#Klee (Genshin Impact) (Japanese) (RVC V2) RMVPE - 280 Epochs
1 messages · Page 1 of 1 (latest)
What pre train did you use and what was the sample rate?
I used default pretrain and 32k
I did it a bit faster, is my first model since 2024 haha
First model since last year? Doesn't sound bad
Thank you so much! Yes i don't do models from march 2024!
How long was the dataset? Also what batch size was it
If i was not wrong it was set to 12, and dataset almost 11 min
Yes it was a bit too high i think the batch size
I used some auto settings
Usually for any audio under maybe 14 minutes I'd do batch 4 since it's such a small dataset
If u wanted u could retrain this exact model on batch 4 to see if it sounds better compared to this one
Yes true i had to select a smaller batch size idk why it was put to 12, i have installed applio for try because before i was train models always on mangio RVC
I think I trained on mangio years ago
Ye I haven't used mangio since rmvpe became a thing
Oh yes, Mangio RVC have rmvpe too, but i think is very good still today because i saw it gives me same performance as applio
that's good
that looks neat
my first ever model was made when rvc v2 wasn't a thing yet
i mean back in may 2023
So cool! My first in RVC was End of June 2023!
This model has been synced with Weights and is ready to use for free!
my first ever model was made June 27th in 2023
it still doesn't sound that bad
mangio is more outdated than the original rvc.
its missing several updates mainline have
i'd highly recommend you to switch to either the og rvc or applio
Oh yes for this i used applio i downloaded recently and trying it a bit
Oh another question, actually do you recommend to use applio or mainline?
In this actual period
that is a good question
in one hand you have mainline rvc, the original rvc, no much i can say about this, it's made by the original rvc team, this is your "true" rvc
and in the other hand you got applio, it's more updated than mainline/original rvc and the code was simplified in order to be more easy to add or delete things
applio should give a extremely similar result if you use the same mel loss between trainers (L1)
the thing applio updated are dependencies and such, no actual code was changed from mainline, for example upgrading fairseq to transformers, but also training is faster and requires less vram compared to the original rvc
Oh so nice! In fact i tried now mainline but the gui is always the same from the start, isn't so much perfect but rvc works the same, maybe applio have a better ui and interface and can be a bit faster, but if in performance they are the same i will keep applio
If is all basically the same
applio it's faster yeah