search

NVlabs / stylegan2-ada-pytorch

StyleGAN2-ADA - Official PyTorch implementation
https://arxiv.org/abs/2006.06676
Other
4.07k stars 1.15k forks source link

Very strange result. About the "augment_pipe.p" #175

Open shoutOutYangJie opened 3 years ago

shoutOutYangJie commented 3 years ago

I find that you didn't limit the ”augment_pipe.p“ at the range of [0, 1). In my training step, the "augment_pipe.p" achieve "21.838".

"sec/kimg 4.23 maintenance 0.0 cpumem 5.71 gpumem 3.16 augment 21.838"

As a result, the generated images are inclined to be in the augmented distribution rather than the original distribution. I train on my custom dataset, whose images contain the cloud and sky. when at 10281 iterations, the generated image like this, and it is normal. image

However, when at 21974 iterations, the generated images are augmented images. The color is not normal. image

I guess the too high "augment_pipe.p" leads to this.

shoutOutYangJie commented 3 years ago

I supply some tensorboard curve. Please help me. image image image

waizai99 commented 2 years ago

Same error, did you solve it?

PDillis commented 2 years ago

Don't use color augment. It barely improves the results and usually leaks as the strength needs to be higher towards the end. I recommend only using bg and you shouldn't see those drastic color changes.

YukiSakuma commented 2 years ago

Don't use color augment. It barely improves the results and usually leaks as the strength needs to be higher towards the end. I recommend only using bg and you shouldn't see those drastic color changes.

Interesting so what kind of dataset color augmentation is good for?

PDillis commented 2 years ago

Don't use color augment. It barely improves the results and usually leaks as the strength needs to be higher towards the end. I recommend only using bg and you shouldn't see those drastic color changes.

Interesting so what kind of dataset color augmentation is good for?

I don't think there is one (from the perspective of having a dataset with a low amount of data). The effects you get from this augmentation are interesting, but you could achieve these manually afterward if you so wished. Also, as per the paper:

image

So using the color augmentation will yield a model that is on par with using no augmentations, hence why the recommended setting is to use bg (but remember, these were tests done on FFHQ, so perhaps your dataset will benefit from this augmentation). I haven't found this to be the case for my datasets, so I am leaning toward these findings being general.

Hong753 commented 1 year ago

Shouldn't the correct implementation limit the upper range of p as well?