caffe 版本 yolo 过程记录

【注】后续再写可训练版本的yolo caffe版
【caffe训练版本已经更新caffe-yolo训练】
YOLO 框的获取

 x = image.cols*(result[idx++]+j)/7;//i表示gride的位置 y = image.rows*(result[idx++]+i)/7;//j表示gride的位置 w = image.cols*result[idx]*result[idx++]; h = image.rows*result[idx]*result[idx];

一、yolo 网络 weight 准备

yolo网络里面含有权重的层分别为卷积层和全链接层，因此，我对yolo里面将每一个含有权重的层进行了存储权重的操作，并将它们分别存储在txt中，下面的程序是loadweights函数实现的是将txt中的权重赋值给net。
下面涉及到的大部分文件都在百度盘中（链接：http://pan.baidu.com/s/1hs8XCW0 密码：8rcu）

二、yolo网络prototxt的准备

这里只是展示一下yolo-tiny-deploy.prototxt。 yolo-deploy.prototxt可以参见百度盘

name: "yolotiny"input: "data"input_dim: 1input_dim: 3input_dim: 448input_dim: 448layer{  name: "conv1"  type: "Convolution"  bottom: "data"  top: "conv1"  convolution_param{    num_output: 16    kernel_size: 3    stride: 1    pad: 1  }}layer{  name: "Leaky1"  type: "Leaky"  bottom: "conv1"  top: "Leaky1"}layer{  name: "pool1"  type: "Pooling"  bottom: "Leaky1"  top: "pool1"  pooling_param {    pool: MAX    kernel_size: 2    stride: 2  }}layer{  name: "conv2"  type: "Convolution"  bottom: "pool1"  top: "conv2"  convolution_param{    num_output: 32    kernel_size: 3    stride: 1    pad: 1  }}layer{  name: "Leaky2"  type: "Leaky"  bottom: "conv2"  top: "Leaky2"}layer{  name: "pool2"  type: "Pooling"  bottom: "Leaky2"  top: "pool2"  pooling_param {    pool: MAX    kernel_size: 2    stride: 2  }}layer{  name: "conv3"  type: "Convolution"  bottom: "pool2"  top: "conv3"  convolution_param{    num_output: 64    kernel_size: 3    stride: 1    pad: 1  }}layer{  name: "Leaky3"  type: "Leaky"  bottom: "conv3"  top: "Leaky3"}layer{  name: "pool3"  type: "Pooling"  bottom: "Leaky3"  top: "pool3"  pooling_param {    pool: MAX    kernel_size: 2    stride: 2  }}layer{  name: "conv4"  type: "Convolution"  bottom: "pool3"  top: "conv4"  convolution_param{    num_output: 128    kernel_size: 3    stride: 1    pad: 1  }}layer{  name: "Leaky4"  type: "Leaky"  bottom: "conv4"  top: "Leaky4"}layer{  name: "pool4"  type: "Pooling"  bottom: "Leaky4"  top: "pool4"  pooling_param {    pool: MAX    kernel_size: 2    stride: 2  }}layer{  name: "conv5"  type: "Convolution"  bottom: "pool4"  top: "conv5"  convolution_param{    num_output: 256    kernel_size: 3    stride: 1    pad: 1  }}layer{  name: "Leaky5"  type: "Leaky"  bottom: "conv5"  top: "Leaky5"}layer{  name: "pool5"  type: "Pooling"  bottom: "Leaky5"  top: "pool5"  pooling_param {    pool: MAX    kernel_size: 2    stride: 2  }}layer{  name: "conv6"  type: "Convolution"  bottom: "pool5"  top: "conv6"  convolution_param{    num_output: 512    kernel_size: 3    stride: 1    pad: 1  }}layer{  name: "Leaky6"  type: "Leaky"  bottom: "conv6"  top: "Leaky6"}layer{  name: "pool6"  type: "Pooling"  bottom: "Leaky6"  top: "pool6"  pooling_param {    pool: MAX    kernel_size: 2    stride: 2  }}layer{  name: "conv7"  type: "Convolution"  bottom: "pool6"  top: "conv7"  convolution_param{    num_output: 1024    kernel_size: 3    stride: 1    pad: 1  }}layer{  name: "Leaky7"  type: "Leaky"  bottom: "conv7"  top: "Leaky7"}layer{  name: "conv8"  type: "Convolution"  bottom: "Leaky7"  top: "conv8"  convolution_param{    num_output: 1024    kernel_size: 3    stride: 1    pad: 1  }}layer{  name: "Leaky8"  type: "Leaky"  bottom: "conv8"  top: "Leaky8"}layer{  name: "conv9"  type: "Convolution"  bottom: "Leaky8"  top: "conv9"  convolution_param{    num_output: 1024    kernel_size: 3    stride: 1    pad: 1  }}layer{  name: "Leaky9"  type: "Leaky"  bottom: "conv9"  top: "Leaky9"}layer {  name: "connect1"  type: "InnerProduct"  bottom: "Leaky9"  top: "connect1"  inner_product_param {    num_output: 256  }}layer {  name: "connect2"  type: "InnerProduct"  bottom: "connect1"  top: "connect2"  inner_product_param {    num_output: 4096  }}layer{  name: "Leaky10"  type: "Leaky"  bottom: "connect2"  top: "Leaky10"}layer {  name: "connect3"  type: "InnerProduct"  bottom: "Leaky10"  top: "connect3"  inner_product_param {    num_output: 1470  }}

三、添加层（Leaky层）

由于caffe没有自带Leaky层，因此需要添加以下层。
leaky_layer.hpp

template <typename Dtype>class LeakyLayer : public NeuronLayer<Dtype>{public:    explicit LeakyLayer(const LayerParameter& param)       : NeuronLayer<Dtype>(param) {}    virtual void LayerSetUp(const vector<Blob<Dtype>*>& bottom,          const vector<Blob<Dtype>*>& top);    virtual inline const char* type() const { return "Leaky"; }    virtual inline int ExactNumBottomBlobs() const { return 1; }    virtual inline int ExactNumTopBlobs() 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){};//暂时还没实现      virtual void Backward_gpu(const vector<Blob<Dtype>*>& top,          const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom){};//暂时还没实现};

leaky_layer.cpp

//add by yangtemplate <typename Dtype>void LeakyLayer<Dtype>::LayerSetUp(const vector<Blob<Dtype>*>& bottom,          const vector<Blob<Dtype>*>& top) {    NeuronLayer<Dtype>::LayerSetUp(bottom, top);      CHECK_NE(top[0], bottom[0]) << this->type() << " Layer does not "        "allow in-place computation.";}template <typename Dtype>void LeakyLayer<Dtype>::Forward_cpu(        const vector<Blob<Dtype>*>& bottom, const vector<Blob<Dtype>*>& top) {      const int count = top[0]->count();      Dtype* top_data = top[0]->mutable_cpu_data();      const Dtype* bottom_data = bottom[0]->cpu_data();      for(int i = 0; i < count; ++i){          if(bottom_data[i] > 0)              top_data[i] = bottom_data[i];          else              top_data[i] = 0.1*bottom_data[i];          //top_data[i] = bottom_data[i] > 0？ bottom_data[i]: (Dtype(0.1)*bottom_data[i]);      }}//add by yang#ifdef CPU_ONLYSTUB_GPU(LeakyLayer);#endifINSTANTIATE_CLASS(LeakyLayer);REGISTER_LAYER_CLASS(Leaky);

