[FIX] add Ci of pytorch 1.10 and comments for bbox clamp (#7081)

.github/workflows/build.yml

@@ -33,7 +33,7 @@ jobs:
     strategy:
       matrix:
         python-version: [3.7]
-        torch: [1.5.1, 1.6.0, 1.7.0, 1.8.0, 1.9.0]
+        torch: [1.5.1, 1.6.0, 1.7.0, 1.8.0, 1.9.0, 1.10.1]
         include:
           - torch: 1.5.1
             torchvision: 0.6.1
@@ -50,6 +50,9 @@ jobs:
           - torch: 1.9.0
             torchvision: 0.10.0
             mmcv: 1.9.0
+          - torch: 1.10.1
+            torchvision: 0.11.2
+            mmcv: 1.10.0
     steps:
       - uses: actions/checkout@v2
       - name: Set up Python ${{ matrix.python-version }}
@@ -160,12 +163,16 @@ jobs:
     strategy:
       matrix:
         python-version: [3.6, 3.7, 3.8, 3.9]
-        torch: [1.9.0+cu102]
+        torch: [1.9.0+cu102, 1.10.1+cu102]
         include:
           - torch: 1.9.0+cu102
             torch_version: torch1.9.0
             torchvision: 0.10.0+cu102
             mmcv: 1.9.0
+          - torch: 1.10.1+cu102
+            torch_version: torch1.10.1
+            torchvision: 0.11.2+cu102
+            mmcv: 1.10.0
 
     steps:
       - uses: actions/checkout@v2

mmdet/models/dense_heads/autoassign_head.py

@@ -197,6 +197,9 @@ def forward_single(self, x, scale, stride):
         # scale the bbox_pred of different level
         # float to avoid overflow when enabling FP16
         bbox_pred = scale(bbox_pred).float()
+        # bbox_pred needed for gradient computation has been modified
+        # by F.relu(bbox_pred) when run with PyTorch 1.10. So replace
+        # F.relu(bbox_pred) with bbox_pred.clamp(min=0)
         bbox_pred = bbox_pred.clamp(min=0)
         bbox_pred *= stride
         return cls_score, bbox_pred, centerness

mmdet/models/dense_heads/fcos_head.py

@@ -154,6 +154,9 @@ def forward_single(self, x, scale, stride):
         # float to avoid overflow when enabling FP16
         bbox_pred = scale(bbox_pred).float()
         if self.norm_on_bbox:
+            # bbox_pred needed for gradient computation has been modified
+            # by F.relu(bbox_pred) when run with PyTorch 1.10. So replace
+            # F.relu(bbox_pred) with bbox_pred.clamp(min=0)
             bbox_pred = bbox_pred.clamp(min=0)
             if not self.training:
                 bbox_pred *= stride