https://uberduck.ai/ logo
Powered by
# machine-learning
  • h
    what batch size and learning rate are you using
  • t
    batch size 32 learning rate is 1e-3 ~3 hours of dataset converted to turkish ipa
  • h
    and you set the symbol set to ipa yes?
  • t
    yeah
  • h
    and are you training from scratch or finagling a base model to work with ipa
  • t
    training from scratch, couldn't find any turkish ipa base model
  • h
    oh that's gonna take ages
  • t
    cannot start from tacotron2_statedict either it gives size mismatch error
  • h
    use an english base model and ignore the 
    embedding.weight
    layer
  • h
    assuming this is pipeline
  • t
    yes uberduck pipeline
  • t
    english base model gives size mismatch error when i put its path into warm_start_name
  • t
    even when i set the ignoring layers
  • h
    but did you set embedding.weight
  • h
    not speaker_embedding.weight that's a different thing
  • h
    just embedding
  • t
    wait a minute
  • t
    f me it worked initial loss is 1.36 it was 40 when i train from scratch
  • t
    but will it converge with turkish?
  • h
    i did this with a simpler language on half an hour of data and it worked decently
  • h
    so should work on 3 hours of turkish yeah
  • t
    thank you so much i want to hug you i got one more question should sample_inference_text be ipa?
  • h
    yes
  • t
    so why ipa is better than just plain alphabet
  • t
    gives more surface to train i think
  • h
    it's more that the network doesn't have to learn as much
  • t
    and is this okay "ignore_layers": ["embedding.weight","speaker_embedding.weight", "spkr_lin.weight", "spkr_lin.bias"],
  • h
    yeah
  • t
    thanks
  • t
    if i need to warm start from latest checkpoint(due to google colab usage limitations) which layers should i include does it cause any problem?
1...918919920...1068Latest