Skip to content

Replacing deconvolutions in GANs with resize-convolutions

Notifications You must be signed in to change notification settings

hvy/gan-resize-convolution

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

14 Commits
 
 
 
 
 
 
 
 

Repository files navigation

GAN Resize Convolution

Comparison between a regular GAN and the resize-convolution propsed in the article http://distill.pub/2016/deconv-checkerboard/ as an alternative to the standard deconvolution (transposed convolution) in the generator to get rid of the checkboard artifacts.

Upsampling

The upsampling was made using the 2d unpooling function in Chainer which is very similar to a nearest-neighbor interpolation.

Dataset

Training dataset of CIFAR-10.

Results

We see that the artifacts are less noticeable in the resize-convolutions and that the images look more natural, but that the difference becomes less obvious as the training goes on.

Left images are images generated using a regular GAN and right images are generated using the resize-convolutions technique.

1 iteration (Left: Regular GAN, Right: Resize-convolution)

1 epoch

2 epochs

3 epochs

4 epochs

5 epochs

6 epochs

7 epochs

8 epochs

9 epochs

10 epochs

About

Replacing deconvolutions in GANs with resize-convolutions

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages