[Fix] fix runtime error of "+=" inplace operation in PyTorch 1.10 (#8439)

* fix runtime error of "+=" inplace operation in PyTorch 1.10

* minor fix

Author: BIGWangYuDong

mmdet/models/dense_heads/rpn_head.py

@@ -62,7 +62,7 @@ def _init_layers(self):
     def forward_single(self, x):
         """Forward feature map of a single scale level."""
         x = self.rpn_conv(x)
-        x = F.relu(x, inplace=True)
+        x = F.relu(x, inplace=False)
         rpn_cls_score = self.rpn_cls(x)
         rpn_bbox_pred = self.rpn_reg(x)
         return rpn_cls_score, rpn_bbox_pred

mmdet/models/dense_heads/solov2_head.py

@@ -141,7 +141,9 @@ def forward(self, feats):
                                                  input_p.device)
                 input_p = torch.cat([input_p, coord_feat], 1)
 
-            feature_add_all_level += self.convs_all_levels[i](input_p)
+            # fix runtime error of "+=" inplace operation in PyTorch 1.10
+            feature_add_all_level = feature_add_all_level + \
+                self.convs_all_levels[i](input_p)
 
         feature_pred = self.conv_pred(feature_add_all_level)
         return feature_pred

mmdet/models/necks/dyhead.py

@@ -109,7 +109,8 @@ def forward(self, x):
             summed_levels = 1
             if level > 0:
                 low_feat = self.spatial_conv_low(x[level - 1], offset, mask)
-                sum_feat += low_feat * self.scale_attn_module(low_feat)
+                sum_feat = sum_feat + \
+                    low_feat * self.scale_attn_module(low_feat)
                 summed_levels += 1
             if level < len(x) - 1:
                 # this upsample order is weird, but faster than natural order
@@ -119,7 +120,8 @@ def forward(self, x):
                     size=x[level].shape[-2:],
                     mode='bilinear',
                     align_corners=True)
-                sum_feat += high_feat * self.scale_attn_module(high_feat)
+                sum_feat = sum_feat + high_feat * \
+                    self.scale_attn_module(high_feat)
                 summed_levels += 1
             outs.append(self.task_attn_module(sum_feat / summed_levels))
 

mmdet/models/necks/pafpn.py

@@ -111,7 +111,8 @@ def forward(self, inputs):
         used_backbone_levels = len(laterals)
         for i in range(used_backbone_levels - 1, 0, -1):
             prev_shape = laterals[i - 1].shape[2:]
-            laterals[i - 1] += F.interpolate(
+            # fix runtime error of "+=" inplace operation in PyTorch 1.10
+            laterals[i - 1] = laterals[i - 1] + F.interpolate(
                 laterals[i], size=prev_shape, mode='nearest')
 
         # build outputs

mmdet/models/roi_heads/htc_roi_head.py

@@ -72,7 +72,7 @@ def forward_dummy(self, x, proposals):
             if self.with_semantic and 'mask' in self.semantic_fusion:
                 mask_semantic_feat = self.semantic_roi_extractor(
                     [semantic_feat], mask_rois)
-                mask_feats += mask_semantic_feat
+                mask_feats = mask_feats + mask_semantic_feat
             last_feat = None
             for i in range(self.num_stages):
                 mask_head = self.mask_head[i]
@@ -133,7 +133,7 @@ def _mask_forward_train(self,
             if mask_semantic_feat.shape[-2:] != mask_feats.shape[-2:]:
                 mask_semantic_feat = F.adaptive_avg_pool2d(
                     mask_semantic_feat, mask_feats.shape[-2:])
-            mask_feats += mask_semantic_feat
+            mask_feats = mask_feats + mask_semantic_feat
 
         # mask information flow
         # forward all previous mask heads to obtain last_feat, and fuse it
@@ -167,7 +167,7 @@ def _bbox_forward(self, stage, x, rois, semantic_feat=None):
             if bbox_semantic_feat.shape[-2:] != bbox_feats.shape[-2:]:
                 bbox_semantic_feat = adaptive_avg_pool2d(
                     bbox_semantic_feat, bbox_feats.shape[-2:])
-            bbox_feats += bbox_semantic_feat
+            bbox_feats = bbox_feats + bbox_semantic_feat
         cls_score, bbox_pred = bbox_head(bbox_feats)
 
