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.

The model was trained on ~2000 recordings of the American Crow, American Robin, Barn Swallow, Bewick’s Wren, Blue Jay, Carolina Wren, Northern Raven, Common Yellowthroat, House Sparrow, House Wren, Mallard, Marsh Wren, Northern Cardinal, Northern Mockingbird, Red Crossbill, Red-winged Blackbird, Song Sparrow, Spotted Towhee, Swainson’s Thrush, and the Western Meadowlark.

The data was split 80:10:10 into train, validation, and test sets. The weighted average precision was 0.88 on the test set.

Built With

• css
• docker
• flask
• heroku
• html
• javascript
• python
• pytorch