This was my Keras implementation of a CNN which classifies the MNIST dataset. I also submitted this code to the related Kaggle competition. The Jupyter Notebook includes my sources as well as extensive comments/documentation in order to explain what each line of code does.

Developed in September/October of 2018. Updated 11/1/2018 for a presentation to the UDSC at the University of Rochester.

My main goal with this project was introduce myself to the Keras library, as well as to submit something of my own through the Kaggle submission process. I added methods to save/load Keras models into the Notebook in order to save ongoing training progress on a given Keras model.

I coded the notebook by primarily following along a tutorial for creating a CNN in Keras for classifying MNIST the dataset. I altered the layers of my CNN from the tutorial's suggestion, simply to add additional possible accuracy to the network. During this process, I eventually had to transition the MNIST import statements to Kaggle's format, which uses .csv files. Near the end of my work, someone suggested that I add the ability to save and reload a trained model, which I then did. The Github repo contains one such model that I trained and reloaded a few times to get the CNN up to 50 epochs over the dataset.

A Jupyter Notebook as well as an ongoing Keras CNN model capable of additional training. My Kaggle submission is in the Top 32% of submissions. My model achieved 99.157% accuracy after 30 epochs. I also presented the final notebook as a Keras workshop during a meeting of the University of Rochester Undergraduate Data Science Council.