[BE] No warning if user sets the log_prob_key explicitly and only one variable is sampled from the ProbTDMod

ghstack-source-id: ccc966d
Pull Request resolved: #1209

Author: Vincent Moens

tensordict/nn/distributions/composite.py

@@ -130,7 +130,9 @@ def __init__(
             dist_params = params.get(name)
             kwargs = extra_kwargs.get(name, {})
             if dist_params is None:
-                raise KeyError(f"no param {name} found in params with keys {params.keys(True, True)}")
+                raise KeyError(
+                    f"no param {name} found in params with keys {params.keys(True, True)}"
+                )
             dist = dist_class(**dist_params, **kwargs)
             dists[write_name] = dist
         self.dists = dists

tensordict/nn/probabilistic.py

@@ -328,6 +328,9 @@ class ProbabilisticTensorDictModule(TensorDictModuleBase):
 
     """
 
+    # To be removed in v0.9
+    _trigger_warning_lpk: bool = False
+
     def __init__(
         self,
         in_keys: NestedKey | List[NestedKey] | Dict[str, NestedKey],
@@ -396,9 +399,11 @@ def __init__(
                 "composite_lp_aggregate is set to True but log_prob_keys were passed. "
                 "When composite_lp_aggregate() returns ``True``, log_prob_key must be used instead."
             )
+        self._trigger_warning_lpk = len(self._out_keys) > 1
         if log_prob_key is None:
             if composite_lp_aggregate(nowarn=True):
                 log_prob_key = "sample_log_prob"
+                self._trigger_warning_lpk = True
             elif len(out_keys) == 1:
                 log_prob_key = _add_suffix(out_keys[0], "_log_prob")
         elif len(out_keys) > 1 and not composite_lp_aggregate(nowarn=True):
@@ -451,13 +456,15 @@ def log_prob_key(self):
                 f"unless there is one and only one element in log_prob_keys (got log_prob_keys={self.log_prob_keys}). "
                 f"When composite_lp_aggregate() returns ``False``, try to use {type(self).__name__}.log_prob_keys instead."
             )
-        if _composite_lp_aggregate.get_mode() is None:
+        if _composite_lp_aggregate.get_mode() is None and self._trigger_warning_lpk:
             warnings.warn(
                 f"You are querying the log-probability key of a {type(self).__name__} where the "
-                f"composite_lp_aggregate has not been set. "
+                f"composite_lp_aggregate has not been set and the log-prob key has not been chosen. "
                 f"Currently, it is assumed that composite_lp_aggregate() will return True: the log-probs will be aggregated "
-                f"in a {self._log_prob_key} entry. From v0.9, this behaviour will be changed and individual log-probs will "
-                f"be written in `('path', 'to', 'leaf', '<sample_name>_log_prob')`. To prepare for this change, "
+                f"in a {self._log_prob_key} entry. "
+                f"From v0.9, this behaviour will be changed and individual log-probs will "
+                f"be written in `('path', 'to', 'leaf', '<sample_name>_log_prob')`. "
+                f"To prepare for this change, "
                 f"call `set_composite_lp_aggregate(mode: bool).set()` at the beginning of your script (or set the "
                 f"COMPOSITE_LP_AGGREGATE env variable). Use mode=True "
                 f"to keep the current behaviour, and mode=False to use per-leaf log-probs.",

test/test_nn.py

@@ -2270,6 +2270,72 @@ def test_index_prob_seq(self):
         assert isinstance(seq[:2], ProbabilisticTensorDictSequential)
         assert isinstance(seq[-2:], ProbabilisticTensorDictSequential)
 
+    def test_no_warning_single_key(self):
+        # Check that there is no warning if the number of out keys is 1 and sample log prob is set
+        torch.manual_seed(0)
+        with set_composite_lp_aggregate(None):
+            mod = ProbabilisticTensorDictModule(
+                in_keys=["loc", "scale"],
+                distribution_class=torch.distributions.Normal,
+                out_keys=[("an", "action")],
+                log_prob_key="sample_log_prob",
+                return_log_prob=True,
+            )
+            td = TensorDict(loc=torch.randn(()), scale=torch.rand(()))
+            mod(td.copy())
+            mod.log_prob(mod(td.copy()))
+            mod.log_prob_key
+
+            # Don't set the key and trigger the warning
+            mod = ProbabilisticTensorDictModule(
+                in_keys=["loc", "scale"],
+                distribution_class=torch.distributions.Normal,
+                out_keys=[("an", "action")],
+                return_log_prob=True,
+            )
+            with pytest.warns(
+                DeprecationWarning, match="You are querying the log-probability key"
+            ):
+                mod(td.copy())
+                mod.log_prob(mod(td.copy()))
+                mod.log_prob_key
+
+            # add another variable, and trigger the warning
+            mod = ProbabilisticTensorDictModule(
+                in_keys=["params"],
+                distribution_class=CompositeDistribution,
+                distribution_kwargs={
+                    "distribution_map": {
+                        "dirich": torch.distributions.Dirichlet,
+                        "categ": torch.distributions.Categorical,
+                    }
+                },
+                out_keys=[("dirich", "categ")],
+                return_log_prob=True,
+            )
+            with pytest.warns(
+                DeprecationWarning, match="You are querying the log-probability key"
+            ), pytest.warns(
+                DeprecationWarning,
+                match="Composite log-prob aggregation wasn't defined explicitly",
+            ):
+                td = TensorDict(
+                    params=TensorDict(
+                        dirich=TensorDict(
+                            concentration=torch.rand(
+                                (
+                                    10,
+                                    11,
+                                )
+                            )
+                        ),
+                        categ=TensorDict(logits=torch.rand((5,))),
+                    )
+                )
+                mod(td.copy())
+                mod.log_prob(mod(td.copy()))
+                mod.log_prob_key
+
 
 class TestEnsembleModule:
     def test_init(self):