hierarchical_sigmoid_op.cc 6.4 KB
Newer Older
Y
Yancey1989 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14
/* 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. */

W
weixing02 已提交
15 16
#include "paddle/fluid/operators/hierarchical_sigmoid_op.h"
#include <vector>
Y
Yancey1989 已提交
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

namespace paddle {
namespace operators {

/**
 * Organize the classes into a binary tree. At each node, a sigmoid function
 * is used to calculate the probability of belonging to the right branch.
 * This idea is from "F. Morin, Y. Bengio (AISTATS 05):
 * Hierarchical Probabilistic Neural Network Language Model."
 *
 * Here we uses a simple way of making the binary tree.
 * Assuming the number of classes C = 6,
 * The classes are organized as a binary tree in the following way:
 *
 * @code{.py}
 * *-*-*- 2
 * | | |- 3
 * | |
 * | |-*- 4
 * |   |- 5
 * |
 * |-*- 0
 *   |- 1
 * @endcode
 *
 * where * indicates an internal node, and each leaf node represents a class.
 * - Node 0 ... C-2 are internal nodes.
 * - Node C-1 ... 2C-2 are leaf nodes.
 * - Class c is represented by leaf node \f$c+C-1\f$.
 *
 * We assign an id for each node:
 * - the id of root be 0.
 * - the left child of a node i is 2*i+1.
 * - the right child of a node i is 2*i+2.
 *
 * It's easy to see that:
 * - the parent of node i is \f$\left\lfloor(i-1)/2\right\rfloor\f$.
 * - the j-th level ancestor of node i is
 * \f$\left\lfloor(i+1)/2^{j+1}\right\rfloor - 1\f$.
 * - A node i is a left child of its parent if \f$(i-1)\%2==0\f$.
 *
 */

class HierarchicalSigmoidOp : public framework::OperatorWithKernel {
 public:
  using framework::OperatorWithKernel::OperatorWithKernel;
  void InferShape(framework::InferShapeContext* ctx) const override {
Y
Yancey1989 已提交
64
    PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null.");
Y
Yancey1989 已提交
65 66
    PADDLE_ENFORCE(ctx->HasInput("Ids"), "Input(Ids) should not be null.");
    PADDLE_ENFORCE(ctx->HasInput("W"), "Input(W) should not be null.");
Y
Yancey1989 已提交
67
    PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) should not be null.");
W
weixing02 已提交
68 69
    PADDLE_ENFORCE(ctx->HasOutput("PreOut"),
                   "Output(PreOut) should not be null.");
Y
Yancey1989 已提交
70
    const int64_t batch_size = ctx->GetInputDim("X")[0];
Y
Yancey1989 已提交
71
    std::vector<int64_t> output_shape({batch_size, 1});
Y
Yancey1989 已提交
72 73
    ctx->SetOutputDim("Out", framework::make_ddim(output_shape));
  }
Y
Yancey1989 已提交
74 75

 protected:
W
weixing02 已提交
76
  framework::OpKernelType GetExpectedKernelType(
Y
Yancey1989 已提交
77 78 79 80 81
      const framework::ExecutionContext& ctx) const override {
    return framework::OpKernelType(
        framework::ToDataType(ctx.Input<framework::Tensor>("X")->type()),
        ctx.GetPlace());
  }
Y
Yancey1989 已提交
82 83
};

