I don't remember

oof

250 epoches probably

k

you trained it until it went stonks right?

probably

I wasn't really paying attention

and this was two months ago now

yeah

hey guys so I was retraining pepper to finetune him and I got a perfect chart, but then I was too late to stop the thing and it did another epoch that was near perfect but dies at the end

so what do I do about that

do I just continue on with fine tuning or re do it?

it picks a random sentence for each chart, so it's less likely that it suddenly got ruined and more likely that it does good charts 90% of the time

plus the chart only shows what the network is looking at when speaking, it doesn't say anything about how good the voice sounds

dang i wish there was a way to generate test audio with a model while it's training maybe i could try adding that to my training notebook copy

that would be among us

just like how gpt2 training does it where it Outputs Samples every 100 or so steps

yeah exactly

probably do it every save instead of every epoch so it doesn't slow down as much

or i could somehow yoink the sample it's generating for the graph and turn that into audio

I'm waiting an update for talknet training notebook for support for voices in other languages such spanish, portuguese, japanese, french, italian, among others. For now only works with voices in English

i have voices

ok so you'll need to get more data then a few lines, then you will need to get rid of the music via an ai splitting tool

@User i was just wondering if you could roll back the training to a specific epoch actually, because i noticed earlier ones were slight better than later ones. that'd be a pretty nice option to have when you go too far.

not possible in the current notebook, though one could add a thing that saves elsewhere every 50 epochs

or every x epochs i suppose. after thinking about it - because it saves the same file. is the output binary? if not i was thinking what if you make each epoch a commit to a Git repo? that way you could just roll the commit back. if it was binary that's a bit more difficult to have diff updates though.

but you could certainly find the saving code and save the epoch state every so often.

how many minutes of transcibed audio is needed and in what format to submit a dataset ?

The Chris griffin model sounds werid

yeah it's a big blob of hundreds of megabytes of data and diffing almost certainly wouldn't work because neural networks change everywhere at once i have heard of google drive having a built-in versioning thing but i don't think it can be accessed when it's mounted as a regular directory, plus i'd rather have separate files tbh