https://uberduck.ai/ logo
Powered by
# machine-learning
  • t
    Keep in mind I cd'd into a deep folder where I keep it all
  • t
    Unless I should write that directory into this command?
  • h
    my code shouldn't depend on cd location
  • h
    i specifically put 
    /content/blah
    everywhere
  • h
    but it seems you've omitted some mkdir things
  • t
    This is for both codes you suggested I use
  • t
    For the file rewrite, I may have to use more code preceding it I guess
  • t
    Besides the elif portion you gave me
  • t
    Also for if loader == "/mlp/ dataset"
  • t
    It said loader not defined
  • t
    So I disabled it to run more only to get the same problem as the other code
  • t
    Tweaked the code to get further and got stuck here
  • t
    [Errno 2] No such file or directory: '/content/tacotron2/filelists/list.txt'
  • t
    So I guess I need to rewrite the cd portion somehow, I guess
  • u

    https://youtu.be/LrcnaLvu5FE▾

  • h
    if you tell me what your directory structure is and what/where you want the output to be (multiple small datasets or one multispeaker one) i can do the adjustments for you
  • z
    #841461499946860576 , not here
  • c
    has anyone actually tried fine-tuning waveglow with their own data
  • c
    because when i see people talk about how waveglow is bad they usually talk about the default ljspeech model
  • h
    the folks in #903712295534882897 do it all the time
  • h
  • h
    they say the waveglow clip is better, i say nothing because my ears are garbo
  • u
    I dunno if this exists already, but I finished coding a notebook that automatically creates an ljtext formatted text file based off of audacity saving label transcripts as title metadata. It's worked for me but I'd love some feedback! https://colab.research.google.com/drive/1V3qxDBQWBs6d1S65NZQkDXanuADyuWG3?usp=sharing
  • u
    it also automatically detects file names so no need to do bulk file renaming or w/e
  • c
    audacity labels are part of my dataset creation workflow so thats definitely. interesting
  • t
    My cd was cd /content/drive/My\ Drive/ML/encoder/russian_single/book/
  • u
    @User would you call this spam? it’s being posted in every unrelated channel repeatedly
  • m
    Ran through a training session over the holidays and epoch 238 was looking baller and saved out. then Epoch 240 looked awful but also saved overwriting 238... My question is was 240 really better than 238? If not should I be looking back to see if version history on the file was saved in drive to recover 238? should I be saving every version and then comparing?
  • m
    This is great. Coulda used this as I was prepping my last dataset that used SRT files converted to audacity labels to chop up long audio.
  • j
    what's a good number of hifi gan steps for a talknet model?
1...643644645...1068Latest