims = input->dims();
int frame_size = static_cast<int>(in_dims[1] / 4);
......@@ -69,17 +71,26 @@ class LSTMKernel : public framework::OpKernel<T> {
}
math::LstmMetaValue<T> lstm_value;
T* bias_data = const_cast<T*>(bias->data<T>());
// the code style in LstmMetaValue will be updated later.
lstm_value.checkIg = bias_data + 4 * frame_size;
lstm_value.checkFg = lstm_value.checkIg + frame_size;
lstm_value.checkOg = lstm_value.checkFg + frame_size;
if (bias) {
T* bias_data = const_cast<T*>(bias->data<T>());
// the code style in LstmMetaValue will be updated later.
lstm_value.checkIg = bias_data + 4 * frame_size;
lstm_value.checkFg = lstm_value.checkIg + frame_size;
lstm_value.checkOg = lstm_value.checkFg + frame_size;
} else {
lstm_value.checkIg = nullptr;
lstm_value.checkFg = nullptr;
lstm_value.checkOg = nullptr;
}
lstm_value.prevStateValue = nullptr;
framework::LoDTensor batch_out, batch_cell, batch_cell_pre_act;
batch_out.mutable_data<T>(dims, ctx.GetPlace());
// Use the local variable as here.
LoDTensor batch_hidden, batch_cell;
auto* batch_cell_pre_act = ctx.Output<LoDTensor>("BatchCellPreAct");
batch_hidden.mutable_data<T>(dims, ctx.GetPlace());
batch_cell.mutable_data<T>(dims, ctx.GetPlace());
batch_cell_pre_act.mutable_data<T>(dims, ctx.GetPlace());
batch_cell_pre_act->mutable_data<T>(dims, ctx.GetPlace());
auto batch_starts = batch_gate->lod()[0];
size_t num_batch = batch_starts.size() - 1;
......@@ -92,18 +103,18 @@ class LSTMKernel : public framework::OpKernel<T> {
int bend = static_cast<int>(batch_starts[n + 1]);
Tensor gate_t = batch_gate->Slice(bstart, bend);
Tensor out_t = batch_out.Slice(bstart, bend);
Tensor out_t = batch_hidden.Slice(bstart, bend);
Tensor cell_t = batch_cell.Slice(bstart, bend);
Tensor cell_pre_act_t = batch_cell_pre_act.Slice(bstart, bend);
Tensor cell_pre_act_t = batch_cell_pre_act->Slice(bstart, bend);
int cur_batch_size = bend - bstart;
if (n != 0) {
int pre_h_start = static_cast<int>(batch_starts[n - 1]);
int pre_h_end = pre_h_start + cur_batch_size;
auto pre_hidden_t = batch_out.Slice(pre_h_start, pre_h_end);
math::matmul<Place, T>(ctx.device_context(), pre_hidden_t, false,
*weight, false, static_cast<T>(1.0), &gate_t,
auto pre_hidden_t = batch_hidden.Slice(pre_h_start, pre_h_end);
math::matmul<Place, T>(device_ctx, pre_hidden_t, false, *weight, false,
static_cast<T>(1.0), &gate_t,
static_cast<T>(1.0));
}
// else if : FIXME support the initial hidden and cell
......@@ -112,27 +123,186 @@ class LSTMKernel : public framework::OpKernel<T> {
lstm_value.outputValue = out_t.data<T>();
lstm_value.stateValue = cell_t.data<T>();
lstm_value.stateActiveValue = cell_pre_act_t.data<T>();
math::LstmUnitFunctor<Place, T>::compute(ctx.device_context(), lstm_value,
math::LstmUnitFunctor<Place, T>::compute(device_ctx, lstm_value,
frame_size, cur_batch_size,
gate_act, cell_act, cand_act);
lstm_value.prevStateValue = lstm_value.stateValue;
}
math::Batch2LoDTensorFunctor<Place, T> to_seq;
batch_out.set_lod(batch_gate->lod());
batch_hidden.set_lod(batch_gate->lod());
// restore the output hidden in LoDTensor from the batch hidden
to_seq(ctx.device_context(), batch_out, *hidden_out);
to_seq(device_ctx, batch_hidden, *hidden_out);
batch_cell.set_lod(batch_gate->lod());
// restore the output cell state in LoDTensor from the batch cell
to_seq(ctx.device_context(), batch_cell, *cell_out);
to_seq(device_ctx, batch_cell, *cell_out);
}
};
template <typename Place, typename T>
class LSTMGradKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {}
void Compute(const framework::ExecutionContext& ctx) const override {
auto* input = ctx.Input<LoDTensor>("Input");
auto* weight = ctx.Input<Tensor>("Weight");
auto* bias = ctx.Input<Tensor>("Bias");
auto* hidden_out = ctx.Input<LoDTensor>("Hidden");
auto* cell_out = ctx.Input<LoDTensor>("Cell");
auto* batch_gate = ctx.Input<LoDTensor>("BatchGate");
auto* batch_cell_pre_act = ctx.Input<LoDTensor>("BatchCellPreAct");
auto* hidden_g = ctx.Input<LoDTensor>(framework::GradVarName("Hidden"));
auto* in_g = ctx.Output<LoDTensor>(framework::GradVarName("Input"));
auto* weight_g = ctx.Output<Tensor>(framework::GradVarName("Weight"));
auto* bias_g = ctx.Output<Tensor>(framework::GradVarName("Bias"));
auto& device_ctx = ctx.device_context();
math::SetConstant<Place, T> zero;
if (weight_g) {
weight_g->mutable_data<T>(ctx.GetPlace());
zero(device_ctx, weight_g, static_cast<T>(0.0));
}
auto in_dims = input->dims();
auto out_dims = hidden_g->dims();
int frame_size = static_cast<int>(in_dims[1] / 4);
PADDLE_ENFORCE_EQ(frame_size, out_dims[1]);
math::LstmMetaValue<T> lstm_value;
if (bias) {
T* bias_data = const_cast<T*>(bias->data<T>());
lstm_value.checkIg = bias_data + 4 * frame_size;
lstm_value.checkFg = lstm_value.checkIg + frame_size;
lstm_value.checkOg = lstm_value.checkFg + frame_size;
} else {
lstm_value.checkIg = nullptr;
lstm_value.checkFg = nullptr;
lstm_value.checkOg = nullptr;
}
math::LstmMetaGrad<T> lstm_grad;
if (bias && bias_g) {
T* bias_g_data = const_cast<T*>(bias_g->mutable_data<T>(ctx.GetPlace()));
zero(device_ctx, bias_g, static_cast<T>(0.0));
lstm_grad.checkIgGrad = bias_g_data + 4 * frame_size;
lstm_grad.checkFgGrad = lstm_grad.checkIgGrad + frame_size;
lstm_grad.checkOgGrad = lstm_grad.checkFgGrad + frame_size;
} else {
lstm_grad.checkIgGrad = nullptr;
lstm_grad.checkFgGrad = nullptr;
lstm_grad.checkOgGrad = nullptr;
}
math::LoDTensor2BatchFunctor<Place, T> to_batch;
// use the local variable as here.