         bbox_results = dict(cls_score=cls_score, bbox_pred=bbox_pred)
@@ -186,7 +186,7 @@ def _mask_forward_test(self, stage, x, bboxes, semantic_feat=None):
             if mask_semantic_feat.shape[-2:] != mask_feats.shape[-2:]:
                 mask_semantic_feat = F.adaptive_avg_pool2d(
                     mask_semantic_feat, mask_feats.shape[-2:])
-            mask_feats += mask_semantic_feat
+            mask_feats = mask_feats + mask_semantic_feat
         if self.mask_info_flow:
             last_feat = None
             last_pred = None
@@ -459,7 +459,7 @@ def simple_test(self, x, proposal_list, img_metas, rescale=False):
                 if self.with_semantic and 'mask' in self.semantic_fusion:
                     mask_semantic_feat = self.semantic_roi_extractor(
                         [semantic_feat], mask_rois)
-                    mask_feats += mask_semantic_feat
+                    mask_feats = mask_feats + mask_semantic_feat
                 last_feat = None
 
                 num_bbox_per_img = tuple(len(_bbox) for _bbox in _bboxes)
@@ -600,7 +600,7 @@ def aug_test(self, img_feats, proposal_list, img_metas, rescale=False):
                                 -2:]:
                             mask_semantic_feat = F.adaptive_avg_pool2d(
                                 mask_semantic_feat, mask_feats.shape[-2:])
-                        mask_feats += mask_semantic_feat
+                        mask_feats = mask_feats + mask_semantic_feat
                     last_feat = None
                     for i in range(self.num_stages):
                         mask_head = self.mask_head[i]

mmdet/models/roi_heads/roi_extractors/generic_roi_extractor.py

@@ -68,7 +68,7 @@ def forward(self, feats, rois, roi_scale_factor=None):
                 roi_feats_t = self.pre_module(roi_feats_t)
             if self.aggregation == 'sum':
                 # and sum them all
-                roi_feats += roi_feats_t
+                roi_feats = roi_feats + roi_feats_t
             else:
                 # and concat them along channel dimension
                 roi_feats[:, start_channels:end_channels] = roi_feats_t

mmdet/models/roi_heads/roi_extractors/single_level_roi_extractor.py

@@ -91,11 +91,11 @@ def forward(self, feats, rois, roi_scale_factor=None):
                 mask = mask.float().unsqueeze(-1)
                 # select target level rois and reset the rest rois to zero.
                 rois_i = rois.clone().detach()
-                rois_i *= mask
+                rois_i = rois_i * mask
                 mask_exp = mask.expand(*expand_dims).reshape(roi_feats.shape)
                 roi_feats_t = self.roi_layers[i](feats[i], rois_i)
-                roi_feats_t *= mask_exp
-                roi_feats += roi_feats_t
+                roi_feats_t = roi_feats_t * mask_exp
+                roi_feats = roi_feats + roi_feats_t
                 continue
             inds = mask.nonzero(as_tuple=False).squeeze(1)
             if inds.numel() > 0:
@@ -109,7 +109,7 @@ def forward(self, feats, rois, roi_scale_factor=None):
                 # in other GPUs and will cause a hanging error.
                 # Therefore, we add it to ensure each feature pyramid is
                 # included in the computation graph to avoid runtime bugs.
-                roi_feats += sum(
+                roi_feats = roi_feats + sum(
                     x.view(-1)[0]
                     for x in self.parameters()) * 0. + feats[i].sum() * 0.
         return roi_feats

mmdet/models/roi_heads/scnet_roi_head.py

@@ -110,7 +110,7 @@ def _bbox_forward(self,
             if bbox_semantic_feat.shape[-2:] != bbox_feats.shape[-2:]:
                 bbox_semantic_feat = adaptive_avg_pool2d(
                     bbox_semantic_feat, bbox_feats.shape[-2:])
-            bbox_feats += bbox_semantic_feat
+            bbox_feats = bbox_feats + bbox_semantic_feat
         if self.with_glbctx and glbctx_feat is not None:
             bbox_feats = self._fuse_glbctx(bbox_feats, glbctx_feat, rois)
         cls_score, bbox_pred, relayed_feat = bbox_head(
@@ -137,7 +137,7 @@ def _mask_forward(self,
             if mask_semantic_feat.shape[-2:] != mask_feats.shape[-2:]:
                 mask_semantic_feat = F.adaptive_avg_pool2d(
                     mask_semantic_feat, mask_feats.shape[-2:])
-            mask_feats += mask_semantic_feat
+            mask_feats = mask_feats + mask_semantic_feat
         if self.with_glbctx and glbctx_feat is not None:
             mask_feats = self._fuse_glbctx(mask_feats, glbctx_feat, rois)
         if self.with_feat_relay and relayed_feat is not None: