Caffe源码解读（十二）：自定义数据输入层

第1，3，4，5步跟上一节的自定义神经层的一样。
数据输入层需要重写三个函数：
1. DataLayerSetUp：定义好从prototxt读入的参数名和容器的规格（设好N,K,H,W）
2. ShuffleImages：打乱顺序
3. load_batch：把图片读入到内存

代码及解读如下：

#ifdef USE_OPENCV#include <opencv2/core/core.hpp>#include <fstream>  // NOLINT(readability/streams)#include <iostream>  // NOLINT(readability/streams)#include <string>#include <utility>#include <vector>#include "caffe/data_transformer.hpp"#include "caffe/layers/base_data_layer.hpp"#include "caffe/layers/image_data_layer.hpp"#include "caffe/util/benchmark.hpp"#include "caffe/util/io.hpp"#include "caffe/util/math_functions.hpp"#include "caffe/util/rng.hpp"namespace caffe {template <typename Dtype>ImageDataLayer<Dtype>::~ImageDataLayer<Dtype>() {  this->StopInternalThread();}//DataLayerSetUp：定义好从prototxt读入的参数名和容器的规格（设好N, K, H, W）template <typename Dtype>void ImageDataLayer<Dtype>::DataLayerSetUp(const vector<Blob<Dtype>*>& bottom,      const vector<Blob<Dtype>*>& top) {    /*        读取在prototxt里面的设置数据:            1、caffe.proto中的LayerParameter中定义了ImageDataParameter类型的image_data_param变量            2、ImageDataParameter类中，定义了new_height、new_width、is_color、root_folder    */    //layer_param_是Layer类中定义的protected变量;    //Layer类在layer.hpp中定义，Layer没有继承任何其他类;    //Layer类中定义了LayerParameter类型的变量layer_param_;    //LayerParameter在caffe.proto定义，LayerParameter中定义了“optional ImageDataParameter image_data_param = 115;”    //ImageDataParameter也在caffe.proto中定义，ImageDataParameter中定义了new_height、new_width、is_color、root_folder  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.";  // Read the file with filenames and labels    /*        读取lmdb文件，并把data和label配对，存储到lines_里面，lines_在image_data_layer.hpp由我们自己定义，不是由prototxt生成。    */    //source跟new_height、new_width一样在ImageDataParameter中定义  const string& source = this->layer_param_.image_data_param().source();  LOG(INFO) << "Opening file " << source;    //读取lmdb文件    //string类的c_str()函数，返回string的内含字符串，创建一个stream类对象infile，从硬盘读数据到内存。  std::ifstream infile(source.c_str());  string line;  size_t pos;  int label;  while (std::getline(infile, line)) {    pos = line.find_last_of(' ');   //find_last_of：查找最近一个空格的位置。每一行的data和label是由空格分开，pos就是空格的位置。    label = atoi(line.substr(pos + 1).c_str()); //substr：取子字符串，从pos+1（也就是label的首字母）到行尾。取label的字符串表示。    lines_.push_back(std::make_pair(line.substr(0, pos), label));   //lines_在ImageDataLayer类中由自己定义的变量;                                                                    //类型为vector<std::pair<std::string, int> >：std::pair主要的作用是将两个数据组合成一个数据                                                                    //make_pair:生成pair对象                                                                    //push_back：vector的函数，把变量装入vector中。  }  CHECK(!lines_.empty()) << "File is empty";    /*        使用shuffle打乱顺序    */  if (this->layer_param_.image_data_param().shuffle()) {    // randomly shuffle data    LOG(INFO) << "Shuffling data";    const unsigned int prefetch_rng_seed = caffe_rng_rand();    //生成一个随机数作为种子，caffe_rng_rand在math_functions.cpp中定义    prefetch_rng_.reset(new Caffe::RNG(prefetch_rng_seed));     //prefetch_rng_在ImageDataLayer类中由自己定义，类型为shared_ptr<Caffe::RNG>                                                                //shared_ptr请参见《笔记.doc》    ShuffleImages();    //本类的函数，根据prefetch_rng_随机数打乱容器lines_的顺序  }  LOG(INFO) << "A total of " << lines_.size() << " images.";    /*        使用skip随机跳过一些图片    */  lines_id_ = 0;    //lines_id_：由在ImageDataLayer类中由自己定义  // Check if we would need to randomly skip a few data points  if (this->layer_param_.image_data_param().rand_skip()) {  //rand_skip指定随机跳过的间隔，跟new_height、new_width一样在ImageDataParameter中定义    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;  }  // Read an image, and use it to initialize the top blob.    /*        代码核心：加载图片    */  cv::Mat cv_img = ReadImageToCVMat(root_folder + lines_[lines_id_].first,  //root_folder：根目录; lines_：类型为<data,label>的容器;first指的是data，也就是图片                                    new_height, new_width, is_color);       //图片的高、宽、通道数  CHECK(cv_img.data) << "Could not load " << lines_[lines_id_].first;  // Use data_transformer to infer the expected blob shape from a cv_image.  vector<int> top_shape = this->data_transformer_->InferBlobShape(cv_img);  //data_transformer_：在BaseDataLayer类中定义，类型为shared_ptr<DataTransformer<Dtype> >                                                                            //DataTransformer类：在data_transformer.hpp中定义，作用是将常用变换应用于输入数据，例如缩放，镜像，减去图像平均值。                                                                            //InferBlobShape函数：在DataTransformer类中定义，推断Blob的shape                                                                            //详情见http://blog.csdn.net/xizero00/article/details/50905685                                                                            //top_shape：以mnist的图片为例，top_shape的值将是：[1,1,28,28]。