一、Data_layers.hpp文件的作用简介
Data_layers.hpp在目前caffe的master分支中已经不能存在了，分散到各个文件中去了。
而之前是存在于cafferoot\include\caffe中。现在已经变成了各个类的名称的头文件了。这里做个提醒
首先给出这个文件中所包含的几个与数据读取有关的类。
分别为：
BaseDataLayer
数据层的基类，继承自通用的类Layer
Batch
Batch实际上就是一个data_和label_类标
BasePrefetchingDataLayer
是预取层的基类，继承自BaseDataLayer和InternalThread，包含能够读取一批数据的能力
DataLayer
DataLayer才是主角，继承自BasePrefetchingDataLayer
使用DataReader来进行数据共享，从而实现并行化
DummyDataLayer
该类是继承自Layer,通过Filler产生数据
HDF5DataLayer
从HDF5中读取，继承自Layer
HDF5OutputLayer
将数据写入到HDF5文件，继承自Layer
ImageDataLayer
从图像文件中读取数据，这个应该比较常用，继承自BasePrefetchingDataLayer
MemoryDataLayer
从内存中读取数据，这里指已经从数据文件或者图像文件中读取到了数据，然后输入到该层，继承自BaseDataLayer
WindowDataLayer
从图像文件的窗口获取数据，需要指定窗口数据文件，继承自BasePrefetchingDataLayer
二、Data_layers文件的的详细介绍
上述类虽然在同一个头文件中进行的定义，但是却都是在不同的cpp文件进行的实现。
下面给出类的实现文件
BaseDataLayer和BasePrefetchingDataLayer
对应于：
base_data_layer.cpp
base_data_layer.cu
DataLayer
对应于：
data_layer.cpp
DummyDataLayer
对应于：
dummy_data_layer.cpp
HDF5DataLayer
HDF5OutputLayer
对应于：
hdf5_data_layer.cpp
hdf5_data_layer.cu
以及
hdf5_output_layer.cpp
hdf5_output_layer.cu
ImageDataLayer
对应于：
image_data_layer.cpp
MemoryDataLayer
对应于：
memory_data_layer.cpp
WindowDataLayer
对应于
window_data_layer.cpp
接下来对这些类进行详细阐述：
（1）BaseDataLayer的类定义以及实现如下：
/**
* @brief Provides base for data layers that feed blobs to the Net.
*
* TODO(dox): thorough documentation for Forward and proto params.
* 数据层的基类
*/
template <typename Dtype>
class BaseDataLayer : public Layer<Dtype> {
public:
// 显式构造函数
explicit BaseDataLayer(const LayerParameter& param);
// LayerSetUp: implements common data layer setup functionality, and calls
// DataLayerSetUp to do special data layer setup for individual layer types.
// This method may not be overridden except by the BasePrefetchingDataLayer.
// 该函数只能被BasePrefetchingDataLayer层进行重载
virtual void LayerSetUp(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top);
// Data layers should be shared by multiple solvers in parallel
// 数据是否需要给多个并行solver进行共享
virtual inline bool ShareInParallel() const { return true; }
// 数据层的初始化
virtual void DataLayerSetUp(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top) {}
// 数据层是没有输入的(即bottoms)，所以reshape只是形式
// Data layers have no bottoms, so reshaping is trivial.
virtual void Reshape(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top) {}
virtual void Backward_cpu(const vector<Blob<Dtype>*>& top,
const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom) {}
virtual void Backward_gpu(const vector<Blob<Dtype>*>& top,
const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom) {}
protected:
// 对输入的数据进行变换的参数，这其中包括是否需要mirror，是否需要crop
// 是否需要减去meanfile，是否需要scale
TransformationParameter transform_param_;
// 实际执行数据变换类的指针(一个Transform函数加上参数即可完成对数据的变换，参数是数据哈)
shared_ptr<DataTransformer<Dtype> > data_transformer_;
bool output_labels_;
};
具体的实现：
// 构造函数就是初始化数据变换参数
template <typename Dtype>
BaseDataLayer<Dtype>::BaseDataLayer(const LayerParameter& param)
: Layer<Dtype>(param),
transform_param_(param.transform_param()) {
}
// 初始化的时候根据top的大小来确定，如果是1表明只输出数据，而不输出类标
template <typename Dtype>
void BaseDataLayer<Dtype>::LayerSetUp(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top) {
if (top.size() == 1) {
output_labels_ = false;
} else {
output_labels_ = true;
}
// 初始化一个DataTransformer实例，便于对数据进行预处理
data_transformer_.reset(
new DataTransformer<Dtype>(transform_param_, this->phase_));
// 初始化种子
data_transformer_->InitRand();
// The subclasses should setup the size of bottom and top
// 执行数据层的初始化
DataLayerSetUp(bottom, top);
}（2）BasePrefetchingDataLayer类的定义以及实现如下：
BasePrefetchingDataLayer类的定义如下：
// BasePrefetchingDataLayer层是继承于BaseDataLayer的
// 是预取层的基类
template <typename Dtype>
class BasePrefetchingDataLayer :
public BaseDataLayer<Dtype>, public InternalThread {
public:
explicit BasePrefetchingDataLayer(const LayerParameter& param);
// LayerSetUp: implements common data layer setup functionality, and calls
// DataLayerSetUp to do special data layer setup for individual layer types.
// This method may not be overridden.
void LayerSetUp(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top);
virtual void Forward_cpu(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top);
virtual void Forward_gpu(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top);
// Prefetches batches (asynchronously if to GPU memory)
static const int PREFETCH_COUNT = 3;
protected:
virtual void InternalThreadEntry();
// 多了load_batch函数，该函数是纯虚函数，继承该函数的类都需要实现的
virtual void load_batch(Batch<Dtype>* batch) = 0;
// 还有prefetch数组,prefetch_free_,prefetch_full_
Batch<Dtype> prefetch_[PREFETCH_COUNT];
BlockingQueue<Batch<Dtype>*> prefetch_free_;
BlockingQueue<Batch<Dtype>*> prefetch_full_;
Blob<Dtype> transformed_data_;
};
BasePrefetchingDataLayer类的具体实现如下：
// 构造函数，初始化预取的队列,free和full
template <typename Dtype>
BasePrefetchingDataLayer<Dtype>::BasePrefetchingDataLayer(
const LayerParameter& param)
: BaseDataLayer<Dtype>(param),
prefetch_free_(), prefetch_full_() {
for (int i = 0; i < PREFETCH_COUNT; ++i) {
prefetch_free_.push(&prefetch_[i]);
}
}
// 进行层的初始化
template <typename Dtype>
void BasePrefetchingDataLayer<Dtype>::LayerSetUp(
const vector<Blob<Dtype>*>& bottom, const vector<Blob<Dtype>*>& top) {
// 首先执行基类BaseDataLayer的层初始化
BaseDataLayer<Dtype>::LayerSetUp(bottom, top);
// Before starting the prefetch thread, we make cpu_data and gpu_data
// calls so that the prefetch thread does not accidentally make simultaneous
// cudaMalloc calls when the main thread is running. In some GPUs this
// seems to cause failures if we do not so.
// 在开启预取线程的时候，需要让cpu数据和gpu数据分配空间
// 这样才能够避免在某些GPU上出现问题
// 首先是CPU
for (int i = 0; i < PREFETCH_COUNT; ++i) {
prefetch_[i].data_.mutable_cpu_data();
if (this->output_labels_) {
prefetch_[i].label_.mutable_cpu_data();
}
}
#ifndef CPU_ONLY
// 然后是GPU
if (Caffe::mode() == Caffe::GPU) {
for (int i = 0; i < PREFETCH_COUNT; ++i) {
prefetch_[i].data_.mutable_gpu_data();
if (this->output_labels_) {
prefetch_[i].label_.mutable_gpu_data();
}
}
}
#endif
DLOG(INFO) << "Initializing prefetch";
// 初始化随机数种子
this->data_transformer_->InitRand();
// 开启线程
StartInternalThread();
DLOG(INFO) << "Prefetch initialized.";
}
// 在StartInternalThread开启线程后就会执行下面自己定义的函数
// 这个就是自己定义的函数，让线程去执行的
template <typename Dtype>
void BasePrefetchingDataLayer<Dtype>::InternalThreadEntry() {
#ifndef CPU_ONLY
cudaStream_t stream;
if (Caffe::mode() == Caffe::GPU) {
// 创建非阻塞流
CUDA_CHECK(cudaStreamCreateWithFlags(&stream, cudaStreamNonBlocking));
}
#endif
try {
while (!must_stop()) {
// 弹出一个batch
Batch<Dtype>* batch = prefetch_free_.pop();
// 装载batch
load_batch(batch);
#ifndef CPU_ONLY
if (Caffe::mode() == Caffe::GPU) {
// 如果GPU模式开始，则推送到GPU
batch->data_.data().get()->async_gpu_push(stream);
// 检查是否成功
CUDA_CHECK(cudaStreamSynchronize(stream));
}
#endif
// 将装好的batch压入full队列
prefetch_full_.push(batch);
}
} catch (boost::thread_interrupted&) {
// Interrupted exception is expected on shutdown
}
#ifndef CPU_ONLY
if (Caffe::mode() == Caffe::GPU) {
// 销毁流
CUDA_CHECK(cudaStreamDestroy(stream));
}
#endif
}
template <typename Dtype>
void BasePrefetchingDataLayer<Dtype>::Forward_cpu(
const vector<Blob<Dtype>*>& bottom, const vector<Blob<Dtype>*>& top) {
// 传递的时候是从full队列中弹出一个数据
Batch<Dtype>* batch = prefetch_full_.pop("Data layer prefetch queue empty");
// Reshape to loaded data.
// 根据batch的形状改变数据形状
top[0]->ReshapeLike(batch->data_);
// Copy the data
// 将batch数据复制到top[0]
caffe_copy(batch->data_.count(), batch->data_.cpu_data(),
top[0]->mutable_cpu_data());
DLOG(INFO) << "Prefetch copied";
if (this->output_labels_) {
// 输出类标的话
// Reshape to loaded labels.
// 根据batch中类标的形状改变top[1]的形状
top[1]->ReshapeLike(batch->label_);
// Copy the labels.
