Open Yifei-Y opened 2 years ago
And I found that this bug is the reason why the loss drops off a cliff at epoch=9 (when nc_loss is included). Is this your special design, or is it really a bug? I think this loss curve is very strange and there is no explanation in the paper.
I have the same confusion
In the class SetCriterion forward function of models/deformable_detr.py, both of the nested for loops use i as the iteration variable (Line 530 and Line 543). And after the internal for loop ends, it still uses the variable i. Isn't this a bug? Could you explain this?