[BugFix] Apply inverse transform to input of TransformedEnv._reset

ghstack-source-id: 5f7c1fbd19b716f2b1602c34cf2ae1362f7bc7f6
Pull Request resolved: #2787

test/test_env.py

@@ -4058,6 +4058,35 @@ def test_chess_tokenized(self):
         assert "fen" in ftd["next"]
         env.check_env_specs()
 
+    @pytest.mark.parametrize("stateful", [False, True])
+    @pytest.mark.parametrize("include_san", [False, True])
+    def test_env_reset_with_hash(self, stateful, include_san):
+        env = ChessEnv(
+            include_fen=True,
+            include_hash=True,
+            include_hash_inv=True,
+            stateful=stateful,
+            include_san=include_san,
+        )
+        cases = [
+            # (fen, num_legal_moves)
+            ("5R1k/8/8/8/6R1/8/8/5K2 b - - 0 1", 1),
+            ("8/8/2kq4/4K3/1R3Q2/8/8/8 w - - 0 1", 2),
+            ("6R1/8/8/4rq2/3pPk2/5n2/8/2B1R2K b - e3 0 1", 2),
+        ]
+        for fen, num_legal_moves in cases:
+            # Load the state by fen.
+            td = env.reset(TensorDict({"fen": fen}))
+            assert td["fen"] == fen
+            assert td["action_mask"].sum() == num_legal_moves
+            # Reset to initial state just to make sure that the next reset
+            # actually changes the state.
+            assert env.reset()["action_mask"].sum() == 20
+            # Load the state by fen hash and make sure it gives the same output
+            # as before.
+            td_check = env.reset(td.select("fen_hash"))
+            assert (td_check == td).all()
+
 
 class TestCustomEnvs:
     def test_tictactoe_env(self):

torchrl/envs/transforms/transforms.py

@@ -1018,6 +1018,10 @@ def _reset(self, tensordict: Optional[TensorDictBase] = None, **kwargs):
             tensordict = tensordict.select(
                 *self.reset_keys, *self.state_spec.keys(True, True), strict=False
             )
+            # Inputs might be transformed, so need to apply inverse transform
+            # before passing to the env reset function.
+            with _set_missing_tolerance(self.transform, True):
+                tensordict = self.transform.inv(tensordict)
         tensordict_reset = self.base_env._reset(tensordict, **kwargs)
         if tensordict is None:
             # make sure all transforms see a source tensordict