Caffe_MultiLabel_Classification

This is a instance of multilabel classification in caffe. Thanks to the tools of multi-label data conversion tool in https://github.com/HolidayXue/CodeSnap by HolidayXue.

Step

1、Recompile

Download convert_multilabel.cpp in https://github.com/HolidayXue/CodeSnap, replace convert_imageset.cpp, then recompile caffe. You will get a new exe :convert_multilabel.exe, which can make multi-label data.

2、Manufacture data

Download my ZnCar data in this page. Try to manufacture your own lmdb by the example command line:

convert_imageset.exe  --resize_height=227  --resize_width=227 ZnCar/ ZnCar/Label.txt ZnCarTrainImage ZnCarTrainLabel 2

We also need mean file:

compute_image_mean.exe ZnCarTrainImage ZnCarTrainMean.binaryproto

Generally speaking, you should get lmdb:ZnCarTrainImage/ZnCarTestImage/ZnCarTrainLabel/ZnCarTestlabel mean_file:ZnCarTrainMean.binaryproto/ZnCarTestMean.binaryproto finally.

3、Finetune AlexNet

We can use pretrained caffemodel:bvlc_alexnet.caffemodel to assign the weights to some of the layers that have no change(conv/pool/fc6/7). Finetune command line:

Build\x64\Release\caffe.exe train --solver=D:\caffe-master\models\bvlc_alexnet\solver.prototxt --weights=D:\caffe-master\models\bvlc_alexnet\bvlc_alexnet.caffemodel

You can refer to some of my parameters in solver.prototxt and multi_label_AlexNet.prototxt.

4.Modify classification.cpp

classification.exe in Caffe only support single label classification, we should modify classification.cpp. You can use my classification_multilabel.cpp and recompile just like step.1

5.Use our model to classify a picture

(deploy.prototxt is necessary.)We can use our model to classify a picture for 2 labels:

classification.exe deploy.prototxt network.caffemodel mean.binaryproto label1.txt label2.txt img.jpg

The source Image:

The result:

DataSet

You can try ZnCar.zip. It includes 3 vehicle types and 2 vehicle surfaces. 60 train imgs and 12 test imgs. http://download.csdn.net/detail/mr_curry/9742578

Note that this project only support 2 classes classification.

The network structure: