Snapshot的存储
概述
Snapshot的存储格式有两种，分别是BINARYPROTO格式和hdf5格式。BINARYPROTO是一种二进制文件，并且可以通过修改shapshot_format来设置存储类型。该项的默认是BINARYPROTO。不管哪种格式，运行的过程是类似的，都是从Solver<Dtype>::Snapshot()函数进入，首先调用Net网络的方法，再操作网络中的每一层，最后再操作每一层中blob，最后调用write函数写入输出。源码入口：
1 void Solver<Dtype>::Snapshot() { 2 CHECK(Caffe::root_solver()); 3 string model_filename; 4 switch (param_.snapshot_format()) { 5 case caffe::SolverParameter_SnapshotFormat_BINARYPROTO: 6 model_filename = SnapshotToBinaryProto(); 7 break; 8 case caffe::SolverParameter_SnapshotFormat_HDF5: 9 model_filename = SnapshotToHDF5(); 10 break; 11 default: 12 LOG(FATAL) << "Unsupported snapshot format."; 13 }
BINARYPROTO格式
如果是BINARYPROTO的存储格式，就执行如下代码：
1 string Solver<Dtype>::SnapshotToBinaryProto() { 2 string model_filename = SnapshotFilename(".caffemodel"); 3 LOG(INFO) << "Snapshotting to binary proto file " << model_filename; 4 NetParameter net_param; 5 net_->ToProto(&net_param, param_.snapshot_diff()); 6 WriteProtoToBinaryFile(net_param, model_filename); 7 return model_filename; 8 }
首先会执行SnapshotFilename(“.caffemodel”)函数，识别出sovler.prototxt文件中snapshot_prefix的内容，作用该snapshot文件的文件名前缀。然后调用net_->ToProto()，具体的代码如下：
1 void Net<Dtype>::ToProto(NetParameter* param, bool write_diff) const { 2 param->Clear(); 3 param->set_name(name_); 4 for (int i = 0; i < net_input_blob_indices_.size(); ++i) { 5 param->add_input(blob_names_[net_input_blob_indices_[i]]); 6 } 7 for (int i = 0; i < layers_.size(); ++i) { 8 LayerParameter* layer_param = param->add_layer(); 9 layers_[i]->ToProto(layer_param, write_diff); 10 } 11 }
获取到网络中的每层的名字等参数后，调用layers_[i]->ToProto()每一层的ToProto方法，接下来
1 void Layer<Dtype>::ToProto(LayerParameter* param, bool write_diff) { 2 param->Clear(); 3 param->CopyFrom(layer_param_); 4 param->clear_blobs(); 5 for (int i = 0; i < blobs_.size(); ++i) { 6 blobs_[i]->ToProto(param->add_blobs(), write_diff); 7 } 8 }
然后调用当前层下的所有blob的ToProto方法，即：
1 void Blob<double>::ToProto(BlobProto* proto, bool write_diff) const { 2 proto->clear_shape(); 3 for (int i = 0; i < shape_.size(); ++i) { 4 proto->mutable_shape()->add_dim(shape_[i]); 5 } 6 proto->clear_double_data(); 7 proto->clear_double_diff(); 8 const double* data_vec = cpu_data(); 9 for (int i = 0; i < count_; ++i) { 10 proto->add_double_data(data_vec[i]); 11 } 12 if (write_diff) { 13 const double* diff_vec = cpu_diff(); 14 for (int i = 0; i < count_; ++i) { 15 proto->add_double_diff(diff_vec[i]); 16 } 17 }
在每一个blob中，会调用add_double_data()函数，把data添加到snapshot文件中，同时会判断是否当前blob参与diff的计算，如果需要当前blob需要diff参数，就调用add_double_diff()添加到snapshot文件中。
调用完所有的blob的ToProto()方法后，会执行WriteProtoToBinaryFile()把该文件写出即可。
