you could scavenge for more clips

since anomaly has a twitch you could probably use his vods as data it might be different quality though

~500

i can get more from his videos

he has some where its just talking no music no nothing

and they all have auto generated subs

Whoah, thats alot, how long it took you to get these and transcribe them?

(this one i didn't transcribe a word, just used auto subs)

i think loss was still going down?

when i stopped it was at like

0.15

Good

so can i continue training it and how?

do i just reconnect to the notebook and re-upload my files there

the notebook can resume training automatically

you just need to unpack the dataset again

so this again?

so do it as if you were training a new voice but leave the output directory to where you want to continue training from

yes

Can anyone explain what the other files (besides dataset) are?

train_ and val_filelist both point to

list.txt

in our case, with more data it's common practice to leave out like 5% and put it only in the val_filelist (used for calculating the loss) so we know when it overfits but you can just keep them the same

and output directory is where it saves the model

Thank you so much! So that's just where it writes that info to!

yeah

hi

i need help

what's the issue

please delete this

actually hm header's even worse

im serious