Skip to content

AudioSet fine-tuning weights

Latest
Compare
Choose a tag to compare
@daisukelab daisukelab released this 25 Mar 13:42
· 14 commits to master since this release

We release the weights fine-tuned on AudioSet (AS2M), which were originally pre-trained with M2D (masking ratio of 0.7).

  • m2d_clap_vit_base-80x1001p16x16-240128_AS-FT_enconly.zip ~ mAP 0.485
  • m2d_as_vit_base-80x1001p16x16-240213_AS-FT_enconly.zip ~ mAP 0.485
  • m2d_as_vit_base-80x1001p16x16p32k-240413_AS-FT_enconly .zip - 0.47998, 32 kHz input
  • m2d_vit_base-80x1001p16x16-221006-mr7_as_46ab246d.zip ~ mAP 0.479

All weights are 16 kHz input unless denoted.