nce_op.h 8.3 KB
Newer Older
W
wanghaoshuang 已提交
1 2
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.

L
Luo Tao 已提交
3 4 5
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
W
wanghaoshuang 已提交
6

L
Luo Tao 已提交
7
    http://www.apache.org/licenses/LICENSE-2.0
W
wanghaoshuang 已提交
8

L
Luo Tao 已提交
9 10 11 12 13
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
wanghaoshuang 已提交
14 15 16

#pragma once

W
wanghaoshuang 已提交
17
#include <math.h>
W
wanghaoshuang 已提交
18 19 20 21 22 23 24
#include <random>
#include "paddle/framework/eigen.h"
#include "paddle/framework/op_registry.h"
#include "unsupported/Eigen/CXX11/Tensor"
namespace paddle {
namespace operators {

25
using Tensor = framework::Tensor;
W
wanghaoshuang 已提交
26 27 28 29 30

template <typename T, int MajorType = Eigen::RowMajor,
          typename IndexType = Eigen::DenseIndex>
using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;

Q
QI JUN 已提交
31
template <typename DeviceContext, typename T>
W
wanghaoshuang 已提交
32 33
void PrepareSamples(const framework::ExecutionContext& context) {
  auto label = context.Input<Tensor>("Label");
W
wanghaoshuang 已提交
34
  const int64_t* label_data = label->data<int64_t>();
W
wanghaoshuang 已提交
35
  auto label_dims = label->dims();
W
wanghaoshuang 已提交
36
  int num_total_classes = context.Attr<int>("num_total_classes");
W
wanghaoshuang 已提交
37
  // for unitest
W
wanghaoshuang 已提交
38 39
  std::vector<int> custom_neg_classes =
      context.Attr<std::vector<int>>("custom_neg_classes");
W
wanghaoshuang 已提交
40 41 42
  // random machine
  std::random_device rd;
  std::mt19937 rng(rd());
W
wanghaoshuang 已提交
43
  std::uniform_int_distribution<int> rand(0, num_total_classes - 1);
W
wanghaoshuang 已提交
44 45 46

  auto sample_labels = context.Output<Tensor>("SampleLabels");
  auto sample_labels_dims = sample_labels->dims();
W
wanghaoshuang 已提交
47 48
  int64_t* sample_labels_data =
      sample_labels->mutable_data<int64_t>(context.GetPlace());
W
wanghaoshuang 已提交
49 50

  int num_label = label_dims.size() == 2 ? label_dims[1] : 1;
W
wanghaoshuang 已提交
51
  int index = 0;
52
  for (int64_t i = 0; i < label_dims[0]; ++i) {
W
wanghaoshuang 已提交
53 54
    int j = 0;
    for (; j < num_label; ++j) {
W
wanghaoshuang 已提交
55
      sample_labels_data[index++] = label_data[i * num_label + j];
W
wanghaoshuang 已提交
56
    }
W
wanghaoshuang 已提交
57 58
    if (custom_neg_classes.size() > 0) {
      for (auto label : custom_neg_classes) {
W
wanghaoshuang 已提交
59 60 61 62
        sample_labels_data[index++] = label;
      }
    } else {
      for (; j < sample_labels_dims[1]; ++j) {
W
wanghaoshuang 已提交
63
        // TODO(wanghaoshuang): support more distribution sampling
W
wanghaoshuang 已提交
64 65
        sample_labels_data[index++] = rand(rng);
      }
W
wanghaoshuang 已提交
66 67 68 69
    }
  }
}

