如何使用Caffe
Caffe教程（http://robots.princeton.edu/courses/COS598/2015sp/slides/Caffe/caffe_tutorial.pdf）
預(yù)備知識
Google Protocol Buffer
https://developers.google.com/protocol-buffers/docs/cpptutorial
Caffe數(shù)據(jù)的讀取庐冯、運算孽亲、存儲都是采用Google Protocol Buffer來進(jìn)行的。PB是一種輕便肄扎、高效的結(jié)構(gòu)化數(shù)據(jù)存儲格式墨林，可以用于結(jié)構(gòu)化數(shù)據(jù)串行化，很適合做數(shù)據(jù)存儲或 RPC 數(shù)據(jù)交換格式犯祠。它可用于通訊協(xié)議旭等、數(shù)據(jù)存儲等領(lǐng)域的語言無關(guān)、平臺無關(guān)衡载、可擴(kuò)展的序列化結(jié)構(gòu)數(shù)據(jù)格式搔耕。是一種效率和兼容性都很優(yōu)秀的二進(jìn)制數(shù)據(jù)傳輸格式，目前提供了 C++痰娱、Java弃榨、Python 三種語言的 API。Caffe采用的是C++和Python的API梨睁。
轉(zhuǎn)載自https://github.com/shicai/Caffe_Manual/blob/master/ReadMe.md
初始化網(wǎng)絡(luò)
#include "caffe/caffe.hpp"
#include <string>
#include <vector>
using namespace caffe;
char *proto = "H:\\Models\\Caffe\\deploy.prototxt"; /* 加載CaffeNet的配置 */
Phase phase = TEST; /* or TRAIN */
Caffe::set_mode(Caffe::CPU);
// Caffe::set_mode(Caffe::GPU);
// Caffe::SetDevice(0);
//! Note: 后文所有提到的net鲸睛，都是這個net
boost::shared_ptr< Net<float> > net(new caffe::Net<float>(proto, phase));
加載已訓(xùn)練好的模型
char *model = "H:\\Models\\Caffe\\bvlc_reference_caffenet.caffemodel";
net->CopyTrainedLayersFrom(model);
讀取模型中的每層的結(jié)構(gòu)配置參數(shù)
char *model = "H:\\Models\\Caffe\\bvlc_reference_caffenet.caffemodel";
NetParameter param;
ReadNetParamsFromBinaryFileOrDie(model, ¶m);
int num_layers = param.layer_size();
for (int i = 0; i < num_layers; ++i)
{
// 結(jié)構(gòu)配置參數(shù):name，type坡贺，kernel size官辈，pad箱舞，stride等
LOG(ERROR) << "Layer " << i << ":" << param.layer(i).name() << "\t" << param.layer(i).type();
if (param.layer(i).type() == "Convolution")
{
ConvolutionParameter conv_param = param.layer(i).convolution_param();
LOG(ERROR) << "\t\tkernel size: " << conv_param.kernel_size()
<< ", pad: " << conv_param.pad()
<< ", stride: " << conv_param.stride();
}
}
讀取圖像均值
char *mean_file = "H:\\Models\\Caffe\\imagenet_mean.binaryproto";
Blob<float> image_mean;
BlobProto blob_proto;
const float *mean_ptr;
unsigned int num_pixel;
bool succeed = ReadProtoFromBinaryFile(mean_file, &blob_proto);
if (succeed)
{
image_mean.FromProto(blob_proto);
num_pixel = image_mean.count(); /* NCHW=1x3x256x256=196608 */
mean_ptr = (const float *) image_mean.cpu_data();
}
根據(jù)指定數(shù)據(jù)，前向傳播網(wǎng)絡(luò)
//! Note: data_ptr指向已經(jīng)處理好（去均值的拳亿，符合網(wǎng)絡(luò)輸入圖像的長寬和Batch Size）的數(shù)據(jù)
void caffe_forward(boost::shared_ptr< Net<float> > & net, float *data_ptr)
{
Blob<float>* input_blobs = net->input_blobs()[0];
switch (Caffe::mode())
{
case Caffe::CPU:
memcpy(input_blobs->mutable_cpu_data(), data_ptr,
sizeof(float) * input_blobs->count());
break;
case Caffe::GPU:
cudaMemcpy(input_blobs->mutable_gpu_data(), data_ptr,
sizeof(float) * input_blobs->count(), cudaMemcpyHostToDevice);
break;
default:
LOG(FATAL) << "Unknown Caffe mode.";
}
net->ForwardPrefilled();
}
根據(jù)Feature層的名字獲取其在網(wǎng)絡(luò)中的Index
//! Note: Net的Blob是指晴股，每個層的輸出數(shù)據(jù)，即Feature Maps
// char *query_blob_name = "conv1";
unsigned int get_blob_index(boost::shared_ptr< Net<float> > & net, char *query_blob_name)
{
std::string str_query(query_blob_name);
vector< string > const & blob_names = net->blob_names();
for( unsigned int i = 0; i != blob_names.size(); ++i )
