mnist

Example code for training on the MNIST dataset using PyTorch, with CUDA enabled.

The MNIST dataset consists of black/white images that display one of ten handwritten numbers (0 to 9). These images are represented by a 28x28 matrix, with each matrix element depicting its corresponding grayscale intensity.

The training set has 60,000 images and the test set has 10,000 images.

This repository includes:

• a custom module mnist_tools.py, which defines the LeNet5 convolutional neural network, my custom convolutional network, the train function, and the test function
• a jupyter notebook mnist.ipynb, which performs training, validation, and testing of the dataset (using LeNet5)
• a python script lenet_main.py which also performs training, validation, and testing of the dataset, with the added bonus of user input and an option to save the fully trained model
• a python script custom_main.py — similar to lenet_main.py, except using a custom CNN that I developed
  • achieves similar accuracy of ~98% for a batch size of 32, trained through 5 epochs

The LeNet5 architecture was based off this wikipedia article: https://en.wikipedia.org/wiki/LeNet