Neural-Network-in-Python-using-Numpy

This is the implementation of a fully customizable neural network with arbitrary no. of hidden layers using naught but NumPy, in Python.

Documentation for the train module

• NN: A Network that uses Sigmoid activation function. Methods defined here:
  • add_layer(self, n_nodes, output_layer=False) : Adds a layer of specified no. of output nodes. For the output layer, the flag output_layer must be True. A network must have an output layer.
  • sigmoid(self, z): Calculates the sigmoid activation function.
  • predict(self, x, predict=True, argmax=True, rand_weights=False): Performs a pass of forwrd propagation on x. If predict is set to True, trained weights are used and predictions are returned in a single vector(if argmax is set to True) with labels from 0 to (n_classes - 1).
  • cost(self, x, y, lamda=0): Calculates the cost for given data and labels, with trained weights.
  • fit(self, data, labels, test=[], test_labels=[], alpha=0.01, lamda=0, epochs=50): Performs specified no. of epoches. One epoch = one pass of forward propagation + one pass of backpropagation.
• plot_confusion_matrix(df_confusion, title='Confusion matrix', cmap=plt.cm.gray_r): Used to plot confusion matrix. (Internal function)