search

w86763777 / pytorch-ddpm

Unofficial PyTorch implementation of Denoising Diffusion Probabilistic Models
Do What The F*ck You Want To Public License
477 stars 60 forks source link

About loss convergence #1

Closed HAN-oQo closed 1 year ago

HAN-oQo commented 2 years ago

I tried to train the model with your train code and 2 gpu. But the training loss seems to converge too fast around 0.04. Is it right? or Did you use different hyperparameter with the original paper?

Thank you

w86763777 commented 2 years ago

It seems correct. This is my loss curve. loss

All the hyperparameters are copied from the official implementation.

HAN-oQo commented 2 years ago

Thank you for your quick answer!

HAN-oQo commented 2 years ago

Can I see your FID, IS score logs while training? (If you have the record, I want to compare with mine)

Thank you!

w86763777 commented 2 years ago

I didn't record the FID and IS during training since it costs too much time to generate images for evaluation. However I can give the example at step 10k, 50k and 100k for your reference.

I actually tested 2 different random seeds and both experiments had similar results.