day/split12.py

import os
import shutil
import random
from pathlib import Path

def repartition_yolo_dataset_safe(
    dataset_root: str,
    train_ratio: float = 0.8,
    val_ratio: float = 0.1,
    test_ratio: float = 0.1,
    seed: int = 42,
    image_exts: tuple = ('.jpg', '.jpeg', '.png'),
    class_names: list = None,
    backup: bool = True
):
    dataset_root = Path(dataset_root)
    images_root = dataset_root / 'images'
    labels_root = dataset_root / 'labels'

    splits = ['train', 'val', 'test']
    
    # 检查原始结构是否存在
    for s in splits:
        if not (images_root / s).exists() or not (labels_root / s).exists():
            raise FileNotFoundError(f"缺少 {s} 目录，请确保数据集结构完整")

    # === 1. 备份原始数据 ===
    backup_dir = dataset_root / 'backup_original'
    if backup and not backup_dir.exists():
        print("📁 正在备份原始数据...")
        shutil.copytree(images_root, backup_dir / 'images')
        shutil.copytree(labels_root, backup_dir / 'labels')
        print("✅ 备份完成")

    # === 2. 收集所有图片路径（来自 train/val/test）===
    all_images = []
    for split in splits:
        split_img_dir = images_root / split
        for f in split_img_dir.iterdir():
            if f.is_file() and f.suffix.lower() in image_exts:
                all_images.append((split, f))  # 保存 (来源split, 文件路径)

    if not all_images:
        raise ValueError("未找到任何图片！")

    print(f"🔍 找到 {len(all_images)} 张图片")

    # === 3. 随机打乱并划分 ===
    random.seed(seed)
    random.shuffle(all_images)

    total = len(all_images)
    train_end = int(total * train_ratio)
    val_end = train_end + int(total * val_ratio)

    new_splits = {
        'train': all_images[:train_end],
        'val': all_images[train_end:val_end],
        'test': all_images[val_end:]
    }

    # === 4. 创建新的目标目录（加 _new 后缀避免冲突）===
    new_images = dataset_root / 'images_new'
    new_labels = dataset_root / 'labels_new'

    for split in splits:
        (new_images / split).mkdir(parents=True, exist_ok=True)
        (new_labels / split).mkdir(parents=True, exist_ok=True)

    # === 5. 复制文件（从原位置 → 新位置）===
    missing_labels = 0
    for split_name, files in new_splits.items():
        print(f"📦 {split_name}: {len(files)} 张")
        for orig_split, img_path in files:
            # 复制图片
            dst_img = new_images / split_name / img_path.name
            shutil.copy2(img_path, dst_img)

            # 复制标签（标签在 labels/orig_split/ 下）
            label_path = labels_root / orig_split / (img_path.stem + '.txt')
            dst_label = new_labels / split_name / (img_path.stem + '.txt')
            if label_path.exists():
                shutil.copy2(label_path, dst_label)
            else:
                print(f"⚠️ 缺失标签: {label_path}")
                missing_labels += 1

    # === 6. 原子性替换：删除旧 images/labels，重命名 new 为正式名 ===
    shutil.rmtree(images_root)
    shutil.rmtree(labels_root)
    shutil.move(str(new_images), str(images_root))
    shutil.move(str(new_labels), str(labels_root))

    # === 7. 生成 dataset.yaml（可选）===
    if class_names:
        yaml_content = f"""path: {dataset_root.resolve()}
train: images/train
val: images/val
test: images/test

names: {class_names}
"""
        with open(dataset_root / 'dataset.yaml', 'w', encoding='utf-8') as f:
            f.write(yaml_content)
        print(f"\n📄 已生成 dataset.yaml")

    print(f"\n✅ 重划分成功！")
    print(f"训练: {len(new_splits['train'])}, 验证: {len(new_splits['val'])}, 测试: {len(new_splits['test'])}")
    if missing_labels:
        print(f"⚠️ 共 {missing_labels} 个标签缺失")

if __name__ == "__main__":
    DATASET_ROOT = r"20251210/20251210"
    #CLASS_NAMES = ["dust"]  # 替换为你的类别

    repartition_yolo_dataset_safe(
        dataset_root=DATASET_ROOT,
        train_ratio=0.8,
        val_ratio=0.1,
        test_ratio=0.1,
        seed=42,
        #class_names=CLASS_NAMES,
        backup=True
    )