add_position_encoding_op.cc 4.6 KB
Newer Older
G
gmcather 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */

#include "paddle/fluid/operators/add_position_encoding_op.h"
16

S
sneaxiy 已提交
17
#include <memory>
G
gmcather 已提交
18 19 20 21 22 23 24 25 26

namespace paddle {
namespace operators {

class AddPositionEncodingOp : public framework::OperatorWithKernel {
 public:
  using framework::OperatorWithKernel::OperatorWithKernel;

  void InferShape(framework::InferShapeContext* ctx) const override {
27
    OP_INOUT_CHECK(ctx->HasInput("X"), "Input", "X", "AddPositionEncoding");
28 29
    OP_INOUT_CHECK(
        ctx->HasOutput("Out"), "Output", "Out", "AddPositionEncoding");
G
gmcather 已提交
30 31 32 33 34

    auto x_dims = ctx->GetInputDim("X");
    ctx->SetOutputDim("Out", x_dims);
    ctx->ShareLoD("X", /*->*/ "Out");
  }
35 36 37 38

 protected:
  framework::OpKernelType GetExpectedKernelType(
      const framework::ExecutionContext& ctx) const override {
39 40 41
    return framework::OpKernelType(
        OperatorWithKernel::IndicateVarDataType(ctx, "X"),
        platform::CPUPlace());
42
  }
G
gmcather 已提交
43 44 45 46 47 48 49 50
};

class AddPositionEncodingOpGrad : public framework::OperatorWithKernel {
 public:
  using framework::OperatorWithKernel::OperatorWithKernel;

  void InferShape(framework::InferShapeContext* ctx) const override {
    if (ctx->HasOutput(framework::GradVarName("X"))) {
S
sneaxiy 已提交
51
      auto out_dims = ctx->GetInputDim(framework::GradVarName("Out"));
G
gmcather 已提交
52 53 54
      ctx->SetOutputDim(framework::GradVarName("X"), out_dims);
    }
  }
55 56 57 58

 protected:
  framework::OpKernelType GetExpectedKernelType(
      const framework::ExecutionContext& ctx) const override {
59 60 61
    return framework::OpKernelType(OperatorWithKernel::IndicateVarDataType(
                                       ctx, framework::GradVarName("Out")),
                                   platform::CPUPlace());
62
  }
G
gmcather 已提交
63 64 65 66 67 68 69 70 71 72
};

class AddPositionEncodingOpMaker : public framework::OpProtoAndCheckerMaker {
 public:
  void Make() override {
    AddInput("X", "Input of AddPositionEncoding operator");
    AddOutput("Out", "Output of AddPositionEncoding operator");
    AddAttr<float>("alpha", "The scale of Original Embedding.")
        .SetDefault(1.0f)
        .AddCustomChecker([](const float& alpha) {
73
          PADDLE_ENFORCE_GE(
74 75
              alpha,
              0.0f,
76 77
              platform::errors::InvalidArgument(
                  "Attribute 'alpha' must be greater than or equal to 0.0."));
G
gmcather 已提交
78 79 80 81
        });
    AddAttr<float>("beta", "The scale of Position Embedding.")
        .SetDefault(1.0f)
        .AddCustomChecker([](const float& beta) {
82
          PADDLE_ENFORCE_GE(
83 84
              beta,
              0.0f,
85 86
              platform::errors::InvalidArgument(
                  "Attribute 'beta' must be greater than or equal to 0.0."));
G
gmcather 已提交
87 88 89 90 91 92 93 94 95 96
        });
    AddComment(R"DOC(
    Add Position Encoding Operator.
    
    The add position encoding calculates the output based on the input, alpha, beta.
    The size of each dimension of the parameters checked in the infer-shape.
  )DOC");
  }
};

H
hong 已提交
97 98
template <typename T>
class AddPositionEncodingGradOpMaker : public framework::SingleGradOpMaker<T> {
S
sneaxiy 已提交
99
 public:
H
hong 已提交
100
  using framework::SingleGradOpMaker<T>::SingleGradOpMaker;
S
sneaxiy 已提交
101 102

 protected:
103
  void Apply(GradOpPtr<T> op) const override {
S
sneaxiy 已提交
104
    op->SetType("add_position_encoding_grad");
H
hong 已提交
105 106 107
    op->SetInput(framework::GradVarName("Out"), this->OutputGrad("Out"));
    op->SetOutput(framework::GradVarName("X"), this->InputGrad("X"));
    op->SetAttrMap(this->Attrs());
S
sneaxiy 已提交
108 109 110
  }
};

G
gmcather 已提交
111 112 113 114 115 116
}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;
namespace plt = paddle::platform;

H
hong 已提交
117
REGISTER_OPERATOR(
118 119
    add_position_encoding,
    ops::AddPositionEncodingOp,
H
hong 已提交
120 121 122
    ops::AddPositionEncodingOpMaker,
    ops::AddPositionEncodingGradOpMaker<paddle::framework::OpDesc>,
    ops::AddPositionEncodingGradOpMaker<paddle::imperative::OpBase>);
G
gmcather 已提交
123 124 125 126 127 128 129 130 131 132 133
REGISTER_OPERATOR(add_position_encoding_grad, ops::AddPositionEncodingOpGrad);

REGISTER_OP_CPU_KERNEL(
    add_position_encoding,
    ops::AddPositionEncodingKernel<plt::CPUDeviceContext, float>,
    ops::AddPositionEncodingKernel<plt::CPUDeviceContext, double>);

REGISTER_OP_CPU_KERNEL(
    add_position_encoding_grad,
    ops::AddPositionEncodingGradKernel<plt::CPUDeviceContext, float>,
    ops::AddPositionEncodingGradKernel<plt::CPUDeviceContext, double>);