# machine-learning
- u
(Dawn) Will Draw Fictional Women
03/15/2022, 7:12 PMi would guess the bad results come from bad alignment or incomplete alignment due to high val loss - u
(Dawn) Will Draw Fictional Women
03/15/2022, 7:12 PMlike sub-0.09 - u
(Dawn) Will Draw Fictional Women
03/15/2022, 7:12 PMbut yeah dont overtrain it - a
Alphabet
03/15/2022, 7:13 PMhmm, i'll try again for taco then. How about talknet, any recommendations for that one for training? - u
(Dawn) Will Draw Fictional Women
03/15/2022, 7:13 PMhmm - a
Alphabet
03/15/2022, 7:13 PMi think talknet is more useful since you can have the characters speak the way you want them - u
(Dawn) Will Draw Fictional Women
03/15/2022, 7:13 PMwhat were your learning rates? - a
Alphabet
03/15/2022, 7:13 PMdefault learning rates - u
(Dawn) Will Draw Fictional Women
03/15/2022, 7:13 PMfor melspec - u
(Dawn) Will Draw Fictional Women
03/15/2022, 7:13 PMthe old defaults or the new ones? - u
(Dawn) Will Draw Fictional Women
03/15/2022, 7:14 PMold ones are in the comments and the new ones were set in place in the latest training notebook - a
Alphabet
03/15/2022, 7:15 PMlearning rate 1e-4 and min learning rate 3e-7 for talknet - a
Alphabet
03/15/2022, 7:15 PMyea, should be the latest training book for talknet - a
Alphabet
03/15/2022, 7:16 PMactually - a
Alphabet
03/15/2022, 7:16 PMi'm using the backup talknet one, so it looks like the learning rates or different than the other one - u
(Dawn) Will Draw Fictional Women
03/15/2022, 7:16 PMtry setting them back to old default - u
(Dawn) Will Draw Fictional Women
03/15/2022, 7:17 PM1e-3 3e-6 - a
Alphabet
03/15/2022, 7:17 PMhmmmm - a
Alphabet
03/15/2022, 7:18 PMalright, let's try that again then. And i'll train taco to sub 0.09 - j
JamesVincentWalsh
03/15/2022, 7:47 PMgod almost every file has rustling in it. No wonder my models sound so robotic x.x Guess I know what I'm spending the next few hours doing 🙃 - m
mega b
03/15/2022, 9:52 PMHey there, if you want to DM me, I'll try to work out a way to give you a big multi speaker model of which you can finetune from. - m
mega b
03/15/2022, 9:53 PMNot exactly. In the pipeline, since it uses enforcement teacher learning, the val loss does not reflect the true quality - m
mega b
03/15/2022, 9:53 PMThe best way to prove quality is by audio samples or the sample inference alignment - u
(Dawn) Will Draw Fictional Women
03/15/2022, 9:55 PMi mean duh - u
(Dawn) Will Draw Fictional Women
03/15/2022, 9:55 PMbut ive found it super reliable - a
Alphabet
03/15/2022, 10:17 PMoh yea, that would be great. And then I can start off of that model instead? - u
⫩ DanielDaBloxGamer ⫩
03/16/2022, 11:47 AMUsing this to generate sentances I will record to train the Tacotron AI - j
JamesVincentWalsh
03/16/2022, 6:30 PMCuda out of memory. F Searching the server isn't really helping much- seems like a lot of people haven't gotten any answers as to solving this, aside from slimming down the dataset, which I don't really want to do. Would deleting one of my models solve the problem? - u
(Dawn) Will Draw Fictional Women
03/16/2022, 6:33 PMcuda out of memory for talknet or taco? - j
JamesVincentWalsh
03/16/2022, 6:36 PMtaco