AutoVision

Enterprise-grade visual model training framework: train → evaluate → export → deploy in one chain. Its flagship model family is GTR (MIT): one shared backbone covering detection, oriented detection, instance / semantic segmentation, depth and pose, with all 26 official weights registered and sha256-pinned. The Python API and autovision-cli share one strict Pydantic validation surface.

Distribution name: the PyPI package is autovision-framework, while the import package stays autovision and the CLI stays autovision-cli. An unrelated autovision 0.0.1 package exists on PyPI — make sure you install autovision-framework.

Links: Source repository · Documentation · Changelog

Installation

Python ≥ 3.10:

pip install autovision-framework

# GTR route: backbone + official weight download + ONNX export (CUDA required)
pip install "autovision-framework[gtr,download,export]"

# CPU-friendly detection route
pip install "autovision-framework[rfdetr,export]"

Model packages load lazily per extra; a missing extra yields an empty registry listing instead of an error. See the repository README for the full extras table and the GTR task matrix.

Quick start (GTR)

Download the official MIT weights, then run inference — no training:

python scripts/download_gtr_assets.py --task det   # or --all for all 26

import autovision

model = autovision.detection("detr/gtrvit-s", pretrained=True)  # COCO AP 53.6
result = model.predict("image.jpg")   # {'bboxes': ..., 'scores': ..., 'labels': ...}
model.export("model.onnx")            # ONNX export, parity gate on by default

Finetune from official weights (pose shown; five more tasks share the shape):

result = autovision.train_pose_estimation(
    out="runs/pose", data=data,       # COCO person_keypoints JSON + images
    model="gtr/s",
    model_args={"checkpoint": "gtrpose_s_coco"},   # omit for random init
)
print(result.best_exported_model_path, result.metrics)

Equivalent CLI (same Pydantic validation):

autovision-cli list-models --task detection --pretrained
autovision-cli train-pose-estimation out=runs/pose data=pose.yaml model=gtr/s \
    model_args='{"checkpoint": "gtrpose_s_coco"}'
autovision-cli predict model=runs/pose/exported_models/exported_best.pt images=dir/ out=preds/

CPU smoke path (no GPU): rtdetr/resnet18 on the bundled coco8 dataset via autovision.detection("rtdetr/resnet18").train(...).

Common extras

GTR exports ONNX / TensorRT / OpenVINO via the pure-ONNX GLA lowering (fixed-size graphs); the eager forward needs the gtr extra on CUDA.

License

Apache-2.0 (AutoVision original code; full text: https://www.apache.org/licenses/LICENSE-2.0). Third-party component and weight provenance ship in NOTICE and THIRD_PARTY_LICENSES.md. GTR weights are MIT. Restricted-license weights (DINOv2 CC-BY-NC 4.0 / DINOv3 License) are not downloaded by default; user-held weights may not be redistributed.

Full documentation (model zoo, benchmarks, export details, known limits) is available on the documentation site and in the docs/ directory of the source repository.