即1张图，单通道，高28,宽28;  this->transformed_data_.Reshape(top_shape);   //更改Blob的维度大小到图片的大小  // Reshape prefetch_data and top[0] according to the batch_size.  const int batch_size = this->layer_param_.image_data_param().batch_size();    //读取batch_size大小，  CHECK_GT(batch_size, 0) << "Positive batch size required";  top_shape[0] = batch_size;    //设置第一个维度的大小为batch_size，即每次迭代有batch_size个图片  for (int i = 0; i < this->PREFETCH_COUNT; ++i) {  //PREFETCH_COUNT：静态变量，预取的batch数，默认为3    this->prefetch_[i].data_.Reshape(top_shape);    //把prefetch_的data_做reshape到top_shape大小                                                    //prefetch_[PREFETCH_COUNT]：类型Batch<Dtype> ，Batch类只有两个public的变量data_和label_  }  top[0]->Reshape(top_shape);   //top即ImageDataLayer要输出的数据，由<data,label>组成，top[0]表示数据，top[1]表示label。data和label都是blob类型。  LOG(INFO) << "output data size: " << top[0]->num() << ","      << top[0]->channels() << "," << top[0]->height() << ","      << top[0]->width();  // label  vector<int> label_shape(1, batch_size);   //label的空间尺寸：1表示1维空间  top[1]->Reshape(label_shape);  for (int i = 0; i < this->PREFETCH_COUNT; ++i) {    this->prefetch_[i].label_.Reshape(label_shape);  }}template <typename Dtype>void ImageDataLayer<Dtype>::ShuffleImages() {  caffe::rng_t* prefetch_rng =      static_cast<caffe::rng_t*>(prefetch_rng_->generator());   //generator()：随机数生成器  shuffle(lines_.begin(), lines_.end(), prefetch_rng);  //shuffle：根据随机数打乱vector的顺序，为什么}// This function is called on prefetch thread//把图片读到内存template <typename Dtype>void ImageDataLayer<Dtype>::load_batch(Batch<Dtype>* batch) {  CPUTimer batch_timer;  batch_timer.Start();  double read_time = 0;  double trans_time = 0;  CPUTimer timer;  CHECK(batch->data_.count());  CHECK(this->transformed_data_.count());  //读取prototxt的文件配置，和SetUp函数操作一致。  ImageDataParameter image_data_param = this->layer_param_.image_data_param();  const int batch_size = image_data_param.batch_size();  const int new_height = image_data_param.new_height();  const int new_width = image_data_param.new_width();  const bool is_color = image_data_param.is_color();  string root_folder = image_data_param.root_folder();  // Reshape according to the first image of each batch  // on single input batches allows for inputs of varying dimension.  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;  // Use data_transformer to infer the expected blob shape from a cv_img.  vector<int> top_shape = this->data_transformer_->InferBlobShape(cv_img);  //计算cv_img大小，和SetUp函数操作一致。  this->transformed_data_.Reshape(top_shape);   //设置blob空间尺寸，和SetUp函数操作一致。  // Reshape batch according to the batch_size.  top_shape[0] = batch_size;    //batch_size张图片，和SetUp函数操作一致。  batch->data_.Reshape(top_shape);  //batch类只有两个public的变量data_和label_，都为Blob类型。  Dtype* prefetch_data = batch->data_.mutable_cpu_data();   //mutable_cpu_data()返回data_的地址  Dtype* prefetch_label = batch->label_.mutable_cpu_data();  // datum scales  const int lines_size = lines_.size(); //样本个数  for (int item_id = 0; item_id < batch_size; ++item_id) {    // get a blob    timer.Start();    CHECK_GT(lines_size, lines_id_);    cv::Mat cv_img = ReadImageToCVMat(root_folder + lines_[lines_id_].first,    //读取第lines_id_个样本的数据，转化为Mat型        new_height, new_width, is_color);    CHECK(cv_img.data) << "Could not load " << lines_[lines_id_].first;    read_time += timer.MicroSeconds();    timer.Start();    // Apply transformations (mirror, crop...) to the image    int offset = batch->data_.offset(item_id);  //获取item_id个图像数据的偏移量    this->transformed_data_.set_cpu_data(prefetch_data + offset);   //set_cpu_data指定数据地址为prefetch_data    this->data_transformer_->Transform(cv_img, &(this->transformed_data_)); //把cv_img数据转换到transformed_data_指定的data地址prefetch_data + offset    trans_time += timer.MicroSeconds();    prefetch_label[item_id] = lines_[lines_id_].second;    // go to the next iter    lines_id_++;    if (lines_id_ >= lines_size) {      // We have reached the end. Restart from the first.      DLOG(INFO) << "Restarting data prefetching from start.";      lines_id_ = 0;      if (this->layer_param_.image_data_param().shuffle()) {        ShuffleImages();      }    }  }  batch_timer.Stop();  DLOG(INFO) << "Prefetch batch: " << batch_timer.MilliSeconds() << " ms.";  DLOG(INFO) << "     Read time: " << read_time / 1000 << " ms.";  DLOG(INFO) << "Transform time: " << trans_time / 1000 << " ms.";}INSTANTIATE_CLASS(ImageDataLayer);REGISTER_LAYER_CLASS(ImageData);}  // namespace caffe#endif  // USE_OPENCV

以ImageDataLayer层的使用：

    layer {        name: "data"        type: "ImageData"  //在ImageDataLayer.hpp中的type函数定义      top: "data"  //由这两个top可知，ImageDataLayer会定义top[0]和top[1]两个输出，top[0]是数据，top[1]是label      top: "label"        transform_param {  //定义了图像数据预处理的操作        mirror: false          crop_size: 227          mean_file: "data/ilsvrc12/imagenet_mean.binaryproto"        }        image_data_param {  //这是我们要定义的source        source: "examples/_temp/file_list.txt"          batch_size: 50          new_height: 256          new_width: 256        }      }