Hailo ONNX to HEF Converter

This docker image can be used to convert an ONNX exported yolov8s model to Hailo's HEF format. It utilizes the Hailo Dataflow Compiler.

Pre-requisites

• Docker
• ONNX exported model (only yolov8s tested so far)
• Calibration images

Usage

1. Train Model and Export

Train your model with Ultralytics and export it to ONNX format.

yolo export model=/path/to/trained/best.pt imgsz=640 format=onnx opset=11

For reference see:

• https://docs.ultralytics.com/modes/export/#why-choose-yolov8s-export-mode
• https://github.com/hailo-ai/hailo_model_zoo/blob/master/training/yolov8/README.rst

2. Clone the repository

Note

The cloning takes longer than usual because of a large whl file.

git clone https://github.com/cyclux/HailoConverter.git
cd HailoConverter

3. Build the docker image

docker build -t hailo_converter .

4. Prepare model and calibration images

Place the ONNX model and calibration images (calibration_imgs folder) in the root directory of the repository.

Note

The calibration set should be real images that are a subset of the training dataset. They should be diverse and representative of the dataset. Max 64 images are used.

cp /path/to/best.onnx best.onnx
cp -r /path/to/calibration_imgs calibration_imgs

5. Run the docker image

Ensure that the current directory is where the ONNX model and calibration images are placed.

Important

They need to be named best.onnx and calibration_imgs.

docker run -v $(pwd):/workspace --gpus all --ipc=host hailo_converter:latest

This will run the following command inside the docker container:

hailomz compile --ckpt best.onnx --hw-arch hailo8l --calib-path calibration_imgs/ --yaml yolov8s_custom.yaml --model-script yolov8s_custom.alls --classes 1

Note

The Using deprecated NumPy API warning at the beginning can be ignored it seems.

Note

The compilation can take a considerable amount of time.

Once successful, the HEF model will be saved in the same directory.