numpy-complete

NumPy-Complete is a neural network to recognise handwritten digits, trained using only NumPy(No TensorFlow, Pytorch) on the MNIST database.

Installation

1. Make sure the following packages are installed:

pip install numpy pandas matplotlib opencv-python

2. Clone the repository

git clone https://github.com/fringewidth/numpy-complete.git

Usage

1. Modify input-image.png to a custom handwritten digit. The image must be 28px $\times$ 28px

2. Run run-model.py

Features

• Recognises a handwritten digit from a 28 $\times$ 28 pixel grid using a feed-forward neural network with two hidden layers, with an accuracy of 90.41%.
• Makes use of only NumPy for numerical processing. All functions ar custom implemented and modifiable.
• Ability to run custom input.
• Training notebook available for customizing the training process. See train.ipynb for details on architecture and implementaion.