kkoutini/PaSST

Implements patch-based dropout during spectrogram processing to reduce training time and memory by 2-3× while maintaining or improving accuracy on audio classification tasks. Uses vision transformer architecture adapted for audio with configurable structured (time-frequency) or unstructured patch dropout, integrated with PyTorch Lightning, Sacred for experiment management, and Weights & Biases for logging. Provides pre-trained checkpoints compatible with the HEAR 2021 NeurIPS API for direct inference or fine-tuning on downstream datasets like AudioSet.

370 stars. No commits in the last 6 months.