Upload ai/utils/obs_adapters.py with huggingface_hub

ai/utils/obs_adapters.py

@@ -0,0 +1,182 @@
+import numpy as np
+
+from engine.game.game_state import GameState
+
+
+class UnifiedObservationEncoder:
+    """
+    Translates current GameState into various historic observation formats.
+    """
+
+    @staticmethod
+    def encode(state: GameState, dim: int, player_idx: int = None) -> np.ndarray:
+        if player_idx is None:
+            player_idx = state.current_player
+
+        if dim == 8192:
+            return UnifiedObservationEncoder._encode_8192(state, player_idx)
+        elif dim == 2048:
+            return UnifiedObservationEncoder._encode_2048(state, player_idx)
+        elif dim == 320:
+            return UnifiedObservationEncoder._encode_320(state, player_idx)
+        elif dim == 128:
+            return UnifiedObservationEncoder._encode_128(state, player_idx)
+        else:
+            raise ValueError(f"Unsupported observation dimension: {dim}")
+
+    @staticmethod
+    def _encode_8192(state: GameState, player_idx: int) -> np.ndarray:
+        from ai.vector_env import VectorGameState as VGS
+        from ai.vector_env import encode_observations_vectorized
+
+        p = state.players[player_idx]
+        opp = state.players[1 - player_idx]
+
+        # Prepare arrays (mirroring VectorGameState layout)
+        batch_hand = np.zeros((1, 60), dtype=np.int32)
+        h_len = min(len(p.hand), 60)
+        for i in range(h_len):
+            batch_hand[0, i] = p.hand[i]
+
+        batch_stage = np.full((1, 3), -1, dtype=np.int32)
+        for i in range(3):
+            if i < len(p.stage):
+                batch_stage[0, i] = p.stage[i]
+
+        batch_energy_count = np.zeros((1, 3), dtype=np.int32)
+        # Assuming p.stage_energy_count exists (based on _encode_2048)
+        if hasattr(p, "stage_energy_count"):
+            for i in range(3):
+                if i < len(p.stage_energy_count):
+                    batch_energy_count[0, i] = p.stage_energy_count[i]
+
+        # NOTE: VectorEnv uses size 16 for tapped (Energy Tapping)
+        batch_tapped = np.zeros((1, 16), dtype=np.int32)
+        if hasattr(p, "tapped_members"):
+            for i in range(min(3, len(p.tapped_members))):
+                batch_tapped[0, i] = 1 if p.tapped_members[i] else 0
+
+        batch_scores = np.array([len(p.success_lives)], dtype=np.int32)
+        opp_scores = np.array([len(opp.success_lives)], dtype=np.int32)
+
+        opp_stage = np.full((1, 3), -1, dtype=np.int32)
+        for i in range(3):
+            if i < len(opp.stage):
+                opp_stage[0, i] = opp.stage[i]
+
+        opp_tapped = np.zeros((1, 16), dtype=np.int32)
+        if hasattr(opp, "tapped_members"):
+            for i in range(min(3, len(opp.tapped_members))):
+                opp_tapped[0, i] = 1 if opp.tapped_members[i] else 0
+
+        # Global Context - FIXED MAPPING
+        # Index 0 (SC): len(p.success_lives)
+        # Index 1 (OS): len(opp.success_lives)
+        # Index 2 (TR): len(p.discard)
+        # Index 3 (HD): len(p.hand)
+        # Index 4 (DI): len(opp.hand)
+        # Index 5 (EN): p.energy_count
+        # Index 6 (DK): len(p.main_deck)
+        # Index 7 (OT): len(opp.discard)
+        # Index 8 (PH): int(state.phase)
+        # Index 9 (OD): len(opp.main_deck)
+        g_ctx = np.zeros((1, 128), dtype=np.int32)
+        g_ctx[0, 0] = len(p.success_lives)
+        g_ctx[0, 1] = len(opp.success_lives)
+        g_ctx[0, 2] = len(p.discard)
+        g_ctx[0, 3] = len(p.hand)
+        g_ctx[0, 4] = len(opp.hand)
+        g_ctx[0, 5] = p.energy_count
+        g_ctx[0, 6] = len(p.main_deck)
+        g_ctx[0, 7] = len(opp.discard)
+        g_ctx[0, 8] = int(state.phase)
+        g_ctx[0, 9] = len(opp.main_deck)
+
+        # Additional Buffers for Updated Signature
+        batch_live = np.zeros((1, 50), dtype=np.int32)
+        batch_opp_history = np.zeros((1, 50), dtype=np.int32)
+
+        # Fill opp history with discard (Top Cards)
+        d_len = min(len(opp.discard), 50)
+        # VectorEnv assumes LIFO/Top-is-last.
+        for i in range(d_len):
+            batch_opp_history[0, i] = opp.discard[-(i + 1)]
+
+        # VGS Cache
+        if not hasattr(UnifiedObservationEncoder, "_vgs_cache"):
+            UnifiedObservationEncoder._vgs_cache = VGS(1)
+        vgs = UnifiedObservationEncoder._vgs_cache
+
+        # Output
+        obs = np.zeros((1, 8192), dtype=np.float32)
+
+        encode_observations_vectorized(
+            1,
+            batch_hand,
+            batch_stage,
+            batch_energy_count,
+            batch_tapped,
+            batch_scores,
+            opp_scores,
+            opp_stage,
+            opp_tapped,
+            vgs.card_stats,
+            g_ctx,
+            batch_live,
+            batch_opp_history,
+            state.turn_number,
+            obs,
+        )
+        return obs[0]
+
+    @staticmethod
+    def _encode_320(state: GameState, player_idx: int) -> np.ndarray:
+        # LEGACY 320 (First Speed-up Era)
+        # Replicates the encoding from ai/vector_env_legacy.py exactly.
+        # This era ONLY saw Self Stage and Self Score. Hand/Opp were 0.
+
+        obs = np.zeros(320, dtype=np.float32)
+        p = state.players[player_idx]
+        max_id_val = 2000.0  # Standard for VectorEnv
+
+        # Phase [5] = 1.0 (Mocking Main Phase index from Legacy VectorEnv)
+        obs[5] = 1.0
+        # Current Player [16]
+        obs[16] = 1.0
+
+        # Stage [168:204] (3 slots * 12 features)
+        # Note: Hand [36:168] remains 0.0 as in legacy training.
+        for i in range(3):
+            cid = p.stage[i]
+            base = 168 + i * 12
+            if cid >= 0:
+                obs[base] = 1.0  # Exist
+                obs[base + 1] = cid / max_id_val
+                # Legacy energy count was normalized by 5.0
+                obs[base + 11] = min(p.stage_energy_count[i] / 5.0, 1.0)
+
+        # Score [270] (Self Score normalized by 5.0 in legacy)
+        obs[270] = min(len(p.success_lives) / 5.0, 1.0)
+
+        return obs
+
+    @staticmethod
+    def _encode_128(state: GameState, player_idx: int) -> np.ndarray:
+        # 128-dim is the global_ctx vector
+        p = state.players[player_idx]
+        opp = state.players[1 - player_idx]
+
+        g_ctx = np.zeros(128, dtype=np.float32)
+        # Standard normalization from AlphaZero era
+        g_ctx[0] = len(p.success_lives) / 3.0
+        g_ctx[1] = len(opp.success_lives) / 3.0
+        g_ctx[2] = len(p.discard) / 50.0
+        g_ctx[3] = len(p.hand) / 50.0  # Normalized to deck size usually
+        g_ctx[5] = p.energy_count / 10.0
+        g_ctx[6] = len(p.main_deck) / 50.0
+
+        # Turn info
+        g_ctx[10] = state.turn_number / 20.0
+        g_ctx[11] = 1.0 if state.current_player == player_idx else 0.0
+
+        return g_ctx