diff --git a/Dockerfile b/Dockerfile index 98f61ba586a681e53b435d592c8e43b1cc964139..136db772cc6a24b8084120fa6bab666bc1eda78e 100644 --- a/Dockerfile +++ b/Dockerfile @@ -10,13 +10,11 @@ RUN /bin/bash -c 'if [[ -n ${UBUNTU_MIRROR} ]]; then sed -i 's#http://archive.ub ARG WITH_GPU ARG WITH_AVX ARG WITH_DOC -ARG WITH_STYLE_CHECK ENV WOBOQ OFF -ENV WITH_GPU=${WITH_GPU:-OFF} +ENV WITH_GPU=${WITH_GPU:-ON} ENV WITH_AVX=${WITH_AVX:-ON} ENV WITH_DOC=${WITH_DOC:-OFF} -ENV WITH_STYLE_CHECK=${WITH_STYLE_CHECK:-OFF} ENV HOME /root # Add bash enhancements diff --git a/doc/howto/dev/build_cn.md b/doc/howto/dev/build_cn.md new file mode 100644 index 0000000000000000000000000000000000000000..0b911f7b7509da4a147c65954acb7e7c38f489da --- /dev/null +++ b/doc/howto/dev/build_cn.md @@ -0,0 +1,124 @@ +# 编译PaddlePaddle和运行单元测试 + +## 需要的软硬件 + +为了开发PaddlePaddle，我们需要 + +1. 一台电脑，可以装的是 Linux, BSD, Windows 或者 MacOS 操作系统，以及 +1. Docker。 + +不需要依赖其他任何软件了。即便是 Python 和 GCC 都不需要，因为我们会把所有编译工具都安装进一个 Docker image 里。 + +## 总体流程 + +1. 获取源码 + + ```bash + git clone https://github.com/paddlepaddle/paddle + ``` + +2. 安装开发工具到 Docker image 里 + + ```bash + cd paddle; docker build -t paddle:dev . + ``` + + 请注意这个命令结尾处的 `.`；它表示 `docker build` 应该读取当前目录下的 [`Dockerfile`文件](https://github.com/PaddlePaddle/Paddle/blob/develop/Dockerfile)，按照其内容创建一个名为 `paddle:dev` 的 Docker image，并且把各种开发工具安装进去。 + +3. 编译 + + 以下命令启动一个 Docker container 来执行 `paddle:dev` 这个 Docker image，同时把当前目录（源码树根目录）映射为 container 里的 `/paddle` 目录，并且运行 `Dockerfile` 描述的默认入口程序 [`build.sh`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/scripts/docker/build.sh)。这个脚本调用 `cmake` 和 `make` 来编译 `/paddle` 里的源码，结果输出到 `/paddle/build`，也就是本地的源码树根目录里的 `build` 子目录。 + + ```bash + docker run --rm -v $PWD:/paddle paddle:dev + ``` + + 上述命令编译出一个 CUDA-enabled 版本。如果我们只需要编译一个只支持 CPU 的版本，可以用 + + ```bash + docker run --rm -e WITH_GPU=OFF -v $PWD:/paddle paddle:dev + ``` + +4. 运行单元测试 + + 用本机的第一个 GPU 来运行包括 GPU 单元测试在内的所有单元测试： + + ```bash + NV_GPU=0 nvidia-docker run --rm -v $PWD:/paddle paddle:dev bash -c "cd /paddle/build; ctest" + ``` + + 如果编译的时候我们用了 `WITH_GPU=OFF` 选项，那么编译过程只会产生 CPU-based 单元测试，那么我们也就不需要 nvidia-docker 来运行单元测试了。我们只需要： + + ```bash + docker run --rm -v $PWD:/paddle paddle:dev bash -c "cd /paddle/build; ctest" + ``` + + 有时候我们只想运行一个特定的单元测试，比如 `memory_test`，我们可以 + + ```bash + nvidia-docker run --rm -v $PWD:/paddle paddle:dev bash -c "cd /paddle/build; ctest -V -R memory_test" + ``` + +5. 清理 + + 有时候我们会希望清理掉已经下载的第三方依赖以及已经编译的二进制文件。此时只需要： + + ```bash + rm -rf build + ``` + +## 为什么要 Docker 呀？ + +- 什么是 Docker? + + 如果您没有听说 Docker，可以把它想象为一个类似 virtualenv 的系统，但是虚拟的不仅仅是 Python 的运行环境。 + +- Docker 还是虚拟机？ + + 有人用虚拟机来类比 Docker。需要强调的是：Docker 不会虚拟任何硬件，Docker container 里运行的编译工具实际上都是在本机的 CPU 和操作系统上直接运行的，性能和把编译工具安装在本机运行一样。 + +- 为什么用 Docker? + + 把工具和配置都安装在一个 Docker image 里可以标准化编译环境。这样如果遇到问题，其他人可以复现问题以便帮助。 + + 另外，对于习惯使用Windows和MacOS的开发者来说，使用Docker就不用配置交叉编译环境了。 + +- 我可以选择不用Docker吗？ + + 当然可以。大家可以用把开发工具安装进入 Docker image 一样的方式，把这些工具安装到本机。这篇文档介绍基于 Docker 的开发流程，是因为这个流程比其他方法都更简便。 + +- 学习 Docker 有多难？ + + 理解 Docker 并不难，大概花十分钟看一下[这篇文章](https://zhuanlan.zhihu.com/p/19902938)。这可以帮您省掉花一小时安装和配置各种开发工具，以及切换机器时需要新安装的辛苦。别忘了 PaddlePaddle 更新可能导致需要新的开发工具。更别提简化问题复现带来的好处了。 + +- 我可以用 IDE 吗？ + + 当然可以，因为源码就在本机上。IDE 默认调用 make 之类的程序来编译源码，我们只需要配置 IDE 来调用 Docker 命令编译源码即可。 + + 很多 PaddlePaddle 开发者使用 Emacs。他们在自己的 `~/.emacs` 配置文件里加两行 + + ```emacs + (global-set-key "\C-cc" 'compile) + (setq compile-command + "docker run --rm -it -v $(git rev-parse --show-toplevel):/paddle paddle:dev") + ``` + + 就可以按 `Ctrl-C` 和 `c` 键来启动编译了。 + +- 可以并行编译吗？ + + 是的。我们的 Docker image 运行一个 [Bash 脚本](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/scripts/docker/build.sh)。这个脚本调用 `make -j$(nproc)` 来启动和 CPU 核一样多的进程来并行编译。 + +## 可能碰到的问题 + +- Docker 需要 sudo + + 如果用自己的电脑开发，自然也就有管理员权限（sudo）了。如果用公用的电脑开发，需要请管理员安装和配置好 Docker。此外，PaddlePaddle 项目在努力开始支持其他不需要 sudo 的集装箱技术，比如 rkt。 + +- 在 Windows/MacOS 上编译很慢 + + Docker 在 Windows 和 MacOS 都可以运行。不过实际上是运行在一个 Linux 虚拟机上。可能需要注意给这个虚拟机多分配一些 CPU 和内存，以保证编译高效。具体做法请参考[这个issue](https://github.com/PaddlePaddle/Paddle/issues/627)。 + +- 磁盘不够 + + 本文中的例子里，`docker run` 命令里都用了 `--rm` 参数，这样保证运行结束之后的 containers 不会保留在磁盘上。可以用 `docker ps -a` 命令看到停止后但是没有删除的 containers。`docker build` 命令有时候会产生一些中间结果，是没有名字的 images，也会占用磁盘。可以参考[这篇文章](https://zaiste.net/posts/removing_docker_containers/)来清理这些内容。 diff --git a/doc/howto/dev/build_en.md b/doc/howto/dev/build_en.md new file mode 100644 index 0000000000000000000000000000000000000000..d0048e3714a5861a503736879d6c0870e5906c95 --- /dev/null +++ b/doc/howto/dev/build_en.md @@ -0,0 +1,124 @@ +# Build PaddlePaddle from Source Code and Run Unit Test + +## What Developers Need + +To contribute to PaddlePaddle, you need + +1. A computer -- Linux, BSD, Windows, MacOS, and +1. Docker. + +Nothing else. Not even Python and GCC, because you can install all build tools into a Docker image. We run all the tools by running this image. + +## General Process + +1. Retrieve source code. + + ```bash + git clone https://github.com/paddlepaddle/paddle + ``` + +2. Install build tools into a Docker image. + + ```bash + cd paddle; docker build -t paddle:dev . + ``` + + Please be aware of the `.` at the end of the command, which refers to the [`./Dockerfile` file](https://github.com/PaddlePaddle/Paddle/blob/develop/Dockerfile). `docker build` follows instructions in this file to create a Docker image named `paddle:dev`, and installs building tools into it. + +3. Build from source. + + This following command starts a Docker container that executes the Docker image `paddle:dev`, mapping the current directory to `/paddle/` in the container, and runs the default entry-point [`build.sh`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/scripts/docker/build.sh) as specified in the Dockefile. `build.sh` invokes `cmake` and `make` to build PaddlePaddle source code, which had been mapped to `/paddle`, and writes outputs to `/paddle/build`, which maps to `build` in the current source directory on the computer. + + ```bash + docker run -v $PWD:/paddle paddle:dev + ``` + + Above command builds a CUDA-enabled version. If we want to build a CPU-only version, we can type + + ```bash + docker run -e WITH_GPU=OFF -v $PWD:/paddle paddle:dev + ``` + +4. Run unit tests. + + To run all unit tests using the first GPU of a node: + + ```bash + NV_GPU=0 nvidia-docker run -v $PWD:/paddle paddle:dev bash -c "cd /paddle/build; ctest" + ``` + + If we used `WITH_GPU=OFF` at build time, it generates only CPU-based unit tests, and we don't need nvidia-docker to run them. We can just run + + ```bash + docker run -v $PWD:/paddle paddle:dev bash -c "cd /paddle/build; ctest" + ``` + + Sometimes we want to run a specific unit test, say `memory_test`, we can run + + ```bash + nvidia-docker run -v $PWD:/paddle paddle:dev bash -c "cd /paddle/build; ctest -V -R memory_test" + ``` + +5. Clean Build. + + Sometimes, we might want to clean all thirt-party dependents and built binaries. To do so, just + + ```bash + rm -rf build + ``` + +## Docker, Or Not? + +- What is Docker? + + If you haven't heard of it, consider it something like Python's virtualenv. + +- Docker or virtual machine? + + Some people compare Docker with VMs, but Docker doesn't virtualize any hardware nor running a guest OS, which means there is no compromise on the performance. + +- Why Docker? + + Using a Docker image of build tools standardizes the building environment, which makes it easier for others to reproduce your problems and to help. + + Also, some build tools don't run on Windows or Mac or BSD, but Docker runs almost everywhere, so developers can use whatever computer they want. + +- Can I choose not to use Docker? + + Sure, you don't have to install build tools into a Docker image; instead, you can install them in your local computer. This document exists because Docker would make the development way easier. + +- How difficult is it to learn Docker? + + It takes you ten minutes to read [an introductory article](https://docs.docker.com/get-started) and saves you more than one hour to install all required build tools, configure them, especially when new versions of PaddlePaddle require some new tools. Not even to mention the time saved when other people trying to reproduce the issue you have. + +- Can I use my favorite IDE? + + Yes, of course. The source code resides on your local computer, and you can edit it using whatever editor you like. + + Many PaddlePaddle developers are using Emacs. They add the following few lines into their `~/.emacs` configure file: + + ```emacs + (global-set-key "\C-cc" 'compile) + (setq compile-command + "docker run --rm -it -v $(git rev-parse --show-toplevel):/paddle paddle:dev") + ``` + + so they could type `Ctrl-C` and `c` to build PaddlePaddle from source. + +- Does Docker do parallel building? + + Our building Docker image runs a [Bash script](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/scripts/docker/build.sh), which calls `make -j$(nproc)` to starts as many processes as the number of your CPU cores. + +## Some Gotchas + +- Docker requires sudo + + An owner of a computer has the administrative privilege, a.k.a., sudo, and Docker requires this privilege to work properly. If you use a shared computer for development, please ask the administrator to install and configure Docker. We will do our best to support rkt, another container technology that doesn't require sudo. + +- Docker on Windows/MacOS builds slowly + + On Windows and MacOS, Docker containers run in a Linux VM. You might want to give this VM some more memory and CPUs so to make the building efficient. Please refer to [this issue](https://github.com/PaddlePaddle/Paddle/issues/627) for details. + +- Not enough disk space + + Examples in this article uses option `--rm` with the `docker run` command. This option ensures that stopped containers do not exist on hard disks. We can use `docker ps -a` to list all containers, including stopped. Sometimes `docker build` generates some intermediate dangling images, which also take disk space. To clean them, please refer to [this article](https://zaiste.net/posts/removing_docker_containers/). diff --git a/doc/howto/dev/new_op_cn.md b/doc/howto/dev/new_op_cn.md new file mode 100644 index 0000000000000000000000000000000000000000..ebd2cf3ff04567e3f34b1707696d025b834c58eb --- /dev/null +++ b/doc/howto/dev/new_op_cn.md @@ -0,0 +1,300 @@ +# 如何写新的Operator + + - [概念简介](#概念简介) + - [实现C++类](#实现C++类) + - [定义ProtoMaker类](#定义ProtoMaker类) + - [定义Operator类](#定义Operator类) + - [定义OpKernel类](#定义OpKernel类) + - [注册类](#注册类) + - [编译](#编译) + - [绑定Python](#绑定Python) + - [实现单元测试](#实现单元测试) + - [前向Operator单测](#前向Operator单测) + - [反向Operator单测](#反向Operator单测) + + +## 概念简介 + +简单介绍需要用到基类，详细介绍请参考设计文档。 + +- `framework::OperatorBase`: Operator(简写，Op)基类。 +- `framework::OpKernel`: Op计算函数的基类，称作Kernel。 +- `framework::OperatorWithKernel`：继承自OperatorBase，Op有计算函数，称作有Kernel。 +- `class OpProtoAndCheckerMaker`：描述该Op的输入、输出、属性、注释,主要用于Python API接口生成 + +依据是否包含kernel，将Op分为两种：包含Kernel的Op和不包含kernel的Op，前者Op的定义继承自`OperatorBase`，后者继承自`OperatorWithKernel`。本教程主要介绍带Kernel的Op如何写，简单总结如下： + +Forward Op需要包含： + + - OpProtoMake定义 + - Op定义 + - Kernel实现 + +与之对应的Backward Op包含： + + - Op定义 + - Kernel实现 + +下面以矩阵乘操作，即[MulOp](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/mul_op.cc)为例来介绍如何写带Kernel的Operator。 + + +## 实现C++类 + + +### 1. 定义ProtoMaker类 + +矩阵乘的公式：$Out = X * Y$, 可见该计算由两个输入，一个输出组成。首先定义`ProtoMaker`来描述该Op的输入、输出及注释： + +``` +class MulOpMaker : public framework::OpProtoAndCheckerMaker { + public: + MulOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "The first input of mul op"); + AddInput("Y", "The second input of mul op"); + AddOutput("Out", "The output of mul op"); + AddComment(R"DOC( +Two Element Mul Operator. +The equation is: Out = X * Y +)DOC"); + } +}; +``` + +[`MulOpMaker`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/mul_op.cc#L43)继承自`framework::OpProtoAndCheckerMaker`，构造函数包括2个： + + - `framework::OpProto` ： 前者存储Op的输入输出和参数属性，将用于Python API接口的生成。 + - `framework::OpAttrChecker` ：后者用于检查参数属性的合法性。 + +构造函数里通过`AddInput`添加输入参数，通过`AddOutput`添加输出参数，通过`AddComment`添加该Op的注释，这些函数会将对应内容添加到`OpProto`中。 + +在`MulOp`中添加两个输入`X`和`Y`，添加了一个输出`Out`，并解释了各自含义，该命名尽可能的规范。 + + +再举个[`ScaleOp`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/scale_op.cc#L37)的例子： + +``` +template +class ScaleOpMaker : public framework::OpProtoAndCheckerMaker { + public: + ScaleOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "The input tensor of scale operator.").NotInGradient(); + AddOutput("Out", "The output tensor of scale operator.").NotInGradient(); + AddComment(R"DOC(Scale operator +The equation is: Out = scale*X +)DOC"); + AddAttr("scale", "scale of scale operator.").SetDefault(1.0); + } +}; +``` + + 在这个例子里，两处不同： + + - `AddInput("X","...").NotInGradient()` : 表示`X`这个输入不参与`ScaleOp`对应的梯度Op计算之中。 + - `AddAttr("scale", "...").SetDefault(1.0);` : 增加`scale`系数，作为参数属性，并且设置默认值为1.0。 + + +### 2. 定义Operator类 + + +```c++ +class MulOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(const framework::InferShapeContext &ctx) const override { + auto dim0 = ctx.Input("X")->dims(); + auto dim1 = ctx.Input("Y")->dims(); + PADDLE_ENFORCE_EQ(dim0.size(), 2, + "input X(%s) should be a tensor with 2 dims, a matrix", + ctx.op_.Input("X")); + PADDLE_ENFORCE_EQ(dim1.size(), 2, + "input Y(%s) should be a tensor with 2 dims, a matrix", + ctx.op_.Input("Y")); + PADDLE_ENFORCE_EQ( + dim0[1], dim1[0], + "First matrix's width must be equal with second matrix's height."); + ctx.Output("Out")->Resize({dim0[0], dim1[1]}); + } +}; +``` + +[`MulOp`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/mul_op.cc#L22)继承自`OperatorWithKernel`。`public`成员： + +```c++ +using framework::OperatorWithKernel::OperatorWithKernel; +``` + +这句表示使用基类`OperatorWithKernel`的构造函数，也可写成： + +```c++ +MulOp(const std::string &type, const framework::VariableNameMap &inputs, + const framework::VariableNameMap &outputs, + const framework::AttributeMap &attrs) + : OperatorWithKernel(type, inputs, outputs, attrs) {} +``` + +还需要重写`InferShape`接口。