Merge remote-tracking branch 'origin/main' into kitti

softgroup/model/softgroup.py

@@ -166,6 +166,12 @@ def point_wise_loss(self, semantic_scores, pt_offsets, semantic_labels, instance
     @force_fp32(apply_to=('cls_scores', 'mask_scores', 'iou_scores'))
     def instance_loss(self, cls_scores, mask_scores, iou_scores, proposals_idx, proposals_offset,
                       instance_labels, instance_pointnum, instance_cls, instance_batch_idxs):
+        if proposals_idx.size(0) == 0 or (instance_cls != self.ignore_label).sum() == 0:
+            cls_loss = cls_scores.sum() * 0
+            mask_loss = mask_scores.sum() * 0
+            iou_score_loss = iou_scores.sum() * 0
+            return dict(cls_loss=cls_loss, mask_loss=mask_loss, iou_score_loss=iou_score_loss)
+
         losses = {}
         proposals_idx = proposals_idx[:, 1].cuda()
         proposals_offset = proposals_offset.cuda()
@@ -370,8 +376,12 @@ def forward_grouping(self,
             if proposals_idx.size(0) > 0:
                 proposals_idx_list.append(proposals_idx)
                 proposals_offset_list.append(proposals_offset)
-        proposals_idx = torch.cat(proposals_idx_list, dim=0)
-        proposals_offset = torch.cat(proposals_offset_list)
+        if len(proposals_idx_list) > 0:
+            proposals_idx = torch.cat(proposals_idx_list, dim=0)
+            proposals_offset = torch.cat(proposals_offset_list)
+        else:
+            proposals_idx = torch.zeros((0, 2), dtype=torch.int32)
+            proposals_offset = torch.zeros((0, ), dtype=torch.int32)
         return proposals_idx, proposals_offset
 
     def forward_instance(self, inst_feats, inst_map):
@@ -393,6 +403,9 @@ def forward_instance(self, inst_feats, inst_map):
     @force_fp32(apply_to=('semantic_scores', 'cls_scores', 'iou_scores', 'mask_scores'))
     def get_instances(self, scan_id, proposals_idx, semantic_scores, cls_scores, iou_scores,
                       mask_scores):
+        if proposals_idx.size(0) == 0:
+            return []
+
         num_instances = cls_scores.size(0)
         num_points = semantic_scores.size(0)
         cls_scores = cls_scores.softmax(1)
@@ -467,6 +480,17 @@ def clusters_voxelization(self,
                               scale,
                               spatial_shape,
                               rand_quantize=False):
+        if clusters_idx.size(0) == 0:
+            # create dummpy tensors
+            coords = torch.tensor(
+                [[0, 0, 0, 0], [0, spatial_shape - 1, spatial_shape - 1, spatial_shape - 1]],
+                dtype=torch.int,
+                device='cuda')
+            feats = feats[0:2]
+            voxelization_feats = spconv.SparseConvTensor(feats, coords, [spatial_shape] * 3, 1)
+            inp_map = feats.new_zeros((1, ), dtype=torch.long)
+            return voxelization_feats, inp_map
+
         batch_idx = clusters_idx[:, 0].cuda().long()
         c_idxs = clusters_idx[:, 1].cuda()
         feats = feats[c_idxs.long()]