https://uberduck.ai/ logo
Powered by
# machine-learning
  • m
    iterations = steps
  • m
    keep training until Loss Total is under 1 or less
  • m
    i think
  • u
    alright then i should stop since i only trained my model for about 100 epochs
  • m
    no
  • m
    keep going
  • u
    my loss total has stayed pretty steady between 40-35, it doesn't really seem to go up or down
  • m
    keep training
  • p
    from scratch? a couple of 100k steps
  • v
    when should i use hifi?
  • v
    how many interations?
  • m
    ???
  • m
    if you have your model trained already, no need for training it on hifi gan
  • p
    that's not enough for these tiny datasets most people make here
  • v
    im working on training it
  • v
    im at like 800 iterations or something like that
  • i
    true
  • p
    for fine tuning 100 -200 epochs should be ok
  • p
    hifigan doesn't have the concept of alignment
  • p
    its just a vocoder net
  • m
    oh alrighty
  • p
    you won't need that
  • p
    unless you have a very unusual voice
  • p
    in most cases a way better quality improvement can be had by increasing the training dataset size for tacotron
  • v
    so just keep running it with new data, that makes sense
  • v
    and unusual voice?
  • v
    as in it sounds robotic?
  • p
    there is no point in doing that if the people training don't understand the concept of validation loss
  • p
    very high pitch/very low pitch/reverbaration
  • v
    hows this? my newest rendition
1...260261262...1068Latest