LoDTensor batch_hidden;
batch_hidden.mutable_data<T>(out_dims, ctx.GetPlace());
batch_hidden.set_lod(batch_gate->lod());
to_batch(device_ctx, *hidden_out, batch_hidden, false);
LoDTensor batch_hidden_g;
batch_hidden_g.mutable_data<T>(out_dims, ctx.GetPlace());
batch_hidden_g.set_lod(batch_gate->lod());
to_batch(device_ctx, *hidden_g, batch_hidden_g, false);
LoDTensor batch_cell;
batch_cell.mutable_data<T>(out_dims, ctx.GetPlace());
batch_cell.set_lod(batch_gate->lod());
to_batch(device_ctx, *cell_out, batch_cell, false);
LoDTensor batch_cell_g;
batch_cell_g.mutable_data<T>(out_dims, ctx.GetPlace());
batch_cell_g.set_lod(batch_gate->lod());
// TODO(qingqing) support the case output cell has gradient.
// to_batch(device_ctx, *cell_g, batch_cell_g, false);
zero(device_ctx, &batch_cell_g, static_cast<T>(0.0));
LoDTensor batch_gate_g;
batch_gate_g.mutable_data<T>(batch_gate->dims(), ctx.GetPlace());
batch_gate_g.set_lod(batch_gate->lod());
auto gate_act = ctx.Attr<std::string>("gateActivation");
auto cell_act = ctx.Attr<std::string>("cellActivation");
auto cand_act = ctx.Attr<std::string>("candidateActivation");
auto batch_starts = batch_gate->lod()[0];
size_t num_batch = batch_starts.size() - 1;
for (int n = static_cast<int>(num_batch) - 1; n >= 0; n--) {
int bstart = static_cast<int>(batch_starts[n]);
int bend = static_cast<int>(batch_starts[n + 1]);
Tensor gate = batch_gate->Slice(bstart, bend);
Tensor cell = batch_cell.Slice(bstart, bend);
Tensor cell_pre_act = batch_cell_pre_act->Slice(bstart, bend);
lstm_value.gateValue = gate.data<T>();
lstm_value.stateValue = cell.data<T>();
lstm_value.stateActiveValue = cell_pre_act.data<T>();
Tensor out_g = batch_hidden_g.Slice(bstart, bend);
Tensor gate_g = batch_gate_g.Slice(bstart, bend);
Tensor cell_g = batch_cell_g.Slice(bstart, bend);
lstm_grad.stateGrad = cell_g.data<T>();
lstm_grad.gateGrad = gate_g.data<T>();
lstm_grad.outputGrad = out_g.data<T>();
if (n) {
int bstart_pre = static_cast<int>(batch_starts[n - 1]);
Tensor cell_pre = batch_cell.Slice(bstart_pre, bstart);
Tensor cell_pre_g = batch_cell_g.Slice(bstart_pre, bstart);
lstm_value.prevStateValue = cell_pre.data<T>();
lstm_grad.prevStateGrad = cell_pre_g.data<T>();
} else {
lstm_value.prevStateValue = nullptr;
lstm_grad.prevStateGrad = nullptr;
}
int cur_batch_size = bend - bstart;
math::LstmUnitGradFunctor<Place, T>::compute(
device_ctx, lstm_value, lstm_grad, frame_size, cur_batch_size,
gate_act, cell_act, cand_act);
if (n != 0) {
int pre_h_start = static_cast<int>(batch_starts[n - 1]);
int pre_h_end = pre_h_start + cur_batch_size;
auto pre_hidden_g = batch_hidden_g.Slice(pre_h_start, pre_h_end);
math::matmul<Place, T>(device_ctx, gate_g, false, *weight, true,
static_cast<T>(1.0), &pre_hidden_g,
static_cast<T>(1.0));
if (weight_g) {
/* backward weight */
auto pre_hidden = batch_hidden.Slice(pre_h_start, pre_h_end);
math::matmul<Place, T>(device_ctx, pre_hidden, true, gate_g, false,
static_cast<T>(1.0), weight_g,
static_cast<T>(1.0));
}
}
}
math::Batch2LoDTensorFunctor<Place, T> to_seq;
if (in_g) {
/* backward data */
in_g->mutable_data<T>(ctx.GetPlace());
to_seq(device_ctx, batch_gate_g, *in_g);
}
if (bias && bias_g) {
/* backward bias */
int m = static_cast<int>(batch_gate_g.dims()[0]);
int n = static_cast<int>(batch_gate_g.dims()[1]);
Tensor ones;
ones.mutable_data<T>({m}, ctx.GetPlace());
math::SetConstant<Place, T> set;
set(device_ctx, &ones, static_cast<T>(1.0));
math::gemv<Place, T>(device_ctx, true, m, n, 1., batch_gate_g.data<T>(),
ones.data<T>(), 0., bias_g->data<T>());
}
}
};
} // namespace operators
......
paddle/operators/math/context_project.h
浏览文件 @ 1e8300fa
......@@ -16,36 +16,36 @@ limitations under the License. */
#include "paddle/framework/eigen.h"
#include "paddle/framework/lod_tensor.h"
#include "paddle/framework/tensor.h"
#include "paddle/operators/math/im2col.h"
namespace paddle {
namespace operators {
namespace math {
using Tensor = framework::Tensor;
using LoDTensor = framework::LoDTensor;
template <typename T, int MajorType = Eigen::RowMajor,
typename IndexType = Eigen::DenseIndex>
using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
/*
* \brief Context projection concatenate features in adjacent time steps in
* \brief Context projection concatenates features in adjacent time-steps in
* a sequence. The i-th row of the output is the concatenation of
* context_length rows of the input. The context_length rows are the
* consecutive rows from the i+shift_start row.
* ContextProjectGradFunctor is the inverse process of ContextProjectFunctor.
*
* \param in Input data.
* \param Shape The shape of Input data,
* [minibatch, number_of_input_features].
* \param type A float LoDTensor.
* \param Shape The shape of Input data:
* [mini-batch, input_hidden_size].
*
* \param padding_data Padding data.
* \param Shape The shape of Padding data,
* [up_pad + down_pad, number_of_input_features].
* \param type A float Tensor.
* \param Shape The shape of Padding data:
* [up_pad + down_pad, input_hidden_size].
*
* \param col Col data.
* \param Shape The shape of Col data,
* [minibatch, context_length * number_of_input_features].
* \param type A float Tensor.
* \param Shape The shape of Col data:
* [mini-batch, context_length * input_hidden_size].
*
* For a mini-batch of 2 variable lengths sentences, containing 3, and 1
* time-steps:
......@@ -63,72 +63,170 @@ using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
* representation is 2.