// 复制类标到top[1]
caffe_copy(batch->label_.count(), batch->label_.cpu_data(),
top[1]->mutable_cpu_data());
}
// 将该batch压入free队列
prefetch_free_.push(batch);
}
// 如果没有GPU的话则在BasePrefetchingDataLayer类中生成一个Forward函数
// 该函数并不前传，而是直接报错
#ifdef CPU_ONLY
STUB_GPU_FORWARD(BasePrefetchingDataLayer, Forward);
#endif
// 初始化层
INSTANTIATE_CLASS(BaseDataLayer);
INSTANTIATE_CLASS(BasePrefetchingDataLayer);（3）DataLayer类的定义以及实现如下：
数据层的主要功能是：
首先给出类的定义
// DataLayer才是主角，继承自BasePrefetchingDataLayer
template <typename Dtype>
class DataLayer : public BasePrefetchingDataLayer<Dtype> {
public:
explicit DataLayer(const LayerParameter& param);
virtual ~DataLayer();
virtual void DataLayerSetUp(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top);
// DataLayer uses DataReader instead for sharing for parallelism
// 多了下面几个
virtual inline bool ShareInParallel() const { return false; }
virtual inline const char* type() const { return "Data"; }
virtual inline int ExactNumBottomBlobs() const { return 0; }
virtual inline int MinTopBlobs() const { return 1; }
virtual inline int MaxTopBlobs() const { return 2; }
protected:
virtual void load_batch(Batch<Dtype>* batch);
DataReader reader_;
};
具体的实现如下：
#ifdef USE_OPENCV
#include <opencv2/core/core.hpp>
#endif // USE_OPENCV
#include <stdint.h>
#include <string>
#include <vector>
#include "caffe/common.hpp"
#include "caffe/data_layers.hpp"
#include "caffe/layer.hpp"
#include "caffe/proto/caffe.pb.h"
#include "caffe/util/benchmark.hpp"
#include "caffe/util/io.hpp"
namespace caffe {
// 初始化DataReader，层参数
template <typename Dtype>
DataLayer<Dtype>::DataLayer(const LayerParameter& param)
: BasePrefetchingDataLayer<Dtype>(param),
reader_(param) {
}
// 析构函数停止内部线程
template <typename Dtype>
DataLayer<Dtype>::~DataLayer() {
this->StopInternalThread();
}
// 数据层的初始化
template <typename Dtype>
void DataLayer<Dtype>::DataLayerSetUp(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top) {
// 从层参数中读取batch_size
const int batch_size = this->layer_param_.data_param().batch_size();
// Read a data point, and use it to initialize the top blob.
// 从reader_中获取一个数据
Datum& datum = *(reader_.full().peek());
// Use data_transformer to infer the expected blob shape from datum.
// 用数据来推断blob的形状存放到top_shape
vector<int> top_shape = this->data_transformer_->InferBlobShape(datum);
this->transformed_data_.Reshape(top_shape);
// Reshape top[0] and prefetch_data according to the batch_size.
// 既然获取了数据的形状(channel,height,width)，那么这里再设置一下batch_size
// top_shape[0]=batch_size
// top_shape[1]=channel
// top_shape[2]=height
// top_shape[3]=width
top_shape[0] = batch_size;
// 根据形状设置top[0]的形状
top[0]->Reshape(top_shape);
// 设置预取数据的形状
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();
// label
// 如果输出类标的话则把top[1]的形状也弄一下
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);
}
}
}
// This function is called on prefetch thread
// 这个函数是在自己定义的线程执行函数内部执行的
template<typename Dtype>
void DataLayer<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());
// Reshape according to the first datum of each batch
// on single input batches allows for inputs of varying dimension.
// 意思是像以下这种做法这样的话，每个batch的数据的维度可以不一样
// 从参数文件获取batch_size
const int batch_size = this->layer_param_.data_param().batch_size();
// 获取第一个数据
Datum& datum = *(reader_.full().peek());
// Use data_transformer to infer the expected blob shape from datum.
// 使用第一个数据推断blob的形状
vector<int> top_shape = this->data_transformer_->InferBlobShape(datum);
this->transformed_data_.Reshape(top_shape);
// Reshape batch according to the batch_size.
top_shape[0] = batch_size;
batch->data_.Reshape(top_shape);
// top_data存数据
Dtype* top_data = batch->data_.mutable_cpu_data();
Dtype* top_label = NULL; // suppress warnings about uninitialized variables
// top_label存类标
if (this->output_labels_) {
top_label = batch->label_.mutable_cpu_data();
}
// 对这批数据进行处理
for (int item_id = 0; item_id < batch_size; ++item_id) {
timer.Start();
// get a datum
Datum& datum = *(reader_.full().pop("Waiting for data"));
read_time += timer.MicroSeconds();
timer.Start();
// Apply data transformations (mirror, scale, crop...)
// 对于给定批的数据获取offset，这里调用的是给定batchid，然后获取offset
int offset = batch->data_.offset(item_id);
this->transformed_data_.set_cpu_data(top_data + offset);
this->data_transformer_->Transform(datum, &(this->transformed_data_));
// Copy label.
// 复制类标
if (this->output_labels_) {
top_label[item_id] = datum.label();
}
// 数据传输时间
trans_time += timer.MicroSeconds();
// 将数据指针压到free队列
reader_.free().push(const_cast<Datum*>(&datum));
}
timer.Stop();
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(DataLayer);
REGISTER_LAYER_CLASS(Data);
} // namespace caffe
（4）DummyDataLayer类的定义与实现介绍：
Dummy数据层的主要功能就是根据所给定的Filler产生数据，然后前向传
首先给出定义
/**
* @brief Provides data to the Net generated by a Filler.
*
* TODO(dox): thorough documentation for Forward and proto params.
* 该类是继承自Layer,通过Filler产生数据
*/
template <typename Dtype>
class DummyDataLayer : public Layer<Dtype> {
public:
explicit DummyDataLayer(const LayerParameter& param)
: Layer<Dtype>(param) {}
virtual void LayerSetUp(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top);
// Data layers should be shared by multiple solvers in parallel
virtual inline bool ShareInParallel() const { return true; }
// Data layers have no bottoms, so reshaping is trivial.
virtual void Reshape(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top) {}
virtual inline const char* type() const { return "DummyData"; }
virtual inline int ExactNumBottomBlobs() const { return 0; }
virtual inline int MinTopBlobs() const { return 1; }
protected:
virtual void Forward_cpu(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top);
virtual void Backward_cpu(const vector<Blob<Dtype>*>& top,
const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom) {}
virtual void Backward_gpu(const vector<Blob<Dtype>*>& top,
const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom) {}
vector<shared_ptr<Filler<Dtype> > > fillers_;
vector<bool> refill_;
};
接下来给出详细的定义：
首先给出FillerParameter的定义，里面指定了值的类型，值是啥，最小是啥，最大是啥，平均值、方差是啥、是否稀疏、以及将扇入个数还是扇出个数还是所有的加起来求均值作为分母
message FillerParameter {
// The filler type.
optional string type = 1 [default = 'constant'];
optional float value = 2 [default = 0]; // the value in constant filler
optional float min = 3 [default = 0]; // the min value in uniform filler
optional float max = 4 [default = 1]; // the max value in uniform filler
optional float mean = 5 [default = 0]; // the mean value in Gaussian filler
optional float std = 6 [default = 1]; // the std value in Gaussian filler
// The expected number of non-zero output weights for a given input in
// Gaussian filler -- the default -1 means don't perform sparsification.
optional int32 sparse = 7 [default = -1];
// Normalize the filler variance by fan_in, fan_out, or their average.
// Applies to 'xavier' and 'msra' fillers.
enum VarianceNorm {
FAN_IN = 0;
FAN_OUT = 1;
AVERAGE = 2;
}
optional VarianceNorm variance_norm = 8 [default = FAN_IN];
}
再看看该类的参数
</pre><pre name="code" class="plain">// DummyDataLayer fills any number of arbitrarily shaped blobs with random
// (or constant) data generated by "Fillers" (see "message FillerParameter").
message DummyDataParameter {
// This layer produces N >= 1 top blobs. DummyDataParameter must specify 1 or N
// shape fields, and 0, 1 or N data_fillers.
//
// If 0 data_fillers are specified, ConstantFiller with a value of 0 is used.
// If 1 data_filler is specified, it is applied to all top blobs. If N are
// specified, the ith is applied to the ith top blob.
repeated FillerParameter data_filler = 1;
repeated BlobShape shape = 6;
// 4D dimensions -- deprecated. Use "shape" instead.
repeated uint32 num = 2;
repeated uint32 channels = 3;
repeated uint32 height = 4;
repeated uint32 width = 5;
}
接下来给出具体的实现
#include <vector>
#include "caffe/filler.hpp"
#include "caffe/layer.hpp"
#include "caffe/vision_layers.hpp"
namespace caffe {
template <typename Dtype>
void DummyDataLayer<Dtype>::LayerSetUp(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top) {
// 输出有几个
const int num_top = top.size();
// 获取该层的参数
const DummyDataParameter& param = this->layer_param_.dummy_data_param();
// 有几个filler
const int num_data_filler = param.data_filler_size();
// 检查filler的个数，要么为0、1、或者等于输出的个数
CHECK(num_data_filler == 0 || num_data_filler == 1 ||
num_data_filler == num_top)
<< "Number of data fillers must be 0, 1 or equal to the number of tops: "
<< num_top << "; you specified " << num_data_filler << " data fillers.";
// 判断是否全部为0
const bool legacy_dims = param.num_size() || param.channels_size() ||
param.height_size() || param.width_size();
// 下面就是检查参数是不是满足要求，1或者0或者等于num_top
if (legacy_dims) {// 如果不是全部为0
CHECK_EQ(0, param.shape_size())
<< "Both shape and legacy fields were specified";
// Using deprecated 4D output dim specifiers.
CHECK(param.num_size() == 1 || param.num_size() == num_top)
<< "Must specify 'num' once, or once per top blob "
<< "(" << num_top << "); specified " << param.num_size() << ".";
CHECK(param.channels_size() == 1 || param.channels_size() == num_top)
<< "Must specify 'channels' once, or once per top blob "
<< "(" << num_top << "); specified " << param.channels_size() << ".";
CHECK(param.height_size() == 1 || param.height_size() == num_top)
<< "Must specify 'height' once, or once per top blob "
<< "(" << num_top << "); specified " << param.height_size() << ".";
CHECK(param.width_size() == 1 || param.width_size() == num_top)
<< "Must specify 'width' once, or once per top blob "
<< "(" << num_top << "); specified " << param.width_size() << ".";
} else {
CHECK(param.shape_size() == 1 || param.shape_size() == num_top)
<< "Must specify 'shape' once, or once per top blob "
<< "(" << num_top << "); specified " << param.shape_size() << ".";
}
// refill_[i] tells Forward i whether or not to actually refill top Blob i.
// If refill_[i] is false, Forward does nothing for Blob i. We use this to
// avoid wastefully refilling "constant" Blobs in every forward pass.
// We first fill refill_ in with the INVERSE of its final values.
// The first time we run Forward from the LayerSetUp method, we'll fill only
// Blobs for which refill_ is normally false. These Blobs will never be
// filled again.
// refill_表明是不是需要填充Blob，如果refill_[i]=false，那么就不会Blob i做任何事
//
refill_.clear();
fillers_.clear();
// 要么是0，要么是1
if (num_data_filler <= 1) {
// 定义了生成数据的参数
// 比如均值、方差等，详细请看其定义
FillerParameter filler_param;
if (num_data_filler == 0) {
// 如果没有指定，那么就是常数值填充
filler_param.set_type("constant");
filler_param.set_value(0);
} else {
// 否则复制filler到filler_param
filler_param.CopyFrom(param.data_filler(0));
}
// Refill on each iteration iff not using a constant filler,
// but use the inverse of this rule for the first run.
