https://uberduck.ai/ logo
Powered by
# machine-learning
  • h
    you can use low batch size (single digits) with a low learning rate (double-click the cell where you set the batch size, change 
    5e-4
    (or 
    1e-3
    ?) to 
    2e-4
    , run it)
  • h
    and instead of looking at epochs look at the loss, around 0.12 is where you should stop
  • h
    also assuming you're using the non-pipeline notebook i have a fancier version here that trains the model to pronounce things better with the pronunciation_dictionary option https://colab.research.google.com/drive/1h9NMKterow9s8ggC87s8wM6iQaD0T8Lh
  • m
    Thank you for helping me. I'm new to this world and it's been amazing to learn more and more about it.
  • d
    Ah so you don't have to wait for it to finish by itself? that's awesome because it is currently at 0.10
  • d
    Okay stopped it, I hope it saved it
  • d
    Hmmmm okay I restarted just in case.
  • u
    it saves it every once in a while, check the output for when it last says "saved model output at " or whatever
  • d
    So it doesn't matter what the graph looks like, as long as the number is below 0.12? Gotcha.
  • h
    it does matter
  • h
    but don't focus on just the most recent graph but rather look at a bunch
  • h
    if most of them are fine then so is your model
  • h
    oh also: 
    /¯
    is normal, it just means the test wav was short
  • u
    if the graphs aren't fine by the time you get to the target val loss, there's most of the time either an issue with the transcription, or an issue with the dataset's wavs
  • u
    sounds good! the artifacting is because of the vocoder; if you finetune hifigan with the notebook in #841437191073955920, that should fix it
  • u
    training a model and that lil skip is worrying me... it appears on every few graphs hopefully it clears up
  • u
    good plan! i really like it
  • u
    if you want to fix the small pronunciation issues, you can train it with the arpabet notebook your choice, though! she already sounds really good
  • d
    and I think I know why no one has ever done animals yet....How the heck would you parse what they say in the List file?
  • d
    I wonder though...I'm gonna try a bunch of cat noises....
  • l
    omg i just did my cupcakke talknet does this sound good or does it need a finetune or train more or just fine
  • u
    is it a problem if during hifigan finetuning it displays all like this rather than displaying any number of steps?
  • u
    it sounds sorta weird but maybe that's how cupcakke sounds? idk you can't finetune it since talknet already does that when you're training (it's the last step) i have no idea what it's saying. can it speak normally?
  • u
    forgot to attach the photo
  • h
    this means you have less than 16 wavs and since batch size is 16 it throws away all of them
  • h
    idk where the setting is to be changed, might be in hifi-gan.py?
  • c
    Lookin' great so far
  • m
    @hecko getting this a lot when training on the pipeline notebook 
    warning: audio amplitude out of range, auto clipped.
    is there a specific amplitude we should aim for when normalising the clips?
  • h
    uhh that probably means your data is at max volume and the ai is overshooting a bit
  • h
    you can bring it down a couple decibels if you want but i don't think it's a big deal
1...859860861...1068Latest