Trouble Converting tddfa_mobilenet_v1 ONNX Model to HEF for Hailo-8L

General
1

Hello Hailo Community,

I’m working on an Edge AI project using the Hailo-8L accelerator and a Raspberry Pi-based dual-camera setup. The goal is to run real-time facial landmark detection using a pre-trained tddfa_mobilenet_v1 model (input shape: 1x3x120x120, output: 1x62) for drowsiness and head pose analysis.

So far:

The ONNX model was verified successfully.
I’m using a GCP Ubuntu 22.04 VM (x86_64, 31 GB RAM, AVX support) to perform model compilation.
I’ve installed hailo_dataflow_compiler-3.31.0 in a clean virtual environment and confirmed the CLI works.

Current blocker: The hailo parser onnx command fails to accept YAML or CLI arguments like --model-path or --output-path. I’ve followed the [Hailo Dataflow Compiler User Guide] closely, but every parsing attempt results in an “unrecognized arguments” error. We’ve tried multiple argument structures and YAML-based approaches without success.

I’m wondering if there’s a syntax change or version-specific quirk we’re missing. This step is crucial so we can continue with optimization and generate a .hef compatible with the Hailo-8L board.

P.S. I came across a pre-compiled .hef file on GitHub that claims to support Hailo-8L for facial landmark inference. However, when inspecting it using hailortcli parse-hef tddfa_mobilenet_v1.hef, it shows:
Architecture HEF was compiled for: HAILO8
This mismatch causes incompatibility on the Hailo-8L hardware.

Any guidance, working example commands, or clarification on parser usage in v3.31.0 would be greatly appreciated.

Thank you in advance for your support!

Best regards,
Elton.

2

I recommend using the Hailo AI Software Suite Docker for easier installation and upgrade.

The Hailo Dataflow Compiler does not accept YAML files. This is a function of the Hailo Model Zoo. Inside the Hailo AI Software Suite Docker call:

hailomz --help

That look like my colleagues uploaded the wrong file. I will check and let them know.

I recommend to run the tutorials in the Hailo AI Software Suite to learn how each step of the standard model conversion flow works. Inside the Docker call:

hailo tutorial

This will open a Jupyter Notebook server with notebooks for each step. This flow is more generic then the Model Zoo flow and will allow you to convert your own models and add your own steps in the script.