softmax_with_cross_entropy_op.cc 9.5 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
2

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
6

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

C
caoying03 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. */
14

Y
Yi Wang 已提交
15
#include "paddle/fluid/operators/softmax_with_cross_entropy_op.h"
S
sneaxiy 已提交
16
#include <memory>
Y
Yu Yang 已提交
17

18 19 20 21 22 23
namespace paddle {
namespace operators {

class SoftmaxWithCrossEntropyOpMaker
    : public framework::OpProtoAndCheckerMaker {
 public:
Y
Yu Yang 已提交
24
  void Make() override {
C
caoying03 已提交
25
    AddInput("Logits",
26
             "(Tensor, default: Tensor<float>), The unscaled log probabilities "
C
caoying03 已提交
27
             "which is a 2-D tensor with shape [N x K]. N is the batch_size, "
28 29
             "and K is the class number.");
    AddInput("Label",
C
caoying03 已提交
30 31 32 33
             "(Tensor) The ground truth which is a 2-D tensor. If soft_label "
             "is set to false, Label is a Tensor<int64> with shape [N x 1]. If "
             "soft_label is set to true, Label is a Tensor<float/double> with "
             "shape [N x K].");
C
caoying03 已提交
34 35
    AddOutput(
        "Softmax",
36
        "(Tensor, default: Tensor<float>), A 2-D tensor with shape [N x K]. "
C
caoying03 已提交
37 38
        "The outputs value of softmax activation by given the input batch, "
        "which will be used in backward calculation.")
C
caoying03 已提交
39
        .AsIntermediate();
C
caoying03 已提交
40
    AddOutput("Loss",
41
              "(Tensor, default: Tensor<float>), A 2-D tensor. The cross "
C
caoying03 已提交
42
              "entropy loss with shape [N x 1].");
C
caoying03 已提交
43
    AddAttr<bool>(
44
        "soft_label",
C
caoying03 已提交
45 46 47
        "(bool, default: false), A flag to indicate whether to interpretate "
        "the given labels as soft labels.")
        .SetDefault(false);
S
sneaxiy 已提交
48 49
    AddAttr<bool>(
        "numeric_stable_mode",
50
        "(bool, default: true), A flag to indicate whether to use more "
S
sneaxiy 已提交
51 52
        "numerically stable algorithm. This flag is only valid when "
        "soft_label is false and GPU is used.")
53
        .SetDefault(true);
54 55 56 57 58 59
    AddAttr<int>(
        "ignore_index",
        "(int, default -100), Specifies a target value that is ignored and"
        "does not contribute to the input gradient. Only valid if soft_label"
        "is set to False")
        .SetDefault(-100);
60
    AddComment(R"DOC(
61 62 63
Softmax With Cross Entropy Operator.

Cross entropy loss with softmax is used as the output layer extensively. This
64
operator computes the softmax normalized values for each row of the input
65
tensor, after which cross-entropy loss is computed. This provides a more
66 67
numerically stable gradient.

68 69 70
Because this operator performs a softmax on logits internally, it expects
unscaled logits. This operator should not be used with the output of
softmax operator since that would produce incorrect results.
71

C
caoying03 已提交
72
When the attribute soft_label is set false, this operators expects mutually
73 74
exclusive hard labels, each sample in a batch is in exactly one class with a
probability of 1.0. Each sample in the batch will have a single label.
75

76
The equation is as follows:
77

78
1) Hard label (one-hot label, so every sample has exactly one class)
79

80
$$Loss_j =  -\text{Logit}_{Label_j} +
81
\log\left(\sum_{i=0}^{K}\exp(\text{Logit}_i)\right),
82
j = 1,..., K$$
C
caoying03 已提交
83

84
2) Soft label (each sample can have a distribution over all classes)
C
caoying03 已提交
85

86
$$Loss_j =  -\sum_{i=0}^{K}\text{Label}_i \left(\text{Logit}_i -
87
\log\left(\sum_{i=0}^{K}\exp(\text{Logit}_i)\right)\right),
88
j = 1,...,K$$
C
caoying03 已提交
89 90

