https://uberduck.ai/ logo
Join Discord
Powered by
# tacotron-2-support
  • u
    I don't know how to read the value of loss or machine learning. Is there a way to know if learning is going well and progressing in general?
  • r
    The straightér that line is, and more yellow, the better the model should be at talking. Loss, depending on your dataset and settings should lower accordingly. Smaller datasets generally need lower loss values.
  • r
    The graph afaik is a chart is wpm
  • u
    thankyou
  • n
    I tried to fix hparams but it doesn't work. If anyone who have higher levels helped me, I will continue.
  • t
    is it legacy or uberduck pipeline
  • t
    oh its legacy
  • c
    You need tensorflow 1.x
  • h
    common voice is noisy isn't it
  • h
    and noise isn't very predictable, meaning it's gonna have a high loss
  • h
    idk how much voicefixer factors into that
  • h
    uhh if you have multiple voices i trust you've separated them into speakers
  • h
    batch size being a divisor doesn't matter as much in pipeline because it allows for incomplete batches
  • h
    sample inference is the most important graph to look at yeah it shows what the network performs like without any handrails whereas for training/validation graphs it's generating one frame at a time, with the actual training/val data as the previous context
  • u
    I tried the Legacy model again.
  • u
    sooooo, are you guys like code masochist's who enjoy having s!@# not work for no reason; or am I just cursed?
  • u
    I have 0 clue on why it just doesn't work anymore
  • j
    I set 'Arpabet' to true when submitting my model. It is Uberduck Pipeline. It sounds a bit monotone, should I have set to false?
  • h
    arpabet doesn't affect tone
  • h
    it might be because of the default vocoder
  • h
    lemme know what your model is and i'll switch it to a better one
  • j
    It is 04ccbd9c-5b86-4eaf-9985-184b5527e690 thank you
  • p
    [Errno 2] No such file or directory: '/content/base_aitch.pt'
  • h
    aitch got ratelimited, you can change the drive id to the url in the code but it'll start at epoch 1250
  • h
    done, see if it helped
  • a
    my godness
  • h
    it seems google drive is struggling because yeah aitch isn't working and the other one stored on drive isn't working either and some versions of my model didn't save
  • j
    That's much better thank you. If I have an updated model how do I get the new one up there? I might train it a bit longer
  • d
    anyone else having issues with the Tacotron2 Hi-Fi Gan Synthesis (CPU) colab right now?
  • t
    Google Drive is having an issue right now and needs to be resolved.
1...737475...158Latest