A versatile, MIT-licensed computer vision library, LibreYOLO provides inference and training capabilities for a wide range of object detection and image classification models. It is designed for engineers and researchers looking to streamline their computer vision workflows, offering compatibility with common YOLO-format datasets and a familiar high-level Python and CLI interface to ease integration and reduce migration effort.

How It Works

LibreYOLO abstracts complex model operations through a user-friendly Python API and command-line interface. It natively supports common YOLO-format datasets and offers pre-trained models for architectures like YOLOv9 and RF-DETR, enabling both efficient inference and straightforward fine-tuning. The library's design emphasizes extensibility, allowing users to install optional "extras" for specific model families, advanced training utilities, and various export backends such as ONNX, TensorRT, and OpenVINO.

Quick Start & Requirements

Installation is typically achieved via pip: pip install libreyolo. For users requiring specific model families (e.g., rfdetr) or export backends (e.g., onnx), the command can be extended: pip install "libreyolo[rfdetr,onnx]". For development purposes, cloning the repository and running pip install -e . enables editable mode. The library also supports image classification, including loading pre-trained classifiers and fine-tuning on ImageFolder datasets. Links to official documentation and contribution guidelines are available within the repository.

Highlighted Details

• Core library is licensed under the permissive MIT license.
• Features YOLOv9 for CNN-based detection and RF-DETR for transformer-based detection/segmentation as flagship models.
• Supports a broad spectrum of export formats, including ONNX, TensorRT, OpenVINO, NCNN, TFLite, and CoreML.
• Integrates built-in experiment loggers (TensorBoard, MLflow, WandB) and universal training hooks via callbacks.

Maintenance & Community

The project actively encourages community involvement through issue reporting, suggestions, and code contributions, with detailed guidelines provided in CONTRIBUTING.md. Specific information regarding core maintainers, sponsorships, or dedicated community channels (e.g., Discord, Slack) is not detailed in the provided README excerpt.

Licensing & Compatibility

The LibreYOLO library itself is distributed under the MIT license, which generally permits commercial use and integration into closed-source projects. However, users must be aware that pre-trained weights may inherit licenses from their original sources; verification on the specific Hugging Face repository for each weight is recommended.

Limitations & Caveats

Certain model families and export functionalities are designated as experimental (e.g., RF-DETR training/export, YOLOv9 export). The compatibility table indicates that some features or model types are not yet supported. Users are responsible for verifying the licensing terms of any pre-trained weights they utilize.