https://uberduck.ai/ logo
Powered by
# machine-learning
  • h
    (anyway tbh yes it would be interesting to make a thing that removes the samples with the highest loss)
  • j
    ok cool, i thought i was in deep trouble or something if it was plotting cosine similarity, but thanks
  • j
    how far should i run the epochs? i have 500 wav files and i currently set it at 1000 epochs
  • j
    i got better at seeing the breaths to expedite the sentence tagging process but it was still painful
  • h
    1000 epochs holy hell
  • h
    i do 250 but that's partly because i'm impatient
  • j
    i could always stop the code given that it saves every epoch
  • j
    but yeah it might be overkill
  • h
    i'd say check every so often with a sentence from the training data, if the intonation is the exact same you've gone too far but checking requires you to stop the notebook and run another one, which can run up your gpu cap quickly
  • j
    can't I just duplicate the output file, make that link public and then run the synthesis notebook using another google login?
  • h
    yeah that would work
  • h
    you don't even need to duplicate it
  • a
    @User I've been in that colaboratory you told me to go check out, and I got this error:
  • h
    uh oh
  • h
    idk what the cause might be
  • h
    if i had to guess maybe people downloaded tacotron too much so it hit a rate limit??
  • a
    ooh ok
  • a
    i clicked the google drive link in the error and got a 404 if that helps
  • h
    hmm
  • h
    $5 says it's a retaliation by 4chan
  • h
    hm
  • h
    step 2 in the training notebook works fine for me
  • h
    what are you using again
  • u
    My initial model sounds better than the one with more samples
  • h
    ohhh no wait it's
  • h
    you put in an incorrect id
  • h
    no 4chan actions have happened
  • a
  • h
    yeah
  • h
    you put in the id wrong
1...440441442...1068Latest