FishServer/FishMeasure/detect_refbox/train_refbox_yolo.py

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
使用 Ultralytics YOLO 训练参考物（ref）单类检测模型。

默认权重：仓库根目录 yolo26n.pt
默认数据：先运行 prepare_refbox_dataset.py 生成 detect_refbox/dataset/refbox.yaml

小样本建议：较轻增强、适中 epoch、较小 batch。
"""

from __future__ import annotations

import argparse
import subprocess
import sys
from datetime import datetime
from pathlib import Path


def repo_root() -> Path:
    return Path(__file__).resolve().parents[1]


def parse_args() -> argparse.Namespace:
    root = repo_root()
    p = argparse.ArgumentParser(description="训练 ref 框 YOLO（yolo26n）")
    p.add_argument(
        "--model",
        type=Path,
        default=root / "yolo26n.pt",
        help="初始权重路径",
    )
    p.add_argument(
        "--data",
        type=Path,
        default=root / "detect_refbox" / "dataset" / "refbox.yaml",
        help="数据集 YAML",
    )
    p.add_argument(
        "--source",
        type=Path,
        default=Path("/home/ubuntu/data/fish/2016-1-22-last_images"),
        help="原始 JSON 图像根目录（仅 --prepare 时）",
    )
    p.add_argument(
        "--dataset-out",
        type=Path,
        default=root / "detect_refbox" / "dataset",
        help="prepare 输出目录",
    )
    p.add_argument("--prepare", action="store_true", help="训练前先执行数据集准备")
    p.add_argument("--val-ratio", type=float, default=0.2, help="验证比例（prepare）")
    p.add_argument("--copy-images", action="store_true", help="prepare 时复制图像")
    p.add_argument("--epochs", type=int, default=200)
    p.add_argument("--batch", type=int, default=8)
    p.add_argument("--imgsz", type=int, default=640)
    p.add_argument("--device", type=str, default="", help="如 0 或 cpu，空则自动")
    p.add_argument(
        "--project",
        type=Path,
        default=root / "detect_refbox" / "runs",
        help="Ultralytics project 目录",
    )
    p.add_argument("--name", type=str, default="", help="运行名，默认带时间戳")
    p.add_argument("--workers", type=int, default=4)
    p.add_argument("--patience", type=int, default=80)
    p.add_argument("--seed", type=int, default=42)
    p.add_argument("--exist-ok", action="store_true")
    return p.parse_args()


def run_prepare(args: argparse.Namespace) -> int:
    prep = Path(__file__).resolve().parent / "prepare_refbox_dataset.py"
    cmd = [
        sys.executable,
        str(prep),
        "--source",
        str(args.source),
        "--out",
        str(args.dataset_out),
        "--val-ratio",
        str(args.val_ratio),
        "--seed",
        str(args.seed),
    ]
    if args.copy_images:
        cmd.append("--copy-images")
    return subprocess.call(cmd)


def main() -> int:
    args = parse_args()
    if args.prepare:
        rc = run_prepare(args)
        if rc != 0:
            return rc

    data_yaml = args.data.expanduser().resolve()
    if not data_yaml.is_file():
        print(
            f"[错误] 未找到 {data_yaml}。请先运行:\n"
            f"  python3 {Path(__file__).parent / 'prepare_refbox_dataset.py'} "
            f"--source {args.source}",
            file=sys.stderr,
        )
        return 1

    model_path = args.model.expanduser().resolve()
    if not model_path.is_file():
        print(f"[错误] 未找到权重: {model_path}", file=sys.stderr)
        return 1

    try:
        from ultralytics import YOLO
    except ImportError as e:
        print("[错误] 需要 ultralytics: pip install ultralytics", file=sys.stderr)
        print(e, file=sys.stderr)
        return 1

    name = args.name or f"refbox_y26n_{datetime.now().strftime('%Y%m%d_%H%M%S')}"
    args.project = args.project.expanduser().resolve()
    args.project.mkdir(parents=True, exist_ok=True)

    # 小数据：关闭强 mosaic/mixup，减轻过拟合风险
    model = YOLO(str(model_path))
    model.train(
        data=str(data_yaml),
        epochs=args.epochs,
        imgsz=args.imgsz,
        batch=args.batch,
        device=args.device if args.device else None,
        project=str(args.project),
        name=name,
        workers=args.workers,
        patience=args.patience,
        seed=args.seed,
        exist_ok=args.exist_ok,
        verbose=True,
        mosaic=0.0,
        mixup=0.0,
        copy_paste=0.0,
        degrees=5.0,
        translate=0.08,
        scale=0.25,
        shear=0.0,
        perspective=0.0,
        flipud=0.0,
        fliplr=0.5,
        hsv_h=0.01,
        hsv_s=0.4,
        hsv_v=0.3,
        close_mosaic=0,
    )

    save_dir = args.project / name
    print(f"\n完成。权重目录: {save_dir / 'weights'}")
    return 0


if __name__ == "__main__":
    raise SystemExit(main())