dot_op.cc 6.0 KB
Newer Older
L
liuwei1031 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160
// Copyright (c) 2020 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/dot_op.h"

namespace paddle {
namespace operators {

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

  void InferShape(framework::InferShapeContext* ctx) const override {
    PADDLE_ENFORCE_EQ(true, ctx->HasInput("X"),
                      platform::errors::PreconditionNotMet(
                          "Input(X) of DotOp should not be null."));
    PADDLE_ENFORCE_EQ(true, ctx->HasInput("Y"),
                      platform::errors::PreconditionNotMet(
                          "Input(Y) of DotOp should not be null."));
    PADDLE_ENFORCE_EQ(true, ctx->HasOutput("Out"),
                      platform::errors::PreconditionNotMet(
                          "Output(Out) of DotOp should not be null."));

    auto x_dims = ctx->GetInputDim("X");
    auto x_rank = (size_t)x_dims.size();
    PADDLE_ENFORCE_EQ(true, 1 == x_rank || 2 == x_rank,
                      platform::errors::PreconditionNotMet(
                          "ShapeError: The dimensions of input tensor X (%s) "
                          "should be 1 or 2",
                          x_dims.to_str()));

    auto y_dims = ctx->GetInputDim("Y");
    PADDLE_ENFORCE_EQ(
        true, x_rank == (size_t)y_dims.size(),
        platform::errors::PreconditionNotMet(
            "ShapeError: The shape of input tensor Y: %s should match with "
            "input tenosr X: %s",
            y_dims.to_str(), x_dims.to_str()));
    bool shape_match = true;
    for (size_t i = 0; i < x_rank; ++i) {
      if (x_dims[i] != y_dims[i]) {
        shape_match = false;
        break;
      }
    }

    PADDLE_ENFORCE_EQ(true, shape_match,
                      platform::errors::PreconditionNotMet(
                          "ShapeError: The shape of input tensor X: %s should "
                          "be exactly the same "
                          "with input tensor Y: %s",
                          x_dims.to_str(), y_dims.to_str()));
    auto dims = vectorize(x_dims);
    dims[dims.size() - 1] = 1;
    ctx->SetOutputDim("Out", framework::make_ddim(dims));
  }

  framework::OpKernelType GetExpectedKernelType(
      const framework::ExecutionContext& ctx) const override {
    return framework::OpKernelType(
        OperatorWithKernel::IndicateVarDataType(ctx, "X"), ctx.GetPlace());
  }
};

class DotOpMaker : public framework::OpProtoAndCheckerMaker {
 public:
  void Make() final {
    AddInput("X", "(Tensor) The first input tensor. ");
    AddInput("Y", "(Tensor) The second input tensor. ");
    AddOutput("Out", "(Tensor) The result tensor.");
    AddComment("");
  }
};

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

  void InferShape(framework::InferShapeContext* ctx) const override {
    PADDLE_ENFORCE_EQ(
        true, ctx->HasInput("X"),
        platform::errors::PreconditionNotMet("Input(X) should not be null."));
    PADDLE_ENFORCE_EQ(
        true, ctx->HasInput("Y"),
        platform::errors::PreconditionNotMet("Input(Y) should not be null."));
    PADDLE_ENFORCE_EQ(true, ctx->HasInput(framework::GradVarName("Out")),
                      platform::errors::PreconditionNotMet(
                          "Input(Out@GRAD) should not be null."));

    auto x_grad_name = framework::GradVarName("X");
    auto y_grad_name = framework::GradVarName("Y");
    if (ctx->HasOutput(x_grad_name)) {
      ctx->ShareDim("X", /*->*/ x_grad_name);
      ctx->ShareLoD("X", /*->*/ x_grad_name);
    }
    if (ctx->HasOutput(y_grad_name)) {
      ctx->ShareDim("Y", /*->*/ y_grad_name);
      ctx->ShareLoD("Y", /*->*/ y_grad_name);
    }
  }

 protected:
  framework::OpKernelType GetExpectedKernelType(
      const framework::ExecutionContext& ctx) const override {
    return framework::OpKernelType(OperatorWithKernel::IndicateVarDataType(
                                       ctx, framework::GradVarName("Out")),
                                   ctx.GetPlace());
  }
};

template <typename T>
class DotOpGradMaker : public framework::SingleGradOpMaker<T> {
 public:
  using framework::SingleGradOpMaker<T>::SingleGradOpMaker;

 protected:
  void Apply(GradOpPtr<T> op) const override {
    op->SetType("dot_grad");

    op->SetInput("X", this->Input("X"));
    op->SetInput("Y", this->Input("Y"));
    op->SetInput(framework::GradVarName("Out"), this->OutputGrad("Out"));
    op->SetAttrMap(this->Attrs());
    op->SetOutput(framework::GradVarName("X"), this->InputGrad("X"));
    op->SetOutput(framework::GradVarName("Y"), this->InputGrad("Y"));
  }
};

}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;

REGISTER_OPERATOR(dot, ops::DotOp, ops::DotOpMaker,
                  ops::DotOpGradMaker<paddle::framework::OpDesc>,
                  ops::DotOpGradMaker<paddle::imperative::OpBase>);

REGISTER_OPERATOR(dot_grad, ops::DotGradOp);

REGISTER_OP_CPU_KERNEL(
    dot, ops::DotKernel<paddle::platform::CPUDeviceContext, float>,
    ops::DotKernel<paddle::platform::CPUDeviceContext, double>,
    ops::DotKernel<paddle::platform::CPUDeviceContext, int>,
    ops::DotKernel<paddle::platform::CPUDeviceContext, int64_t>);
REGISTER_OP_CPU_KERNEL(
    dot_grad, ops::DotGradKernel<paddle::platform::CPUDeviceContext, float>,
    ops::DotGradKernel<paddle::platform::CPUDeviceContext, double>,
    ops::DotGradKernel<paddle::platform::CPUDeviceContext, int>,
    ops::DotGradKernel<paddle::platform::CPUDeviceContext, int64_t>);