search

facebookresearch / supervision-by-registration

Supervision-by-Registration: An Unsupervised Approach to Improve the Precision of Facial Landmark Detectors
Other
762 stars 165 forks source link

landmarks still shake after 10 epochs #34

Closed Racha1992 closed 5 years ago

Racha1992 commented 5 years ago

Issue description

Hi! How. many epochs did you train? I trained for about 10 epochs on 50 300VW videos, but the actual effect is not obvious, landmarks still shake

D-X-Y commented 5 years ago

Hi, would you mind to show me the log? what kinds of data do you use to training the model? I train SBR more than 50 epochs by using 300W+300VW+YouTube-Face.

Racha1992 commented 5 years ago

I use 300W+300VW. Loss drops by 25%. Maybe it's not enough... but training procedure is so slow TAT

D-X-Y commented 5 years ago

Unfortunately, the codes are not optimized for efficiency. We are preparing for releasing an accelerated version and the new codes will come in two or three months.

Racha1992 commented 5 years ago

Thanks for reply! Looking forward for it

Racha1992 commented 5 years ago

Excuse me, shall you share your lk train log? I change the base network and the loss so that it produces the coordinates instead of landmark heatmaps. May I refer to your log so that I can know my project is right or not. Thanks!

D-X-Y commented 5 years ago

Hi @wangbingcan1990, it has been almost one year since I trained LK. I tried but did not find the log. Would you mind to show me your training log ? and maybe I can give some suggestions.

Racha1992 commented 5 years ago

@D-X-Y Thanks for your 'trying out' haha. Would you mind give me your email? Maybe sending my log files to you is a better way as there is too much content to display here.

Racha1992 commented 5 years ago

I can describe it roughly here. At the begining of training with sbr, lk-loss is about 0.60, and it drops to 0.520 after 18 epochs. From epoch 12 to epoch 18 it decreases from 0.541 to 0.521. I doubt if the rate of loss reduction is right. lr is 1e-4, the same with yours, and lk_loss accounts for about 50% of total loss. Trainset is 300W, 300VW and YouTubeFaces.

D-X-Y commented 5 years ago

Hi @wangbingcan1990 , my email is Xuanyi.Dong@student.uts.edu.au, feel free to contact me. How about your detection loss? In addition, YouTubeFaces has a lot of noise, and you should better filter some low-resolution videos.

Racha1992 commented 5 years ago

@D-X-Y Hi, my detection loss is normalized l2 loss, different from your heatmap loss. When train with SBR, detection loss rises a little. I agree with your sugesstion, and I will filter the videos. Thank u!

D-X-Y commented 5 years ago

You are welcome~ What is the number of your L2 loss? Since you use a different loss, one should also adjust the loss weight.

D-X-Y commented 5 years ago

Close because of no further response in 7 days.