fix(model-builder-rel): bbox delta normalization uses (+2 more)

[andrewwhitecdw]

This PR addresses 3 issues in lib/modeling_rel/model_builder_rel.py.

Fix 1

fix: bbox delta normalization uses ndarray.view with tensor shape

Fix: Replace:

            if cfg.TRAIN.BBOX_NORMALIZE_TARGETS_PRECOMPUTED:
                # (legacy) Optionally normalize targets by a precomputed mean and stdev
                box_deltas = box_deltas.view(-1, 4) * cfg.TRAIN.BBOX_NORMALIZE_STDS \
                             + cfg.TRAIN.BBOX_NORMALIZE_MEANS

with:

            if cfg.TRAIN.BBOX_NORMALIZE_TARGETS_PRECOMPUTED:
                # (legacy) Optionally normalize targets by a precomputed mean and stdev
                box_deltas = box_deltas.reshape(-1, 4) * cfg.TRAIN.BBOX_NORMALIZE_STDS \
                             + cfg.TRAIN.BBOX_NORMALIZE_MEANS

Fix 2

fix: DEDUP_BOXES branch references undefined inv_index

Fix: Replace:

        if cfg.DEDUP_BOXES > 0 and not cfg.MODEL.FASTER_RCNN:
            # Map scores and predictions back to the original set of boxes
            scores = scores[inv_index, :]
            pred_boxes = pred_boxes[inv_index, :]

with:

        if cfg.DEDUP_BOXES > 0 and not cfg.MODEL.FASTER_RCNN:
            raise NotImplementedError(
                'DEDUP_BOXES requires inv_index, which is not available here')

Fix 3

fix: prepare_det_rois uses training batch size at test time

Fix: Replace:

        for im_i in range(cfg.TRAIN.IMS_PER_BATCH):

with:

        for im_i in range(im_info.shape[0]):

Files changed

• lib/modeling_rel/model_builder_rel.py