神经网络：caffe特征可视化的代码样例

caffe特征可视化的代码样例

不少读者看了我前面两篇文章

总结一下用caffe跑图片数据的研究流程

deep learning实践经验总结2--准确率再次提升，到达0.8，再来总结一下

之后，想知道我是怎么实现特征可视化的。

简单来说，其实就是让神经网络正向传播一次，然后把某层的特征值给取出来，然后转换为图片保存。

下面我提供一个demo，大家可以根据自己的需求修改。

先看看我的demo的使用方法。

visualize_features.bin net_proto pretrained_net_proto iterations  [CPU/GPU]  img_list_file dstdir laydepth

visualize_features.bin是cpp编译出来的可执行文件

下面看看各参数的意义：

1 net_proto：caffe规定的一种定义网络结构的文件格式，后缀名为".prototxt"。这个文件定义了网络的输入，已经相关参数，还有就是整体的网络结构。

2 pretrained_net_proto：这个是已经训练好了的模型

3 iterations：迭代次数

4 [CPU/GPU]：cpu还是gpu模式

5 img_list_file：待测试的文件名列表。我这里需要这个主要是为了得到图片的类名。

6 dstdir：图片输出的文件夹

7 laydepth：需要输出哪一层的特征

下面是一个实例例子：

./visualize_features.bin /home/linger/linger/caffe-action/caffe-master/examples/cifar10/cifar10_full_test.prototxt /home/linger/linger/caffe-action/caffe-master/examples/cifar10/cifar10_full_iter_60000 20 GPU /home/linger/linger/testfile/skirt_test_attachment/image_filename /home/linger/linger/testfile/innerproduct/ 7

下面是源代码：

// Copyright 2013 Yangqing Jia//// This is a simple script that allows one to quickly test a network whose// structure is specified by text format protocol buffers, and whose parameter// are loaded from a pre-trained network.// Usage://    test_net net_proto pretrained_net_proto iterations [CPU/GPU]#include <cuda_runtime.h>#include <fstream>#include <iostream>#include <cstring>#include <cstdlib>#include <algorithm>#include <vector>#include <utility>#include "caffe/caffe.hpp"#include <opencv2/highgui/highgui.hpp>#include <opencv2/highgui/highgui_c.h>#include <opencv2/imgproc/imgproc.hpp>using std::make_pair;using std::pair;using namespace caffe;  // NOLINT(build/namespaces)using namespace std;vector<string> fileNames;char * filelist;/* * 读入的文件的内容格式类似这样子的：全局id 类名_所在类的id.jpg0 一步裙_0.jpg1 一步裙_1.jpg2 一步裙_10.jpg */void readFile(){if(fileNames.empty()){ifstream read(filelist);//"/home/linger/linger/testfile/test_attachment/image_filename"// "/home/linger/imdata/test_files_collar.txt"//  "/home/linger/linger/testfile/testfilename"if(read.is_open()){while(!read.eof()){string name;int id;read>>id>>name;fileNames.push_back(name);}}}}/* * 根据图片id获取类名 */string getClassNameById(int id){readFile();int index = fileNames[id].find_last_of('_') ;return fileNames[id].substr(0, index);}void writeBatch(const float* data,int num,int channels,int width,int height,int startID,const char*dir){for(int id = 0;id<num;id++){for(int channel=0;channel<channels;channel++){cv::Mat mat(height,width, CV_8UC1);//高宽vector<vector<float> > vec;vec.resize(height);float max = -1;float min = 999999;for(int row=0;row<height;row++){vec[row].resize(width);for(int col=0;col<width;col++){vec[row][col] =data[id*channels*width*height+channel*width*height+row*width+col];if(max<vec[row][col]){max = vec[row][col];}if(min>vec[row][col]){min = vec[row][col];}}}for(int row=0;row<height;row++){  for(int col=0;col<width;col++){vec[row][col] = 255*((float)(vec[row][col]-min))/(max-min);uchar& img = mat.at<uchar>(row,col);img= vec[row][col];}}char filename[100];string label = getClassNameById(startID+id);string file_reg =dir;file_reg+="%s%05d_%05d.png";snprintf(filename, 100, file_reg.c_str(), label.c_str(),startID+id,channel);//printf("%s\n",filename);cv::imwrite(filename, mat);}}}int main(int argc, char** argv){  if (argc < 4)  {    LOG(ERROR) << "visualize_features.bin net_proto pretrained_net_proto iterations "        << "[CPU/GPU] img_list_file dstdir laydepth";    return 0;  }  /*  ./visualize_features.bin /home/linger/linger/caffe-action/caffee-ext/Caffe_MM/prototxt/triplet/triplet_test_simple.prototxt /home/linger/linger/caffe-action/caffee-ext/Caffe_MM/snapshorts/_iter_100000 8 GPU /home/linger/linger/testfile/test_attachment/image_filename /home/linger/linger/testfile/innerproduct/ 6  */  filelist = argv[5];  cudaSetDevice(0);  Caffe::set_phase(Caffe::TEST);  if (argc == 5 && strcmp(argv[4], "GPU") == 0)  {    LOG(ERROR) << "Using GPU";    Caffe::set_mode(Caffe::GPU);  }  else  {    LOG(ERROR) << "Using CPU";    Caffe::set_mode(Caffe::CPU);  }  NetParameter test_net_param;  ReadProtoFromTextFile(argv[1], &test_net_param);  Net<float> caffe_test_net(test_net_param);  NetParameter trained_net_param;  ReadProtoFromBinaryFile(argv[2], &trained_net_param);  caffe_test_net.CopyTrainedLayersFrom(trained_net_param);  int total_iter = atoi(argv[3]);  LOG(ERROR) << "Running " << total_iter << " Iterations.";  double test_accuracy = 0;  vector<Blob<float>*> dummy_blob_input_vec;  int startID = 0;  int nums;  int dims;  int batchsize = test_net_param.layers(0).layer().batchsize();  int laynum = caffe_test_net.bottom_vecs().size();  printf("num of layers:%d\n",laynum);  for (int i = 0; i < total_iter; ++i)  {    const vector<Blob<float>*>& result =        caffe_test_net.Forward(dummy_blob_input_vec);    int laydepth = atoi(argv[7]);    Blob<float>* features = (*(caffe_test_net.bottom_vecs().begin()+laydepth))[0];//调整第几层即可    nums = features->num();    dims= features->count()/features->num();    int num = features->num();    int channels = features->channels();    int width = features->width();    int height = features->height();    printf("channels:%d,width:%d,height:%d\n",channels,width,height);    writeBatch(features->cpu_data(),num,channels,width,height,startID,argv[6]);    startID += nums;  }  return 0;}