Edge ML: TensorFlow Lite Micro Implementation for ESP8266

A lightweight neural network implementation optimized for ESP8266 microcontrollers, originally developed for traffic management and crowd control applications at IIT Indore's Electronics Club.

Overview

This repository contains a quantized TensorFlow Lite model converted for edge deployment, implementing a two-layer neural network (784-128-10 architecture) trained on the Fashion MNIST dataset. Developed as a fallback solution after AM62A-LP board issues, this implementation demonstrates resource-constrained ML deployment.

Key Features

• Memory-efficient design: 1.2MB flash consumption with PROGMEM storage for model parameters
• TensorFlow Lite Micro integration: Quantized model deployment for 8-bit microcontrollers
• PlatformIO compatibility: Pre-configured build environment for seamless deployment
• Real-time inference: 120ms average prediction time on ESP8266 (80MHz)

Repository Structure

Edge-ML/
├── CONTRIBUTORS.md
├── README.md
├── platformio.ini
└── src/
    ├── NeuralNetwork.cpp
    ├── NeuralNetwork.h
    ├── dense1_bias.h
    ├── dense1_weights.h
    ├── dense2_bias.h
    ├── dense2_weights.h
    ├── main.cpp
    ├── model_data.cc
    ├── model_data.h
    └── test_images.h

Installation

1. Prerequisites:

   • PlatformIO Core (VSCode extension recommended)
   • ESP8266 board support

2. Clone repository:

git clone https://github.com/Stonky-Boi/Edge-ML.git
cd Edge-ML

3. Build & Upload:

pio run -t upload

Usage

#include "NeuralNetwork.h"

void setup() {
  Serial.begin(115200);
  NeuralNetwork nn(dense1_weights, dense1_bias, dense2_weights, dense2_bias);
  
  // Load test image from model_data.cc
  std::vector input(test_image, test_image + 784);
  
  auto prediction = nn.predict(input);
  Serial.print("Predicted class: ");
  Serial.println(std::distance(prediction.begin(), 
                 std::max_element(prediction.begin(), prediction.end())));
}

Contributors

See CONTRIBUTORS.md for the complete list of developers and mentors.

Training Reference

For model training and quantization: Fashion MNIST Training Notebook