COVID Cough Predictor

Hackacthon bitsxlamarato submission: Telegram bot that classifies between cough and non-cough audio files and then COVID-cough and non-COVID-cough

The goal of this project is to develop a tool that can classify a cough audio message between COVID and non-COVID cough. This has been achieved by training a first predictor on cough and non-cough data and a second predictor on COVID and non-COVID cough data.

The main challenges we ahve had to tackle are the difference in the length of the audio files, as well as the heavy unbalance between classes.

Preprocessing

To work with different length audio signals we obtain the spectrogram and reshape it to a standard size. This preserves the features of the sound files while allowing us to feed the networks with images of the same size:

To solve the unbalanced class problem we have performed data augmentation by changing the sampling frequency of the audio signals. We have also tunned the class weights in the loss function.

Learning Models

The final model uses trained weights from the VGG16 network trained over ImageNet. The feature extraction layers have been frozen, and we have fine-tuned an MLP on top of them. This approach is much faster than learning a model from scratch, and aids in avoiding the over-fitting problem.

Installation

In order to correctly use the programs provided, it is necessary to have installed the following libraries:

• numpy
• torch
• torchvision
• torchaudio
• tqdm
• matplotlib
• soundfile
• pandas
• shutil
• sklearn

These can be installed using the pip command in the command line:

pip install name_of_package

Architecture

The repository folder contains several files, one being this README. The rest of them are:

• The preprocess_pipeline.py inside the preprocessing folder, along with other auxiliary notebooks.
• CNN+LSTM_covid.ipynb, CNN_scratch.ipynb, CNN_transfer.ipynb, CNN_transfer_covid.ipynb inside the model model_training folder, notebooks that contain the training and validation of the different models.
• cough_detection_model_from_scratch.pt, fine_tuned_transfer.pt, fine_tuned_transfer_augmented.pt, fine_tuned_transferv2.pt inside the models folder.
• Presentation with slides showing an overview of the project.

The datasets required have not been uploaded due to their size. The repository is for educational purpose only. Note: The rest of the files in the master branch are auxiliary or license related

Team

This project was developed by:

Carlos Hurtado	Elías Abad	Alex Martí	Arnau Turch

Students of Data Science and Engineering at UPC.

License

MIT License