ONNX to HEF Conversion Requirements and Cloud Workflow Clarification

General Discussion
,
1

Hi All,

I am currently using the Ultralytics cloud platform primarily because of the computing resources required for model training. Once we are confident that our models and AI architecture are performing well, we plan to move training to a local GPU environment. However, for now, we intend to continue using the cloud-based solution.

The primary output format from Ultralytics is ONNX, which we then need to convert to HEF for deployment on Hailo hardware. My question is: how resource-intensive is the ONNX-to-HEF conversion process? Am I correct in understanding that it requires significantly fewer resources than the original model training?

From what I understand, I need to follow the Hailo Model Zoo examples. If the conversion process is not too resource-heavy, would it be sufficient to run a Linux virtual environment on a standard PC?

I also noticed in one of the examples that during steps 4–5, both the dataset and the ONNX file are required. If the model has already been trained, could you clarify why the dataset is still needed at this stage?

Thank you in advance for your guidance.

2

Generally speaking, the resources needed for ONNX to HEF conversion are indeed fewer than the resources needed for training.

  • GPU is recommended but not mandatory. Some of the advanced quantization algorithms will be skipped without a GPU.
  • The model quantization process requires a few dozens of sample images, but it doesn’t require a full dataset.