need some explanation about optimization

General
1

Hi. I’m trying to use moblienet_v2 model from this repo.

I used 3.30 Dataflow complier.

1. Is it necessary to optimize .har file? if not is there any way to compile it to hef without optimization.

I converted onnx to har by using following command :

hailo parser onnx gaze.onnx

and optimized

hailo optimize gaze.har --use-random-calib-set

image
image2159×872 79.9 KB

I used “use-random-calib-set” command without any reason. what does that command do?
seems like I have to do something when optimization. and I have no idea what those command means.

2

Yes, optimization is a mandatory step in the model conversion process. The main part of the optimization is quantizing the network weights. Because some optimization levels go beyond quantizing the weights, the step was renamed from quantization to optimization.

The quantization step requires images from the dataset for calibration. When accuracy is not important, you can use random data for the calibration. This for quick conversions where you want to see how fast the network can run on Hailo.

I recommend you run the tutorials in the AI Software Suite Docker. Call the following command to start a Jupyter Notebook server with notebooks for each step of the model conversion process.

hailo tutorial