openai / improved-diffusion

Release for Improved Denoising Diffusion Probabilistic Models

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

About model training

xinyuz8270 opened this issue · comments

commented

Do I need to take any other steps during the training process? After training for a few days, the loss did not continue to decrease and the sampling results obtained were also very blurry. Do anyone have any training suggestions?
Here are some of my sampling results. I trained it for 250K iterations.
real_data

Hi there. Did you solve your problem? I had the same issue. How many images do u have for the training?

I had the same issue. How many images do u have for the training?

My dataset is only 1000 images, but after checking the code, I found that we should use random crop instead of using center crop (because only 1k images). In my case, If I train from scratch (4 GPUs A100), after 100k steps the generated images will be clearer. If I use a pre-trained model to train, after about 50k steps the generated would be better.

Hi there. Did you solve your problem? I had the same issue. How many images do u have for the training?

Sorry for not replying in time. The ImageNet dataset I used for training has about 1.2 million images. After checking my code, I found that my problem was caused by a coding error in the sampling process. So I'm sorry I can't answer your question.