tokenizer.py

"""
Custom Chess Tokenizer for the Chess Challenge.

This tokenizer tokenizes each move into 4 tokens using the extended UCI notation
from the Lichess dataset (e.g., WPe2e4, BNg8f6).

4-token scheme per move:
1) Side: W / B
2) Piece: P/N/B/R/Q/K
3) Source square: e2
4) Destination square + any suffix (capture/check/mate/promo/castling markers)
"""

from __future__ import annotations

import json
import os
import re
from typing import Dict, List, Optional, Tuple

from transformers import PreTrainedTokenizer


class ChessTokenizer(PreTrainedTokenizer):
    """
    A custom tokenizer for chess moves using extended UCI notation.

    It splits each move into 4 tokens and builds a vocabulary from the dataset
    so that training-time tokens have IDs.

    Example move:
      WPe2e4  -> ["W", "P", "e2", "e4"]
      BNg8f6  -> ["B", "N", "g8", "f6"]
      WPe7e8=Q -> ["W", "P", "e7", "e8=Q"]   (promotion kept in 4th token)
      WKe1g1(O) -> ["W", "K", "e1", "g1(O)"] (suffix kept in 4th token)
    """

    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]"

    # Regex to parse a standard extended-UCI move token:
    #  side (W/B), piece (P/N/B/R/Q/K), src square, dst square, optional suffix
    MOVE_RE = re.compile(r"^([WB])([PNBRQK])([a-h][1-8])([a-h][1-8])(.*)$")

    def __init__(
        self,
        vocab_file: Optional[str] = None,
        vocab: Optional[Dict[str, int]] = None,
        **kwargs,
    ):
        # 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 = dict(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:
            self._vocab = self._create_default_vocab()

        # 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,
        )

        # Safety: ensure special tokens exist
        for tok in [self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN, self.UNK_TOKEN]:
            if tok not in self._vocab:
                raise ValueError(f"Special token {tok} missing from vocab.")

    def _create_default_vocab(self) -> Dict[str, int]:
        special_tokens = [self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN, self.UNK_TOKEN]
        return {token: idx for idx, token in enumerate(special_tokens)}

    @classmethod
    def _move_to_4tokens(cls, move: str) -> List[str]:
        """
        Convert a move string into exactly 4 subtokens.
        If parsing fails, returns 4x UNK_TOKEN.
        """
        m = cls.MOVE_RE.match(move)
        if not m:
            return [cls.UNK_TOKEN, cls.UNK_TOKEN, cls.UNK_TOKEN, cls.UNK_TOKEN]
        side, piece, src, dst, suffix = m.groups()
        return [side, piece, src, dst + (suffix or "")]

    @classmethod
    def build_vocab_from_iterator(
        cls,
        iterator,
        min_frequency: int = 1,
    ) -> "ChessTokenizer":
        """
        Build a tokenizer vocabulary from an iterator of game strings.

        IMPORTANT: since we tokenize each move into 4 tokens, we must count
        those subtokens here (not the raw full move strings).
        """
        from collections import Counter

        token_counts = Counter()

        for game in iterator:
            for move in str(game).strip().split():
                subtokens = cls._move_to_4tokens(move)
                token_counts.update(subtokens)

        # Filter by frequency
        tokens = [tok for tok, count in token_counts.items() if count >= min_frequency]

        # Sort for reproducibility
        tokens = sorted(tokens)

        # Build vocab with special tokens first
        special_tokens = [cls.PAD_TOKEN, cls.BOS_TOKEN, cls.EOS_TOKEN, cls.UNK_TOKEN]
        tokens = [t for t in tokens if t not in set(special_tokens)]
        vocab = {token: idx for idx, token in enumerate(special_tokens + tokens)}

        return cls(vocab=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.
        """
        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)

    @property
    def vocab_size(self) -> int:
        return len(self._vocab)

    def get_vocab(self) -> Dict[str, int]:
        return dict(self._vocab)

    def _tokenize(self, text: str) -> List[str]:
        """
        Tokenize a space-separated game string into a flat list of subtokens,
        using exactly 4 tokens per move.
        """
        out: List[str] = []
        for move in str(text).strip().split():
            out.extend(self._move_to_4tokens(move))
        return out

    def _convert_token_to_id(self, token: str) -> int:
        # Always fall back to unk_token_id (never silently to PAD)
        return self._vocab.get(token, self.unk_token_id)

    def _convert_id_to_token(self, index: int) -> str:
        return self._ids_to_tokens.get(index, self.unk_token)

    def convert_tokens_to_string(self, tokens: List[str]) -> str:
        """
        Convert tokens back to a string (space-separated).
        We drop PAD/BOS/EOS; keep UNK for debugging.
        """
        drop = {self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN}
        return " ".join(t for t in tokens if t not in drop)

    def save_vocabulary(
        self,
        save_directory: str,
        filename_prefix: Optional[str] = None,
    ) -> Tuple[str]:
        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.

    NOTE: This counts the 4-subtoken scheme (not whole moves).
    """
    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:
        for move in str(example[column]).strip().split():
            token_counts.update(ChessTokenizer._move_to_4tokens(move))

    return dict(token_counts)