Audio Spectrogram Contrastive Learning

Self-supervised contrastive embeddings of audio via SimCLR on log-mel spectrograms.

Project Structure

• data/: download & preprocess scripts
• notebooks/: EDA & visualization
• src/:
  • dataset.py — contrastive-dataset
  • model.py — SimCLR model
  • loss.py — InfoNCE loss
  • train.py — training script
  • eval.py — downstream evaluation
• configs/: hyperparameter files
• blog/: draft write-ups
• Dockerfile: reproducible env

Dataset

Place the ESC-50 dataset under data/esc50 so that the directory looks like:

data/
  esc50/
    audio/          # WAV files
    meta/esc50.csv  # metadata

src/dataset.py expects a list of file paths pointing to the WAV files.