loss=nan

[lee-vius]

When I tried to train on a partial scannet dataset (contructed as you guys did), sometimes the logging shows that "loss =nan". I loaded the checkpoint you provided.

Note that when I initialize the model without checkpoint and do the training, the logging easily shows that "loss =nan". Any idea what's the reason?

[lee-vius]

It seems in the "loftr_loss.py", the function "_compute_fine_loss_l2_std" sometimes output a nan loss when there doesn't exist ground truth matches for input image pair.

Not sure how to solve it.