https://uberduck.ai/ logo
Powered by
# machine-learning
  • h
    show me the output of the cell numbered 1
  • m
    nvidia-smi
    or the google drive mount?
  • m
    I've got a K80 and the drive mounted successfully
  • m
    GPU 0: Tesla K80 (UUID: GPU-1f6664c7-10c0-d4d1-5fa1-2e2e74282edb)
  • h
    the installation one
  • h
    with a lot of output
  • m
    ah I think i see it. i'm missing 
    merged.dict.txt
  • h
    i think the arpabet patch might not have downloaded
  • h
    i did update it a while ago but i don't see why that would break things
  • h
    hmm at the top of the cell put 
    !pip install --upgrade gdown
    and run it again
  • m
    Copy code
    Applying cmudict patch
Downloading...
From: https://drive.google.com/uc?id=1gatrzuyxEcYeUnd8eSB1zmtnnHBJi8tn
To: /content/tacotron2-cmudict-patch.zip
100%|██████████| 1.73M/1.73M [00:00<00:00, 27.5MB/s]Archive:  /content/tacotron2-cmudict-patch.zip
  inflating: /content/tacotron2/text/cleaners.py  
  inflating: /content/tacotron2/text/cmudict.py  
  inflating: /content/tacotron2/text/LICENSE  
  inflating: /content/tacotron2/text/merged.dict.txt  
  inflating: /content/tacotron2/text/README  
  inflating: /content/tacotron2/text/__init__.py
  • m
    think it's working fine now. didn't need to use 
    !pip install --upgrade gdown
    but I've put that there anyway.
  • m
    this
    Copy code
    print(text.sequence_to_text(text.text_to_sequence("We must capture an Earth creature, K 9, and return it back with us to Mars.", ["cmudict_cleaners", "english_cleaners"]
    outputs this
    Copy code
    /content/tacotron2
{W IY1} {M AH1 S T} {K AE1 P CH ER0} {AH0 N} {ER1 TH} {K R IY1 CH ER0} , {K EY1} nine , {AE1 N D} {R IY0 T ER1 N} {IH1 T} {B AE1 K} {W IH1 TH} {AH1 S} {T AH0} {M AA1 R Z} .
  • h
    yeah that means it's good
  • h
    though the 
    nine
    is a funky bug that i'll need to fix
  • h
    ,,if i even can fix it since it seems it runs cmudict before digit→word
  • h
    no big deal though
  • m
    in almost all the notebooks i run 
    merged.dict.txt
    fails to download. do you know what this file is?
  • h
    it's a dictionary used to convert from arpabet to phonemes
  • h
    in this case it's part of the cmudict patch
  • m
    ran into another error checking for missing files
    Copy code
    #@markdown ## **4** Check for missing files

# ---- Replace .wav with .npy in filelists ----
!sed -i -- 's,.wav|,.npy|,g' {hparams.training_files}; sed -i -- 's,.wav|,.npy|,g' {hparams.validation_files}

check_dataset(hparams)
    Copy code
    Checking Training Files
 
[WARNING] does not exist.
---------------------------------------------------------------------------
IndexError                                Traceback (most recent call last)
<ipython-input-13-959e2a02f2c3> in <module>()
      4 get_ipython().system("sed -i -- 's,.wav|,.npy|,g' {hparams.training_files}; sed -i -- 's,.wav|,.npy|,g' {hparams.validation_files}")
      5 
----> 6 check_dataset(hparams)

1 frames
<ipython-input-7-443c4424a8da> in check_dataset(hparams)
    497     print("Checking Training Files")
    498     audiopaths_and_text = load_filepaths_and_text(hparams.training_files) # get split lines from training_files text file.
--> 499     check_arr(audiopaths_and_text)
    500     print("Checking Validation Files")
    501     audiopaths_and_text = load_filepaths_and_text(hparams.validation_files) # get split lines from validation_files text file.

<ipython-input-7-443c4424a8da> in check_arr(filelist_arr)
    485             if (not os.path.exists(file[0])):
    486                 print("|".join(file), "\n[WARNING] does not exist.")
--> 487             if len(file[1]) < 3:
    488                 print("|".join(file), "\n[info] has no/very little text.")
    489             if not ((file[1].strip())[-1] in r"!?,.;:"):

IndexError: list index out of range
  • h
    you probably have an empty line at the end
  • m
    training now!
  • m
    another question i have is why the training and validation sets are the same?
  • h
    wooo
  • h
    either because we usually don't have enough data to do a proper separation or because whoever coded that didn't feel like doing it properly
  • m
    @hecko you posted two notebook last week. the arpabet one i'm using now and a 'pipeline'. what's the difference? does the pipeline use arpabet?
  • h
    the current one uses the official tacotron 2 code by nvidia, whereas pipeline uses uberduck's custom reimplementation
  • h
    i don't think you can submit pipeline models just yet, i've only used it for experiments with its fancy features (mostly multispeaker models)
  • h
    and yeah pipeline uses arpabet by default, and i believe it converts everything instead of just the words it knows
1...834835836...1068Latest