I found that the val_loss can not converge while the train_loss can converge

[whytin]

Thank you first for your nice work! I use this model to train the new datasets (Penn Action), and I found that the val_loss can not converge while the train_loss can converge. I did not modify the config and the process. I want to figure out what causes this problem. Thanks! The train_loss convergence: The val_loss convergence:

[Superlee506]

@whytin In your example, over-fitting happens. You can try the following methods: 1) Add more training data or data augment; 2) increase the weight decay; 3) simplify the model

[whytin]

I will try, Thanks!