如何在程序中调用CAFFE做图像分类

【原文：http://www.cnblogs.com/waring/p/4471183.html】

Caffe是目前深度学习比较优秀好用的一个开源库，采样c++和CUDA实现，具有速度快，模型定义方便等优点。学习了几天过后，发现也有一个不方便的地方，就是在我的程序中调用Caffe做图像分类没有直接的接口。Caffe的数据层可以从数据库（支持leveldb、lmdb、hdf5）、图片、和内存中读入。我们要在程序中使用，当然得从内存中读入。参见http://caffe.berkeleyvision.org/tutorial/layers.html#data-layers和MemoryDataLayer源码，我们首先在模型定义文件中定义数据层：

layers {
name: "mydata"
type: MEMORY_DATA
top: "data"
top: "label"
transform_param {
scale: 0.00390625
}
memory_data_param {
batch_size: 10
channels: 1
height: 24
width: 24
}
}

这里必须设置memory_data_param中的四个参数，对应这些参数可以参见源码中caffe.proto文件。现在，我们可以设计一个Classifier类来封装一下：

#ifndef CAFFE_CLASSIFIER_H
#define CAFFE_CLASSIFIER_H

#include <string>
#include <vector>
#include "caffe/net.hpp"
#include "caffe/data_layers.hpp"
#include <opencv2/core.hpp>
using cv::Mat;

namespace caffe {

template <typename Dtype>
class Classifier {
public:
explicit Classifier(const string& param_file, const string& weights_file);
Dtype test(vector<Mat> &images, vector<int> &labels, int iter_num);
virtual ~Classifier() {}
inline shared_ptr<Net<Dtype> > net() { return net_; }
void predict(vector<Mat> &images, vector<int> *labels);
void predict(vector<Dtype> &data, vector<int> *labels, int num);
void extract_feature(vector<Mat> &images, vector<vector<Dtype>> *out);

protected:
shared_ptr<Net<Dtype> > net_;
MemoryDataLayer<Dtype> *m_layer_;
int batch_size_;
int channels_;
int height_;
int width_;

DISABLE_COPY_AND_ASSIGN(Classifier);
};
}//namespace
#endif //CAFFE_CLASSIFIER_H

构造函数中我们通过模型定义文件（.prototxt）和训练好的模型（.caffemodel）文件构造一个Net对象，并用m_layer_指向Net中的memory data层，以便待会调用MemoryDataLayer中AddMatVector和Reset函数加入数据。

#include <cstdio>

#include <algorithm>
#include <string>
#include <vector>

#include "caffe/net.hpp"
#include "caffe/proto/caffe.pb.h"
#include "caffe/util/io.hpp"
#include "caffe/util/math_functions.hpp"
#include "caffe/util/upgrade_proto.hpp"
#include "caffe_classifier.h"

namespace caffe {

template <typename Dtype>
Classifier<Dtype>::Classifier(const string& param_file, const string& weights_file) : net_()
{
net_.reset(new Net<Dtype>(param_file, TEST));
net_->CopyTrainedLayersFrom(weights_file);
//m_layer_ = (MemoryDataLayer<Dtype>*)net_->layer_by_name("mnist").get();
m_layer_ = (MemoryDataLayer<Dtype>*)net_->layers()[0].get();
batch_size_ = m_layer_->batch_size();
channels_ = m_layer_->channels();
height_ = m_layer_->height();
width_ = m_layer_->width();
}

template <typename Dtype>
Dtype Classifier<Dtype>::test(vector<Mat> &images, vector<int> &labels, int iter_num)
{
m_layer_->AddMatVector(images, labels);
//
int iterations = iter_num;
vector<Blob<Dtype>* > bottom_vec;

vector<int> test_score_output_id;
vector<Dtype> test_score;
Dtype loss = 0;
for (int i = 0; i < iterations; ++i) {
Dtype iter_loss;
const vector<Blob<Dtype>*>& result =
net_->Forward(bottom_vec, &iter_loss);
loss += iter_loss;
int idx = 0;
for (int j = 0; j < result.size(); ++j) {
const Dtype* result_vec = result[j]->cpu_data();
for (int k = 0; k < result[j]->count(); ++k, ++idx) {
const Dtype score = result_vec[k];
if (i == 0) {
test_score.push_back(score);
test_score_output_id.push_back(j);
} else {
test_score[idx] += score;
}
const std::string& output_name = net_->blob_names()[
net_->output_blob_indices()[j]];
LOG(INFO) << "Batch " << i << ", " << output_name << " = " << score;
}
}
}
loss /= iterations;
LOG(INFO) << "Loss: " << loss;
return loss;
}

