C++ Web Server and Neural Networks

Purpose

• I wanted to get a more fundamental understanding of neural networks and web servers, which we use all the time using abstracted frameworks like Node.js or PyTorch
• This web server take requests and respond with neural network data (weights, biases, loss, prediction) after each epoch
• A frontend will make requests to the server, and display the data, visualizing how the various parameters of a neural network change over time.

Versions

• On the main branch is the implementation of the standard feed forward neural network
• I tried with a convolutional neural network, and that needs some work. It is on the cnn branch

Saving Training and Inference Data

• After each epoch, the weights and biases are saved in binary
• After inference, the output probabilities are also saved

Running

• Build & Run Training/Inference:

  (cd backend/networking/NN && make clean && make && ./train.out && ./inference.out)

• Build & Run Server:

  (cd backend/networking && make clean && make && ./server.exe)

• Run Frontend:

  (cd frontend && npm run dev)

Socket and Servers

A socket is a software endpoint that enables communication between two computers over a network.

Socket classes

• SimpleSocket: Creates a raw socket.
• BindingSocket
  • Inherits from SimpleSocket
  • Binds the socket to an IP and port.
  • A raw socket does not do anything on its own. Binding allows clients to know where to connect.
• ListeningSocket
  • Inherits from BindingSocket
  • Bound sockets cannot accept connections until it's actively listening
  • listen() sets up a queue for incoming connections
• ConnectingSocket: A client-side socket that connects to a server. This is not used right now as I'm just connecting to the server from my Next.js frontend

Server Classes

• SimpleServer: Owns the listening socket and handles client requests.
• TestServer: A specific implementation of SimpleServer that processes requests.

Web Server Flow

• Create a socket
• Bind it to an address and port
• Listen for incoming connections
• Accept client connections and respond

MNIST Image File Structure

• Header (16 bytes): Magic number, image count, rows, columns (big-endian).
• Pixel data of the actual images follows the header.

Data Download

1. Kaggle API Token: Download from Kaggle.
2. Download Dataset:

   kaggle datasets download hojjatk/mnist-dataset