1 void WriteProtoToBinaryFile(const Message& proto, const char* filename) { 2 fstream output(filename, ios::out | ios::trunc | ios::binary); 3 CHECK(proto.SerializeToOstream(&output)); 4 }
在该方法里调用FStream的output方法进行输出。
Hdf5格式
Hdf5格式的运行过程和BINARYPROTO格式的过程类似，首先会调用SnapshotToHDF5()函数，即：
1 string Solver<Dtype>::SnapshotToHDF5() { 2 string model_filename = SnapshotFilename(".caffemodel.h5"); 3 LOG(INFO) << "Snapshotting to HDF5 file " << model_filename; 4 net_->ToHDF5(model_filename, param_.snapshot_diff()); 5 return model_filename; 6 }
首先会执行SnapshotFilename(“.caffemodel.h5”)函数，识别出sovler.prototxt文件中snapshot_prefix的内容，作用该snapshot文件的文件名前缀。然后调用net_->ToHDF5()，即：
1 void Net<Dtype>::ToHDF5(const string& filename, bool write_diff) const { 2 hid_t file_hid = H5Fcreate(filename.c_str(), H5F_ACC_TRUNC, H5P_DEFAULT, 3 H5P_DEFAULT); 4 hid_t data_hid = H5Gcreate2(file_hid, "data", H5P_DEFAULT, H5P_DEFAULT, 5 H5P_DEFAULT); 6 hid_t diff_hid = -1; 7 if (write_diff) { 8 diff_hid = H5Gcreate2(file_hid, "diff", H5P_DEFAULT, H5P_DEFAULT, 9 H5P_DEFAULT); 10 } 11 for (int layer_id = 0; layer_id < layers_.size(); ++layer_id) { 12 const LayerParameter& layer_param = layers_[layer_id]->layer_param(); 13 string layer_name = layer_param.name(); 14 hid_t layer_data_hid = H5Gcreate2(data_hid, layer_name.c_str(), 15 hid_t layer_diff_hid = -1; 16 if (write_diff) { 17 layer_diff_hid = H5Gcreate2(diff_hid, layer_name.c_str(), 18 H5P_DEFAULT, H5P_DEFAULT, H5P_DEFAULT); 19 } 20 int num_params = layers_[layer_id]->blobs().size(); 21 for (int param_id = 0; param_id < num_params; ++param_id) { 22 ostringstream dataset_name; 23 dataset_name << param_id; 24 const int net_param_id = param_id_vecs_[layer_id][param_id]; 25 if (param_owners_[net_param_id] == -1) { 26 hdf5_save_nd_dataset<Dtype>(layer_data_hid, dataset_name.str(), 27 *params_[net_param_id]); 28 } 29 if (write_diff) { 30 hdf5_save_nd_dataset<Dtype>(layer_diff_hid, dataset_name.str(), 31 *params_[net_param_id], true); 32 } 33 ............... 34 H5Fclose(file_hid); 35 }
该函数首先调用H5Fcreate()创建一个file文件，然后循环调用每一层，通过调用每一层的H5Gcreate2函数记录出该层的data_hid或者diff_hid(如果该层需要参与计算)，然后进入每一层内部的blob，然后在当前blob内调用hdf5_save_nd_dataset()或hdf5_save_nd_dataset()(如果当前blob需要参与计算diff)，将data添加到hdf5格式的文件中，最后调用H5Fclose(file_hid)函数，输出该文件。
Snapshot的恢复
概述
想在已经训练好的网络上继续训练，那么需要调用Restore()方法从snapshot的文件中恢复成网络，从而缩短了训练时间。方法的入口是Solver<Dtype>::Restore(const char* state_file)函数，即：
1 void Solver<Dtype>::Restore(const char* state_file) { 2 CHECK(Caffe::root_solver()); 3 string state_filename(state_file); 4 if (state_filename.size() >= 3 && 5 state_filename.compare(state_filename.size() - 3, 3, ".h5") == 0) { 6 RestoreSolverStateFromHDF5(state_filename); 7 } else { 8 RestoreSolverStateFromBinaryProto(state_filename); 9 }
