VAR icon indicating copy to clipboard operation
VAR copied to clipboard

How can I change the scale for training?

Open AlexzQQQ opened this issue 1 year ago • 4 comments

Thanks for the excellent work, I tried to change the default scale to others scale(e.g like 3 6 10),it can't work and the output is noise. and I wonder if it is ok to change the default scale without retrain for vqvae

AlexzQQQ avatar Dec 05 '24 06:12 AlexzQQQ

when i tried i change var.py class VAR's patch_nums, class VARHF's patch_nums,and arg_util.py pn

AlexzQQQ avatar Dec 05 '24 07:12 AlexzQQQ

It's OK to slightly change the scale schedule for vqvae since it adopts a CNN architecture. It could still encodes and decodes images normally but with a slight performance drop. However, VAR relys on the predefined scale schedule in the training stage. If you change the scale schedule without training VAR, it may generates noisy results. We suggest you fine-tune VAR if you change the scale schedule. The results will improves quickly.

JeyesHan avatar Dec 13 '24 08:12 JeyesHan

@AlexzQQQ You can also train var with dynamic scale schedule, which means that you can random drop some scales, when testing, you can use dynamic scale schedule. However i strongly recommend you to fine-tune VAR if you change the scale schedule.

enjoyyi00 avatar Dec 15 '24 07:12 enjoyyi00

@iFighting thanks a lot! I am working on retrain var for inpainting research, hope i can get a good result :)

AlexzQQQ avatar Dec 16 '24 08:12 AlexzQQQ

@AlexzQQQ how is the performance in inpainting?

egbertYeah avatar Dec 18 '24 07:12 egbertYeah

still trying    ------------------ Original ------------------ From: @.>; Date:  Wed, Dec 18, 2024 03:47 PM To: @.>; Cc:  "张昱辉Zhang @.>; @.>; Subject:  Re: [FoundationVision/VAR] How can I change the scale for training? (Issue #97)

 

@AlexzQQQ how is the performance in inpainting?

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

AlexzQQQ avatar Dec 19 '24 06:12 AlexzQQQ