leaky_layer.cu

//add by yangtemplate <typename Dtype>__global__ void LeakyForward(const int n, const Dtype* in, Dtype* out){    CUDA_KERNEL_LOOP(index, n){        out[index] = in[index] > 0 ? in[index] : in[index]*0.1;    }}template <typename Dtype>void LeakyLayer<Dtype>::Forward_gpu(    const vector<Blob<Dtype>*>& bottom, const vector<Blob<Dtype>*>& top) {  const int count = bottom[0]->count();  const Dtype* bottom_data = bottom[0]->gpu_data();  Dtype* top_data = top[0]->mutable_gpu_data();  LeakyForward<Dtype><<<CAFFE_GET_BLOCKS(count), CAFFE_CUDA_NUM_THREADS>>>(          count, bottom_data, top_data);  CUDA_POST_KERNEL_CHECK;}INSTANTIATE_LAYER_GPU_FUNCS(LeakyLayer);

四、使用

将第六步的代码复制到caffe根目录下面的tools文件夹，make通过之后。

argv[1]网络的prototxtargv[2]权重文件(如果编译时候使用的是loadweight，argv[2]对应的是txt文件夹，如果使用的是CopyTrainedLayersFromBinaryProto，argv[2]对应的是caffemodel)argv[3]图片文件

