Line 868: noise = th.randn_like(x_start) remains unchanged
KevinWang676 opened this issue · 5 comments
Hi, I found that in line 204 and 868 of guided_diffusion/gaussian_diffusion.py, the code noise = th.randn_like(x_start) remains unchanged. Shouldn't we have new_noise = noise + gamma * th.randn_like(noise) in line 868 as well? Or it doesn't matter. Thanks!
In line 204, I think it's okay to remain unchanged, but in line 868 I'm not sure if it's necessary to update the code (add new_noise).
Also, I wonder if ImageNet 32x32 dataset for training is the one I highlighted in the picture below. Thank you!
Hi, I found that in line 204 and 868 of guided_diffusion/gaussian_diffusion.py, the code
noise = th.randn_like(x_start)remains unchanged. Shouldn't we havenew_noise = noise + gamma * th.randn_like(noise)in line 868 as well? Or it doesn't matter. Thanks!
The only change our method makes and should only make is inside the training_losses function since our method is a training regularization
Also, I wonder if ImageNet 32x32 dataset for training is the one I highlighted in the picture below. Thank you!
@KevinWang676 Correct
Thanks for your reply!