Hi, I want to convert yolov9c or yolov9s model to hef model and I know that for this I have to convert pt model to onnx model first. Also I have installed hailo_dataflow_compiler-3.30.0 and hailo_model_zoo-2.15.0 and I was able to create har model with the following command:
hailo parser onnx ./yolov9c.onnx --end-node-names “Conv_1058” “Conv_1065” “Conv_1088” “Conv_1095” “Conv_1118” “Conv_1125” --hw-arch “hailo8l”
Now I want to know what other commands I should use and how?
I apologize for my repeated question because I could not find a comprehensive and suitable solution.
I recommend using the Hailo AI Software Suite Docker for easy installation.
Next please work trough the built in tutorials. Inside the Docker use the following command to start a Jupyter Notebook server with notebooks for each step of the conversion process.
hailo tutorial
You can also use the ALLS script from the Model Zoo as basis for your model script.
Thank you for your guidance.
Hello,
After parsing and creating the .har file you can run the following 2 commands:
hailo optimize <har-file> --hw-arch <hailo-board> --model-script <.alls-file> --calib-set-path <calibration-dataset>
hailo compiler <optimized-har-file> --hw-arch <hailo-board>
You can run the following command to see the available arguments:
hailo -h
1 Like
Hello, thank you very much for your guidance.
I followed the instructions and the hef model was created, but I have a few questions:
1- Shouldn’t the calibrated images that we convert to npy format have labels?
2- I checked the hef model that I created based on image calibration (–calib-set-path) with the model that was calibrated randomly(–use-random-calib-set) and there was no difference in the output of the models. Shouldn’t the output of the model calibrated with the image be better than the output of the model that was calibrated randomly? Because the model in hailo model zoo has much better performance.
3- If we have a model that has two inputs, how can we determine which npy images are related to which input?
number of images(.npy): 64
No problem. Glad to hear that it was useful
- No. You can find a more detail explanation here.
- If you wanna increase your performance you could try: increasing your calibration dataset or modify your .alls file. More info for option 1 and option 2. Note that both solutions will increase your compilation time and demand more resources
- Sorry, I didn’t understand the question
Hope that helps