Q
QI JUN 已提交
70
template <typename DeviceContext, typename T>
W
wanghaoshuang 已提交
71 72 73
class NCEKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& context) const override {
Q
QI JUN 已提交
74
    PrepareSamples<DeviceContext, T>(context);
W
wanghaoshuang 已提交
75
    auto sample_labels = context.Output<Tensor>("SampleLabels");
W
wanghaoshuang 已提交
76
    const int64_t* sample_labels_data = sample_labels->data<int64_t>();
W
wanghaoshuang 已提交
77 78 79 80 81 82 83 84
    auto sample_out = context.Output<Tensor>("SampleLogits");
    T* sample_out_data = sample_out->mutable_data<T>(context.GetPlace());
    auto label = context.Input<Tensor>("Label");
    auto sample_weight = context.Input<Tensor>("SampleWeight");
    const T* sample_weight_data = nullptr;
    if (sample_weight != nullptr) {
      sample_weight_data = sample_weight->data<T>();
    }
W
wanghaoshuang 已提交
85
    auto out = context.Output<Tensor>("Cost");
W
wanghaoshuang 已提交
86
    T* out_data = out->mutable_data<T>(context.GetPlace());
W
wanghaoshuang 已提交
87 88
    int num_neg_samples = context.Attr<int>("num_neg_samples");
    int num_total_classes = context.Attr<int>("num_total_classes");
89
    int64_t num_true_class = 1;
W
wanghaoshuang 已提交
90 91 92
    if (label != nullptr) {
      num_true_class = label->dims()[1];
    }
W
wanghaoshuang 已提交
93
    T b = 1. / num_total_classes * num_neg_samples;
W
wanghaoshuang 已提交
94
    // forward bias
W
wanghaoshuang 已提交
95
    auto bias = context.Input<Tensor>("Bias");
W
wanghaoshuang 已提交
96 97
    if (bias != nullptr) {
      const T* bias_data = bias->data<T>();
98
      for (int64_t i = 0; i < sample_labels->numel(); ++i) {
W
wanghaoshuang 已提交
99 100 101
        sample_out_data[i] = bias_data[sample_labels_data[i]];
      }
    } else {
102
      for (int64_t i = 0; i < sample_labels->numel(); ++i) {
W
wanghaoshuang 已提交
103 104 105 106
        sample_out_data[i] = 0;
      }
    }
    // forward mul
W
wanghaoshuang 已提交
107 108
    auto input_mat = EigenMatrix<T>::From(*(context.Input<Tensor>("Input")));
    auto weight_mat = EigenMatrix<T>::From(*(context.Input<Tensor>("Weight")));
109
    for (int64_t i = 0; i < sample_labels->numel(); ++i) {
110
      Eigen::Tensor<T, 0, Eigen::RowMajor, Eigen::DenseIndex> result =
W
wanghaoshuang 已提交
111 112 113 114
          (input_mat.chip((int)(i / sample_labels->dims()[1]), 0) *
           weight_mat.chip(sample_labels_data[i], 0))
              .sum();
      sample_out_data[i] += result(0);
W
wanghaoshuang 已提交
115
      sample_out_data[i] = (1. / (1. + exp(-sample_out_data[i])));
W
wanghaoshuang 已提交
116 117
    }
    // forward cost
118 119
    for (int64_t i = 0; i < sample_labels->dims()[0]; ++i) {
      int64_t j = 0;
W
wanghaoshuang 已提交
120 121
      out_data[i] = 0;
      T w = sample_weight == nullptr ? 1. : sample_weight_data[i];
W
wanghaoshuang 已提交
122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137
      // for true classes
      for (; j < num_true_class; ++j) {
        T o = sample_out_data[i * sample_out->dims()[1] + j];
        T cost = -log(o / (o + b));
        out_data[i] += w * cost;
      }
      // for sampled neg classes
      for (; j < sample_labels->dims()[1]; ++j) {
        T o = sample_out_data[i * sample_out->dims()[1] + j];
        T cost = -log(b / (o + b));
        out_data[i] += w * cost;
      }
    }
  }
};

