manigen 📝

manigen is a fast and reliable CLI tool designed to generate file manifests (lists of file paths) for Machine Learning and Deep Learning datasets.

Whether you are preparing a small local dataset or parsing a massive image corpus like ImageNet, manigen handles recursive scanning, multithreading, path formatting, and Train/Val/Test splitting out of the box.

✨ Features

• ⏱️ Efficient & Multithreaded: Uses a thread pool to parallelize I/O operations, significantly speeding up the scanning of large and deeply nested directory trees compared to sequential scripts.
• ✂️ Portable Manifests: Generate relative paths by stripping absolute prefixes (--strip-prefix), making it easy to move datasets between local machines and cloud servers.
• 🔀 ML-Ready Splits: Built-in shuffling and automatic Train/Validation/Test dataset splitting directly into separate files (--split).
• 🛡️ Robust Architecture: Built with modern Python, featuring thread-safe list operations, strict input validation, and clear error handling.

🎯 Motivation

While working on Super-Resolution Deep Learning projects, I found myself repeatedly copying the same massive datasets across multiple project directories. To save disk space, I decided to store all datasets in a single central location (e.g., ~/.local/share/datasets) and feed the models using simple text files containing absolute paths to the images.

Initially, I wrote a bash script for this task. However, generating a manifest for the ImageNet dataset took about 30 minutes. By rewriting the tool in Python and leveraging multithreading, manigen can now generate a manifest for ImageNet (1,281,167 images) in 12 seconds.

📦 Installation

You can install manigen directly from PyPI using pip:

pip install manigen

Or, if you use uv (recommended for CLI tools):

uv tool install manigen

🚀 Quick Start

Generate a manifest of all images in a dataset directory:

manigen -i ./datasets/ImageNet/train -o manifest.txt

💡 Advanced Usage Examples

1. Multithreaded Scanning

Speed up scanning for datasets with heavily nested directories (like ImageNet) by utilizing multiple threads and recursive search:

manigen -i ./datasets/ImageNet/train -o train_paths.txt -t 8 -r

2. Making Paths Portable (Relative Paths)

If your absolute path is /Users/ml_engineer/projects/data/images/cat.jpg, but you want the manifest to only contain data/images/cat.jpg:

manigen -i /Users/ml_engineer/projects/data -o dataset.txt --strip-prefix /Users/ml_engineer/projects/

3. Creating Train/Val/Test Splits

Automatically shuffle the dataset and split it into training (70%), validation (20%), and testing (10%) sets:

manigen -i ./dataset -o manifest.txt --shuffle --split 0.7 0.2 0.1

4. Custom File Extensions

Override the default extensions to scan for audio, text, or any other formats:

manigen -i ./audio_dataset -o audio_manifest.txt -e wav mp3 flac

🛠️ CLI Reference

🤝 Contributing

1. Clone the repository

git clone https://github.com/ash1ra/manigen
cd manigen

2. Install dependencies using uv

uv sync
# On Windows
.venv\Scripts\activate
# on Unix or MacOS
source .venv/bin/activate

3. Format and lint the code

uv run ruff format .
uv run ruff check .

4. Run the tests

uv run pytest tests/ -v

5. Submit a pull request

If you'd like to contribute, please fork the repository and open a pull request to the main branch.