template <typename Dtype>
void Classifier<Dtype>::predict(vector<Mat> &images, vector<int> *labels)
{
int original_length = images.size();
if(original_length == 0)
return;
int valid_length = original_length / batch_size_ * batch_size_;
if(original_length != valid_length)
{
valid_length += batch_size_;
for(int i = original_length; i < valid_length; i++)
{
images.push_back(images[0].clone());
}
}
vector<int> valid_labels, predicted_labels;
valid_labels.resize(valid_length, 0);
m_layer_->AddMatVector(images, valid_labels);
vector<Blob<Dtype>* > bottom_vec;
for(int i = 0; i < valid_length / batch_size_; i++)
{
const vector<Blob<Dtype>*>& result = net_->Forward(bottom_vec);
const Dtype * result_vec = result[1]->cpu_data();
for(int j = 0; j < result[1]->count(); j++)
{
predicted_labels.push_back(result_vec[j]);
}
}
if(original_length != valid_length)
{
images.erase(images.begin()+original_length, images.end());
}
labels->resize(original_length, 0);
std::copy(predicted_labels.begin(), predicted_labels.begin() + original_length, labels->begin());
}

template <typename Dtype>
void Classifier<Dtype>::predict(vector<Dtype> &data, vector<int> *labels, int num)
{
int size = channels_*height_*width_;
CHECK_EQ(data.size(), num*size);
int original_length = num;
if(original_length == 0)
return;
int valid_length = original_length / batch_size_ * batch_size_;
if(original_length != valid_length)
{
valid_length += batch_size_;
for(int i = original_length; i < valid_length; i++)
{
for(int j = 0; j < size; j++)
data.push_back(0);
}
}
vector<int> predicted_labels;
Dtype * label_ = new Dtype[valid_length];
memset(label_, 0, valid_length);
m_layer_->Reset(data.data(), label_, valid_length);
vector<Blob<Dtype>* > bottom_vec;
for(int i = 0; i < valid_length / batch_size_; i++)
{
const vector<Blob<Dtype>*>& result = net_->Forward(bottom_vec);
const Dtype * result_vec = result[1]->cpu_data();
for(int j = 0; j < result[1]->count(); j++)
{
predicted_labels.push_back(result_vec[j]);
}
}
if(original_length != valid_length)
{
data.erase(data.begin()+original_length*size, data.end());
}
delete [] label_;
labels->resize(original_length, 0);
std::copy(predicted_labels.begin(), predicted_labels.begin() + original_length, labels->begin());
}
template <typename Dtype>
void Classifier<Dtype>::extract_feature(vector<Mat> &images, vector<vector<Dtype>> *out)
{
int original_length = images.size();
if(original_length == 0)
return;
int valid_length = original_length / batch_size_ * batch_size_;
if(original_length != valid_length)
{
valid_length += batch_size_;
for(int i = original_length; i < valid_length; i++)
{
images.push_back(images[0].clone());
}
}
vector<int> valid_labels;
valid_labels.resize(valid_length, 0);
m_layer_->AddMatVector(images, valid_labels);
vector<Blob<Dtype>* > bottom_vec;
out->clear();
for(int i = 0; i < valid_length / batch_size_; i++)
{
const vector<Blob<Dtype>*>& result = net_->Forward(bottom_vec);
const Dtype * result_vec = result[0]->cpu_data();
const int dim = result[0]->count(1);
for(int j = 0; j < result[0]->num(); j++)
{
const Dtype * ptr = result_vec + j * dim;
vector<Dtype> one_;
for(int k = 0; k < dim; ++k)
one_.push_back(ptr[k]);
out->push_back(one_);
}
}
if(original_length != valid_length)
{
images.erase(images.begin()+original_length, images.end());
out->erase(out->begin()+original_length, out->end());
}
}
INSTANTIATE_CLASS(Classifier);
}  // namespace caffe

由于加入的数据个数必须是batch_size的整数倍，所以我们在加入数据时采用填充的方式。

CHECK_EQ(num % batch_size_, 0) <<
"The added data must be a multiple of the batch size.";  //AddMatVector

在模型文件的最后，我们把训练时的loss层改为argmax层：

layers {
name: "predicted"
type: ARGMAX
bottom: "prob"
top: "predicted"
}

作者：waring  出处：http://www.cnblogs.com/waring 
欢迎转载或分享，但请务必声明文章出处。