From a828c66fbaae491e568b1fe1d7b39e37a7a0bf86 Mon Sep 17 00:00:00 2001 From: chengduozh Date: Sun, 14 Apr 2019 18:00:00 +0800 Subject: [PATCH] update new_op.cm --- .../development/new_op/new_op.md | 264 ++++++++++++------ 1 file changed, 172 insertions(+), 92 deletions(-) diff --git a/doc/fluid/advanced_usage/development/new_op/new_op.md b/doc/fluid/advanced_usage/development/new_op/new_op.md index b1d89954e..7eabd3f50 100644 --- a/doc/fluid/advanced_usage/development/new_op/new_op.md +++ b/doc/fluid/advanced_usage/development/new_op/new_op.md @@ -7,9 +7,14 @@ - `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 +- `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,可以将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`OperatorWithKernel`,后者继承自`OperatorBase`。本教程主要介绍带Kernel的Op如何写,简单总结Op需要包含的内容如下: +根据是否包含Kernel,可以将Op分为两种:包含Kernel的Op和不包含kernel的Op: + +- 包含Kernel的Op继承自`OperatorWithKernel`,这类Op的功能实现与输入的数据类型、数据布局、数据所在的设备以及Op实现所调用第三方库等有关。比如ConvOp,如果使用CPU计算,一般通过调用mkl库中的矩阵乘操作实现,如果使用GPU计算,一般通过调用cublas库中的矩阵乘操作实现,或者直接调用cudnn库中的卷积操作。 +- 不包含Kernel的Op继承自`OperatorBase`,因为这类Op的功能实现与设备以及输入的数据不相关。比如WhileOp、IfElseOp等。 + +本教程主要介绍带Kernel的Op如何写,简单总结Op需要包含的内容如下: @@ -21,7 +26,7 @@ - + @@ -38,16 +43,11 @@
OpProtoMake定义 .cc 文件,Backward Op不需要定义OpProtoMake .cc 文件
Op定义
- -实现新的op都添加至目录[paddle/fluid/operators](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/fluid/operators)下,文件命名以`*_op.h`(如有) 、 `*_op.cc` 、`*_op.cu`(如有)结尾。**系统会根据文件名自动构建op和其对应的Python扩展。** - +实现新的op都添加至目录[paddle/fluid/operators](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/fluid/operators)下,文件命名以`*_op.h`(如有)、`*_op.cc` 、`*_op.cu`(如有)结尾。**系统会根据文件名自动构建op和其对应的Python扩展。** 下面以矩阵乘操作,即[MulOp](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/operators/mul_op.cc)为例来介绍如何写带Kernel的Operator。 - ## 实现C++类 - - ### 定义ProtoMaker类 矩阵乘法的公式:$Out = X * Y$, 可见该计算由两个输入,一个输出组成。 @@ -57,73 +57,94 @@ ```cpp class MulOpMaker : public framework::OpProtoAndCheckerMaker { public: - MulOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("X", "(Tensor), 2D tensor of size (M x K)"); - AddInput("Y", "(Tensor), 2D tensor of size (K x N)"); - AddOutput("Out", "(Tensor), 2D tensor of size (M x N)"); + void Make() override { + AddInput("X", "(Tensor), The first input tensor of mul op."); + AddInput("Y", "(Tensor), The second input tensor of mul op."); + AddOutput("Out", "(Tensor), The output tensor of mul op."); + AddAttr( + "x_num_col_dims", + R"DOC((int, default 1), The mul_op can take tensors with more than two + dimensions as its inputs. If the input $X$ is a tensor with more + than two dimensions, $X$ will be flattened into a two-dimensional + matrix first. The flattening rule is: the first `num_col_dims` + will be flattened to form the first dimension of the final matrix + (the height of the matrix), and the rest `rank(X) - num_col_dims` + dimensions are flattened to form the second dimension of the final + matrix (the width of the matrix). As a result, height of the + flattened matrix is equal to the product of $X$'s first + `x_num_col_dims` dimensions' sizes, and width of the flattened + matrix is equal to the product of $X$'s last `rank(x) - num_col_dims` + dimensions' size. For example, suppose $X$ is a 6-dimensional + tensor with the shape [2, 3, 4, 5, 6], and `x_num_col_dims` = 3. + Thus, the flattened matrix will have a shape [2 x 3 x 4, 5 x 6] = + [24, 30]. + )DOC") + .SetDefault(1) + .EqualGreaterThan(1); + AddAttr( + "y_num_col_dims", + R"DOC((int, default 1), The mul_op can take tensors with more than two, + dimensions as its inputs. If the input $Y$ is a tensor with more + than two dimensions, $Y$ will be flattened into a two-dimensional + matrix first. The attribute `y_num_col_dims` determines how $Y$ is + flattened. See comments of `x_num_col_dims` for more details. + )DOC") + .SetDefault(1) + .EqualGreaterThan(1); AddComment(R"DOC( -Two Element Mul Operator. -The equation is: Out = X * Y -)DOC"); - } -}; -``` +Mul Operator. -[`MulOpMaker`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/operators/mul_op.cc#L76-L127)继承自`framework::OpProtoAndCheckerMaker`,构造函数含有2个参数: +This operator is used to perform matrix multiplication for input $X$ and $Y$. - - `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Op的注释。这些函数会将对应内容添加到`OpProto`中。 - -上面的代码在`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,命名请遵守[命名规范](https://github.com/PaddlePaddle/FluidDoc/blob/release/1.2/doc/fluid/dev/name_convention.md)。 +The equation is: +$$Out = X * Y$$ -再以[`ScaleOp`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/operators/scale_op.cc#L38-L55)为例: +Both the input $X$ and $Y$ can carry the LoD (Level of Details) information, +or not. But the output only shares the LoD information with input $X$. -```cpp -template -class ScaleOpMaker : public framework::OpProtoAndCheckerMaker { - public: - ScaleOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("X", "(Tensor) Input tensor of scale operator."); - AddOutput("Out", "(Tensor) Output tensor of scale operator."); - AddComment(R"DOC( -Scale operator -$$Out = scale*X$$ )DOC"); - AddAttr("scale", - "(float, default 1.0)" - "The scaling factor of the scale operator.") - .SetDefault(1.0); } }; ``` -这个例子有`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。 +[`MulOpMaker`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/operators/mul_op.cc)继承自`framework::OpProtoAndCheckerMaker`。 + +开发者通过覆盖`framework::OpProtoAndCheckerMaker`中的`Make`函数来定义Op所对应的Proto,通过`AddInput`添加输入参数,通过`AddOutput`添加输出参数,通过`AddAttr`添加属性参数,通过`AddComment`添加Op的注释。这些函数会将对应内容添加到`OpProto`中。 + +上面的代码在`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,命名请遵守[命名规范](https://github.com/PaddlePaddle/FluidDoc/blob/release/1.2/doc/fluid/dev/name_convention.md)。 ### 定义GradProtoMaker类 -每个Op的必须有一个对应的GradProtoMaker,若未定制对应前向Op的GradProtoMaker,fluid提供了DefaultGradProtoMaker,默认注册会使用全部输入输出,包括Input, Output, Output@Grad等,使用不需要的变量的会造成显存浪费。 -下面示例定义了ScaleOp的GradProtoMaker。 +通常情况下,每个Op的会有一个对应的GradProtoMaker,为方便代码编写,fluid提供了默认的GradProtoMaker,即:`DefaultGradProtoMaker`。`DefaultGradProtoMaker`会使用前向Op的全部输入输出以及输出变量所对应的梯度(`Output@Grad`)作为反向Op的输入,将前向Op的输入变量所对应的的梯度(`Input@Grad`)作为输出。**注意:不要将反向Op不会用到的变量放到反向Op的输入列表中,这样会导致这些不会被反向Op用到的变量的空间不能够及时回收,进而有可能导致用到该Op的模型可以设置的最大batch_size较低。** + +下面示例定义了`MulOp`的GradProtoMaker。 ```cpp -class ScaleGradMaker : public framework::SingleGradOpDescMaker { +class MulOpGradMaker : public framework::SingleGradOpDescMaker { public: using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; + protected: std::unique_ptr Apply() const override { - auto *grad_op = new framework::OpDesc(); - grad_op->SetType("scale"); - grad_op->SetInput("X", OutputGrad("Out")); - grad_op->SetOutput("Out", InputGrad("X")); - grad_op->SetAttr("scale", GetAttr("scale")); - return std::unique_ptr(grad_op); + std::unique_ptr retv(new framework::OpDesc()); + retv->SetType("mul_grad"); + retv->SetInput("X", Input("X")); + retv->SetInput("Y", Input("Y")); + retv->SetInput(framework::GradVarName("Out"), OutputGrad("Out")); + retv->SetOutput(framework::GradVarName("X"), InputGrad("X")); + retv->SetOutput(framework::GradVarName("Y"), InputGrad("Y")); + retv->SetAttrMap(Attrs()); + return retv; } }; ``` +**注意:** + +- 有些Op的前向逻辑和反向逻辑是一样的,比如[`ScaleOp`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/operators/scale_op.cc).