Create engine/evaluate.py

engine/evaluate.py

@@ -0,0 +1,193 @@
+"""
+Nexus-Core Position Evaluator
+Pure ResNet-20 CNN with 12-channel input
+
+Research References:
+- He et al. (2016) - Deep Residual Learning for Image Recognition
+- Silver et al. (2017) - AlphaZero position evaluation
+"""
+
+import onnxruntime as ort
+import numpy as np
+import chess
+import logging
+from pathlib import Path
+from typing import Dict
+
+logger = logging.getLogger(__name__)
+
+
+class NexusCoreEvaluator:
+    """
+    Nexus-Core neural network evaluator
+    12-channel CNN input (simpler than Synapse-Base)
+    """
+    
+    # Stockfish piece values for material calculation
+    PIECE_VALUES = {
+        chess.PAWN: 100,
+        chess.KNIGHT: 320,
+        chess.BISHOP: 330,
+        chess.ROOK: 500,
+        chess.QUEEN: 900,
+        chess.KING: 0
+    }
+    
+    def __init__(self, model_path: str, num_threads: int = 2):
+        """Initialize evaluator with ONNX model"""
+        
+        self.model_path = Path(model_path)
+        if not self.model_path.exists():
+            raise FileNotFoundError(f"Model not found: {model_path}")
+        
+        # ONNX Runtime session
+        sess_options = ort.SessionOptions()
+        sess_options.intra_op_num_threads = num_threads
+        sess_options.inter_op_num_threads = num_threads
+        sess_options.execution_mode = ort.ExecutionMode.ORT_SEQUENTIAL
+        sess_options.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL
+        
+        logger.info(f"Loading Nexus-Core model from {model_path}...")
+        self.session = ort.InferenceSession(
+            str(self.model_path),
+            sess_options=sess_options,
+            providers=['CPUExecutionProvider']
+        )
+        
+        self.input_name = self.session.get_inputs()[0].name
+        self.output_name = self.session.get_outputs()[0].name
+        
+        logger.info(f"✅ Model loaded: {self.input_name} -> {self.output_name}")
+    
+    def fen_to_12_channel_tensor(self, board: chess.Board) -> np.ndarray:
+        """
+        Convert board to 12-channel tensor
+        Channels: 6 white pieces + 6 black pieces
+        
+        Args:
+            board: chess.Board object
+        
+        Returns:
+            numpy array of shape (1, 12, 8, 8)
+        """
+        tensor = np.zeros((1, 12, 8, 8), dtype=np.float32)
+        
+        piece_to_channel = {
+            chess.PAWN: 0, 
+            chess.KNIGHT: 1, 
+            chess.BISHOP: 2,
+            chess.ROOK: 3, 
+            chess.QUEEN: 4, 
+            chess.KING: 5
+        }
+        
+        # Fill piece positions
+        for square, piece in board.piece_map().items():
+            rank, file = divmod(square, 8)
+            channel = piece_to_channel[piece.piece_type]
+            
+            # White pieces: channels 0-5
+            # Black pieces: channels 6-11
+            if piece.color == chess.BLACK:
+                channel += 6
+            
+            tensor[0, channel, rank, file] = 1.0
+        
+        return tensor
+    
+    def evaluate_neural(self, board: chess.Board) -> float:
+        """
+        Neural network evaluation
+        
+        Args:
+            board: chess.Board object
+        
+        Returns:
+            Evaluation score (centipawns from white's perspective)
+        """
+        # Convert to tensor
+        input_tensor = self.fen_to_12_channel_tensor(board)
+        
+        # Run inference
+        outputs = self.session.run(
+            [self.output_name],
+            {self.input_name: input_tensor}
+        )
+        
+        # Extract value (tanh output in range [-1, 1])
+        raw_value = float(outputs[0][0][0])
+        
+        # Convert to centipawns (scale by 400)
+        centipawns = raw_value * 400.0
+        
+        return centipawns
+    
+    def evaluate_material(self, board: chess.Board) -> int:
+        """
+        Classical material evaluation
+        
+        Args:
+            board: chess.Board object
+        
+        Returns:
+            Material balance in centipawns
+        """
+        material = 0
+        
+        for piece_type in [chess.PAWN, chess.KNIGHT, chess.BISHOP, 
+                          chess.ROOK, chess.QUEEN]:
+            white_count = len(board.pieces(piece_type, chess.WHITE))
+            black_count = len(board.pieces(piece_type, chess.BLACK))
+            
+            material += (white_count - black_count) * self.PIECE_VALUES[piece_type]
+        
+        return material
+    
+    def evaluate_hybrid(self, board: chess.Board) -> float:
+        """
+        Hybrid evaluation: 90% neural + 10% material
+        
+        Args:
+            board: chess.Board object
+        
+        Returns:
+            Final evaluation score
+        """
+        # Neural evaluation (primary)
+        neural_eval = self.evaluate_neural(board)
+        
+        # Material evaluation (safety check)
+        material_eval = self.evaluate_material(board)
+        
+        # Blend: 90% neural, 10% material
+        hybrid_eval = 0.90 * neural_eval + 0.10 * material_eval
+        
+        # Flip for black's perspective
+        if board.turn == chess.BLACK:
+            hybrid_eval = -hybrid_eval
+        
+        return hybrid_eval
+    
+    def evaluate_mobility(self, board: chess.Board) -> int:
+        """
+        Mobility evaluation (number of legal moves)
+        
+        Args:
+            board: chess.Board object
+        
+        Returns:
+            Mobility score
+        """
+        current_mobility = board.legal_moves.count()
+        
+        # Flip turn to count opponent mobility
+        board.push(chess.Move.null())
+        opponent_mobility = board.legal_moves.count()
+        board.pop()
+        
+        # Mobility difference
+        return (current_mobility - opponent_mobility) * 5
+    
+    def get_model_size_mb(self) -> float:
+        """Get model size in MB"""
+        return self.model_path.stat().st_size / (1024 * 1024)