alibaba/animate-anything

Finetuning animate-anything VRAM issue

mylee95 opened this issue · 2 comments

mylee95 commented

I tried to finetune animate-anything model with single sample video preprocessed with Video-BLIP2-Preprocessor Script, using example/train_mask_motion.yaml config file and single A6000 GPU(48GB), but it keeps on getting CUDA out of memory error trying to allocate more than 48GB.

Then I tried to reduce resolution in config file from 512x512 to 256x256, but it says tensor size does not match while forwarding through unet_3d_blocks.

It would be nice if you can share the config file and GPU that you used to finetune animate-anything with around 30GB VRAM, as it is written in README, and what do we need to do if we are going to downscale the input resolution.

efzero commented

You need to change batch size from 8 to around 2 for getting enough VRAM

  • 👍1
daizuozhuo commented

I update the code for smaller VRAM requirement.

  • 👍1