MNIST Classifier with Modified LeNet-5

This repository contains a PyTorch implementation of a MNIST digit classifier based on a modified LeNet-5 architecture. The model has slight adjustments compared to the original LeNet-5, including dropout layers after fully connected layers and max pooling instead of average pooling to improve generalization.

Features

• Accuracy: 99.27%
• Architecture: LeNet-5 with small modifications for better performance.
• Dropout: Added after FC1 and FC2 to reduce overfitting.
• Max Pooling: Replaces average pooling in original LeNet-5.
• Training Utilities:
  • Adam optimizer with weight decay
  • OneCycleLR scheduler with cosine annealing
  • Cross-entropy loss
• Data Handling:
  • MNIST dataset with 90/10 train-validation split
  • DataLoader with configurable batch size