【caffe配置】在vs2013用已经训练好的Alexnet网络提取fc6特征

首先请参照我的上一篇博文，在windows下配置好caffe-windows这个过程不算太复杂。把ubuntu下的caffe环境也配置好，这样我们就可以在linux环境下训练网络，并且把训练好的网络放在windows下的vs工程中进行一次前馈来提取特征了。

在ubuntu的caffe根目录下把下面4个文件拷出来：
a. 找到caffe-master/models/bvlc_alexnet/readme.md，这里有alexnet的caffemodel下载地址可以直接下载：http://dl.caffe.berkeleyvision.org/bvlc_alexnet.caffemodel。这个文件是一个已经训练好的网络参数
b. caffe-master/models/下的deploy.prototxt。这个文件说明了网络的结构
c. caffe-master/examples/imagenet下的imagenet_mean.binaryproto。这个文件记录了输入数据的均值
c. caffe-master/examples/imagenet下的synset_words.txt。这个文件记录了网络分类的标签含义
将这几个文件拷出来之后就可以回到windows了，新建一个vs2013工程，将刚才拷出来的4个文件，以及任意一张你要分类的图片放在工程的一个子目录下。如下图所示：

之后将工程调整为release x64，我们在编译caffe-windows时只编译了release x64的版本，所以debug下，或者是win32下都是无法跑这个demo的。

新建头文件head.h

//head.h
#include "caffe/common.hpp"  
#include "caffe/layers/input_layer.hpp"  
#include "caffe/layers/inner_product_layer.hpp"  
#include "caffe/layers/conv_layer.hpp"  
#include "caffe/layers/relu_layer.hpp"  
#include "caffe/layers/pooling_layer.hpp"  
#include "caffe/layers/softmax_layer.hpp"  
namespace caffe
 extern INSTANTIATE_CLASS(DataLayer);
 extern INSTANTIATE_CLASS(InputLayer);
 extern INSTANTIATE_CLASS(InnerProductLayer);
 //REGISTER_LAYER_CLASS(Dropout);
 extern INSTANTIATE_CLASS(ConvolutionLayer);
 extern INSTANTIATE_CLASS(ReLULayer);
 extern INSTANTIATE_CLASS(PoolingLayer);
 extern INSTANTIATE_CLASS(LRNLayer);
 extern INSTANTIATE_CLASS(SoftmaxLayer);
}

新建头文件Classifier.h

//Classifier.h
#define USE_OPENCV
//#define CPU_ONLY
#include "caffe/common.hpp"
#include <opencv2/core/core.hpp>
#include <opencv2/highgui/highgui.hpp>
#include <opencv2/imgproc/imgproc.hpp>
#include <iosfwd>
#include <string>
#include <vector>
#include <fstream>
//using namespace std;
using namespace caffe; // NOLINT(build/namespaces)
/* Pair (label, confidence) representing a prediction. */
class Classifier {
 Classifier(const string& model_file,
  const string& mean_file,
  const string& label_file);
 std::ofstream ofile;
 std::vector<Prediction> Classify(const cv::Mat& img, int N = 5);
 void Extract_Feature(const cv::Mat& img, vector<float>& feature);
 std::vector<float> Predict(const cv::Mat& img);
 void WrapInputLayer(std::vector<cv::Mat>* input_channels);
  std::vector<cv::Mat>* input_channels);
 shared_ptr<Net<float> > net_;
 cv::Mat mean_;
 std::vector<string> labels_;
};