// 如果
refill_.resize(1);
refill_[0] = (strcmp(filler_param.type().c_str(), "constant") == 0);
fillers_.resize(1);
// 实例化填充器
fillers_[0].reset(GetFiller<Dtype>(filler_param));
} else {// 如果等于=num_top
refill_.resize(num_top);
fillers_.resize(num_top);
for (int i = 0; i < num_top; ++i) {
fillers_[i].reset(GetFiller<Dtype>(param.data_filler(i)));
// Refill on each iteration iff not using a constant filler,
// but use the inverse of this rule for the first run.
refill_[i] =
(strcmp(param.data_filler(i).type().c_str(), "constant") == 0);
}
}
// 改变形状
for (int i = 0; i < num_top; ++i) {
if (legacy_dims) {
const int num = (param.num_size() == 1) ? param.num(0) : param.num(i);
const int channels =
(param.channels_size() == 1) ? param.channels(0) : param.channels(i);
const int height =
(param.height_size() == 1) ? param.height(0) : param.height(i);
const int width =
(param.width_size() == 1) ? param.width(0) : param.width(i);
top[i]->Reshape(num, channels, height, width);
} else {
const int shape_index = (param.shape_size() == 1) ? 0 : i;
top[i]->Reshape(param.shape(shape_index));
}
}
// Run Forward once, with refill_ inverted, to fill the constant Blobs.
// 执行forward_cpu
this->Forward(bottom, top);
// Invert the inverted refill_ values to refill the desired (non-constant)
// Blobs in every usual forward pass.
for (int i = 0; i < refill_.size(); ++i) {
refill_[i] = !refill_[i];
}
}
// Forward里调用了该函数
template <typename Dtype>
void DummyDataLayer<Dtype>::Forward_cpu(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top) {
// 调用fillers_来进行錐ill
for (int i = 0; i < top.size(); ++i) {
const int filler_id = (fillers_.size() > 1) ? i : 0;
if (refill_[filler_id]) {
fillers_[filler_id]->Fill(top[i]);
}
}
}
// 初始化类
// 注册类
INSTANTIATE_CLASS(DummyDataLayer);
REGISTER_LAYER_CLASS(DummyData);
} // namespace caffe（5）HDF5DataLayer类的定义以及实现如下：
HDF5数据层的主要功能是从给定的HDF5文件列表读取数据，然后设置top，即向前传播的数据。
首先给出类的定义：
template <typename Dtype>
class HDF5DataLayer : public Layer<Dtype> {
public:
explicit HDF5DataLayer(const LayerParameter& param)
: Layer<Dtype>(param) {}
virtual ~HDF5DataLayer();
virtual void LayerSetUp(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top);
// Data layers should be shared by multiple solvers in parallel
virtual inline bool ShareInParallel() const { return true; }
// Data layers have no bottoms, so reshaping is trivial.
virtual void Reshape(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top) {}
virtual inline const char* type() const { return "HDF5Data"; }
virtual inline int ExactNumBottomBlobs() const { return 0; }
virtual inline int MinTopBlobs() const { return 1; }
protected:
virtual void Forward_cpu(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top);
virtual void Forward_gpu(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top);
virtual void Backward_cpu(const vector<Blob<Dtype>*>& top,
const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom) {}
virtual void Backward_gpu(const vector<Blob<Dtype>*>& top,
const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom) {}
// 从HDF5文件读取数据
virtual void LoadHDF5FileData(const char* filename);
std::vector<std::string> hdf_filenames_;
unsigned int num_files_;
unsigned int current_file_;
hsize_t current_row_;
std::vector<shared_ptr<Blob<Dtype> > > hdf_blobs_;
// 存放的是数据的索引,可以对索引进行shuffle
std::vector<unsigned int> data_permutation_;
// 存放的是文件名字的索引，可以对索引进行shuffle
std::vector<unsigned int> file_permutation_;
};
接下来给出类的具体实现：
给出实现之前先给出HDF5的操作
头文件：
#ifndef CAFFE_UTIL_HDF5_H_
#define CAFFE_UTIL_HDF5_H_
#include <string>
#include "hdf5.h"
#include "hdf5_hl.h"
#include "caffe/blob.hpp"
namespace caffe {
// 获取HDF5文件的信息以及数据的维度
template <typename Dtype>
void hdf5_load_nd_dataset_helper(
hid_t file_id, const char* dataset_name_, int min_dim, int max_dim,
Blob<Dtype>* blob);
// float类型的获取数据维度和信息的包裹函数
template <typename Dtype>
void hdf5_load_nd_dataset(
hid_t file_id, const char* dataset_name_, int min_dim, int max_dim,
Blob<Dtype>* blob);
// double类型的获取数据维度和信息的包裹函数
template <typename Dtype>
void hdf5_save_nd_dataset(
const hid_t file_id, const string& dataset_name, const Blob<Dtype>& blob,
bool write_diff = false);
// 读取int和存储int，读取字符串和存储字符串到文件
int hdf5_load_int(hid_t loc_id, const string& dataset_name);
void hdf5_save_int(hid_t loc_id, const string& dataset_name, int i);
string hdf5_load_string(hid_t loc_id, const string& dataset_name);
void hdf5_save_string(hid_t loc_id, const string& dataset_name,
const string& s);
// 获取链接数
int hdf5_get_num_links(hid_t loc_id);
// 根据名字找到索引
string hdf5_get_name_by_idx(hid_t loc_id, int idx);
} // namespace caffe
#endif // CAFFE_UTIL_HDF5_H_
cpp文件：
#include "caffe/util/hdf5.hpp"
#include <string>
#include <vector>
namespace caffe {
// Verifies format of data stored in HDF5 file and reshapes blob accordingly.
// 获取HDF5文件的信息以及数据的维度
template <typename Dtype>
void hdf5_load_nd_dataset_helper(
hid_t file_id, const char* dataset_name_, int min_dim, int max_dim,
Blob<Dtype>* blob) {
// Verify that the dataset exists.
// 检查是否存在
CHECK(H5LTfind_dataset(file_id, dataset_name_))
<< "Failed to find HDF5 dataset " << dataset_name_;
// Verify that the number of dimensions is in the accepted range.
herr_t status;
int ndims;
// 获取数据维度
status = H5LTget_dataset_ndims(file_id, dataset_name_, &ndims);
CHECK_GE(status, 0) << "Failed to get dataset ndims for " << dataset_name_;
CHECK_GE(ndims, min_dim);
CHECK_LE(ndims, max_dim);
// Verify that the data format is what we expect: float or double.
std::vector<hsize_t> dims(ndims);
H5T_class_t class_;
// 获取数据信息
status = H5LTget_dataset_info(
file_id, dataset_name_, dims.data(), &class_, NULL);
CHECK_GE(status, 0) << "Failed to get dataset info for " << dataset_name_;
switch (class_) {
case H5T_FLOAT:
LOG_FIRST_N(INFO, 1) << "Datatype class: H5T_FLOAT";
break;
case H5T_INTEGER:
LOG_FIRST_N(INFO, 1) << "Datatype class: H5T_INTEGER";
break;
case H5T_TIME:
LOG(FATAL) << "Unsupported datatype class: H5T_TIME";
case H5T_STRING:
LOG(FATAL) << "Unsupported datatype class: H5T_STRING";
case H5T_BITFIELD:
LOG(FATAL) << "Unsupported datatype class: H5T_BITFIELD";
case H5T_OPAQUE:
LOG(FATAL) << "Unsupported datatype class: H5T_OPAQUE";
case H5T_COMPOUND:
LOG(FATAL) << "Unsupported datatype class: H5T_COMPOUND";
case H5T_REFERENCE:
LOG(FATAL) << "Unsupported datatype class: H5T_REFERENCE";
case H5T_ENUM:
LOG(FATAL) << "Unsupported datatype class: H5T_ENUM";
case H5T_VLEN:
LOG(FATAL) << "Unsupported datatype class: H5T_VLEN";
case H5T_ARRAY:
LOG(FATAL) << "Unsupported datatype class: H5T_ARRAY";
default:
LOG(FATAL) << "Datatype class unknown";
}
// 设置blob的维度
vector<int> blob_dims(dims.size());
for (int i = 0; i < dims.size(); ++i) {
blob_dims[i] = dims[i];
}
blob->Reshape(blob_dims);
}
// float类型的获取数据维度和信息的包裹函数
template <>
void hdf5_load_nd_dataset<float>(hid_t file_id, const char* dataset_name_,
int min_dim, int max_dim, Blob<float>* blob) {
hdf5_load_nd_dataset_helper(file_id, dataset_name_, min_dim, max_dim, blob);
herr_t status = H5LTread_dataset_float(
file_id, dataset_name_, blob->mutable_cpu_data());
CHECK_GE(status, 0) << "Failed to read float dataset " << dataset_name_;
}
// double类型的获取数据维度和信息的包裹函数
template <>
void hdf5_load_nd_dataset<double>(hid_t file_id, const char* dataset_name_,
int min_dim, int max_dim, Blob<double>* blob) {
hdf5_load_nd_dataset_helper(file_id, dataset_name_, min_dim, max_dim, blob);
herr_t status = H5LTread_dataset_double(
file_id, dataset_name_, blob->mutable_cpu_data());
CHECK_GE(status, 0) << "Failed to read double dataset " << dataset_name_;
}
// 存放float类型到hdf5文件
template <>
void hdf5_save_nd_dataset<float>(
const hid_t file_id, const string& dataset_name, const Blob<float>& blob,
bool write_diff) {
// blob信息放到dims
int num_axes = blob.num_axes();
hsize_t *dims = new hsize_t[num_axes];
for (int i = 0; i < num_axes; ++i) {
dims[i] = blob.shape(i);
}
// 获取数据指针
const float* data;
if (write_diff) {
data = blob.cpu_diff();
} else {
data = blob.cpu_data();
}
// 存放数据到hdf5
herr_t status = H5LTmake_dataset_float(
file_id, dataset_name.c_str(), num_axes, dims, data);
CHECK_GE(status, 0) << "Failed to make float dataset " << dataset_name;
delete[] dims;
}
// 存放double类型到hdf5文件
template <>
void hdf5_save_nd_dataset<double>(
hid_t file_id, const string& dataset_name, const Blob<double>& blob,
bool write_diff) {
int num_axes = blob.num_axes();
hsize_t *dims = new hsize_t[num_axes];
for (int i = 0; i < num_axes; ++i) {
dims[i] = blob.shape(i);
}
const double* data;
if (write_diff) {
data = blob.cpu_diff();
} else {
data = blob.cpu_data();
}
herr_t status = H5LTmake_dataset_double(
file_id, dataset_name.c_str(), num_axes, dims, data);
CHECK_GE(status, 0) << "Failed to make double dataset " << dataset_name;
delete[] dims;
}
// 读取string到字符串
string hdf5_load_string(hid_t loc_id, const string& dataset_name) {
// Get size of dataset
size_t size;
H5T_class_t class_;
herr_t status = \
H5LTget_dataset_info(loc_id, dataset_name.c_str(), NULL, &class_, &size);
CHECK_GE(status, 0) << "Failed to get dataset info for " << dataset_name;
char *buf = new char[size];
status = H5LTread_dataset_string(loc_id, dataset_name.c_str(), buf);
CHECK_GE(status, 0)
<< "Failed to load int dataset with name " << dataset_name;
string val(buf);
delete[] buf;
return val;
}
// 保存string到字符串
void hdf5_save_string(hid_t loc_id, const string& dataset_name,
const string& s) {
herr_t status = \
H5LTmake_dataset_string(loc_id, dataset_name.c_str(), s.c_str());
CHECK_GE(status, 0)
<< "Failed to save string dataset with name " << dataset_name;
}
// 载入int类型
int hdf5_load_int(hid_t loc_id, const string& dataset_name) {
int val;
herr_t status = H5LTread_dataset_int(loc_id, dataset_name.c_str(), &val);
CHECK_GE(status, 0)
<< "Failed to load int dataset with name " << dataset_name;
return val;
}
// 存储int类型
void hdf5_save_int(hid_t loc_id, const string& dataset_name, int i) {
hsize_t one = 1;
herr_t status = \
H5LTmake_dataset_int(loc_id, dataset_name.c_str(), 1, &one, &i);
CHECK_GE(status, 0)
<< "Failed to save int dataset with name " << dataset_name;
}
// 获取链接数
int hdf5_get_num_links(hid_t loc_id) {
H5G_info_t info;
herr_t status = H5Gget_info(loc_id, &info);
CHECK_GE(status, 0) << "Error while counting HDF5 links.";
return info.nlinks;
}
// 通过名字找到索引
string hdf5_get_name_by_idx(hid_t loc_id, int idx) {
ssize_t str_size = H5Lget_name_by_idx(
loc_id, ".", H5_INDEX_NAME, H5_ITER_NATIVE, idx, NULL, 0, H5P_DEFAULT);
CHECK_GE(str_size, 0) << "Error retrieving HDF5 dataset at index " << idx;
char *c_str = new char[str_size+1];
ssize_t status = H5Lget_name_by_idx(
loc_id, ".", H5_INDEX_NAME, H5_ITER_NATIVE, idx, c_str, str_size+1,
H5P_DEFAULT);
CHECK_GE(status, 0) << "Error retrieving HDF5 dataset at index " << idx;
string result(c_str);
delete[] c_str;
return result;
}
} // namespace caffe
给出具体实现：
/*
TODO:
- load file in a separate thread ("prefetch")
- can be smarter about the memcpy call instead of doing it row-by-row
:: use util functions caffe_copy, and Blob->offset()
:: don't forget to update hdf5_daa_layer.cu accordingly
- add ability to shuffle filenames if flag is set
*/
#include <fstream> // NOLINT(readability/streams)
#include <string>
#include <vector>
#include "hdf5.h"
#include "hdf5_hl.h"
#include "stdint.h"
#include "caffe/data_layers.hpp"
#include "caffe/layer.hpp"
#include "caffe/util/hdf5.hpp"
namespace caffe {
template <typename Dtype>
HDF5DataLayer<Dtype>::~HDF5DataLayer<Dtype>() { }
// Load data and label from HDF5 filename into the class property blobs.
