search

vlfeat / matconvnet

MatConvNet: CNNs for MATLAB
Other
1.4k stars 753 forks source link

Why all the parameters come to NaN when using small batchSize? #547

Closed jinluyang closed 7 years ago

jinluyang commented 8 years ago

I've been training on the ORL face database for face recognition use.I tried different batchSize and found that when using batchSize smallerthan 40,all the weights of the net turn into NaN after some epochs.the objective curve stops drawing and the error curve stays at 0.975 and 0.875.Can anyone tell me why?

jinluyang commented 8 years ago

I just found that when using a large learning rate,the same thing happens.

hyassine commented 7 years ago

Hi jinluyang

Which classifier are you using ?

Regards

jinluyang commented 7 years ago

@hyassine vl_nnconv It was about a year ago .I don't know whether the newest version has the same problem.

hyassine commented 7 years ago

@jinluyang would you mind send me your scripts (and the version of matconvnet used) or publish them to debug the issue ?

jinluyang commented 7 years ago

@hyassine I just tried the code and found that the problem doesn't occur no matter what batchsize I use. So I think maybe I should close this issue.