Neural-Networks

Description

UNM CS 529 Project 3: Creation of Multilayer Perceptron (MLP) and Convolutional Neural Network (CNN) classifiers using Pytorch and Pytorch-Lightning libraries respectively.

Instructions for Use

Install Dependencies

Create a Python virtual environment

python -m venv YOURVENV

Activate the environment by running this command on Windows...

YOURENV/Scripts/activate

or this command on Linux/MacOS

source ./YOURENV/bin/activate

Install required dependecies with the command below (all platforms):

pip install -r requirements.txt

Note that some of the dependencies require Python 3.10 to install correctly and will not install on more recent versions of Python.

Train MLP

1. Specify values for the variables below in utils/consts.py:

   • training_data_path: the path to your training data directory (audio files)
   • testing_data_path: the path to your kaggle testing data directory (audio files)

2. Run python -m MLP.training from the top level directory.

• A hyperparameter search will be started with Ray Tune; this will take a while.
• After training is complete, the best performing model will be used to generate a predictions file in MLP/.

Train CNN

1. Specify values for the variables below in utils/consts.py:

   • training_data_path: the path to your training data directory (audio files)
   • testing_data_path: the path to your kaggle testing data directory (audio files)

2. Run python -m CNN.training from the top level directory.

• A hyperparameter search will be started with Ray Tune; this will take a while.
• After training is complete, the best performing model will be used to generate a predictions file in CNN/.

Train Transfer Learning

1. Specify values for the variables below in utils/consts.py:

   • training_data_path: the path to your training data directory (audio files)
   • testing_data_path: the path to your kaggle testing data directory (audio files)

2. Run python -m Transfer.training from the top level directory.

• A hyperparameter search will be started with Ray Tune; this will take a while.
• After training is complete, the best performing model will be used to generate a predictions file in Transfer/.

Code Manifest

File Name	Description
CNN/training.py	This file contains the training algorithm for the CNN.
MLP/training.py	This file contains the training algorithm for the MLP.
Transfer/training.py	This file contains the training algorithm for our transfer learning model.
utils/consts.py	This file has constants used throughout the library.
utils/spectrograms.py	This file contains our functions to convert audio files to spectrograms.
requirements.txt	This file contains our project's Python dependencies.

Developer Contributions

Prasanth Reddy Guvvala

• Implemented CNN architecture.
• Implemented spectrogram transforms.
• Implemented transfer learning.
• Implemented data augmentation.

Thomas Fisher

• Implemented MLP architecture.
• Implemented Ray Tune hyperparameter searching.
• Implemented model evaluation.
• Implemented report framework.

kaggle Submission

Leaderboard position 11 achieved with accuracy 68% on May 6th (team name: Fisher & Guvvala).