The reproduce obama video looks blurry
pensz opened this issue · 0 comments
pensz commented
env :
Ubuntu 20.04.5 LTS
libcudnn8 8.7.0.84-1+cuda11.8
I have fixed some compile error
- fa = face_alignment.FaceAlignment(face_alignment.LandmarksType._2D, flip_input=False)
+ fa = face_alignment.FaceAlignment(face_alignment.LandmarksType.TWO_D, flip_input=False)
- c_flags = ['-O3', '-std=c++14']
+ c_flags = ['-O3', '-std=c++17']
after i test the train result. the video looks very blurry.
when i use the pretrained model, It works fine.
Is the epoch too small?
[INFO] Trainer: ngp | 2023-12-09_14-18-09 | cuda | fp16 | trial_obama/
[INFO] #parameters: 3024277
[INFO] Loading latest checkpoint ...
[WARN] No checkpoint found, model randomly initialized.
[INFO] load 100 val frames.
[INFO] load aud_features: torch.Size([7999, 44, 16])
Loading val data: 100%| █ | 100/100 [00:00<00:00, 195.24it/s]
[INFO] eye_area: 0.1254321038722992 - 0.2622222304344177
[INFO] max_epoch = 44
Any suggestions would be appreciated.