*
* - Case1:
* If context_start is -1 and padding_trainable is false, we use zero to pad
* instead of learned weight to pad,
* and the context_lenth is 3, the output (Out) is:
* If context_start is -1 and padding_trainable is false, we use zero to pad
* instead of learned weight to pad,
* and the context_length is 3, the output (Out) is:
*
* Out =[[0, 0, a1, a2, b1, b2;
* a1, a2, b1, b2, c1, c2;
* b1, b2, c1, c2, 0, 0 ]
* [0, 0, d1, d2, 0, 0 ]]
* Out =[[0, 0, a1, a2, b1, b2;
* a1, a2, b1, b2, c1, c2;
* b1, b2, c1, c2, 0, 0 ]
* [0, 0, d1, d2, 0, 0 ]]
*
* - Case2:
* If context_start is -1 and padding_trainable is true, we use learned weight
* to pad,
* and the context_lenth is 3, the output (Out) is:
* If context_start is -1 and padding_trainable is true, we use learned weight
* to pad,
* and the context_length is 3, the output (Out) is:
*
* Out = [[w1, w2, a1, a2, b1, b2;
* a1, a2, b1, b2, c1, c2;
* b1, b2, c1, c2, w3, w4]
* [w1, w2, d1, d2, w3, w4]]
* Out = [[w1, w2, a1, a2, b1, b2;
* a1, a2, b1, b2, c1, c2;
* b1, b2, c1, c2, w3, w4]
* [w1, w2, d1, d2, w3, w4]]
*
*/
template <typename Place, typename T>
class ContextProjectFunctor {
public:
void operator()(const platform::DeviceContext& context,
framework::LoDTensor& in, framework::Tensor& padding_data,
framework::Tensor& col, bool padding_trainable,
void operator()(const platform::DeviceContext& context, const LoDTensor& in,
const Tensor& padding_data, Tensor& col,
bool padding_trainable, int context_start, int context_length,
int context_stride, int up_pad, int down_pad) {
auto lod_level_0 = in.lod()[0];
math::Im2ColFunctor<math::ColFormat::kOCF, Place, float> im2col_ocf;
int input_row_begin, input_row_end;
int sequence_height, sequence_width;
sequence_width = in.dims()[1];
for (int i = 0; i < static_cast<int>(lod_level_0.size()) - 1; ++i) {
input_row_begin = (context_start > 0)
? static_cast<int>(lod_level_0[i]) + context_start
: static_cast<int>(lod_level_0[i]);
input_row_end = static_cast<int>(lod_level_0[i + 1]);
Tensor out_t = col.Slice(static_cast<int>(lod_level_0[i]),
static_cast<int>(lod_level_0[i + 1]));
sequence_height = static_cast<int>(out_t.dims()[0]);
if (input_row_begin < input_row_end) {
Tensor in_t = in.Slice(input_row_begin, input_row_end);
std::vector<int64_t> output_shape(
{sequence_height, 1, 1, context_length,
sequence_width}); // output_height, output_width,
// input_channels, filter_height, filter_width
out_t.Resize(framework::make_ddim(output_shape));
std::vector<int64_t> input_shape(
{1, input_row_end - input_row_begin,
sequence_width}); // input_channels, input_height, input_width
in_t.Resize(framework::make_ddim(input_shape));
im2col_ocf(context, in_t, out_t,
/*stride_height*/ context_stride, /*stride_width*/ 1, up_pad,
down_pad, 0, 0);
out_t.Resize({sequence_height, context_length * sequence_width});
}
}
if (padding_trainable) {
for (int i = 0; i < static_cast<int>(lod_level_0.size()) - 1; ++i) {
Tensor out_t = col.Slice(static_cast<int>(lod_level_0[i]),
static_cast<int>(lod_level_0[i + 1]));
sequence_height = static_cast<int>(out_t.dims()[0]);
// add up trainable data
out_t.Resize({sequence_height * context_length, sequence_width});
if (up_pad > 0) { // add up pad
int padding_rows = std::min(
up_pad, static_cast<int>(lod_level_0[i + 1] - lod_level_0[i]));
for (int k = 0; k < padding_rows; ++k) {
int padding_size =
k + context_length < up_pad ? context_length : up_pad - k;
Tensor out_t_sub = out_t.Slice(k * context_length,
k * context_length + padding_size);
Tensor w_sub = padding_data.Slice(k, k + padding_size);
auto out_t_sub_e = EigenMatrix<T>::From(out_t_sub);
auto w_sub_e = EigenMatrix<T>::From(w_sub);
out_t_sub_e.device(*context.GetEigenDevice<Place>()) = w_sub_e;
}
}
if (down_pad > 0) { // add down pad
int down_pad_begin_row =
std::max(0,
(sequence_height - context_start - context_length) + 1) +
1;
int padding_begin = std::max(0, context_start - sequence_height);
int padding_size =
sequence_height - context_start >= context_length
? 1
: context_length - (sequence_height - context_start);
if (context_start >= sequence_height) padding_size = context_length;
int padding_idx = padding_begin;
for (int t = 0; t + down_pad_begin_row <= sequence_height;
++t, ++padding_size) {
if (context_start >= sequence_height) padding_size = context_length;
if (padding_size > context_length) {
padding_size = context_length;
padding_idx++;
}
if (padding_begin > 0 || sequence_height == context_start)
padding_idx = padding_begin + t;
Tensor out_t_sub = out_t.Slice(
(down_pad_begin_row + t) * context_length - padding_size,
(down_pad_begin_row + t) * context_length);
Tensor w_sub = padding_data.Slice(
up_pad + padding_idx, up_pad + padding_idx + padding_size);
auto out_t_sub_e = EigenMatrix<T>::From(out_t_sub);
auto w_sub_e = EigenMatrix<T>::From(w_sub);
out_t_sub_e.device(*context.GetEigenDevice<Place>()) = w_sub_e;
}
}
out_t.Resize({sequence_height, context_length * sequence_width});
}
}
}
};
template <typename Place, typename T>
class ContextProjectGradFunctor {
public:
void operator()(const platform::DeviceContext& context, LoDTensor& in,
Tensor& padding_data, Tensor& col, bool padding_trainable,
int context_start, int context_length, int context_stride,
int up_pad, int down_pad, bool gradient, bool input_grad,
bool pad_grad) {
int up_pad, int down_pad, bool input_grad, bool pad_grad) {
auto lod_level_0 = in.lod()[0];
paddle::operators::math::Im2ColFunctor<
paddle::operators::math::ColFormat::kOCF, Place, float>
im2col_ocf;
paddle::operators::math::Col2ImFunctor<
paddle::operators::math::ColFormat::kOCF, Place, float>
col2im_ocf;
math::Col2ImFunctor<math::ColFormat::kOCF, Place, float> col2im_ocf;
int input_row_begin, input_row_end;
int sequence_height, sequence_width;
sequence_width = in.dims()[1];
input_grad = gradient && input_grad;
pad_grad = gradient && pad_grad;
if (!gradient || input_grad) {
if (input_grad) {
for (int i = 0; i < static_cast<int>(lod_level_0.size()) - 1; ++i) {
input_row_begin = (context_start > 0)
? static_cast<int>(lod_level_0[i]) + context_start
: static_cast<int>(lod_level_0[i]);
input_row_end = static_cast<int>(lod_level_0[i + 1]);
framework::Tensor out_t =
col.Slice(static_cast<int>(lod_level_0[i]),
static_cast<int>(lod_level_0[i + 1]));
Tensor out_t = col.Slice(static_cast<int>(lod_level_0[i]),
static_cast<int>(lod_level_0[i + 1]));
sequence_height = static_cast<int>(out_t.dims()[0]);
if (input_row_begin < input_row_end) {
framework::Tensor in_t = in.Slice(input_row_begin, input_row_end);
Tensor in_t = in.Slice(input_row_begin, input_row_end);
std::vector<int64_t> output_shape(
{sequence_height, 1, 1, context_length,
sequence_width}); // output_height, output_width,
// input_channels, filter_height, filter_width
out_t.Resize(framework::make_ddim(output_shape));
std::vector<int64_t> input_shape(
......@@ -136,53 +234,39 @@ class ContextProjectFunctor {
sequence_width}); // input_channels, input_height, input_width
in_t.Resize(framework::make_ddim(input_shape));
if (gradient) {
col2im_ocf(context, in_t, out_t,
/*stride_height*/ context_stride, /*stride_width*/ 1,
up_pad, down_pad, 0, 0);
} else {
im2col_ocf(context, in_t, out_t,
/*stride_height*/ context_stride, /*stride_width*/ 1,
up_pad, down_pad, 0, 0);
}
col2im_ocf(context, in_t, out_t,
/*stride_height*/ context_stride, /*stride_width*/ 1,
up_pad, down_pad, 0, 0);
out_t.Resize({sequence_height, context_length * sequence_width});
}
}
}
if (!gradient || pad_grad) {
if (pad_grad) {
if (padding_trainable) {
for (int i = 0; i < static_cast<int>(lod_level_0.size()) - 1; ++i) {
framework::Tensor out_t =
col.Slice(static_cast<int>(lod_level_0[i]),
static_cast<int>(lod_level_0[i + 1]));
Tensor out_t = col.Slice(static_cast<int>(lod_level_0[i]),
static_cast<int>(lod_level_0[i + 1]));
sequence_height = static_cast<int>(out_t.dims()[0]);
// add up trainable data
out_t.Resize({sequence_height * context_length, sequence_width});
if (up_pad > 0) { // add up pad
if (up_pad > 0) {
int padding_rows = std::min(
up_pad, static_cast<int>(lod_level_0[i + 1] - lod_level_0[i]));
for (int k = 0; k < padding_rows; ++k) {
int padding_size =
k + context_length < up_pad ? context_length : up_pad - k;
framework::Tensor out_t_sub = out_t.Slice(
k * context_length, k * context_length + padding_size);
framework::Tensor w_sub = padding_data.Slice(k, k + padding_size);
// in this block, using EigenVector<T>::Flatten is ok too.
