search

hzwer / ECCV2022-RIFE

ECCV2022 - Real-Time Intermediate Flow Estimation for Video Frame Interpolation
MIT License
4.41k stars 438 forks source link

about training the RIFE_HDv2 #147

Closed tqyunwuxin closed 2 years ago

tqyunwuxin commented 3 years ago

hello , I train the RIFE_HD successfully ,but training RIFE_HDv2 failed. the optical flow labels is produced by liteflownet with Vimeo90K dataset.

to train RIFE_HDv2, I use " 3e-4*mul" in function "get_learning_rate", batch size 16, and loss "loss_G = loss_l1 + loss_cons + loss_ter" , optimG "weight_decay=1e-5"

Is there anything else to pay special attention to？ it is a little difficult to converge because the output of flownet is four channels？

hzwer commented 3 years ago

Hi, I set weight_decay=1e-4 and get_learning_rate to "1e-4*mul" as paper discribed. The training should be more stable.