caffe accuracy 学习

首先我们先看一下accuracy （在caffe.proto里面）的类定义

message AccuracyParameter {
  // When computing accuracy, count as correct by comparing the true label to
  // the top k scoring classes.  By default, only compare to the top scoring
  // class (i.e. argmax).
  optional uint32 top_k = 1 [default = 1];//默认只取前得分最高的值位置作为标签，如果大于1，就选择值最高的top_k个，如果标签值在这top_k中，就认为识别正确。


  // The "label" axis of the prediction blob, whose argmax corresponds to the
  // predicted label -- may be negative to index from the end (e.g., -1 for the
  // last axis).  For example, if axis == 1 and the predictions are
  // (N x C x H x W), the label blob is expected to contain N*H*W ground truth
  // labels with integer values in {0, 1, ..., C-1}.
  optional int32 axis = 2 [default = 1];


  // If specified, ignore instances with the given label.
  optional int32 ignore_label = 3;//忽略的标签，出现这类标签就不进行统计。
}

accuracy_layer.cpp

template <typename Dtype>

void AccuracyLayer<Dtype>::LayerSetUp(
  const vector<Blob<Dtype>*>& bottom, const vector<Blob<Dtype>*>& top) {
  top_k_ = this->layer_param_.accuracy_param().top_k();


  has_ignore_label_ =
    this->layer_param_.accuracy_param().has_ignore_label();
  if (has_ignore_label_) {
    ignore_label_ = this->layer_param_.accuracy_param().ignore_label();
  }

}

template <typename Dtype>
void AccuracyLayer<Dtype>::Reshape(
  const vector<Blob<Dtype>*>& bottom, const vector<Blob<Dtype>*>& top) {

//bottom[0]是前一层的输入 bottom[1]是label的输入

  CHECK_LE(top_k_, bottom[0]->count() / bottom[1]->count())
      << "top_k must be less than or equal to the number of classes.";
  label_axis_ =
      bottom[0]->CanonicalAxisIndex(this->layer_param_.accuracy_param().axis());

/*

 inline int count(int start_axis, int end_axis) const {
    CHECK_LE(start_axis, end_axis);
    CHECK_GE(start_axis, 0);
    CHECK_GE(end_axis, 0);
    CHECK_LE(start_axis, num_axes());
    CHECK_LE(end_axis, num_axes());
    int count = 1;
    for (int i = start_axis; i < end_axis; ++i) {
      count *= shape(i);
    }
    return count;
  }
  inline int count(int start_axis) const {
    return count(start_axis, num_axes());
  }

*/
  outer_num_ = bottom[0]->count(0, label_axis_);//如果默认的情况就是label_axis_ = 1，outer_num_ = N
  inner_num_ = bottom[0]->count(label_axis_ + 1);//如果默认的情况就是label_axis_ = 1，inner_num_ = W*H

  CHECK_EQ(outer_num_ * inner_num_, bottom[1]->count())
      << "Number of labels must match number of predictions; "
      << "e.g., if label axis == 1 and prediction shape is (N, C, H, W), "
      << "label count (number of labels) must be N*H*W, "
      << "with integer values in {0, 1, ..., C-1}.";
  vector<int> top_shape(0);  // Accuracy is a scalar; 0 axes.??为什么这样设置，不知？
  top[0]->Reshape(top_shape);
}

template <typename Dtype>
void AccuracyLayer<Dtype>::Forward_cpu(const vector<Blob<Dtype>*>& bottom,
    const vector<Blob<Dtype>*>& top) {
  Dtype accuracy = 0;
  const Dtype* bottom_data = bottom[0]->cpu_data();
  const Dtype* bottom_label = bottom[1]->cpu_data();
  const int dim = bottom[0]->count() / outer_num_;
  const int num_labels = bottom[0]->shape(label_axis_);
  vector<Dtype> maxval(top_k_+1);
  vector<int> max_id(top_k_+1);
  int count = 0;
  for (int i = 0; i < outer_num_; ++i) {
    for (int j = 0; j < inner_num_; ++j) {
      const int label_value =
          static_cast<int>(bottom_label[i * inner_num_ + j]);
      if (has_ignore_label_ && label_value == ignore_label_) {
        continue;
      }
      DCHECK_GE(label_value, 0);
      DCHECK_LT(label_value, num_labels);
      // Top-k accuracy
      std::vector<std::pair<Dtype, int> > bottom_data_vector;

      //将前一层输入进来的信息进行pair，pair为值和其位置k。
      for (int k = 0; k < num_labels; ++k) {
        bottom_data_vector.push_back(std::make_pair(
            bottom_data[i * dim + k * inner_num_ + j], k));
      }

     //排序
      std::partial_sort(
          bottom_data_vector.begin(), bottom_data_vector.begin() + top_k_,
          bottom_data_vector.end(), std::greater<std::pair<Dtype, int> >());
      // check if true label is in top k predictions
      for (int k = 0; k < top_k_; k++) {
        if (bottom_data_vector[k].second == label_value) {
          ++accuracy;
          break;
        }
      }
      ++count;
    }
  }


  // LOG(INFO) << "Accuracy: " << accuracy;
  top[0]->mutable_cpu_data()[0] = accuracy / count;
  // Accuracy layer should not be used as a loss function.
}