Tensor out_t_sub = out_t.Slice(k * context_length,
k * context_length + padding_size);
Tensor w_sub = padding_data.Slice(k, k + padding_size);
auto out_t_sub_e = EigenMatrix<T>::From(out_t_sub);
auto w_sub_e = EigenMatrix<T>::From(w_sub);
if (gradient) {
w_sub_e.device(*context.GetEigenDevice<Place>()) =
w_sub_e + out_t_sub_e;
} else {
out_t_sub_e.device(*context.GetEigenDevice<Place>()) = w_sub_e;
}
w_sub_e.device(*context.GetEigenDevice<Place>()) =
w_sub_e + out_t_sub_e;
}
}
if (down_pad > 0) { // add down pad
if (down_pad > 0) {
int down_pad_begin_row =
std::max(
0, (sequence_height - context_start - context_length) + 1) +
......@@ -204,19 +288,16 @@ class ContextProjectFunctor {
}
if (padding_begin > 0 || sequence_height == context_start)
padding_idx = padding_begin + t;
framework::Tensor out_t_sub = out_t.Slice(
Tensor out_t_sub = out_t.Slice(
(down_pad_begin_row + t) * context_length - padding_size,
(down_pad_begin_row + t) * context_length);
framework::Tensor w_sub = padding_data.Slice(
Tensor w_sub = padding_data.Slice(
up_pad + padding_idx, up_pad + padding_idx + padding_size);
auto out_t_sub_e = EigenMatrix<T>::From(out_t_sub);
auto w_sub_e = EigenMatrix<T>::From(w_sub);
if (gradient) {
w_sub_e.device(*context.GetEigenDevice<Place>()) =
w_sub_e + out_t_sub_e;
} else {
out_t_sub_e.device(*context.GetEigenDevice<Place>()) = w_sub_e;
}
w_sub_e.device(*context.GetEigenDevice<Place>()) =
w_sub_e + out_t_sub_e;
}
}
out_t.Resize({sequence_height, context_length * sequence_width});
......
paddle/operators/math/detail/lstm_cpu_kernel.h
浏览文件 @ 1e8300fa
......@@ -26,10 +26,7 @@ namespace detail {
template <class T, class Op>
void naive_lstm_forward_one_sequence(Op op, LstmMetaValue<T> value,
int frameSize,
activation_mode_t active_node,
activation_mode_t active_gate,
activation_mode_t active_state) {
int frameSize) {
T rValueIn;
T rValueIg;
T rValueFg;
......@@ -60,10 +57,8 @@ void naive_lstm_forward_one_sequence(Op op, LstmMetaValue<T> value,
rPrevState = value.prevStateValue[i];
}
hppl::cpu::ForwardAct<T> act;
op(rValueIn, rValueIg, rValueFg, rValueOg, rPrevState, rState, rStateAtv,
rOut, rCheckI, rCheckF, rCheckO, act(active_node), act(active_gate),
act(active_state));
rOut, rCheckI, rCheckF, rCheckO);
valueIn[i] = rValueIn;
valueIg[i] = rValueIg;
......@@ -77,10 +72,7 @@ void naive_lstm_forward_one_sequence(Op op, LstmMetaValue<T> value,
template <class T, class Op>
void naive_lstm_backward_one_sequence(Op op, LstmMetaValue<T> value,
LstmMetaGrad<T> grad, int frameSize,
activation_mode_t active_node,
activation_mode_t active_gate,
activation_mode_t active_state) {
LstmMetaGrad<T> grad, int frameSize) {
T rValueIn;
T rValueIg;
T rValueFg;
......@@ -127,11 +119,10 @@ void naive_lstm_backward_one_sequence(Op op, LstmMetaValue<T> value,
rPrevState = value.prevStateValue[i];
}
hppl::cpu::BackwardAct<T> act;
op(rValueIn, rValueIg, rValueFg, rValueOg, rGradIn, rGradIg, rGradFg,
rGradOg, rPrevState, rPrevStateGrad, rState, rStateGrad, rStateAtv,
rOutputGrad, rCheckI, rCheckF, rCheckO, rCheckIGrad, rCheckFGrad,
rCheckOGrad, act(active_node), act(active_gate), act(active_state));
rCheckOGrad);
gradIn[i] = rGradIn;
gradIg[i] = rGradIg;
......@@ -283,8 +274,7 @@ void cpu_lstm_forward(Op op, LstmMetaValue<T> value, int frameSize,
avx_lstm_forward_one_sequence<T>(op, value, frameSize, active_node,
active_gate, active_state);
} else {
naive_lstm_forward_one_sequence<T>(op, value, frameSize, active_node,
active_gate, active_state);
naive_lstm_forward_one_sequence<T>(op, value, frameSize);
}
}
......@@ -297,8 +287,7 @@ void cpu_lstm_backward(Op op, LstmMetaValue<T> value, LstmMetaGrad<T> grad,
avx_lstm_backward_one_sequence<T>(op, value, grad, frameSize, active_node,
active_gate, active_state);
} else {
naive_lstm_backward_one_sequence<T>(op, value, grad, frameSize, active_node,
active_gate, active_state);
naive_lstm_backward_one_sequence<T>(op, value, grad, frameSize);
}
}
......
