loss

[Lqm26]

Thank you very much for your work. We have a question about the loss in the code. Line 89 of pr_trainer.py uses loss_event + loss_image + kl_loss +cls_loss. We would like to ask what the cls_loss is and why it is added. The paper only says that loss_event + loss_image + kl_loss is used.

[Yan98]

Thank you for your interest. Please refer to here. cls_loss is used to monitor pre-training only, and gradients from labels to the backbone are prevented by detach.

[Lqm26]

Thanks for your answer. I have another question. Why do you need to perform the (data["event1"] - 0.5) * 2 operation on the event in lines 64 and 65 of pr_trainer.py?

[Yan98]

There is no specific reason for this, and it should also work if you do not perform the scaling.