// 读取HDF5文件数据到hdf_blobs
template <typename Dtype>
void HDF5DataLayer<Dtype>::LoadHDF5FileData(const char* filename) {
DLOG(INFO) << "Loading HDF5 file: " << filename;
// 打开文件
hid_t file_id = H5Fopen(filename, H5F_ACC_RDONLY, H5P_DEFAULT);
if (file_id < 0) {
LOG(FATAL) << "Failed opening HDF5 file: " << filename;
}
int top_size = this->layer_param_.top_size();
hdf_blobs_.resize(top_size);
const int MIN_DATA_DIM = 1;
const int MAX_DATA_DIM = INT_MAX;
for (int i = 0; i < top_size; ++i) {
hdf_blobs_[i] = shared_ptr<Blob<Dtype> >(new Blob<Dtype>());
// message LayerParameter {
// optional string name = 1; // the layer name
// optional string type = 2; // the layer type
// repeated string bottom = 3; // the name of each bottom blob
// repeated string top = 4; // the name of each top blob
hdf5_load_nd_dataset(file_id, this->layer_param_.top(i).c_str(),
MIN_DATA_DIM, MAX_DATA_DIM, hdf_blobs_[i].get());
}
herr_t status = H5Fclose(file_id);
CHECK_GE(status, 0) << "Failed to close HDF5 file: " << filename;
// MinTopBlobs==1 guarantees at least one top blob
CHECK_GE(hdf_blobs_[0]->num_axes(), 1) << "Input must have at least 1 axis.";
const int num = hdf_blobs_[0]->shape(0);
for (int i = 1; i < top_size; ++i) {
CHECK_EQ(hdf_blobs_[i]->shape(0), num);
}
// Default to identity permutation.
data_permutation_.clear();
data_permutation_.resize(hdf_blobs_[0]->shape(0));
for (int i = 0; i < hdf_blobs_[0]->shape(0); i++)
data_permutation_[i] = i;
// Shuffle if needed.
// 将数据索引映射表进行shuffle
if (this->layer_param_.hdf5_data_param().shuffle()) {
std::random_shuffle(data_permutation_.begin(), data_permutation_.end());
DLOG(INFO) << "Successully loaded " << hdf_blobs_[0]->shape(0)
<< " rows (shuffled)";
} else {
DLOG(INFO) << "Successully loaded " << hdf_blobs_[0]->shape(0) << " rows";
}
}
// 主要的功能就是读取HDF5文件，并且设置top blob的形状
template <typename Dtype>
void HDF5DataLayer<Dtype>::LayerSetUp(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top) {
// Refuse transformation parameters since HDF5 is totally generic.
CHECK(!this->layer_param_.has_transform_param()) <<
this->type() << " does not transform data.";
// Read the source to parse the filenames.
// 读取HDF列表文件
const string& source = this->layer_param_.hdf5_data_param().source();
LOG(INFO) << "Loading list of HDF5 filenames from: " << source;
hdf_filenames_.clear();
std::ifstream source_file(source.c_str());
if (source_file.is_open()) {
std::string line;
while (source_file >> line) {
hdf_filenames_.push_back(line);
}
} else {
LOG(FATAL) << "Failed to open source file: " << source;
}
source_file.close();
num_files_ = hdf_filenames_.size();
current_file_ = 0;
LOG(INFO) << "Number of HDF5 files: " << num_files_;
CHECK_GE(num_files_, 1) << "Must have at least 1 HDF5 filename listed in "
<< source;
file_permutation_.clear();
file_permutation_.resize(num_files_);
// 文件名字是否shuffle
// Default to identity permutation.
for (int i = 0; i < num_files_; i++) {
file_permutation_[i] = i;
}
// Shuffle if needed.
if (this->layer_param_.hdf5_data_param().shuffle()) {
std::random_shuffle(file_permutation_.begin(), file_permutation_.end());
}
// Load the first HDF5 file and initialize the line counter.
// 从给定的文件名列表中的第一个文件名读取数据到hdf_blobs
LoadHDF5FileData(hdf_filenames_[file_permutation_[current_file_]].c_str());
// 设置行指针
current_row_ = 0;
// Reshape blobs.
// 根据读取的hdf_blobs形状改变top的形状
const int batch_size = this->layer_param_.hdf5_data_param().batch_size();
const int top_size = this->layer_param_.top_size();
vector<int> top_shape;
for (int i = 0; i < top_size; ++i) {
top_shape.resize(hdf_blobs_[i]->num_axes());
top_shape[0] = batch_size;
for (int j = 1; j < top_shape.size(); ++j) {
top_shape[j] = hdf_blobs_[i]->shape(j);
}
top[i]->Reshape(top_shape);
}
}
template <typename Dtype>
void HDF5DataLayer<Dtype>::Forward_cpu(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top) {
const int batch_size = this->layer_param_.hdf5_data_param().batch_size();
for (int i = 0; i < batch_size; ++i, ++current_row_) {
// 因为SetUp里面已经读取了第一个文件的数据了
if (current_row_ == hdf_blobs_[0]->shape(0)) {
if (num_files_ > 1) {// 如果文件数目大于1
++current_file_;
// 如果current_file是最后一个文件的索引编号则
if (current_file_ == num_files_) {
current_file_ = 0;// 重置
// 打乱文件索引，再来一遍
if (this->layer_param_.hdf5_data_param().shuffle()) {
std::random_shuffle(file_permutation_.begin(),
file_permutation_.end());
}
DLOG(INFO) << "Looping around to first file.";
}
// 读取数据到hdf_blobs
LoadHDF5FileData(
hdf_filenames_[file_permutation_[current_file_]].c_str());
}// end of if (current_row_
current_row_ = 0;
// 打乱数据顺序索引
if (this->layer_param_.hdf5_data_param().shuffle())
std::random_shuffle(data_permutation_.begin(), data_permutation_.end());
}
// 复制数据到top
for (int j = 0; j < this->layer_param_.top_size(); ++j) {
int data_dim = top[j]->count() / top[j]->shape(0);
caffe_copy(data_dim,
&hdf_blobs_[j]->cpu_data()[data_permutation_[current_row_]
* data_dim], &top[j]->mutable_cpu_data()[i * data_dim]);
}
}
}
#ifdef CPU_ONLY
STUB_GPU_FORWARD(HDF5DataLayer, Forward);
#endif
INSTANTIATE_CLASS(HDF5DataLayer);
REGISTER_LAYER_CLASS(HDF5Data);
} // namespace caffe（6）HDF5OutputLayer类的定义以及实现如下：
HDF5输出层主要就是将传递过来的数据存储到HDF5文件，并没有向前传播数据啥的，也没有反传，仅仅是将前一层传输过来的bottom存储到文件。
HDF5输出层的定义：
/**
* @brief Write blobs to disk as HDF5 files.
*
* TODO(dox): thorough documentation for Forward and proto params.
