Closed missTL closed 3 months ago
We also found this issue recently, but have not figured out where it comes from. It would happen one or two times during the whole training, thus we didn't notice before. We did some experiments and found it would appear when combing the preposed GKT-h and SGQ, although they both work well seperately.
Delete the following may solve this problem. https://github.com/HXMap/MapQR/blob/7dc705740fe026b7908b10e60fcd2846f9bb83d9/projects/mmdet3d_plugin/bevformer/modules/encoder.py#L149-L150
During training, grad_norm suddenly becomes nan