File size: 6,926 Bytes
f7b47a4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
import os
import chess
import torch
import numpy as np
from tqdm import tqdm
import random

SHARD_SIZE = 250_000
VAL_FRACTION = 0.1
MOVE_BUCKETS = [
    (0, 10), (10, 20), (20, 30),
    (30, 40), (40, 50), (50, 70), (70, 10_000)
]
BUCKET_NAMES = ["1-10", "11-20", "21-30", "31-40", "41-50", "51-70", "71+"]
ALL_BUCKETS = BUCKET_NAMES + ["mating"]
NUM_INPUT_PLANES = 17

def fen_to_tensor_perspective(fen: str) -> torch.Tensor:
    board = chess.Board(fen)
    tensor = np.zeros((NUM_INPUT_PLANES, 8, 8), dtype=np.float32)

    if board.turn == chess.WHITE:
        piece_to_plane = {
            'P': 0, 'N': 1, 'B': 2, 'R': 3, 'Q': 4, 'K': 5,
            'p': 6, 'n': 7, 'b': 8, 'r': 9, 'q': 10, 'k': 11
        }
        f_castling, e_castling = (12, 13), (14, 15)
        flip_board = False
    else:
        piece_to_plane = {
            'P': 6, 'N': 7, 'B': 8, 'R': 9, 'Q': 10, 'K': 11,
            'p': 0, 'n': 1, 'b': 2, 'r': 3, 'q': 4, 'k': 5
        }
        f_castling, e_castling = (12, 13), (14, 15)
        flip_board = True

    rows = fen.split()[0].split('/')
    for rank_index, row in enumerate(rows):
        file_index = 0
        for char in row:
            if char.isdigit():
                file_index += int(char)
            else:
                if char in piece_to_plane:
                    tensor[piece_to_plane[char], rank_index, file_index] = 1.0
                file_index += 1

    castling_str = fen.split()[2]
    if castling_str != "-":
        if ('K' in castling_str and board.turn == chess.WHITE) or ('k' in castling_str and board.turn == chess.BLACK):
            tensor[f_castling[0], :, :] = 1.0
        if ('Q' in castling_str and board.turn == chess.WHITE) or ('q' in castling_str and board.turn == chess.BLACK):
            tensor[f_castling[1], :, :] = 1.0
        if ('K' in castling_str and board.turn == chess.BLACK) or ('k' in castling_str and board.turn == chess.WHITE):
            tensor[e_castling[0], :, :] = 1.0
        if ('Q' in castling_str and board.turn == chess.BLACK) or ('q' in castling_str and board.turn == chess.WHITE):
            tensor[e_castling[1], :, :] = 1.0

    if board.ep_square is not None:
        rank = 7 - (board.ep_square // 8)
        file = board.ep_square % 8
        tensor[16, rank, file] = 1.0

    if flip_board:
        tensor = np.flip(tensor, axis=(1, 2)).copy()

    return torch.tensor(tensor, dtype=torch.float32)

def load_game_data(fen_path, move_path):
    result = None
    fens, moves = [], []
    with open(fen_path, "r") as f_fen, open(move_path, "r") as f_move:
        first_fen_line = f_fen.readline().strip()
        first_move_line = f_move.readline().strip()
        if first_fen_line.startswith("# Result") and first_move_line.startswith("# Result"):
            result = first_fen_line.split()[-1]
        else:
            f_fen.seek(0)
            f_move.seek(0)
        for fen_line, move_line in zip(f_fen, f_move):
            fen, move = fen_line.strip(), move_line.strip()
            if fen and move:
                fens.append(fen)
                moves.append(move)
    assert len(fens) == len(moves)
    return result, fens, moves

def get_buckets_for_position(idx, total_moves):
    buckets = []
    if total_moves - idx <= 10:
        buckets.append("mating")
    for (start, end), name in zip(MOVE_BUCKETS, BUCKET_NAMES):
        if start <= idx < end:
            buckets.append(name)
            break
    return buckets

def generate_samples(fen_dir, move_dir, val_files_set):
    game_files = sorted(f for f in os.listdir(fen_dir) if f.endswith(".txt"))
    for filename in tqdm(game_files, desc="Streaming"):
        fen_path = os.path.join(fen_dir, filename)
        move_path = os.path.join(move_dir, filename.replace(".txt", ".moves.txt"))
        try:
            result, fens, moves = load_game_data(fen_path, move_path)
        except Exception as e:
            print(f"⚠️ {filename}: {e}")
            continue
        if result is None:
            continue
        winner = "white" if result == "1-0" else "black" if result == "0-1" else "draw"
        is_val = filename in val_files_set
        for idx, (fen, move) in enumerate(zip(fens, moves)):
            if move == "none":
                continue
            try:
                tensor = fen_to_tensor_perspective(fen)
            except Exception as e:
                print(f"⚠️ Tensor conversion error in {filename}: {e}")
                continue
            bucket = get_buckets_for_position(idx, len(fens))[0]
            side = "white" if fen.split()[1] == "w" else "black"
            if winner == "draw":
                outcome = "draw"
            else:
                outcome = "win" if side == winner else "loss"
            yield is_val, (tensor, move, outcome, bucket, side)

class ShardWriter:
    def __init__(self, output_dir, prefix, shard_size=250_000):
        self.output_dir = output_dir
        self.prefix = prefix
        self.shard_size = shard_size
        self.shard_count = 0
        self.buffer = []

    def add(self, sample):
        self.buffer.append(sample)
        if len(self.buffer) >= self.shard_size:
            self.flush()

    def flush(self):
        if not self.buffer:
            return
        tensors, moves, values, buckets, sides = zip(*self.buffer)
        out_path = os.path.join(self.output_dir, f"{self.prefix}_{self.shard_count:03d}.pt")
        torch.save({
            "inputs": torch.stack(tensors),
            "policy": list(moves),
            "value": list(values),
            "buckets": list(buckets),
            "side_to_move": list(sides)
        }, out_path)
        print(f"💾 Saved {out_path} ({len(self.buffer)} samples)")
        self.shard_count += 1
        self.buffer.clear()

    def finalize(self):
        self.flush()

def merge_to_shards(fen_dir, move_dir, output_dir, max_games=None):
    os.makedirs(output_dir, exist_ok=True)
    game_files = sorted(f for f in os.listdir(fen_dir) if f.endswith(".txt"))
    if max_games:
        game_files = game_files[:max_games]

    random.shuffle(game_files)
    split = int(len(game_files) * (1 - VAL_FRACTION))
    train_files, val_files = game_files[:split], game_files[split:]
    val_files_set = set(val_files)

    print(f"Splitting games → Train: {len(train_files)} | Val: {len(val_files)}")

    train_writer = ShardWriter(output_dir, prefix="shard", shard_size=SHARD_SIZE)
    val_writer = ShardWriter(output_dir, prefix="val", shard_size=SHARD_SIZE)

    for is_val, sample in generate_samples(fen_dir, move_dir, val_files_set):
        (val_writer if is_val else train_writer).add(sample)

    train_writer.finalize()
    val_writer.finalize()

    print(f"\n✅ Done. Train shards: {train_writer.shard_count}, Val shards: {val_writer.shard_count}")

if __name__ == "__main__":
    merge_to_shards(fen_dir="fens/", move_dir="moves/", output_dir="shards/", max_games=None)