bpr_loss_op.cc 6.2 KB
Newer Older
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
/* Copyright (c) 2016 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/bpr_loss_op.h"

namespace paddle {
namespace operators {

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

  void InferShape(framework::InferShapeContext* ctx) const override {
    PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should be not null.");
26 27
    PADDLE_ENFORCE(ctx->HasInput("LabelPos"),
                   "Input(LabelPos) should be not null.");
28 29 30
    PADDLE_ENFORCE(ctx->HasOutput("Y"), "Output(Y) should be not null.");

    auto x_dims = ctx->GetInputDim("X");
31
    auto label_Pos_dims = ctx->GetInputDim("LabelPos");
32
    int rank = x_dims.size();
33 34
    PADDLE_ENFORCE_EQ(rank, label_Pos_dims.size(),
                      "Input(X) and Input(LabelPos) shall have the same rank.");
35 36
    PADDLE_ENFORCE_EQ(framework::slice_ddim(x_dims, 0, rank - 1),
                      framework::slice_ddim(label_Pos_dims, 0, rank - 1),
37
                      "Input(X) and Input(LabelPos) shall have the same shape "
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
                      "except the last dimension.");

    auto y_dims = x_dims;
    y_dims[rank - 1] = 1;
    ctx->SetOutputDim("Y", y_dims);
    ctx->ShareLoD("X", /*->*/ "Y");
  }

 protected:
  // Explicitly set that the data type of computation kernel of Seq-bpr
  // is determined by its input "X".
  framework::OpKernelType GetExpectedKernelType(
      const framework::ExecutionContext& ctx) const override {
    return framework::OpKernelType(
        framework::ToDataType(ctx.Input<Tensor>("X")->type()),
        platform::CPUPlace());
  }
};

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

  void InferShape(framework::InferShapeContext* ctx) const override {
    PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should be not null.");
63 64
    PADDLE_ENFORCE(ctx->HasInput("LabelPos"),
                   "Input(LabelPos) should be not null.");
65 66 67 68 69 70
    PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Y")),
                   "Input(Y@GRAD) shoudl be not null.");
    PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("X")),
                   "Output(X@GRAD) should be not null.");

    auto x_dims = ctx->GetInputDim("X");
71
    auto label_pos_dims = ctx->GetInputDim("LabelPos");
72 73 74 75 76 77
    auto dy_dims = ctx->GetInputDim(framework::GradVarName("Y"));
    int rank = x_dims.size();
    PADDLE_ENFORCE_EQ(dy_dims.size(), rank,
                      "Input(Y@Grad) and Input(X) should have the same rank.");
    PADDLE_ENFORCE_EQ(
        label_pos_dims.size(), rank,
78
        "Input(LabelPos) and Input(X) should have the same rank.");
79 80
    PADDLE_ENFORCE_EQ(framework::slice_ddim(x_dims, 0, rank - 1),
                      framework::slice_ddim(label_pos_dims, 0, rank - 1),
81
                      "The Input(X) and Input(LabelPos) should have the same "
82 83 84 85 86 87 88 89
                      "shape except the last dimension.");
    PADDLE_ENFORCE_EQ(framework::slice_ddim(x_dims, 0, rank - 1),
                      framework::slice_ddim(dy_dims, 0, rank - 1),
                      "The Input(X) and Input(Y@Grad) should have the same "
                      "shape except the last dimension.");
    PADDLE_ENFORCE_EQ(dy_dims[rank - 1], 1,
                      "The last dimension of Input(Y@Grad) should be 1.");
    PADDLE_ENFORCE_EQ(label_pos_dims[rank - 1], 1,
90
                      " the last dimension of Input(LabelPos) should be 1.");
91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113
    ctx->SetOutputDim(framework::GradVarName("X"), x_dims);
    ctx->ShareLoD("X", framework::GradVarName("X"));
  }

 protected:
  // Explicitly set that the data type of computation kernel of cross_entropy
  // is determined by its input "X".
  framework::OpKernelType GetExpectedKernelType(
      const framework::ExecutionContext& ctx) const override {
    return framework::OpKernelType(
        framework::ToDataType(ctx.Input<Tensor>("X")->type()),
        platform::CPUPlace());
  }
};

class BprLossOpMaker : public framework::OpProtoAndCheckerMaker {
 public:
  void Make() override {
    AddInput("X",
             "(Tensor, default Tensor<float>), a tensor whose last dimension "
             "size is equal to the number of classes. This input is a "
             "real number.");
    AddInput(
114
        "LabelPos",
115 116 117 118 119 120 121 122
        "(Tensor), the tensor which represents the ground truth. It has the "
        "same shape with 'X' except the last dimension. the last dimension "
        "size is 1.");
    AddOutput("Y",
              "(Tensor, default Tensor<float>), a tensor whose shape is same "
              "with 'X' except that the last dimension size is 1. It "
              "represents the sequence bpr loss.");
    AddComment(R"DOC(
123
Bayesian Personalized Ranking Loss Operator.
124

125 126
This operator belongs to pairwise ranking loss. LabelPos is the desired item.
The loss at a given point in one session is defined as:
127 128 129
$Y[i] = -\frac{1}{N_{i}} * \sum_{j=0}^{N_{i}}\log(\sigma(X[i, Label[i]]-X[i, j]))$

Learn more details by reading paper <session-based recommendations with recurrent
130
neural networks>(https://arxiv.org/abs/1511.06939)
131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148

)DOC");
  }
};
}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;
using CPUCtx = paddle::platform::CPUDeviceContext;

REGISTER_OPERATOR(bpr_loss, ops::BprLossOp, ops::BprLossOpMaker,
                  paddle::framework::DefaultGradOpDescMaker<true>);
REGISTER_OPERATOR(bpr_loss_grad, ops::BprLossGradientOp);
REGISTER_OP_CPU_KERNEL(bpr_loss, ops::BprLossOpKernel<CPUCtx, float>,
                       ops::BprLossOpKernel<CPUCtx, double>);
REGISTER_OP_CPU_KERNEL(bpr_loss_grad,
                       ops::BprLossGradientOpKernel<CPUCtx, float>,
                       ops::BprLossGradientOpKernel<CPUCtx, double>);