)DOC");
91 92 93 94 95 96 97
  }
};

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

98
  void InferShape(framework::InferShapeContext* ctx) const override {
Q
qiaolongfei 已提交
99 100 101 102 103 104 105 106 107 108
    PADDLE_ENFORCE(ctx->HasInput("Logits"),
                   "Input(Logits) should be not null.");
    PADDLE_ENFORCE(ctx->HasInput("Label"), "Input(Label) should be not null.");

    PADDLE_ENFORCE(ctx->HasOutput("Softmax"),
                   "Output(Softmax) should be not null.");
    PADDLE_ENFORCE(ctx->HasOutput("Loss"), "Output(Loss) should be not null.");

    auto logits_dims = ctx->GetInputDim("Logits");
    auto labels_dims = ctx->GetInputDim("Label");
109 110

    int rank = logits_dims.size();
C
caoying03 已提交
111
    PADDLE_ENFORCE_EQ(
112 113
        rank, labels_dims.size(),
        "Input(logits) and Input(Label) shall have the same rank.");
P
phlrain 已提交
114 115
    bool check = ctx->IsRuntime() || (framework::product(logits_dims) > 0 &&
                                      framework::product(labels_dims) > 0);
116 117 118 119 120 121
    if (check) {
      PADDLE_ENFORCE_EQ(framework::slice_ddim(logits_dims, 0, rank - 1),
                        framework::slice_ddim(labels_dims, 0, rank - 1),
                        "Input(X) and Input(Label) shall have the same shape "
                        "except the last dimension.");
    }
122

123
    if (ctx->Attrs().Get<bool>("soft_label")) {
124 125 126 127 128
      if (check) {
        PADDLE_ENFORCE_EQ(logits_dims[rank - 1], labels_dims[rank - 1],
                          "If Attr(soft_label) == true, the last dimension of "
                          "Input(X) and Input(Label) should be equal.");
      }
129
    } else {
130 131
      PADDLE_ENFORCE_EQ(labels_dims[rank - 1], 1UL,
                        "If Attr(softLabel) == false, the last dimension of "
132 133 134
                        "Input(Label) should be 1.");
    }

Q
qiaolongfei 已提交
135
    ctx->SetOutputDim("Softmax", logits_dims);
136 137 138
    auto loss_dims = logits_dims;
    loss_dims[rank - 1] = 1;
    ctx->SetOutputDim("Loss", loss_dims);
139

Q
qiaolongfei 已提交
140 141
    ctx->ShareLoD("Logits", /*->*/ "Softmax");
    ctx->ShareLoD("Logits", /*->*/ "Loss");
C
caoying03 已提交
142
  }
Y
Yu Yang 已提交
143

144
 protected:
145
  framework::OpKernelType GetExpectedKernelType(
Y
Yu Yang 已提交
146
      const framework::ExecutionContext& ctx) const override {
Y
Yu Yang 已提交
147 148
    return framework::OpKernelType(ctx.Input<Tensor>("Logits")->type(),
                                   ctx.device_context());
Y
Yu Yang 已提交
149
  }
C
caoying03 已提交
150 151 152 153 154 155
};

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

156
  void InferShape(framework::InferShapeContext* ctx) const override {
Q
qiaolongfei 已提交
157 158 159 160 161 162 163 164 165 166
    PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Loss")),
                   "Input(Loss@Grad) should not be null.");
    PADDLE_ENFORCE(ctx->HasInput("Softmax"),
                   "Input(Softmax) should be not null.");
    PADDLE_ENFORCE(ctx->HasInput("Label"), "Input(Label) should be not null.");
    PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("Logits")),
                   "Output(Logits@Grad) should be not null.");

    auto softmax_dims = ctx->GetInputDim("Softmax");
    auto labels_dims = ctx->GetInputDim("Label");
167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183

    int rank = softmax_dims.size();
    PADDLE_ENFORCE_EQ(
        rank, labels_dims.size(),
        "Input(logits) and Input(Label) shall have the same rank.");
    bool check = true;
    if ((!ctx->IsRuntime()) && (framework::product(softmax_dims) <= 0 ||
                                framework::product(labels_dims) <= 0)) {
      check = false;
    }
    if (check) {
      PADDLE_ENFORCE_EQ(
          framework::slice_ddim(softmax_dims, 0, rank - 1),
          framework::slice_ddim(labels_dims, 0, rank - 1),
          "Input(Softmax) and Input(Label) shall have the same shape "
          "except the last dimension.");
    }
184

