Practical-RIFE icon indicating copy to clipboard operation
Practical-RIFE copied to clipboard

Published 20 hours ago • verified publisher icon hzwer

Fine tuning

Open laugam opened this issue 3 years ago • 4 comments

Hi, thank you for making your code available. The latest model release (v4.6) works really well for my project. I would like to fine tune the model on my dataset. Do you plan to make the script for the fine tuning available any time soon? Thank you.

laugam avatar Sep 28 '22 08:09 laugam

https://github.com/megvii-research/ECCV2022-RIFE I think if you can train original RIFE, it will be not hard to finetune practical-RIFE v4.6. If you meet any problem, just raise the issue.

hzwer avatar Oct 01 '22 14:10 hzwer

Thank you for your reply. The train and evaluate functions that you defined in the file train.py don't seem to be compatible with the new files RIFE_HDv3.py and IFNet_HDv3.py, for instance the loss functions are different. Do you have any suggestions on how to modify those functions?

Many thanks.

laugam avatar Oct 07 '22 10:10 laugam

Hello, I am also trying to retrain or fine-tune the v4.6 model on my own data set recently. I also found that RIFE_HDv3.py and IFNet_HDv3.py are different from the original version. I modified a lot of codes, such as removing After using loss_cons, I found that the effect is very poor after successful training. Did you fine-tune successfully? How did you solve these problems? Thanks!

xsp1 avatar Jul 03 '23 03:07 xsp1

Hello, unfortunately I did not manage to train/fine-tune the v4.6 model, I'm using the model as it is for the moment. Maybe you will find useful the comments on this issue: https://github.com/megvii-research/ECCV2022-RIFE/issues/293#issue-1472024315

I asked for the training code there but they didn't reply yet. Please let me know if you succeed in this. Thanks!

laugam avatar Jul 06 '23 10:07 laugam