Caffe学习--Net分析
Caffe_Net
1.基本数据
vector<shared_ptr<Layer<Dtype> > > layers_; // 记录每一层的layer参数
vector<vector<Blob<Dtype>*> > bottom_vecs_;
vector<vector<int> > bottom_id_vecs_;
vector<vector<bool> > bottom_need_backward_;
/// top_vecs stores the vectors containing the output for each layer
vector<vector<Blob<Dtype>*> > top_vecs_;
vector<vector<int> > top_id_vecs_;
vector<vector<int> > param_id_vecs_;
vector<string> layer_names_;
//learnable_params_[learnable_param_ids_[i]] == params_[i].get()
vector<Blob<Dtype>*> learnable_params_;//层间权重与bias
2. 常用的函数
介绍了Caffe内的Net的常用函数:
const string& name(){return name_;}//网络的名称
const vector<string>& layer_names{return layer_names_;}// net每层的layer名称
// net内每层的layer的Blob名称
const vector<string>& blob_names(){return blob_names_;}
//net内层次间的权值与bias
const vector<shared_ptr<Blob<Dtype>>>& blobs(){return blob_;};
//net内的layers
const vector<shared_ptr<Layer<Dtype>>>& layers(){return layers_;};
//net->bottom_vecs() 返回该layer的输入，输出向量，
//以及具体的 top_id_vecs[layer_id][top_id];
const vector<vector<Blob<Dtype>*> >& bottom_vecs(){ return bottom_vecs_;}
const vector<vector<Blob<Dtype>*> >& top_vecs() { return top_vecs_;}
const vector<vector<int> >& bottom_id_vecs(){ return bottom_id_vecs_;}
const vector<vector<int> >& top_id_vecs() { return top_id_vecs_;}
void CopyTrainedLayersFrom(const string trained_filename);//加载权重
//网络的输入输出
//感觉等效于bottom_vecs_[0]
const vector<Blob<Dtype>*>& input_blobs(){return net_input_blobs_;}
const vector<Blob<Dtype>*>& output_blobs()
{return net_output_blobs;}//top_vecs[top_vecs.size()-1];
const int num_input(){return net_input_blobs_.size()};//输入blob的size
//has_blob()然后find return
const shared_ptr<Blob<Dtype>>blob_by_name(const string& blob_name);
// 前向计算loss和网络的输出
const vector<Blob<Dtype>*>& forward(Dtype* loss = NULL);
// --- *loss = ForwardFromTo(0.layers_.size()-1);
// --- 此处调用 Dtype* Net<Dtype>::ForwardFrom(int start,int end)
for (size_t i = start; i < end; i++){
//多态，调用具体的Layer的Forward函数,并返回该层次的loss
Dtype layer_loss = layers_[i]->Forward(bottom_vecs_[i],top_vecs_[i]);
loss += layer_loss;
}
return loss;
// backward反向，更新权值
void Net<Dtype>::Backward(){ //
BackwardFromTo(layers_size()-1,0); // 具体函数实现如第三部分
if (debug_info_) {
/*层次的参数*/
}
}
3.具体函数实现
template <typename Dtype>
const int Net<Dtype>::AppendBottom(const NetParamter& param, int layer_id,
int bottom_id,set<string>* availabel_blobs,map<string,int>* blob_name_to_idx){
const LayerParammeter& layer_param = param.layer(layer_id);
const string& blob_name = layer_param.bottom(bottom_id);
const int blob_id = (*blob_name_to_idx)[blob_name];
//layer输入的shape等
bottom_vecs_[layer_id].push_back(blobs_[blob_id].get());
bottom_id_vecs_[layer_id].push_back(blob_id);
//LOG CONV<--data 等,只要是丢入输入
}
// learnable_params_
//conv的shape一般为num_output*input_channels*kernel_width*kernel_height
//bias的shape一般为Num_output
template <typename Dtype>
void Net<Dtype>::AppendParam(const NetParameter& param, const int layer_id,
const int param_id) {
const int learnable_param_id = learnable_params_.size();
learnable_params_.push_back(params_[net_param_id].get());
learnable_param_ids_.push_back(learnable_param_id);
has_params_lr_.push_back(param_spec->has_lr_mult());
has_params_decay_.push_back(param_spec->has_decay_mult());
params_lr_.push_back(param_spec->lr_mult());
params_weight_decay_.push_back(param_spec->decay_mult());
}
template <typename Dtype>
void Net<Dtype>::BackwardFromTo(int start,int end){
for(int i = start;i >= end;--i){
//backward 调用各层次的backward更新权值和bias
layers_[i].Backward(top_vecs_[i],bottom_need_backward_[i],
bottom_vecs_[i]);
}
}
4.基本流程
基本流程：Net构造函数开始
// 递归更新变量
vectot<string>*stage ;
int level;
//起始调用
net_.reset(new Net<float>(model_file, TEST));
//送入prototxt文件和Test OR Train
explicit Net(const string& param_file,Phase phase,const int level = 0,
vector<string>* stage = NULL,const Net* root_net = NULL);
// 解析保存在NetParamter param内,这里用到了
//protobuf::TextFormat::Parse(FileInputStream*,param)
ReadNetParamsFromTextFileOrDie(param_file,¶m);
// 读取了NetParamter 后需要进行整个网络的初始化工作
Init(param); //初始化网络的接口，下续为具体实现
FilterNet(param, &filtered_param);// 打印网络结构
/*内部会完成split added 如果有必要(残差结构),记录层与层之间的联系关系与层次的名称
等，是否有loss_weight，layer的size等*/
InsertSplits(filtered_param,¶m);
for (size_t i = 0; i < param.layer_size(); i++) { //遍历setupLayer
const LayerParammeter& layer_param = param.layer(i);//层次的参数
layers_.push_back(LayerRegistry<Dtype>::CreateLayer(layer_param));
// CreateLayer会走layer_factory的CreateLayer的注册 ,比如input,conv,bn...
