caffe accuracy层以及blob的梳理

基础概念

blob的reshape

reshape接收的是一个vector的参数，目的是要给这个blob分配相应的内存空间。这个vector参数，表示了不同的维度，一般就是N*C*H*W，如[0]表示N，[1]表示C，等。
直观的解析就是：有N张图，每张图有C个channel，每张图的大小是H*W，如果每个像素都是float类型，那么，
计算总共有多少像素（元素）的公式就是：N*C*H*W，就是blob的count的值。
计算用来存储这批数据的blob所需的内存的大小为：N*C*H*W*sizeof(float)。
相关源码：

template <typename Dtype>
void Blob<Dtype>::Reshape(const int num, const int channels, const int height,
    const int width) {
  vector<int> shape(4);
  shape[0] = num;
  shape[1] = channels;
  shape[2] = height;
  shape[3] = width;
  Reshape(shape);
}

template <typename Dtype>
void Blob<Dtype>::Reshape(const vector<int>& shape) {
  CHECK_LE(shape.size(), kMaxBlobAxes);
  count_ = 1;
  shape_.resize(shape.size());
  if (!shape_data_ || shape_data_->size() < shape.size() * sizeof(int)) {
    shape_data_.reset(new SyncedMemory(shape.size() * sizeof(int)));
  }
  int* shape_data = static_cast<int*>(shape_data_->mutable_cpu_data());
  for (int i = 0; i < shape.size(); ++i) {
    CHECK_GE(shape[i], 0);
    if (count_ != 0) {
      CHECK_LE(shape[i], INT_MAX / count_) << "blob size exceeds INT_MAX";
    }
    count_ *= shape[i];  //累计count的数量，记住是N*C*H*W这样计算总共有多少元素
    shape_[i] = shape[i]; //同时记住各维度的大小
    shape_data[i] = shape[i];
  }
  if (count_ > capacity_) {
    capacity_ = count_;
    data_.reset(new SyncedMemory(capacity_ * sizeof(Dtype))); //如果reshape的大小比以前的大，要重新分配空间
    diff_.reset(new SyncedMemory(capacity_ * sizeof(Dtype)));
  }
}

blob的count函数

caffe中表示数据都是以N*C*H*W这4个维度来表示的，count函数用来计算某些范围的大小。

/**
   * @brief Compute the volume of a slice; i.e., the product of dimensions
   *        among a range of axes.
   *
   * @param start_axis The first axis to include in the slice.
   *
   * @param end_axis The first axis to exclude from the slice.
   */
  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;
  }

caffe源码中的accuracy_layer.cpp的相关梳理

首先明确accuracy层的功能。
accuracy层是通过对比预测的结果与输入的label，通过统计预测正确的数量与总共要预测的数量的比值得到的。
accuracy层需要两个输入源，一个是经过网络预测的数值，另一个是最开始输入的label至，分别对应了bottom[0]和bottom[1].
对于一批输入N*C*H*W来说，图像表示为：

AccuracyLayer的Reshape函数分析

template <typename Dtype>
void AccuracyLayer<Dtype>::Reshape(
  const vector<Blob<Dtype>*>& bottom, const vector<Blob<Dtype>*>& top) {
  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()); //label_axis默认是1（查看caffe.probo文件)
  outer_num_ = bottom[0]->count(0, label_axis_); //结合上图，该outer_num是N
  inner_num_ = bottom[0]->count(label_axis_ + 1); //结合上图，该inner_num是H*W
  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}.";                      
//上面的check表示的意思为：
//先结合上图理解outer_num与inner_num的数量，以及bottom[1]的结构。
//如果C表示的是预测的种类的数量（就是label有多少种），那么label的数量有多少?
//就是N*H*W，一个输入是H*W大，对于H*W的每个元素都有一个label值，所以，一个输入就会有H*W个label，
//N个输入（一批），就会有N*H*W个label数量。
//很多网络最后一层都是全连接层，即一张h*w的图片进去，到最后就成为了一列单个的元素，N张图进去就是变成N列（计算上来说），
//所以一般输入给accuracy层的bottom[0]都是N*C*1*1.
//对于分类来说，这个特别好理解，每张图片对应一个label，那么就会有N个label
  vector<int> top_shape(0);  // Accuracy is a scalar; 0 axes.
  top[0]->Reshape(top_shape);
  if (top.size() > 1) {
    // Per-class accuracy is a vector; 1 axes.
    vector<int> top_shape_per_class(1);
    top_shape_per_class[0] = bottom[0]->shape(label_axis_);
    top[1]->Reshape(top_shape_per_class);
    nums_buffer_.Reshape(top_shape_per_class);
  }
}

AccuracyLayer的Forward_cpu函数分析

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);
  if (top.size() > 1) {
    caffe_set(nums_buffer_.count(), Dtype(0), nums_buffer_.mutable_cpu_data());
    caffe_set(top[1]->count(), Dtype(0), top[1]->mutable_cpu_data());
  }
  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;
      }
      if (top.size() > 1) ++nums_buffer_.mutable_cpu_data()[label_value];
      DCHECK_GE(label_value, 0);
      DCHECK_LT(label_value, num_labels);
      // Top-k accuracy
      std::vector<std::pair<Dtype, int> > bottom_data_vector;
      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));
      }
      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;
          if (top.size() > 1) ++top[1]->mutable_cpu_data()[label_value];
          break;
        }
      }
      ++count;
    }
  }

  // LOG(INFO) << "Accuracy: " << accuracy;
  top[0]->mutable_cpu_data()[0] = accuracy / count;
  if (top.size() > 1) {
    for (int i = 0; i < top[1]->count(); ++i) {
      top[1]->mutable_cpu_data()[i] =
          nums_buffer_.cpu_data()[i] == 0 ? 0
          : top[1]->cpu_data()[i] / nums_buffer_.cpu_data()[i];
    }
  }
  // Accuracy layer should not be used as a loss function.
}