I am receiving the RPi5 AI kit in the following days and wanted to get an overview on how to run (custom) models on the AI kit.
From the documentation I understood, that a hef file format is required and it can be compiled using the dataflow compiler. But I cannot find it anywhere…
Where can I get the compiler to transform my custom network?
I did notice the topic but are wondering what solutions there are.
Offering an AI accelerator without the capability to run anything except a few pretrained vision based networks…it is hard for me to understand the usecase and intention.
I figured out that the model zoo whl* includes some kind of compiler for limited vision based models. So far I understood you can finetune the networks and compile them, even though it clearly mentions that you need the data flow compiler in the requirements. It looks like you don’t actually need it for some of the networks.
But still, beeing limited to a few selected networks for very specific use-cases defeats the purpose of the AI kit.
I just saw that there is a post mentioning that Hailo is about to work on a solution. Hopefully that is a matter of days and not months…
*you have to downgrade pip to 22.0.x to install the whl, otherwise it will fail (e.g. with the latest pip version)
My current knowledge is, that the dataflow compiler is unfortunately not available as of now.
You can run the pretrained networks from the model zoo. That’s about it. You can not get your retrained or your custom network running on the device as the format is proprietary and the compiler not provided.
Hailo seems to be working on it, but no word on timeline.