search

kpandey008 / DiffuseVAE

Official implementation of "DiffuseVAE: Efficient, Controllable and High-Fidelity Generation from Low-Dimensional Latents"
MIT License
344 stars 32 forks source link

FID score on cifar10 dataset #16

Open backwards123 opened 1 year ago

backwards123 commented 1 year ago

Hello, I trained your code, but after I trained 2850 epochs on cifar10, using form2 and not using GMM, the FID@50k I got was only 4.00, which is much different from the 2.86 in your paper. I'm wondering why this is happening? Can you give your complete list of hyperparameters during training? Thank you！

backwards123 commented 1 year ago

It may also be that there is a problem with the hyperparameters in my evaluation. If possible, please tell me the complete list of hyperparameters in your test, thank you

kpandey008 commented 1 year ago

Hi, Thanks for pointing this out. I noticed that the link to newer checkpoints for CIFAR-10 is missing. I will add a link to the new checkpoints soon.

[Edit] Furthermore, when training from scratch (only for CIFAR-10) you might want to use the rework-exact-repro branch instead since we found that using the Unet model used in the DDIM repo (https://github.com/ermongroup/ddim) gives best result both for the baseline DDPM and consequently DiffuseVAE

backwards123 commented 1 year ago

Thanks Reply! I know where the problem is.