{
if( str_query == blob_names[i] )
{
return i;
}
}
LOG(FATAL) << "Unknown blob name: " << str_query;
}
讀取網(wǎng)絡(luò)指定Feature層數(shù)據(jù)
//! Note: 根據(jù)CaffeNet的deploy.prototxt文件肺魁，該Net共有15個Blob电湘，從data一直到prob
char *query_blob_name = "conv1"; /* data, conv1, pool1, norm1, fc6, prob, etc */
unsigned int blob_id = get_blob_index(net, query_blob_name);
boost::shared_ptr<Blob<float> > blob = net->blobs()[blob_id];
unsigned int num_data = blob->count(); /* NCHW=10x96x55x55 */
const float *blob_ptr = (const float *) blob->cpu_data();
根據(jù)文件列表，獲取特征鹅经，并存為二進(jìn)制文件
詳見get_features.cpp文件：
主要包括三個步驟
- 生成文件列表寂呛，格式與訓(xùn)練用的類似，每行一個圖像
包括文件全路徑瞬雹、空格昧谊、標(biāo)簽（沒有的話，可以置0） - 根據(jù)train_val或者deploy的prototxt酗捌，改寫生成feat.prototxt
主要是將輸入層改為image_data層呢诬，最后加上prob和argmax（為了輸出概率和Top1/5預(yù)測標(biāo)簽） - 根據(jù)指定參數(shù)，運行程序后會生成若干個二進(jìn)制文件胖缤，可以用MATLAB讀取數(shù)據(jù)尚镰，進(jìn)行分析
根據(jù)Layer的名字獲取其在網(wǎng)絡(luò)中的Index
//! Note: Layer包括神經(jīng)網(wǎng)絡(luò)所有層，比如哪廓，CaffeNet共有23層
// char *query_layer_name = "conv1";
unsigned int get_layer_index(boost::shared_ptr< Net<float> > & net, char *query_layer_name)
{
std::string str_query(query_layer_name);
vector< string > const & layer_names = net->layer_names();
for( unsigned int i = 0; i != layer_names.size(); ++i )
{
if( str_query == layer_names[i] )
{
return i;
}
}
LOG(FATAL) << "Unknown layer name: " << str_query;
}
讀取指定Layer的權(quán)重數(shù)據(jù)
//! Note: 不同于Net的Blob是Feature Maps狗唉，Layer的Blob是指Conv和FC等層的Weight和Bias
char *query_layer_name = "conv1";
const float *weight_ptr, *bias_ptr;
unsigned int layer_id = get_layer_index(net, query_layer_name);
boost::shared_ptr<Layer<float> > layer = net->layers()[layer_id];
std::vector<boost::shared_ptr<Blob<float> >> blobs = layer->blobs();
if (blobs.size() > 0)
{
weight_ptr = (const float *) blobs[0]->cpu_data();
bias_ptr = (const float *) blobs[1]->cpu_data();
}
//! Note: 訓(xùn)練模式下，讀取指定Layer的梯度數(shù)據(jù)涡真，與此相似分俯，唯一的區(qū)別是將cpu_data改為cpu_diff
修改某層的Weight數(shù)據(jù)
const float* data_ptr; /* 指向待寫入數(shù)據(jù)的指針， 源數(shù)據(jù)指針*/
float* weight_ptr = NULL; /* 指向網(wǎng)絡(luò)中某層權(quán)重的指針哆料，目標(biāo)數(shù)據(jù)指針*/
unsigned int data_size; /* 待寫入的數(shù)據(jù)量 */
char *layer_name = "conv1"; /* 需要修改的Layer名字 */
unsigned int layer_id = get_layer_index(net, query_layer_name);
boost::shared_ptr<Blob<float> > blob = net->layers()[layer_id]->blobs()[0];
CHECK(data_size == blob->count());
switch (Caffe::mode())
{
case Caffe::CPU:
weight_ptr = blob->mutable_cpu_data();
break;
case Caffe::GPU:
weight_ptr = blob->mutable_gpu_data();
break;
default:
LOG(FATAL) << "Unknown Caffe mode";
}
caffe_copy(blob->count(), data_ptr, weight_ptr);
//! Note: 訓(xùn)練模式下缸剪，手動修改指定Layer的梯度數(shù)據(jù)，與此相似
// mutable_cpu_data改為mutable_cpu_diff东亦，mutable_gpu_data改為mutable_gpu_diff
保存新的模型
char* weights_file = "bvlc_reference_caffenet_new.caffemodel";
NetParameter net_param;
net->ToProto(&net_param, false);
WriteProtoToBinaryFile(net_param, weights_file);
Caffe中添加新的層
https://github.com/BVLC/caffe/wiki/Development
這里寫圖片描述
用預(yù)訓(xùn)練網(wǎng)絡(luò)參數(shù)初始化
caffe的參數(shù)初始化是根據(jù)名字從caffemodel讀取的杏节，只要修改名字，自己想要修改的層就能隨機(jī)初始化典阵。
- 修改名字奋渔，保留前面幾層的參數(shù)，同時后面的參數(shù)設(shè)置較高的學(xué)習(xí)率壮啊，基礎(chǔ)學(xué)習(xí)率大概0.00001左右嫉鲸。
