DataForge is a CLI tool designed to automate the preparation and management of machine learning datasets. It helps you transform raw data (like videos and unsorted images) into clean, balanced, and ready-to-train datasets with minimal effort.

Read the full documentation here

• Parallel Processing: Uses multiprocessing to handle thousands of files quickly.
• Vectorized Calculations: Employs NumPy for ultra-fast image comparison and hashing.
• Smart Caching: Incremental caching (MD5-based) allows working with large datasets on NAS or local storage without re-calculating existing data.
• Flexible Configuration: Built with Pydantic v2 for safe settings via config.json or CLI arguments.

• move — Move files from source to target directory based on specific patterns.
• slice — Convert video files into sequences of images. Use --remove to delete the source video after a successful slice.
• delete — Safely remove files matching specific patterns.
• dedup — Find and remove visual duplicates using dHash.
  • Threshold: Similarity limit (0-100%).
  • Core Size: Higher values (e.g., 32) detect small changes; lower values (e.g., 8) ignore noise.
• clean-annotations — Automatically find and delete "orphan" annotation files (XML/TXT) that do not have a corresponding image.
• convert-annotations — Convert dataset labels between formats (e.g., Pascal VOC to YOLO).
• stats — Advanced Dataset Analytics & Health Check This command performs a deep-dive into your dataset to identify biases and feature correlations before you start training.
  • Analytics Highlights:
    • Class Distribution: Visualizes object counts to detect imbalances.
    • Spatial Density Heatmaps: Identifies "positional bias" for each class using 3x3 grids.
    • Correlation Analysis: Global and per-class matrices showing relationships between features.
    • Dataset Manifold (UMAP): 2D projection to identify "representation gaps" and object clusters.
    • Quality Metrics: Analysis of object areas, aspect ratios, brightness, contrast, and blur.
    • Outlier Detection: Automatically marks extreme data points using the IQR method.
  • Outputs: Technical console summary, high-resolution PNG plots, and unified PDF reports.

Usage Example:

python data_forge.py stats --src ./data/train --target_format yolo --report_path ./reports/v1

By default, commands run once. To monitor a folder and process files as they appear, use these flags:

• -r: Run the command in a continuous cycle.
• -s: Set the delay (in seconds) between cycles.

1. Clone the repository:

git clone https://github.com/SeregaCodit/DataForge.git
cd DataForge

2. Setup environment:

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

3. Check usage:

python data_forge.py --help             # See all available commands
python data_forge.py {command} --help   # See arguments for a specific command

For multiple tasks, you can modify start_all_tasks.sh and run them in the background:

bash start_all_tasks.sh

To stop all running DataForge processes:

pkill -f data_forge.py

You can manage default settings in config.json. DataForge follows this priority: CLI Arguments > config.json > Internal Defaults.