The network training is fine without adding any contrast normalization layer
(all types), but ones add the contrast normalization layers, after several
iterations the net gets nan values. I tried different values of the size,
scale and pow values, and tried to place the layer before and after pooling
layer.
Original issue reported on code.google.com by zybzm...@gmail.com on 8 May 2013 at 7:44
Original issue reported on code.google.com by
zybzm...@gmail.com
on 8 May 2013 at 7:44