kyotovision-public / dynamic-3d-gaze-from-afar

MIT License
49 stars 6 forks source link

Val loss does not decrease #10

Open Ulrich-1995 opened 1 year ago

Ulrich-1995 commented 1 year ago

Hi, there. I have tried training your model on GAFA dataset from scratch, and there seems to be an overfitting. The direction_loss and kappa_loss continue to decrease on the training dataset, however, the val_loss and val_mae hit mininum at the end of first epoch, and then keep increasing. The learning rates for both direction_loss and kappa_loss start at 1e-4. How did you avoid the overfit when you trained your model? Thank you so much in advance