* 将数据写入到HDF5文件
*/
template <typename Dtype>
class HDF5OutputLayer : public Layer<Dtype> {
public:
explicit HDF5OutputLayer(const LayerParameter& param)
: Layer<Dtype>(param), file_opened_(false) {}
virtual ~HDF5OutputLayer();
virtual void LayerSetUp(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top);
// Data layers should be shared by multiple solvers in parallel
virtual inline bool ShareInParallel() const { return true; }
// Data layers have no bottoms, so reshaping is trivial.
virtual void Reshape(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top) {}
virtual inline const char* type() const { return "HDF5Output"; }
// TODO: no limit on the number of blobs
virtual inline int ExactNumBottomBlobs() const { return 2; }
virtual inline int ExactNumTopBlobs() const { return 0; }
inline std::string file_name() const { return file_name_; }
protected:
// HDF5输出层不前向传也不反向传，只是将前一层传递过来的数据写入HDF5文件
virtual void Forward_cpu(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top);
virtual void Forward_gpu(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top);
virtual void Backward_cpu(const vector<Blob<Dtype>*>& top,
const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom);
virtual void Backward_gpu(const vector<Blob<Dtype>*>& top,
const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom);
// 将bottom的数据存储到文件
virtual void SaveBlobs();
bool file_opened_;
std::string file_name_;
hid_t file_id_;
Blob<Dtype> data_blob_;
Blob<Dtype> label_blob_;
};
HDF5输出层的实现如下：
#include <vector>
#include "hdf5.h"
#include "hdf5_hl.h"
#include "caffe/blob.hpp"
#include "caffe/common.hpp"
#include "caffe/layer.hpp"
#include "caffe/util/hdf5.hpp"
#include "caffe/vision_layers.hpp"
namespace caffe {
template <typename Dtype>
void HDF5OutputLayer<Dtype>::LayerSetUp(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top) {
// 参数文件中的文件名
file_name_ = this->layer_param_.hdf5_output_param().file_name();
// 打开文件
file_id_ = H5Fcreate(file_name_.c_str(), H5F_ACC_TRUNC, H5P_DEFAULT,
H5P_DEFAULT);
CHECK_GE(file_id_, 0) << "Failed to open HDF5 file" << file_name_;
file_opened_ = true;// 设置文件打开标志
}
template <typename Dtype>
HDF5OutputLayer<Dtype>::~HDF5OutputLayer<Dtype>() {
if (file_opened_) {
herr_t status = H5Fclose(file_id_);
CHECK_GE(status, 0) << "Failed to close HDF5 file " << file_name_;
}
}
// 将blob存放到hdf5文件
// 数据和类标
template <typename Dtype>
void HDF5OutputLayer<Dtype>::SaveBlobs() {
// TODO: no limit on the number of blobs
LOG(INFO) << "Saving HDF5 file " << file_name_;
CHECK_EQ(data_blob_.num(), label_blob_.num()) <<
"data blob and label blob must have the same batch size";
hdf5_save_nd_dataset(file_id_, HDF5_DATA_DATASET_NAME, data_blob_);
hdf5_save_nd_dataset(file_id_, HDF5_DATA_LABEL_NAME, label_blob_);
LOG(INFO) << "Successfully saved " << data_blob_.num() << " rows";
}
// 实际上就是从bottom将输入过来的数据存放到hdf5文件
template <typename Dtype>
void HDF5OutputLayer<Dtype>::Forward_cpu(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top) {
CHECK_GE(bottom.size(), 2);
CHECK_EQ(bottom[0]->num(), bottom[1]->num());
// 改变data_blob_的形状以及label_blob_的形状
data_blob_.Reshape(bottom[0]->num(), bottom[0]->channels(),
bottom[0]->height(), bottom[0]->width());
label_blob_.Reshape(bottom[1]->num(), bottom[1]->channels(),
bottom[1]->height(), bottom[1]->width());
const int data_datum_dim = bottom[0]->count() / bottom[0]->num();
const int label_datum_dim = bottom[1]->count() / bottom[1]->num();
// 从bottom[0]和[1]复制到data_blob_和label_blob_
for (int i = 0; i < bottom[0]->num(); ++i) {
caffe_copy(data_datum_dim, &bottom[0]->cpu_data()[i * data_datum_dim],
&data_blob_.mutable_cpu_data()[i * data_datum_dim]);
caffe_copy(label_datum_dim, &bottom[1]->cpu_data()[i * label_datum_dim],
&label_blob_.mutable_cpu_data()[i * label_datum_dim]);
}
// 存放到文件
SaveBlobs();
}
// 不反传
template <typename Dtype>
void HDF5OutputLayer<Dtype>::Backward_cpu(const vector<Blob<Dtype>*>& top,
const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom) {
return;
}
#ifdef CPU_ONLY
STUB_GPU(HDF5OutputLayer);
#endif
INSTANTIATE_CLASS(HDF5OutputLayer);
REGISTER_LAYER_CLASS(HDF5Output);
} // namespace caffe
（7）ImageDataLayer类的定义以及实现如下：
该层主要的功能是，从参数中给定的列表文件读取图像列表以及类标，读取图像的时候会进行预处理，然后前向传。
首先给出该层的参数的定义：
message ImageDataParameter {
// Specify the data source.
// 列表文件包含图像的路径和对应的类标，以空格隔开
optional string source = 1;
// Specify the batch size.
// 批大小
optional uint32 batch_size = 4 [default = 1];
// The rand_skip variable is for the data layer to skip a few data points
// to avoid all asynchronous sgd clients to start at the same point. The skip
// point would be set as rand_skip * rand(0,1). Note that rand_skip should not
// be larger than the number of keys in the database.
// 随机调过一些数据
optional uint32 rand_skip = 7 [default = 0];
// 是否需要打乱数据顺序
// Whether or not ImageLayer should shuffle the list of files at every epoch.
optional bool shuffle = 8 [default = false];
// It will also resize images if new_height or new_width are not zero.
// 将图像resize到新的高度的宽度
optional uint32 new_height = 9 [default = 0];
optional uint32 new_width = 10 [default = 0];
// Specify if the images are color or gray
// 图像是否是彩色的
optional bool is_color = 11 [default = true];
// DEPRECATED. See TransformationParameter. For data pre-processing, we can do
// simple scaling and subtracting the data mean, if provided. Note that the
// mean subtraction is always carried out before scaling.
// 是否需要对图像进行缩放
optional float scale = 2 [default = 1];
// 均值文件
optional string mean_file = 3;
// DEPRECATED. See TransformationParameter. Specify if we would like to randomly
// crop an image.
// crop的大小
optional uint32 crop_size = 5 [default = 0];
// DEPRECATED. See TransformationParameter. Specify if we want to randomly mirror
// data.
// 是否需要对图像进行镜像，所谓镜像就是左边复制到右边
optional bool mirror = 6 [default = false];
// 图像的根目录
optional string root_folder = 12 [default = ""];
}
首先给出类的定义：
/**
* @brief Provides data to the Net from image files.
*
* TODO(dox): thorough documentation for Forward and proto params.
* 从图像文件中读取数据，这个应该比较常用
* 从一个列表文件读取图像的路径和类标，列表文件的路径在层参数的配置文件中指定
*/
template <typename Dtype>
class ImageDataLayer : public BasePrefetchingDataLayer<Dtype> {
public:
explicit ImageDataLayer(const LayerParameter& param)
: BasePrefetchingDataLayer<Dtype>(param) {}
virtual ~ImageDataLayer();
virtual void DataLayerSetUp(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top);
virtual inline const char* type() const { return "ImageData"; }
virtual inline int ExactNumBottomBlobs() const { return 0; }
virtual inline int ExactNumTopBlobs() const { return 2; }
protected:
shared_ptr<Caffe::RNG> prefetch_rng_;
// 对图像索引进行打乱
virtual void ShuffleImages();
virtual void load_batch(Batch<Dtype>* batch);
// 图像路径和类标的vector
vector<std::pair<std::string, int> > lines_;
// 随机跳过的图像的个数，也就是调过之后的一开始的图像的id
int lines_id_;
};
下面给出具体的实现细节：
#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_layers.hpp"
#include "caffe/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();
}
template <typename Dtype>
void ImageDataLayer<Dtype>::DataLayerSetUp(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top) {
// 根据参数文件设置参数
// 图像的高度、宽度、是否彩色图像、图像目录
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();
// 当前只支持读取高度和宽度同样大小的图像
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
// 读取存放图像文件名和类标的列表文件
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;
// lines_存放文件名和类标的pair
while (infile >> filename >> label) {
lines_.push_back(std::make_pair(filename, label));
}
// 是否需要打乱文件的顺序
if (this->layer_param_.image_data_param().shuffle()) {
// randomly shuffle data
LOG(INFO) << "Shuffling data";
const unsigned int prefetch_rng_seed = caffe_rng_rand();
prefetch_rng_.reset(new Caffe::RNG(prefetch_rng_seed));
ShuffleImages();
}
LOG(INFO) << "A total of " << lines_.size() << " images.";
// 随机跳过的图像，调过的图像个数在[0, rand_skip-1]之间
lines_id_ = 0;
// Check if we would need to randomly skip a few data points
// 如果参数中的rand_skip大于1，则随机跳过[0,rand_skip-1]个图片
//
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;
}
// Read an image, and use it to initialize the top blob.
// 读取文件名到Mat
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_image.
// 对数据的形状进行推断
vector<int> top_shape = this->data_transformer_->InferBlobShape(cv_img);
// 设置transformed_data_的形状
this->transformed_data_.Reshape(top_shape);
// Reshape prefetch_data and top[0] according to the batch_size.
// 设置batch_size
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;
// 设置预取数组中数据的形状
for (int i = 0; i < this->PREFETCH_COUNT; ++i) {
this->prefetch_[i].data_.Reshape(top_shape);
}
// 设置输出的数据的形状
top[0]->Reshape(top_shape);
LOG(INFO) << "output data size: " << top[0]->num() << ","
<< top[0]->channels() << "," << top[0]->height() << ","
<< top[0]->width();
// label
// 设置输出的类标的形状
vector<int> label_shape(1, batch_size);
top[1]->Reshape(label_shape);
// 设置预取数组中类标的形状
for (int i = 0; i < this->PREFETCH_COUNT; ++i) {
this->prefetch_[i].label_.Reshape(label_shape);
}
}
// 产生打乱图像顺序的数组
template <typename Dtype>
void ImageDataLayer<Dtype>::ShuffleImages() {
caffe::rng_t* prefetch_rng =
static_cast<caffe::rng_t*>(prefetch_rng_->generator());
shuffle(lines_.begin(), lines_.end(), prefetch_rng);
}
// This function is called on prefetch thread
// 该函数会被内部的线程调用
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());
// 获取层参数，具体参见层参数的定义的解释
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.
// 读取跳过之后的第一幅图像，然后根据该图像设置相撞
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.
// 推断图像形状
vector<int> top_shape = this->data_transformer_->InferBlobShape(cv_img);
// 设置transformed_data_形状
this->transformed_data_.Reshape(top_shape);
// Reshape batch according to the batch_size.
// 设置batch_size
top_shape[0] = batch_size;
batch->data_.Reshape(top_shape);
Dtype* prefetch_data = batch->data_.mutable_cpu_data();
Dtype* prefetch_label = batch->label_.mutable_cpu_data();
// datum scales
// 读取一批图像，并进行预处理
const int lines_size = lines_.size();
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,
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
// 进行预处理
// 根据图像的批次获得图像数据的偏移量
int offset = batch->data_.offset(item_id);
// 设置图像数据的指针到transformed_data_
this->transformed_data_.set_cpu_data(prefetch_data + offset);
// 进行预处理
this->data_transformer_->Transform(cv_img, &(this->transformed_data_));
trans_time += timer.MicroSeconds();//统计预处理时间
// 复制类标到prefetch_label
prefetch_label[item_id] = lines_[lines_id_].second;
// go to the next iter
lines_id_++;
// 是否是图像目录中的最后一个图像
if (lines_id_ >= lines_size) {
// We have reached the end. Restart from the first.
DLOG(INFO) << "Restarting data prefetching from start.";
lines_id_ = 0;
// 打乱图像索引的顺序
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.";
// 预处理时间
DLOG(INFO) << "Transform time: " << trans_time / 1000 << " ms.";
}
INSTANTIATE_CLASS(ImageDataLayer);
REGISTER_LAYER_CLASS(ImageData);
} // namespace caffe
#endif // USE_OPENCV（8）MemoryDataLayer 类的定义以及实现如下：
该类主要就是对于读取好的Datum或者OpenCV读取的Mat的Vector进行预处理（图像的crop、scale等），然后前传。
首先给出该类的定义
/**
* @brief Provides data to the Net from memory.
