A custom neural network written from scratch in Python with no external libraries (only exception being pillow, used for converting images to pixel arrays).
The model takes in a 10x10 image representing any digit as an input, and outputs the most likely digit to be represented in the image. The whole architecture was written in less than a day in Python, but is not meant to be optimized at all. The goal of this project was to challenge myself and see how fast I could write a custom neural network from scratch, with no prior experience in machine learning.
The model has 4 layers:
- An input layer consisting of 100 neurons, representing a 10x10 image (thus a total of 100 pixels, one neuron per pixel). The activation of a neuron in the input layer will be 1 if the corresponding pixel is part of the digit's shape to indentify, else it will be 0
- Two hidden layers each consisting of 81 neurons
- An output layer, where the most active neuron will be the recognized digit
The model utilizes a custom dataset I have made myself using a custom pixel art drawer created with Pygame. The images in the dataset are black and white and pixel-perfect, thus making the model bad at recognizing very blurry / colorful images.
You can use your own dataset (or add your own images to the existing one) as long as you modify the img_to_pixels function to convert your images to the format required that is specified in the input layer's architecture.
The model is currently trained with a learning rate of 0.01 across 1000 epochs. I heavily recommend the usage of PyPy when running train.py to substantially speed up this process.
The model is not very accurate as of now (~78%) due to a lack of data, even though the dataset is constantly growing.
The model is quite fast, taking less than 40 milliseconds in total to read the weights, run the model, and print the result.