usage-of-caffe

来源：互联网发布：sql 附加数据库编辑：程序博客网时间：2024/06/07 10:05

----------------------------train------fine-turn-----------------------------------------------------

1. training stage

!/usr/bin/env shcaff_root=xxxsolver_file=xxxcaffe_file=caffe_root+'built/tools/caffe'$caffe train -solver $solver_file -weights xxx.caffemodel -gpu 1,2  2>&1 | tee $log_file-snapshot xxx.solverstate

----------------------------plot-acc-loss--------------------------------------

caff_root/tools/extra/plot_ [0-7] xxx.png xx.log

------------------------plot-net-------------------------------------------------

./python/draw_net.py xxx.prototxt xxx.png xx

-------------------------------fine turn------------------------------

在套用模型的时候会将层名相同的参数值作为初始值赋值给相应层的参数，那么就要求层名相同的的层的定义要完全一致，假设输出不一致，那么赋值失败。如果在原模型中没有找到相应的层，那么该层的参数初始值与普通训练无异。

------------------0-----------------label-------------

#!/bin/bashIMAGE_ROOT=xxxrm -rf train.txtfind IMAGE_ROOT/image/ -name *xx.jpg | cut -d '/' -f3 |sed "s/$/ 1/" >> train.txtfine IMAGE_ROOT/image -name *yy.png | cut -d '/' -f3 | sed "s/$/ 2/" >> train.txtecho "done..."

--------------------1--------------------------lmdb----

#!/bin/bashrm -rf xxx_lmdbbuild/tools/convert_imageset --shuffle --resize_height=xxx --resize_width=xxx  xx/images  train_label.txt  xxx_lmdbecho "done...."

label.txt 包含图片的名字空格类标

--------------------2-------------------------mean-----------------

build/tools/compute_image_mean xxx/train_lmdb xxx.binaryproto

----------------------------------------------------------------------------

//根据layer名字复制,layer的参数必须一样否则会报错...

//fineturn有些层发生儿变化必须要改layer名字

I0329 09:13:13.990504  5366 net.cpp:261] This network produces output loss1/loss1I0329 09:13:13.990509  5366 net.cpp:261] This network produces output loss1/top-1I0329 09:13:13.990512  5366 net.cpp:261] This network produces output loss1/top-5I0329 09:13:13.990516  5366 net.cpp:261] This network produces output loss2/loss1I0329 09:13:13.990520  5366 net.cpp:261] This network produces output loss2/top-1I0329 09:13:13.990523  5366 net.cpp:261] This network produces output loss2/top-5I0329 09:13:13.990527  5366 net.cpp:261] This network produces output loss3/loss3I0329 09:13:13.990530  5366 net.cpp:261] This network produces output loss3/top-1I0329 09:13:13.990533  5366 net.cpp:261] This network produces output loss3/top-5I0329 09:13:13.990656  5366 net.cpp:274] Network initialization done.I0329 09:13:13.991480  5366 solver.cpp:60] Solver scaffolding done.I0329 09:13:13.995544  5366 caffe.cpp:129] Finetuning from ./sub_tianchi_models/finetuning_37w.caffemodel     I0329 09:13:19.320947  5366 net.cpp:753] Ignoring source layer loss1/classifier                                                           I0329 09:13:19.325259  5366 net.cpp:753] Ignoring source layer loss2/classifier        I0329 09:13:19.327749  5366 net.cpp:753] Ignoring source layer loss3/classifierI0329 09:13:25.825736  5366 net.cpp:753] Ignoring source layer loss1/classifierI0329 09:13:25.830339  5366 net.cpp:753] Ignoring source layer loss2/classifierI0329 09:13:25.833221  5366 net.cpp:753] Ignoring source layer loss3/classifier

-----------------------------feature--extractor----------------------------------

mean.binaryproto -> mean.npy

-- > https://github.com/BVLC/caffe/issues/290

---> http://www.zhihu.com/question/28102651/answer/39397649

def mean_npy():
    mean_path = root_path + 'mean.binaryproto'
    blob = caffe.proto.caffe_pb2.BlobProto()
    data = open(mean_path,'rb').read()
    blob.ParseFromString(data)
    arr = np.array(caffe.io.blobproto_to_array(blob))
    np.save( root_path +'mean.npy', arr[0])

mean_blob = blob_pb2.BlobProto()

data =open(os.path.join(LMDB_PATH,"mean.binaryproto"),'rb').read()

mean_blob.ParseFromString(data)

img_mean = np.array(mean_blob.data).reshape(mean_blob.num,mean_blob.channels,mean_blob.height,mean_blob.width)

---------------------------------------------------------------

num_epochs = num_iterations * batch_size / set_siz

Reference:

[1] http://www.cnblogs.com/denny402/category/759199.html

[2] http://chrischoy.github.io/research/reading-protobuf-db-in-python/

0 0