hailo-ai/hailo_model_zoo

The Hailo Model Zoo includes pre-trained models and a full building and evaluation environment

What it solves

This project provides a collection of pre-trained deep learning models optimized for high-performance execution on Hailo AI hardware. It simplifies the process of taking a model from a framework like TensorFlow or ONNX and deploying it to an edge device by handling the translation, optimization, and compilation into a hardware-specific binary format.

How it works

The Model Zoo integrates with the Hailo Dataflow Compiler (DFC) to move models through a deployment pipeline:

  1. Parsing: Translates input models (ONNX/TF) into Hailo's internal representation.
  2. Profiling: Generates reports on expected performance on Hailo hardware.
  3. Optimization: Compresses the model into an integer representation for efficient inference.
  4. Compilation: Converts the optimized model into a Hailo Executable Format (HEF) binary file.
  5. Evaluation: Measures accuracy using either the Hailo Emulator or actual Hailo hardware.

Who it’s for

Developers and engineers building AI applications for Hailo-10, Hailo-15, Hailo-8, and Hailo-8L hardware who need ready-to-use models or a streamlined path to deploy custom models.

Highlights

  • Pre-compiled Binaries: Provides HEF files for immediate deployment to Hailo devices.
  • Broad Task Support: Includes models for classification, object detection, and semantic segmentation.
  • ** eigenvalues**: Supports both public models trained on open datasets and specialized in-house models.
  • Retraining Capabilities: Includes instructions for retraining models on custom datasets.
  • Claude Code Integration: Ships with specialized skills to automate the parsing, optimization, and compilation flow.