hzwer / ECCV2022-RIFE

ECCV2022 - Real-Time Intermediate Flow Estimation for Video Frame Interpolation
MIT License
4.44k stars 441 forks source link

Is the teacher model has trained parameter when load pretrained model? #284

Closed guoyilin closed 2 years ago

guoyilin commented 2 years ago

Hi, thanks for your exciting work, my problem is when I finetune the pretrained model on my small data(mostly contain slow human action), the l1-loss is oscillation ,not decrease, I was wondering if the pretrained model (Loaded ArXiv-RIFE model) contained the trained teacher parameter? or maybe my data is too hard for train(contain the hand move)?

hzwer commented 2 years ago

contained the trained teacher parameter? I think yes.

image

Hello, this is the result of my training on vimeo90K for 1000 epochs. It can be seen that the descending curve of the VFI model is very flat.

guoyilin commented 2 years ago

thanks, @hzwer , thanks for your reply, maybe i should contain training.