How to get the trained_ddpm_checkpoint of custom data?

Open wangxiaodong1021 opened this issue 3 years ago • 3 comments

Jan 07 '23 15:01 wangxiaodong1021

If no trained_ddpm_checkpoint exists for your data, you may need to train it by yourself using UNet. We will update the code to support regular ddpm training later.

Jan 08 '23 01:01 ckczzj

Could you share the code for regular ddpm training? I want to train this model with my custom dataset, thanks

Mar 21 '23 12:03 minkyujeon

We have updated the regular ddpm training code.

To train your own ddpm, you should add or modify following files:

dataset/your_dataset.py
config/your_dataset_regular.yaml
model/denoise_fn/your_dataset.py
model/denoise_fn/init.py
trainer/train_regular.py

After training, you may extract checkpoint.pt ("ema_denoise_fn") and config.yml into pre-trained-dpms folder, like other existing ones.

Mar 21 '23 14:03 ckczzj