faustomilletari/VNet

How to change the loss?

sagarhukkire opened this issue · 5 comments

sagarhukkire commented

hi @faustomilletari

I see Vnet has been tested on softmax with loss ,if i want to use it instead of dice for testing then which changes I have to consider?

Thanks

faustomilletari commented
sagarhukkire commented

i just want to try softmax with loss, I went through doc , so if I comment dice loss in prototxt and write softmax with loss ,it will work? as it is provided by caffe itself. What your input

faustomilletari commented
sagarhukkire commented

hi @faustomilletari

I was just checking prototxt in Hough-CNN ,and comparing your comment in Vnet.py

 #solver.net.blobs['labelWeight'].data[...] = batchWeight.astype(dtype=np.float32)
            #use only if you do softmax with loss

then it means for softmax with loss layer in Prototxt, there will be three bottom ? as you have in hough cnn? I do not get this , or can I continue with two bottom ( pred,label) and one top (loss)

layer {
  name: "loss"
  type: "SoftmaxWithLoss"
  bottom: "pred"
  bottom: "label"
  bottom: "weight"
  top: "loss"
}
sagarhukkire commented

@faustomilletari
By the way great job man !! you added sample weight which is not available in normal caffe. Hope they will add it
After so much findings ,why there are three bottom I found out following blog

http://deepdish.io/2014/11/04/caffe-with-weighted-samples/

I guess that's reason you added three bottom in SoftmaxWithLoss