dlgenai — Audio Genre Classifier

EfficientNet-V7 trained on log-mel spectrograms to classify 10 music genres under distribution shift (clean stems → noisy mashups). Built with PyTorch Lightning and W&B.

Kaggle private LB: 0.93030

Setup

Requirements: Python 3.13, uv

uv sync

Without uv:

python3.13 -m venv .venv
source .venv/bin/activate
pip install -e .

Data

Place the competition data under messy_mashup/:

messy_mashup/
├── genres_stems/         # per-genre stem audio files (training)
├── mashups/              # mixed audio files (test)
├── test.csv
└── ESC-50-master/audio/  # background noise for augmentation

Train

wandb login   # first time only
uv run train.py

Checkpoints are saved locally by Lightning. The best checkpoint is also uploaded to Kaggle Hub (nevrohelios/genre-classifier/pyTorch/best-checkpoint).

Config

All hyperparameters live in config.py (CFG class). See report.pdf for architecture details, augmentation design, and full results.