Caffe源碼解析4： Data_layer

轉載請注明出處，樓燚(yì)航的blog，http://home.cnblogs.com/louyihang-loves-baiyan/

data_layer應該是網絡的最底層，主要是將數據送給blob進入到net中，在data_layer中存在多個跟data_layer相關的類

• BaseDataLayer
• BasePrefetchingDataLayer
• DataLayer
• DummyDataLayer
• HDF5DataLayer
• HDF5OutputLayer
• ImageDataLayer
• MemoryDataLayer
• WindowDataLayer
• Batch

這里首先說明一下這幾個類之間的區別。
首先Layer是基類，這個之前就已經提到過了。其次看HDF5相關的類有兩個，一個是HDF5DataLayer，另一個是HDF5OutputLayer，主要是基於HDF5數據格式的讀取和存儲

留意到這個data_layer的頭文件還include了不少頭文件

#include <string>
#include <utility>
#include <vector>
#include "hdf5.h"

#include "caffe/blob.hpp"
#include "caffe/common.hpp"
#include "caffe/data_reader.hpp"
#include "caffe/data_transformer.hpp"
#include "caffe/filler.hpp"
#include "caffe/internal_thread.hpp"
#include "caffe/layer.hpp"
#include "caffe/proto/caffe.pb.h"
#include "caffe/util/blocking_queue.hpp"
#include "caffe/util/db.hpp"

hdf5就是之前說到的一種主要用於科學數據記錄、能自我描述的數據格式。
還有幾個跟data相關的頭文件比如data_read.hpp,data_transformer.hpp
其中data_reader主要是負責數據的讀取，傳送到data layer中。並且對於每一個source，都會開一一起獨立的reading thread讀取線程，幾十有多個solver在並行的跑。比如在多GPU訓練的時候，可以保證對於數據庫的讀取是順序的

data_transformer.hpp里面的DataTransformer這個類，這個類我們要關注一下，這個類主要能對input data 執一些預處理操作，比如縮放、鏡像、減去均值。同時還支持一些隨機的操作。
其核心的函數如下，這里總共有5個常在的Transform函數，其中所有函數的第二部分是相同的，都是一個目標blob，而輸入根據輸入的情況可以有所選擇，可以是blob,也可以是opencv的mat 結構，或者proto中定義的datum結構。

void Transform(const Datum& datum, Blob<Dtype>* transformed_blob);
void Transform(const vector<Datum> & datum_vector, Blob<Dtype>* transformed_blob);
void Transform(const vector<cv::Mat> & mat_vector, Blob<Dtype>* transformed_blob);
void Transform(const cv::Mat& cv_img, Blob<Dtype>* transformed_blob);
void Transform(Blob<Dtype>* input_blob, Blob<Dtype>* transformed_blob);

TransformationParameter是該類構造器中需要傳入的一些變形參數，相關的操作定義在proto中，摘錄如下,可以看到總共有sacle,mirror,crop_size,mean_file,mean_value,force_color,force_grey共7個相關操作

message TransformationParameter {
  optional float scale = 1 [default = 1];
  optional bool mirror = 2 [default = false];
  optional uint32 crop_size = 3 [default = 0];
  optional string mean_file = 4;
  repeated float mean_value = 5;
  optional bool force_color = 6 [default = false];
  optional bool force_gray = 7 [default = false];
}

首先對於dat_layer，里面根據繼承關系最后的幾個子類分別是ImageDataLayer,DataLayer,WindowDataLayer,MemoryDataLayer,HDF5以及Dummy這里暫時先不做分析。
其實最重要的就是類面的layerSetup.首先我們來看DataLayer的DataLayerSetUp

void DataLayer<Dtype>::DataLayerSetUp(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top) {
  const int batch_size = this->layer_param_.data_param().batch_size();
  //獲得相應的datum，用來初始化top blob
  Datum& datum = *(reader_.full().peek());
  //使用data_transformer 來計算根據datum的期望blob的shape
  vector<int> top_shape = this->data_transformer_->InferBlobShape(datum);
  this->transformed_data_.Reshape(top_shape);
  //首先reshape top[0]，再根據batch的大小進行預取
  top_shape[0] = batch_size;
  top[0]->Reshape(top_shape);
  for (int i = 0; i < this->PREFETCH_COUNT; ++i) {
    this->prefetch_[i].data_.Reshape(top_shape);
  }
  LOG(INFO) << "output data size: " << top[0]->num() << ","
      << top[0]->channels() << "," << top[0]->height() << ","
      << top[0]->width();
  // 同樣reshape label的blob的shape
  if (this->output_labels_) {
    vector<int> label_shape(1, batch_size);
    top[1]->Reshape(label_shape);
    for (int i = 0; i < this->PREFETCH_COUNT; ++i) {
      this->prefetch_[i].label_.Reshape(label_shape);
    }
  }
}

