I have built a neural network from scratch in Python using numpy and object-oriented programming. This network takes a 28x28 pixel image of a handwritten digit and correctly classifies the digit as 0, 1, 2, 3, 4, 5, 6, 7, 8, or 9
This Project is written in Python and requires the use of numpy, pillow, math, and pathlib. You can install these via an IDE or with "pip install" in the command line.
SUMMARY:
Utilizes machine learning algorithms/concepts like forward/backpropagation, activation functions, and loss calculation for intelligent learning. Trained the network on 60,000 images, which allowed the network to classify 97.64% of 10,000 new images correctly it had never seen before.
The 3 main files to run are: train.py, test.py, and customTest.py
train.py: Trains the neural network with images extracted from the .npz file in the data folder. First Check to see if model.txt currently has data. If it does, the script will load the model weights into the network and start training based on the provided model weights. Otherwise, the weights between the layers in our network will be initialized with random weights and trained from scratch. Clear the model.txt file to retrain the network from the beginning Forward propagates pixel data through the network while implementing an Activation function at the end of each neuron. Calculates loss at the end of the network based on the correct answer provided by training data, and back propagates with a learning rate of 0.01 Prints the learning progress of the network in the console.
test.py: Extracts different images from the .npz file in the data folder than train.py These images are new images that the network has never seen before Loads the model.txt weights into the neural network if the data exists. If the data doesn't exist, the script will run the test images through a randomized network Running test images through randomized networks shows that my ML algorithms and neural network work!! At the conclusion, the script prints the Accuracy of the model
customTest.py: Does the same as test.py, HOWEVER, it uses custom images of my handwritten digits