# machine-learning
- u
(Dawn) Will Draw Fictional Women
03/16/2022, 10:12 PMdid you restart runtime and rerun all cells - u
(Dawn) Will Draw Fictional Women
03/16/2022, 10:13 PModd - u
(Dawn) Will Draw Fictional Women
03/16/2022, 10:30 PMi gotchu - u
(Dawn) Will Draw Fictional Women
03/16/2022, 10:40 PMhttps://media.discordapp.net/attachments/364889805898776576/953784703280898068/dpVTRDdC.gif▾
- j
JamesVincentWalsh
03/16/2022, 11:10 PMMe watching the val loss increase with every iteration: 😨 - j
JamesVincentWalsh
03/16/2022, 11:12 PMAaand booted. Got it down to .10 in one sitting. That feels like some kinda record... - j
JamesVincentWalsh
03/16/2022, 11:23 PMTrained this one from the beginning with arpabet and the noisiest wavs denoised as best as premiere pro can manage (which is pretty poor tbh). Still very robotic and now it's got a weird accent...??????? Starting to seem like training it with a mix of arpabet and normal text might yield a better result. - a
Alphabet
03/17/2022, 1:15 AMso retrained the tacotron Aerith model. Initially itwas trained to 200 epoch and val loss of .15. Retrained to 400 epoch and val loss .11, and 600 epoch with val loss .12. So the .11 to .12 range was the max it would go - a
Alphabet
03/17/2022, 1:16 AMthere wasn't much difference between 400epoh and 600 epoch. The main issue with tacotron is that result you get has random acting for how they say the lines. So tacotron will always be limited in usage since you can't get the models to act the way you want them to - a
Alphabet
03/17/2022, 1:17 AMTalknet seems to be more fun, but results from talknet usually seem worse. Time to test different talknet values to see if there's any difference - u
(Dawn) Will Draw Fictional Women
03/17/2022, 1:18 AMtacotron's always a nice ease-of-use option - a
Alphabet
03/17/2022, 1:20 AMtrue. tacotron is much easier to use - a
Alphabet
03/17/2022, 1:21 AMand you don't need to speak a bunch of wav files - a
Alphabet
03/17/2022, 1:22 AMactually, you know what? I found something that might help with tacotron - a
Alphabet
03/17/2022, 1:23 AMthat program lets you change pitch and inflection. So you might be able to change the tacotron results to howyou want them to act - m
mega b
03/17/2022, 1:23 AMim not sure if it is possible just yet to extract GSTs from audio - a
Alphabet
03/17/2022, 1:24 AMGST? - m
mega b
03/17/2022, 1:27 AMGlobal Style Tokens, it is the thing we train TorchMoji on your transcript - m
mega b
03/17/2022, 1:27 AMit allows emotional synthesis like - m
mega b
03/17/2022, 1:27 AM - a
Alphabet
03/17/2022, 1:28 AMoh right, it mentioned that torchmoji could help with emotions - a
Alphabet
03/17/2022, 1:28 AMbut i could never figure out how it did that. Is there a tutorial somewhere to how that can be done? - m
mega b
03/17/2022, 1:30 AMit is already done during training - m
mega b
03/17/2022, 1:30 AMin the notebook, you can see cells "Download Torchmoji" - m
mega b
03/17/2022, 1:30 AMand it is already set inside the config - a
Alphabet
03/17/2022, 1:38 AMoh, but I mean on the output. Like is there a way to control it in the output result? Like {HAPPY} will make the next sentence sound happy? Or maybe i'm misunderstanding what torchmoji does - u
(Dawn) Will Draw Fictional Women
03/17/2022, 1:40 AMyou can override the sentence it grabs emotion from - u
(Dawn) Will Draw Fictional Women
03/17/2022, 1:40 AMso if you want a happy output you give it a happy sounding tweet as an override