MemoryDataLayer

void MemoryDataLayer<Dtype>::DataLayerSetUp(const vector<Blob<Dtype>*>& bottom,
     const vector<Blob<Dtype>*>& top) {
  //直接通過memory_data_param類設置layer的相關參數
  batch_size_ = this->layer_param_.memory_data_param().batch_size();
  channels_ = this->layer_param_.memory_data_param().channels();
  height_ = this->layer_param_.memory_data_param().height();
  width_ = this->layer_param_.memory_data_param().width();
  size_ = channels_ * height_ * width_;
  CHECK_GT(batch_size_ * size_, 0) <<
      "batch_size, channels, height, and width must be specified and"
      " positive in memory_data_param";
  //這里跟datalayer一樣都是先設置top[0]，然后對label進行reshape
  vector<int> label_shape(1, batch_size_);
  top[0]->Reshape(batch_size_, channels_, height_, width_);
  top[1]->Reshape(label_shape);
  added_data_.Reshape(batch_size_, channels_, height_, width_);
  added_label_.Reshape(label_shape);
  data_ = NULL;
  labels_ = NULL;
  added_data_.cpu_data();
  added_label_.cpu_data();
}

ImageDataLayer,它的DataLayerSetUP函數

void ImageDataLayer<Dtype>::DataLayerSetUp(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top) {
  const int new_height = this->layer_param_.image_data_param().new_height();
  const int new_width  = this->layer_param_.image_data_param().new_width();
  const bool is_color  = this->layer_param_.image_data_param().is_color();
  string root_folder = this->layer_param_.image_data_param().root_folder();

  CHECK((new_height == 0 && new_width == 0) ||
      (new_height > 0 && new_width > 0)) << "Current implementation requires "
      "new_height and new_width to be set at the same time.";
  //讀取圖像文件和相應的label
  const string& source = this->layer_param_.image_data_param().source();
  LOG(INFO) << "Opening file " << source;
  std::ifstream infile(source.c_str());
  string filename;
  int label;
  while (infile >> filename >> label) {
    lines_.push_back(std::make_pair(filename, label));
  }

  if (this->layer_param_.image_data_param().shuffle()) {
    // randomly shuffle data
    LOG(INFO) << "Shuffling data";
    const unsigned int prefetch_rng_seed = caffe_rng_rand();
    prefetch_rng_.reset(new Caffe::RNG(prefetch_rng_seed));
    ShuffleImages();
  }
  LOG(INFO) << "A total of " << lines_.size() << " images.";

  lines_id_ = 0;
  //check是否需要隨機跳過一些圖像
  if (this->layer_param_.image_data_param().rand_skip()) {
    unsigned int skip = caffe_rng_rand() %
        this->layer_param_.image_data_param().rand_skip();
    LOG(INFO) << "Skipping first " << skip << " data points.";
    CHECK_GT(lines_.size(), skip) << "Not enough points to skip";
    lines_id_ = skip;
  }
  //使用Opencv來讀進圖像，然后使用它初始化相應的top blob
  cv::Mat cv_img = ReadImageToCVMat(root_folder + lines_[lines_id_].first,
                                    new_height, new_width, is_color);
  CHECK(cv_img.data) << "Could not load " << lines_[lines_id_].first;
  //這里的步驟和上面相同，使用transformer來做reshape
  vector<int> top_shape = this->data_transformer_->InferBlobShape(cv_img);
  this->transformed_data_.Reshape(top_shape);
  //之后部分跟前面差不多，初始化top[0]
  const int batch_size = this->layer_param_.image_data_param().batch_size();
  CHECK_GT(batch_size, 0) << "Positive batch size required";
  top_shape[0] = batch_size;
  for (int i = 0; i < this->PREFETCH_COUNT; ++i) {
    this->prefetch_[i].data_.Reshape(top_shape);
  }
  top[0]->Reshape(top_shape);

  LOG(INFO) << "output data size: " << top[0]->num() << ","
      << top[0]->channels() << "," << top[0]->height() << ","
      << top[0]->width();
  //reshape label
  vector<int> label_shape(1, batch_size);
  top[1]->Reshape(label_shape);
  for (int i = 0; i < this->PREFETCH_COUNT; ++i) {
    this->prefetch_[i].label_.Reshape(label_shape);
  }
}

WindowDataLayer的DataLayerSetUp，這個函數標比較長，我只列出了其中主要的部分，之前的Image相當於是已經剪裁過的一個圖像，也就是說你的目標基本上是充棉了整個畫面，而Window File是用於原始圖的，也就是說有background和object，這個window file 的格式如下

window_file format
repeated:
   # image_index
   img_path (abs path)
   channels
   height
   width
   num_windows
   class_index overlap x1 y1 x2 y2

//讀取每一個box
int num_windows;
infile >> num_windows;
const float fg_threshold =
    this->layer_param_.window_data_param().fg_threshold();
const float bg_threshold =
    this->layer_param_.window_data_param().bg_threshold();
for (int i = 0; i < num_windows; ++i) {
  int label, x1, y1, x2, y2;
  float overlap;
  infile >> label >> overlap >> x1 >> y1 >> x2 >> y2;

  vector<float> window(WindowDataLayer::NUM);
  window[WindowDataLayer::IMAGE_INDEX] = image_index;
  window[WindowDataLayer::LABEL] = label;
  window[WindowDataLayer::OVERLAP] = overlap;
  window[WindowDataLayer::X1] = x1;
  window[WindowDataLayer::Y1] = y1;
  window[WindowDataLayer::X2] = x2;
  window[WindowDataLayer::Y2] = y2;