这种情况下,前向Op和反向Op的Kernel可以为同一个。 +- 有些前向Op所对应的反向Op可能有多个,比如[`SumOp`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/operators/sum_op.cc),这种情况下,`GradMaker`需要继承`framework::GradOpDescMakerBase`。 +- 有些Op的反向对应另一个Op的前向,比如[`SplitOp`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/operators/split_op.h),这种情况下,[`SplitGradMaker`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/operators/split_op.h#L52)中定义的`SplitOp`反向Op的Type就是`concat`, + ### 定义Operator类 下面实现了MulOp的定义: @@ -134,20 +155,53 @@ class MulOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - //never use Input or Output if you want a to get a LoDTensor. - 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]}); + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) of MulOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Y"), "Input(Y) of MulOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of MulOp should not be null."); + + auto x_dims = ctx->GetInputDim("X"); + auto y_dims = ctx->GetInputDim("Y"); + + int x_num_col_dims = ctx->Attrs().Get("x_num_col_dims"); + int y_num_col_dims = ctx->Attrs().Get("y_num_col_dims"); + + VLOG(3) << "mul operator x.shape=" << x_dims << " y.shape=" << y_dims + << " x_num_col_dims=" << x_num_col_dims + << " y_num_col_dims=" << y_num_col_dims; + + PADDLE_ENFORCE_GT( + x_dims.size(), x_num_col_dims, + "The input tensor X's rank of MulOp should be larger than " + "x_num_col_dims."); + PADDLE_ENFORCE_GT( + y_dims.size(), y_num_col_dims, + "The input tensor Y's rank of MulOp should be larger than " + "y_num_col_dims: %ld vs %ld", + y_dims.size(), y_num_col_dims); + + auto x_mat_dims = framework::flatten_to_2d(x_dims, x_num_col_dims); + auto y_mat_dims = framework::flatten_to_2d(y_dims, y_num_col_dims); + + PADDLE_ENFORCE_EQ(x_mat_dims[1], y_mat_dims[0], + "First matrix's width must be equal with second matrix's " + "height. %s, %s", + x_mat_dims[1], y_mat_dims[0]); + std::vector output_dims; + output_dims.reserve( + static_cast(x_num_col_dims + y_dims.size() - y_num_col_dims)); + + for (int i = 0; i < x_num_col_dims; ++i) { + output_dims.push_back(x_dims[i]); + } + + for (int i = y_num_col_dims; i < y_dims.size(); ++i) { + output_dims.push_back(y_dims[i]); + } + + ctx->SetOutputDim("Out", framework::make_ddim(output_dims)); + ctx->ShareLoD("X", /*->*/ "Out"); } }; ``` @@ -167,10 +221,10 @@ MulOp(const std::string &type, const framework::VariableNameMap &inputs, : 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: +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`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: - 做检查, 尽早报错:检查输入数据维度、类型等是否合法。 - - 设置输出Tensor的形状。 + - 设置输出Tensor的形状以及LoD信息。 通常`OpProtoMaker`和`Op`类的定义写在`.cc`文件中,和下面将要介绍的注册函数一起放在`.cc`中 @@ -178,7 +232,7 @@ MulOp(const std::string &type, const framework::VariableNameMap &inputs, `MulKernel`继承自`framework::OpKernel`,带有下面两个模板参数: -- `typename DeviceContext`: 表示设备类型,不同设备(CPU、CUDA)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/fluid/operators/cross_entropy_op.h#L43)。 +- `typename DeviceContext`: 表示设备类型,不同设备(CPU、CUDA)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[`SGDOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/operators/optimizers/sgd_op.h)。 - `typename T` : 表示数据类型,如`float`, `double`等。 @@ -192,33 +246,53 @@ MulOp(const std::string &type, const framework::VariableNameMap &inputs, Op的输入和输出可分别通过`ExecutionContext::Input()`和`ExecutionContext::Output()`获得。 -**注意:** 若op的输入/输出的变量类型是`LoDTensor`(fluid默认所有的Tensor默认都是LoDTensor类型),请写成`ExecutionContext::Input()`和`ExecutionContext::Output()`,不要写`ExecutionContext::Input()`和`ExecutionContext::Output()`。因为若实际的变量类型为`SelectedRows`,`Input()`和`Output()`方法会将`SelectedRows`类型特化为`Tensor`,导致潜在的错误。 +**注意:** 若op的输入/输出的变量类型是`LoDTensor`(fluid默认所有的`Tensor`默认都是`LoDTensor`类型),请写成`ExecutionContext::Input()`和`ExecutionContext::Output()`,不要写`ExecutionContext::Input()`和`ExecutionContext::Output()`。因为若实际的变量类型为`SelectedRows`,`Input()`和`Output()`方法会将`SelectedRows`类型特化为`Tensor`,导致潜在的错误。 下面是 `MulKernel` `Compute`的实现: - ```cpp - template - class MulKernel : public framework::OpKernel { - public: +```cpp +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 = context.template device_context(); - math::matmul(*X, false, *Y, false, 1, Z, 0, device_context); + const Tensor* x = context.Input("X"); + const Tensor* y = context.Input("Y"); + Tensor* z = context.Output("Out"); + const Tensor x_matrix = + x->dims().size() > 2 + ? framework::ReshapeToMatrix( + *x, context.template Attr("x_num_col_dims")) + : *x; + const Tensor y_matrix = + y->dims().size() > 2 + ? framework::ReshapeToMatrix( + *y, context.template Attr("y_num_col_dims")) + : *y; + + z->mutable_data(context.GetPlace()); + auto z_dim = z->dims(); + if (z_dim.size() != 2) { + z->Resize({x_matrix.dims()[0], y_matrix.dims()[1]}); + } + + auto blas = math::GetBlas(context); + + blas.MatMul(x_matrix, y_matrix, z); + if (z_dim.size() != 2) { + z->Resize(z_dim); + } } - }; - ``` +}; +``` 需要注意:**不同设备(CPU、CUDA)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、CUDA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/fluid/operators/cross_entropy_op.h#L43)。 +`MulOp`的CPU、CUDA实现共享同一个`Kernel`。`OpKernel`不共享的例子可以参考:[`SGDOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/operators/optimizers/sgd_op.h)。 为了使`OpKernel`的计算过程书写更加简单,并且CPU、CUDA的代码可以复用,我们通常借助 Eigen unsupported Tensor模块来实现`Compute`接口。关于在PaddlePaddle中如何使用Eigen库,请参考[使用文档](https://github.com/PaddlePaddle/FluidDoc/blob/release/1.2/doc/fluid/dev/use_eigen_cn.md)。 到此,前向Op实现完成。接下来,需要在`.cc`文件中注册该op和kernel。 -反向Op类的定义,反向OpKernel的定义与前向Op类似,这里不再赘述。**但需注意反向Op没有`ProtoMaker`**。 +反向Op类的定义,反向OpKernel的定义与前向Op类似,这里不再赘述。 ### 注册Operator @@ -227,11 +301,14 @@ Op的输入和输出可分别通过`ExecutionContext::Input()`和`ExecutionCo ```cpp namespace ops = paddle::operators; REGISTER_OPERATOR(mul, ops::MulOp, ops::MulOpMaker, - paddle::framework::DefaultGradOpDescMaker) + ops::MulOpGradMaker) REGISTER_OPERATOR(mul_grad, ops::MulGradOp) - REGISTER_OP_CPU_KERNEL(mul, ops::MulKernel); + REGISTER_OP_CPU_KERNEL(mul, + ops::MulKernel, + ops::MulKernel); REGISTER_OP_CPU_KERNEL(mul_grad, - ops::MulGradKernel); + ops::MulGradKernel, + ops::MulGradKernel); ``` 在上面的代码中: @@ -250,9 +327,12 @@ Op的输入和输出可分别通过`ExecutionContext::Input()`和`ExecutionCo #define EIGEN_USE_GPU namespace ops = paddle::operators; - REGISTER_OP_CUDA_KERNEL(mul, ops::MulKernel); + REGISTER_OP_CUDA_KERNEL(mul, + ops::MulKernel, + ops::MulKernel); REGISTER_OP_CUDA_KERNEL(mul_grad, - ops::MulGradKernel); + ops::MulGradKernel, + ops::MulGradKernel); ``` ### 编译 @@ -273,7 +353,7 @@ make mul_op ### 前向Operator单测 -Op单元测试继承自`OpTest`。各项更加具体的单元测试在`TestMulOp`里完成。测试Operator,需要: +Op单元测试继承自`OpTest`。各项具体的单元测试在`TestMulOp`里完成。测试Operator,需要: 1. 在`setUp`函数定义输入、输出,以及相关的属性参数。 2. 生成随机的输入数据。 @@ -365,7 +445,7 @@ PADDLE_ENFORCE_EQ(比较对象A, 比较对象B, 错误提示信息) #### 总体原则 -任何使用了PADDLE_ENFORCE与PADDLE_ENFORCE_**检查的地方,必须有详略得当的备注解释!**错误提示信息**不能为空! +任何使用了PADDLE_ENFORCE与PADDLE_ENFORCE_**检查的地方,必须有详略得当的备注解释!错误提示信息**不能为空! #### 提示信息书写标准 -- GitLab