When drop_last is True in the DataLoader (train_loader), the lengths of annotations and sorted_segments in run.py are different. Then the predicted time and ground truth time are not matching, and the evalutation results (tIoU) are totally wrong.
The error can be fixed by selecting these annotations whose index are in the sorted_segments_dict.
When
drop_lastisTruein the DataLoader (train_loader), the lengths ofannotationsandsorted_segmentsinrun.pyare different. Then the predicted time and ground truth time are not matching, and the evalutation results (tIoU) are totally wrong.The error can be fixed by selecting these
annotationswhoseindexare in thesorted_segments_dict.