the loss is 0 when training model with custom dataset

[OpenSorceYCW]

Thanks for your excellent work. I meet the problem when I train the model with custom dataset, can you give me some advice to solve it. thank you, I am looking for you reply!

[dingfengshi]

HI! There are many reasons that might cause this situation. First, is the format of the data annotation JSON consistent with other datasets? Second, do the input features correspond to the parameters in your configuration file? Is the ground truth annotated correctly? You may need to further track intermediate outputs to find out where the issue lies.

[OpenSorceYCW]

Thank your very much! Another problem, I am puzzled about function of truncate_feats in https://github.com/dingfengshi/TriDet/blob/master/libs/datasets/thumos14.py#L182, whether to increase sample diversity? can you explain it detail, thank you!

[dingfengshi]

Thank your very much! Another problem, I am puzzled about function of truncate_feats in https://github.com/dingfengshi/TriDet/blob/master/libs/datasets/thumos14.py#L182, whether to increase sample diversity? can you explain it detail, thank you!

Oh, That's question haha! That is an implementation inherited from Actionformer, which I think is to prevent some long videos from consuming too much memory during training and has nothing to do with the features themselves.

[OpenSorceYCW]

ok, thank you for your reply!

[dingfengshi]

Issue has been sloved