layer_names_.push_back(layer_param.name());
//开始继续遍历每层输入的具体的细节,第i个layer的第botom_id个输入
for (size_t bottom_id = 0; bottom_id < layer_param.bottom_size();
bottom_id++) {
const int blob_id =
AppendBottom(param,i,bottom_id,&availabel_blobs,&blob_name_to_idx);
}
//开始继续遍历每层输出的具体细节，第i个layer的第 top_id的输出
for (size_t top_id = 0; top_id < layer_param.top_size();
top_id++) {
AppendTop(param,i,top_id,&availabel_blobs,&blob_name_to_idx);
if (layer_param.type()== "Input") {//输入
const int blob_id = blobs_.size() - 1;
net_input_blob_indices_.push_back(blob_id);
net_input_blobs_.push_back(blobs_[blob_id].get());
}
}
//多态，具体调用具体的layer的Setup函数
layers_[layer_id]->SetUp(bottom_vecs_[layer_id], top_vecs_[layer_id]);
//每个输出遍历
for (size_t top_id = 0; top_id < top_vecs_[layer_id].size();
top_id++) {
/*完成层次的blob_loss_weights,并统计memory_used_*/;
memory_used_ += top_vecs_[layer_id][top_id]->count();
}
//总的memory_used_: memory_used_*sizeof(Dtype);
//如果层次间有学习权值和偏置，则需要再次设置，比如conv
//num_param_blobs weights And bias
for (int param_id = 0; param_id < num_param_blobs; ++param_id) {
AppendParam(param, layer_id, param_id);
}
}
/*接下来需要研究网络的backwards问题，决定哪些层次对loss有贡献，并且检查哪些层次
不需要back_propagate_down操作，遍历是反向的操作*/
for (size_t layer_id = layers_.size()-1; layer_id >= 0; --layer_id){
bool layer_contributes_loss = false;//默认是无贡献的
bool layer_skip_propagate_down = true;// 默认不参与backwards的loss贡献
//Layer内的输出遍历
for (size_t top_id = 0; top_id < top_vecs_[layer_id].size();
top_id++) {
//blob_name_[index]名字
string& blob_name = blob_names_[top_id_vecs_[layer_id][top_id]];
if (layer_[layer_id]->loss(top_id)||
blobs_under_loss.find(blob_name) != blobs_under_loss.end()) {
//该层次的layerloss不为0或者loss_weight = 1;
layer_contributes_loss = true;
}
if (blobs_skip_backp.find(blob_name) == blobs_skip_backp.end()) {
layer_skip_propagate_down = false;
}
}
//同理 Layer内的输入遍历
for (size_t bottom_id = 0; bottom_id < bottom_vecs_[layer_id].size();
bottom_id++) {
if (layer_contributes_loss) {
string* blob_name = blob_names_[bottom_id_vecs_[layer_id][bottom_id]];
blobs_under_loss.insert(blob_name);
}
else{
bottom_need_backward_[layer_id][bottom_id] = false;
}
if (!bottom_need_backward_[layer_id][bottom_id]) {
string&blob_name = blob_names_[bottom_id_vecs_[layer_id][bottom_id]];
blok_skip_backp.insert(blob_name);
}
}
/*code*/
}//init函数尾
