label_smooth_op.cc 5.8 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
Y
Yibing Liu 已提交
2 3 4 5 6 7 8 9 10 11 12 13 14

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. */

Y
Yi Wang 已提交
15
#include "paddle/fluid/operators/label_smooth_op.h"
S
sneaxiy 已提交
16
#include <memory>
17
#include <string>
Y
Yibing Liu 已提交
18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35

namespace paddle {
namespace operators {

class LabelSmoothOp : public framework::OperatorWithKernel {
 public:
  LabelSmoothOp(const std::string &type,
                const framework::VariableNameMap &inputs,
                const framework::VariableNameMap &outputs,
                const framework::AttributeMap &attrs)
      : OperatorWithKernel(type, inputs, outputs, attrs) {}

  void InferShape(framework::InferShapeContext *ctx) const override {
    PADDLE_ENFORCE(ctx->HasInput("X"),
                   "Input(X) of LabelSmoothOp should not be null.");
    PADDLE_ENFORCE(ctx->HasOutput("Out"),
                   "Output(Out) of LabelSmoothOp should not be null.");
    auto in_dims = ctx->GetInputDim("X");
36 37 38 39 40 41 42 43
    if (ctx->HasInput("PriorDist")) {
      auto noise_dims = ctx->GetInputDim("PriorDist");
      auto noise_numel = paddle::framework::product(noise_dims);
      PADDLE_ENFORCE(
          in_dims[1] == noise_numel,
          "The number of elements in Input(PriorDist) must be equal to the "
          "dimension of each label.");
    }
Y
Yibing Liu 已提交
44 45 46 47 48 49 50
    ctx->ShareLoD("X", /*->*/ "Out");
    ctx->SetOutputDim("Out", in_dims);
  }
};

class LabelSmoothOpMaker : public framework::OpProtoAndCheckerMaker {
 public:
Y
Yu Yang 已提交
51
  void Make() override {
52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67
    AddInput("X",
             "(LoDTensor) The input labels of LabelSmooth operator. This "
             "input can be batched labels in one-hot encoding or output from "
             "softmax, with shape [N x K], where N is the batch size and K is "
             "the number of classes");
    AddInput("PriorDist",
             "(Tensor, optional)"
             "The prior distribution to be added to the smoothed label. It is "
             "fixed during training and the number of elements should be equal "
             "to the dimension K of each label. Default is uniform "
             "distribution and each element will be set to 1/K if not provided "
             "in input.")
        .AsDispensable();
    AddOutput("Out",
              "(loDTensor) The smoothed label of LabelSmooth operator. It has"
              "the same shape and LoD with the Input(LoDTensor).");
Y
Yibing Liu 已提交
68 69 70 71 72 73 74
    AddAttr<float>("epsilon",
                   "(float, default 0.0f)"
                   "The smoothing parameter of LabelSmooth operator.")
        .SetDefault(0.0f);
    AddComment(R"DOC(
LabelSmooth Operator.

75 76 77 78 79 80 81 82
Label smoothing is a mechanism to regularize the classifier layer. In machine 
learning, optimizing the log-likelihood of the correct label directly may 
cause two problems. First, it may result in overfitting: if the model learns 
to assign full probability to the ground-truth label for each training example,
it is not guaranteed to generalize. Second, it encourages the differences 
between the largest logit and all others to become large, reducing the ability 
of the model to adapt. Label smoothing is proposed to encourage the model to 
be less confident, which replaces the ground-truth label $y$ with the weighted 
Y
Yibing Liu 已提交
83
sum of itself and some fixed distribution $\mu$, i.e.
84 85 86 87 88 89 90 91 92 93 94 95

$$
    \tilde{y} = (1 - \epsilon) * y + \epsilon * \mu,
$$

where $(1 - \epsilon)$ and $\epsilon$ are the weights respectively, and 
$\tilde{y}$ is the smoothed label. Usually uniform distribution is used for 
$\mu$. This change in the ground-truth label is called label-smoothing 
regularization or LSR.

See more details about label smoothing in https://arxiv.org/abs/1512.00567.

Y
Yibing Liu 已提交
96 97 98 99 100 101 102 103 104 105 106 107 108
)DOC");
  }
};

class LabelSmoothGradOp : public framework::OperatorWithKernel {
 public:
  LabelSmoothGradOp(const std::string &type,
                    const framework::VariableNameMap &inputs,
                    const framework::VariableNameMap &outputs,
                    const framework::AttributeMap &attrs)
      : OperatorWithKernel(type, inputs, outputs, attrs) {}

  void InferShape(framework::InferShapeContext *ctx) const override {
S
sneaxiy 已提交
109 110 111 112 113
    ctx->SetOutputDim(framework::GradVarName("X"),
                      ctx->GetInputDim(framework::GradVarName("Out")));
  }
};

H
hong 已提交
114 115
template <typename T>
class LabelSmoothGradMaker : public framework::SingleGradOpMaker<T> {
S
sneaxiy 已提交
116
 public:
H
hong 已提交
117
  using framework::SingleGradOpMaker<T>::SingleGradOpMaker;
S
sneaxiy 已提交
118 119

 protected:
H
hong 已提交
120 121
  std::unique_ptr<T> Apply() const override {
    std::unique_ptr<T> op(new T());
S
sneaxiy 已提交
122
    op->SetType("label_smooth_grad");
H
hong 已提交
123 124 125
    op->SetInput(framework::GradVarName("Out"), this->OutputGrad("Out"));
    op->SetOutput(framework::GradVarName("X"), this->InputGrad("X"));
    op->SetAttrMap(this->Attrs());
S
sneaxiy 已提交
126
    return op;
Y
Yibing Liu 已提交
127 128 129 130 131 132 133
  }
};

}  // namespace operators
}  // namespace paddle
namespace ops = paddle::operators;

Y
Yang Yang 已提交
134
REGISTER_OPERATOR(label_smooth, ops::LabelSmoothOp, ops::LabelSmoothOpMaker,
H
hong 已提交
135 136
                  ops::LabelSmoothGradMaker<paddle::framework::OpDesc>,
                  ops::LabelSmoothGradMaker<paddle::imperative::OpBase>);
137
REGISTER_OPERATOR(label_smooth_grad, ops::LabelSmoothGradOp);
Y
Yibing Liu 已提交
138 139 140 141 142 143 144 145
REGISTER_OP_CPU_KERNEL(
    label_smooth,
    ops::LabelSmoothKernel<paddle::platform::CPUDeviceContext, float>,
    ops::LabelSmoothKernel<paddle::platform::CPUDeviceContext, double>);
REGISTER_OP_CPU_KERNEL(
    label_smooth_grad,
    ops::LabelSmoothGradKernel<paddle::platform::CPUDeviceContext, float>,
    ops::LabelSmoothGradKernel<paddle::platform::CPUDeviceContext, double>);