Did anyone get good CIFAR10 results?

Hi, thanks for providing this code. I’m trying to reproduce the CIFAR10 results from the original DDPM paper. I use 3x32x32 images, all the CIFAR data (50k frames), 2000 epochs (but I check every 100 epochs how it looks like), and I get some similar results, but not as good as the paper. This is the result that I get: I’m also attaching the training results that I get (the divergent one is the validation loss): My training schedule is similar to the original except that I maximize the batch size on my GPUs. I’m using image size of 32, and U-Net options dim=64, dim_mults=(1,2,4,8).

Was anyone more successful and can share their results and tips? I think that this result is far from perfect. Thanks very much, I hope you could help me find what I’m missing.

[yiyixuxu]

you can try p2_loss_weight_gamma = 1. , my result with cifar10 wasn’t great either with the default setting but I think you can see a big difference with p2 weighting - the original paper used a reweighted loss too

[tcapelle]

I do here, but with a different codebase: https://wandb.ai/capecape/train_sd/reports/How-to-Train-a-Conditional-Diffusion-Model-from-Scratch–VmlldzoyNzIzNTQ1

Sent from ProtonMail mobile

-------- Original Message -------- On Nov 5, 2022, 6:30 PM, Michael Albergo < @.***> wrote:

does someone have example code for mnist/cifar10?

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: @.***>