* 从内存中读取数据，这里指已经从数据文件或者图像文件中读取到了数据，然后输入到该层
* TODO(dox): thorough documentation for Forward and proto params.
*/
template <typename Dtype>
class MemoryDataLayer : public BaseDataLayer<Dtype> {
public:
explicit MemoryDataLayer(const LayerParameter& param)
: BaseDataLayer<Dtype>(param), has_new_data_(false) {}
virtual void DataLayerSetUp(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top);
virtual inline const char* type() const { return "MemoryData"; }
virtual inline int ExactNumBottomBlobs() const { return 0; }
virtual inline int ExactNumTopBlobs() const { return 2; }
// 将内存中的数据加入added_data_和added_label_(数据和类标)
virtual void AddDatumVector(const vector<Datum>& datum_vector);
#ifdef USE_OPENCV
// 如果有opencv则将opencv读取到的Mat,并且将labels加入added_data_和added_label_(数据和类标)
virtual void AddMatVector(const vector<cv::Mat>& mat_vector,
const vector<int>& labels);
#endif // USE_OPENCV
// Reset should accept const pointers, but can't, because the memory
// will be given to Blob, which is mutable
// Reset函数实际上是将data、label、以及batchsize(n)设置到内部的变量里面去
void Reset(Dtype* data, Dtype* label, int n);
void set_batch_size(int new_size);
int batch_size() { return batch_size_; }
int channels() { return channels_; }
int height() { return height_; }
int width() { return width_; }
protected:
virtual void Forward_cpu(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top);
int batch_size_, channels_, height_, width_, size_;
Dtype* data_;
Dtype* labels_;
// batch_size
int n_;
size_t pos_;
// 内部的数据和类标
Blob<Dtype> added_data_;
Blob<Dtype> added_label_;
// 是否有新的数据
bool has_new_data_;
};
下面给出具体的实现细节：
#ifdef USE_OPENCV
#include <opencv2/core/core.hpp>
#endif // USE_OPENCV
#include <vector>
#include "caffe/data_layers.hpp"
#include "caffe/layer.hpp"
#include "caffe/util/io.hpp"
namespace caffe {
template <typename Dtype>
void MemoryDataLayer<Dtype>::DataLayerSetUp(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top) {
// 从参数文件获取参数
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";
// 设置top的形状
vector<int> label_shape(1, batch_size_);
top[0]->Reshape(batch_size_, channels_, height_, width_);
top[1]->Reshape(label_shape);
// 设置内部变量added_data_和added_label_的形状
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();
}
// 将Datum的vector放入到added_data_和added_label_
// 并进行预处理
template <typename Dtype>
void MemoryDataLayer<Dtype>::AddDatumVector(const vector<Datum>& datum_vector) {
CHECK(!has_new_data_) <<
"Can't add data until current data has been consumed.";
size_t num = datum_vector.size();
CHECK_GT(num, 0) << "There is no datum to add.";
CHECK_EQ(num % batch_size_, 0) <<
"The added data must be a multiple of the batch size.";
// 改变形状
added_data_.Reshape(num, channels_, height_, width_);
added_label_.Reshape(num, 1, 1, 1);
// Apply data transformations (mirror, scale, crop...)
// 对数据进行预处理
this->data_transformer_->Transform(datum_vector, &added_data_);
// Copy Labels
// 复制类标到top_label
Dtype* top_label = added_label_.mutable_cpu_data();
for (int item_id = 0; item_id < num; ++item_id) {
top_label[item_id] = datum_vector[item_id].label();
}
// num_images == batch_size_
Dtype* top_data = added_data_.mutable_cpu_data();
// 将数据、类标以及数据个数设置到该类的内部变量
Reset(top_data, top_label, num);
// 设置标记为true
has_new_data_ = true;
}
// 如果定义OPENCV，则对数据进行处理存放到added_data_和added_label_
#ifdef USE_OPENCV
template <typename Dtype>
void MemoryDataLayer<Dtype>::AddMatVector(const vector<cv::Mat>& mat_vector,
const vector<int>& labels) {
size_t num = mat_vector.size();
CHECK(!has_new_data_) <<
"Can't add mat until current data has been consumed.";
CHECK_GT(num, 0) << "There is no mat to add";
CHECK_EQ(num % batch_size_, 0) <<
"The added data must be a multiple of the batch size.";
added_data_.Reshape(num, channels_, height_, width_);
added_label_.Reshape(num, 1, 1, 1);
// Apply data transformations (mirror, scale, crop...)
// 预处理
this->data_transformer_->Transform(mat_vector, &added_data_);
// Copy Labels
Dtype* top_label = added_label_.mutable_cpu_data();
for (int item_id = 0; item_id < num; ++item_id) {
top_label[item_id] = labels[item_id];
}
// num_images == batch_size_
Dtype* top_data = added_data_.mutable_cpu_data();
Reset(top_data, top_label, num);
has_new_data_ = true;
}
#endif // USE_OPENCV
// 将数据和类标设置到内部的变量
// data_、labels_、n_
// 并且设置位置pos_=0
template <typename Dtype>
void MemoryDataLayer<Dtype>::Reset(Dtype* data, Dtype* labels, int n) {
CHECK(data);
CHECK(labels);
CHECK_EQ(n % batch_size_, 0) << "n must be a multiple of batch size";
// Warn with transformation parameters since a memory array is meant to
// be generic and no transformations are done with Reset().
if (this->layer_param_.has_transform_param()) {
LOG(WARNING) << this->type() << " does not transform array data on Reset()";
}
data_ = data;
labels_ = labels;
n_ = n;// batch_size
pos_ = 0;
}
// 设置内内部变量added_data_和added_label_的批数
template <typename Dtype>
void MemoryDataLayer<Dtype>::set_batch_size(int new_size) {
CHECK(!has_new_data_) <<
"Can't change batch_size until current data has been consumed.";
batch_size_ = new_size;
added_data_.Reshape(batch_size_, channels_, height_, width_);
added_label_.Reshape(batch_size_, 1, 1, 1);
}
// 将内部变量added_data_和added_label_复制到top传递给下一层
template <typename Dtype>
void MemoryDataLayer<Dtype>::Forward_cpu(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top) {
CHECK(data_) << "MemoryDataLayer needs to be initalized by calling Reset";
// 这里直接使用内部变量将数据复制到top[0]、将类标复制到top[1]
top[0]->Reshape(batch_size_, channels_, height_, width_);
top[1]->Reshape(batch_size_, 1, 1, 1);
top[0]->set_cpu_data(data_ + pos_ * size_);
top[1]->set_cpu_data(labels_ + pos_);
pos_ = (pos_ + batch_size_) % n_;
if (pos_ == 0)
has_new_data_ = false;// 传过一次之后，就没有新数据啦
}
INSTANTIATE_CLASS(MemoryDataLayer);
REGISTER_LAYER_CLASS(MemoryData);
} // namespace caffe（9）WindowDataLayer类的定义以及实现如下：
该类主要就是对于读取好的Datum或者OpenCV读取的Mat的Vector进行预处理（图像的crop、scale等），然后前传。
首先给出窗口数据文件的格式，便于自己训练
窗口文件的格式如下:
# 图像索引(举例:# 1就表示第一个图像,注意#号与数字之间有空格)
图像的路径
图像通道数
图像高度
图像宽度
窗口数目
类标,与前景目标的重叠率,x1,y1,x2,y2
注:x1,y1,x2,y2是窗口的左上和右下的坐标
为了理解的更清楚我这里举个例子：
# 1 /1.jpg 3 720 480 100 1 1 0 0 100 100 2 30 100 1500 1500
上述的例子表示一个编号为1的图像相对路径为/1.jpg，通道为3，高度为720
宽度为480，窗口数目为100，类标为1，与前景目标的重叠率为0.8，类标为1窗口的左上坐标为(0,0),右下坐标为(100,100)
类标为2的窗口的左上角坐标为(30,100)，右下角的坐标为(1500,1500)。有多少窗口后面就这么继续写下去
接下来给出该层的参数：
message WindowDataParameter {
// Specify the data source.
// 装有窗口数据的列表文件
optional string source = 1;
// For data pre-processing, we can do simple scaling and subtracting the
// data mean, if provided. Note that the mean subtraction is always carried
// out before scaling.
// 是否需要缩放图像中的像素值，注意哈这不是缩放图像的大小，是拿图像的像素值乘以这个
optional float scale = 2 [default = 1];
// 平均值文件路径
optional string mean_file = 3;
// Specify the batch size.
// 批大小
optional uint32 batch_size = 4;
// Specify if we would like to randomly crop an image.
// 随机crop的图像块的大小
optional uint32 crop_size = 5 [default = 0];
// Specify if we want to randomly mirror data.
// 是否随机镜像图像
optional bool mirror = 6 [default = false];
// Foreground (object) overlap threshold
// 前景重叠阈值
optional float fg_threshold = 7 [default = 0.5];
// Background (non-object) overlap threshold
// 背景重叠阈值
optional float bg_threshold = 8 [default = 0.5];
// Fraction of batch that should be foreground objects
// 每一批中有多少比例应该是前景(也就是是你要检测的物体)
optional float fg_fraction = 9 [default = 0.25];
// Amount of contextual padding to add around a window
// (used only by the window_data_layer)
// 是否需要在窗口周围padding
optional uint32 context_pad = 10 [default = 0];
// Mode for cropping out a detection window
// warp: cropped window is warped to a fixed size and aspect ratio
// square: the tightest square around the window is cropped
// crop的模式,square还是warp
optional string crop_mode = 11 [default = "warp"];
// cache_images: will load all images in memory for faster access
// 是否将文件缓冲到内存
optional bool cache_images = 12 [default = false];
// append root_folder to locate images
// 图像文件根目录
optional string root_folder = 13 [default = ""];
}
我们给出该类的定义：
/**
* @brief Provides data to the Net from windows of images files, specified
* by a window data file.
* 从图像文件的窗口获取数据，需要指定窗口数据文件
* TODO(dox): thorough documentation for Forward and proto params.
