safe_text_image_pairs_ja_en_txt_per_image.py

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
日本語文字画像 + 英語説明文（CC0）
- 単語は外部ファイルから読み込み（UTF-8, 1行1語）
- 白背景・黒文字
- 説明文に色指定
- PNG+TXTペアを train/ に出力
- 最後に train.tar.gz にまとめる
"""

import csv, random, argparse, hashlib, datetime, tarfile, shutil
from pathlib import Path
from dataclasses import dataclass
from PIL import Image, ImageDraw, ImageFont
import numpy as np

# --------------------
# 定数
# --------------------
DEFAULT_FONTS_DIR = Path("./fonts")
TRAIN_DIR         = Path("./train")
LINE_SPACING      = 1.25

@dataclass
class Example:
    jp_text: str
    en_desc: str

# --------------------
# 外部ファイルから単語読み込み
# --------------------
def load_words(file_path: Path):
    if not file_path.exists():
        raise FileNotFoundError(f"単語ファイルが見つかりません: {file_path}")
    with open(file_path, "r", encoding="utf-8") as f:
        words = [line.strip() for line in f if line.strip()]
    if not words:
        raise ValueError(f"単語ファイルが空です: {file_path}")
    return words

# --------------------
# データ生成
# --------------------
def gen_examples(n:int, seed:int, words:list[str]):
    random.seed(seed)
    exs = []
    for _ in range(n):
        jp = random.choice(words)
        desc = f'This image is saying "{jp}". The background is white. The letter is black.'
        exs.append(Example(jp_text=jp, en_desc=desc))
    return exs

# --------------------
# フォント
# --------------------
def list_fonts(font_dir:Path):
    fonts = [p for p in font_dir.glob("*") if p.suffix.lower() in (".ttf",".otf",".ttc",".otc")]
    if not fonts:
        raise FileNotFoundError(f"フォントが見つかりません: {font_dir} にOFL/PDの日本語フォントを置いてください")
    return fonts

# --------------------
# 描画（背景：白、文字：黒）
# --------------------
def draw_horizontal(text, font_path:Path, size, max_font_size, min_font_size, margin_px):
    W,H = size
    img = Image.new("RGB", (W,H), (255,255,255))  # 白背景
    draw = ImageDraw.Draw(img)
    for fs in range(max_font_size, min_font_size-1, -2):
        font = ImageFont.truetype(str(font_path), fs)
        bbox = draw.textbbox((0,0), text, font=font)
        w, h = bbox[2] - bbox[0], bbox[3] - bbox[1]
        if w <= W - 2*margin_px and h <= H - 2*margin_px:
            break
    x = (W - w)//2
    y = (H - h)//2
    draw.text((x, y), text, font=font, fill=(0,0,0))  # 黒
    return img

def draw_vertical(text, font_path:Path, size, max_font_size, min_font_size, margin_px):
    W,H = size
    img = Image.new("RGB", (W,H), (255,255,255))  # 白背景
    draw = ImageDraw.Draw(img)
    for fs in range(max_font_size, min_font_size-1, -2):
        font = ImageFont.truetype(str(font_path), fs)
        line_h = font.getbbox("Hg")[3] - font.getbbox("Hg")[1]
        step = int(line_h * LINE_SPACING)
        total_h = len(text) * step if text else step
        if text:
            widths = []
            for c in text:
                cb = draw.textbbox((0,0), c, font=font)
                widths.append(cb[2] - cb[0])
            col_w = max(widths)
        else:
            cb = draw.textbbox((0,0), "あ", font=font)
            col_w = cb[2] - cb[0]
        if col_w <= W - 2*margin_px and total_h <= H - 2*margin_px:
            break
    x = (W - col_w)//2
    y = (H - total_h)//2
    for i, ch in enumerate(text):
        draw.text((x, y + i*step), ch, font=font, fill=(0,0,0))  # 黒
    return img

# --------------------
# その他
# --------------------
def sha1_of_text(t:str)->str:
    import hashlib
    return hashlib.sha1(t.encode("utf-8")).hexdigest()[:16]

def write_license_file():
    text = f"""Dataset License (CC0, JP glyphs + EN descriptions)

