https://uberduck.ai/ logo
Join Discord
Powered by
# machine-learning
  • r
    Tbh can't tell what Coqui is supposed to be cause there was a paper that says it's it's own tts archi but other things say it's just a collection of others
  • r
    Unless I just apparently forgot how to understand English
  • u
    It's got its own tts architecture but it's also a toolkit for a bunch of other things
  • w
    If I were to have a talknet model that I made where I trained all of the steps exept HifiGan, could I add onto the hifigan versions of my model?
  • w
    Wait i need to explain better
  • w
    I've got V1 of a model. Let's say that that model which was completely trained with hifigan and all had 1 hour of audio. Let's then say I made a addition to that model and made a V2 with more audio data. Could I just take the hifigan data I've already got from V1 and add onto it and train it more? Or would that not work since there's new data?
  • u
    That would work, hifigan models are not data-specific
  • w
    Oh really? If that's how that works than why isn't there a pretrained hifigan thing that it just applies to every dataset?
  • w
    Or is it specific to the dataset?
  • w
    But that's contrary to what you just said here
  • u
    There is a pretrained hifigan model, it's used in the hifigan training step It's just that a hifigan model finetuned on specific data works better on that specific data
  • w
    Oh okay
  • w
    So how would I add onto a hifigan thing?
  • w
    Because I am adding new data to the dataset, would it do anything bad introducing new data for it to learn off of?
  • a
    I thinking my training model is not training good
  • m
    100 epochs in and still at validation loss 1?? something is definitely wrong
  • a
    What's wrong?
  • m
    it shouldn't be above val loss 1... it should've gone below at least before the 10th or so epoch... it might be a problem with the dataset? i'm unsure
  • a
    I changed the epoch to 200
  • a
    I thought that the epoch would vary according to the amount of wavs
  • u
    It wouldn't do anything bad And i'd just recommend training it again, it's not super hard
  • w
    I'm just wondering somtimes I'll train a model with one account and do the hifigan to 7,000 steps. I just did that. Because I ran out of GPU credits I moved accounts and flipped all of the files to that account's drive. After going through the steps again to finally save it all, the Hifi-Gan started from step 0 for some reason!
  • w
    And it appears it's not just a visual error either that's actaully what happened it sounds terrible testing it
  • w
    How would you transfer a pretrained hifigan thing between models? Just take that folder and copy paste? My issue I just had might say I can't do that
  • g
    The problem is that your batch size is 1
  • g
    Change it to 8 or 4
  • a
    Okay. Thanks
  • u
    Yeah It might sound better or worse between models but it shouldn't sound horrible
  • w
    I might try. I've been having issues with the overall quality of my models and I can't decipher if it's me not training the model enough or my actual dataset.. Would it be ok if I put up a link to my dataset for someone to look over to see about if it's the dataset itself and what I should remove?
  • u
    Yeah sure
1...101310141015...1068Latest