185
    if (ctx->Attrs().Get<bool>("soft_label")) {
186 187 188 189 190
      if (check) {
        PADDLE_ENFORCE_EQ(softmax_dims[rank - 1], labels_dims[rank - 1],
                          "If Attr(soft_label) == true, the last dimension of "
                          "Input( Softmax) and Input(Label) should be equal.");
      }
191
    } else {
192 193
      PADDLE_ENFORCE_EQ(labels_dims[rank - 1], 1UL,
                        "If Attr(softLabel) == false, the last dimension of "
194 195
                        "Input(Label) should be 1.");
    }
C
caoying03 已提交
196

Q
qiaolongfei 已提交
197 198
    ctx->SetOutputDim(framework::GradVarName("Logits"),
                      ctx->GetInputDim("Softmax"));
199
  }
Y
Yu Yang 已提交
200

201
 protected:
202
  framework::OpKernelType GetExpectedKernelType(
Y
Yu Yang 已提交
203
      const framework::ExecutionContext& ctx) const override {
Y
Yu Yang 已提交
204
    return framework::OpKernelType(
Y
Yu Yang 已提交
205
        ctx.Input<Tensor>(framework::GradVarName("Loss"))->type(),
Y
Yu Yang 已提交
206
        ctx.device_context());
Y
Yu Yang 已提交
207
  }
208 209
};

210 211 212 213 214
class SoftmaxGradMaker : public framework::SingleGradOpDescMaker {
 public:
  using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;

 protected:
Y
Yu Yang 已提交
215 216
  std::unique_ptr<framework::OpDesc> Apply() const override {
    auto* grad_op = new framework::OpDesc();
Y
Yu Yang 已提交
217 218 219 220 221 222 223
    grad_op->SetType("softmax_with_cross_entropy_grad");
    grad_op->SetInput("Label", Input("Label"));
    grad_op->SetInput("Softmax", Output("Softmax"));
    grad_op->SetInput(framework::GradVarName("Softmax"), OutputGrad("Softmax"));
    grad_op->SetInput(framework::GradVarName("Loss"), OutputGrad("Loss"));
    grad_op->SetOutput(framework::GradVarName("Logits"), InputGrad("Logits"));
    grad_op->SetAttrMap(Attrs());
Y
Yu Yang 已提交
224
    return std::unique_ptr<framework::OpDesc>(grad_op);
225 226 227
  }
};

228 229 230 231 232
}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;

233
REGISTER_OPERATOR(softmax_with_cross_entropy, ops::SoftmaxWithCrossEntropyOp,
Y
Yu Yang 已提交
234
                  ops::SoftmaxWithCrossEntropyOpMaker, ops::SoftmaxGradMaker);
235 236
REGISTER_OPERATOR(softmax_with_cross_entropy_grad,
                  ops::SoftmaxWithCrossEntropyOpGrad);
237
REGISTER_OP_CPU_KERNEL(softmax_with_cross_entropy,
C
caoying03 已提交
238 239
                       ops::SoftmaxWithCrossEntropyKernel<float>,
                       ops::SoftmaxWithCrossEntropyKernel<double>);
240
REGISTER_OP_CPU_KERNEL(softmax_with_cross_entropy_grad,
C
caoying03 已提交
241 242
                       ops::SoftmaxWithCrossEntropyGradKernel<float>,
                       ops::SoftmaxWithCrossEntropyGradKernel<double>);