PyTorch example clients into TinyML

Flower Help - Beginners
1

Hi,
I’m using the pytorch example project and I want to use TinyML [1]. How can I do this? Is this something I can do with flower directly or do I have to go another way?

[1] Federated learning for IoT devices: Enhancing TinyML with on-board training - ScienceDirect

2

Welcome to the Flower community, @neonduck!

Great question. While Flower’s PyTorch example runs on devices with full Python (like laptops, Raspberry Pi, or Jetson), true TinyML requires tweaks.

Microcontrollers typically can’t run PyTorch or perform on-device training due to memory constraints (usually <256KB SRAM).

Flower does support clients in any language (like C++), but you’d need to implement a lightweight client and use a compatible TinyML framework for inference. Here’s a good reference on running FL on embedded devices.

Let us know more about your setup and we can help point you in the right direction!

1 Like
3

Thanks for the answer!

So I have to use some other ML framework that is also supported by Flower? As far as I understand is TensorFlow a better solution for that?

The goal of my experiment would be a FL model that is able to run on Smart Home devices that have similar ressources like devices that are used in the real-world. I have a Raspberry Pi 5 that acts like a gateway (DiyHue), an RPi that acts like a voice assisstant but also a motion sensor (Sonoff SNZB-03), Sonoff NS Panel Pro 120, Android 13 Device, and a Sonoff IP Security Camera.

4

@neonduck Thanks for clarifying your setup!

If your goal is to eventually deploy models to resource-constrained devices like the Sonoff units, you’re right that TensorFlow Lite Micro (TFLM) is often the go-to for true TinyML on microcontrollers. PyTorch itself can’t run directly on such low-memory devices, but you can:

  1. Train in PyTorch (or another full ML framework) on more capable devices (like your Raspberry Pi 5 gateway or other edge nodes).

  2. Export the trained model to a format supported by your target TinyML runtime (e.g., convert PyTorch → ONNX → TFLM, or train in TF and export directly to TFLM).

  3. Run inference on the microcontrollers with TFLM (or similar) while keeping the heavier FL client logic on devices that can handle it.

  4. Use the Pi(s) as Flower clients — they can participate in federated training, then push converted models down to your smaller devices for deployment.

So in short: Flower can still orchestrate the FL part for your capable devices, and you can integrate a conversion + deployment step for the smaller ones. TensorFlow (via TFLM) is indeed a solid choice for the final microcontroller stage, but you don’t have to switch your whole training pipeline to TF unless you want to simplify that conversion.

5

I figured out that PyTorch has a C++ framework. Could this be a possible solution with Flower’s C++ to use less resources? Further, it seems that PyTorch has a TinyML integration

6

sorry for delayed reply,

both links could provide possible solutions and we do need to update our C++ SDK anyway.

Did you try out if the any of the suggested links work? (some sort of minimal example) @neonduck

7

I will create one in the next weeks but I have to work on a paper before.

1 Like