Loss not decreasing in stage 1

[Bismuth209]

After training stage-1 for 30000 steps on TikTok dataset I'm getting the following loss curve and images from validation_pipeline is this correct?

[guoqincode]

I also encountered this problem during the initial training. Could you tell me your detailed training config?

[Bismuth209]

It's not much different from what you used.

https://github.com/Bismuth209/AnimateAnyone-unofficial/blob/main/configs/training/my_train_stage_1.yaml

[guoqincode]

guoqin@stu.pku.edu.cn

[Bismuth209]

I have a couple of questions

1. Were you able to train stage 1 with decreasing loss?
2. In stage 1 training, I see that you are not loading motion modules. wouldn't it be beneficial to load them from animate-diff?

[guoqincode]

1. The loss is high at the beginning and becomes low after a few hundred iterations
2. The first stage does not need to train the motion module, just the image level pre-training

[Bismuth209]

How do you verify if the first stage has passed?

[guoqincode]

Test the results on the validation set.

[Bismuth209]

can you show sample results that you received? I have shared what I was getting at the top of the comments

[tdlhyj]

I use the latest training code, but the loss does not decrease