Fix "to_standard_module" for Ghost Clipping (meta-pytorch#754)

Summary:

Issue [meta-pytorch#749](meta-pytorch#749)

Under FGC, we use ``del p.grad_sample`` to release used per-sample gradients. However, when running ``to_standard_module``, w/o new changes, it will lead to attribute not found errors.

Reviewed By: iden-kalemaj

Differential Revision: D74019695

Author: HuanyuZhang

opacus/grad_sample/gsm_base.py

@@ -24,7 +24,12 @@
 
 logger = logging.getLogger(__name__)
 
-OPACUS_PARAM_MONKEYPATCH_ATTRS = ["_forward_counter", "_current_grad_sample"]
+OPACUS_PARAM_MONKEYPATCH_ATTRS = [
+    "grad_sample",
+    "_forward_counter",
+    "_current_grad_sample",
+    "_norm_sample",
+]
 
 
 class AbstractGradSampleModule(nn.Module, ABC):
@@ -131,18 +136,15 @@ def to_standard_module(self) -> nn.Module:
         return self._module
 
     def _close(self):
-        self.del_grad_sample()
-        self._clean_up_attributes()
-
-    def __repr__(self):
-        return f"{type(self).__name__}({self._module.__repr__()})"
-
-    def _clean_up_attributes(self):
+        # Clean up attributes
         for attr in OPACUS_PARAM_MONKEYPATCH_ATTRS:
             for p in self.parameters():
                 if hasattr(p, attr):
                     delattr(p, attr)
 
+    def __repr__(self):
+        return f"{type(self).__name__}({self._module.__repr__()})"
+
     def forbid_grad_accumulation(self):
         """
         Sets a flag to detect gradient accumulation (multiple forward/backward passes

opacus/tests/grad_sample_module_test.py

@@ -130,7 +130,7 @@ def test_to_standard_module(self):
             self.original_model.state_dict(),
             strict=True,
         )
-        new_grad_sample_module = GradSampleModule(
+        new_grad_sample_module = self.CLS(
             copy_of_original_model, batch_first=True, loss_reduction="mean"
         )