might switch to uhh the cruise seals iunno

✅ larger dataset ✅ more layers unlocked ❌ good

idk how much of this is the fault of the model vs the dataset vs the me

-ope wait it's doing something

Deepponies? Is that what all the MLP TTS voices are called?

Can Transformers models be called Deepformers?

Deeponies is an app It has more than just ponies but the ponies are very hq

uhm, first spectrogram that came out during training. good or bad?

im assuming bad

its the first one

yes

cant call it yet

wait til 30

set it to this

manually stop training at a good val loss

"best val losses imo sub-0.14 for ridiculously large datasets (like 30 minutes or more), sub-0.10 for 150+ wav datasets, sub-0.09 for datasets between 100-30, sub-0.07 for 30 below "

i have around 300 voicelines

2nd one

That's a good sign, actually If the graph is like that at only the second time, that means it's learning well

Once it gets to below 0.15, check the graphs again They'll be better You'll know it's ready when you get to the target val loss for your dataset size and your graphs look like a diagonal line

@User is it ok if I just let it finish by itself?

or do I have to manually stop it

If you know how many epochs will work, you can set it to that and let it execute until it stops It's really hard to know what exactly will be enough, though, so it's best to let it run and pass the time while still being able to observe it every once in a while You can stop it when it gets below a validation loss of 0.1. Before stopping it, check the graphs and make sure they look like a diagonal line (theres examples in the notebook i think)

thanks im new to this lol also is more epochs good or bad

i have 354 voiceliens

voicelines

Epochs just need to be high enough for your model to get to the right val loss

What's important is to stop it when it's okay

hmmm ok so if its at 250 and i have 354 voicelines that isnt good?

i set it to 1000 💀