Classifier.cpp

#include "Classifier.h"
Classifier::Classifier(const string& model_file,
 const string& mean_file,
 const string& label_file) {
 Caffe::set_mode(Caffe::CPU);
 Caffe::set_mode(Caffe::GPU);
 /* Load the network. */
 ofile.open("feature.txt");
 net_.reset(new Net<float>(model_file, TEST));
 CHECK_EQ(net_->num_inputs(), 1) << "Network should have exactly one input.";
 CHECK_EQ(net_->num_outputs(), 1) << "Network should have exactly one output.";
 num_channels_ = input_layer->channels();
 CHECK(num_channels_ == 3 || num_channels_ == 1)
  << "Input layer should have 1 or 3 channels.";
 SetMean(mean_file);
 /* Load labels. */
 CHECK(labels) << "Unable to open labels file " << label_file;
  labels_.push_back(string(line));
 Blob<float>* output_layer = net_->output_blobs()[0];
  << "Number of labels is different from the output layer dimension.";
static bool PairCompare(const std::pair<float, int>& lhs,
 return lhs.first > rhs.first;
/* Return the indices of the top N values of vector v. */
static std::vector<int> Argmax(const std::vector<float>& v, int N) {
 for (size_t i = 0; i < v.size(); ++i)
  pairs.push_back(std::make_pair(v[i], i));
 std::partial_sort(pairs.begin(), pairs.begin() + N, pairs.end(), PairCompare);
  result.push_back(pairs[i].second);
/* Return the top N predictions. */
std::vector<Prediction> Classifier::Classify(const cv::Mat& img, int N) {
 N = std::min<int>(labels_.size(), N);
 std::vector<int> maxN = Argmax(output, N);
 for (int i = 0; i < N; ++i) {
  predictions.push_back(std::make_pair(labels_[idx], output[idx]));
/* Load the mean file in binaryproto format. */
void Classifier::SetMean(const string& mean_file) {
 ReadProtoFromBinaryFileOrDie(mean_file.c_str(), &blob_proto);
 Blob<float> mean_blob;
 mean_blob.FromProto(blob_proto);
 CHECK_EQ(mean_blob.channels(), num_channels_)
  << "Number of channels of mean file doesn't match input layer.";
 /* The format of the mean file is planar 32-bit float BGR or grayscale. */
 float* data = mean_blob.mutable_cpu_data();
  /* Extract an individual channel. */
  cv::Mat channel(mean_blob.height(), mean_blob.width(), CV_32FC1, data);
  data += mean_blob.height() * mean_blob.width();
 /* Merge the separate channels into a single image. */
 /* Compute the global mean pixel value and create a mean image
 cv::Scalar channel_mean = cv::mean(mean);
 mean_ = cv::Mat(input_geometry_, mean.type(), channel_mean);
unsigned int get_blob_index(boost::shared_ptr< Net<float> > & net, char *query_blob_name)
 vector< string > const & blob_names = net->blob_names();
 {
  {
   return i;
 LOG(FATAL) << "Unknown blob name: " << str_query;
std::vector<float> Classifier::Predict(const cv::Mat& img) {
 input_layer->Reshape(1, num_channels_,
  input_geometry_.height, input_geometry_.width);
 net_->Reshape();
 std::vector<cv::Mat> input_channels;
 Preprocess(img, &input_channels);
 //char *query_blob_name = "fc6"; /* data, conv1, pool1, norm1, fc6, prob, etc */
 //boost::shared_ptr<Blob<float> > blob = net_->blobs()[blob_id];
 //std::ofstream ofile("feature.txt");
 //const float *blob_ptr = (const float *)blob->cpu_data();
 // ofile << *(blob_ptr + i)<<" ";
 //ofile.close();
 /* Copy the output layer to a std::vector */
 Blob<float>* output_layer = net_->output_blobs()[0];
 const float* end = begin + output_layer->channels();
/* Wrap the input layer of the network in separate cv::Mat objects
* don't need to rely on cudaMemcpy2D. The last preprocessing
* operation will write the separate channels directly to the input
void Classifier::WrapInputLayer(std::vector<cv::Mat>* input_channels) {
 int width = input_layer->width();
 int height = input_layer->height();
 float* input_data = input_layer->mutable_cpu_data();
  cv::Mat channel(height, width, CV_32FC1, input_data);
  input_data += width * height;
void Classifier::Preprocess(const cv::Mat& img,
 /* Convert the input image to the input image format of the network. */
  cv::cvtColor(img, sample, cv::COLOR_BGR2GRAY);
 else if (img.channels() == 4 && num_channels_ == 1)
 else if (img.channels() == 4 && num_channels_ == 3)
 else if (img.channels() == 1 && num_channels_ == 3)
 else
  sample = img;
 cv::Mat sample_resized;
 if (sample.size() != input_geometry_)
  cv::resize(sample, sample_resized, input_geometry_);
 cv::Mat sample_float;
 if (num_channels_ == 3)
  sample_resized.convertTo(sample_float, CV_32FC3);
  sample_resized.convertTo(sample_float, CV_32FC1);
 cv::subtract(sample_float, mean_, sample_normalized);
 /* This operation will write the separate BGR planes directly to the
 * objects in input_channels. */
 cv::split(sample_normalized, *input_channels);
 CHECK(reinterpret_cast<float*>(input_channels->at(0).data)
  << "Input channels are not wrapping the input layer of the network.";
void Classifier::Extract_Feature(const cv::Mat& img, vector<float>& feature)
 input_layer->Reshape(1, num_channels_,
  input_geometry_.height, input_geometry_.width);
 net_->Reshape();
 std::vector<cv::Mat> input_channels;
 Preprocess(img, &input_channels);
 char *query_blob_name = "fc6"; /* data, conv1, pool1, norm1, fc6, prob, etc */
 boost::shared_ptr<Blob<float> > blob = net_->blobs()[blob_id];
 const float *blob_ptr = (const float *)blob->cpu_data();
  //ofile << *(blob_ptr + i) << " ";
 }
 //ofile << std::endl;
}

