questions about atst-frame evaluated on audioset_strong_eval about audiossl HOT 3 CLOSED

Hi,

Appreciate a lot for your interest!

The experiments on the Strongly labeled AudioSet is meaningful for evaluating the SED capacity of models under real applications. But using it is indeed troublesome. Of course, we met the problems you mentioned in our development stage.

I would answer your questions one by one:

1. Yes, the "w/o var-pen" in our table means the $\alpha_{\text{st}} = 0$.
2. We also found the computations of PSDS1 and PSDS2 are very slow, due to the high CPU demanding. We accelerated the process by parallelizing the computation process. You could refer to our updating codes to see the implementation (still need some days...). We reduced the entire inference time to around 10 - 30mins, which is acceptable for us.
3. You are right. We only use a strongly-labeled audio clips + BCE loss to train the model.

Hope these answers make sense to you! If you have any further questions, please let us know.

Nian

from audiossl.

And also, please notice that we calculated PSDS scores by the psds_eval package (used in the DCASE 2022) instead of the sed_scores_eval package (used in the DCASE 2023). The latter provides a smoothen curve and gives higher scores.

from audiossl.

Thank you for your advice and I will give it a shot. : )

from audiossl.