Download:
https://huggingface.co/V-Mikey-V/TF2-Engineer/resolve/main/TF2-Engineer.zip?download=true
Dataset: 20:41
Sample Rate: 40k
Pretrain: OG Pretrain
Batch size: 8
"I've seen better sides of beef, been run over by a combine."
#TF2 Engineer (RVC V2, RMVPE, 100 Epochs, 8,200 Steps)
vast summit
keen stream
i feel like this model would sound better with more epochs
shut geode
true
vast summit
the graph went flat pretty early
first time this has happened this quick really
I'll try to run it further and see if i get any more progress on it
the inference doesn't sound full enough but when using Vonovox it sounds pretty good.
shut geode
don't use the graphs to let u know when a model is ready, it's best to test until the model sounds just right to you
your ears are your best friend
haven't used tensorboard in a year
unborn jasper
the graph went flat pretty early
don’t judge how good a model is just by looking at the graphs
graphs can look fine, but the model behind it might still perform poorly
use the graphs to spot potential unstability and issues in the training, thats what theyre meant for
vast summit
trained for 200 more epochs, it does sound a bit better but still on inference it doesn't sound as good as the voice changer
gonna try a new training session using a different pretrain to see if i can get better results
shut geode
If you're using any pretrain that isn't og use legacy v1.5
vast summit
If you're using any pretrain that isn't og use legacy v1.5
this one wasn't using any pretrain
its as the desc says
vast summit
If you're using any pretrain that isn't og use legacy v1.5
this is the one i was gonna try next actually
shut geode
this one wasn't using any pretrain
Yea default is set to og pretrain
Which is the one u used
vast summit
ye
vast summit
wow this new training session is already way lower than this model at only 40 epochs so far
shut geode
wow this new training session is already way lower than this model at only 40 ep...
legacy tends to sound better quicker for whatever reason
unborn jasper
legacy tends to sound better quicker for whatever reason
what, the model learns the features too soon?
shut geode
not that, it just seems to me it sounds good at around 100 epochs normally
unborn jasper
aaaa ok i got worried for a second lol, so it learns normally, oof
vast summit
ok yeah this new version using leg 1.5 is a lot better
sounds much smoother and less robotic on inference
and it seems high pitches are better
also effs and esses aren't getting confused with each other in vc like this model
shut geode
legacy 1.5 is better than og :3
keen stream
legacy 1.5 is better than og :3
only sometimes
