https://uberduck.ai/ logo
Powered by
# machine-learning
  • m
    synthesis notebook inside #841437191073955920
  • m
    above is ram on the colab session
  • m
    below is gpu ram (circled)
  • m
    okay just wrote this up
  • m
    Copy code
    a = !grep MemTotal /proc/meminfo
RAM = int(a[0].replace('MemTotal:       ','')[:-3]) / 1000000

if RAM > 14:
  print("High ram detected -> Batch size 25 is recommended")
else:
  print("Low ram detected -> batch size 14 is recommended")

print(RAM)
  • m
    maybe 25 batch size is too much
  • m
    change that to like 20 or something
  • m
    this is from colab PRO high ram runtime
  • m
    colab free (low ram runtime)
  • h
    ah
  • h
    well thank you for correcting the misinformation
  • m
    you can definitely make a "Auto Select" option for batch size
  • h
    -actually hm i've realized this is your notebook
  • h
    though might still be my fault for misreading it
  • h
    i need some help, for some reason, tacotron2 isn't saving the model after training, and it restarts at epoch 0 if i try to restart the training
  • m
    which colab are you using?
  • h
    the one in #841437191073955920
  • h
    it doesn't save every single epoch, only every so often
  • h
    i think i had this issue last time too and that trained for atleast 40 minutes
  • h
    hm
  • h
    that's interesting
  • h
    how long does each epoch take
  • m
    sounds like its somewhere you are saving
  • m
    what directory are you saving the model in?
  • h
    /content/drive/My Drive/colab/outdir
  • h
    nvm i just switched to another recipe i used before and now it works, still don't know why it won't work though
  • f
    I made a dataset of Chuck E. Cheese (1977 John Widelock voice), and I just started training him. Right now it looks like he got the s t o n k s... (don't worry guys, I'll continue making the voice of Maurice (from Madagascar) soon)
  • f
  • f
    @User I'd like to inquire, who would I badger for a script to turn these all into numbers and the filenames into a transcript file?
  • m
    shouldn't be too hard
1...461462463...1068Latest