paddle/operators/math/detail/lstm_gpu_kernel.h
浏览文件 @ 1e8300fa
......@@ -32,9 +32,7 @@ namespace detail {
*/
template <class T, class Op, bool isBatch>
__global__ void KeLstmForward(Op op, LstmMetaValue<T> value, int frameSize,
int batchSize, activation_mode_t active_node,
activation_mode_t active_gate,
activation_mode_t active_state) {
int batchSize) {
const int frameIdx = blockIdx.x * blockDim.x + threadIdx.x;
if (frameIdx >= frameSize) return;
......@@ -70,10 +68,8 @@ __global__ void KeLstmForward(Op op, LstmMetaValue<T> value, int frameSize,
rPrevState = value.prevStateValue[frameIdx];
}
hppl::gpu::ForwardAct<T> act;
op(rValueIn, rValueIg, rValueFg, rValueOg, rPrevState, rState, rStateAtv,
rOut, rCheckI, rCheckF, rCheckO, act(active_node), act(active_gate),
act(active_state));
rOut, rCheckI, rCheckF, rCheckO);
value.gateValue[frameIdx] = rValueIn;
value.gateValue[frameIdx + frameSize] = rValueIg;
......@@ -92,9 +88,7 @@ __global__ void KeLstmForward(Op op, LstmMetaValue<T> value, int frameSize,
template <class T, class Op, bool isBatch>
__global__ void KeLstmBackward(Op op, LstmMetaValue<T> value,
LstmMetaGrad<T> grad, int frameSize,
int batchSize, activation_mode_t active_node,
activation_mode_t active_gate,
activation_mode_t active_state) {
int batchSize) {
const int frameIdx = blockIdx.x * blockDim.x + threadIdx.x;
if (frameIdx >= frameSize) return;
......@@ -145,11 +139,9 @@ __global__ void KeLstmBackward(Op op, LstmMetaValue<T> value,
rPrevState = value.prevStateValue[frameIdx];
}
hppl::gpu::BackwardAct<T> act;
op(rValueIn, rValueIg, rValueFg, rValueOg, rGradIn, rGradIg, rGradFg, rGradOg,
rPrevState, rPrevStateGrad, rState, rStateGrad, rStateAtv, rOutputGrad,
rCheckI, rCheckF, rCheckO, rCheckIGrad, rCheckFGrad, rCheckOGrad,
act(active_node), act(active_gate), act(active_state));
rCheckI, rCheckF, rCheckO, rCheckIGrad, rCheckFGrad, rCheckOGrad);
grad.gateGrad[frameIdx] = rGradIn;
grad.gateGrad[frameIdx + frameSize] = rGradIg;
......@@ -205,13 +197,11 @@ void gpu_lstm_forward(const platform::DeviceContext& context, Op op,
if (batchSize == 1) {
KeLstmForward<T, Op,
/* isBatch= */ false><<<grid, threads, 0, stream>>>(
op, value, frameSize, batchSize, active_node, active_gate,
active_state);
op, value, frameSize, batchSize);
} else {
KeLstmForward<T, Op,
/* isBatch= */ true><<<grid, threads, 0, stream>>>(
op, value, frameSize, batchSize, active_node, active_gate,
active_state);
op, value, frameSize, batchSize);
}
}
......@@ -240,13 +230,11 @@ void gpu_lstm_backward(const platform::DeviceContext& context, Op op,
if (batchSize == 1) {
KeLstmBackward<T, Op,
/* isBatch= */ false><<<grid, threads, 0, stream>>>(
op, value, grad, frameSize, batchSize, active_node, active_gate,
active_state);
op, value, grad, frameSize, batchSize);
} else {
KeLstmBackward<T, Op,
/* isBatch= */ true><<<grid, threads, 0, stream>>>(
op, value, grad, frameSize, batchSize, active_node, active_gate,
active_state);
op, value, grad, frameSize, batchSize);
}
}
......
paddle/operators/math/detail/lstm_kernel.h
浏览文件 @ 1e8300fa
......@@ -24,15 +24,29 @@ namespace detail {
namespace forward {
template <typename T>
DEVICE inline T sigmoid(const T a) {
const T min = SIGMOID_THRESHOLD_MIN;
const T max = SIGMOID_THRESHOLD_MAX;
T tmp = (a < min) ? min : ((a > max) ? max : a);
return static_cast<T>(1.0) / (static_cast<T>(1.0) + exp(-tmp));
}
template <typename T>
DEVICE inline T tanh(const T a) {
T tmp = -2.0 * a;
tmp = (tmp > EXP_MAX_INPUT) ? EXP_MAX_INPUT : tmp;
return (2.0 / (1.0 + exp(tmp))) - 1.0;
}
template <class T>
class lstm {
public:
HOSTDEVICE void operator()(T &valueIn, T &valueIg, T &valueFg, T &valueOg,
T &prevState, T &state, T &stateAtv, T &output,
T &checkI, T &checkF, T &checkO,
typename hppl::ForwardActType<T>::type actInput,
typename hppl::ForwardActType<T>::type actGate,
typename hppl::ForwardActType<T>::type actState) {
T &checkI, T &checkF, T &checkO) {
#if 0
// TODO(qingqing) support to activation speficed by users
valueIn = actInput(valueIn);
valueIg = actGate(valueIg + prevState * checkI);
valueFg = actGate(valueFg + prevState * checkF);
......@@ -40,6 +54,15 @@ class lstm {
valueOg = actGate(valueOg + state * checkO);
stateAtv = actState(state);
output = valueOg * stateAtv;
#else
valueIn = tanh<T>(valueIn);
valueIg = sigmoid<T>(valueIg + prevState * checkI);
valueFg = sigmoid<T>(valueFg + prevState * checkF);
state = valueIn * valueIg + prevState * valueFg;
valueOg = sigmoid<T>(valueOg + state * checkO);
stateAtv = tanh<T>(state);
output = valueOg * stateAtv;
#endif
}
#ifndef __NVCC__
#ifndef __AVX__ // If not compiled with AVX instructs. Disable AVX by default
......@@ -72,6 +95,16 @@ class lstm {
namespace backward {
template <typename T>
DEVICE inline T sigmoid(const T a, const T b) {
return a * b * (1.0 - b);
}
template <typename T>
DEVICE inline T tanh(const T a, const T b) {
return a * (1.0 - b * b);
}
template <class T>
class lstm {
public:
......@@ -80,10 +113,9 @@ class lstm {
T &prevState, T &prevStateGrad, T &state,
T &stateGrad, T &stateAtv, T &outputGrad,
T &checkI, T &checkF, T &checkO, T &checkIGrad,
T &checkFGrad, T &checkOGrad,
typename hppl::BackwardActType<T>::type actInput,
typename hppl::BackwardActType<T>::type actGate,
typename hppl::BackwardActType<T>::type actState) {
T &checkFGrad, T &checkOGrad) {
#if 0
// TODO(qingqing) support to activation speficed by users
gradOg = actGate(outputGrad * stateAtv, valueOg);
stateGrad += actState(outputGrad * valueOg, stateAtv) + gradOg * checkO;
gradIn = actInput(stateGrad * valueIg, valueIn);
......@@ -93,6 +125,17 @@ class lstm {
checkIGrad = gradIg * prevState;
checkFGrad = gradFg * prevState;
checkOGrad = gradOg * state;
#else
gradOg = sigmoid<T>(outputGrad * stateAtv, valueOg);
stateGrad += tanh<T>(outputGrad * valueOg, stateAtv) + gradOg * checkO;
gradIn = tanh<T>(stateGrad * valueIg, valueIn);
gradIg = sigmoid<T>(stateGrad * valueIn, valueIg);
gradFg = sigmoid<T>(stateGrad * prevState, valueFg);
prevStateGrad = gradIg * checkI + gradFg * checkF + stateGrad * valueFg;
checkIGrad = gradIg * prevState;
checkFGrad = gradFg * prevState;
checkOGrad = gradOg * state;
#endif
}
#ifndef __NVCC__
#ifndef __AVX__ // If not compiled with AVX instructs. Disable AVX by default
......
