emlearn-micropython

Micropython integration for the emlearn Machine Learning library for microcontrollers.

It enables MicroPython applications to run efficient Machine Learning models on microcontroller, without having to touch any C code.

scikit-learn for Microcontrollers

This is a TinyML library, particularly well suited for low-compexity and low-power classification tasks. It can be combined with feature preprocessing, including neural networks to address more complex tasks.

Status

Minimally useful, on some MicroPython ports

• Tested working on x64 (Unix port) and armv7emsp (Cortex M4F/M7 / STM32).
• Not working on armv6m (Cortex M0 / RP2040). Issue
• Not working on xtensawin (ESP32). Issue

Features

• Classification with RandomForest/DecisionTree models
• Classification and on-device learning with K-Nearest Neighbors (KNN)
• Classification with Convolutional Neural Network (CNN), using TinyMaix library.
• Fast Fourier Transform (FFT) for feature preprocessing, or general DSP
• Infinite Impulse Response (IIR) filters for feature preprocessing, or general DSP
• Clustering using K-means
• Scaling and data type transformations for array, using emlearn_arrayutils.
• Load/save Numpy .npy files using micropython-npyfile
• Installable as a MicroPython native module. No rebuild/flashing needed
• Operates on standard array.array data structures
• Models can be loaded at runtime from a file in disk/flash
• Highly efficient. Inference times down to 100 microseconds, RAM usage <2 kB, FLASH usage <2 kB
• Pre-built binaries available for most architectures.

Examples

• xor_trees. A "Hello World", using RandomForest.
• mnist_cnn. Basic image classification, using Convolutional Neural Network.
• har_trees. Accelerometer-based Human Activity Recognition, using Random Forest
• soundlevel_iir. Sound Level Meter, using Infinite Impulse Response (IIR) filters.

Prerequisites

Minimally you will need

• Python 3.10+ on host
• MicroPython 1.23+ running onto your device

Download repository

Download the repository with examples etc

git clone https://github.com/emlearn/emlearn-micropython

Usage

Start with the instructions in XOR example.

Find architecture and .mpy version

The correct .mpy files to use depend on the CPU architecture of your microcontroller, as well as the MicroPython version.

MicroPython version	.mpy version
1.23.x	6.3

Identify which CPU architecture your device uses. You need to specify ARCH to install the correct module version.

ARCH	Description	Examples
x64	x86 64 bit	PC
x86	x86 32 bit
armv6m	ARM Thumb (1)	Cortex-M0
armv7m	ARM Thumb 2	Cortex-M3
armv7emsp	ARM Thumb 2, single float	Cortex-M4F, Cortex-M7
armv7emdp	ARM Thumb 2, double floats	Cortex-M7
xtensa	non-windowed	ESP8266
xtensawin	windowed with window size 8	ESP32

Information is also available in the official documentation: MicroPython: .mpy files

Benchmarks

UCI handwriting digits

UCI ML hand-written digits datasets dataset from sklearn.datasets.load_digits. 8x8 image, 64 features. Values are 4-bit integers (16 levels). 10 classes.

Running with a very simple RandomForest, 7 trees. Reaches approx 86% accuracy. Tested on Raspberry PI Pico, with RP2040 microcontroller (ARM Cortex M0 @ 133 MHz).

NOTE: over half of the time for emlearn case, is spent on converting the Python lists of integers into a float array. Removing that bottleneck would speed up things considerably.

Developing locally

Prerequisites

These come in addition to the prequisites described above.

Make sure you have the dependencies needed to build for your platform. See MicroPython: Building native modules.

We assume that micropython is installed in the same place as this repository. If using another location, adjust MPY_DIR accordingly.

You should be using the latest MicroPython 1.23 (or newer).

Build

Build the .mpy native module

make dist ARCH=armv6m MPY_DIR=../micropython

Install it on device

mpremote cp dist/armv6m*/emlearn_trees.mpy :emlearn_trees.mpy

Run tests

To build and run tests on host

make check

Citations

If you use emlearn-micropython in an academic work, please reference it using:

@misc{emlearn_micropython,
  author       = {Jon Nordby},
  title        = {{emlearn-micropython: Efficient Machine Learning engine for MicroPython}},
  month        = aug,
  year         = 2023,
  doi          = {10.5281/zenodo.8212731},
  url          = {https://doi.org/10.5281/zenodo.8212731}
}