Fix torch.load() in model_utils.py (#696)

Summary:
Pull Request resolved: #696

Addressing Github issue #690 (#690).

Since Opacus is primarily intended for academic use cases, it is considered low risk to utilize torch.load() with the setting "weights_only=False". Additionally, reminders have been added to the function description.

Reviewed By: iden-kalemaj

Differential Revision: D66999322

fbshipit-source-id: 6eddc7b5a0390809ef25fc9fe9e5d28bf6b55130

opacus/utils/module_utils.py

@@ -89,6 +89,8 @@ def clone_module(module: nn.Module) -> nn.Module:
     """
     Handy utility to clone an nn.Module. PyTorch doesn't always support copy.deepcopy(), so it is
     just easier to serialize the model to a BytesIO and read it from there.
+    When ``weights_only=False``, ``torch.load()`` uses "pickle" module implicity, which is known to be insecure.
+    Only load the model you trust.
 
     Args:
         module: The module to clone
@@ -99,7 +101,7 @@ def clone_module(module: nn.Module) -> nn.Module:
     with io.BytesIO() as bytesio:
         torch.save(module, bytesio)
         bytesio.seek(0)
-        module_copy = torch.load(bytesio)
+        module_copy = torch.load(bytesio, weights_only=False)
     next_param = next(
         module.parameters(), None
     )  # Eg, InstanceNorm with affine=False has no params