Python-Perceptron

An Basic implementation of the perceptron, the build block a neural net.

Table of contents

• Installation
• Example
• Docs
• Contributing

Installation

pip install PyPerceptron

Example

Here's how to instanitate the Perceptron

from Perceptron.perceptron import Perceptron

p = Perceptron(number_inputs, learning_rate, Activation_fn, Loss_fn)

from Perceptron.perceptron import Perceptron
from Perceptron.functions.activationFunctions.heaviside import Heaviside
from Perceptron.functions.lossFunctions.quadratic_loss import QuadraticLoss

dataset = [[2.7810836, 2.550537003, 0],
               [1.465489372, 2.362125076, 0],
               [3.396561688, 4.400293529, 0],
               [1.38807019, 1.850220317, 0],
               [3.06407232, 3.005305973, 0],
               [7.627531214, 2.759262235, 1],
               [5.332441248, 2.088626775, 1],
               [6.922596716, 1.77106367, 1],
               [8.675418651, -0.242068655, 1],
               [7.673756466, 3.508563011, 1]]


p = Perceptron(2, 0.1, Heaviside(), QuadraticLoss()) # number of inputs, learning rate, activation function, loss funciton
p.train(dataset, 3, 30)

for d in dataset:
    assert p.evaluate(d[0], d[1]) == d[2]

To find out more about the math behind the perceptron, check out the notebook with the fully explanation.

If you wanna see more about how to use the perceptron checkout the demos.

Documentation

• Perceptron
• Function
• Activation Function
• Loss Function

Perceptron

Here's how to create a perceptron instance

from Perceptron.perceptron import Perceptron

p = Perceptron(no_inputs: int, lr: float, act_fn:Function, loss_fn:Function)

• no_inputs: are the number of inputs of the perceptron
• lr: the learning rate of the preceptron
• act_fn: the activation function of the perceptron
• loss_fn: the loss function of the perceptron

Note both act_fn and loss_fn must be instance of the class Function

Attributes

• no_input (int) : the number of inputs of the perceptron
• bias (float): the bias of the perceptron
• weights (list): the weights of the perceptron
• act_fn (Function): the activation function of the perceptron
• loss_fn (Function): the loss function of the perceptron
• lr (float): the learning rate of the perceptron

---

p.evaluate(inputs: list)

Return the prediction of the perceptron

• inputs: the inputs list that the vector have to evaluate, and the last element must be the prediced value

Note: the length of the inputs must be len(p.no_inputs) + 1

---

p.train(training_data: list, mini_batches_size: int, n_epoch=30: int)

Train the perceptron using mini batch stocastic gradient descend

• training_data: the data used to train the preceptron that will be divide in mini batches
• mini_batches_size: the size of the mini batch
• n_epoch: number of iteration

Note: the length of the training_data must be len(p.no_inputs)

---

Function

The class function is just a abstract class that represent a mathematical function

The only attribute that it has is:

• is_diff: True if the function is differentiable false otherwise

---

fn.compute(a) -> float

Abstract method that every child of the Function class implement, and it return the computed value of the given function

• a: the input of the function, the type can vary on the function (most of the time is an int or a tuple)

---

fn.compute_derivative(a) -> float

Abstract method that every child of the Function class implement, and it return the computed value of the derivative of the given function

• a: the input of the function, the type can vary on the function (most of the time is an int or a tuple)

---

Activation Function

The activation function already included in the package are:

• Heaviside (1/0 step function)
• Identity (the identity function)
• Sign
• ReLU
• Leaky ReLU
• Smooth ReLU
• Sigmoid
• Tanh (Hyperbolic tangent)
• Softmax

from Perceptron.functions.activationFunctions.sigmoid import Sigmoid

fn = Sigmoid()

print(fn.compute(0)) # 0.5

Loss Function

The loss functions already included in the package are:

• Quadratic Loss
• Cross Entropy Loss

Contributing

Feel free report issues and contribute to the project, making it better.

Author

Paolo D'Elia

License

MIT