# machine-learning
- h
hecko
09/02/2022, 10:11 PMgoing from e.g. 1e-3 to 1e-4 decreases it by a factor of 10 yeah - h
hecko
09/02/2022, 10:12 PMi've heard that 1e-3 then 5e-4 (/2) then 1e-5 (/50) is a good progression curve but idk how scientific that is - c
Couch
09/02/2022, 10:16 PMThe Exponential Learning Rate Decay - g
Gamma Prime
09/02/2022, 11:08 PMI'm trying a much lower rate on Astrotrain now. - g
Gamma Prime
09/02/2022, 11:08 PMIf he fails again, should I just scrap the dataset? - p
PopcornBubblegum
09/03/2022, 1:32 AMwords - p
PUMPKINEATER
09/04/2022, 4:21 AMMy Thoughts on Keanu Reeves but Pruane got his balls kicked in (testing Koe Recast) - p
PixPrucer
09/04/2022, 12:38 PMDidn't know the pipeline notebook can train 44.1kHz audio just fine, interesting - p
PixPrucer
09/04/2022, 12:38 PMThe only problem is that the training audio is lowed down by a factor of 2 - p
PixPrucer
09/04/2022, 12:39 PM(or that's just an error from my side not converting to 22.05kHz beforehand...) - h
hecko
09/04/2022, 12:44 PMactually iirc it's kinda weird - p
PixPrucer
09/04/2022, 12:45 PMIt's wacky af LMAO - p
PixPrucer
09/04/2022, 12:46 PM(upsampled) - h
hecko
09/04/2022, 12:46 PMfor one there's
andsample_rate
and only one of them is correctsampling_rate - h
hecko
09/04/2022, 12:46 PMfor two the pitch issue yeah - h
hecko
09/04/2022, 12:46 PMsomehow i could only get the correct pitch by doing 22√2khz - p
PixPrucer
09/04/2022, 12:46 PMokay lemme redo the training rq - p
PixPrucer
09/04/2022, 12:46 PM(this time with the proper sampling rate) - p
PixPrucer
09/04/2022, 1:09 PMdo we know after which step the output gets better? It's very muffly compared to the training data - p
PixPrucer
09/04/2022, 1:10 PM(might this be because of the basemodel...?) - p
PixPrucer
09/04/2022, 1:14 PMalso I'm only listening to the audio from tensorboard, so it might not be the accurate representation of the model - h
hecko
09/04/2022, 2:56 PMyeah tensorboard uses griffin-lim which is very warbly - h
hecko
09/04/2022, 2:56 PMvs the synthesis notebook which uses hifi-gan - p
PixPrucer
09/04/2022, 2:56 PMexplains - e
Epic toad
09/05/2022, 2:32 AMYo that you? - p
PixPrucer
09/05/2022, 6:45 AMI finally managed to train it It struggles with intonation heavily and the pronunciation is sometimes messed up (I wonder if that's just general lack of epochs/training data contributing to the monotone intonation...) - t
TonyTony
09/05/2022, 6:46 AMWhat is this - p
PixPrucer
09/05/2022, 6:46 AMUbeduck pipeline PL model - t
TonyTony
09/05/2022, 6:48 AMPipeline - t
TonyTony
09/05/2022, 6:48 AMPL, Poland?