search

HKUST-Aerial-Robotics / Stereo-RCNN

Code for 'Stereo R-CNN based 3D Object Detection for Autonomous Driving' (CVPR 2019)
MIT License
690 stars 177 forks source link

loss is negative #21

Closed xueyushan closed 5 years ago

xueyushan commented 5 years ago

Why is the loss negative? What do the plus or minus of the loss represent? Thanks very much!

xueyushan commented 5 years ago

@Peiliang

PeiliangLi commented 5 years ago

It's normal. Because each loss is weighted by their uncertainty: loss = sum(loss_i * torch.exp(-uncert[i]) + uncert[i]). As the loss_i becomes smaller, the uncertainty will become negative.

xueyushan commented 5 years ago

The larger the absolute value of the loss,the network is in convergence?

PeiliangLi commented 5 years ago

The optimizing variable is the loss instead of abs(loss).