[BugFix] Fix params.clone (#509)

tensordict/nn/params.py

@@ -440,7 +440,35 @@ def cuda(self, device=None):
         return TensorDictParams(params)
 
     def clone(self, recurse: bool = True) -> TensorDictBase:
-        return TensorDictParams(self._param_td.clone(recurse=recurse))
+        """Clones the TensorDictParams.
+
+        The effect of this call is different from a regular torch.Tensor.clone call
+        in that it will create a TensorDictParams instance with a new copy of the
+        parameters and buffers __detached__ from the current graph.
+
+        See :meth:`tensordict.TensorDictBase.clone` for more info on the clone
+        method.
+
+        """
+        if not recurse:
+            return TensorDictParams(self._param_td.clone(False), no_convert=True)
+        out = {}
+        for key, val in self._param_td.items(True, True):
+            if isinstance(val, nn.Parameter):
+                out[key] = nn.Parameter(
+                    val.data.clone(), requires_grad=val.requires_grad
+                )
+            else:
+                out[key] = Buffer(val.data.clone(), requires_grad=val.requires_grad)
+        return TensorDictParams(
+            TensorDict(
+                out,
+                batch_size=self._param_td.batch_size,
+                device=self._param_td.device,
+                names=self._param_td.names,
+            ),
+            no_convert=True,
+        )
 
     @_fallback
     def chunk(self, chunks: int, dim: int = 0) -> tuple[TensorDictBase, ...]:

test/test_nn.py

@@ -28,7 +28,7 @@
 from tensordict.nn.ensemble import EnsembleModule
 from tensordict.nn.functional_modules import is_functional, make_functional
 from tensordict.nn.probabilistic import InteractionType, set_interaction_type
-from tensordict.nn.utils import set_skip_existing, skip_existing
+from tensordict.nn.utils import Buffer, set_skip_existing, skip_existing
 from torch import nn
 from torch.distributions import Normal
 
@@ -2875,6 +2875,27 @@ def test_td_params_post_hook(self):
         assert not param_td.get("e").requires_grad
         assert not param_td.get(("a", "b", "c")).requires_grad
 
+    def test_tdparams_clone(self):
+        td = TensorDict(
+            {
+                "a": {
+                    "b": {"c": nn.Parameter(torch.zeros((), requires_grad=True))},
+                    "d": Buffer(torch.zeros((), requires_grad=False)),
+                },
+                "e": nn.Parameter(torch.zeros((), requires_grad=True)),
+                "f": Buffer(torch.zeros((), requires_grad=False)),
+            },
+            [],
+        )
+        td = TensorDictParams(td, no_convert=True)
+        tdclone = td.clone()
+        assert type(tdclone) == type(td)  # noqa
+        for key, val in tdclone.items(True, True):
+            assert type(val) == type(td.get(key))  # noqa
+            assert val.requires_grad == td.get(key).requires_grad
+            assert val.data_ptr() != td.get(key).data_ptr()
+            assert (val == td.get(key)).all()
+
 
 if __name__ == "__main__":
     args, unknown = argparse.ArgumentParser().parse_known_args()