Avoid checkerboard pattern

[sheerun]

Hey,

Your algorithm has been mentioned in recent article about checkerboard artifacts. It says you could avoid them by "switching deconvolutional layers for resize-convolution layers". Not sure what it means, but I thought I could mention it..

[sheerun]

The link to article: http://distill.pub/2016/deconv-checkerboard/

[xlvector]

I find in https://github.com/jcjohnson/fast-neural-style/blob/master/fast_neural_style/models.lua#L96, it has already support nearest neighbor upsampling to fix this problem.

[htoyryla]

Nice find @xlvector . I am now training with

-arch c9s1-16,d32,d64,R64,R64,R64,R64,R64,U2,U2,c9s1-3

and it looks good so far.

[jcjohnson]

@htoyryla U2 is just a nearest-neighbor upsampling; you may get better results if you stick a stride-1 convolution between the two upsampling layers:

-arch c9s1-16,d32,d64,R64,R64,R64,R64,R64,U2,c3s1-32,U2,c9s1-3

[htoyryla]

Thanks for the tip! Even if I didn't have any problems with it. BTW in the udnie test today I used this architecture.

[fehmi]

What about vertical lines?