onehot_cross_entropy_op.h 2.9 KB
Newer Older
Q
Qiao Longfei 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
/* 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. */

#pragma once
D
dongzhihong 已提交
16
#include "paddle/framework/op_registry.h"
Q
Qiao Longfei 已提交
17 18 19 20

namespace paddle {
namespace operators {

D
dongzhihong 已提交
21 22
using Tensor = framework::Tensor;

23
template <typename T>
D
dangqingqing 已提交
24
inline T tolerable_value(const T x) {
25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40
  static_assert(std::is_floating_point<T>::value,
                "tolerable_value works only on float, "
                "double and double double.");

  const T kApproInf = 1e20;

  if (x == INFINITY) {
    return kApproInf;
  }

  if (x == -INFINITY) {
    return -kApproInf;
  }

  return x;
}
Y
Yan Chunwei 已提交
41

42
template <typename T>
D
dongzhihong 已提交
43
class OnehotCrossEntropyOpKernel : public framework::OpKernel {
44
 public:
D
dongzhihong 已提交
45
  void Compute(const framework::ExecutionContext& ctx) const override {
46 47 48
    PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()),
                   "It must use CPUPlace.");

Y
Yan Chunwei 已提交
49 50
    auto X = ctx.Input<Tensor>("X");
    const T* Xdata = X->data<T>();
Y
Yu Yang 已提交
51
    const int* label_data = ctx.Input<Tensor>("label")->data<int>();
Y
Yan Chunwei 已提交
52
    auto Y = ctx.Output<Tensor>("Y");
Q
Qiao Longfei 已提交
53

54
    Y->mutable_data<T>(ctx.GetPlace());
Q
Qiao Longfei 已提交
55

Y
Yan Chunwei 已提交
56
    T* Ydata = Y->data<T>();
Q
Qiao Longfei 已提交
57

58 59
    int batch_size = X->dims()[0];
    int class_num = X->dims()[1];
Q
Qiao Longfei 已提交
60 61

    for (int i = 0; i < batch_size; ++i) {
62 63
      int index = i * class_num + label_data[i];
      Ydata[i] = -tolerable_value(std::log(Xdata[index]));
Y
Yan Chunwei 已提交
64 65 66 67
    }
  }
};

68
template <typename T>
D
dongzhihong 已提交
69
class OnehotCrossEntropyGradientOpKernel : public framework::OpKernel {
Y
Yan Chunwei 已提交
70
 public:
D
dongzhihong 已提交
71
  void Compute(const framework::ExecutionContext& ctx) const override {
72 73 74
    PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()),
                   "It must use CPUPlace.");

Y
Yan Chunwei 已提交
75 76 77 78 79 80 81 82 83 84 85 86 87
    auto X = ctx.Input<Tensor>("X");
    auto dX = ctx.Output<Tensor>(framework::GradVarName("X"));
    auto dY = ctx.Input<Tensor>(framework::GradVarName("Y"));
    auto label = ctx.Input<Tensor>("label");

    auto* dXdata = dX->template mutable_data<T>(ctx.GetPlace());
    auto* dYdata = dY->template data<T>();
    auto* Xdata = X->template data<T>();
    auto* label_data = label->data<int>();

    const int batch_size = X->dims()[0];
    const int class_num = X->dims()[1];

88
    // TODO(qingqing): make zero setting an common function.
89
    memset(dXdata, 0, sizeof(T) * batch_size * class_num);
Y
Yan Chunwei 已提交
90
    for (int i = 0; i < batch_size; ++i) {
91 92
      int index = i * class_num + label_data[i];
      dXdata[index] = -tolerable_value(dYdata[i] / Xdata[index]);
Q
Qiao Longfei 已提交
93 94 95 96 97 98
    }
  }
};

}  // namespace operators
}  // namespace paddle