search

lucidrains / lightweight-gan

Implementation of 'lightweight' GAN, proposed in ICLR 2021, in Pytorch. High resolution image generations that can be trained within a day or two
MIT License
1.63k stars 222 forks source link

what is the recommended --aug-prob ? #83

Open rom1504 opened 3 years ago

rom1504 commented 3 years ago

The default seems to be 0/None Shouldn't 0.25 be better ? (referring to https://github.com/lucidrains/stylegan2-pytorch/#low-amounts-of-training-data )

philhoefer commented 3 years ago

I don't know where you got the 0/None but normally it should set the augmentation probability automatically. This should be the code that does it:

if not exists(self.aug_prob) and num_samples < 1e5:
    self.aug_prob = min(0.5, (1e5 - num_samples) * 3e-6)
    print(f'autosetting augmentation probability to {round(self.aug_prob * 100)}%')

So this means that if you have less than 100,000 samples it sets the augmentation probability to (100,000 - num_samples) * 0.0003% but never higher than 50%. For example: If you have 10,000 samples the augmentation probability will be set to 27%.

Where this formula comes from and how well it works, I do not know. (But it seems reasonable enough.)