Loss problem

[jzhang538]

I use my own caffe and find the loss doesn't decrease a lot. After a period of training, nan happens. The data layer is same as predict process provided by you while label lnput is normalized to 0-1. I find that there is a Crop layer, which reshape output salient map to the shape of 'data'. However, salient map has 1 channel while data has 3 channels, why we need to make the shapes same. Is there any problem with it? Also, when I use shape 500,500 as width and height of input, nan happens very quickly while I use shape 512,512. the loss doesn't decrease a lot. It quite strange.