PaSST Versions Save

Added a pre-trained model, trained with KD using the setup in https://github.com/fschmid56/EfficientAT

Pre-trained PaSST-U and PaSST-B on Audioset

• Pre-trained models on the 5 folds of ESC-50.
• The pre-processed ESC50 dataset for fine tuning.

• fsd50k-passt-s-n-f128-p16-s16-ap.642.pt pre-trained on FSD50K with structured patchout and no overlap map=0.642.
• fsd50k-passt-s-f128-p16-s10-ap.655.pt pre-trained on FSD50K with structured patchout map=0.655.
• openmic-passt-s-f128-10sec-p16-s10-ap.85.pt pre-trained on OpenMIC-2008 with structured patchout map=0.85.
• passt-s-f128-30sec-p16-s10-ap.473-swa.pt pre-trained on Audioset but supports inference up to 30-seconds.
• passt-s-f128-20sec-p16-s10-ap.474-swa.pt pre-trained on Audioset but supports inference up to 20-seconds.

Pre-trained models with a smaller STFT hop

Added more pretrained models