This project uses Object Oriented Programming for modularity. The core of this project is the class NeuralNetworkClassifier whose constructor can take the following parameters:

1. no_of_neurons - A list which represents the architecture of the neural network. (i.e) number of neurons in each of the hidden layer
2. n_class - Number of classes
3. alpha - L2 regularization weight decay parameter. Default value is 0 which means no L2 regularization.
4. activation - Activation function to be used. Default value is 'sigmoid'.
5. output - Output function to be used. Default value is 'softmax'
6. loss - Loss function to be used. Default value is 'cross_entropy'
7. optimizer - Optimizer to be used. Default is None. Supported optimizers are sgd, momentum, nesterov, rmsprop, adam, nadam

This class has three useful methods namely fit, predict and accuracy.

fit(x, y, batch_size=BATCH_SIZE, epochs=100, eta=0.01, weight_initializer = None)

To train the neural network classifier.

x - training data
y - training labels
batch_size - batch size used for training
epochs - number of epochs until which the model will be trained
eta - learning rate
weight_initializer - weight initialization method to be used. By default it is None for which weights are initialized randomly. 'Xavier' method is supported.

Returns None

predict(X_test)

To predict the classes for given data

X_test - Data for which predictions are to be made

Returns an array of predictions

accuracy(X_test, y_test)

To calculate the accuracy of the model given some data

X_test - Data for which the accuracy of the model need to be evaluated

y_test - Corresponding class labels of the data

Returns the accuracy score

Example:

model = nn.NeuralNetworkClassifier([64, 64, 64], 10, alpha = 0.05, optimizer = 'adam')

model.fit(X_trainval, y_trainval, batch_size = 256, epochs = 150, eta = 0.001, 
          weight_initializer = 'Xavier')

y_predictions = model.predict(X_test)
acc = model.accuracy(X_test, y_test)

This example creates a neural network with 3 hidden layers each having 64 neurons with 10 output classes. The model is trained by calling the fit method.