Classification loss does not converge

[zhuzzjlu]

Thanks for your code!
I noticed that you did not record the classified loss. After adding the log of classification loss, I noticed that the classification loss does not converge. This leads to a bad effect of top1 performance.
Do you have any suggestions on the non convergence of classification loss.
Looking forward to your reply~

[ZikangZhou]

Yeah, this is indeed an issue, and I feel that cross-entropy loss is not a good choice. For the classification loss, you can try to optimize the NLL of the goal point's mixture of Laplace (detaching the gradient of y_hat and optimizing the classification score only). For the regression part, you can still use the winner-take-all LaplaceNLL loss. I found that this can improve the Brier score a lot. But ultimately, ensembling is the best way to improve the classification score. Some top-ranking methods like WayFormer even ensemble 15 models. However, it is impractical to do ensembling on a real autonomous car, and using ensembling causes unfair comparisons.