Hi everyone
Last week, I bought a Cortex-M board (NUCLEO STM32F746ZG) to run my small deep learning network on this board. Because MDK toolchain is very expensive, I tried to use ARM-NN to convert from Tensorflow model to a CMSIS-NN model. I found an example using ARM-NN with Cortex-A board. The output of ARM-NN is a .so file and can load easily to Cortex-A board. However, Cortex-M board doesn't support any system to load .so file
So, I would like to know is it possible to use ARM-NN for converting a Tensorflow model to a CMSIS-NN model and run on a Cortex-M board. If it is possible, please let me know the way to do that?
Thank you so much.
Many thanks for your question. Arm NN currently supports Cortex-A CPUs and Mali GPUs, and specifically uses the Arm Compute Library for the underlying kernels. We are actively considering how best to provide similar support for Cortex-M CPUs, but I don't have news to share on that today. In the meantime - if you haven't seen it already - it might be worth looking at TensorFlow Lite for Microcontrollers (https://github.com/tensorflow/tensorflow/tree/master/tensorflow/lite/experimental/micro). I believe this is in the early stages, but could provide a mechanism for getting TF models to run in the short term.