`InferShape`为const函数，不能修改Op的成员变量，参数为`const framework::InferShapeContext &ctx`，通过该参数可获取到输入输出以及属性。它的功能是： + - 1). 做检查， 尽早报错：检查输入数据维度、类型等是否合法 + - 2). 设置输出Tensor的形状 + +通常`OpProtoMaker`和`Op`类的定义写在`.cc`文件中，和要讲到的注册函数一起放在`.cc`中 + +### 3. 定义OpKernel类 + +```C++ +template +class MulKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto* X = context.Input("X"); + auto* Y = context.Input("Y"); + auto* Z = context.Output("Out"); + Z->mutable_data(context.GetPlace()); + auto* device_context = + const_cast(context.device_context_); + math::matmul(*X, false, *Y, false, 1, Z, 0, device_context); + } +}; +``` + +`MulKernel`继承自`framework::OpKernel`，带有模板参数: + + - `typename Place`: 表示设备类型，不同设备(CPU、GPU)共享同一个Kernel时，需加该模板参数，不共享则不加，一个不共享的例子是[`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43)。 + + - `typename T` : 表示数据类型，如`float`, `double`等。 + +`MulKernel`需要重写`Compute`接口，该接口参数为`const framework::ExecutionContext& context`, `ExecutionContext`相比`InferShapeContext`增加了设备类型，同样可获取到输入输出和属性参数，`Compute`函数里写具体实现时。 + +注意，不同设备(CPU、GPU)共享一个Op定义，是否则共享同一个`OpKernel`，取决于`Compute`调用的函数是否支持不同设备。`MulOp`的CPU、GPU实现共享同一个`Kernel`，`OpKernel`不共享的例子可以参考[`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43)。 + +到此前向Op实现完成，需要在`.cc`文件中注册该op和kernel。反向Op类的定义和Kernel定义与前向Op类似，这里不再重复。但注意，反向Op没有`ProtoMaker`。 + +### 4. 注册类 + +在`.cc`文件中注册前向、反向Op类，注册CPU Kernel。 + +```c++ +namespace ops = paddle::operators; +REGISTER_OP(mul, ops::MulOp, ops::MulOpMaker, mul_grad, ops::MulOpGrad); +REGISTER_OP_CPU_KERNEL(mul, ops::MulKernel); +REGISTER_OP_CPU_KERNEL(mul_grad, + ops::MulGradKernel); +``` + + - `REGISTER_OP` ： 注册`ops::MulOp`类，类型名为`mul`，该类的`ProtoMaker`为`ops::MulOpMaker`，注册`ops::MulOpGrad`，类型名为`mul_grad`， + - `REGISTER_OP_WITHOUT_GRADIENT` ： 用于注册没有反向的Op。 + - `REGISTER_OP_CPU_KERNEL` ：注册`ops::MulKernel`类，并特化模板参数为`paddle::platform::CPUPlace`和`float`类型，同理，注册`ops::MulKernel`类。 + +在 `.cu`文件中注册GPU Kernel。 + +```c++ +namespace ops = paddle::operators; +REGISTER_OP_GPU_KERNEL(mul, ops::MulKernel); +REGISTER_OP_GPU_KERNEL(mul_grad, + ops::MulGradKernel); +``` + +### 5. 编译 + +在[paddle/operators/CMakeLists.txt](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/CMakeLists.txt)文件中添加编译。 + +``` +op_library(mul_op SRCS mul_op.cc mul_op.cu DEPS math_function) +``` + +下面命令可以编译： + +``` +make mul_op +``` + +## 绑定Python + +- 绑定Python + + 在 [`paddle/pybind/pybind.cc +`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/pybind/pybind.cc)文件中添加该类： + + ``` + USE_OP(mul); + ``` + 如果只实现了CPU版本，则使用`USE_CPU_ONLY_OP`: + + ``` + USE_CPU_ONLY_OP(gather); + ``` + + 使用`USE_OP`告知编译器需要链接该Op的目标文件，具体解释参考[代码注释](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/op_registry.h#L81)。 + + + - 生成库 + + 在 [`paddle/pybind/CMakeLists.txt`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/pybind/CMakeLists.txt)文件添加类到`DEPS`中，使得该Op可以链接到生成的lib库中。 + + ``` + if(WITH_PYTHON) + cc_library(paddle_pybind SHARED + SRCS pybind.cc + DEPS pybind python backward + mul_op + minus_op) + endif(WITH_PYTHON) + ``` + +## 实现单元测试 + +单测包括对比前向Op不同设备(CPU、GPU)的实现、对比反向OP不同设备(CPU、GPU)的实现、反向Op的梯度测试。下面介绍介绍[`MulOp`的单测](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/framework/tests/test_mul_op.py)。 + +### 前向Operator单测 + +前向Op单测继承自`unittest.TestCase`，并定义元类`__metaclass__ = OpTestMeta`，具体单测流程在`OpTestMeta`里完成。需在`setUp`函数定义输入输出和属性参数，以及Python对比的输出值。 + +``` +import unittest +import numpy as np +from gradient_checker import GradientChecker, create_op +from op_test_util import OpTestMeta + +class TestMulOp(unittest.TestCase): + __metaclass__ = OpTestMeta + + def setUp(self): + self.type = "mul" + self.inputs = { + 'X': np.random.random((32, 84)).astype("float32"), + 'Y': np.random.random((84, 100)).astype("float32") + } + self.outputs = {'Out': np.dot(self.inputs['X'], self.inputs['Y'])} +``` + 首先需要`import`必要的包,下面详细解释其他值： + + - `self.type = "mul" ` : 定义类型，和注册的类型一致。 + - `self.inputs` : 定义输入，类型为Numpy.array，并初始化。 + - `self.outputs` : 定义输出，并得到Python结算结果。 + + +### 反向Operator单测 + +反向Op单测继承自`GradientChecker`，而`GradientChecker`集成自`unittest.TestCase`，所以反向单测函数需要`test_`开头。 + + ``` + class MulGradOpTest(GradientChecker): + def test_mul(self): + op = create_op("mul") + inputs = { + 'X': np.random.random((32, 84)).astype("float32"), + 'Y': np.random.random((84, 100)).astype("float32") + } + self.compare_grad(op, inputs) + # mul op will enlarge the relative error + self.check_grad( + op, inputs, set(["X", "Y"]), "Out", max_relative_error=0.5) + ``` + + - 调用`create_op("mul")`创建反向Op对应的前向Op。 + - 定义输入`inputs`。 + - 调用`compare_grad`函数对比CPU、GPU计算结果。 + - 调用`check_grad`检查梯度稳定性。 diff --git a/doc/howto/index_cn.rst b/doc/howto/index_cn.rst index 26449a6365843b526b3ac3111b337d2f17524c9d..0608aa30968b0e8474eef330e4d2cc63c9def97d 100644 --- a/doc/howto/index_cn.rst +++ b/doc/howto/index_cn.rst @@ -19,6 +19,7 @@ .. toctree:: :maxdepth: 1 + dev/build_cn.rst dev/write_docs_cn.rst dev/contribute_to_paddle_cn.md diff --git a/doc/howto/index_en.rst b/doc/howto/index_en.rst index 1fbfcd260b912078f00ed5b720ed607db725c4e2..1b6034be4edffd2cbc822018b733b9a3836ea84a 100644 --- a/doc/howto/index_en.rst +++ b/doc/howto/index_en.rst @@ -18,6 +18,7 @@ Development .. toctree:: :maxdepth: 1 + dev/build_en.rst dev/new_layer_en.rst dev/contribute_to_paddle_en.md diff --git a/paddle/framework/backward.md b/paddle/framework/backward.md index 74c001b06a9e7b2279abf998604f2acf1b1168e4..c8fa3fefe5632a36d9044b4bccfd3dbb7c64dbf6 100644 --- a/paddle/framework/backward.md +++ b/paddle/framework/backward.md @@ -21,18 +21,32 @@ grad_op_builder(fengjiayi) given a forward network, it generates the backward network. We only care about the Gradients—`OutputGradients`,`InputGradients`. -1. bla bla bla (yuyang) +1. Op + + when the input forward network is a Op, return its gradient Operator Immediately. 2. NetOp - when the input forward network is a NetOp, it need to call the sub NetOp/Operators backward function recursively and ensure them done. During the process, we need to collect the `OutputGradients` name. + when the input forward network is a NetOp, it need to call the sub NetOp/Operators backward function recursively. During the process, we need to collect the `OutputGradients` name according to forward NetOp. + + **shared variable**. As illustrated in the pictures, two operator's `Output` `Gradient` will overwirte their shared input variable. + +

+
+ + 1. shared variable in two operators. + +

+ + Share variable between operators or same input variable used in multiple operators lead to a duplicate gradient variable. As demo show above, we need to rename gradient name recursively, and add a generic add operator replace the overwirte links. + +

+
- We share variable in the same scope, as a result, duplicate operator `OutputGradients` will overwirte then duplicate variable. + 2. replace shared variable gradient with `Add` Operator - ![./images/duplicate_op]() +

