Academic-DeepNeuralNetsFromScratch

A framework that constructs deep neural networks, autoencoders, logistic regressors, and linear networks without the use of any outside machine learning libraries - all from scratch.

This project was constructed for the Introduction to Machine Learning course, class 605.649 section 84 at Johns Hopkins University. FranceLab4 is a machine learning toolkit that implements several algorithms for classification and regression tasks. Specifically, the toolkit coordinates a linear network, a logistic regressor, an autoencoder, and a neural network that implements backpropagation; it also leverages data structures built in the preceding labs. FranceLab4 is a software module written in Python 3.7 that facilitates such algorithms.

##Notes for Graders All files of concern for this project (with the exception of main.py) may be found in the Linear_Network, Logistic_Regression, and Neural_Network folders. I kept most of my files from Projects 1, 2, and 3 because I ended up using cross validation, encoding, and other helper methods. However, these three folders contains the neural network algorithms of interest.

I have created blocks of code for you to test and run each algorithm if you choose to do so. In __main__.py scroll to the bottom and find the main function. Simply comment or uncomment blocks of code to test if desired.

Each neural network and autoencoder constructed are sub-classed / inherited from the NeuralNet class in neural_net.py. I simply initialize the class differently in order to construct an autoencoder, a feed-forward neural network, or a combination of both.

Data produced in my paper were run with KFCV. However within the main program, you may notice that the number of folds k has been reduced to 2 to make the analysis quicker and the console output easier to follow.

The construction of a linear network begins on line 84 in __main__.py.

The construction of a logistic regressor begins on line 102 in __main__.py.

The construction of an autoencoder only begins on line 128 in __main__.py.

The construction of a feed-forward neural network only begins on line 141 in __main__.py.

The construction of an autoencoder that is trained, the decoder removed, and the encoder attached to a new hidden layer with a prediction layer attached to form a new neural network begins on line 221 in __main__.py.

The code for the weight updates and backward and forward propagation may be found in the following files within the Neural_Network folder:

• layer.py
• optimizer_function.py
• neural_net.py

__main__.py is the driver behind importing the dataset, cleaning the data, coordinating KFCV, and initializing each of the neural network algorithms.

Running FranceLab4

1. Ensure Python 3.7 is installed on your computer.
2. Navigate to the Lab4 directory. For example, cd User\Documents\PythonProjects\FranceLab4. Do NOT cd into the Lab4 module.
3. Run the program as a module: python3 -m Lab4.
4. Input and output files ar located in the io_files subdirectory.

FranceLab4 Usage

usage: python3 -m Lab4