该函数会解析snapshot文件是BINARYPROTO格式还是Hdf5格式，如果是BINARYPROTO格式的话就调用RestoreSolverStateFromBinaryProto()函数，如果格式Hdf5的格式，就执行RestoreSolverStateFromHDF5()。
BINARYPROOTO格式
如果是BINARYPROTO格式，则执行下列代码：
1 void SGDSolver<Dtype>::RestoreSolverStateFromBinaryProto( 2 const string& state_file) { 3 SolverState state; 4 ReadProtoFromBinaryFile(state_file, &state); 5 this->iter_ = state.iter(); 6 if (state.has_learned_net()) { 7 NetParameter net_param; 8 ReadNetParamsFromBinaryFileOrDie(state.learned_net().c_str(), &net_param); 9 this->net_->CopyTrainedLayersFrom(net_param); 10 } 11 this->current_step_ = state.current_step(); 12 CHECK_EQ(state.history_size(), history_.size()) 13 << "Incorrect length of history blobs."; 14 for (int i = 0; i < history_.size(); ++i) { 15 history_[i]->FromProto(state.history(i)); 16 } 17 }
该函数会大量调用google的protobuf包内的函数，首先会通过ReadProtoFromBinaryFile()函数读取BINARYPROTO格式的文件来返回是否可以成功读取。然后判断该snapshot是否有曾经训练过的网络，如果有，则调用函数ReadNetParamsFromBinaryFileOrDie()读取出该Net网络，然后调用函数CopyTrainedLayersFrom(net_param)具体恢复该网络的每一层以及当前层内的所有blob，具体数据恢复的工作就是CopyTrainedLayersFrom()函数内部变量调用FromProto()函数来实现blob复制的。然后会通过函数current_step()来判断上次训练的位置(迭代到多少次)，然后通过循环把训练过的data数据通过FromProto()完成数据的复制。
Hdf5格式
1 void SGDSolver<Dtype>::RestoreSolverStateFromHDF5(const string& state_file) { 2 hid_t file_hid = H5Fopen(state_file.c_str(), H5F_ACC_RDONLY, H5P_DEFAULT); 3 CHECK_GE(file_hid, 0) << "Couldn't open solver state file " << state_file; 4 this->iter_ = hdf5_load_int(file_hid, "iter"); 5 if (H5LTfind_dataset(file_hid, "learned_net")) { 6 string learned_net = hdf5_load_string(file_hid, "learned_net"); 7 this->net_->CopyTrainedLayersFrom(learned_net); 8 } 9 this->current_step_ = hdf5_load_int(file_hid, "current_step"); 10 hid_t history_hid = H5Gopen2(file_hid, "history", H5P_DEFAULT); 11 CHECK_GE(history_hid, 0) << "Error reading history from " << state_file; 12 int state_history_size = hdf5_get_num_links(history_hid); 13 CHECK_EQ(state_history_size, history_.size()) 14 << "Incorrect length of history blobs."; 15 for (int i = 0; i < history_.size(); ++i) { 16 ostringstream oss; 17 oss << i; 18 hdf5_load_nd_dataset<Dtype>(history_hid, oss.str().c_str(), 0, 19 kMaxBlobAxes, history_[i].get()); 20 } 21 H5Gclose(history_hid); 22 H5Fclose(file_hid); 23 }
该函数会识别hdf5格式存储的snapshot文件的file_hid编号，会判断是否存在之前训练过的网络，如果存在则执行CopyTrainedLayersFrom()函数，完成网络的每层以及每层内的blob的数据的恢复复制，然后或取上一次的训练位置(进行的迭代)，并且调用函数hdf5_load_nd_dataset()具体把每次迭代的数据恢复复制，最后再调用H5Fclose()关闭。