五、结果展示

六、代码

#include <string>#include <vector>#include "boost/algorithm/string.hpp"#include "google/protobuf/text_format.h"#include <opencv2/opencv.hpp>#include "caffe/blob.hpp"#include "caffe/layer.hpp"#include "caffe/common.hpp"#include "caffe/net.hpp"#include "caffe/proto/caffe.pb.h"#include "caffe/util/db.hpp"#include "caffe/util/format.hpp"#include "caffe/util/io.hpp"#include <stdio.h>#include <malloc.h>#include <fstream>#include <boost/progress.hpp>#include <boost/math/special_functions/next.hpp>#include <boost/random.hpp>#include <limits>#include "caffe/common.hpp"#include "caffe/util/math_functions.hpp"#include "caffe/util/rng.hpp"//#include "caffe/util/math_functions.hpp"using caffe::Blob;using caffe::Caffe;using caffe::Datum;using caffe::Net;using caffe::Layer;using std::string;namespace db = caffe::db;template<typename Dtype>Dtype lap(Dtype x1_min,Dtype x1_max,Dtype x2_min,Dtype x2_max){    if(x1_min < x2_min){        if(x1_max < x2_min){            return 0;        }else{            if(x1_max > x2_min){                if(x1_max < x2_max){                    return x1_max - x2_min;                }else{                    return x2_max - x2_min;                }            }else{                return 0;            }        }    }else{        if(x1_min < x2_max){            if(x1_max < x2_max)                return x1_max-x1_min;            else                return x2_max-x1_min;        }else{            return 0;        }    }}template int lap(int x1_min, int x1_max, int x2_min, int x2_max);template float lap(float x1_min, float x1_max, float x2_min, float x2_max);void loadweights(boost::shared_ptr<Net<float> >& net,char* argv);void loaddata(boost::shared_ptr<Net<float> >& net, std::string image_path);void getbox(std::vector<float> result,float* pro_obj,int* idx_class,std::vector<std::vector<int> >& bboxs,float thresh,cv::Mat image);//int lap(int x1_min,int x1_max,int x2_min,int x2_max);int main(int argc, char** argv){  //boost::progress_timer t;  char *labelname[] = {"aeroplane", "bicycle", "bird", "boat", "bottle", "bus", "car", "cat", "chair", "cow", "diningtable", "dog", "horse", "motorbike", "person", "pottedplant", "sheep", "sofa", "train", "tvmonitor"};  Caffe::set_mode(Caffe::GPU);  boost::shared_ptr<Net<float> > net(new Net<float>(argv[1], caffe::TEST));  //loadweights(net,argv[2]);//这行代码是还没caffemodel的时候使用的。  net->CopyTrainedLayersFromBinaryProto(argv[2]);//有caffemodel的时候就可以使用这行代码，跟上面一行互补使用。  loaddata(net,std::string(argv[3]));  //std::cout<<"load weights and data 's time = "<< t.elapsed() <<std::endl;  net->Forward();  //std::cout<<"to process finish time = "<< t.elapsed() <<std::endl;  Blob<float>* output_layer = net->output_blobs()[0];  const float* begin = output_layer->cpu_data();  const float* end = begin + output_layer->channels();  std::vector<float> result(begin, end);  //接下来就是生成框。  std::vector<std::vector<int> > bboxs;  float pro_obj[49][2];  int idx_class[49];  cv::Mat image = cv::imread(argv[3]);  float overlap;  float overlap_thresh = 0.4;  //cv::resize(image,image,cv::Size(448,448));  getbox(result, &pro_obj[0][0],idx_class,bboxs,0.2,image);  std::vector<bool> mark(bboxs.size(),true);  for(int i = 0; i < bboxs.size(); ++i){      for(int j = i+1; j < bboxs.size(); ++j){          int overlap_x = lap(bboxs[i][0],bboxs[i][2],bboxs[j][0],bboxs[j][2]);          int overlap_y = lap(bboxs[i][1],bboxs[i][3],bboxs[j][1],bboxs[j][3]);          overlap = (overlap_x*overlap_y)*1.0/((bboxs[i][0]-bboxs[i][2])*(bboxs[i][1]-bboxs[i][3])+(bboxs[j][0]-bboxs[j][2])*(bboxs[j][1]-bboxs[j][3])-(overlap_x*overlap_y));          if(overlap > overlap_thresh){              if(bboxs[i][4] > bboxs[j][4]){                  mark[j] = false;              }else{                  mark[i] = false;              }          }      }  }  for(int i = 0; i < bboxs.size();++i){      if(mark[i]){          cv::Point point1(bboxs[i][1],bboxs[i][2]);          cv::Point point2(bboxs[i][3],bboxs[i][4]);          cv::rectangle(image, cv::Rect(point1,point2),cv::Scalar(0,bboxs[i][0]/20.0*225,255),bboxs[i][5]/8);          char ch[100];          sprintf(ch,"%s %.2f",labelname[bboxs[i][0]-1], bboxs[i][5]*1.0/100);          std::string temp(ch);          cv::putText(image,temp,point1,CV_FONT_HERSHEY_COMPLEX,0.4,cv::Scalar(255,255,255));      }  }  //输出结果，画框。。  