Neural Network

About

Light neural network library that can create, train, and test different models.

Usage

Classes

• Model

  • Members

    • layers ~ all layers within current model object

  • Methods

    • initialize(self, n, seed) ~ initializes all weights and biases within all layers. n is # of units in input.
    • forward_propagate(self, x_in) ~ takes a training input and propragate through all layers.
    • back_propagate(self, m, x_in, y_out, alpha) ~ performs backpropagation across all layers for a single epoch. y_out holds actual y values for current training set.
    • fit(self, X_train, Y_train, X_test, Y_test, alpha, epochs, seed, reg_rate) ~ trains data across all epochs. Returns training cost, test cost history, train accuracy, and test accuracy
    • predict(self, x_test) ~ returns forward_propagate results for x_test
    • summarize(self) ~ summarizes the model

• Layer

  • Members

    • activation ~ holds activation function class for current layer
    • units ~ number of neurons in current layer
    • neurons ~ np.array holding activation values for each neuron
    • W_l ~ weight matrix for current layer
    • B_l ~ bias vector for current layer

  • Methods

    • initialize(self, prev_layer_units) ~ initializes weights and biases within current layer.
    • feed_forward(self, a_in) ~ forward propagates through current layer. Weighted sum and activation calculations are performed here.
    • back_prop(self, del_J_z, prev_layer, alpha) ~ performs backprop through current layer. note, del_J_z represents partial derivative of cost W.R.T weighted sum, z.
    • summarize(self) ~ summarizes current layer

Example Usage