In this release, we upload pre-trained model checkpoints. Specifically, for all five transformers in the paper (ATST-Frame, fPaSST, BEATs, ASiT, M2D), we upload the SSL/ImageNet, weak and strong checkpoints that we used.
Beyond the scope of the paper, we upload two low-complexity models (customized MobileNets with widths of 0.6 and 1.0), which are trained with a similar Knowledge Distillation procedure: frame_mn06_strong_1 and frame_mn10_strong_1.
All credits for SSL/ImageNet and weak checkpoints (except for fPaSST and ASiT weak checkpoints) go to the authors of the respective papers (as referenced in the README file).
There are also some audio files downloaded from Freesound that may be used as examples to test model inference. If you use them, checkout the attributions.