caffe-----使用C++ 提取网络中间层特征数据
最近实验，想要在c++下知道网络中间某一层的特征数据情况，查找了相关资料，记录一下。
其实在caffe框架里面是包含这种操作的，可以模仿tools/extract_features.cpp中的操作来得到网络中间的特征数据。
首先看下extract_features.cpp是如何写的。
template<typename Dtype> int feature_extraction_pipeline(int argc, char** argv) { ::google::InitGoogleLogging(argv[0]); const int num_required_args = 7; if (argc < num_required_args) { LOG(ERROR)<< "This program takes in a trained network and an input data layer, and then" " extract features of the input data produced by the net.\n" "Usage: extract_features pretrained_net_param" " feature_extraction_proto_file extract_feature_blob_name1[,name2,...]" " save_feature_dataset_name1[,name2,...] num_mini_batches db_type" " [CPU/GPU] [DEVICE_ID=0]\n" "Note: you can extract multiple features in one pass by specifying" " multiple feature blob names and dataset names separated by ','." " The names cannot contain white space characters and the number of blobs" " and datasets must be equal."; return 1; } int arg_pos = num_required_args; arg_pos = num_required_args; if (argc > arg_pos && strcmp(argv[arg_pos], "GPU") == 0) { LOG(ERROR)<< "Using GPU"; int device_id = 0; if (argc > arg_pos + 1) { device_id = atoi(argv[arg_pos + 1]); CHECK_GE(device_id, 0); } LOG(ERROR) << "Using Device_id=" << device_id; Caffe::SetDevice(device_id); Caffe::set_mode(Caffe::GPU); } else { LOG(ERROR) << "Using CPU"; Caffe::set_mode(Caffe::CPU); } arg_pos = 0; // the name of the executable std::string pretrained_binary_proto(argv[++arg_pos]); // Expected prototxt contains at least one data layer such as // the layer data_layer_name and one feature blob such as the // fc7 top blob to extract features. /* layers { name: "data_layer_name" type: DATA data_param { source: "/path/to/your/images/to/extract/feature/images_leveldb" mean_file: "/path/to/your/image_mean.binaryproto" batch_size: 128 crop_size: 227 mirror: false } top: "data_blob_name" top: "label_blob_name" } layers { name: "drop7" type: DROPOUT dropout_param { dropout_ratio: 0.5 } bottom: "fc7" top: "fc7" } */ std::string feature_extraction_proto(argv[++arg_pos]); boost::shared_ptr<Net<Dtype> > feature_extraction_net( new Net<Dtype>(feature_extraction_proto, caffe::TEST)); feature_extraction_net->CopyTrainedLayersFrom(pretrained_binary_proto);//初始化网络 std::string extract_feature_blob_names(argv[++arg_pos]); std::vector<std::string> blob_names; boost::split(blob_names, extract_feature_blob_names, boost::is_any_of(",")); std::string save_feature_dataset_names(argv[++arg_pos]); std::vector<std::string> dataset_names; boost::split(dataset_names, save_feature_dataset_names, boost::is_any_of(",")); CHECK_EQ(blob_names.size(), dataset_names.size()) << " the number of blob names and dataset names must be equal"; size_t num_features = blob_names.size(); for (size_t i = 0; i < num_features; i++) { CHECK(feature_extraction_net->has_blob(blob_names[i])) << "Unknown feature blob name " << blob_names[i] << " in the network " << feature_extraction_proto; } int num_mini_batches = atoi(argv[++arg_pos]); std::vector<boost::shared_ptr<db::DB> > feature_dbs; std::vector<boost::shared_ptr<db::Transaction> > txns; const char* db_type = argv[++arg_pos]; for (size_t i = 0; i < num_features; ++i) { LOG(INFO)<< "Opening dataset " << dataset_names[i]; boost::shared_ptr<db::DB> db(db::GetDB(db_type)); db->Open(dataset_names.at(i), db::NEW); feature_dbs.push_back(db); boost::shared_ptr<db::Transaction> txn(db->NewTransaction()); txns.push_back(txn); } LOG(ERROR)<< "Extracting Features"; Datum datum; std::vector<int> image_indices(num_features, 0); for (int batch_index = 0; batch_index < num_mini_batches; ++batch_index) { feature_extraction_net->Forward();//首先进行前传 这样才能有中间数据 for (int i = 0; i < num_features; ++i) { const boost::shared_ptr<Blob<Dtype> > feature_blob = feature_extraction_net->blob_by_name(blob_names[i]);//通过名字查找blob int batch_size = feature_blob->num(); int dim_features = feature_blob->count() / batch_size; const Dtype* feature_blob_data; for (int n = 0; n < batch_size; ++n) { datum.set_height(feature_blob->height()); datum.set_width(feature_blob->width()); datum.set_channels(feature_blob->channels()); datum.clear_data(); datum.clear_float_data(); feature_blob_data = feature_blob->cpu_data() + feature_blob->offset(n); for (int d = 0; d < dim_features; ++d) { datum.add_float_data(feature_blob_data[d]);//将feature_blob的数据都保存到datum里 } string key_str = caffe::format_int(image_indices[i], 10); string out; CHECK(datum.SerializeToString(&out));//将datum保存到本地 txns.at(i)->Put(key_str, out); ++image_indices[i]; if (image_indices[i] % 1000 == 0) { txns.at(i)->Commit(); txns.at(i).reset(feature_dbs.at(i)->NewTransaction()); LOG(ERROR)<< "Extracted features of " << image_indices[i] << " query images for feature blob " << blob_names[i]; } } // for (int n = 0; n < batch_size; ++n) } // for (int i = 0; i < num_features; ++i) } // for (int batch_index = 0; batch_index < num_mini_batches; ++batch_index) // write the last batch for (int i = 0; i < num_features; ++i) { if (image_indices[i] % 1000 != 0) { txns.at(i)->Commit(); } LOG(ERROR)<< "Extracted features of " << image_indices[i] << " query images for feature blob " << blob_names[i]; feature_dbs.at(i)->Close(); } LOG(ERROR)<< "Successfully extracted the features!"; return 0; }
主要三个核心步骤：
1.初始化网络，并前传，
net->Forward()
2.通过blob的名字（prototxt中的name）来得到blob数据，
const boost::shared_ptr<Blob<Dtype> > feature_blob = net->blob_by_name(blob_names[i])
3.blob里面已经保存了所有的特征数据，按照需求取出来就好了。
count = feature_blob->channels() * feature_blob->height() * feature_blob->width(); float* feature_array = new float[count]; const float* feature_blob_data = feature_blob->cpu_data() + feature_blob->offset(n); // feature data generated from // the nth input image within a batch memcpy(feature_array, feature_blob_data, count * sizeof(float)); ...// other operations delete [] feature_array;
如下是做实验时候的一个例子，提取出了blstm_input中的数据，并保存到了txt里。
Blob<float>* input_layer = m_net->input_blobs()[0]; input_layer->Reshape(1, m_channelNum, m_inputGeometry.height, m_inputGeometry.width); m_net->Reshape(); std::vector<cv::Mat> input_channels; wrapInputLayer(&input_channels); preprocess(img, &input_channels); m_net->Forward(); Blob<float>* output_layer = m_net->output_blobs()[0]; int alphabet_size=output_layer->shape(2); int time_step=output_layer->shape(0); vector<int> shape; const boost::shared_ptr<Blob<float> > blstm_input = m_net->blob_by_name("blstm_input"); shape = blstm_input->shape(); for(int i = 0; i < shape.size(); i++) { cout<<" blstm_input shape:"<<i<<" :"<<shape[i]<<endl; } const boost::shared_ptr<Blob<float> > lstm1 = m_net->blob_by_name("lstm1"); shape = lstm1->shape(); for(int i = 0; i < shape.size(); i++) { cout<<" lstm1 shape:"<<i<<" :"<<shape[i]<<endl; } cout<<"==============blob info======="<<endl; ofstream of("blstm.txt"); for(int h = 0; h < 192; h++) { int count = blstm_input->channels() * blstm_input->height() * blstm_input->width(); // cout<<"blstm_input->channels():"<<blstm_input->channels()<<" blstm_input->height():"<<blstm_input->height() // <<" blstm_input->width():"<<blstm_input->width()<<endl; float* feature_array = new float[count]; const float* feature_blob_data = blstm_input->cpu_data() + blstm_input->offset(h); // feature data generated from the nth input image within a batch memcpy(feature_array, feature_blob_data, count * sizeof(float)); for(int i = 0; i < count; i++ ) { if(i && i % 512 == 0) { of<<endl; } of<<" ["<< h<< ","<<i % 512<< "]:"<<feature_blob_data[i]; } of<<endl; delete [] feature_array; } of.close();
参考：
