https://uberduck.ai/ logo
Join Discord
Powered by
# tacotron-2-support
  • t
    Tbh I don't really understand how you can tell if the training went well with pipeline
  • t
    I didn't see any graphs
  • t
    Plus it takes too long
  • t
    If someone could tell me what I'm doing wrong, that would be extremely helpful
  • t
    ...am I supposed to put the dataset under the sample data folder?
  • s
    You Can Stop Epoch At Anytime
  • t
    And how do I know if the training results went well?
  • c
    It takes around two times less than legacy
  • c
    check images tab and audio tab
  • t
    Alright, Imma try with Pipeline and see what happens
  • t
    Where is that tab located?
  • c
    tensorboard
  • t
    Alright, I'm setting that up
  • t
    what does this mean?
  • t
    alright it doesn't work
  • c
    As I've said
  • c
    images and sound tab
  • c
    forget that
  • c
    check the loss too
  • c
    at the training part
  • t
    I have finally figured out how to train a model, but just to confirm, if I set to 200 epochs, should I wait until it reaches 200 epochs, or stop it at less than 0.2 validation loss?
  • u
    i prefer stopping manually
  • u
    just set epochs to a ridiculous number
  • w
    Has anyone ever trained tacotron from scratch? At what epoch can I expect to see decent alignment?
  • u
    from scratch? like without a base model?
  • w
    yea, I am trying to train model for my language
  • t
    So let's say I have 67 wav files and is set at 200 epochs. Should I stop at 150 epochs if the validation loss is less than 0.2?
  • u
    @hecko time to shine
  • u
    yes?
  • h
    you can train a model for a new language and use the english bases, just put 
    embedding.weight
    in the ignored layers if you're using a different symbol set
1...109110111...158Latest