Add Tokenizer

tokenizer.py

@@ -0,0 +1,340 @@
+"""
+Custom Chess Tokenizer for the Chess Challenge.
+
+This tokenizer treats each move as a single token using the extended UCI notation
+from the Lichess dataset (e.g., WPe2e4, BNg8f6).
+
+The dataset format uses:
+- W/B prefix for White/Black
+- Piece letter: P=Pawn, N=Knight, B=Bishop, R=Rook, Q=Queen, K=King
+- Source and destination squares (e.g., e2e4)
+- Special suffixes: (x)=capture, (+)=check, (+*)=checkmate, (o)/(O)=castling
+"""
+
+from __future__ import annotations
+
+import json
+import os
+from pathlib import Path
+from typing import Dict, List, Optional
+
+from transformers import PreTrainedTokenizer
+
+
+class ChessTokenizer(PreTrainedTokenizer):
+    """
+    A custom tokenizer for chess moves using extended UCI notation.
+    
+    This tokenizer maps each possible chess move to a unique token ID.
+    The vocabulary is built from the training dataset to ensure all moves
+    encountered during training have a corresponding token.
+    
+    Example:
+        >>> tokenizer = ChessTokenizer()
+        >>> tokenizer.encode("WPe2e4 BPe7e5")
+        [1, 42, 87, 2]  # [BOS, e2e4, e7e5, EOS]
+    """
+    
+    model_input_names = ["input_ids", "attention_mask"]
+    vocab_files_names = {"vocab_file": "vocab.json"}
+    
+    # Special tokens
+    PAD_TOKEN = "[PAD]"
+    BOS_TOKEN = "[BOS]"
+    EOS_TOKEN = "[EOS]"
+    UNK_TOKEN = "[UNK]"
+    
+    def __init__(
+        self,
+        vocab_file: Optional[str] = None,
+        vocab: Optional[Dict[str, int]] = None,
+        **kwargs,
+    ):
+        """
+        Initialize the chess tokenizer.
+        
+        Args:
+            vocab_file: Path to a JSON file containing the vocabulary mapping.
+            vocab: Dictionary mapping tokens to IDs (alternative to vocab_file).
+            **kwargs: Additional arguments passed to PreTrainedTokenizer.
+        """
+        # Initialize special tokens
+        self._pad_token = self.PAD_TOKEN
+        self._bos_token = self.BOS_TOKEN
+        self._eos_token = self.EOS_TOKEN
+        self._unk_token = self.UNK_TOKEN
+
+        # Remove any duplicate special-token entries passed through kwargs
+        # to avoid "multiple values for keyword" errors when loading from disk.
+        kwargs.pop("pad_token", None)
+        kwargs.pop("bos_token", None)
+        kwargs.pop("eos_token", None)
+        kwargs.pop("unk_token", None)
+        
+        # Load or create vocabulary
+        if vocab is not None:
+            self._vocab = vocab
+        elif vocab_file is not None and os.path.exists(vocab_file):
+            with open(vocab_file, "r", encoding="utf-8") as f:
+                self._vocab = json.load(f)
+        else:
+            # Create a minimal vocabulary with just special tokens
+            # The full vocabulary should be built from the dataset
+            self._vocab = self._create_default_vocab()
+        
+        # Create reverse mapping
+        self._ids_to_tokens = {v: k for k, v in self._vocab.items()}
+        
+        # Call parent init AFTER setting up vocab
+        super().__init__(
+            pad_token=self._pad_token,
+            bos_token=self._bos_token,
+            eos_token=self._eos_token,
+            unk_token=self._unk_token,
+            **kwargs,
+        )
+    
+    def _create_default_vocab(self) -> Dict[str, int]:
+        """
+        Create a fixed structured vocabulary (no dataset-dependent move tokens).
