Request for Guidance: Integrating TensorFlow Lite Micro (C++) with Custom MicroPython Port (C-based Firmware)

[Biancaa-R]

Hello maintainers,

I had worked on a custom MicroPython port for a RISC-V based secureIoT SoC. The firmware is fully written in C, and I’ve successfully integrated drivers for most of the peripherals. https://github.com/Mindgrove-Technologies/micropython

Now, I’d like to add TensorFlow Lite Micro (TFLM) support to the same port for on-device inference. However, since TFLM is written primarily in C++, and MicroPython’s core and most ports are written in C, I’m looking for the right way to bridge the two.

My setup:

Target: RISC-V 32-bit bare-metal SoC (no OS)

Toolchain: riscv32-unknown-elf-gcc

Current MicroPython port: fully functional (C-based)

Goal: Integrate a subset of TFLite Micro kernels and expose an API (possibly under a tflm module) to run inferences from MicroPython scripts

What I’ve considered so far:

Compiling the TFLM source with riscv32-unknown-elf-g++ and linking against MicroPython’s build.

Creating a C wrapper layer around C++ TFLM APIs so that MicroPython’s C module system can interact with it.

Modifying the port Makefile to support mixed C/C++ compilation (adding CXX = riscv32-unknown-elf-g++ and appropriate flags).

Questions:

Is there an established or recommended way to integrate C++ libraries (like TFLM) into a MicroPython port that’s written purely in C?

Should I expose the TFLM APIs through a C wrapper (e.g., tflm_wrapper.c) or is it better to directly compile and link .cpp sources alongside MicroPython?

Are there known examples of existing MicroPython ports that have successfully used C++ libraries (e.g., for ML or DSP acceleration)?

Any advice on memory allocation — since both MicroPython and TFLM use custom allocators — to avoid conflicts in a bare-metal environment?

PS. Any advice would be helpful !

[andrewleech]

OpenMV has integrated Tensorflow Lite https://docs.openmv.io/library/omv.tf.html

I'm not sure quite how it's integrated, there'a few pieces:

• https://github.com/openmv/libtflm
• https://github.com/openmv/openmv/blob/master/lib/tflm/tflm_backend.cc

[Biancaa-R]

Oh okay ,this is nice I'll look into it :) ..
so you are suggesting me to try openMV instead of Micropython right? It would be more direct to use this ...
But since its a custom board I have to setup the entire drivers from scratch again instead of using micropython.
It seems they use C-style wrappers using the extern "C" linkage specification here ,I can probably try something similar .Thank you.

[jonnor]

You can probably adapt what OpenMV has done to standard MicroPython. It would be great to have a TFLM module that one could just include as a USER_C_MODULE in a typical build.

[jonnor]

If you do not need TFlite specifically, you can use emlearn-micropython to run models defined in Keras. https://github.com/emlearn/emlearn-micropython

Disclaimer: I am the maintainer

[Biancaa-R]

If you do not need TFlite specifically, you can use emlearn-micropython to run models defined in Keras. https://github.com/emlearn/emlearn-micropython

Disclaimer: I am the maintainer

Oh okay ,thats great thanks @jonnor