W
weixing02 已提交
84
template <typename AttrType>
Y
Yancey1989 已提交
85 86
class HierarchicalSigmoidOpMaker : public framework::OpProtoAndCheckerMaker {
 public:
W
weixing02 已提交
87
  void Make() override {
Y
Yancey1989 已提交
88
    AddInput("X",
Y
Yancey1989 已提交
89 90 91
             "(Tensor, required) The input Tensor, which the shape is"
             "[N * D], which N is the size of mini-batch,"
             "D is the embded size");
Y
Yancey1989 已提交
92
    AddInput("W",
Y
Yancey1989 已提交
93
             "(Tensor, required), The parameters of hierarchical "
Y
Yancey1989 已提交
94
             "sigmoid operator, each of them is s a 3-D tensor, the shape is"
W
weixing02 已提交
95
             "[num_classes - 1, D]");
Y
Yancey1989 已提交
96
    AddInput("Ids",
Y
Yancey1989 已提交
97
             "(Tensor, required), The labels of training data. It's a"
Y
Yancey1989 已提交
98
             "1-D tensor, which the shape is [1, N]");
Y
Yancey1989 已提交
99 100
    AddInput("Bias",
             "(Tensor, optional), The bias is a 1-D tensor, "
Y
Yancey1989 已提交
101 102 103 104 105
             "which is applied to the output, the shape is"
             "[1, num_classes -1]");
    AddOutput("Out",
              "(Tensor, required) The output of hierarchical sigmoid operator."
              "the shape is [N, 1]");
W
weixing02 已提交
106 107 108 109 110
    AddOutput("PreOut",
              "(Tensor, required) A intermedia 2-D Tensor, which the shape is "
              "[batch_size, code_length]")
        .AsIntermediate();
    AddAttr<AttrType>("num_classes", "(int, required), The number of classes")
Y
Yancey1989 已提交
111
        .SetDefault(2);
Y
Yancey1989 已提交
112 113
    AddComment(R"DOC(
The hierarchical sigmoid operator organize the classes into a binary tree.
W
weixing02 已提交
114 115 116
At each node, a sigmoid function is used to caculate the probability of
belonging to the right branch. This idea is from
"F. Morin, Y. Bengio (AISTATS 05):
Y
Yancey1989 已提交
117 118 119 120 121
Hierarchical Probabilistic Neural Network Language Model."
      )DOC");
  }
};

W
weixing02 已提交
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
class HierarchicalSigmoidGradOp : public framework::OperatorWithKernel {
 public:
  using framework::OperatorWithKernel::OperatorWithKernel;
  void InferShape(framework::InferShapeContext* ctx) const override {
    PADDLE_ENFORCE(ctx->HasInput("W"), "Input(W) should not be null.");
    PADDLE_ENFORCE(ctx->HasInput("Ids"), "Input(Ids) should not be null.");
    PADDLE_ENFORCE(ctx->HasInput("PreOut"),
                   "Input(Preout) should not be null.");
    PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("W")),
                   "Output(W@Grad should not be null.)");
    PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("X")));
    if (ctx->HasOutput(framework::GradVarName("Bias"))) {
      ctx->SetOutputDim(framework::GradVarName("Bias"),
                        ctx->GetInputDim("Bias"));
    }
    ctx->SetOutputDim(framework::GradVarName("W"), ctx->GetInputDim("W"));
    ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X"));
  }

 protected:
  framework::OpKernelType GetExpectedKernelType(
      const framework::ExecutionContext& ctx) const override {
    return framework::OpKernelType(
        framework::ToDataType(ctx.Input<framework::Tensor>("X")->type()),
        ctx.GetPlace());
  }
};

Y
Yancey1989 已提交
150 151 152 153
}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;
W
weixing02 已提交
154 155 156 157
REGISTER_OPERATOR(hierarchical_sigmoid, ops::HierarchicalSigmoidOp,
                  ops::HierarchicalSigmoidOpMaker<int>,
                  paddle::framework::DefaultGradOpDescMaker<true>);
REGISTER_OPERATOR(hierarchical_sigmoid_grad, ops::HierarchicalSigmoidGradOp);
Y
Yancey1989 已提交
158 159 160 161 162 163
REGISTER_OP_CPU_KERNEL(hierarchical_sigmoid,
                       ops::HierarchicalSigmoidOpKernel<
                           paddle::platform::CPUDeviceContext, float>);
REGISTER_OP_CPU_KERNEL(hierarchical_sigmoid_grad,
                       ops::HierarchicalSigmoidGradOpKernel<
                           paddle::platform::CPUDeviceContext, float>);