loss is negative and cannot overfit on mini-set of nuScenes

[XinchaoGou]

Thank you for your great work and sharing the code!
I'm trying to train a model to overfit on the mini-set for the first step. But it cannot overfit on the mini-set even after 1000 epochs.

According to the code, you have two losses for each task, one is "factor*loss_fn" and the other is for "uncertainty". I'm confused about the part "uncertainty" and I think it may hinder the training.

For example, loss for instance_offset is about 0.2-0.6.

But loss for instance_uncertainty is negative and contribute more (about -1.4 at 14k) to the final loss. So the model may just blindly decrease "uncertainty".

I wonder why you design the loss this why. I'll appreciate it If you have any reference or document to help me understand this part.

Best Regards,
Xinchao

[Anthony Hu]

Hi Xinchao,

The uncertainty loss corresponds to the Multitask uncertainty to automatically weight the difference losses instead of tuning them, as described in "Multi-Task Learning Using Uncertainty to Weigh Losses for Scene Geometry and Semantics, Kendall et al.".

Best,
Anthony