+
+        Tokens:
+        - Special: [PAD], [BOS], [EOS], [UNK]
+        - Color: [W], [B]
+        - Pieces: [P], [N], [BISHOP], [R], [Q], [K]
+        - Squares: [a1]..[h8]
+        - Suffixes: [x], [+], [#]
+        - Castling: [O-O], [O-O-O]
+        - Promotions: [prom_Q], [prom_R], [prom_B], [prom_N]
+        - Move separator: [MOVE_END]
+        """
+        special = [self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN, self.UNK_TOKEN]
+        colors = ["[W]", "[B]"]
+        pieces = ["[P]", "[N]", "[BISHOP]", "[R]", "[Q]", "[K]"]
+
+        files = "abcdefgh"
+        ranks = "12345678"
+        squares = [f"[{f}{r}]" for r in ranks for f in files]  # a1..h8
+
+        suffixes = ["[x]", "[+]", "[#]"]
+        castling = ["[O-O]", "[O-O-O]"]
+        promotions = ["[prom_Q]", "[prom_R]", "[prom_B]", "[prom_N]"]
+        move_end = ["[MOVE_END]"]
+
+        tokens = special + colors + pieces + squares + suffixes + castling + promotions + move_end
+        return {tok: i for i, tok in enumerate(tokens)}
+
+    @classmethod
+    def build_vocab_from_iterator(cls, iterator, min_frequency: int = 1) -> "ChessTokenizer":
+        # Structured tokenizer uses a fixed vocab; iterator is unused.
+        return cls(vocab=cls().get_vocab())
+    
+    # @classmethod
+    # def build_vocab_from_dataset(
+    #     cls,
+    #     dataset_name: str = "dlouapre/lichess_2025-01_1M",
+    #     split: str = "train",
+    #     column: str = "text",
+    #     min_frequency: int = 500,
+    #     max_samples: Optional[int] = 100000,
+    # ) -> "ChessTokenizer":
+    #     """
+    #     Build a tokenizer vocabulary from a Hugging Face dataset.
+        
+    #     Args:
+    #         dataset_name: Name of the dataset on Hugging Face Hub.
+    #         split: Dataset split to use.
+    #         column: Column containing the game strings.
+    #         min_frequency: Minimum frequency for a token to be included (default: 500).
+    #         max_samples: Maximum number of samples to process (default: 100k).
+        
+    #     Returns:
+    #         A ChessTokenizer with the built vocabulary.
+    #     """
+    #     from datasets import load_dataset
+        
+    #     dataset = load_dataset(dataset_name, split=split)
+        
+    #     if max_samples is not None:
+    #         dataset = dataset.select(range(min(max_samples, len(dataset))))
+        
+    #     def game_iterator():
+    #         for example in dataset:
+    #             yield example[column]
+        
+    #     return cls.build_vocab_from_iterator(game_iterator(), min_frequency=min_frequency)
+
+    @classmethod
+    def build_vocab_from_dataset(cls,dataset_name: str = "dlouapre/lichess_2025-01_1M",split: str = "train",column: str = "text",min_frequency: int = 500,max_samples: Optional[int] = 100000,) -> "ChessTokenizer":
+        # Structured tokenizer uses a fixed vocab; dataset params are unused.
+        return cls(vocab=cls().get_vocab())
+    
+    @property
+    def vocab_size(self) -> int:
+        """Return the size of the vocabulary."""
+        return len(self._vocab)
+    
+    def get_vocab(self) -> Dict[str, int]:
+        """Return the vocabulary as a dictionary."""
+        return dict(self._vocab)
+    
+    def _move_to_tokens(self, move: str) -> List[str]:
+        """
+        Convert one extended-UCI move string to structured tokens.
+
+        Examples:
+        "WPe2e4" -> ["[W]","[P]","[e2]","[e4]"]
+        "WBb5c6(x+)" -> ["[W]","[BISHOP]","[b5]","[c6]","[x]","[+]"]
+        "BKe8g8(o)" -> ["[B]","[O-O]"]
+        "WPa7a8(Q)" -> ["[W]","[P]","[a7]","[a8]","[prom_Q]"]
+        """
+        toks: List[str] = []
+
+        if not move:
+            return [self.UNK_TOKEN]
+
+        # Color
+        color = move[0]
+        toks.append("[W]" if color == "W" else "[B]")
+
+        # Basic fields
+        # move[1] is piece letter in dataset (P,N,B,R,Q,K)
+        piece_char = move[1] if len(move) > 1 else ""
+        piece_map = {"P": "[P]", "N": "[N]", "B": "[BISHOP]", "R": "[R]", "Q": "[Q]", "K": "[K]"}
+        toks.append(piece_map.get(piece_char, self.UNK_TOKEN))
+
+        # Source and destination squares assumed at positions 2:4 and 4:6
+        # e.g. WPe2e4 -> from=e2 to=e4
+        if len(move) >= 6:
+            from_sq = move[2:4]
+            to_sq = move[4:6]
+            toks.append(f"[{from_sq}]")
+            toks.append(f"[{to_sq}]")
+        else:
+            # malformed
+            toks.append(self.UNK_TOKEN)
+            toks.append(self.UNK_TOKEN)
+
+        # --- Castling ---
+        # Dataset mentions (o)/(O)=castling, sometimes attached to king moves.
