NeuralNetworks

This project builds and compares multiple neural network architectures for CIFAR-10 image classification using PyTorch. It trains each model, evaluates test performance, and generates visual artifacts so you can quickly see how architecture choices affect accuracy and prediction behavior.

PyTorch-based CIFAR-10 image classification project that trains and compares three models:

• MLP (fully connected baseline)
• SimpleCNN (2 convolution blocks)
• ImprovedCNN (deeper CNN with dropout)

The training script saves performance plots and prediction examples for each model, making it easy to compare behavior and accuracy.

Project Structure

• cifar.py - main training and evaluation script
• requirements.txt - Python dependencies
• results_*.png - training loss and test accuracy plots per model
• examples_*.png - one correct and one incorrect prediction example per model
• results.pdf - exported report/summary artifact

Dataset

This project uses the CIFAR-10 dataset from torchvision.datasets.CIFAR10.

• Training set and test set are downloaded automatically to ./data on first run.
• Images are normalized with mean/std (0.5, 0.5, 0.5).

Installation

Use Python 3.10+ (recommended) and install dependencies:

pip install -r requirements.txt

Run Training

python cifar.py

The script will:

1. Load CIFAR-10
2. Train MLP, SimpleCNN, and ImprovedCNN
3. Evaluate each model on the test set
4. Save:
   • results_MLP.png, results_SimpleCNN.png, results_ImprovedCNN.png
   • examples_MLP.png, examples_SimpleCNN.png, examples_ImprovedCNN.png

Notes on Training Logic

• Optimizer: SGD (lr=0.001, momentum=0.9)
• Loss: Cross-entropy
• Maximum epochs: 10
• Early stop condition: training stops early if average training loss increases between epochs

Dependencies

• torch
• torchvision
• matplotlib