MAS-DETR

MAS-DETR is a state-of-the-art Python library for real-time object detection and segmentation based on DETR (DEtection TRansformer) architecture with ultra-fast inference and lightweight backbones.

🚀 Installation

Install the package via pip or uv:

pip install masdetr

Or using uv:

uv add masdetr

Optional Extras

Install additional capabilities for training, ONNX export, or TensorRT deployment:

# Training support (PyTorch Lightning, COCO eval, PEFT)
pip install "masdetr[train]"

# ONNX export support
pip install "masdetr[onnx]"

# TensorRT deployment support
pip install "masdetr[tensorrt]"

# Full installation
pip install "masdetr[train,onnx,tensorrt]"

⚡ Quick Start

1. Model Initialization & Inference

from masdetr import MASDETRNano, MASDETRSmall, MASDETRMedium, MASDETRLarge

# Option A: Train from scratch (no pretrained weights required)
model = MASDETRNano(pretrain_weights=None)

# Option B: Load from a local PyTorch checkpoint
model = MASDETRSmall.from_checkpoint("path/to/checkpoint.pth")

# Predict on an image file, URL, or PIL Image
results = model.predict("image.jpg")
print(results)

2. Model Training

Train an MAS-DETR model on a custom dataset in COCO format:

from masdetr import MASDETRSmall

model = MASDETRSmall(pretrain_weights=None)

# Train on COCO dataset
model.train(
    dataset_dir="path/to/coco_dataset",
    dataset_file="coco",
    epochs=50,
    batch_size=8,
    lr=1e-4,
)

3. Model Export

Export your trained model to ONNX or TensorRT:

from masdetr import MASDETRNano

model = MASDETRNano(pretrain_weights=None)

# Export to ONNX
model.export("model.onnx", format="onnx")

🛠️ Command Line Interface (CLI)

MAS-DETR includes a full-featured command-line tool:

# View general CLI options
masdetr --help

# Run training via CLI
masdetr train --config config.yaml

# Export model via CLI
masdetr export --model masdetr-nano --format onnx

📑 Model Variants

📄 License

This project is licensed under the Apache License 2.0.
Portions of the codebase are derived from RF-DETR by Roboflow, Inc., licensed under Apache 2.0.