Q
QI JUN 已提交
138
template <typename DeviceContext, typename T>
W
wanghaoshuang 已提交
139 140 141
class NCEGradKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& context) const override {
W
wanghaoshuang 已提交
142 143
    auto d_out = context.Input<Tensor>(framework::GradVarName("Cost"));
    const T* d_out_data = d_out->data<T>();
W
wanghaoshuang 已提交
144 145 146 147
    auto label = context.Input<Tensor>("Label");
    auto sample_out = context.Input<Tensor>("SampleLogits");
    const T* sample_out_data = sample_out->data<T>();
    auto sample_labels = context.Input<Tensor>("SampleLabels");
W
wanghaoshuang 已提交
148
    const int64_t* sample_labels_data = sample_labels->data<int64_t>();
W
wanghaoshuang 已提交
149 150 151 152 153
    auto sample_weight = context.Input<Tensor>("SampleWeight");
    const T* sample_weight_data = nullptr;
    if (sample_weight != nullptr) {
      sample_weight_data = sample_weight->data<T>();
    }
W
wanghaoshuang 已提交
154 155
    int num_neg_samples = context.Attr<int>("num_neg_samples");
    int num_total_classes = context.Attr<int>("num_total_classes");
W
wanghaoshuang 已提交
156 157 158 159
    int num_true_class = 1;
    if (label != nullptr) {
      num_true_class = label->dims()[1];
    }
W
wanghaoshuang 已提交
160
    T b = 1. / num_total_classes * num_neg_samples;
W
wanghaoshuang 已提交
161 162 163 164
    Tensor sample_grad;  // tmp tensor
    T* sample_grad_data =
        sample_grad.mutable_data<T>(sample_labels->dims(), context.GetPlace());
    // backward cost
165
    for (int64_t i = 0; i < sample_labels->numel(); ++i) {
W
wanghaoshuang 已提交
166 167 168 169 170
      T o = sample_out_data[i];
      T w = sample_weight == nullptr
                ? 1
                : sample_weight_data[i / sample_labels->dims()[1]];
      sample_grad_data[i] = (i % sample_labels->dims()[1]) < num_true_class
W
wanghaoshuang 已提交
171 172 173
                                ? w * (b / (o + b)) * (o - 1)
                                : w * (o * (1 - o) / (o + b));
      sample_grad_data[i] *= d_out_data[i / sample_labels->dims()[1]];
W
wanghaoshuang 已提交
174 175
    }
    // get d_bias
W
wanghaoshuang 已提交
176
    auto d_bias = context.Output<Tensor>(framework::GradVarName("Bias"));
W
wanghaoshuang 已提交
177 178
    if (d_bias != nullptr) {
      T* d_bias_data = d_bias->mutable_data<T>(context.GetPlace());
W
wanghaoshuang 已提交
179
      std::fill(d_bias_data, d_bias_data + d_bias->numel(), 0.0);
180
      for (int64_t i = 0; i < sample_labels->numel(); ++i) {
W
wanghaoshuang 已提交
181 182 183 184
        d_bias_data[sample_labels_data[i]] += sample_grad_data[i];
      }
    }
    // get d_w
W
wanghaoshuang 已提交
185
    auto d_w = context.Output<Tensor>(framework::GradVarName("Weight"));
W
wanghaoshuang 已提交
186
    if (d_w != nullptr) {
W
wanghaoshuang 已提交
187 188
      auto d_w_data = d_w->mutable_data<T>(context.GetPlace());
      std::fill(d_w_data, d_w_data + d_w->numel(), 0.0);
W
wanghaoshuang 已提交
189
      auto d_w_matrix = EigenMatrix<T>::From(*d_w);
W
wanghaoshuang 已提交
190
      auto x_matrix = EigenMatrix<T>::From(*(context.Input<Tensor>("Input")));
191
      for (int64_t i = 0; i < sample_labels->numel(); ++i) {
W
wanghaoshuang 已提交
192
        d_w_matrix.chip(sample_labels_data[i], 0) +=
W
wanghaoshuang 已提交
193 194 195 196 197
            x_matrix.chip((int)(i / sample_labels->dims()[1]), 0) *
            sample_grad_data[i];
      }
    }
    // get d_x
W
wanghaoshuang 已提交
198
    auto d_x = context.Output<Tensor>(framework::GradVarName("Input"));
W
wanghaoshuang 已提交
199
    if (d_x != nullptr) {
Y
Yang Yu 已提交
200 201
      auto* d_x_data = d_x->mutable_data<T>(context.GetPlace());
      std::fill(d_x_data, d_x_data + d_x->numel(), 0.0);
W
wanghaoshuang 已提交
202
      auto d_x_matrix = EigenMatrix<T>::From(*d_x);
W
wanghaoshuang 已提交
203
      auto w_matrix = EigenMatrix<T>::From(*(context.Input<Tensor>("Weight")));
204
      for (int64_t i = 0; i < sample_labels->numel(); ++i) {
W
wanghaoshuang 已提交
205 206 207 208 209 210 211 212
        d_x_matrix.chip((int)(i / sample_labels->dims()[1]), 0) +=
            w_matrix.chip(sample_labels_data[i], 0) * sample_grad_data[i];
      }
    }
  }
};
}  // namespace operators
}  // namespace paddle