What should I change if I change the number of class in training?

[heesu-pia]

I'm trying to train my own data with ZF model. There is a difference in the number of class with VOC2007. VOC2007 has 20 classes and my own datasets has 3 classes. So I've got this error because of the unmatched number of class.

I0106 20:31:24.703429 1352 net.cpp:746] Copying source layer relu7 I0106 20:31:24.703429 1352 net.cpp:746] Copying source layer drop7 I0106 20:31:24.703429 1352 net.cpp:743] Ignoring source layer fc8 I0106 20:31:24.703429 1352 net.cpp:743] Ignoring source layer prob F0106 20:31:25.450764 1352 smooth_L1_loss_layer.cpp:32] Check failed: bottom[0]->channels() == bottom[1]->channels() (84 vs. 16) F0106 20:31:25.450764 1352 smooth_L1_loss_layer.cpp:32] Check failed: bottom[0]->channels() == bottom[1]->channels() (84 vs. 16)

I think that the 84 means (20+1)*4 and the 16 means (3+1)*4. The 20 is the number of class in VOC2007 and the 3 is the number of class in my own datasets.

So what should I change to solve this problem?

[abner2015]

you can change in the file of VOCinit.m . the dir is VOCdevkit\VOCcode