Upload ai/utils/verify_parity.py with huggingface_hub

ai/utils/verify_parity.py

@@ -0,0 +1,72 @@
+import os
+import sys
+
+import numpy as np
+import torch
+
+# Add project root to path
+sys.path.append(os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__)))))
+
+import engine_rust
+
+from ai.models.training_config import POLICY_SIZE
+from ai.training.train import AlphaNet
+
+
+def verify_parity():
+    # 1. Setup Models
+    model_path_pt = "ai/models/alphanet_best.pt"
+    model_path_onnx = "ai/models/alphanet.onnx"
+
+    device = torch.device("cpu")
+    checkpoint = torch.load(model_path_pt, map_location=device)
+    state_dict = checkpoint["model_state"] if "model_state" in checkpoint else checkpoint
+
+    model_pt = AlphaNet(policy_size=POLICY_SIZE)
+    model_pt.load_state_dict(state_dict)
+    model_pt.eval()
+
+    nmcts_rust = engine_rust.PyNeuralMCTS(model_path_onnx)
+
+    # 2. Setup Game
+    with open("engine/data/cards_compiled.json", "r", encoding="utf-8") as f:
+        db_content = f.read()
+    db = engine_rust.PyCardDatabase(db_content)
+    game = engine_rust.PyGameState(db)
+
+    # Simple init
+    p0_deck = [124, 127] * 20
+    p1_deck = [124, 127] * 20
+    p0_lives = [1024, 1025, 1027]
+    p1_lives = [1024, 1025, 1027]
+    game.initialize_game(p0_deck, p1_deck, [20000] * 10, [20000] * 10, p0_lives, p1_lives)
+
+    # 3. Compare Encoding
+    obs_rust = np.array(game.get_observation(), dtype=np.float32)
+
+    # 4. Compare Inference
+    with torch.no_grad():
+        logits_pt, val_pt = model_pt(torch.from_numpy(obs_rust).unsqueeze(0))
+        logits_pt = logits_pt.numpy()[0]
+        val_pt = val_pt.numpy()[0][0]
+
+    print(f"Observation Size: {len(obs_rust)}")
+    print(f"Value (Python): {val_pt:.6f}")
+
+    # Note: We need a way to get raw evaluate results from Rust for deep parity
+    # But for now, we'll check suggestions as a proxy
+    num_sims = 100
+    suggestions = nmcts_rust.get_suggestions(game, num_sims)
+
+    print("\nTop Python Actions (Logits):")
+    top_pt = np.argsort(logits_pt)[::-1][:5]
+    for i in top_pt:
+        print(f"  Action {i:4d}: {logits_pt[i]:.4f}")
+
+    print("\nRust Suggestions (1 simulation):")
+    for action, score, visits in suggestions[:5]:
+        print(f"  Action {action:4d}: Score {score:.4f}, Visits {visits}")
+
+
+if __name__ == "__main__":
+    verify_parity()