CMSC421 Project 3 Solved

Description

What you’ll need to do

• Implement and train a basic neural network with backpropagation
  • We’ll give you part of the code, you need to fill in the details
• We’ll give you a ZIP archive containing two files:
  • py: the class and method outlines

○    studenttest.py: code to train and test your neural network

• py contains some, but not all, of the necessary methods
  • You’ll need to write the others yourself

activation(z):

• This is the activation function used in the feed-forward compution of the network.
• The included function is the sigmoid, but you can change this as you see fit.

Here is a graph showing visually how the sigmoid function works:

sigderiv(z):

This is the derivative of the sigmoid function

• You’ll need to use it to update the values when training your neural network
• If you use an activation function other than the sigmoid, then you’ll need to use a different derivative than the one returned by sigderiv

Class: neuralnetwork:

__init__(self, size, seed):

• This function will initialize the weights and biases for a neural network of the size specified by the ‘size’ parameter. ● The ‘size’ parameter is a list of the form
  • inputsize, hidden layer 1 size, … , hidden layer n size, output size]
• Example on next page

Example of __init__

Suppose size = [3, 4, 4, 1]

• then __init__ creates the network shown at right
• It initializes a variable weights = [ [v₁₁ v₁₂ v₁₃]

[v21 v22 v23] [v31 v32 v33]

[v41 v42 v43] …]

• each v_ij is the weight on the connection to unit i in the first hidden layer from unit j in the input layer

Methods/Classes provided in studNet.py (continued) forward(self, input):

Given a vector of input parameters of form:

[ [parameter 11, p12, …, p1n] [p21, p22, …, p2n] … ] This method will return a 3-tuple:

• the output value(s) of each example (the variable ‘a’ in the source code)

○    The values before activation was applied after the input was weighted (the variable ‘pre_activations’)

○   The values after activation for all layers (the variable ‘activations’)

• The reason it returns ‘pre_activations’ and ‘activations’ is because you’ll need them for updating the network

What you have to do:

• Implement a method called train(…)
  • test_train calls it to train the neural network on the data set

○   To do this training, you will need to perform back propagation and calculate deltas

• We’ve provided method headers for train(), backpropagate(), and calcDeltas()
  • If you want to use them, you’ll need to write the method bodies

○    But you don’t have to use them if you want to implement your network differently

• The only things that must stay the same for testing:
  • you MUST have the test_train method and predict as provided

○    test_train must return an instance of your neural network that can be used to call predict(a).