paddle/operators/math/math_function.cc
浏览文件 @ 1e8300fa
......@@ -211,6 +211,26 @@ void batched_gemm<platform::CPUPlace, double>(
}
#endif
template <>
void gemv<platform::CPUPlace, float>(const platform::DeviceContext& context,
const bool trans_a, const int M,
const int N, const float alpha,
const float* A, const float* B,
const float beta, float* C) {
CBLAS_TRANSPOSE transA = (trans_a == false) ? CblasNoTrans : CblasTrans;
cblas_sgemv(CblasRowMajor, transA, M, N, alpha, A, N, B, 1, beta, C, 1);
}
template <>
void gemv<platform::CPUPlace, double>(const platform::DeviceContext& context,
const bool trans_a, const int M,
const int N, const double alpha,
const double* A, const double* B,
const double beta, double* C) {
CBLAS_TRANSPOSE transA = (trans_a == false) ? CblasNoTrans : CblasTrans;
cblas_dgemv(CblasRowMajor, transA, M, N, alpha, A, N, B, 1, beta, C, 1);
}
template struct SetConstant<platform::CPUPlace, float>;
} // namespace math
......
paddle/operators/math/math_function.cu
浏览文件 @ 1e8300fa
......@@ -203,6 +203,33 @@ void batched_gemm<platform::GPUPlace, double>(
&beta, C, ldc, strideC, batchCount));
}
template <>
void gemv<platform::GPUPlace, float>(const platform::DeviceContext& context,
const bool trans_a, const int M,
const int N, const float alpha,
const float* A, const float* B,
const float beta, float* C) {
cublasOperation_t cuTransA = (trans_a == false) ? CUBLAS_OP_T : CUBLAS_OP_N;
PADDLE_ENFORCE(platform::dynload::cublasSgemv(
reinterpret_cast<const platform::CUDADeviceContext&>(context)
.cublas_handle(),
cuTransA, N, M, &alpha, A, N, B, 1, &beta, C, 1));
}
template <>
void gemv<platform::GPUPlace, double>(const platform::DeviceContext& context,
const bool trans_a, const int M,
const int N, const double alpha,
const double* A, const double* B,
const double beta, double* C) {
cublasOperation_t cuTransA = (trans_a == false) ? CUBLAS_OP_T : CUBLAS_OP_N;
PADDLE_ENFORCE(platform::dynload::cublasDgemv(
reinterpret_cast<const platform::CUDADeviceContext&>(context)
.cublas_handle(),
cuTransA, N, M, &alpha, A, N, B, 1, &beta, C, 1));
}
template struct SetConstant<platform::GPUPlace, float>;
} // namespace math
......
paddle/operators/math/math_function.h
浏览文件 @ 1e8300fa
......@@ -93,6 +93,11 @@ void batched_gemm(const platform::DeviceContext& context,
const T* A, const T* B, const T beta, T* C,
const int batchCount, const int strideA, const int strideB);
template <typename Place, typename T>
void gemv(const platform::DeviceContext& context, const bool trans_a,
const int M, const int N, const T alpha, const T* A, const T* B,
const T beta, T* C);
template <typename Place, typename T>
struct SetConstant {
void operator()(const platform::DeviceContext& context,
......
paddle/operators/math/math_function_test.cc
浏览文件 @ 1e8300fa
......@@ -89,3 +89,53 @@ TEST(math_function, zero) {
EXPECT_EQ(t[2], 1);
EXPECT_EQ(t[3], 1);
}
template <typename T>
void GemvTest(int m, int n, bool trans) {
paddle::framework::Tensor mat_a;
paddle::framework::Tensor vec_b;
paddle::framework::Tensor vec_c;
auto* cpu_place = new paddle::platform::CPUPlace();
int b_num = trans ? m : n;
int c_num = trans ? n : m;
T* data_a = mat_a.mutable_data<T>({m, n}, *cpu_place);
T* data_b = vec_b.mutable_data<T>({b_num}, *cpu_place);
T* data_c = vec_c.mutable_data<T>({c_num}, *cpu_place);
for (int i = 0; i < mat_a.numel(); ++i) {
data_a[i] = static_cast<T>(i);
}
for (int i = 0; i < vec_b.numel(); ++i) {
data_b[i] = static_cast<T>(i);
}
paddle::platform::CPUDeviceContext context(*cpu_place);
paddle::operators::math::gemv<paddle::platform::CPUPlace, T>(
context, trans, static_cast<int>(m), static_cast<int>(n), 1., data_a,
data_b, 0., data_c);
if (!trans) {
for (int i = 0; i < m; ++i) {
T sum = 0.0;
for (int j = 0; j < n; ++j) {
sum += data_a[i * n + j] * data_b[j];
}
ASSERT_FLOAT_EQ(data_c[i], sum);
}
} else {
for (int i = 0; i < n; ++i) {
T sum = 0.0;
for (int j = 0; j < m; ++j) {
sum += data_a[j * n + i] * data_b[j];
}
ASSERT_FLOAT_EQ(data_c[i], sum);
}
}
}
TEST(math_function, gemv) {
GemvTest<float>(3, 13, false);
GemvTest<double>(4, 5, false);
GemvTest<float>(12, 7, true);
GemvTest<double>(7, 9, true);
}
paddle/operators/math/math_function_test.cu
浏览文件 @ 1e8300fa
......@@ -177,3 +177,65 @@ TEST(math_function, gemm_trans_cublas) {
EXPECT_EQ(input3_ptr[7], 99);
delete gpu_place;
}
template <typename T>
void GemvTest(int m, int n, bool trans) {
paddle::framework::Tensor mat_a;
paddle::framework::Tensor vec_b;
paddle::framework::Tensor vec_c;
auto* cpu_place = new paddle::platform::CPUPlace();
T* data_a = mat_a.mutable_data<T>({m, n}, *cpu_place);
T* data_b = vec_b.mutable_data<T>({trans ? m : n}, *cpu_place);
T* data_c = vec_c.mutable_data<T>({trans ? n : m}, *cpu_place);
auto* gpu_place = new paddle::platform::GPUPlace(0);
paddle::framework::Tensor g_mat_a;
paddle::framework::Tensor g_vec_b;
paddle::framework::Tensor g_vec_c;
T* g_data_a = g_mat_a.mutable_data<T>(mat_a.dims(), *gpu_place);
T* g_data_b = g_vec_b.mutable_data<T>(vec_b.dims(), *gpu_place);
T* g_data_c = g_vec_c.mutable_data<T>(vec_c.dims(), *gpu_place);
for (int i = 0; i < mat_a.numel(); ++i) {
data_a[i] = static_cast<T>(i);
}
for (int i = 0; i < vec_b.numel(); ++i) {
data_b[i] = static_cast<T>(i);
}
paddle::platform::CUDADeviceContext context(*gpu_place);
g_mat_a.CopyFrom(mat_a, *gpu_place, context);
g_vec_b.CopyFrom(vec_b, *gpu_place, context);
paddle::operators::math::gemv<paddle::platform::GPUPlace, T>(
context, trans, static_cast<int>(m), static_cast<int>(n), 1., g_data_a,
