Project Overview

The mnist_digit_classifier.py script builds, trains, and evaluates a deep learning model for image classification. It's a foundational example for understanding CNNs with PyTorch.

Script Description: scripts/mnist_digit_classifier.py

This script performs the following key steps:

1. Data Loading and Preprocessing:
   • Downloads the MNIST dataset (training and test sets) from torchvision.datasets.
   • Automatically transforms images into PyTorch tensors using ToTensor().
   • Creates DataLoader objects for efficient batch processing, with shuffling for the training data.
   • Prints the shape of a sample batch to verify data dimensions.
2. Device Configuration:
   • Automatically detects and utilizes a GPU (CUDA) if available; otherwise, it defaults to CPU.
3. Neural Network Definition:
   • Defines a NeuralNetwork class inheriting from nn.Module.
   • It consists of:
     • A conv_stack (convolutional layers) with nn.Conv2d, nn.ReLU activations, and nn.MaxPool2d for feature extraction and downsampling. This stack includes three convolutional blocks.
     • A classifier (fully connected layers) with nn.Flatten, nn.Linear layers, nn.ReLU activations, and nn.Dropout for classification and regularization.
   • The forward method defines the pass through the network.
4. Model Initialization:
   • Instantiates the NeuralNetwork model and moves it to the selected device (GPU/CPU).
5. Loss Function and Optimizer:
   • Uses nn.CrossEntropyLoss() as the loss function, suitable for multi-class classification.
   • Employs the Adam optimizer with a learning rate of 1e-3 and weight_decay for regularization.
6. Training and Testing Functions:
   • train(dataloader, model, loss_fn, optimizer): Iterates over the training data, performs forward and backward passes, updates model weights, and tracks training loss and accuracy per epoch.
   • test(dataloader, model, loss_fn): Evaluates the model's performance on the test set (without gradient calculation), tracking test loss and accuracy.
7. Model Training Loop:
   • Trains the model for a specified number of epochs (default 8), calling the train and test functions for each epoch.
   • Prints progress including loss and accuracy for both training and testing.
8. Model Saving:
   • Saves the trained model's state dictionary to mnist_model.pth.
9. Results Visualization:
   • Plots the training and test loss over epochs.
   • Plots the training and test accuracy over epochs.
   • Displays 9 random test images along with their predicted and true labels to visually assess performance.
10. Confusion Matrix:
    • Generates and plots a confusion matrix using sklearn.metrics.confusion_matrix to provide a detailed breakdown of correct and incorrect classifications for each digit.

Dataset

This project uses the MNIST dataset, which is a large database of handwritten digits commonly used for training various image processing systems. It consists of:

• 60,000 training images.
• 10,000 test images.
• Each image is a 28x28 pixel grayscale image of a handwritten digit (0-9).