seems counter intuitve

this is the default batch size for Uberduck's training notebook

I have no clue what batch size you can get away with

what dataset are you training on?

I'm retraining Hammond.

I feel like I didn't train him long enough because his voice sounds really gravely

I assume if you enable fp16 you should be able to get about 35 in VRAM

maybe more

FP16?

Whats that

a data format

I already made a Skipper voice model (it's on the site now), so I made a Private (Christopher Knights) voice model and submitted it on Uberduck. I'll think of doing a Kowalski (Chris Miller or Jeff Bennett) or a Private (James Patrick Stuart) voice model later to complete the Penguin team (I can't do Rico, because he just speaks gibberish).

imagine rico speaking perfect english

who?

from Madagascar/Penguins of Madagascar

oh ok

Pi, how do I enable fp16

I tried googling it, but got no results

the on the nvidia implementation?

so...?

yes?

he don't know

Pi, I'm an illustrator, not a tech savvy programmer that knows about DNNs

I'm trying to understand what you mean.

oh thx 🙂

link to the colab you're using?

It's the uberduck tacotron training colab.