import argparse import os import random import shutil from tqdm import tqdm import webdataset as wds from huggingface_hub import HfApi, HfFolder def sample_and_copy_files(source_dir, output_dir, num_samples, seed, scan_report_interval): """ Scans a source directory, randomly samples image files, and copies them to an output directory, preserving the original structure. IMPORTANT: This script is currently hardcoded to only sample images from corruption severity level 5. """ print(f"1. Scanning for image files in '{source_dir}'...") all_files = [] count = 0 for root, _, files in os.walk(source_dir): # HARDCODED FILTER: Only include files from severity level 5 directories. # We check if '5' is a component of the directory path. path_parts = root.split(os.sep) if '5' in path_parts: for file in files: if file.lower().endswith(('.png', '.jpg', '.jpeg', '.gif', '.bmp')): all_files.append(os.path.join(root, file)) count += 1 if count > 0 and count % scan_report_interval == 0: print(f" ...scanned {count} images (from severity 5)...") if not all_files: print(f"Error: No image files found for severity level 5 in '{source_dir}'. Exiting.") return False print(f"Found {len(all_files)} total images.") # Sort the file list to ensure reproducibility across different systems. # The order of files returned by os.walk is not guaranteed to be the same. all_files.sort() print(f"2. Randomly sampling {num_samples} images (seed={seed})...") # Set the random seed for reproducibility. # Note: For 100% identical results, the same version of Python should be used # (e.g., this script was created using Python 3.10.14), as the underlying # algorithm for the 'random' module can change between versions. random.seed(seed) num_to_sample = min(num_samples, len(all_files)) sampled_files = random.sample(all_files, num_to_sample) print(f"Selected {len(sampled_files)} files to copy.") print(f"3. Copying files to '{output_dir}'...") if os.path.exists(output_dir): print(f"Output directory '{output_dir}' already exists. Removing it.") shutil.rmtree(output_dir) os.makedirs(output_dir) for file_path in tqdm(sampled_files, desc="Copying files"): relative_path = os.path.relpath(file_path, source_dir) dest_path = os.path.join(output_dir, relative_path) os.makedirs(os.path.dirname(dest_path), exist_ok=True) shutil.copy(file_path, dest_path) print("File copying complete.") return True def generate_readme_content(repo_id, seed, python_version, tar_filename, script_filename): """ Generates the content for the README.md file for the Hugging Face dataset. """ return f"""--- license: mit tags: - image-classification - computer-vision - imagenet-c --- # Nano ImageNet-C (Severity 5) This is a randomly sampled subset of the ImageNet-C dataset, containing 5,000 images exclusively from corruption **severity level 5**. It is designed for efficient testing and validation of model robustness. 这是一个从 ImageNet-C 数据集中随机抽样的子集，包含 5000 张仅来自损坏等级为 **5** 的图像。它旨在用于高效地测试和验证模型的鲁棒性。 ## How to Generate / 如何生成 This dataset was generated using the `{script_filename}` script included in this repository. To ensure reproducibility, the following parameters were used: 本数据集使用此仓库中包含的 `{script_filename}` 脚本生成。为确保可复现性，生成时使用了以下参数： - **Source Dataset / 源数据集**: The full ImageNet-C dataset is required. / 需要完整的 ImageNet-C 数据集。 - **Random Seed / 随机种子**: `{seed}` - **Python Version / Python 版本**: `{python_version}` ## Dataset Structure / 数据集结构 The dataset is provided as a single `.tar` file named `{tar_filename}` in the `webdataset` format. The internal structure preserves the original ImageNet-C hierarchy: `corruption_type/class_name/image.jpg`. 数据集以 `webdataset` 格式打包在名为 `{tar_filename}` 的单个 `.tar` 文件中。其内部结构保留了原始 ImageNet-C 的层次结构：`corruption_type/class_name/image.jpg`。 ## Citation / 引用 If you use this dataset, please cite the original ImageNet-C paper: 如果您使用此数据集，请引用原始 ImageNet-C 的论文： ```bibtex @inproceedings{{danhendrycks2019robustness, title={{Benchmarking Neural Network Robustness to Common Corruptions and Perturbations}}, author={{Dan Hendrycks and Thomas Dietterich}}, booktitle={{International Conference on Learning Representations}}, year={{2019}}, url={{https://openreview.net/forum?id=HJz6tiCqYm}}, }} ``` """ def package_with_webdataset(output_dir, tar_path): """ Packages the contents of a directory into a single .tar file using webdataset. """ print(f"4. Packaging '{output_dir}' into '{tar_path}'...") with wds.TarWriter(tar_path) as sink: for root, _, files in tqdm(list(os.walk(output_dir)), desc="Packaging files"): for file in files: file_path = os.path.join(root, file) with open(file_path, "rb") as stream: content = stream.read() relative_path = os.path.relpath(file_path, output_dir) key, ext = os.path.splitext(relative_path) extension = ext.lstrip('.') sink.write({ "__key__": key, extension: content }) print("Packaging complete.") def upload_to_hf(tar_path, readme_path, script_path, repo_id): """ Uploads a file to a specified Hugging Face Hub repository. """ print(f"5. Uploading files to Hugging Face Hub repository: {repo_id}...") if HfFolder.get_token() is None: print("Hugging Face token not found. Please log in using `huggingface-cli login` first.") return try: api = HfApi() print(f"Creating repository '{repo_id}' (if it doesn't exist)...") api.create_repo(repo_id, repo_type="dataset", exist_ok=True) print("Uploading README.md...") api.upload_file( path_or_fileobj=readme_path, path_in_repo="README.md", repo_id=repo_id, repo_type="dataset" ) script_filename = os.path.basename(script_path) print(f"Uploading generation script '{script_filename}'...") api.upload_file( path_or_fileobj=script_path, path_in_repo=script_filename, repo_id=repo_id, repo_type="dataset" ) tar_filename = os.path.basename(tar_path) print(f"Uploading dataset file '{tar_filename}'...") api.upload_file( path_or_fileobj=tar_path, path_in_repo=tar_filename, repo_id=repo_id, repo_type="dataset" ) print("Upload successful!") print(f"Dataset available at: https://huggingface.co/datasets/{repo_id}") except Exception as e: print(f"An error occurred during upload: {e}") def main(): """ Main function to orchestrate the dataset creation, packaging, and upload process. """ parser = argparse.ArgumentParser(description="Create, package, and upload a smaller version of an image dataset.") parser.add_argument("--source_dir", type=str, default="./data/ImageNet-C", help="Path to the source dataset.") parser.add_argument("--output_dir", type=str, default="./data/nano-ImageNet-C", help="Path to save the new sampled dataset.") parser.add_argument("--num_samples", type=int, default=5000, help="Number of images to sample.") parser.add_argument("--seed", type=int, default=7600, help="Random seed for reproducibility.") parser.add_argument("--repo_id", type=str, default="niuniandaji/nano-imagenet-c", help="The Hugging Face Hub repository ID.") parser.add_argument("--tar_path", type=str, default="./data/nano-ImageNet-C.tar", help="Path to save the final webdataset archive.") parser.add_argument("--scan_report_interval", type=int, default=50000, help="How often to report progress during file scanning.") args = parser.parse_args() print("--- Starting Dataset Creation Process ---") print("IMPORTANT: The script is configured to sample ONLY from severity level 5.") # Generate README content script_filename = os.path.basename(__file__) readme_content = generate_readme_content( args.repo_id, args.seed, "3.10.14", os.path.basename(args.tar_path), script_filename ) # Write README to a local file readme_path = "README.md" with open(readme_path, "w", encoding="utf-8") as f: f.write(readme_content) print(f"Generated README.md for the dataset.") if sample_and_copy_files(args.source_dir, args.output_dir, args.num_samples, args.seed, args.scan_report_interval): package_with_webdataset(args.output_dir, args.tar_path) upload_to_hf(args.tar_path, readme_path, script_filename, args.repo_id) print("--- Process Finished ---") if __name__ == "__main__": main()