nait 3.0.0

Neural network library

---

Nait

---

Neural network module

Simple but powerful

Network - class

---

Description

class for creating a neural network

containing everything needed for training, using and testing a neural network

Syntax

network()

Values

• weights

• biases

• activation_function

Methods

• train

• predict

• save

• load

• evaluate

• values

train - method

---

Description

function for training a network to improve at a given task

comes with a large variety of customization options for the training process

Syntax

train(x=[[1, 1, 1, 1]], y=None, structure=(4, 4, 4), activation_function="linear", generate_network=True, learning_rate=0.01, batch_size=10, sample_size=None, loss_function=None, epochs=100, backup=False, verbose=True)

Arguments

• x - training inputs

• y - training outputs

• structure - array of layer sizes

• activation_function - function applied to the output of each layer (linear / relu / step / sigmoid / leaky_relu)

• generate_network - if the program should generate a new structure or try to train the existing one

• learning_rate - how drastically the network will try to improve

• batch_size - how many variations the network will try each epoch

• sample_size - how much of the dataset the network will train on each epoch - if set to none the network will use the whole dataset

• loss_function - the function used for calculating loss - if set to none the program will use output to y difference - more information lower on the page

• epochs - number of epochs the network should train for

• backup - if the network should backup itself while it trains

• verbose - if the network should output additional information to the screen while training

predict - method

---

Description

function for passing a single input array through the network

Syntax

predict(input)

Arguments

• input - input array

save - method

---

Description

function for exporting the network into a json file

which can late be imported with load()

Syntax

save(file="model.json")

Arguments

• file - where to save

load - method

---

Description

function for importing a network json file exported with save()

Syntax

save(file="model.json")

Arguments

• file - where to save

evaluate - method

---

Description

function to get a loss and average loss of a network

with completely new inputs and outputs without changing the network

Syntax

evaluate(x=[[1, 1, 1, 1]], y=None, loss_function=None, output_to_screen=True)

Arguments

• x - testing inputs

• y - testing outputs

• loss_function - the function used for calculating loss - if set to none the program will use output to y difference - more information lower on the page

• output_to_screen - if the network should output the result of the evaulation to the console

values - method

---

Description

function for printing the network values

in a readable format

Syntax

values()

Additional information

Creating a loss function

---

to create a loss function for the 'train' and 'evaluate' function

you create a function which takes three arguments: forward, x, y

• forward - a class that can be used pass an input through the network - usage: forward.predict(input)

• x - the wanted inputs

• y - the wanted outputs

Network structure

---

you can create a network structure with the 'train' function in the 'structure' argument

you then pass in a tuple with the layer sizes that you want, example: (2, 6, 3)

the first and last values are the input and output size, so they have to mach the size of x and y

Nait v2.0.4 - Change Log

---

• Added more structure control

• Added 'values' function

• Added 'sample_size' argument to the train function

• Added 'verbose' argument to the train function

• Added function documentation

• Changed display

• Removed 'layer_size' and 'layers' from the train function

---