Respect `strict=False` when loading detection models (#5841)

* Convert weights only if `old_key` is in `state_dict`

* Fix linter

torchvision/models/detection/mask_rcnn.py

@@ -317,6 +317,7 @@ class MaskRCNNHeads(nn.Sequential):
                 for type in ["weight", "bias"]:
                     old_key = f"{prefix}mask_fcn{i+1}.{type}"
                     new_key = f"{prefix}{i}.0.{type}"
+                    if old_key in state_dict:
                         state_dict[new_key] = state_dict.pop(old_key)
 
         super()._load_from_state_dict(

torchvision/models/detection/retinanet.py

@@ -45,6 +45,7 @@ def _v1_to_v2_weights(state_dict, prefix):
         for type in ["weight", "bias"]:
             old_key = f"{prefix}conv.{2*i}.{type}"
             new_key = f"{prefix}conv.{i}.0.{type}"
+            if old_key in state_dict:
                 state_dict[new_key] = state_dict.pop(old_key)
 
 

torchvision/models/detection/rpn.py

@@ -56,6 +56,7 @@ class RPNHead(nn.Module):
             for type in ["weight", "bias"]:
                 old_key = f"{prefix}conv.{type}"
                 new_key = f"{prefix}conv.0.0.{type}"
+                if old_key in state_dict:
                     state_dict[new_key] = state_dict.pop(old_key)
 
         super()._load_from_state_dict(

torchvision/ops/feature_pyramid_network.py

@@ -128,6 +128,7 @@ class FeaturePyramidNetwork(nn.Module):
                     for type in ["weight", "bias"]:
                         old_key = f"{prefix}{block}.{i}.{type}"
                         new_key = f"{prefix}{block}.{i}.0.{type}"
+                        if old_key in state_dict:
                             state_dict[new_key] = state_dict.pop(old_key)
 
         super()._load_from_state_dict(