Birdcall Recognition

Final project of McGill AI Society Intro to ML Bootcamp (Winter 2022). Demo hosted on Heroku.

Project Description

This project is a web application that classifies birdcall recordings of twenty different species using mel spectrograms and a convolutional neural network. The classifier utilized the efficientnet_pytorch implementation of EfficientNet-B7 and was trained on a subset of the Cornell Birdcall Identification dataset. The model was built using PyTorch, and the web app using Flask. Data preprocessing and augmentation was done through the torchvision package.

Running the Application

Clone or otherwise download the repository. You may need to install Git Large File Storage.

Using Docker

Using a terminal, navigate to the /app subdirectory and run

docker build -t birdcall-recognition .

Start the app container with

docker run -dp 8080:8080 birdcall-recognition -e PORT="8080"

Without Docker

Using a terminal, navigate to the /app subdirectory and run

On Linux/Unix/MacOS

python3 -m venv .venv
source .venv/bin/activate
pip3 install -r requirements.txt
waitress-serve app:app

On Windows

py -m venv .venv
.\.venv\Scripts\activate
py -m pip install -r requirements.txt
waitress-serve app:app

To use the Flask development server, instead of waitress-serve app:app, run python3 app.py on Linux/Unix/MacOS, or py app.py on Windows. Using a web browser, navigate to http://localhost:8080 to view the application.

Repository Organization

This repository contains the source code used to train the model and build the web application.

• app/ - The source code for the web application
  • static/ - Static CSS and JavaScript files
  • templates/ - Jinja2 template for landing page
  • .dockerignore - Files/directories to ignore when building Docker images
  • Dockerfile - Commands used to build Docker image
  • app.py - The Flask web application
  • audio_utils.py - Audio preprocessing functions
  • predictor.py - The classification model
  • requirements.txt - Python dependencies
  • weights.py - Pretrained model weights
• deliverables/ - Deliverables submitted to bootcamp organizers
• training/ - Contains the Jupyter notebook used to train the model