mltoarduino 0.0.7

See readMe.ma

Convert a Python Machine Learning Model to Arduino Code (C++)

Introduction

Motivation

What ?
This project demonstrates the conversion of Python machine learning (ML) models to Arduino C++ code.
We will use some ML models purely as examples; the goal is not to find the best model or achieve minimal error.
Why ?
In certain applications, such as embedded systems, small microcontrollers with limited memory and computing resources are used. The idea is to train a machine learning model in a Python environment and then convert the trained model to C++ for deployment on a microcontroller.
In this project, we will use the Arduino Uno as an example, but the approach can be applied to other microcontrollers as well.
How ?
Follow the step-by-step guide below, or go directly to the PyPi package mltoarduino

Hardware

In this project, the Arduino Uno was used, but you can use other boards like Arduino Nano or Micro, Miga 2560, ESP32...
below a comparaison of some Arduino boards:

Feature	Arduino Uno	Arduino Nano	Arduino Micro	Arduino Mega 2560	ESP32
Microcontroller	ATmega328P	ATmega328P	ATmega32U4	ATmega2560	Tensilica Xtensa LX6
Operating Voltage	5V	5V	5V	5V	3.3V
Input Voltage	7-12V	7-12V	7-12V	7-12V	5V via USB or 7-12V
Digital I/O Pins	14 (6 PWM)	14 (6 PWM)	20 (7 PWM)	54 (15 PWM)	34
Analog Input Pins	6	8	12	16	18
Flash Memory	32 KB	32 KB	32 KB	256 KB	Up to 16 MB
SRAM	2 KB	2 KB	2.5 KB	8 KB	520 KB
EEPROM	1 KB	1 KB	1 KB	4 KB	None
Clock Speed	16 MHz	16 MHz	16 MHz	16 MHz	240 MHz (dual-core)
Connectivity	UART, I2C, SPI	UART, I2C, SPI	UART, I2C, SPI	UART, I2C, SPI	Wi-Fi, Bluetooth
USB Interface	USB-B	Mini USB	Micro USB	USB-B	Micro USB
Dimensions	68.6 x 53.4 mm	45 x 18 mm	48 x 18 mm	101.52 x 53.3 mm	51 x 25.5 mm
Power Consumption	~50 mA	~50 mA	~50 mA	~70 mA	Varies (~80-240 mA)
Special Features	Simple and robust	Compact	USB HID support	High I/O count	Wi-Fi and BLE
Price Range	Low	Low	Medium	Medium	Medium-High

How to use the package

# Install the package 
!pip install mltoarduino

# Import all functions 
from mltoarduino import *

Load an example data

FileName="..\data\processed\df_price_4inputs.csv"
dfnew= pd.read_csv(FileName).astype("float32")
print("Df columns: ", list(dfnew.columns))
X= dfnew.iloc[:,:4].values
y= dfnew.iloc[:,4].values
# Split into training and test sets
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)

Df columns:  ['Gearbox_auto', 'HorseP', 'Euro_stand', 'km', 'price']

Example of Linear Regression model

# Load the model
LR_model = joblib.load(r'../models/LinearReg/LR_model.pkl')

# Sub inputs/outpusts to test the arduino model: 10 samples
sub_X=X_train[:10]
sub_y=LR_model.predict(sub_X)

# Convert the model to arduino C++
arduino_code= LinearRegToC (LR_model, sub_X, sub_y)

print("100 first charaters of arduino code \n",arduino_code[:100])

100 first charaters of arduino code 
 

const int Nv = 10;
const int dimX = 4;

/////// Xy ////// 
const float X [] PROGMEM  = {0.0, 116.0

You can use a code like below to save the arduino code.ino To avoid error in arduino environnement, if your file name is Example.ino this file must be stored in the same folder name "Example/Example.ino"

ino_file="ArduinoCode/LinearReg.ino" # Path of the file
ino_file=ino_file.replace(".ino" ,"")
current_directory = os.getcwd()
new_directory_path = os.path.join(current_directory, ino_file)
try:
    os.makedirs(new_directory_path)
except: pass

path=ino_file+"/"+ino_file.split("/")[-1]+".ino"
with open(path,'w+') as f:
    f.write(arduino_code)
    
    print(path, "saved")

The arduino memory usnig

Sketch uses 3906 bytes (12%) of program storage space. Maximum is 30720 bytes. Global variables use 252 bytes (12%) of dynamic memory, leaving 1796 bytes for local variables. Maximum is 2048 bytes.

# The arduino serial print result 
serialPrint="""
Cal_Ardui,Expected,Delta_time(us)
19074.86,19074.861328,68
22590.43,22590.433593,76
18458.25,18458.253906,80
20624.41,20624.408203,76
20219.45,20219.445312,76
29240.26,29240.261718,80
32525.00,32525.000000,84
27160.86,27160.859375,80
25408.60,25408.605468,80
26429.56,26429.554687,88
====The End====="""

# Convert the serial result to DF 
data = serialPrint.split("\n")[1:-1]
data=[x.split(",") for x in data]
DF_serial= pd.DataFrame( data[1:], columns= data[0]).astype("float32")
DF_serial

<style scoped> .dataframe tbody tr th:only-of-type { vertical-align: middle; }

.dataframe tbody tr th {
    vertical-align: top;
}

.dataframe thead th {
    text-align: right;
}

</style>

	Cal_Ardui	Expected	Delta_time(us)
0	19074.859375	19074.861328	68.0
1	22590.429688	22590.433594	76.0
2	18458.250000	18458.253906	80.0
3	20624.410156	20624.408203	76.0
4	20219.449219	20219.445312	76.0
5	29240.259766	29240.261719	80.0
6	32525.000000	32525.000000	84.0
7	27160.859375	27160.859375	80.0
8	25408.599609	25408.605469	80.0
9	26429.560547	26429.554688	88.0

print("The AVG prediction time of one input is", 
      (DF_serial['Delta_time(us)'].mean()/1000).round(2), 
      "ms"
     ) 

The AVG prediction time of one input is 0.08 ms

# Ploting 
DF_serial.plot.scatter(x='Expected', y='Cal_Ardui',  marker='o', label="Arduino calculation")
xx=[DF_serial['Expected'].min(), DF_serial['Expected'].max()]
plt.plot(xx,xx, c='r', label="equal")
plt.legend()
plt.xlabel("Tenforflow model prediction")
plt.ylabel("Arduino model prediction")
plt.grid()
plt.show()

# The arduino and python model have the same result.

Other models

You can use the same logic to convert other models to ARDUINO C++

Function Name	Model Description
convert_DecTree_To_C	Decision Tree Regression Model
convert_RandForest_To_C	Random Forest Regression Model
XGBOOST_to_CPP	XGBoost Regression Model
tf_model_to_arduino_code	DNN TensorFlow Model

For more information you can see the home page
https://bouz1.github.io/fils/MLModelToArduinoCpp/MLModelToArduinoCpp.html
Or the github repository
https://github.com/bouz1/ML-Model-To-Arduino-Cpp