This repository provides a curated collection of pre-trained, state-of-the-art machine learning models in the ONNX format, aimed at developers, researchers, and enthusiasts seeking to leverage diverse AI capabilities across various frameworks and runtimes. It simplifies the adoption of advanced models by offering them in a standardized, interoperable format.

How It Works

The ONNX Model Zoo sources models from prominent open-source repositories like timm, torchvision, and transformers, converting them to the ONNX format using the TurnkeyML toolchain. Models are categorized into Computer Vision, Natural Language Processing, Generative AI, and Graph Machine Learning, with validated models available for immediate use. Large model files are managed via Git LFS. Intel® Neural Compressor is integrated for model quantization, offering both web UI and code-based options.

Quick Start & Requirements

• Installation: Models are accessed via Git LFS (git lfs pull) or direct download from GitHub. Git LFS requires pip install git-lfs.
• Dependencies: Python environment for running models with an ONNX backend. Test data is provided in .pb or .npz formats for validation.
• Resources: Model files can be large. Quantization via Intel® Neural Compressor may require specific tooling.
• Links: Contribution Guidelines, Netron for visualization.

Highlighted Details

• Extensive coverage across Computer Vision (classification, detection, segmentation, face analysis, image manipulation), NLP (comprehension, translation, language modeling, VQA), and Speech/Audio processing.
• Includes optimized versions like MobileNet, ShuffleNet, and YOLO variants for efficient deployment.
• Integration with Intel® Neural Compressor for INT8 quantization and accuracy tuning.
• Provides starter Python code and test data for validating model accuracy with any ONNX backend.

Maintenance & Community

The project encourages community contributions. Details on community channels or specific maintainers are not explicitly listed in the README.

Licensing & Compatibility

• License: Apache License v2.0.
• Compatibility: The Apache 2.0 license is permissive and generally compatible with commercial and closed-source applications.

Limitations & Caveats

The README notes that new models are continuously added and rigorously validated, implying some models may be in a pre-validation or experimental state. Git clone without Git LFS is not recommended due to large file sizes.