How to account for padded inputs in LEAF?

[PrabhjotKaurGosal]

Hello,

I have a dataset where the audio files are not of the same length. Therefore, I pad the shorter sequences to have the overall shape as (batch_size, num_samples). It does not appear that LEAF handles padded input any differently than the non-padded input.

I would like to mask the padded portion of the input sequence for my downstream classification model (example, transformer). Is there any way I can take special care of the padded inputs when it passes through the LEAF architecture, so that the downstream model knows which portion of the input belongs to padded portion of the original sequence?

Thanks!