Fully Connected Neural Networks

Fully Conected Neural Networks module.

Contains Dense and Dropout layers along with ReLU and SoftMax activation functions.

This module is based on the syntax that Tensorflow/Keras uses.

from NN import Model
from NN.layers import Dense, Dropout
from NN.activations import ReLU, SoftMax
from NN.loss import CatCrossEntr
from NN.optimizers import SGDOptimizer

# the X and y datasets
X = a_dataset
y = another_dataset

# create loss and optimizer objects
loss = CatCrossEntr()
optimizer = SGDOptimizer(lr=0.1, decay=1E-3, momentum=0.6)

# create Model object
model = Model()
model.add(Dense(2, 32))
model.add(ReLU())
model.add(Dropout(rate=0.2))
model.add(Dense(32, 3))
model.add(SoftMax())

# finalize the model
model.finalize(loss=loss, optimizer=optimizer)

# train the model
model.fit(X, y, epochs=300,validation_data=None, verbose=2)

The Model object

class mapyl.NN.Model

Model for sequential, fully connected layers

Returns:

self: The Model class

add(layer)

adds a layer to the layer list

finalize(*, loss, optimizer)

Finalizes the model, adds loss and optimizer

parameters:

loss (Loss): Loss object

optimizer (Optimizer): Optimizer object

Returns:

self: the finalized Model object

fit(X, y, epochs, validation_data=None, verbose=0)

trains and validades the Model instance

Parameters:

X (ndarray): ndarray of shape (num_samples, num_features) of the input

y (ndarray): ndarray of shape (num_samples, num_features) of the input

epochs (int): The number of epochs

validation_data (tuple): Validation data of shape (X ndarray , y ndarray). Defaults to None.

verbose (int): The verbose, can be 0, 1 and 2. Defaults to 0.

Returns:

self: The fitted instance

The Layer objects

class mapyl.NN.layers.Dense(num_inputs, num_neurons)

Dense layer instance

Parameters:

num_inputs (int): The number of inputs (neurons of the previous layer or the model inputs)

num_neurons (int): The number of neurons of the layer (also teh number of outputs)

class mapyl.NN.layers.Dropout(rate)

Dropout layer class, deactivates neurons according to rate

Parameter:

rate (float): the percentage of deactivated neurons

The Optimizer objects

class mapyl.NN.optimizers.SGDOptimizer(lr=0.1, decay=0, momentum=0)

Stochastic Gradient Descent optimizer

Parameters:

lr (float): The learning rate decay (float): The decay of the learning rate momentum (float): The momentum of the learning rate

The Loss objects

class mapyl.NN.loss.CatCrossEntr

Categorical Cross Entropy loss object

The Activation Function objects

class mapyl.NN.activations.ReLU

Rectified Linear Unit activation funtion

class mapyl.NN.activations.SoftMax

SoftMax activation function, classifies the inputs and returns confidence scores for each class