lib/roi_data/fast_rcnn_rel.py [401:465]: - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - replace=False) rel_fg_inds = np.append(rel_pos_inds, rel_gt_inds) # duplicate low-shot predicates to increase their chances to be chosen if cfg.TRAIN.OVERSAMPLE2: rel_pos_labels = rel_gt_labels[gt_assignment_pair_min[rel_fg_inds]] - 1 low_shot_inds = \ np.array([rel_fg_inds[i] for i, p in enumerate(rel_pos_labels) if low_shot_helper.check_low_shot_p([-1, p, -1])], dtype=np.int32) rel_fg_inds = np.append(low_shot_inds, rel_fg_inds) if rel_fg_inds.size > fg_rois_per_image: rel_fg_inds = npr.choice(rel_fg_inds, size=fg_rois_per_image, replace=False) rel_bg_inds = np.where((max_overlaps_pair_min < cfg.TRAIN.BG_THRESH_HI))[0] rel_bg_rois_per_this_image = min(rois_per_image - rel_fg_inds.size, rois_per_image - fg_rois_per_image, rel_bg_inds.size) if rel_bg_inds.size > 0: rel_bg_inds = npr.choice(rel_bg_inds, size=rel_bg_rois_per_this_image, replace=False) # This oversampling method has redundant computation on those # low-shot ROIs, but it's flexible in that those low-shot ROIs # can be fed into the oversampler immediately after ROI-pooling, # instead of as late as after fc7 if cfg.TRAIN.OVERSAMPLE: # Only consider low-shot on P rel_pos_labels = rel_gt_labels[gt_assignment_pair_min[rel_fg_inds]] - 1 # low_shot_inds contains one dummy ROI at the very beginning # This is to make sure that low_shot ROIs are never empty low_shot_inds = \ np.array([rel_fg_inds[i] for i, p in enumerate(rel_pos_labels) if low_shot_helper.check_low_shot_p([-1, p, -1])], dtype=np.int32) rel_fg_inds = np.append(low_shot_inds, rel_fg_inds) if cfg.TRAIN.ADD_LOSS_WEIGHTS: # low-shot on P rel_pos_labels = rel_gt_labels[gt_assignment_pair_min[rel_fg_inds]] - 1 rel_pos_weights = np.ones_like(rel_pos_labels, dtype=np.float32) low_shot_idx = \ np.array([i for i, p in enumerate(rel_pos_labels) if low_shot_helper.check_low_shot_p([-1, p, -1])], dtype=np.int32) rel_pos_weights[low_shot_idx] *= 2.0 rel_pos_weights /= np.mean(rel_pos_weights) if cfg.TRAIN.ADD_LOSS_WEIGHTS_SO: # low-shot on S sbj_pos_weights = np.ones_like(sbj_pos_labels, dtype=np.float32) low_shot_idx = \ np.array([i for i, s in enumerate(sbj_pos_labels) if low_shot_helper.check_low_shot_s([s, -1, -1])], dtype=np.int32) sbj_pos_weights[low_shot_idx] *= 2.0 sbj_pos_weights /= np.mean(sbj_pos_weights) # low-shot on O obj_pos_weights = np.ones_like(obj_pos_labels, dtype=np.float32) low_shot_idx = \ np.array([i for i, o in enumerate(obj_pos_labels) if low_shot_helper.check_low_shot_o([-1, -1, o])], dtype=np.int32) obj_pos_weights[low_shot_idx] *= 2.0 obj_pos_weights /= np.mean(obj_pos_weights) rel_keep_inds = np.append(rel_fg_inds, rel_bg_inds) rel_rois_sbj = rel_all_rois_sbj[rel_keep_inds] rel_rois_obj = rel_all_rois_obj[rel_keep_inds] rel_rois_prd = rel_all_rois_prd[rel_keep_inds] - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - lib/roi_data/fast_rcnn_rel.py [624:688]: - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - replace=False) rel_fg_inds = np.append(rel_pos_inds, rel_gt_inds) # duplicate low-shot predicates to increase their chances to be chosen if cfg.TRAIN.OVERSAMPLE2: rel_pos_labels = rel_gt_labels[gt_assignment_pair_min[rel_fg_inds]] - 1 low_shot_inds = \ np.array([rel_fg_inds[i] for i, p in enumerate(rel_pos_labels) if low_shot_helper.check_low_shot_p([-1, p, -1])], dtype=np.int32) rel_fg_inds = np.append(low_shot_inds, rel_fg_inds) if rel_fg_inds.size > fg_rois_per_image: rel_fg_inds = npr.choice(rel_fg_inds, size=fg_rois_per_image, replace=False) rel_bg_inds = np.where((max_overlaps_pair_min < cfg.TRAIN.BG_THRESH_HI))[0] rel_bg_rois_per_this_image = min(rois_per_image - rel_fg_inds.size, rois_per_image - fg_rois_per_image, rel_bg_inds.size) if rel_bg_inds.size > 0: rel_bg_inds = npr.choice(rel_bg_inds, size=rel_bg_rois_per_this_image, replace=False) # This oversampling method has redundant computation on those # low-shot ROIs, but it's flexible in that those low-shot ROIs # can be fed into the oversampler immediately after ROI-pooling, # instead of as late as after fc7 if cfg.TRAIN.OVERSAMPLE: # Only consider low-shot on P rel_pos_labels = rel_gt_labels[gt_assignment_pair_min[rel_fg_inds]] - 1 # low_shot_inds contains one dummy ROI at the very beginning # This is to make sure that low_shot ROIs are never empty low_shot_inds = \ np.array([rel_fg_inds[i] for i, p in enumerate(rel_pos_labels) if low_shot_helper.check_low_shot_p([-1, p, -1])], dtype=np.int32) rel_fg_inds = np.append(low_shot_inds, rel_fg_inds) if cfg.TRAIN.ADD_LOSS_WEIGHTS: # low-shot on P rel_pos_labels = rel_gt_labels[gt_assignment_pair_min[rel_fg_inds]] - 1 rel_pos_weights = np.ones_like(rel_pos_labels, dtype=np.float32) low_shot_idx = \ np.array([i for i, p in enumerate(rel_pos_labels) if low_shot_helper.check_low_shot_p([-1, p, -1])], dtype=np.int32) rel_pos_weights[low_shot_idx] *= 2.0 rel_pos_weights /= np.mean(rel_pos_weights) if cfg.TRAIN.ADD_LOSS_WEIGHTS_SO: # low-shot on S sbj_pos_weights = np.ones_like(sbj_pos_labels, dtype=np.float32) low_shot_idx = \ np.array([i for i, s in enumerate(sbj_pos_labels) if low_shot_helper.check_low_shot_s([s, -1, -1])], dtype=np.int32) sbj_pos_weights[low_shot_idx] *= 2.0 sbj_pos_weights /= np.mean(sbj_pos_weights) # low-shot on O obj_pos_weights = np.ones_like(obj_pos_labels, dtype=np.float32) low_shot_idx = \ np.array([i for i, o in enumerate(obj_pos_labels) if low_shot_helper.check_low_shot_o([-1, -1, o])], dtype=np.int32) obj_pos_weights[low_shot_idx] *= 2.0 obj_pos_weights /= np.mean(obj_pos_weights) rel_keep_inds = np.append(rel_fg_inds, rel_bg_inds) rel_rois_sbj = rel_all_rois_sbj[rel_keep_inds] rel_rois_obj = rel_all_rois_obj[rel_keep_inds] rel_rois_prd = rel_all_rois_prd[rel_keep_inds] - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -