https://uberduck.ai/ logo
Powered by
# machine-learning
  • t
    I think I understand the format now
  • t
    From the link I downloaded for the English dataset on the GitHub page of my code, I took a file for voices and another for text and metadata
  • r
    Lol yeah my Chao dataset doesn't contain any actual meaningful words but it does have the right sounds to do multiple languages like EN and JP, which I mean I might could use to update the model and shit, but I do feel like it'd really offset any natural tones by just using 1 sound each.
  • t
    Can't find a Japanese dataset to properly download though, at least to my Drive
  • h
    it might actually also be too big for the colab vm itself
  • h
    in which case you'd have to resort to extreme trickery like
  • h
    downloading a random 5 gb chunk of each dataset, but a different chunk each time you resume training
  • h
    idk if that's possible to then unpack with tar
  • h
    only other thing i can think of is begging zwf to do it on one of his machines
  • g
    Should I scrap?
  • m
    bro is that the size of your desktop, how small is your monitor :0
  • g
    It's a laptop.
  • g
    It's at 79 there.
  • g
    It's a laptop, and i have bad vision so I lower the resolution.
  • m
    whats validation loss at?
  • m
    if its still high, you can continue till its 0.15 and see if its okay
  • g
    0.308992
  • m
    yeah its still very high, i wouldn't scrap it just yet...
  • g
    I have it set to go to 250.
  • m
    i normally dont go with the epochs, i just wait to validation loss is 0.15, and then wait for next save and then test it...
  • g
    It would be nice if the training notebooks allowed you to set stopping points based on validation loss instead of epoch.
  • h
    the ideal would probably be based on when it stops improving
  • m
    reworking on my smash announcer model and the graphs are weird but i sometimes get a graph that looks like this so hopefully the model is better than the one on the site right now...
  • g
    177
  • g
    Is there even a way to save those graph images in Kaggle?
  • t
    I remember that being an issue at one point but I think that was for a previous step. Unless I'm remembering incorrectly.
  • t
    Right now I have a gz file sitting in my drive for English datasets that I can't unzip probably because it's too big.
  • t
    And in regards to Japanese, I haven't found a good new dataset Though I think I still have this one of 100 speakers I used the first time around
  • t
    Unlike the ones I tried downloading before this one has wav files
  • t
    But I don't think I know how to keep track of the metadata and that was what stopped me before
1...571572573...1068Latest