+        # We'll map based on king destination:
+        if "(o)" in move or "(O)" in move:
+            # King ends on g-file => O-O ; on c-file => O-O-O
+            if len(move) >= 6:
+                to_sq = move[4:6]
+                if to_sq[0] == "g":
+                    return [toks[0], "[O-O]"]
+                if to_sq[0] == "c":
+                    return [toks[0], "[O-O-O]"]
+
+        # --- Promotion ---
+        if "(Q)" in move:
+            toks.append("[prom_Q]")
+        elif "(R)" in move:
+            toks.append("[prom_R]")
+        elif "(B)" in move:
+            toks.append("[prom_B]")
+        elif "(N)" in move:
+            toks.append("[prom_N]")
+
+        # --- Capture / check / mate ---
+        # Capture patterns: "(x)" "(x+)" "(x+*)" etc.
+        if "(x" in move:
+            toks.append("[x]")
+
+        # Checkmate sometimes written (+*) or similar
+        if "(+*)" in move:
+            toks.append("[#]")
+        elif "(+)" in move or "(x+)" in move:
+            toks.append("[+]")
+
+        return toks
+
+    def _tokenize(self, text: str) -> List[str]:
+        """
+        Tokenize a game string into structured tokens.
+
+        Each move becomes:
+        [W]/[B], [PIECE], [from], [to], optional flags, then [MOVE_END]
+        """
+        moves = text.strip().split()
+        out: List[str] = []
+        for mv in moves:
+            out.extend(self._move_to_tokens(mv))
+            out.append("[MOVE_END]")
+        return out
+    
+    def _convert_token_to_id(self, token: str) -> int:
+        """Convert a token to its ID."""
+        return self._vocab.get(token, self._vocab.get(self.UNK_TOKEN, 0))
+    
+    def _convert_id_to_token(self, index: int) -> str:
+        """Convert an ID to its token."""
+        return self._ids_to_tokens.get(index, self.UNK_TOKEN)
+    
+    def convert_tokens_to_string(self, tokens: List[str]) -> str:
+        special = {self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN, self.UNK_TOKEN}
+        return " ".join(t for t in tokens if (t not in special and t != "[MOVE_END]"))
+    
+    def save_vocabulary(
+        self,
+        save_directory: str,
+        filename_prefix: Optional[str] = None,
+    ) -> tuple:
+        """
+        Save the vocabulary to a JSON file.
+        
+        Args:
+            save_directory: Directory to save the vocabulary.
+            filename_prefix: Optional prefix for the filename.
+        
+        Returns:
+            Tuple containing the path to the saved vocabulary file.
+        """
+        if not os.path.isdir(save_directory):
+            os.makedirs(save_directory, exist_ok=True)
+        
+        vocab_file = os.path.join(
+            save_directory,
+            (filename_prefix + "-" if filename_prefix else "") + "vocab.json",
+        )
+        
+        with open(vocab_file, "w", encoding="utf-8") as f:
+            json.dump(self._vocab, f, ensure_ascii=False, indent=2)
+        
+        return (vocab_file,)
+
+
+def count_vocab_from_dataset(
+    dataset_name: str = "dlouapre/lichess_2025-01_1M",
+    split: str = "train",
+    column: str = "text",
+    max_samples: Optional[int] = 10000,
+) -> Dict[str, int]:
+    """
+    Count token frequencies in a dataset (useful for vocabulary analysis).
+    
+    Args:
+        dataset_name: Name of the dataset on Hugging Face Hub.
+        split: Dataset split to use.
+        column: Column containing the game strings.
+        max_samples: Maximum number of samples to process.
+    
+    Returns:
+        Dictionary mapping tokens to their frequencies.
+    """
+    from collections import Counter
+    from datasets import load_dataset
+    
+    dataset = load_dataset(dataset_name, split=split)
+    
+    if max_samples is not None:
+        dataset = dataset.select(range(min(max_samples, len(dataset))))
+    
+    token_counts = Counter()
+    
+    for example in dataset:
+        moves = example[column].strip().split()
+        token_counts.update(moves)
+    
+    return dict(token_counts)