PaSST models trained downstream on TAU Urban Acoustic Scenes 2022 Mobile, development dataset:
- passt_253.pt : patchout_f=6, mixstyle_alpha=0.6, mixstyle_p=0.5; val. loss: 1.047
- passt_254.pt : patchout_f=6, mixstyle_alpha=0.8, mixstyle_p=0.3; val. loss: 1.055
- passt_255.pt : patchout_f=6, mixstyle_alpha=0.6, mixstyle_p=0.5; val. loss: 1.034
- passt_256.pt : patchout_f=6, mixstyle_alpha=0.8, mixstyle_p=0.7; val. loss: 1.074
Predictions on reassembled 10-second files of a PaSST model:
- preds_227.pt : patchout_f=8; val. loss: 0.655