This project demonstrates the training and evaluation of a neural network model on the MNIST dataset, a popular dataset of hand-written digits. The goal is to classify each image into one of the ten possible digit classes (0-9) using deep learning techniques. The project includes data loading, dataset creation, model definition, and training using k-fold cross-validation.

Before running the code, make sure you have the necessary libraries installed. You can install them using pip (ideally in a python virtual environment or using conda):

pip install torch scikit-learn imbalanced-learn matplotlib tqdm

The project is structured as follows:

• README.md: This document provides an overview of the project.
• mnist.py: The Python script containing the code for data loading, dataset creation, model definition, and training.
• data/: A directory that should contain the MNIST dataset files (training.pt and test.pt).

The project begins by importing the necessary libraries, including PyTorch, scikit-learn, imbalanced-learn, matplotlib, and tqdm. It also checks for the availability of a CUDA-compatible GPU and sets the device accordingly.

The MNIST dataset is loaded from the provided data files training.pt and test.pt.

A custom CTDataset class is defined to preprocess the dataset. It normalizes the pixel values to the range [0, 1] and provides a convenient way to access the data.

The neural network architecture (MyNeuralNet) is defined, consisting of three connected layers with ReLU activation functions. The input size is 28x28 pixels (784), and the output size is 10 (for the ten digit classes).

The train_model_kfold function performs k-fold cross-validation training on the dataset. It uses Stochastic Gradient Descent (SGD) optimization with a learning rate of 0.01 and Cross-Entropy Loss. For each fold:

1. Train and validation subsets are created.
2. Data oversampling is performed using RandomOverSampler to handle class imbalance.
3. Dataloaders are created with a batch size of 20.
4. The model is trained for a specified number of epochs.
5. Validation accuracy is calculated and printed for each fold.

The script provides validation accuracy for each fold during k-fold cross-validation. The final results include accuracy scores for all folds, helping assess the model's performance.

This plot displays predicted digits for a subset of images from the MNIST dataset.