https://uberduck.ai/ logo
Powered by
# machine-learning
  • e
    Havent experienced yet with these synthesizer models due to just starting to mess with them but I'd imagine itd just say the exact phrases from the learning data
  • e
    You usually see it when your loss gets insanely low
  • e
    Suspiciously low
  • e
    Transfer learning can help provide better results on tiny amounts of data without overfitting
  • e
    Tested that out with SV2TTS
  • a
    Epoch: 142 Validation loss 1144: 0.093023 Time: 0.9m LR: 0.000500
  • a
    i dont know how to read that - nor am i experienced enough to imagine what "baseline" is
  • e
    It varies model to model
  • e
    And also your learning rate matters
  • e
    If it over fits too quickly you can decrease learning rate
  • a
    that all makese sense - thanks for the knowledge dump ❤️
  • e
    I'm super inexperienced with ML tho
  • a
    where did you learn your wizardry?
  • e
    Need to actually watch some lectures + read books
  • e
    Literally just screwing around with models in my free time
  • e
    And I technically have built a simple neural network in hardware
  • a
    cool, that's me as of three days ago
  • u
    books arent fun tho :/
  • e
    But without understanding things
  • a
    havent built a neural network in hardware yet
  • e
    Make em fun
  • e
    I do not recommend it
  • e
    You can optimize networks for embedded systems via something like Tensorflow Lite Micro or deepC
  • e
    But hardware networks are pain, like most hardware chip design
  • e
    I think some FPGA support converting models onto boards now tho
  • e
    But for example you could port an optimized single speaker SV2TTS model onto an arduino and make some portable text to speech device or animatronic
  • u
    Have you guys seen NVIDIA Jarvis/NeMo
  • e
    Ye
  • u
    Looks pretty fun
  • e
    Was looking into.. dang what's it called
1...353354355...1068Latest