Caffe使用教程

Caffe官网：http://caffe.berkeleyvision.org/

 

初始化网络

 1 #include "caffe/caffe.hpp"
 2 #include <string>
 3 #include <vector>
 4 using namespace caffe;
 5  
 6 char *proto = "H:\\Models\\Caffe\\deploy.prototxt"; /* 加载CaffeNet的配置 */
 7 Phase phase = TEST; /* or TRAIN */
 8 Caffe::set_mode(Caffe::CPU);
 9 // Caffe::set_mode(Caffe::GPU);
10 // Caffe::SetDevice(0);
11  
12 //! Note: 后文所有提到的net，都是这个net
13 boost::shared_ptr< Net<float> > net(new caffe::Net<float>(proto, phase));

 

加载已训练好的模型

1 char *model = "H:\\Models\\Caffe\\bvlc_reference_caffenet.caffemodel";   
2 net->CopyTrainedLayersFrom(model);

 

读取图像均值

 1 char *mean_file = "H:\\Models\\Caffe\\imagenet_mean.binaryproto";
 2 Blob<float> image_mean;
 3 BlobProto blob_proto;
 4 const float *mean_ptr;
 5 unsigned int num_pixel;
 6  
 7 bool succeed = ReadProtoFromBinaryFile(mean_file, &blob_proto);
 8 if (succeed)
 9 {
10     image_mean.FromProto(blob_proto);
11     num_pixel = image_mean.count(); /* NCHW=1x3x256x256=196608 */
12     mean_ptr = (const float *) image_mean.cpu_data();
13 }

 

根据指定数据，前向传播网络

 1 //! Note: data_ptr指向已经处理好（去均值的，符合网络输入图像的长宽和Batch Size）的数据
 2 void caffe_forward(boost::shared_ptr< Net<float> > & net, float *data_ptr)
 3 {
 4     Blob<float>* input_blobs = net->input_blobs()[0];
 5     switch (Caffe::mode())
 6     {
 7     case Caffe::CPU:
 8         memcpy(input_blobs->mutable_cpu_data(), data_ptr,
 9             sizeof(float) * input_blobs->count());
10         break;
11     case Caffe::GPU:
12         cudaMemcpy(input_blobs->mutable_gpu_data(), data_ptr,
13             sizeof(float) * input_blobs->count(), cudaMemcpyHostToDevice);
14         break;
15     default:
16         LOG(FATAL) << "Unknown Caffe mode.";
17     }
18     net->ForwardPrefilled();
19 }

 

根据Feature层的名字获取其在网络中的Index

 1 //! Note: Net的Blob是指，每个层的输出数据，即Feature Maps
 2 // char *query_blob_name = "conv1";
 3 unsigned int get_blob_index(boost::shared_ptr< Net<float> > & net, char *query_blob_name)
 4 {
 5     std::string str_query(query_blob_name);   
 6     vector< string > const & blob_names = net->blob_names();
 7     for( unsigned int i = 0; i != blob_names.size(); ++i )
 8     {
 9         if( str_query == blob_names[i] )
10         {
11             return i;
12         }
13     }
14     LOG(FATAL) << "Unknown blob name: " << str_query;
15 }

 

读取网络指定Feature层数据

1 //! Note: 根据CaffeNet的deploy.prototxt文件，该Net共有15个Blob，从data一直到prob   
2 char *query_blob_name = "conv1"; /* data, conv1, pool1, norm1, fc6, prob, etc */
3 unsigned int blob_id = get_blob_index(net, query_blob_name);
4  
5 boost::shared_ptr<Blob<float> > blob = net->blobs()[blob_id];
6 unsigned int num_data = blob->count(); /* NCHW=10x96x55x55 */
7 const float *blob_ptr = (const float *) blob->cpu_data();

 

根据Layer的名字获取其在网络中的Index

 1 //! Note: Layer包括神经网络所有层，比如，CaffeNet共有23层
 2 // char *query_layer_name = "conv1";
 3 unsigned int get_layer_index(boost::shared_ptr< Net<float> > & net, char *query_layer_name)
 4 {
 5     std::string str_query(query_layer_name);   
 6     vector< string > const & layer_names = net->layer_names();
 7     for( unsigned int i = 0; i != layer_names.size(); ++i )
 8     {
 9         if( str_query == layer_names[i] )
10         {
11             return i;
12         }
13     }
14     LOG(FATAL) << "Unknown layer name: " << str_query;
15 }

 

读取指定Layer的权重数据

 1 //! Note: 不同于Net的Blob是Feature Maps，Layer的Blob是指Conv和FC等层的Weight和Bias
 2 char *query_layer_name = "conv1";
 3 const float *weight_ptr, *bias_ptr;
 4 unsigned int layer_id = get_layer_index(net, query_layer_name);
 5 boost::shared_ptr<Layer<float> > layer = net->layers()[layer_id];
 6 std::vector<boost::shared_ptr<Blob<float>  >> blobs = layer->blobs();
 7 if (blobs.size() > 0)
 8 {
 9     weight_ptr = (const float *) blobs[0]->cpu_data();
10     bias_ptr = (const float *) blobs[1]->cpu_data();
11 }
12  
13 //! Note: 训练模式下，读取指定Layer的梯度数据，与此相似，唯一的区别是将cpu_data改为cpu_diff

 

修改某层的Weight数据

 1 const float* data_ptr;          /* 指向待写入数据的指针， 源数据指针*/
 2 float* weight_ptr = NULL;       /* 指向网络中某层权重的指针，目标数据指针*/
 3 unsigned int data_size;         /* 待写入的数据量 */
 4 char *layer_name = "conv1";     /* 需要修改的Layer名字 */
 5  
 6 unsigned int layer_id = get_layer_index(net, query_layer_name);   
 7 boost::shared_ptr<Blob<float> > blob = net->layers()[layer_id]->blobs()[0];
 8  
 9 CHECK(data_size == blob->count());
10 switch (Caffe::mode())
11 {
12 case Caffe::CPU:
13     weight_ptr = blob->mutable_cpu_data();
14     break;
15 case Caffe::GPU:
16     weight_ptr = blob->mutable_gpu_data();
17     break;
18 default:
19     LOG(FATAL) << "Unknown Caffe mode";
20 }
21 caffe_copy(blob->count(), data_ptr, weight_ptr);
22  
23 //! Note: 训练模式下，手动修改指定Layer的梯度数据，与此相似
24 // mutable_cpu_data改为mutable_cpu_diff，mutable_gpu_data改为mutable_gpu_diff

 

保存新的模型

1 char* weights_file = "bvlc_reference_caffenet_new.caffemodel";
2 NetParameter net_param;
3 net->ToProto(&net_param, false);
4 WriteProtoToBinaryFile(net_param, weights_file);