# machine-learning
- u
user
11/05/2021, 12:17 AMstonkin' out 🥵 - y
yuananf
11/05/2021, 2:50 AMIs there any training notebook of Fastpitch? Like the tacotron2 one. - u
user
11/05/2021, 10:01 AMstonks - w
Waltman13 (semi-active)
11/05/2021, 1:12 PMthe stonk that could've been... - u
(Edd) bruh moment
11/05/2021, 1:15 PMwho cut the stonks - w
watertail
11/05/2021, 6:02 PMhowdy fellers, anyone here familiar with the tacotrong 2 training notebook by adam does stuff? - w
watertail
11/05/2021, 6:02 PMi just bit the bullet and paid google the 10 dollars so i can finally get a p100 to use, anyone know if i can run all 10k epochs and still come back to the process still running? or will i be kicked off? - u
(Dawn) Will Draw Fictional Women
11/05/2021, 6:08 PMif you stay for the catcpha - u
(Dawn) Will Draw Fictional Women
11/05/2021, 6:08 PMidk why you wanna do 10000 epochs because that will 100% overtrain unless youre doing some learning rate wizardry - w
watertail
11/05/2021, 6:09 PMso more epochs doesnt equal better results? - u
(Dawn) Will Draw Fictional Women
11/05/2021, 6:09 PMnot at all - w
watertail
11/05/2021, 6:09 PM - u
(Dawn) Will Draw Fictional Women
11/05/2021, 6:09 PMabsolutely minuscule validation loss generally means better results - w
watertail
11/05/2021, 6:10 PMwhat would you say is good for a batch of 79 audio files? - w
watertail
11/05/2021, 6:10 PMis there a formula people use? - u
(Dawn) Will Draw Fictional Women
11/05/2021, 6:10 PMthe best batch size is both a (preferably clean) division of the dataset size and something cuda’s memory wont cry about - u
(Dawn) Will Draw Fictional Women
11/05/2021, 6:11 PM(which is generally under 40 for colab pro chads like me) - w
watertail
11/05/2021, 6:11 PMwait youre getting good results using just 40 audio files??? - u
(Dawn) Will Draw Fictional Women
11/05/2021, 6:11 PMthats not what i said - w
watertail
11/05/2021, 6:12 PMoh sorry for misunderstanding - u
(Dawn) Will Draw Fictional Women
11/05/2021, 6:12 PMbut yes that is something that has happened easily - u
(Dawn) Will Draw Fictional Women
11/05/2021, 6:12 PMtheres been good models with just 19 - w
watertail
11/05/2021, 6:13 PMjust for clarification my audio files are within 2 to 10 seconds on average, ive seen people do things like using files that are minutes long - w
watertail
11/05/2021, 6:14 PMis it possible to continue upgrading an already pretrained model with new audio samples using tacotron 2? or do i have to just recreate the model from scratch using the updated audio library? - u
(Dawn) Will Draw Fictional Women
11/05/2021, 6:15 PMim not entirely sure actually - w
watertail
11/05/2021, 6:19 PMeither way thank you for taking the time to help me with my questions :) - w
watertail
11/05/2021, 6:26 PMis there a script that is curated specifically for use in voice synthesis? as in a script someone would read for audio samples - u
(Dawn) Will Draw Fictional Women
11/05/2021, 6:27 PMi guess the most ideal thing would be ljspeech's dataset transcript - u
(Dawn) Will Draw Fictional Women
11/05/2021, 6:27 PMwherever that may be