[Feature,Refactor] Chess improvements: fen, pgn, pixels, san

ghstack-source-id: c3899e8
Pull Request resolved: #2702

Author: Vincent Moens

test/test_env.py

@@ -3291,6 +3291,10 @@ def test_batched_dynamic(self, break_when_any_done):
         )
         del env_no_buffers
         gc.collect()
+        # print(dummy_rollouts)
+        # print(rollout_no_buffers_serial)
+        # # for a, b in zip(dummy_rollouts.exclude("action").unbind(0), rollout_no_buffers_serial.exclude("action").unbind(0)):
+        #     assert_allclose_td(a, b)
         assert_allclose_td(
             dummy_rollouts.exclude("action"),
             rollout_no_buffers_serial.exclude("action"),
@@ -3386,35 +3390,107 @@ def test_partial_rest(self, batched):
 
 # fen strings for board positions generated with:
 # https://lichess.org/editor
-@pytest.mark.parametrize("stateful", [False, True])
 @pytest.mark.skipif(not _has_chess, reason="chess not found")
 class TestChessEnv:
-    def test_env(self, stateful):
-        env = ChessEnv(stateful=stateful)
-        check_env_specs(env)
+    @pytest.mark.parametrize("include_pgn", [False, True])
+    @pytest.mark.parametrize("include_fen", [False, True])
+    @pytest.mark.parametrize("stateful", [False, True])
+    def test_env(self, stateful, include_pgn, include_fen):
+        with pytest.raises(
+            RuntimeError, match="At least one state representation"
+        ) if not stateful and not include_pgn and not include_fen else contextlib.nullcontext():
+            env = ChessEnv(
+                stateful=stateful, include_pgn=include_pgn, include_fen=include_fen
+            )
+            check_env_specs(env)
 
-    def test_rollout(self, stateful):
-        env = ChessEnv(stateful=stateful)
-        env.rollout(5000)
+    def test_pgn_bijectivity(self):
+        np.random.seed(0)
+        pgn = ChessEnv._PGN_RESTART
+        board = ChessEnv._pgn_to_board(pgn)
+        pgn_prev = pgn
+        for _ in range(10):
+            moves = list(board.legal_moves)
+            move = np.random.choice(moves)
+            board.push(move)
+            pgn_move = ChessEnv._board_to_pgn(board)
+            assert pgn_move != pgn_prev
+            assert pgn_move == ChessEnv._board_to_pgn(ChessEnv._pgn_to_board(pgn_move))
+            assert pgn_move == ChessEnv._add_move_to_pgn(pgn_prev, move)
+            pgn_prev = pgn_move
+
+    def test_consistency(self):
+        env0_stateful = ChessEnv(stateful=True, include_pgn=True, include_fen=True)
+        env1_stateful = ChessEnv(stateful=True, include_pgn=False, include_fen=True)
+        env2_stateful = ChessEnv(stateful=True, include_pgn=True, include_fen=False)
+        env0_stateless = ChessEnv(stateful=False, include_pgn=True, include_fen=True)
+        env1_stateless = ChessEnv(stateful=False, include_pgn=False, include_fen=True)
+        env2_stateless = ChessEnv(stateful=False, include_pgn=True, include_fen=False)
+        torch.manual_seed(0)
+        r1_stateless = env1_stateless.rollout(50, break_when_any_done=False)
+        torch.manual_seed(0)
+        r1_stateful = env1_stateful.rollout(50, break_when_any_done=False)
+        torch.manual_seed(0)
+        r2_stateless = env2_stateless.rollout(50, break_when_any_done=False)
+        torch.manual_seed(0)
+        r2_stateful = env2_stateful.rollout(50, break_when_any_done=False)
+        torch.manual_seed(0)
+        r0_stateless = env0_stateless.rollout(50, break_when_any_done=False)
+        torch.manual_seed(0)
+        r0_stateful = env0_stateful.rollout(50, break_when_any_done=False)
+        assert (r0_stateless["action"] == r1_stateless["action"]).all()
+        assert (r0_stateless["action"] == r2_stateless["action"]).all()
+        assert (r0_stateless["action"] == r0_stateful["action"]).all()
+        assert (r1_stateless["action"] == r1_stateful["action"]).all()
+        assert (r2_stateless["action"] == r2_stateful["action"]).all()
+
+    @pytest.mark.parametrize(
+        "include_fen,include_pgn", [[True, False], [False, True], [True, True]]
+    )
+    @pytest.mark.parametrize("stateful", [False, True])
+    def test_rollout(self, stateful, include_pgn, include_fen):
+        torch.manual_seed(0)
+        env = ChessEnv(
+            stateful=stateful, include_pgn=include_pgn, include_fen=include_fen
+        )
+        r = env.rollout(500, break_when_any_done=False)
+        assert r.shape == (500,)
 
-    def test_reset_white_to_move(self, stateful):
-        env = ChessEnv(stateful=stateful)
+    @pytest.mark.parametrize(
+        "include_fen,include_pgn", [[True, False], [False, True], [True, True]]
+    )
+    @pytest.mark.parametrize("stateful", [False, True])
+    def test_reset_white_to_move(self, stateful, include_pgn, include_fen):
+        env = ChessEnv(
+            stateful=stateful, include_pgn=include_pgn, include_fen=include_fen
+        )
         fen = "5k2/4r3/8/8/8/1Q6/2K5/8 w - - 0 1"
         td = env.reset(TensorDict({"fen": fen}))
         assert td["fen"] == fen
+        if include_fen:
+            assert env.board.fen() == fen
         assert td["turn"] == env.lib.WHITE
         assert not td["done"]
 
-    def test_reset_black_to_move(self, stateful):
-        env = ChessEnv(stateful=stateful)
+    @pytest.mark.parametrize("include_fen,include_pgn", [[True, False], [True, True]])
+    @pytest.mark.parametrize("stateful", [False, True])
+    def test_reset_black_to_move(self, stateful, include_pgn, include_fen):
+        env = ChessEnv(
+            stateful=stateful, include_pgn=include_pgn, include_fen=include_fen
+        )
         fen = "5k2/4r3/8/8/8/1Q6/2K5/8 b - - 0 1"
         td = env.reset(TensorDict({"fen": fen}))
         assert td["fen"] == fen
+        assert env.board.fen() == fen
         assert td["turn"] == env.lib.BLACK
         assert not td["done"]
 
-    def test_reset_done_error(self, stateful):
-        env = ChessEnv(stateful=stateful)
+    @pytest.mark.parametrize("include_fen,include_pgn", [[True, False], [True, True]])
+    @pytest.mark.parametrize("stateful", [False, True])
+    def test_reset_done_error(self, stateful, include_pgn, include_fen):
+        env = ChessEnv(
+            stateful=stateful, include_pgn=include_pgn, include_fen=include_fen
+        )
         fen = "1R3k2/2R5/8/8/8/8/2K5/8 b - - 0 1"
         with pytest.raises(ValueError) as e_info:
             env.reset(TensorDict({"fen": fen}))
@@ -3425,12 +3501,19 @@ def test_reset_done_error(self, stateful):
     @pytest.mark.parametrize(
         "endstate", ["white win", "black win", "stalemate", "50 move", "insufficient"]
     )
-    def test_reward(self, stateful, reset_without_fen, endstate):
+    @pytest.mark.parametrize("include_pgn", [False, True])
+    @pytest.mark.parametrize("include_fen", [True])
+    @pytest.mark.parametrize("stateful", [False, True])
+    def test_reward(
+        self, stateful, reset_without_fen, endstate, include_pgn, include_fen
+    ):
         if stateful and reset_without_fen:
             # reset_without_fen is only used for stateless env
             return
 
-        env = ChessEnv(stateful=stateful)
+        env = ChessEnv(
+            stateful=stateful, include_pgn=include_pgn, include_fen=include_fen
+        )
 
         if endstate == "white win":
             fen = "5k2/2R5/8/8/8/1R6/2K5/8 w - - 0 1"
@@ -3443,28 +3526,28 @@ def test_reward(self, stateful, reset_without_fen, endstate):
             fen = "5k2/6r1/8/8/8/8/7r/1K6 b - - 0 1"
             expected_turn = env.lib.BLACK
             move = "Rg1#"
-            expected_reward = -1
+            expected_reward = 1
             expected_done = True
 
         elif endstate == "stalemate":
             fen = "5k2/6r1/8/8/8/8/7r/K7 b - - 0 1"
             expected_turn = env.lib.BLACK
             move = "Rb7"
-            expected_reward = 0
+            expected_reward = 0.5
             expected_done = True
 
         elif endstate == "insufficient":
             fen = "5k2/8/8/8/3r4/2K5/8/8 w - - 0 1"
             expected_turn = env.lib.WHITE
             move = "Kxd4"
-            expected_reward = 0
+            expected_reward = 0.5
             expected_done = True
 
         elif endstate == "50 move":
             fen = "5k2/8/1R6/8/6r1/2K5/8/8 b - - 99 123"
             expected_turn = env.lib.BLACK
             move = "Kf7"
-            expected_reward = 0
+            expected_reward = 0.5
             expected_done = True
 
         elif endstate == "not_done":
@@ -3483,8 +3566,7 @@ def test_reward(self, stateful, reset_without_fen, endstate):
             td = env.reset(TensorDict({"fen": fen}))
             assert td["turn"] == expected_turn
 