*/
template <typename Dtype>
class WindowDataLayer : public BasePrefetchingDataLayer<Dtype> {
public:
explicit WindowDataLayer(const LayerParameter& param)
: BasePrefetchingDataLayer<Dtype>(param) {}
virtual ~WindowDataLayer();
virtual void DataLayerSetUp(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top);
virtual inline const char* type() const { return "WindowData"; }
virtual inline int ExactNumBottomBlobs() const { return 0; }
virtual inline int ExactNumTopBlobs() const { return 2; }
protected:
virtual unsigned int PrefetchRand();
virtual void load_batch(Batch<Dtype>* batch);
shared_ptr<Caffe::RNG> prefetch_rng_;
vector<std::pair<std::string, vector<int> > > image_database_;
// 窗口类中所使用的窗口数据的枚举
// 就是定义个vector<float>，然后里面按顺序存放下面这些类型的数据
enum WindowField { IMAGE_INDEX, LABEL, OVERLAP, X1, Y1, X2, Y2, NUM };
vector<vector<float> > fg_windows_;
vector<vector<float> > bg_windows_;
Blob<Dtype> data_mean_;
vector<Dtype> mean_values_;
bool has_mean_file_;
bool has_mean_values_;
bool cache_images_;
vector<std::pair<std::string, Datum > > image_database_cache_;
};
然后给出该类的实现
#ifdef USE_OPENCV
#include <opencv2/highgui/highgui_c.h>
#include <stdint.h>
#include <algorithm>
#include <map>
#include <string>
#include <utility>
#include <vector>
#include "opencv2/core/core.hpp"
#include "opencv2/highgui/highgui.hpp"
#include "opencv2/imgproc/imgproc.hpp"
#include "caffe/common.hpp"
#include "caffe/data_layers.hpp"
#include "caffe/layer.hpp"
#include "caffe/util/benchmark.hpp"
#include "caffe/util/io.hpp"
#include "caffe/util/math_functions.hpp"
#include "caffe/util/rng.hpp"
// caffe.proto > LayerParameter > WindowDataParameter
// 'source' field specifies the window_file
// 'crop_size' indicates the desired warped size
namespace caffe {
template <typename Dtype>
WindowDataLayer<Dtype>::~WindowDataLayer<Dtype>() {
this->StopInternalThread();
}
// 读取窗口数据文件的信息,并设置各个数据结构的形状
template <typename Dtype>
void WindowDataLayer<Dtype>::DataLayerSetUp(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top) {
// LayerSetUp runs through the window_file and creates two structures
// that hold windows: one for foreground (object) windows and one
// for background (non-object) windows. We use an overlap threshold
// to decide which is which.
// window_file format
// repeated:
// # image_index
// img_path (abs path)
// channels
// height
// width
// num_windows
// class_index overlap x1 y1 x2 y2
// 窗口文件的格式如下:
// # 图像索引(举例:# 1就表示第一个图像,注意#号与数字之间有空格)
// 图像的路径
// 图像通道数
// 图像高度
// 图像宽度
// 窗口数目
// 类标,overlap,x1,y1,x2,y2
// 注:x1,y1,x2,y2是窗口的左上和右下的坐标
// 我这里举个例子
// # 1 /1.jpg 3 720 480 100 1 1 0 0 100 100
// 上述的例子即使表示一个编号为1的图像相对路径为/1.jpg，通道为3，高度为720
// 宽度为480，窗口数目为100，类标为1，overlap为1，窗口的左上坐标为(0,0),右下坐标为(100,100)
LOG(INFO) << "Window data layer:" << std::endl
<< " foreground (object) overlap threshold: "
<< this->layer_param_.window_data_param().fg_threshold() << std::endl
<< " background (non-object) overlap threshold: "
<< this->layer_param_.window_data_param().bg_threshold() << std::endl
<< " foreground sampling fraction: "
<< this->layer_param_.window_data_param().fg_fraction() << std::endl
<< " cache_images: "
<< this->layer_param_.window_data_param().cache_images() << std::endl
<< " root_folder: "
<< this->layer_param_.window_data_param().root_folder();
cache_images_ = this->layer_param_.window_data_param().cache_images();
string root_folder = this->layer_param_.window_data_param().root_folder();
// 根据参数文件中是否需要进行左右mirror，或者是否进行crop，
// 来判断是否需要初始化随机数种子
const bool prefetch_needs_rand =
this->transform_param_.mirror() ||
this->transform_param_.crop_size();
if (prefetch_needs_rand) {
const unsigned int prefetch_rng_seed = caffe_rng_rand();
prefetch_rng_.reset(new Caffe::RNG(prefetch_rng_seed));
} else {
prefetch_rng_.reset();
}
// 打开窗口文件
std::ifstream infile(this->layer_param_.window_data_param().source().c_str());
CHECK(infile.good()) << "Failed to open window file "
<< this->layer_param_.window_data_param().source() << std::endl;
// 这个是类标与类标出现的次数之间的映射
// 这里称之为类标直方图
map<int, int> label_hist;
label_hist.insert(std::make_pair(0, 0));
string hashtag;
int image_index, channels;
// 先从窗口文件中读取一个图像索引测试一下是否为空
if (!(infile >> hashtag >> image_index)) {
LOG(FATAL) << "Window file is empty";
}
do {
// 检查是否# 开头
CHECK_EQ(hashtag, "#");
// read image path
string image_path;
// 接下来读取图像的相对路径
// 将该路径与根目录路径拼接
infile >> image_path;
image_path = root_folder + image_path;
// read image dimensions
vector<int> image_size(3);
// 读取图像的维度信息，分别为channel，height , width
infile >> image_size[0] >> image_size[1] >> image_size[2];
channels = image_size[0];
// 将图像路径和图像大小压入到image_database_中
image_database_.push_back(std::make_pair(image_path, image_size));
// 如果需要缓存图像到内存的话，则用image_database_cache_进行存储
if (cache_images_) {
Datum datum;
// 将图像数据读取到Datum这个结构
if (!ReadFileToDatum(image_path, &datum)) {
LOG(ERROR) << "Could not open or find file " << image_path;
return;
}
// 将Datum结构的图像缓存到到image_database_cache_
image_database_cache_.push_back(std::make_pair(image_path, datum));
}
// read each box
int num_windows;
// 读取窗口个数
infile >> num_windows;
// 从参数文件获取前景和背景阈值
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;
// 读取 类标,与前景目标的重叠率,x1,y1,x2,y2
infile >> label >> overlap >> x1 >> y1 >> x2 >> y2;
// 按照顺序放在window这个数据结构里头
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
// 下面是将窗口的前景和背景都装入到fg_windows_和bg_windows_中去
// 如果重叠的比例大于前景阈值，那么就认为是前景
if (overlap >= fg_threshold) {
int label = window[WindowDataLayer::LABEL];
// 类标必须大于0，因为重叠区域已经大于前景阈值了
// 此时如果类标不大于0，表明数据有误!
CHECK_GT(label, 0);
fg_windows_.push_back(window);
// 该类的直方图+1
label_hist.insert(std::make_pair(label, 0));
label_hist[label]++;
} else if (overlap < bg_threshold) {
// 如果重叠阈值小于背景阈值则认为是背景
// background window, force label and overlap to 0
window[WindowDataLayer::LABEL] = 0;
window[WindowDataLayer::OVERLAP] = 0;
bg_windows_.push_back(window);
// 0类的直方图(也就是背景的直方图)+1
label_hist[0]++;
}
}
// 每处理100个就显示一瞎
if (image_index % 100 == 0) {
LOG(INFO) << "num: " << image_index << " "
<< image_path << " "
<< image_size[0] << " "
<< image_size[1] << " "
<< image_size[2] << " "
<< "windows to process: " << num_windows;
}
} while (infile >> hashtag >> image_index);
// 读取完毕后输出图像的个数
LOG(INFO) << "Number of images: " << image_index+1;
// 输出统计的每个类别的个数
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();
// image
// 获取crop_size
const int crop_size = this->transform_param_.crop_size();
CHECK_GT(crop_size, 0);
// 获取batch_size
const int batch_size = this->layer_param_.window_data_param().batch_size();
// 将top[0]设置为batch_size,channels, crop_size, crop_size大小的
top[0]->Reshape(batch_size, channels, crop_size, crop_size);
// 将prefetch_中的数据形状也这么设置
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();
// label
// 将top[1]设置为类标大小
vector<int> label_shape(1, batch_size);
top[1]->Reshape(label_shape);
// 将prefetch_中的类标形状也这么设置
for (int i = 0; i < this->PREFETCH_COUNT; ++i) {
this->prefetch_[i].label_.Reshape(label_shape);
}
// data mean
// 是否有均值文件或者有均值
has_mean_file_ = this->transform_param_.has_mean_file();
has_mean_values_ = this->transform_param_.mean_value_size() > 0;
if (has_mean_file_) {// 有均值文件就读
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_) {// 有均值就直接从参数中获取
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));
}
// 检查均值是不是等于1，或者等于图像的通道数
// 也就是要么所有通道都使用同一个均值
// 要么每个通道用一个均值
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]);
}
}
}
}
// 随机数生成器进行初始化并生成随机数
template <typename Dtype>
unsigned int WindowDataLayer<Dtype>::PrefetchRand() {
CHECK(prefetch_rng_);
caffe::rng_t* prefetch_rng =
static_cast<caffe::rng_t*>(prefetch_rng_->generator());
return (*prefetch_rng)();
}
// 因为继承BasePrefetchingDataLayer所以要实现load_batch
// 以供线程调用
// This function is called on prefetch thread
template <typename Dtype>
void WindowDataLayer<Dtype>::load_batch(Batch<Dtype>* batch) {
// At each iteration, sample N windows where N*p are foreground (object)
// windows and N*(1-p) are background (non-object) windows
CPUTimer batch_timer;
batch_timer.Start();
double read_time = 0;
double trans_time = 0;
CPUTimer timer;
// top数据和类标
Dtype* top_data = batch->data_.mutable_cpu_data();
Dtype* top_label = batch->label_.mutable_cpu_data();
// 缩放尺度
const Dtype scale = this->layer_param_.window_data_param().scale();
// batch_size
const int batch_size = this->layer_param_.window_data_param().batch_size();
// 上下文填充
const int context_pad = this->layer_param_.window_data_param().context_pad();
// crop_size
const int crop_size = this->transform_param_.crop_size();
// 是否镜像
const bool mirror = this->transform_param_.mirror();
// 前景比例
const float fg_fraction =
this->layer_param_.window_data_param().fg_fraction();
Dtype* mean = NULL;
int mean_off = 0;
int mean_width = 0;
int mean_height = 0;
// 如果有平均值文件则
if (this->has_mean_file_) {
mean = this->data_mean_.mutable_cpu_data();
// 经过crop之后的平均值图像的中心
mean_off = (this->data_mean_.width() - crop_size) / 2;
mean_width = this->data_mean_.width();
mean_height = this->data_mean_.height();
}
cv::Size cv_crop_size(crop_size, crop_size);
// 获取crop的模式，是warp还是square
const string& crop_mode = this->layer_param_.window_data_param().crop_mode();
bool use_square = (crop_mode == "square") ? true : false;
// zero out batch
caffe_set(batch->data_.count(), Dtype(0), top_data);
// 根据前景比例获得前景图像的数目
const int num_fg = static_cast<int>(static_cast<float>(batch_size)
* fg_fraction);
// 样本数量，是前景还是背景?[0]是背景[1]是前景
const int num_samples[2] = { batch_size - num_fg, num_fg };
int item_id = 0;
// sample from bg set then fg set
// 先对背景进行采样
// 再对前景进行采样
for (int is_fg = 0; is_fg < 2; ++is_fg) {
for (int dummy = 0; dummy < num_samples[is_fg]; ++dummy) {
// sample a window
timer.Start();
// 生成一个随机数
const unsigned int rand_index = PrefetchRand();
// fg_windows_和bg_windows_存储的是对应的窗口信息
// 在SetUp中读取的窗口数据文件的时候获得的
// 从该图像的若干窗口中去随机选择一个窗口
vector<float> window = (is_fg) ?
