tokenizer.py

"""
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
"""
Custom Chess Tokenizer - Normalized Version
"""
import re

# Regex pour extraire case départ, arrivée et promotion
MOVE_RE = re.compile(r"([a-h][1-8])([a-h][1-8])")
PROMO_RE = re.compile(r"=([NBRQ])")

def normalize_move(tok: str) -> str:
    """Transforme 'WPe2e4(x)' en 'WPe2e4' pour réduire le vocabulaire."""
    # 1. Garder les infos de base
    m = MOVE_RE.search(tok)
    if not m:
        return tok # Fallback (sera probablement UNK)
    
    fr, to = m.group(1), m.group(2)
    
    # 2. Gérer la promotion
    promo = ""
    pm = PROMO_RE.search(tok)
    if pm:
        promo = "=" + pm.group(1)
        
    # 3. Reconstruire le token standardisé
    # On garde le préfixe WP/BN (chars 0 et 1) pour garder l'info couleur/pièce
    # mais on supprime les suffixes (x), (+), etc.
    prefix = tok[:2] if len(tok) >= 2 else "WP"
    return f"{prefix}{fr}{to}{promo}"

class ChessTokenizer(PreTrainedTokenizer):
    model_input_names = ["input_ids", "attention_mask"]
    
    PAD_TOKEN = "[PAD]"
    BOS_TOKEN = "[BOS]"
    EOS_TOKEN = "[EOS]"
    UNK_TOKEN = "[UNK]"
    
    def __init__(self, vocab_file=None, vocab=None, **kwargs):
        self._pad_token = self.PAD_TOKEN
        self._bos_token = self.BOS_TOKEN
        self._eos_token = self.EOS_TOKEN
        self._unk_token = self.UNK_TOKEN
        
        # Nettoyage kwargs
        for t in ["pad_token", "bos_token", "eos_token", "unk_token"]:
            kwargs.pop(t, None)
            
        if vocab:
            self._vocab = vocab
        elif vocab_file:
            with open(vocab_file, "r", encoding="utf-8") as f:
                self._vocab = json.load(f)
        else:
            self._vocab = {t: i for i, t in enumerate([self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN, self.UNK_TOKEN])}
            
        self._ids_to_tokens = {v: k for k, v in self._vocab.items()}
        super().__init__(pad_token=self.PAD_TOKEN, bos_token=self.BOS_TOKEN, eos_token=self.EOS_TOKEN, unk_token=self.UNK_TOKEN, **kwargs)

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

    def get_vocab(self):
        return dict(self._vocab)

    def _tokenize(self, text):
        # C'est ICI que la magie opère : on normalise à la volée
        return [normalize_move(t) for t in text.strip().split()]

    def _convert_token_to_id(self, token):
        return self._vocab.get(token, self._vocab.get(self.UNK_TOKEN))

    def _convert_id_to_token(self, index):
        return self._ids_to_tokens.get(index, self.UNK_TOKEN)

    def convert_tokens_to_string(self, tokens):
        return " ".join(t for t in tokens if t not in [self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN, self.UNK_TOKEN])
    
    def save_vocabulary(self, save_directory, filename_prefix=None):
        if not os.path.exists(save_directory):
            os.makedirs(save_directory)
        path = os.path.join(save_directory, (filename_prefix + "-" if filename_prefix else "") + "vocab.json")
        with open(path, "w") as f:
            json.dump(self._vocab, f, indent=2)
        return (path,)

    @classmethod
    def build_vocab_from_dataset(cls, dataset_name, min_frequency=2, max_vocab_size=1200, **kwargs):
        """Construit un vocabulaire compact et dense."""
        from datasets import load_dataset
        from collections import Counter
        
        # On charge en streaming pour aller vite
        ds = load_dataset(dataset_name, split="train", streaming=True)
        ds = ds.take(50000) # 50k parties suffisent pour voir tous les coups possibles
        
        counter = Counter()
        for ex in ds:
            # On normalise avant de compter !
            moves = [normalize_move(t) for t in ex["text"].split()]
            counter.update(moves)
            
        # On garde les tokens spéciaux + les N plus fréquents
        special = [cls.PAD_TOKEN, cls.BOS_TOKEN, cls.EOS_TOKEN, cls.UNK_TOKEN]
        most_common = counter.most_common(max_vocab_size - len(special))
        
        vocab = {t: i for i, t in enumerate(special + [t for t, c in most_common])}
        return cls(vocab=vocab)