notiho
notiho
I now also added logging for validation images and their prediction. How do you feel about adding another parameter to control the image logging?
If anyone is interested in trying out U-net in kraken, I just made #499
Here is my vgsl spec for a UFCN style net (with the limitation regarding normalization you observed): https://gist.github.com/notiho/53acb1db6fa885e206ee3caa728ccfe3
(Spaces replaced by newlines for better legibility)
Hi, I just had a look at your VGSL spec again and noticed that you added normalization after the up-convolution layers. While this is the way it is done in...
An alternative could be to pass that hint to all layers in a serial block, however, this would break if you have more than one transposed convolution in a series.
So in principle the Pytorch implementation of transposed convolution is able to handle inputs of any size without requiring any rescaling to multiples of the scale factor, as long as...