precision_recall_op.cc 8.4 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
Y
yangyaming 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. */

W
Wu Yi 已提交
15
#include "paddle/fluid/operators/metrics/precision_recall_op.h"
Y
yangyaming 已提交
16

Y
yangyaming 已提交
17 18 19 20 21 22 23 24
namespace paddle {
namespace operators {

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

  void InferShape(framework::InferShapeContext *ctx) const override {
25 26 27 28
    PADDLE_ENFORCE(ctx->HasInput("MaxProbs"),
                   "Input(MaxProbs) should not be null.");
    PADDLE_ENFORCE(ctx->HasInput("Indices"),
                   "Input(Indices) should not be null.");
Y
yangyaming 已提交
29 30 31 32 33 34 35 36 37
    PADDLE_ENFORCE(ctx->HasInput("Labels"),
                   "Input(Labels) should not be null.");
    PADDLE_ENFORCE(ctx->HasOutput("BatchMetrics"),
                   "Output(BatchMetrics) should not be null.");
    PADDLE_ENFORCE(ctx->HasOutput("AccumMetrics"),
                   "Output(AccumMetrics) should not be null.");
    PADDLE_ENFORCE(ctx->HasOutput("AccumStatesInfo"),
                   "Output(AccumStatesInfo) should not be null.");

38 39 40
    int64_t cls_num =
        static_cast<int64_t>(ctx->Attrs().Get<int>("class_number"));
    auto max_probs_dims = ctx->GetInputDim("MaxProbs");
Y
yangyaming 已提交
41 42
    auto labels_dims = ctx->GetInputDim("Labels");

43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58
    if (ctx->IsRuntime()) {
      PADDLE_ENFORCE_EQ(max_probs_dims[1], 1,
                        "Each instance contains one max probability, so the "
                        "shape of Input(MaxProbs) should be [batch_size, 1].");
      PADDLE_ENFORCE_EQ(
          ctx->GetInputDim("Indices"), max_probs_dims,
          "The shape of Input(Indices) should bes same with max_probs_dims");
      PADDLE_ENFORCE_EQ(
          max_probs_dims[0], labels_dims[0],
          "The 1st dimension of Input(MaxProbs) and "
          "Input(Labels) both are batch_size and the shape should "
          "be the same.");
      PADDLE_ENFORCE_EQ(labels_dims[1], 1,
                        "The 2nd dimension of Input(Labels) contains instance "
                        "label and the shape should be equal to 1.");
    }
Y
yangyaming 已提交
59 60
    if (ctx->HasInput("Weights")) {
      auto weights_dims = ctx->GetInputDim("Weights");
61 62 63 64 65 66 67

      if (ctx->IsRuntime()) {
        PADDLE_ENFORCE_EQ(weights_dims,
                          framework::make_ddim({max_probs_dims[0], 1}),
                          "The shape of Input(Weights) should be "
                          "[batch_size, 1].");
      }
Y
yangyaming 已提交
68 69 70
    }
    if (ctx->HasInput("StatesInfo")) {
      auto states_dims = ctx->GetInputDim("StatesInfo");
71 72 73 74 75 76

      if (ctx->IsRuntime()) {
        PADDLE_ENFORCE_EQ(states_dims, framework::make_ddim({cls_num, 4}),
                          "The shape of Input(StatesInfo) should be "
                          "[class_number, 4].");
      }
Y
yangyaming 已提交
77 78 79 80 81 82 83 84 85 86 87 88
    }

    // Layouts of BatchMetrics and AccumMetrics both are:
    // [
    //  macro average precision, macro average recall, macro average F1 score,
    //  micro average precision, micro average recall, micro average F1 score
    // ]
    ctx->SetOutputDim("BatchMetrics", {6});
    ctx->SetOutputDim("AccumMetrics", {6});
    // Shape of AccumStatesInfo is [class_number, 4]
    // The layout of each row is:
    // [ TP, FP, TN, FN ]
89
    ctx->SetOutputDim("AccumStatesInfo", {cls_num, 4});
Y
yangyaming 已提交
90
  }
Y
yangyaming 已提交
91 92

 protected:
93
  framework::OpKernelType GetExpectedKernelType(
Y
yangyaming 已提交
94
      const framework::ExecutionContext &ctx) const override {
95 96 97
    return framework::OpKernelType(
        OperatorWithKernel::IndicateVarDataType(ctx, "MaxProbs"),
        ctx.device_context());
Y
yangyaming 已提交
98
  }
Y
yangyaming 已提交
99 100 101 102
};