-        moves = env.get_legal_moves(None if stateful else td)
-        td["action"] = moves.index(move)
+        td["action"] = env._san_moves.index(move)
         td = env.step(td)["next"]
         assert td["done"] == expected_done
         assert td["reward"] == expected_reward

torchrl/envs/batched_envs.py

@@ -718,17 +718,13 @@ def _create_td(self) -> None:
             env_output_keys = set()
             env_obs_keys = set()
             for meta_data in self.meta_data:
-                env_obs_keys = env_obs_keys.union(
-                    key
-                    for key in meta_data.specs["output_spec"][
-                        "full_observation_spec"
-                    ].keys(True, True)
-                )
-                env_output_keys = env_output_keys.union(
-                    meta_data.specs["output_spec"]["full_observation_spec"].keys(
-                        True, True
-                    )
+                keys = meta_data.specs["output_spec"]["full_observation_spec"].keys(
+                    True, True
                 )
+                keys = list(keys)
+                env_obs_keys = env_obs_keys.union(keys)
+
+                env_output_keys = env_output_keys.union(keys)
             env_output_keys = env_output_keys.union(self.reward_keys + self.done_keys)
             self._env_obs_keys = sorted(env_obs_keys, key=_sort_keys)
             self._env_input_keys = sorted(env_input_keys, key=_sort_keys)
@@ -1003,7 +999,12 @@ def _reset(self, tensordict: TensorDictBase, **kwargs) -> TensorDictBase:
         for i, _env in enumerate(self._envs):
             if not needs_resetting[i]:
                 if out_tds is not None and tensordict is not None:
-                    out_tds[i] = tensordict[i].exclude(*self._envs[i].reset_keys)
+                    ftd = _env.observation_spec.zero()
+                    if self.device is None:
+                        ftd.clear_device_()
+                    else:
+                        ftd = ftd.to(self.device)
+                    out_tds[i] = ftd
                 continue
             if tensordict is not None:
                 tensordict_ = tensordict[i]

torchrl/envs/common.py

@@ -2505,11 +2505,26 @@ def reset(
         Returns:
             a tensordict (or the input tensordict, if any), modified in place with the resulting observations.
 
+        .. note:: `reset` should not be overwritten by :class:`~torchrl.envs.EnvBase` subclasses. The method to
+            modify is :meth:`~torchrl.envs.EnvBase._reset`.
+
         """
         if tensordict is not None:
             self._assert_tensordict_shape(tensordict)
 
-        tensordict_reset = self._reset(tensordict, **kwargs)
+        select_reset_only = kwargs.pop("select_reset_only", False)
+        if select_reset_only and tensordict is not None:
+            # When making rollouts with step_and_maybe_reset, it can happen that a tensordict has
+            # keys that are used by reset to optionally set the reset state (eg, the fen in chess). If that's the
+            # case and we don't throw them away here, reset will just be a no-op (put the env in the state reached
+            # during the previous step).
+            # Therefore, maybe_reset tells reset to temporarily hide the non-reset keys.
+            # To make step_and_maybe_reset handle custom reset states, some version of TensorDictPrimer should be used.
+            tensordict_reset = self._reset(
+                tensordict.select(*self.reset_keys, strict=False), **kwargs
+            )
+        else:
+            tensordict_reset = self._reset(tensordict, **kwargs)
         # We assume that this is done properly
         # if reset.device != self.device:
         #     reset = reset.to(self.device, non_blocking=True)
@@ -3292,7 +3307,7 @@ def maybe_reset(self, tensordict: TensorDictBase) -> TensorDictBase:
             else:
                 any_done = False
             if any_done:
-                tensordict._set_str(
+                tensordict = tensordict._set_str(
                     "_reset",
                     done.clone(),
                     validated=True,
@@ -3306,7 +3321,7 @@ def maybe_reset(self, tensordict: TensorDictBase) -> TensorDictBase:
                 key="_reset",
             )
         if any_done:
-            tensordict = self.reset(tensordict)
+            return self.reset(tensordict, select_reset_only=True)
         return tensordict
 
     def empty_cache(self):