If I understand correctly, it is theoretically possible to compile models that are not in the Hailo Model Zoo, and even self-built architecture, into .hef format and deploy onto Hailo-8 chip.
Just for an example, say if I add some decoder layer to yolov8n to achieve depth estimation function.
Are there some experience anyone could share on this kind of task? Or is there like a general tutorial, procedure or cautions for this task?
Hi @ShingHin.Hung this is definitely possible!
If you have installed the SW-suite of the DFC, you can laucnh the tutorials:
hailo tutorial
This would give you a bunch of ipynb based tutorials that show how to conert a resnet model. This is a good basic starting point. You might want to also take a look at the supported layers sections in the DFC user guide.
Yes, it’s possible.
Out of the entire model, extract as much as the range supported by hailo and convert it into hef. Fill in the rest with the code typing.
Based on the Windows os environment and the C++ language, both the custom detection and classification models are found to be functioning normally.