segmentation/prepare_cityscapes_folds.py

import argparse
import glob
import os
import random

from sklearn.model_selection import KFold


def prepare_folds(cityscapes_path, output_dir, n_splits=3):
    """
    Prepares k-fold cross-validation splits for the Cityscapes dataset.

    Args:
        cityscapes_path (str): Path to the root Cityscapes directory
                                (containing leftImg8bit and gtFine).
        output_dir (str): Directory to save the split files.
        n_splits (int): Number of folds for cross-validation.
    """
    leftimg8bit_path = os.path.join(cityscapes_path, "leftImg8bit")
    train_img_dir = os.path.join(leftimg8bit_path, "train")
    val_img_dir = os.path.join(leftimg8bit_path, "val")

    train_files = []
    # Check if train_img_dir exists before listing its contents
    if not os.path.exists(train_img_dir):
        print(f"Error: Training image directory not found: {train_img_dir}")
        print(
            f"Please ensure '{cityscapes_path}' is the correct root and contains 'leftImg8bit/train'."
        )
        return
    for city_folder in os.listdir(train_img_dir):
        city_path = os.path.join(train_img_dir, city_folder)
        if os.path.isdir(city_path):
            train_files.extend(glob.glob(os.path.join(city_path, "*.png")))

    val_files = []
    # Check if val_img_dir exists
    if not os.path.exists(val_img_dir):
        print(f"Error: Validation image directory not found: {val_img_dir}")
        print(
            f"Please ensure '{cityscapes_path}' is the correct root and contains 'leftImg8bit/val'."
        )
        pass
    elif os.path.exists(val_img_dir):
        for city_folder in os.listdir(val_img_dir):
            city_path = os.path.join(val_img_dir, city_folder)
            if os.path.isdir(city_path):
                val_files.extend(glob.glob(os.path.join(city_path, "*.png")))

    if not train_files and not val_files:
        print(f"Error: No image files found in {train_img_dir} or {val_img_dir}.")
        print("Please check your Cityscapes dataset structure and path.")
        return

    all_files = train_files + val_files

    # 关键修改：生成正确的相对路径格式
    all_files_relative = []
    for f in all_files:
        rel_path = os.path.relpath(f, leftimg8bit_path)
        # 去掉 _leftImg8bit.png 后缀
        if rel_path.endswith("_leftImg8bit.png"):
            rel_path = rel_path[: -len("_leftImg8bit.png")]
        all_files_relative.append(rel_path)

    all_files_relative = sorted(all_files_relative)

    # Ensure consistent shuffling for reproducibility if needed
    random.seed(42)
    random.shuffle(all_files_relative)

    kf = KFold(n_splits=n_splits, shuffle=False)  # Shuffle is already done

    os.makedirs(output_dir, exist_ok=True)

    for i, (train_index, val_index) in enumerate(kf.split(all_files_relative)):
        fold_train_files = [all_files_relative[k] for k in train_index]
        fold_val_files = [all_files_relative[k] for k in val_index]

        train_file_path = os.path.join(output_dir, f"fold_{i + 1}_train_split.txt")
        val_file_path = os.path.join(output_dir, f"fold_{i + 1}_val_split.txt")

        with open(train_file_path, "w") as f:
            for item in fold_train_files:
                f.write(f"{item}\n")  # 修复：使用单个 \n

        with open(val_file_path, "w") as f:
            for item in fold_val_files:
                f.write(f"{item}\n")  # 修复：使用单个 \n

        print(f"Fold {i + 1}: {len(fold_train_files)} train, {len(fold_val_files)} val")

        # 添加调试信息
        print("Sample train files:")
        for sample in fold_train_files[:3]:
            print(f"  {sample}")
        print("Sample val files:")
        for sample in fold_val_files[:3]:
            print(f"  {sample}")

    print(f"Split files saved to: {os.path.abspath(output_dir)}")


if __name__ == "__main__":
    parser = argparse.ArgumentParser(description="Prepare Cityscapes k-fold splits.")
    parser.add_argument(
        "cityscapes_path",
        type=str,
        help="Absolute path to the Cityscapes dataset directory.",
    )
    parser.add_argument(
        "--output_dir",
        type=str,
        default=None,
        help="Directory to save the split files. If not provided, a 'splits' folder will be created inside the cityscapes_path.",
    )
    parser.add_argument(
        "--n_splits", type=int, default=3, help="Number of folds (default: 3)."
    )

    args = parser.parse_args()

    if args.output_dir is None:
        effective_output_dir = os.path.join(args.cityscapes_path, "splits")
    else:
        effective_output_dir = args.output_dir

    abs_cityscapes_path = os.path.abspath(args.cityscapes_path)
    if not os.path.isdir(abs_cityscapes_path):
        print(
            f"Error: Cityscapes path not found or is not a directory: {abs_cityscapes_path}"
        )
        exit(1)

    prepare_folds(abs_cityscapes_path, effective_output_dir, args.n_splits)

    script_name = os.path.basename(__file__)
    print("\nTo run this script again, for example:")
    print(f"python {script_name} /path/to/your/cityscapes")
    if args.output_dir is not None:
        print(
            f"python {script_name} /path/to/your/cityscapes --output_dir {args.output_dir}"
        )
    print(
        "Replace '/path/to/your/cityscapes' with the actual path to your Cityscapes dataset."
    )