Cornell Birdcall Identification competition

This is the code for Cornell Birdcall Identification challenge hosted on Kaggle

Data

Librosa library is pretty slow for reading and transforming audio. So, I read data using librosa and saved it as HDF5 file. More about that you can read here.

Script for transforming .mp3 to hdf5: create/read_and_transform_audio.py

Augmentations

Augmentations are useful for better models generalization. I've used albumentations library and this Kaggle notebook to build augmentations for spectrograms transforming.

Code for this part tou can find here: modules/data/augmentations

Model

I've used CNN for image classification for this task. Family of EfficientNet models is the SOTA for image classification now, so I chose it. Also I've used PyTorch Lightning to build training pipeline.

Model part: modules/model

Built With

• PyTorch - Neural networks framework used
• PyTorch Lightning - For training pipeline
• Albumentations - Fot spectrogram augmentaions

Authors

• Vadim Titko aka Vadbeg - GitHub | LinkedIn