Pariente Manuel

I'm undecided about this. On the one hand, having data prep and eval script for Chime4 is cool, and it's an example for other models but we could probably make...

I guess it will be clear once people have the data from CHiME3 downloaded.

> Is getting the masker "working size" useful for all filterbanks? Seems so, yes. > The code requires the fix_input_dims method that's only defined for DCUNet right now, shall we...

> Another crazy idea is to have a mode ("training mode") that just doesn't re-pad anything? (Both in the masker and models) So that the loss is only computed on...

Yes, you're right

I don't feel that adding it to the `forward` is the right way to do. The `get_output_size` option on the masker seems better to me.

Thanks for the bug report, it indeed looks like a bug, and thanks for the code. Would you have time to look into it deeper?

Can you have a look [here](https://github.com/mpariente/pb_bss/blob/master/pb_bss_eval/evaluation/module_si_sdr.py) for differences please?

That's what I expected, and this is a good news: the metrics is good, and the only problem is regarding the permutations. We usually don't compute it in the recipes,...

This does make sense. Is it a batching problem maybe? Or just that the optimal permutation for SDR is not the same? Do you have an idea how to fix...