  // add window to foreground list or background list// read each box
int num_windows;
infile >> num_windows;
const float fg_threshold =
    this->layer_param_.window_data_param().fg_threshold();
const float bg_threshold =
    this->layer_param_.window_data_param().bg_threshold();
for (int i = 0; i < num_windows; ++i) {
  int label, x1, y1, x2, y2;
  float overlap;
  infile >> label >> overlap >> x1 >> y1 >> x2 >> y2;

  vector<float> window(WindowDataLayer::NUM);
  window[WindowDataLayer::IMAGE_INDEX] = image_index;
  window[WindowDataLayer::LABEL] = label;
  window[WindowDataLayer::OVERLAP] = overlap;
  window[WindowDataLayer::X1] = x1;
  window[WindowDataLayer::Y1] = y1;
  window[WindowDataLayer::X2] = x2;
  window[WindowDataLayer::Y2] = y2;

  //首先計算得到overlap,根據Overlap與fg_threshold的比較載添加到fg的list中
  if (overlap >= fg_threshold) {
    int label = window[WindowDataLayer::LABEL];
    CHECK_GT(label, 0);
    fg_windows_.push_back(window);
    label_hist.insert(std::make_pair(label, 0));
    label_hist[label]++;
  } else if (overlap < bg_threshold) {
    // background window, force label and overlap to 0
    window[WindowDataLayer::LABEL] = 0;
    window[WindowDataLayer::OVERLAP] = 0;
    bg_windows_.push_back(window);
    label_hist[0]++;
  }
}
=-
  if (overlap >= fg_threshold) {
    int label = window[WindowDataLayer::LABEL];
    CHECK_GT(label, 0);
    fg_windows_.push_back(window);
    label_hist.insert(std::make_pair(label, 0));
    label_hist[label]++;
  } else if (overlap < bg_threshold) {
    //background的label和overlap都是0
    window[WindowDataLayer::LABEL] = 0;
    window[WindowDataLayer::OVERLAP] = 0;
    bg_windows_.push_back(window);
    label_hist[0]++;
  }
}


..............
for (map<int, int>::iterator it = label_hist.begin();
      it != label_hist.end(); ++it) {
    LOG(INFO) << "class " << it->first << " has " << label_hist[it->first]
              << " samples";
  }

  LOG(INFO) << "Amount of context padding: "
      << this->layer_param_.window_data_param().context_pad();

  LOG(INFO) << "Crop mode: "
      << this->layer_param_.window_data_param().crop_mode();

  //這里之后的步驟就差不多了，同樣是對transform的一些操作
  const int crop_size = this->transform_param_.crop_size();
  CHECK_GT(crop_size, 0);
  const int batch_size = this->layer_param_.window_data_param().batch_size();
  top[0]->Reshape(batch_size, channels, crop_size, crop_size);
  for (int i = 0; i < this->PREFETCH_COUNT; ++i)
    this->prefetch_[i].data_.Reshape(
        batch_size, channels, crop_size, crop_size);

  LOG(INFO) << "output data size: " << top[0]->num() << ","
      << top[0]->channels() << "," << top[0]->height() << ","
      << top[0]->width();
  // 對label進行reshape
  vector<int> label_shape(1, batch_size);
  top[1]->Reshape(label_shape);
  for (int i = 0; i < this->PREFETCH_COUNT; ++i) {
    this->prefetch_[i].label_.Reshape(label_shape);
  }

  //做減均值的操作
  has_mean_file_ = this->transform_param_.has_mean_file();
  has_mean_values_ = this->transform_param_.mean_value_size() > 0;
  if (has_mean_file_) {
    const string& mean_file =
          this->transform_param_.mean_file();
    LOG(INFO) << "Loading mean file from: " << mean_file;
    BlobProto blob_proto;
    ReadProtoFromBinaryFileOrDie(mean_file.c_str(), &blob_proto);
    data_mean_.FromProto(blob_proto);
  }
  if (has_mean_values_) {
    CHECK(has_mean_file_ == false) <<
      "Cannot specify mean_file and mean_value at the same time";
    for (int c = 0; c < this->transform_param_.mean_value_size(); ++c) {
      mean_values_.push_back(this->transform_param_.mean_value(c));
    }
    CHECK(mean_values_.size() == 1 || mean_values_.size() == channels) <<
     "Specify either 1 mean_value or as many as channels: " << channels;
    if (channels > 1 && mean_values_.size() == 1) {
      // Replicate the mean_value for simplicity
      for (int c = 1; c < channels; ++c) {
        mean_values_.push_back(mean_values_[0]);
      }
    }
  }