Neural Network Simulation

Pygame visualization of a neural network while it trains on MNIST.

The window shows the graph on the left (nodes, edges, live weights) and a sidebar on the right with the current prediction, the actual digit, and training accuracy. The trainer is a small fully connected network written in NumPy; a separate PyTorch MNIST classifier lives in the same repo for comparison.

Features

• Layered graph of nodes and edges (784 → 10 → 10)
• Weights drawn inside nodes and updated as training runs
• Output node highlighted for the model's current prediction
• Sidebar: predicted digit, actual digit, and accuracy
• Pan and zoom the graph (mouse-centered zoom)
• One gradient-descent step about once per second

Architecture

Visualizer (package/sim): Network owns Layers of Nodes connected by Edges. Camera pan/zoom maps world coordinates to the screen.

From-scratch trainer (package/mnist): two-layer network (784 → 10 ReLU → 10 softmax), forward pass, backprop, and gradient descent in NumPy. Simulation copies averaged weights onto the graph and runs one training step per second.

PyTorch (package/pytorch): a 784 → 128 → 64 → 10 MLP trained with Adam, plus a short tensor tutorial.

Requirements

• Python 3.10+
• Windows for the main sim (pywinstyles is used for the window title bar)
• MNIST train.csv from Kaggle Digit Recognizer

Place the CSV at:

package/data/train.csv

Setup

From the repository root:

python -m venv .venv
.venv\Scripts\activate
pip install pygame pywinstyles numpy pandas matplotlib

Optional, for the PyTorch scripts:

pip install torch torchvision

Run

Always run from the repository root so package/ imports and package/data/train.csv resolve.

Simulation (NumPy trainer + Pygame):

python -m package.sim.main

PyTorch MNIST classifier:

python -m package.pytorch.mnist_classifier

Camera pan/zoom sandbox (no network):

python -m package.sim.test_camera

Controls

Input	Action
Left-drag on the graph	Pan
Mouse wheel	Zoom toward the cursor (1×–10×)
R	Reset pan
0	Reset zoom to 1×

Project layout

package/
  sim/          Pygame graph, camera, and main loop
  mnist/        NumPy MNIST trainer (forward, backprop, GD)
  pytorch/      PyTorch classifier and tensor notes
  data/         train.csv (not committed; download from Kaggle)