cv::imshow("yolo",image);  std::string save_name(argv[3]);  save_name = save_name.substr(0, save_name.find('.'));  save_name = save_name + "_yolo_detect.jpg";  std::cout << save_name << std::endl;  cv::imwrite(save_name, image);  cv::waitKey(0);  //下面这三行注释的代码的作用，当你只有txt存储的weight的时候，需要将weight转化为caffemodel的时候就使用这三行代码。  //caffe::NetParameter net_param; // net->ToProto(&net_param,false);  //WriteProtoToBinaryFile(net_param, "/home/yang/yolo2caffe/yolo.caffemodel");  return 1;}void loadweights(boost::shared_ptr<Net<float> >& net,char* argv){  char txt_name[200];  strcat(txt_name,argv);  char path[200];  const std::vector<boost::shared_ptr<Layer<float> > > layers = net->layers();  int convolution_n = 0;  int connect_n = 0;  FILE* fp;  char* name = (char*)malloc(sizeof(char)*100);  boost::shared_ptr<Layer<float> > layer;  std::vector<boost::shared_ptr<Blob<float> > > blobs;  for(int i = 0; i < layers.size(); ++i){    layer = layers[i];    blobs = layer->blobs();    if(layer->type() == std::string("Convolution")){        ++convolution_n;        std::cout << "convolution" << convolution_n <<std::endl;        sprintf(path,"%s/convolution%d.txt",argv,convolution_n);        //std::cout << path << std::endl;        //sprintf(name,"/home/yang/yolo2caffe/yolo/yolo_convolution%d.txt",convolution_n);        fp = fopen(path,"r");        fread(blobs[1]->mutable_cpu_data(), sizeof(float), blobs[1]->count(), fp);        fread(blobs[0]->mutable_cpu_data(), sizeof(float), blobs[0]->count(), fp);    }    else{      if(layer->type() == std::string("InnerProduct")){        ++connect_n;        std::cout << "Connect" << connect_n <<std::endl;        sprintf(path,"%s/connect%d.txt",argv,connect_n);        //std::cout << path << std::endl;        fp = fopen(path,"r");        fread(blobs[1]->mutable_cpu_data(), sizeof(float), blobs[1]->count(), fp);        fread(blobs[0]->mutable_cpu_data(), sizeof(float), blobs[0]->count(), fp);      }    }  }  if(fp != NULL)    fclose(fp);  delete []name;}void loaddata(boost::shared_ptr<Net<float> >& net, std::string image_path){  Blob<float>* input_layer = net->input_blobs()[0];  int width, height;  width = input_layer->width();  height = input_layer->height();  int size = width*height;  cv::Mat image = cv::imread(image_path,-1);  cv::Mat image_resized;  cv::resize(image, image_resized, cv::Size(height, width));  float* input_data = input_layer->mutable_cpu_data();  int temp,idx;  for(int i = 0; i < height; ++i){    uchar* pdata = image_resized.ptr<uchar>(i);    for(int j = 0; j < width; ++j){      temp = 3*j;      idx = i*width+j;      input_data[idx] = (pdata[temp+2]/127.5)-1;      input_data[idx+size] = (pdata[temp+1]/127.5)-1;      input_data[idx+2*size] = (pdata[temp+0]/127.5)-1;    }  }  //cv::imshow("image",image_resized);}void getbox(std::vector<float> result,float* pro_obj,int* idx_class,std::vector<std::vector<int> >& bboxs,float thresh,cv::Mat image){  float pro_class[49];  int idx;  float max_idx;  float max;  for(int i = 0; i < 7; ++i){    for(int j = 0; j < 7;++j){      max = 0;      max_idx = 0;      idx = 20*(i*7+j);      for(int k = 0; k < 20; ++k){        if (result[idx+k] > max){          max = result[idx+k];          max_idx = k+1;        }      }      idx_class[i*7+j] = max_idx;      pro_class[i*7+j] = max;      pro_obj[(i*7+j)*2] = max*result[7*7*20+(i*7+j)*2];      pro_obj[(i*7+j)*2+1] = max*result[7*7*20+(i*7+j)*2+1];    }  }  std::vector<int> bbox;  int x_min,x_max,y_min,y_max;  float x,y,w,h;  for(int i = 0; i < 7;++i){    for(int j = 0; j < 7;++j){      for(int k = 0; k < 2; ++k){          if(pro_obj[(i*7+j)*2 + k] > thresh){              //std::cout << "(" << i << "," << j << "," << k << ")" << " prob="<<pro_obj[(i*7+j)*2 + k] << " class="<<idx_class[i*7+j]<<std::endl;              idx = 49*20 + 49*2 + ((i*7+j)*2+k)*4;              x = image.cols*(result[idx++]+j)/7;              y = image.rows*(result[idx++]+i)/7;              w = image.cols*result[idx]*result[idx++];              h = image.rows*result[idx]*result[idx];              //std::cout << x <<" "<< y << " " << w <<" "<< h <<std::endl;              x_min = x - w/2;              y_min = y - h/2;              x_max = x + w/2;              y_max = y + h/2;              bbox.clear();              bbox.push_back(idx_class[i*7+j]);              bbox.push_back(x_min);              bbox.push_back(y_min);              bbox.push_back(x_max);              bbox.push_back(y_max);              bbox.push_back(int(pro_obj[(i*7+j)*2 + k]*100));              bboxs.push_back(bbox);          }      }    }  }}