caffe官網:
1 #include "caffe/caffe.hpp
"2 #include
3 #include 4
using
namespace
caffe;56
char *proto = "
h:\\models\\caffe\\deploy.prototxt
"; /*
載入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);
1112
//! note: 後文所有提到的net,都是這個net
13 boost::shared_ptr< net > net(new caffe::net(proto, phase));
1char *model = "
h:\\models\\caffe\\bvlc_reference_caffenet.caffemodel
";
2 net->copytrainedlayersfrom(model);
1char *mean_file = "
h:\\models\\caffe\\imagenet_mean.binaryproto";
2 blobimage_mean;
3blobproto blob_proto;
4const
float *mean_ptr;
5 unsigned int
num_pixel;67
bool succeed = readprotofrombinaryfile(mean_file, &blob_proto);8if
(succeed)
9
1//! note: data_ptr指向已經處理好(去均值的,符合網路輸入影象的長寬和batch size)的資料
2void caffe_forward(boost::shared_ptr< net > & net, float *data_ptr)
318 net->forwardprefilled();
19 }
1//! note: net的blob是指,每個層的輸出資料,即feature maps2//
char *query_blob_name = "conv1";
3 unsigned int get_blob_index(boost::shared_ptr< net > & net, char *query_blob_name)413
}14 log(fatal) << "
unknown blob name:
"<15 }
1//! note: 根據caffenet的deploy.prototxt檔案,該net共有15個blob,從data一直到prob
2char *query_blob_name = "
conv1
"; /*
data, conv1, pool1, norm1, fc6, prob, etc
*/3 unsigned int blob_id =get_blob_index(net, query_blob_name);
45 boost::shared_ptrfloat> > blob = net->blobs()[blob_id];
6 unsigned int num_data = blob->count(); /*
nchw=10x96x55x55 */7
const
float *blob_ptr = (const
float *) blob->cpu_data();
1//! note: layer包括神經網路所有層,比如,caffenet共有23層2//
char *query_layer_name = "conv1";
3 unsigned int get_layer_index(boost::shared_ptr< net > & net, char *query_layer_name)413
}14 log(fatal) << "
unknown layer name:
"<15 }
1//! note: 不同於net的blob是feature maps,layer的blob是指conv和fc等層的weight和bias
2char *query_layer_name = "
conv1";
3const
float *weight_ptr, *bias_ptr;
4 unsigned int layer_id =get_layer_index(net, query_layer_name);
5 boost::shared_ptrfloat> > layer = net->layers()[layer_id];
6 std::vectorfloat> >> blobs = layer->blobs();
7if (blobs.size() > 0)8
1213
//! note: 訓練模式下,讀取指定layer的梯度資料,與此相似,唯一的區別是將cpu_data改為cpu_diff
1const
float* data_ptr; /*
指向待寫入資料的指標, 源資料指標*/2
float* weight_ptr = null; /*
指向網路中某層權重的指標,目標資料指標
*/3 unsigned int data_size; /*
待寫入的資料量 */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_ptrfloat> > blob = net->layers()[layer_id]->blobs()[0];8
9 check(data_size == blob->count());
10switch
(caffe::mode())
1121 caffe_copy(blob->count(), data_ptr, weight_ptr);
2223
//! note: 訓練模式下,手動修改指定layer的梯度資料,與此相似
24//
mutable_cpu_data改為mutable_cpu_diff,mutable_gpu_data改為mutable_gpu_diff
1char* weights_file = "
bvlc_reference_caffenet_new.caffemodel";
2netparameter net_param;
3 net->toproto(&net_param, false
);4 writeprototobinaryfile(net_param, weights_file);
Caffe使用教程
by shicai yang 星空下的巫師 on 2015 08 06 include caffe caffe.hpp include include using namespace caffe char proto h models caffe deploy.prototxt 載入caffenet...
caffe教程翻譯 Loss
與絕大多數的機器學習引擎一樣,caffe是由乙個lossfunction 損失函式 驅動的。loss function也稱為 error function,cost function或者objective function。loss function利用模型中的引數 比如模型中網路的weights ...
Caffe簡明教程1 Caffe簡介
您可以檢視所有文章的索引 caffe簡明教程0 文章列表 caffe是乙個很常用的深度學習框架,官網 在我個人的經歷中,經常遇到 作者在學術 中使用caffe來作為實驗框架。目前,caffe由伯克利ai研究所 bair 以及開源社群成員進行開發和維護。其原始作者是賈揚清。另外,當選擇乙個框架時,我們...