mcahny
commented
3 years ago Hi,
For training, I would recommend you to do multi-stage training. Also, in the training dataset, you will need to include easy samples such as Figure3-(b) in the paper https://openaccess.thecvf.com/content_CVPR_2019/papers/Kim_Deep_Video_Inpainting_CVPR_2019_paper.pdf .
1st stage training is the 2-tower network (1 reference, 1 target) which is the easiest stage. 2nd stage you may scale up to 5-tower network (1 ref, 4 tar), 3rd stage, you can put the recurrent feedback, 6-tower (1 ref, 5 ref) , where the last reference frame is the recurrent feedback.
Best, Dahun
2021년 3월 19일 (금) 오전 11:33, Crocs512 @.***>님이 작성:
Thank you very much for the given code, but I have a few questions about the details of the training.
After I have rewritten the training code according to the given training demo, I found that the results during the training process were vague and the loss was very unstable.
I used YouTube’s Vos dataset, I don’t know if it’s my implementation problem or what? So I look forward to the author can tell me more implementation details！！
Thanks！
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/mcahny/Deep-Video-Inpainting/issues/30, or unsubscribe https://github.com/notifications/unsubscribe-auth/AEX44LWZABNKVQAQF7IFMADTEKZZBANCNFSM4ZN5BB2A .
Thank you very much for the given code, but I have a few questions about the details of the training.
After I have rewritten the training code according to the given training demo, I found that the results during the training process were vague and the loss was very unstable.
I used YouTube’s Vos dataset, I don’t know if it’s my implementation problem or what? So I look forward to the author can tell me more implementation details！！
Thanks！