Generated Images have noise in them

[shreyasbapat]

I am getting a lot of a patterned noise in the generated dataset. Any known reason?

[3DTOPO]

It usually improves after a number of epochs. How many epochs was that trained for?

[shreyasbapat]

It's 200 epochs!

[3DTOPO]

How many images are in your training set?

[shreyasbapat]

1500. Every image size is around 390x190

[shreyasbapat]

@junyanz Can you help?

[junyanz]

I am not sure what caused it.

[shreyasbapat]

Can more epochs fix it?

[prash-p]

I get the same noise patterns in my checkpoint images when using 128x350 images, more epochs don't seem to fix the problem.

[junyanz]

If you keep getting the checkboard artifacts, you can try different upsampling layers. See this post for more details. There is a PR related to this.

[shreyasbapat]

Right! My solution also lead to similar direction. Meanwhile that PR has some merge conflicts!

[CaelynCheung1996]

I have the same patterns when training the CycleGAN. However, it only happens in one direction (translating StyleA to StyleB). Any idea what might cause this? My Style A images are in higher resolution and B is lower resolution. Thx!

[HolicAnn]

@shreyasbapat Hello, I would like to ask why the result of your rectangular image is OK and the image has not changed. When I was training, I used --preprocess crop --crop_size 256. When I used --preprocess none during testing, an error was reported.