Update

[ghstack-poisoned]

tensordict/nn/probabilistic.py

@@ -354,6 +354,16 @@ def __init__(
         if self.return_log_prob and self.log_prob_key not in self.out_keys:
             self.out_keys.append(self.log_prob_key)
 
+    @property
+    def dist_params_keys(self) -> List[NestedKey]:
+        """Returns all the keys pointing at the distribution params."""
+        return list(self.in_keys)
+
+    @property
+    def dist_sample_keys(self) -> List[NestedKey]:
+        """Returns all the keys pointing at the distribution samples."""
+        return [key for key in self.out_keys if key is not self.log_prob_key]
+
     def get_dist(self, tensordict: TensorDictBase) -> D.Distribution:
         """Creates a :class:`torch.distribution.Distribution` instance with the parameters provided in the input tensordict.
 
@@ -389,6 +399,8 @@ def get_dist(self, tensordict: TensorDictBase) -> D.Distribution:
                 raise err
         return dist
 
+    build_dist_from_params = get_dist
+
     def log_prob(
         self,
         tensordict,
@@ -944,6 +956,22 @@ def get_dist_params(
         with set_interaction_type(type):
             return tds(tensordict, tensordict_out, **kwargs)
 
+    @property
+    def dist_params_keys(self) -> List[NestedKey]:
+        """Returns all the keys pointing at the distribution params."""
+        result = []
+        for m in reversed(list(self._module_iter())):
+            result.extend(getattr(m, "dist_params_keys", []))
+        return result
+
+    @property
+    def dist_sample_keys(self) -> List[NestedKey]:
+        """Returns all the keys pointing at the distribution samples."""
+        result = []
+        for m in reversed(list(self._module_iter())):
+            result.extend(getattr(m, "dist_sample_keys", []))
+        return result
+
     @property
     def num_samples(self):
         num_samples = ()

test/test_nn.py

@@ -1992,6 +1992,8 @@ def test_probabilistic_n_samples(self, return_log_prob):
         assert td["sample"].shape == (2, 3, 4)
         if return_log_prob:
             assert "sample_log_prob" in td
+        assert prob.dist_sample_keys == ["sample"]
+        assert prob.dist_params_keys == ["loc"]
 
     @pytest.mark.parametrize("return_log_prob", [True, False])
     @pytest.mark.parametrize("inplace", [True, False])
@@ -2042,6 +2044,8 @@ def test_probabilistic_seq_n_samples(
             assert isinstance(log_prob, torch.Tensor)
         else:
             assert isinstance(log_prob, TensorDict)
+        assert seq.dist_sample_keys == ["sample"]
+        assert seq.dist_params_keys == ["loc"]
 
     @pytest.mark.parametrize("return_log_prob", [True, False])
     @pytest.mark.parametrize("inplace", [True, False])
@@ -2093,6 +2097,8 @@ def test_intermediate_probabilistic_seq_n_samples(
             assert isinstance(log_prob, torch.Tensor)
         else:
             assert isinstance(log_prob, TensorDict)
+        assert seq.dist_sample_keys == ["sample"]
+        assert seq.dist_params_keys == ["loc"]
 
     @pytest.mark.parametrize(
         "log_prob_key",
@@ -2118,6 +2124,8 @@ def test_nested_keys_probabilistic_delta(self, log_prob_key):
             return_log_prob=True,
             log_prob_key=log_prob_key,
         )
+        assert module.dist_sample_keys == [("data", "action")]
+        assert module.dist_params_keys == [("data", "param")]
         td_out = module(policy_module(td))
         assert td_out["data", "action"].shape == (3, 4, 1)
         if log_prob_key:
@@ -2132,6 +2140,8 @@ def test_nested_keys_probabilistic_delta(self, log_prob_key):
             return_log_prob=True,
             log_prob_key=log_prob_key,
         )
+        assert module.dist_sample_keys == [("data", "action")]
+        assert module.dist_params_keys == [("data", "param")]
         td_out = module(policy_module(td))
         assert td_out["data", "action"].shape == (3, 4, 1)
         if log_prob_key:
@@ -2912,6 +2922,8 @@ def test_prob_module(self, interaction, return_log_prob, map_names):
             default_interaction_type=interaction,
             return_log_prob=return_log_prob,
         )
+        assert module.dist_sample_keys == out_keys
+        assert module.dist_params_keys == in_keys
         if not return_log_prob:
             assert module.out_keys[-2:] == out_keys
         else:
@@ -3010,6 +3022,8 @@ def test_prob_module_nested(self, interaction, map_names):
         )
         # loosely checks that the log-prob keys have been added
         assert module.out_keys[-2:] != out_keys
+        assert module.dist_sample_keys == out_keys
+        assert module.dist_params_keys == in_keys
 
         sample = module(params)
         key_logprob0 = (
@@ -3077,6 +3091,8 @@ def test_prob_module_seq(self, interaction, return_log_prob, ordereddict):
                 )
             ]
         module = ProbabilisticTensorDictSequential(*args)
+        assert module.dist_sample_keys == out_keys
+        assert module.dist_params_keys == in_keys
         sample = module(params)
         if return_log_prob:
             assert "cont_log_prob" in sample.keys()
@@ -3156,6 +3172,8 @@ def test_prob_module_seq_nested(self, interaction):
                 log_prob_key=log_prob_key,
             ),
         )
+        assert module.dist_sample_keys == out_keys
+        assert module.dist_params_keys == in_keys
         sample = module(params)
         assert "cont_log_prob" in sample.keys()
         assert ("nested", "cont_log_prob") in sample.keys(True)