- Share variable between operators or same input variable used in multiple operators lead to a duplicate gradient variable. As demo show above, we need to rename gradient name recursively, and add a generic add operator instead. -![./images/duplicate_op2]() -​ Then collect the sub graph OutputGradients/InputGradients as the NetOp's and return it. +​ Then collect the sub graph `OutputGradients`/`InputGradients` as the NetOp's and return it. diff --git a/paddle/operators/CMakeLists.txt b/paddle/operators/CMakeLists.txt index f466dbc79a2059faa1e3d4ad6ede3f2394580842..f0fd12f1b5276d033ea086c60c80616fb1be7585 100644 --- a/paddle/operators/CMakeLists.txt +++ b/paddle/operators/CMakeLists.txt @@ -47,6 +47,7 @@ cc_test(gather_test SRCS gather_test.cc DEPS tensor) op_library(gather_op SRCS gather_op.cc gather_op.cu) cc_test(scatter_test SRCS scatter_test.cc DEPS tensor) +op_library(scatter_op SRCS scatter_op.cc scatter_op.cu) cc_library(net_op SRCS net_op.cc DEPS op_registry) cc_test(net_op_test SRCS net_op_test.cc DEPS net_op) diff --git a/paddle/operators/scatter_op.cc b/paddle/operators/scatter_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..f901edefa22dc9a252e87116df756d04767a7162 --- /dev/null +++ b/paddle/operators/scatter_op.cc @@ -0,0 +1,86 @@ +/* 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/scatter_op.h" +#include "paddle/framework/ddim.h" + +namespace paddle { +namespace operators { + +class ScatterOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(const framework::InferShapeContext &ctx) const override { + PADDLE_ENFORCE_EQ(ctx.Input("Index")->dims().size(), 1, + "Update Index should be 1-D."); + PADDLE_ENFORCE_EQ(ctx.Input("Ref")->dims().size(), + ctx.Input("Updates")->dims().size(), + "Reference and Updates should have the same shape size"); + PADDLE_ENFORCE_EQ(ctx.Input("Updates")->dims()[0], + ctx.Input("Index")->dims()[0], + "Updates and Index should have same batch-size."); + framework::DDim data_dim(ctx.Input("Updates")->dims()); + for (int i = 1; i < data_dim.size(); ++i) + PADDLE_ENFORCE_EQ(data_dim[i], ctx.Input("Updates")->dims()[i]); + ctx.Output("Out")->Resize(ctx.Input("Ref")->dims()); + } +}; + +class ScatterGradOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(const framework::InferShapeContext &ctx) const override { + auto *dUpdates = ctx.Output(framework::GradVarName("Updates")); + auto *Updates = ctx.Input("Updates"); + auto *dRef = ctx.Output(framework::GradVarName("Ref")); + auto *Ref = ctx.Input("Ref"); + + dRef->Resize(Ref->dims()); + dUpdates->Resize(Updates->dims()); + } +}; + +class ScatterOpMaker : public framework::OpProtoAndCheckerMaker { + public: + ScatterOpMaker(framework::OpProto *proto, + framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("Ref", "The source input of scatter op"); + AddInput("Index", + "The index input of scatter op where Ref will be updated"); + AddInput("Updates", "The updated value of updates op"); + AddOutput("Out", "The output of add op"); + AddComment(R"DOC( +Scatter Operator by selecting from the first axis, + +Out = Ref +Out[Index] = Ref[Index] + Updates +)DOC"); + } +}; +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP(scatter, ops::ScatterOp, ops::ScatterOpMaker, scatter_grad, + ops::ScatterGradOp); +REGISTER_OP_CPU_KERNEL(scatter, + ops::ScatterOpKernel); +REGISTER_OP_CPU_KERNEL( + scatter_grad, + ops::ScatterGradientOpKernel); diff --git a/paddle/operators/scatter_op.cu b/paddle/operators/scatter_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..6716b478833ff3adb6112cdb1ee25b7f1744ea1f --- /dev/null +++ b/paddle/operators/scatter_op.cu @@ -0,0 +1,20 @@ +/* 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. */ + +#define EIGEN_USE_GPU +#include "paddle/operators/scatter_op.h" + +namespace ops = paddle::operators; +REGISTER_OP_GPU_KERNEL(scatter, + ops::ScatterOpKernel); diff --git a/paddle/operators/scatter_op.h b/paddle/operators/scatter_op.h new file mode 100644 index 0000000000000000000000000000000000000000..e9595638a86a4a4536ddad4e6f20fd80a54b1608 --- /dev/null +++ b/paddle/operators/scatter_op.h @@ -0,0 +1,60 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include "gather.h" +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" +#include "scatter.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; + +template +class ScatterOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &ctx) const override { + auto *Ref = ctx.Input("Ref"); + auto *Index = ctx.Input("Index"); + auto *Updates = ctx.Input("Updates"); + auto *Out = ctx.Output("Out"); + + // In place output: Out = Ref, Out[Index] += Updates + Out->ShareDataWith(*Ref); + // Apply ScatterUpdate: Out[index] += Updates[:] + ScatterUpdate(ctx.GetPlace(), Updates, Index, Out); + } +}; + +template +class ScatterGradientOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &ctx) const override { + auto *dRef = ctx.Output(framework::GradVarName("Ref")); + auto *dUpdates = ctx.Output(framework::GradVarName("Updates")); + auto *Index = ctx.Input("Index"); + auto *dOut = ctx.Input(framework::GradVarName("Out")); + + // In place gradient: dRef = dO + dRef->ShareDataWith(*dOut); + dUpdates->mutable_data(ctx.GetPlace()); + // Gradient by Gather: dUpdates += dO[Index] + Gather(ctx.GetPlace(), dOut, Index, dUpdates); + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/pybind/CMakeLists.txt b/paddle/pybind/CMakeLists.txt index abb9c248eee9c59e8e6b9fa9d1878fec5dd67569..37e186a408ff5f560b5878e3e51ea81ca5810bc7 100644 --- a/paddle/pybind/CMakeLists.txt +++ b/paddle/pybind/CMakeLists.txt @@ -4,6 +4,7 @@ cc_library(paddle_pybind SHARED DEPS pybind python backward sgd_op gather_op + scatter_op add_op mul_op rowwise_add_op diff --git a/paddle/pybind/pybind.cc b/paddle/pybind/pybind.cc index 8fa8be2cef5fff04ed61ac726e5d8111e30c8a09..3bc150ccb7af2885439cc2344aa0db9ba3b1ca03 100644 --- a/paddle/pybind/pybind.cc +++ b/paddle/pybind/pybind.cc @@ -47,6 +47,7 @@ USE_OP(scale); USE_OP_ITSELF(identity); USE_OP(minus); USE_CPU_ONLY_OP(gather); +USE_CPU_ONLY_OP(scatter); namespace paddle { namespace framework { diff --git a/paddle/scripts/docker/build.sh b/paddle/scripts/docker/build.sh index 2941662f349baf57d1fe8188e88ce21d5de07750..17986420220fec173bbf3ecff240d4c504f8adbd 100644 --- a/paddle/scripts/docker/build.sh +++ b/paddle/scripts/docker/build.sh @@ -38,7 +38,7 @@ Configuring cmake in /paddle/build ... -DWITH_SWIG_PY=${WITH_SWIG_PY:-ON} -DCUDNN_ROOT=/usr/ -DWITH_STYLE_CHECK=${WITH_STYLE_CHECK:-OFF} - -DWITH_TESTING=${WITH_TESTING:-OFF} + -DWITH_TESTING=${WITH_TESTING:-ON} -DCMAKE_EXPORT_COMPILE_COMMANDS=ON ======================================== EOF @@ -56,19 +56,18 @@ cmake .. \ -DWITH_C_API=${WITH_C_API:-OFF} \ -DWITH_PYTHON=${WITH_PYTHON:-ON} \ -DCUDNN_ROOT=/usr/ \ - -DWITH_STYLE_CHECK=${WITH_STYLE_CHECK:-OFF} \ - -DWITH_TESTING=${WITH_TESTING:-OFF} \ + -DWITH_STYLE_CHECK=${WITH_STYLE_CHECK:-ON} \ + -DWITH_TESTING=${WITH_TESTING:-ON} \ -DCMAKE_EXPORT_COMPILE_COMMANDS=ON cat <