resize-convolution instead transposed-convolution to avoid checkerboard artifacts #64
Conversation
…heckerboard artifacts
|
Did you also try PixelShuffle? |
|
I am now testing a transformer-net alternative with PixelShuffle using Fast-Neural-Style loss, CycleGAN still have some convergence issues, and the problem seems to be invariant of the capacity |
|
Thanks for your PR, @botcs ! I took a brief look at the code. This PR contains much more than what you have described. You should definitely clean up stuff like normalization, video, printing, etc. Some questions:
By the way, nn.Upsample is available now. You may also want to uncomment this line and remove the now deprecated nn.UpsamplingBilinear2d? :) |
|
Yeah, sorry for the inconvenience I will clean the code up in two weeks, and make another PR, with all the listed details - if that is acceptable for you |
|
Is there any activity on this topic? I'm also interested to implement this feature - at least to get it work ;-) |
|
@beniroquai I have an implementation here: #190 (bottom of the thread) |
|
Wow that was quick! Thanks a lot. Did you plan to make a PR in this repo as well? So the only thing is to exchange the resnet code part, right? |
|
@beniroquai that's right. Hope that it helps! :) |
|
Thanks! It's working way better now? Maybe stupid question, but Would it
make sense to do the same in the unet?
…On 9 Feb 2018 20:35, "Tongzhou Wang" ***@***.***> wrote:
@beniroquai <https://github.com/beniroquai> that's right. Hope that it
helps! :)
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
<#64 (comment)>,
or mute the thread
<https://github.com/notifications/unsubscribe-auth/AEJOuEo4EI6GOHR3zXebIt8U4ggeGOx3ks5tTJ36gaJpZM4OctGj>
.
|
|
@beniroquai Maybe, you can try adept the code to unet. It’d be cool if it
also helps unet. If you do some experiment, please let me know the results.
:)
…On Sat, Feb 10, 2018 at 12:28 beniroquai ***@***.***> wrote:
Thanks! It's working way better now? Maybe stupid question, but Would it
make sense to do the same in the unet?
On 9 Feb 2018 20:35, "Tongzhou Wang" ***@***.***> wrote:
> @beniroquai <https://github.com/beniroquai> that's right. Hope that it
> helps! :)
>
> —
> You are receiving this because you were mentioned.
> Reply to this email directly, view it on GitHub
> <
#64 (comment)
>,
> or mute the thread
> <
https://github.com/notifications/unsubscribe-auth/AEJOuEo4EI6GOHR3zXebIt8U4ggeGOx3ks5tTJ36gaJpZM4OctGj
>
> .
>
—
You are receiving this because you commented.
Reply to this email directly, view it on GitHub
<#64 (comment)>,
or mute the thread
<https://github.com/notifications/unsubscribe-auth/AFaWZV0w1lXfzsGgZrlgJBapM_yQ--KOks5tTdGogaJpZM4OctGj>
.
|
|
Thanks for getting back! Would you suggest to exchange the convTranspose2D
or do I need to do anything else? I'm not that familiar with the coding in
pytorch though. I'll have a try! Thanks!
|
|
Seems nearest neighbour upsampling not solving all issues, but bring another one - random color edges: |
| @@ -149,21 +163,40 @@ def __init__(self, input_nc, output_nc, ngf=64, norm_layer=nn.BatchNorm2d, use_d | |||
| for i in range(n_downsampling): | |||
| mult = 2**i | |||
| model += [nn.Conv2d(ngf * mult, ngf * mult * 2, kernel_size=3, | |||
| stride=2, padding=1), | |||
| stride=1, padding=1), | |||
| nn.MaxPool2d(2), | |||
There was a problem hiding this comment.
Hi! Why Pooling is used for avoiding checkerboard artifacts if
" Max pooling was previously linked to high-frequency artifacts in [12].)" https://distill.pub/2016/deconv-checkerboard/
See the reasoning behind the reimplementation:
https://distill.pub/2016/deconv-checkerboard/
I became aware of some checkerboard effects while training the CyclicGAN network:
http://i.imgur.com/uTgsC2l.png
because it resembled this:
http://i.imgur.com/M8r47Bp.png
now it looks like this (during training on Van Gogh set):
http://i.imgur.com/wf7B9sT.png