Getting detections in training phase

[MichaelMonashev]

Why did you disable output['detections'] calculation here?

efficientdet-pytorch/effdet/bench.py

Line 136 in c5b694a

if not self.training:

It will useful to get bounding boxes and scores to calculate metrics while training.

[rwightman]

@MichaelMonashev it's rather expensive when it has no impact on training or model selection (via evaluation only). Easy enough to hack it or add a force flag.

[rwightman]

Also, to calculate meaningful metrics from the detections you have to do the full iou calcs, on train set size data (for larger datasets), that is REALLY slow and no way you'd want to have it active given that it doesn't actually factor into proper model selection.