class PrecisionRecallOpMaker : public framework::OpProtoAndCheckerMaker {
 public:
Y
Yu Yang 已提交
103
  void Make() override {
104
    AddInput("MaxProbs",
105
             "(Tensor, default Tensor<float>) A 2-D tensor with shape N x 1, "
106 107 108 109
             "where N is the batch size. Each row contains the max probability "
             "of an instance which computed by the previous top_k (k=1) "
             "operator.");
    AddInput("Indices",
110
             "(Tensor, default Tensor<int>) A 2-D tensor with shape N x 1, "
111 112
             "where N is the batch size. Each row contains the corresponding "
             "index which computed by the previous top_k (k=1) operator.");
Y
yangyaming 已提交
113
    AddInput("Labels",
114
             "(Tensor, default Tensor<int>) A 2-D tensor with shape N x 1, "
Y
yangyaming 已提交
115 116 117
             "where N is the batch size. Each element is a label and the "
             "value should be in [0, class_number - 1].");
    AddInput("Weights",
118
             "(Tensor, default Tensor<float>) A 2-D tensor with shape N x 1, "
Y
yangyaming 已提交
119 120 121 122
             "where N is the batch size. This input is optional. If provided, "
             "weight of instance would be considered when computing metrics.")
        .AsDispensable();
    AddInput("StatesInfo",
123
             "(Tensor, default Tensor<int>) A 2-D tensor with shape D x 4, "
Y
yangyaming 已提交
124 125
             "where D is the number of classes. This input is optional. If "
             "provided, current state will be accumulated to this state and "
126
             "the accumulation state will be the output state.")
Y
yangyaming 已提交
127
        .AsDispensable();
Y
yangyaming 已提交
128
    AddOutput("BatchMetrics",
129 130
              "(Tensor, default Tensor<float>) A 1-D tensor with shape {6}. "
              "This output tensor contains metrics for current batch data. "
Y
yangyaming 已提交
131 132
              "The layout is [macro average precision, macro average recall, "
              "macro f1 score, micro average precision, micro average recall, "
133
              "micro f1 score].");
Y
yangyaming 已提交
134
    AddOutput("AccumMetrics",
135 136
              "(Tensor, default Tensor<float>) A 1-D tensor with shape {6}. "
              "This output tensor contains metrics for accumulated data. "
Y
yangyaming 已提交
137 138
              "The layout is [macro average precision, macro average recall, "
              "macro f1 score, micro average precision, micro average recall, "
139
              "micro f1 score].");
Y
yangyaming 已提交
140
    AddOutput("AccumStatesInfo",
141
              "(Tensor, default Tensor<float>) A 2-D tensor with shape D x 4, "
Y
yangyaming 已提交
142 143 144 145
              "where D is equal to class number. This output tensor contains "
              "accumulated state variables used to compute metrics. The layout "
              "for each class is [true positives, false positives, "
              "true negatives, false negatives].");
146
    AddAttr<int>("class_number", "(int) Number of classes to be evaluated.");
Y
yangyaming 已提交
147
    AddComment(R"DOC(
148 149 150
Precision Recall Operator.

When given Input(Indices) and Input(Labels), this operator can be used
Y
yangyaming 已提交
151
to compute various metrics including:
152 153 154 155 156 157
1. macro average precision
2. macro average recall
3. macro f1 score
4. micro average precision
5. micro average recall
6. micro f1 score
Y
yangyaming 已提交
158

159
To compute the above metrics, we need to do statistics for true positives,
160
false positives and false negatives. Here the count of true negatives is not
161
necessary, but counting it may provide potential usage and the cost is
162
trivial, so the operator also provides the count of true negatives.
Y
yangyaming 已提交
163

164
We define state as a 2-D tensor with shape [class_number, 4]. Each row of a
Y
yangyaming 已提交
165 166
state contains statistic variables for corresponding class. Layout of each row
is: TP(true positives), FP(false positives), TN(true negatives),
167 168
FN(false negatives). If Input(Weights) is provided, TP, FP, TN, FN will be
calculated by given weight instead of the instance count.
Y
yangyaming 已提交
169 170

This operator also supports metrics computing for cross-batch situation. To
171 172
achieve this, Input(StatesInfo) should be provided. State of current batch
data will be accumulated to Input(StatesInfo) and Output(AccumStatesInfo)
Y
yangyaming 已提交
173 174
is the accumulation state.

175 176
Output(BatchMetrics) is metrics of current batch data while
Output(AccumStatesInfo) is metrics of accumulation data.
Y
yangyaming 已提交
177

Y
yangyaming 已提交
178 179 180 181 182 183 184 185
)DOC");
  }
};

}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;
186 187 188 189
REGISTER_OPERATOR(
    precision_recall, ops::PrecisionRecallOp, ops::PrecisionRecallOpMaker,
    paddle::framework::EmptyGradOpMaker<paddle::framework::OpDesc>,
    paddle::framework::EmptyGradOpMaker<paddle::imperative::OpBase>);
Y
yangyaming 已提交
190 191 192
REGISTER_OP_CPU_KERNEL(
    precision_recall,
    ops::PrecisionRecallKernel<paddle::platform::CPUPlace, float>,
Y
yangyaming 已提交
193
    ops::PrecisionRecallKernel<paddle::platform::CPUPlace, double>);
新手
引导
客服 返回
顶部