Austin

> Brainstorming: what if we had an optional `prep_for_metrics` method? We wouldn't enforce that `forward -> metrics` worked, but we would enforce that `forward, batch -> prep_for_metrics -> metrics` worked....

I can see defining a model level weight initialization scheme in one function as a callable but what do you mean by "in addition to the enum values"? I'm having...

Maybe? Perhaps with some kind of initializer registry? I'm still skeptical that these would be used across models but it may be good to cover the case where someone explicitly...

Hi Daniel! This is quite a lot of renaming. Though, I'm not convinced it's adding clarity. To me, the crux of this issue seems to be that data on a...

Thanks a lot for the context. Makes more sense now. I don't mean to come in and disrupt an already discussed design doc... but I'm going to sort of do...

@dakinggg haha thanks for your patience. I've been in computer vision land. Given that the "switching to one file path instead of many and just using the prefix to determine...

Hi @PetrochukM, Thanks for your interest but there's quite a lot of training still to be done on both the spectrogram network and the wavenet before any audio samples will...

Hey @Rayhane-mamah, Thanks for the compliment. Though this is still very much a work in progress. I'll take a look at your work soon! The attention mechanism doesn't seem to...

Hmm, that's what I did. previous context --> conv1d layers --> add. To me "cumulative" would be some weighted sum on previous context vectors though I suppose that information is...

My implementation appears to be working now. I ended up using the last decoder rnn hidden state as the query vector rather than the output of the prenet and fixed...