test.cpp

#include "Classifier.h"
using namespace cv;
int main(int argc, char** argv) {
  std::cerr << "Usage: " << argv[0]
   << " deploy.prototxt network.caffemodel"
   << " mean.binaryproto labels.txt img.jpg" << std::endl;
 string model_file = argv[1];
 string trained_file = argv[2];
 string label_file = argv[4];
 Classifier classifier(model_file, trained_file, mean_file, label_file);
  << file << " ----------" << std::endl;
 CHECK(!img.empty()) << "Unable to decode image " << file;
 std::vector<Prediction> predictions = classifier.Classify(img);
 for (size_t i = 0; i < predictions.size(); ++i) {
  std::cout << std::fixed << std::setprecision(4) << p.second << " - \""
}

接下来再配置一下工程的一些路径：

1. 配置属性->调试->命令参数 添加./imagenet/deploy.prototxt ./imagenet/bvlc_alexnet.caffemodel ./imagenet/imagenet_mean.binaryproto ./imagenet/synset_words.txt ./imagenet/car.jpg
2. 配置属性->VC++目录->包含目录 添加自己的boost、caffe、cuda目录


3. 配置属性->VC++目录->库目录 添加相应的lib文件所在目录

4. 链接器->附加库目录

5. 链接器->输入->附加依赖项 添加：
caffelib.lib
ntdll.lib
kernel32.lib
user32.lib
gdi32.lib
winspool.lib
shell32.lib
ole32.lib
oleaut32.lib
uuid.lib
comdlg32.lib
advapi32.lib
cudart.lib
cublas.lib
curand.lib
libprotobuf.lib
hdf5_tools.lib
hdf5_hl_fortran.lib
hdf5_fortran.lib
hdf5_hl_f90cstub.lib
hdf5_f90cstub.lib
hdf5_cpp.lib
hdf5_hl_cpp.lib
hdf5_hl.lib
hdf5.lib
zlib.lib
szip.lib
opencv_world300.lib
shlwapi.lib
leveldb.lib
cublas_device.lib
cuda.lib
libglog.lib
lmdb.lib
cudnn.lib
libopenblas.dll.a
libgflags.lib

这样应该就可以跑这个工程了。

程序的结果：

可以看到分类的前5名，还是非常准确的。

期间遇到的问题：
1. Check failed: registry.count(t ype) == 1 (0 vs. 1) Unknown layer type
这个问题纠结了好久。。。参考这个解决的：http://blog.youkuaiyun.com/fangjin_kl/article/details/50936952，就添加了head.h中的那几句话就行了
2. 如何提取fc6层的特征：
我在Classifier.cpp的Predict函数中注释了几行，那个是我提取fc6层特征并且将特征保存在文件里的代码，给大家作参考。
3. 一开始我以为不用自己再装一个boost，因为caffe-windows的第三方库里已经自带boost了，后来程序老提示我找不到boost的一个1.58版本的lib，我就自己配置了一下boost1.59然后没有那个错误了。