search

zhaoweicai / mscnn

Caffe implementation of our multi-scale object detection framework
404 stars 211 forks source link

Issues with changing batch size #92

Closed ChongYunPeng closed 6 years ago

ChongYunPeng commented 6 years ago

Hi,

The default model parameters worked well for me. However, when I changed the batch size for trainval_2nd.prototxt train phase from 4 to 2, I got the following losses which is strange. (With batch size 4 the loss is usually less than 1)

Train net output #38: loss_bbox = nan (* 1 = nan loss)
Train net output #39: loss_cls = 87.3365 (* 1 = 87.3365 loss)

Any idea what is the issue?

ChongYunPeng commented 6 years ago

Nevermind I think I found my mistake. I have to change the num_img_per_batch parameter in ProposalTarget.