Neural network from scratch with numpy, and matplotlib for visualization

Tested on python 3.10 - the package may not function as intended on other python versions Pull from github, then create a python 3.10.9 venv and pip install requirements.txt

$ mkdir <dir>
$ cd <dir>
$ git clone https://github.com/tayo4pf/axon.git
$ python3.10 -m venv <env>
$ pip install -r requirements.txt

The API offers 4 modules for use - activation, loss, Network, and optimizer

Activation functions are defined using the module name

import axon

linear = axon.activation.Identity
relu = axon.activation.Relu
leakyrelu = axon.activation.LeakyRelu
sigmoid = axon.activation.Sigmoid
softmax = axon.activation.Softmax

Loss functions are similarly defined using the module name

mse = axon.loss.MSE
logistic = axon.loss.logistic

Optimizers can be defined using the enum from the Optimizer module

sgd = axon.Optimizer.SGD
sgd_with_momentum = axon.Optimizer.SGD_WM
nag = axon.Optimizer.NAG
ada_grad = axon.Optimizer.AdaGrad
ada_delta = axon.Optimizer.AdaDelta
adam = axon.Optimizer.AdaM

Networks can be instantiated with their shape, the activations for each layer, and the loss function for the network

nn = axon.Network([4, 6, 1], [leakyrelu, softmax], logistic)

They can then be trained on a dataset with

Network.train(data, labels, optimizer, learning_rate, momentum = 0.9, epsilon = 0.01, gamma = 0.8, fresh = True)

Performance on a test set can then be computed and input values can be predicted with the model

losses = nn.train(data, labels, adam, 0.03)
perf = nn.test(data, labels)
predictions = nn.predict(inputs)

Model weights and biases can be visualized

nn.visualize()

Models can be saved and loaded from csvs

filename = 'model'
nn.write_to(filename)
copy = axon.Network.read(filename)