update

open-mmlab · Aug 15, 2023 · bef8e50 · bef8e50
1 parent 6b93792
commit bef8e50
Show file tree

Hide file tree

Showing 3 changed files with 4 additions and 7 deletions.
diff --git a/mmdeploy/codebase/mmdet/deploy/object_detection_model.py b/mmdeploy/codebase/mmdet/deploy/object_detection_model.py
@@ -205,7 +205,7 @@ def forward(self,
 
             bboxes = dets[:, :4]
             scores = dets[:, 4]
-            print(img_metas[i])
+
             # perform rescale
             if rescale and 'scale_factor' in img_metas[i]:
                 scale_factor = img_metas[i]['scale_factor']

diff --git a/mmdeploy/codebase/mmdet/models/dense_heads/rtmdet_ins_head.py b/mmdeploy/codebase/mmdet/models/dense_heads/rtmdet_ins_head.py
@@ -142,9 +142,7 @@ def _nms_with_mask_static(self,
         pre_top_k=pre_top_k,
         keep_top_k=keep_top_k,
         output_index=True)
-    # remove padded
-    dets = dets[:, :-1, :]
-    labels = labels[:, :-1]
+
     batch_size = bboxes.shape[0]
     batch_inds = torch.arange(batch_size, device=bboxes.device).view(-1, 1)
     kernels = kernels[batch_inds, inds, :]

diff --git a/mmdeploy/mmcv/ops/nms.py b/mmdeploy/mmcv/ops/nms.py
@@ -242,7 +242,7 @@ def _select_nms_index(scores: torch.Tensor,
         pre_inds = pre_inds.unsqueeze(0).repeat(batch_size, 1)
         pre_inds = pre_inds.where((batch_inds == batch_template.unsqueeze(1)),
                                   pre_inds.new_zeros(1))
-        pre_inds = torch.cat((pre_inds, pre_inds.new_zeros((N, 1))), 1)
+        pre_inds = torch.cat((pre_inds, -pre_inds.new_ones((N, 1))), 1)
     # sort
     is_use_topk = keep_top_k > 0 and \
         (torch.onnx.is_in_onnx_export() or keep_top_k < batched_dets.shape[1])
@@ -258,7 +258,6 @@ def _select_nms_index(scores: torch.Tensor,
     if output_index:
         if pre_inds is not None:
             topk_inds = pre_inds[topk_batch_inds, topk_inds, ...]
-        topk_inds = topk_inds[:, :-1]
         return batched_dets, batched_labels, topk_inds
     # slice and recover the tensor
     return batched_dets, batched_labels
@@ -285,7 +284,7 @@ def _multiclass_nms(boxes: Tensor,
     iou_threshold = torch.tensor([iou_threshold], dtype=torch.float32)
     score_threshold = torch.tensor([score_threshold], dtype=torch.float32)
     batch_size = scores.shape[0]
-
+    topk_inds = None
     if pre_top_k > 0:
         max_scores, _ = scores.max(-1)
         _, topk_inds = max_scores.topk(pre_top_k)