g_data_b, 0., g_data_c);
vec_c.CopyFrom(g_vec_c, paddle::platform::CPUPlace(), context);
if (!trans) {
for (int i = 0; i < m; ++i) {
T sum = 0.0;
for (int j = 0; j < n; ++j) {
sum += data_a[i * n + j] * data_b[j];
}
ASSERT_FLOAT_EQ(data_c[i], sum);
}
} else {
for (int i = 0; i < n; ++i) {
T sum = 0.0;
for (int j = 0; j < m; ++j) {
sum += data_a[j * n + i] * data_b[j];
}
ASSERT_FLOAT_EQ(data_c[i], sum);
}
}
}
TEST(math_function, gemv) {
GemvTest<float>(3, 13, false);
GemvTest<double>(3, 13, false);
GemvTest<float>(3, 13, true);
GemvTest<double>(3, 13, true);
}
paddle/operators/math/sequence2batch.h
浏览文件 @ 1e8300fa
......@@ -53,7 +53,18 @@ class LoDTensor2BatchFunctor {
public:
void operator()(const platform::DeviceContext& context,
const framework::LoDTensor& lod_tensor,
framework::LoDTensor& batch, bool is_reverse) const {
framework::LoDTensor& batch, bool is_cal_batch_lod,
bool is_reverse = false) const {
if (!is_cal_batch_lod) {
auto lods = batch.lod();
PADDLE_ENFORCE_EQ(lods.size(), 2UL);
PADDLE_ENFORCE_EQ(lods[1].size(),
static_cast<size_t>(lod_tensor.dims()[0]));
CopyMatrixRowsFunctor<Place, T> to_batch;
to_batch(context, lod_tensor, lods[1].data(), batch, true);
return;
}
auto lods = lod_tensor.lod();
PADDLE_ENFORCE_EQ(lods.size(), 1UL, "Only support one level sequence now.");
auto lod = lods[0];
......@@ -101,10 +112,10 @@ class LoDTensor2BatchFunctor {
size_t* batch_starts = batch_lods[0].data();
size_t* seq2batch_idx = batch_lods[1].data();
batch_starts[0] = 0;
for (size_t n = 0; n < num_batch; n++) {
for (int n = 0; n < num_batch; n++) {
auto batch_id = static_cast<int>(batch_starts[n]);
for (size_t i = 0; i < seq_info.size(); ++i) {
size_t seq_len = seq_info[i].length;
int seq_len = seq_info[i].length;
int start = seq_info[i].start;
if (n < seq_len) {
seq2batch_idx[batch_id] =
......@@ -132,11 +143,8 @@ class Batch2LoDTensorFunctor {
auto in_lod = batch.lod();
PADDLE_ENFORCE_EQ(in_lod.size(), 2UL,
"The LoD size of input `batch` should be 2.");
auto out_lod = lod_tensor.lod()[0];
auto num = out_lod[out_lod.size() - 1];
PADDLE_ENFORCE_EQ(num, lod_tensor.dims()[0]);
PADDLE_ENFORCE_EQ(num, in_lod[1].size());
PADDLE_ENFORCE_EQ(num, batch.dims()[0]);
PADDLE_ENFORCE_EQ(in_lod[1].size(),
static_cast<size_t>(lod_tensor.dims()[0]));
CopyMatrixRowsFunctor<Place, T> to_seq;
size_t* index = in_lod[1].data();
to_seq(context, batch, index, lod_tensor, false);
......
paddle/operators/nccl_op_test.cu
浏览文件 @ 1e8300fa
......@@ -185,7 +185,7 @@ TEST_F(NCCLTester, ncclAllReduceOp) {
recv_tensor.numel() * sizeof(float),
static_cast<p::CUDADeviceContext *>(dev_ctxs[i])->stream());
for (size_t j = 0; j < f::product(kDims); ++j) {
for (int64_t j = 0; j < f::product(kDims); ++j) {
ASSERT_NEAR(ct[j], result, 1e-5);
}
}
......@@ -234,7 +234,7 @@ TEST_F(NCCLTester, ncclReduceOp) {
recv_tensor.numel() * sizeof(float),
static_cast<p::CUDADeviceContext *>(dev_ctxs[kRoot])->stream());
for (int j = 0; j < f::product(kDims); ++j) {
for (int64_t j = 0; j < f::product(kDims); ++j) {
ASSERT_NEAR(ct[j], result, 1e-5);
}
}
......@@ -282,7 +282,7 @@ TEST_F(NCCLTester, ncclBcastOp) {
recv_tensor.numel() * sizeof(float),
static_cast<p::CUDADeviceContext *>(dev_ctxs[idx])->stream());
for (size_t j = 0; j < f::product(kDims); ++j) {
for (int64_t j = 0; j < f::product(kDims); ++j) {
ASSERT_NEAR(ct[j], result, 1e-5);
}
}
......
paddle/operators/pool_cudnn_op.cu
浏览文件 @ 1e8300fa
......@@ -43,6 +43,7 @@ class PoolCudnnOpKernel : public framework::OpKernel<T> {
std::vector<int> paddings = ctx.Attr<std::vector<int>>("paddings");
if (ctx.Attr<bool>("globalPooling")) {
for (size_t i = 0; i < ksize.size(); ++i) {
paddings[i] = 0;
ksize[i] = static_cast<int>(input->dims()[i + 2]);
}
}
......@@ -97,8 +98,10 @@ class PoolCudnnGradOpKernel : public framework::OpKernel<T> {
std::vector<int> paddings = ctx.Attr<std::vector<int>>("paddings");
if (ctx.Attr<bool>("globalPooling")) {
for (size_t i = 0; i < ksize.size(); ++i)
for (size_t i = 0; i < ksize.size(); ++i) {
paddings[i] = 0;
ksize[i] = static_cast<int>(input->dims()[i + 2]);
}
}
const T *input_data = input->data<T>();
......
paddle/operators/pool_op.cc
浏览文件 @ 1e8300fa
......@@ -39,8 +39,10 @@ void PoolOp::InferShape(framework::InferShapeContext *ctx) const {
if (ctx->Attrs().Get<bool>("globalPooling")) {
ksize.resize(static_cast<size_t>(in_x_dims.size()) - 2);
for (size_t i = 0; i < ksize.size(); ++i)
for (size_t i = 0; i < ksize.size(); ++i) {
paddings[i] = 0;
ksize[i] = static_cast<int>(in_x_dims[i + 2]);
}
}
PADDLE_ENFORCE(in_x_dims.size() - ksize.size() == 2U,
......@@ -84,15 +86,16 @@ Pool2dOpMaker::Pool2dOpMaker(framework::OpProto *proto,
"(string), pooling type, can be \"max\" for max-pooling "
"and \"avg\" for average-pooling.")
.InEnum({"max", "avg"});
AddAttr<std::vector<int>>(
"ksize",
"(vector ), the pooling window size(height, width) of pooling operator."
"If globalPooling = true, ksize is ignored and need not be "
"specified."); // TODO(Chengduo): Add checker. (Currently,
AddAttr<std::vector<int>>("ksize",
"(vector ), the pooling window size(height, width) "
"of pooling operator."
"If globalPooling = true, ksize and paddings will "
"be ignored."); // TODO(Chengduo): Add checker.
// (Currently,
// TypedAttrChecker don't support vector type.)
AddAttr<bool>("globalPooling",
"(bool default: false), whether to use the global pooling."
"If globalPooling = true, ksize is ignored.")
"If globalPooling = true, ksize and paddings will be ignored.")
.SetDefault(false);
AddAttr<std::vector<int>>(
"strides",
......@@ -101,7 +104,8 @@ Pool2dOpMaker::Pool2dOpMaker(framework::OpProto *proto,
// TypedAttrChecker don't support vector type.)
AddAttr<std::vector<int>>(
"paddings",
"(vector defalut:{0,0}), paddings(height, width) of pooling operator.")
"(vector defalut:{0,0}), paddings(height, width) of pooling operator."
"If globalPooling = true, paddings and ksize will be ignored.")
.SetDefault({0, 0}); // TODO(Chengduo): Add checker. (Currently,
// TypedAttrChecker don't support vector type.)
......@@ -145,25 +149,28 @@ Pool3dOpMaker::Pool3dOpMaker(framework::OpProto *proto,
"(string), pooling type, can be \"max\" for max-pooling "
"and \"avg\" for average-pooling.")
.InEnum({"max", "avg"});
AddAttr<std::vector<int>>(
"ksize",
"(vector ), the pooling window size(depth, height, width) of pooling "
"operator."
"If globalPooling = true, ksize is ignored and need not be "
"specified."); // TODO(Chengduo): Add checker. (Currently,
// TypedAttrChecker don't support vector type.)
AddAttr<std::vector<int>>("ksize",
"(vector ), the pooling window size(depth, height, "
"width) of pooling "
"operator."
"If globalPooling = true, ksize and paddings wille "
"be ignored."); // TODO(Chengduo): Add checker.
// (Currently,
// TypedAttrChecker don't support vector type.)
AddAttr<bool>("globalPooling",
"(bool default: false), whether to use the global pooling."
"If globalPooling = true, ksize is ignored.")
"If globalPooling = true, ksize and paddings wille be ignored.")
.SetDefault(false);
AddAttr<std::vector<int>>("strides",
"(vector, default:{1,1,1}), strides(depth, height, "
"width) of pooling operator.")
.SetDefault({1, 1, 1}); // TODO(Chengduo): Add checker. (Currently,
// TypedAttrChecker don't support vector type.)
AddAttr<std::vector<int>>("paddings",
"(vector defalut:{0,0,0}), paddings(depth, height, "
"width) of pooling operator.")
AddAttr<std::vector<int>>(
"paddings",
"(vector defalut:{0,0,0}), paddings(depth, height, "
"width) of pooling operator."
"If globalPooling = true, ksize and paddings wille be ignored.")
.SetDefault({0, 0, 0}); // TODO(Chengduo): Add checker. (Currently,
// TypedAttrChecker don't support vector type.)
......
paddle/operators/pool_op.h
浏览文件 @ 1e8300fa
......@@ -63,6 +63,7 @@ class PoolKernel : public framework::OpKernel<T> {
std::vector<int> paddings = context.Attr<std::vector<int>>("paddings");
if (context.Attr<bool>("globalPooling")) {
for (size_t i = 0; i < ksize.size(); ++i) {
paddings[i] = 0;
ksize[i] = static_cast<int>(in_x->dims()[i + 2]);
}
}
......@@ -103,6 +104,7 @@ class PoolKernel : public framework::OpKernel<T> {
paddings, pool_process);
}
} break;
default: { PADDLE_THROW("Pool op only supports 2D and 3D input."); }
}
}
};
......@@ -123,8 +125,10 @@ class PoolGradKernel : public framework::OpKernel<T> {
std::vector<int> paddings = context.Attr<std::vector<int>>("paddings");
if (context.Attr<bool>("globalPooling")) {
for (size_t i = 0; i < ksize.size(); ++i)
for (size_t i = 0; i < ksize.size(); ++i) {
paddings[i] = 0;
ksize[i] = static_cast<int>(in_x->dims()[i + 2]);
}
}
if (in_x_grad) {
......@@ -164,6 +168,7 @@ class PoolGradKernel : public framework::OpKernel<T> {
*out_grad, ksize, strides, paddings, pool_process);
}
} break;
default: { PADDLE_THROW("Pool op only supports 2D and 3D input."); }
}
}
}
......
paddle/operators/reshape_op.cc
浏览文件 @ 1e8300fa
......@@ -36,7 +36,7 @@ class ReshapeOp : public framework::OperatorWithKernel {
PADDLE_ENFORCE(shape.size() > 0, "Attr(shape) shouldn't be empty.");
auto x_dims = ctx->GetInputDim("X");
// TODO(qiao) change batch_size
for (int i = 1; i < shape.size(); ++i) {
for (size_t i = 1; i < shape.size(); ++i) {
PADDLE_ENFORCE(shape[i] > 0,
"Each dimension of shape "
"must be positiv except the first.");
......
paddle/operators/seq_expand_op.cc 0 → 100644
浏览文件 @ 1e8300fa
此差异已折叠。
paddle/operators/seq_expand_op.cu 0 → 100644
浏览文件 @ 1e8300fa
此差异已折叠。
paddle/operators/seq_expand_op.h 0 → 100644
浏览文件 @ 1e8300fa
此差异已折叠。
paddle/operators/sequence_conv_op.h
浏览文件 @ 1e8300fa
此差异已折叠。
paddle/operators/sequence_pool_op.h
浏览文件 @ 1e8300fa
此差异已折叠。
paddle/operators/top_k_op.cc
浏览文件 @ 1e8300fa
此差异已折叠。
paddle/operators/top_k_op.cu
浏览文件 @ 1e8300fa
此差异已折叠。
paddle/operators/top_k_op.h
浏览文件 @ 1e8300fa
此差异已折叠。
paddle/platform/CMakeLists.txt
浏览文件 @ 1e8300fa
......@@ -9,7 +9,6 @@ cc_test(place_test SRCS place_test.cc DEPS place glog gflags)
add_subdirectory(dynload)
cc_test(enforce_test SRCS enforce_test.cc DEPS stringpiece)
cc_test(environment_test SRCS environment_test.cc DEPS stringpiece)
IF(WITH_GPU)
set(GPU_CTX_DEPS dynload_cuda dynamic_loader)
......
paddle/platform/environment.h 已删除 100644 → 0
浏览文件 @ 3eb42bfd
此差异已折叠。
paddle/platform/environment_test.cc 已删除 100644 → 0
浏览文件 @ 3eb42bfd
此差异已折叠。
paddle/platform/gpu_info.cc
浏览文件 @ 1e8300fa
此差异已折叠。
paddle/pybind/protobuf.cc
浏览文件 @ 1e8300fa
此差异已折叠。
paddle/pybind/pybind.cc
浏览文件 @ 1e8300fa
此差异已折叠。
paddle/trainer/MergeModel.cpp
浏览文件 @ 1e8300fa
此差异已折叠。
proto/CMakeLists.txt
浏览文件 @ 1e8300fa
此差异已折叠。
python/paddle/v2/dataset/imdb.py
浏览文件 @ 1e8300fa
此差异已折叠。
python/paddle/v2/framework/initializer.py 0 → 100644
浏览文件 @ 1e8300fa
此差异已折叠。
python/paddle/v2/framework/nets.py
浏览文件 @ 1e8300fa
此差异已折叠。
python/paddle/v2/framework/tests/test_initializer.py 0 → 100644
浏览文件 @ 1e8300fa
此差异已折叠。
python/paddle/v2/framework/tests/test_seq_expand.py 0 → 100644
浏览文件 @ 1e8300fa
此差异已折叠。
python/paddle/v2/trainer.py
浏览文件 @ 1e8300fa
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