The Hailo Model Zoo provides a comprehensive environment for deep learning applications on Hailo's hardware accelerators. It offers pre-trained models, accuracy evaluation tools (full precision, quantized, and on-device), and the capability to generate Hailo Executable Format (HEF) binaries. This accelerates development and deployment of high-performance AI solutions for engineers and researchers working with Hailo technology.

How It Works

This project utilizes the Hailo Dataflow Compiler (DFC) as its core engine for model processing. It supports models in ONNX and TensorFlow formats, enabling users to parse, optimize, compile, and evaluate them. The workflow includes measuring full precision accuracy, quantized accuracy via the Hailo Emulator, and on-device accuracy, culminating in the generation of HEF files optimized for Hailo's hardware. The zoo also provides models trained on public and internal datasets, alongside retraining instructions.

Quick Start & Requirements

• Primary Install: Clone the repository, navigate to the directory, and run pip install -e .. A sample run command is hailomz info mobilenet_v1.
• Prerequisites: Python 3.10-3.12, TensorFlow 2.19.1, CUDA 12.5.1, Hailo Dataflow Compiler v5.4.0, and optionally HailoRT v5.4.0. Access to Hailo Dataflow Compiler and HailoRT requires contacting Hailo.ai.
• Links: Installation Guide (docs/GETTING_STARTED.rst), DFC User Guide (https://hailo.ai/developer-zone/), Benchmarks (https://hailo.ai/developer-zone/benchmarks/).

Highlighted Details

• Offers pre-trained models for Hailo-15H, -10H, -15L, -8, and -8L devices, covering Classification, Object Detection, Segmentation, and other tasks.
• Models are available in ONNX/TF formats and as pre-compiled HEF binaries.
• Includes detailed retraining instructions for custom datasets.
• Features three Claude Code skills (/hailo-parse, /hailo-optimize, /hailo-compile) to guide users through the DFC workflow.

Maintenance & Community

Issue reporting is disabled in the repository; users are directed to the Hailo Community Forum (https://community.hailo.ai/) for bug reports, feature requests, and discussions. The core stack has been updated to Dataflow Compiler v5.4.0 and HailoRT v5.4.0. A changelog is available at docs/CHANGELOG.rst.

Licensing & Compatibility

The Hailo Model Zoo is released under the MIT license. Compatibility is branch-dependent: the v2.x branch supports Hailo-8/8L devices with DFC v3.x, while the master branch is intended for Hailo-10 and Hailo-15 devices.

Limitations & Caveats

Access to essential components like the Hailo Dataflow Compiler and HailoRT requires direct engagement with Hailo.ai. Device support is explicitly tied to specific repository branches, necessitating careful selection based on the target hardware. All support and issue resolution are channeled through the community forum, with direct repository issue tracking disabled.