nnet

A Neural Network implementation using gradient descent in Python (sans NumPy)

Using nnet

net.py - the main Python file dataprep - folder where data can be located

Loading Data

In net.py, one may adjust the imports to laod in a data file. For example, for an example data file called example.py (located in the dataprep folder), the import statement import dataprep.irisdata as DATA should be changed to import dataprep.example as DATA. (line 5, net.py). The data for dataprep/irisdata.py is actually a CSV file, but the library csv is used to process it into the required format.

Adjustable parameters

One may adjust all the parameters of the neural network directly from the data files. See /dataprep/ANDdata.py for an example.

• Network architecture: create a variable layout to describe the network architecture as a list, with each element determining the number of neurons in the layer. The 1st entry is the input layer, the last one is the output layer.
• Training data: all input values for training data should be in the list x, each training example being an list within in the list with as many elements as input neurons. The correct outputs should be lists within list y and, with each example having the same length as the number of output nodes.
• Initial random range of weights: the variable rp determines the range of random inital weights (optional parameter. Setting rp=1 usually works fine).
• Learning rate of network: This can be adjusted by variable lrate.
• Iterations of neural network: The variable iters determines the number of iterations over the training data done by the network.
• Regularisation: Regularisation works to combat overfitting to the data by discouraging complecated hypothesis functions. A high value of reg corresponds to more regularisation in the network (ie: simple hypotheses are prefered).
• Feature names: List fnames is a list of the names of the features (as a string) in the neural network. Must be the same legnth as a single training example in x.
• Output names: onames is a list of the names of the individual outputs in the neural network (see /dataprep/irisdata.py for an example). Must be the same length as the number of output neurons.

Functionality

Upon running the file net.py with all data properly configured, the network starts training and after it is finished, a GUI will appear in a new window. The cost of the network (the error between predictions and true values) is graphed over the number of iterations allowing you to track the progress of the network. The sliders on the side allow you to adjust different values for the features (as named in fnames) and the network's prediction for those values along with the certainty for its prediction is shown above the graph.

Requirements

• Python 3.7 or above
• Matplotlib (can be installed via pip)