Copyright (c) {datetime.date.today().year} YOUR_NAME
All images (Japanese text) and English descriptions are synthetic and authored by the dataset creator.
Released under CC0-1.0 (Public Domain Dedication).
"""
    Path("./LICENSE.txt").write_text(text, encoding="utf-8")

def append_assets_registry(font_paths):
    reg_path = Path("./provenance/assets_registry.csv")
    reg_path.parent.mkdir(parents=True, exist_ok=True)
    new = not reg_path.exists()
    with open(reg_path, "a", newline="", encoding="utf-8") as f:
        w = csv.writer(f)
        if new:
            w.writerow(["asset_type","path","license","notes"])
        for p in font_paths:
            w.writerow(["font", str(p), "SIL Open Font License (assumed)", "同梱時は各フォントのLICENSEを添付"])

def make_tarfile(source_dir: Path, output_filename: Path, remove_source=False):
    with tarfile.open(output_filename, "w:gz") as tar:
        tar.add(source_dir, arcname=source_dir.name)
    if remove_source:
        shutil.rmtree(source_dir)

# --------------------
# メイン
# --------------------
def main(n_train, seed, mode, img_size, max_font_size, min_font_size, margin_px, words_file, archive, remove_source):
    random.seed(seed); np.random.seed(seed)
    TRAIN_DIR.mkdir(parents=True, exist_ok=True)
    write_license_file()
    fonts = list_fonts(DEFAULT_FONTS_DIR)
    append_assets_registry(fonts)

    words = load_words(words_file)
    exs = gen_examples(n_train, seed, words)

    def render(jp_text, font_path, writing_mode):
        if writing_mode == "horizontal":
            return draw_horizontal(jp_text, font_path, img_size, max_font_size, min_font_size, margin_px)
        elif writing_mode == "vertical":
            return draw_vertical(jp_text, font_path, img_size, max_font_size, min_font_size, margin_px)

    for i, ex in enumerate(exs):
        font = random.choice(fonts)
        writing_mode = random.choice(["horizontal","vertical"]) if mode=="both" else mode
        img = render(ex.jp_text, font, writing_mode)
        uid = sha1_of_text(f"{i}-{ex.jp_text}-{font.name}-{writing_mode}-{img_size}-{max_font_size}-{min_font_size}")
        img_path = TRAIN_DIR/f"{uid}.png"
        txt_path = TRAIN_DIR/f"{uid}.txt"
        img.save(img_path)
        txt_path.write_text(ex.en_desc, encoding="utf-8")

    print(f"Generated {n_train} samples in {TRAIN_DIR}")

    if archive:
        tar_path = Path("./train.tar.gz")
        make_tarfile(TRAIN_DIR, tar_path, remove_source=remove_source)
        print(f"Created archive: {tar_path}")

if __name__ == "__main__":
    ap = argparse.ArgumentParser()
    ap.add_argument("--n_train", type=int, default=1000)
    ap.add_argument("--seed", type=int, default=0)
    ap.add_argument("--mode", type=str, default="both", choices=["horizontal","vertical","both"])
    ap.add_argument("--img_size", type=int, nargs=2, metavar=("WIDTH","HEIGHT"), default=(640,640))
    ap.add_argument("--max_font_size", type=int, default=54)
    ap.add_argument("--min_font_size", type=int, default=28)
    ap.add_argument("--margin_px", type=int, default=28)
    ap.add_argument("--words_file", type=Path, required=True, help="日本語単語リストファイル（UTF-8, 1行1語）")
    ap.add_argument("--archive", action="store_true", help="train/ を tar.gz にまとめる")
    ap.add_argument("--remove_source", action="store_true", help="tar作成後に train/ を削除")
    args = ap.parse_args()
    main(args.n_train, args.seed, args.mode, tuple(args.img_size), args.max_font_size, args.min_font_size,
         args.margin_px, args.words_file, args.archive, args.remove_source)