since the gpu would have to load bits of it

and afaik it runs through pretty much the entire model on each timestep so that'd be a lot of swapping

do you know if its possible to shrink a model?

a more practical solution is to reduce the precision of the model's parameters from 32 bit to 16 bit

it's what tensorvox/deltavox does

but also tensorvox runs on cpu so

yea, the only issue I have right now with CPU is it uses so much

I have a pretty good CPU, and it hit 100% usage for one sentence

I was really hoping it could render sentences with like 10% usage per sentence or something

you could try to set priority to low in task manager

idk how much that does

wonder if ml frameworks like pytorch and tensorflow could possibly have support for other brands of gpus for radeon friends

even more so, I wonder if there is a way to kinda do what tensorvox does but a little smaller scale. Just shink the data make it more performant. Im pretty sure its converting pytorch to tensorflow

because at the end of the day you have to load each model into ram, right?

So if you have a bunch of model, Lets say each model is crazy high or something, 700mb, and you want 10-15 models. you are limited to the GPU's ram

where as the CPU could have a bunch of models (limited to RAM not VRAM), but inference takes forever

wait is a gpu parallel enough to synthesize with this many models at once? if not then it can just load them when needed

I'm trying to test the limits, and on a 1070 TI I've loaded 10 of the same models into the GPU, its using around 4gb of the ram. I did a batch of 10 requests at the same time and they didn't have any issue rendering them. Im guessing thats synthesizing in parallel, no?

check how fast it is compared to running them sequentially

actually don't know, ill have to run that

current voice project: Nolan from Donut Media

Another thing about my model. I can't tell what I should do for Japanese. I assume the input would use hiragana and katakana because of my character list allowing it, but I want to be sure considering how TALQu uses Romaji

Of course — Train for as many steps as it can be trained for.

retraining Sheen in TalkNet for like the third time now

gonna let him get to at least like

4000 steps

well he's far from perfect, but it's at least an improvement

how do i create a dataset?

also, how much snippets of a voice do i need?

see tutorial in #841422801965416538 and it's best to get as many snippets as you have the patience to but about 5-10 minutes is a good start