Neural Network Inference on ESP8266

This project demonstrates running a trained neural network (MNIST digit classifier) on an ESP8266 microcontroller. The neural network is trained in Python, converted to C-compatible format, and deployed on the ESP8266 to perform real-time digit classification.

Project Overview

The project consists of three main components:

1. Python-based neural network training and weight conversion
2. ESP8266 firmware for inference
3. Client-server communication over WiFi

Hardware Requirements

• ESP8266 Generic Module
• USB cable for programming and power
• WiFi network access

Software Components

Python Training and Conversion

• convert.ipynb: Jupyter notebook for training the neural network and converting weights to C format
• predict.ipynb: Jupyter notebook for testing the model
• Generated weight files:
  • mnist_model_weights.h: Original weight file
  • mnist_model_weights_int.h: Quantized weights for ESP8266

ESP8266 Firmware

• NNonEsp32.ino: Main Arduino sketch implementing:
  • WiFi server setup
  • Neural network inference
  • Client communication handling

Network Architecture

The implemented neural network has the following structure:

• Input layer: 784 neurons (28x28 MNIST image)
• Hidden layer: 32 neurons with ReLU activation
• Output layer: 10 neurons (digits 0-9) with softmax activation

Setup Instructions

Make Sure the folder name is NNonEsp32

1. Configure WiFi Settings

   • Open NNonEsp32.ino
   • Update the WiFi credentials:

     const char* ssid = "YOUR_WIFI_SSID";
     const char* password = "YOUR_WIFI_PASSWORD";

2. Upload the Firmware

   • Open the Arduino IDE
   • Select the ESP8266 board
   • Upload the NNonEsp32.ino sketch

3. Connect to the Server

   • The ESP8266 will create a TCP server on port 5000
   • Note the IP address displayed in the Serial Monitor

Communication Protocol

The server expects data in the following format:

1. 2-byte header indicating the number of pixel pairs
2. For each pixel pair:
   • 2 bytes: pixel index
   • 1 byte: pixel value (0-255)

The server responds with a JSON object containing:

• predicted_label: The classified digit (0-9)
• probabilities: Array of probabilities for each digit

Performance Considerations

• Weights are quantized to 8-bit integers for memory efficiency
• The model uses a scale factor of 0.01 for weight conversion
• Inference is performed in real-time on the ESP8266

Dependencies

• ESP8266WiFi library
• Arduino core for ESP8266

Acknowledgments

• MNIST dataset
• ESP8266 community
• Arduino platform