libreyolo 1.1.0

Libre YOLO - An open source YOLO library with MIT license.

LibreYOLO

MIT-licensed object detection library with training and inference support across YOLOv9 (t, s, m, c), YOLOX (n, t, s, m, l, x), YOLO-NAS (s, m, l), RF-DETR (n, s, m, l), and D-FINE (n, s, m, l, x).

Installation

pip install libreyolo

For optional runtime and export dependencies such as ONNX Runtime, OpenVINO, TensorRT, NCNN, and RF-DETR, see the full docs.

Inference Backend and Export Support

Format Status

Format	Export	Runtime backend	Status	Precision	Notes
ONNX	Yes	ONNX Runtime	Supported	FP32, FP16	Release-blocking export path and the broadest tested runtime target.
TorchScript	Yes	PyTorch JIT	Experimental	FP32, FP16	Useful compatibility target, but not a release gate.
TensorRT (tensorrt / trt)	Yes	TensorRT	Experimental	FP32, FP16, INT8	CUDA-only path. INT8 requires calibration data.
OpenVINO	Yes	OpenVINO	Experimental	FP32, FP16, INT8	Runtime-specific path with CPU-oriented deployment coverage.
NCNN	Yes	NCNN	Experimental	FP32, FP16	Highest maintenance overhead today. No INT8 path, and some DETR-family models are not supported.

The e2e suite mirrors this policy with pytest markers: supported_backend for ONNX and experimental_backend for the other export backends.

Model Family Matrix

✓ supported, ~ supported with caveats, — intentionally unsupported

Model family	ONNX	TorchScript	TensorRT	OpenVINO	NCNN
YOLOX	✓	✓	✓	✓	✓
YOLOv9	✓	✓	✓	✓	✓
YOLO-NAS	✓	✓	✓	✓	✓
RF-DETR	✓	~	✓	✓	~
D-FINE	✓	✓	✓	✓	—
RT-DETR	✓	✓	✓	✓	—

Notes:

• RF-DETR TorchScript export exists, but tracing can still be brittle on some checkpoints and shapes.
• NCNN is intentionally blocked for D-FINE and RT-DETR because the runtime lacks required DETR query-selection ops.
• RF-DETR on NCNN is not blocked at export time, but current e2e coverage still tracks known runtime limitations.

Quick Start

from libreyolo import LibreYOLO, SAMPLE_IMAGE

# Auto-detect family and size from the checkpoint name
model = LibreYOLO("LibreYOLOXs.pt")
result = model(SAMPLE_IMAGE, save=True)

print(f"Detected {len(result)} objects")
print(result.boxes.xyxy)
print(result.saved_path)

Documentation

Full documentation at libreyolo.com/docs.

License

• Code: MIT License
• Weights: Pre-trained weights may inherit licensing from the original source