fg_windows_[rand_index % fg_windows_.size()] :
bg_windows_[rand_index % bg_windows_.size()];
// 随机选择是否需要镜像
bool do_mirror = mirror && PrefetchRand() % 2;
// load the image containing the window
// 载入图像的路径以及类标
pair<std::string, vector<int> > image =
image_database_[window[WindowDataLayer<Dtype>::IMAGE_INDEX]];
// 读取图像
cv::Mat cv_img;
if (this->cache_images_) {
// 如果图像缓冲到内存则获得对应图像的Datum
pair<std::string, Datum> image_cached =
image_database_cache_[window[WindowDataLayer<Dtype>::IMAGE_INDEX]];
// 将图像的Datum解码为OpenCV的Mat
cv_img = DecodeDatumToCVMat(image_cached.second, true);
} else {
// 否则直接读取
cv_img = cv::imread(image.first, CV_LOAD_IMAGE_COLOR);
if (!cv_img.data) {
LOG(ERROR) << "Could not open or find file " << image.first;
return;
}
}
read_time += timer.MicroSeconds();
timer.Start();
const int channels = cv_img.channels();
// crop window out of image and warp it
// 窗口坐标
int x1 = window[WindowDataLayer<Dtype>::X1];
int y1 = window[WindowDataLayer<Dtype>::Y1];
int x2 = window[WindowDataLayer<Dtype>::X2];
int y2 = window[WindowDataLayer<Dtype>::Y2];
int pad_w = 0;
int pad_h = 0;
// context_pad也是个大小，具体什么含义，我没有具体研究
// 毕竟不是搞检测的
// context_scale = crop_size / (crop_size - 2*context_pad)
if (context_pad > 0 || use_square) {
// scale factor by which to expand the original region
// such that after warping the expanded region to crop_size x crop_size
// there's exactly context_pad amount of padding on each side
Dtype context_scale = static_cast<Dtype>(crop_size) /
static_cast<Dtype>(crop_size - 2*context_pad);
// compute the expanded region
// 高度的一半
Dtype half_height = static_cast<Dtype>(y2-y1+1)/2.0;
// 宽度的一半
Dtype half_width = static_cast<Dtype>(x2-x1+1)/2.0;
// x中心
Dtype center_x = static_cast<Dtype>(x1) + half_width;
// y中心
Dtype center_y = static_cast<Dtype>(y1) + half_height;
if (use_square) {// 如果使用正方形形状则将较大的那个赋值给小的
if (half_height > half_width) {
half_width = half_height;
} else {
half_height = half_width;
}
}
// 获取经过处理之后的x1,y1,x2,y2
x1 = static_cast<int>(round(center_x - half_width*context_scale));
x2 = static_cast<int>(round(center_x + half_width*context_scale));
y1 = static_cast<int>(round(center_y - half_height*context_scale));
y2 = static_cast<int>(round(center_y + half_height*context_scale));
// the expanded region may go outside of the image
// so we compute the clipped (expanded) region and keep track of
// the extent beyond the image
// 经过处理之后的窗口如果不在图像内部是有问题的
// 这里对窗口的坐标进行处理
// 使得窗口的左上角不超过图像的左上角
// 窗口的右下角不超过图像的右下角
// 所以这里叫clip bounds嘛
int unclipped_height = y2-y1+1;
int unclipped_width = x2-x1+1;
int pad_x1 = std::max(0, -x1);
int pad_y1 = std::max(0, -y1);
int pad_x2 = std::max(0, x2 - cv_img.cols + 1);
int pad_y2 = std::max(0, y2 - cv_img.rows + 1);
// clip bounds
x1 = x1 + pad_x1;
x2 = x2 - pad_x2;
y1 = y1 + pad_y1;
y2 = y2 - pad_y2;
CHECK_GT(x1, -1);
CHECK_GT(y1, -1);
CHECK_LT(x2, cv_img.cols);
CHECK_LT(y2, cv_img.rows);
// 经过clip之后的高度和宽度
int clipped_height = y2-y1+1;
int clipped_width = x2-x1+1;
// scale factors that would be used to warp the unclipped
// expanded region
// scale_x/scale_y=crop_size除以未经clip之后的宽度/高度
Dtype scale_x =
static_cast<Dtype>(crop_size)/static_cast<Dtype>(unclipped_width);
Dtype scale_y =
static_cast<Dtype>(crop_size)/static_cast<Dtype>(unclipped_height);
// size to warp the clipped expanded region to
// 用clip的宽度和高度乘以scale_x或者scale_y得到crop_size中的宽度和高度
cv_crop_size.width =
static_cast<int>(round(static_cast<Dtype>(clipped_width)*scale_x));
cv_crop_size.height =
static_cast<int>(round(static_cast<Dtype>(clipped_height)*scale_y));
// 再对pad的边界进行处理
pad_x1 = static_cast<int>(round(static_cast<Dtype>(pad_x1)*scale_x));
pad_x2 = static_cast<int>(round(static_cast<Dtype>(pad_x2)*scale_x));
pad_y1 = static_cast<int>(round(static_cast<Dtype>(pad_y1)*scale_y));
pad_y2 = static_cast<int>(round(static_cast<Dtype>(pad_y2)*scale_y));
pad_h = pad_y1;
// if we're mirroring, we mirror the padding too (to be pedantic)
// 如果需要镜像填充的部分也要镜像
if (do_mirror) {
pad_w = pad_x2;
} else {
pad_w = pad_x1;
}
// ensure that the warped, clipped region plus the padding fits in the
// crop_size x crop_size image (it might not due to rounding)
// 确保大小是在crop_size x crop_size以内的
if (pad_h + cv_crop_size.height > crop_size) {
cv_crop_size.height = crop_size - pad_h;
}
if (pad_w + cv_crop_size.width > crop_size) {
cv_crop_size.width = crop_size - pad_w;
}
}
cv::Rect roi(x1, y1, x2-x1+1, y2-y1+1);
// 进行crop
cv::Mat cv_cropped_img = cv_img(roi);
// 使用线性插值进行缩放，缩放到cv_crop_size
cv::resize(cv_cropped_img, cv_cropped_img,
cv_crop_size, 0, 0, cv::INTER_LINEAR);
// horizontal flip at random
if (do_mirror) {
// 对图像进行镜像
cv::flip(cv_cropped_img, cv_cropped_img, 1);
}
// copy the warped window into top_data
for (int h = 0; h < cv_cropped_img.rows; ++h) {
const uchar* ptr = cv_cropped_img.ptr<uchar>(h);
int img_index = 0;
for (int w = 0; w < cv_cropped_img.cols; ++w) {
for (int c = 0; c < channels; ++c) {
int top_index = ((item_id * channels + c) * crop_size + h + pad_h)
* crop_size + w + pad_w;
// int top_index = (c * height + h) * width + w;
Dtype pixel = static_cast<Dtype>(ptr[img_index++]);
if (this->has_mean_file_) {// 有均值文件减去均值文件中对应的数值
int mean_index = (c * mean_height + h + mean_off + pad_h)
* mean_width + w + mean_off + pad_w;
top_data[top_index] = (pixel - mean[mean_index]) * scale;
} else {
if (this->has_mean_values_) {// 有均值则减去均值
top_data[top_index] = (pixel - this->mean_values_[c]) * scale;
} else {
top_data[top_index] = pixel * scale;// 像素值进行缩放
}
}
}
}
}
trans_time += timer.MicroSeconds();
// get window label
top_label[item_id] = window[WindowDataLayer<Dtype>::LABEL];
#if 0
// useful debugging code for dumping transformed windows to disk
string file_id;
std::stringstream ss;
ss << PrefetchRand();
ss >> file_id;
std::ofstream inf((string("dump/") + file_id +
string("_info.txt")).c_str(), std::ofstream::out);
inf << image.first << std::endl
<< window[WindowDataLayer<Dtype>::X1]+1 << std::endl
<< window[WindowDataLayer<Dtype>::Y1]+1 << std::endl
<< window[WindowDataLayer<Dtype>::X2]+1 << std::endl
<< window[WindowDataLayer<Dtype>::Y2]+1 << std::endl
<< do_mirror << std::endl
<< top_label[item_id] << std::endl
<< is_fg << std::endl;
inf.close();
std::ofstream top_data_file((string("dump/") + file_id +
string("_data.txt")).c_str(),
std::ofstream::out | std::ofstream::binary);
for (int c = 0; c < channels; ++c) {
for (int h = 0; h < crop_size; ++h) {
for (int w = 0; w < crop_size; ++w) {
top_data_file.write(reinterpret_cast<char*>(
&top_data[((item_id * channels + c) * crop_size + h)
* crop_size + w]),
sizeof(Dtype));
}
}
}
top_data_file.close();
#endif
item_id++;
}
}
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(WindowDataLayer);
REGISTER_LAYER_CLASS(WindowData);
} // namespace caffe
#endif // USE_OPENCV
最后提醒一下该类并没有重载前传函数，而是调用了基类的前传，我把对应的代码贴出来便于你整体进行理解
template <typename Dtype>
void BasePrefetchingDataLayer<Dtype>::Forward_cpu(
const vector<Blob<Dtype>*>& bottom, const vector<Blob<Dtype>*>& top) {
// 传递的时候是从full队列中弹出一个数据
Batch<Dtype>* batch = prefetch_full_.pop("Data layer prefetch queue empty");
// Reshape to loaded data.
// 根据batch的形状改变数据形状
top[0]->ReshapeLike(batch->data_);
// Copy the data
// 将batch数据复制到top[0]
caffe_copy(batch->data_.count(), batch->data_.cpu_data(),
top[0]->mutable_cpu_data());
DLOG(INFO) << "Prefetch copied";
if (this->output_labels_) {
// 输出类标的话
// Reshape to loaded labels.
// 根据batch中类标的形状改变top[1]的形状
top[1]->ReshapeLike(batch->label_);
// Copy the labels.
// 复制类标到top[1]
caffe_copy(batch->label_.count(), batch->label_.cpu_data(),
top[1]->mutable_cpu_data());
}
// 将该batch压入free队列
prefetch_free_.push(batch);
}三、总结
首先理顺类与类之间的关系：
Layer类是所有神经网络层的基类，BaseDataLayer继承自该类，BasePrefetchingDataLayer继承自BaseDataLayer，DataLayer继承自BasePrefetchingDataLayer。
有了上述几个基础的类之后，其他的类都是从这几个类进行派生。
比如DummyDataLayer，HDF5Layer和HDF5OutputLayer都是直接继承自Layer。
MemoryDataLayer则是继承自BaseDataLayer
凡是涉及到直接读取数据文件的一般都是继承自BasePrefetchingDataLayer，这样可以有效地读数据进行预取。
比如：ImageDataLayer、WindowDataLayer
继承自BasePrefetchingDataLayer需要实现load_batch函数以供内部的线程进行调用，实现数据预取。
此外每一个网络层的类（因为所有的网络层都继承自Layer类嘛）都需要实现SetUp，这个是必须的。
这一次的量还真有点大。。。
注释的代码可以从以下位置下载：
http://download.csdn.net/detail/xizero00/9474806
参考：
[1]HDF5格式的介绍
1029
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
