Why isn't negative sampling used during the training of ssd detector?

[allgrom]

Hi! Why isn't negative sampling used during the training of ssd detector?

self._target_generator = SSDTargetGenerator(
            iou_thresh=iou_thresh, stds=box_norm, negative_mining_ratio=-1, **kwargs)

[zhreshold]

TLDR. It's moved to MultiBoxLoss. Reason: OHEM is dependent on prediction value, but we want to pre-compute all targets using cpu cores during data loading. Instead we can use cpu to compute all targets and mask out negative samples in MultiBoxLoss

[allgrom]

Thank you!