Kernel width is 5 in the paper, but 4 in the code. about dcgan.torch

[jianglai3303]

I have the same question， is there anyone explained that？

[aneesh3108]

(Just my two cents)

I think its a typo of sorts since there is no way the image is getting up-scaled to 64x64 if kernal size of 5 is used.

Kernal size 4 cuts it just right given the parameter definitions and the formula used here:
https://github.com/torch/nn/blob/master/doc/convolution.md#nn.SpatialFullConvolution

[soumith]

there's a way to use kernel=5 and get 64x64, you just force-pad the edges. We did kernel size of 5 in the theano code (because of the way the cudnn Transpose was implemented with Theano), and we did ksz=4 with Torch code.

[aneesh3108]

Apologies.
Thanks for this info.
I was referring to the Torch code as such - didn't think Theano was in context too.