NumpAI

A hard-coded NumPy implementation of a Convolutional Neural Network (CNN) for the MNIST dataset.

4730 Machine Learning Fall 2022 Project I Repository -- We misread the assignment and thought we were supposed to implement a CNN from scratch. We were not. But now we have a hard-coded NumPy implementation of a CNN for the MNIST dataset. It is not the most accurate (~70%), but it works.

This is a collaborative project between Mathew Pellarin and I.

Running the Project

There are two CNNs implemented here:

1. main.py is hard coded in NumPy. It is less accurate, but works enough that we are happy with it, and is completely done from scratch.
2. phase_1.ipynb, which uses a sequential Keras model. It is more accurate(>98%), but is not completely done from scratch. This is an example of how to use Keras to implement a CNN.

Installing Requirements

To install the requirements, run the following command:

pip install -r requirements.txt

Notes on the dataset

• These images are 28x28 pixels, grayscale, and centered.