I am trying to compile my own tensor flow model that i have trained, quantized to int8 and converted to onnx. However i am running on apple silicon, so the compiler doesn’t work. Also it does not work on an linux x86 emulator either when running on apple, since the compiler requires operations specific to x86. Does anyone know if the dataflow compiler will be compatible with ARM processors any time soon? Also do you know how i could compile this on apple silicon?
Welcome to the Hailo Community!
You must not quantize your model before exporting it to ONNX. The Hailo model conversion requires models to be in floating point format. The quantization is one part of the model optimization. The Hailo quantization scheme has some unique properties and can therefore only be done by the Dataflow Compiler.
There are currently no plans to port the Dataflow compiler to the ARM architecture. We have many other topics we are working on right now and this would require a large amount of development resources.
No the Dataflow Compiler is not open-source. The HailoRT runtime however is open-source so you could run inference using your Apple silicon as a host.
1 Like