permutation invariant training

[yoyolicoris]

For bleeding/label noise tracks

[ben-hayes]

• linear assignment loss wrapper (allows source-level PIT)
• frame-level tPIT (as in) https://arxiv.org/abs/2102.04945
• (maybe) clustering method as in same paper

[ben-hayes]

to keep memory requirement in O(n^2) rather than O(n!), better to adapt FLoss and TLoss to accommodate dim parameters to allow permutations to be evaluated in parallel

[ben-hayes]

the way some loss fns are currently implemented (e.g. the sdr-combination loss in sdr_loss) are bound to specific shapes and numbers of targets, so either I refactor all these to accommodate the general PIT wrapper, or I scrap the wrapper approach and just hard-code PIT variants of loss functions we're interested in

[yoyolicoris]

I prefer the latter approach since the loss fns are meant to be replaceable (using configs), and we can make as many as we want. :)

[ben-hayes]

yeah fair point! will switch it over.