The model.py code for dealing with COCO crowds was handling the list of negative class ids, by checking 'torch.nonzero(gt_class_ids < 0).size()'; this can lead to bugs as 'torch.nonzero([-1 -2 -3] < 0).size()' is actually True. A better way of checking is doing 'len(torch.nonzero([-1 -2 -3] < 0).size()) > 0', which now works.
The model.py code for dealing with COCO crowds was handling the list of negative class ids, by checking 'torch.nonzero(gt_class_ids < 0).size()'; this can lead to bugs as 'torch.nonzero([-1 -2 -3] < 0).size()' is actually True. A better way of checking is doing 'len(torch.nonzero([-1 -2 -3] < 0).size()) > 0', which now works.