RVC2 working colab link ? (noUI)

[unkempt sail]

rvcdisconnected does not work anymore. Applio is working but needs much more time to train. Anyone got a working noUI link for RVC2 training?

[fleet pagoda]

what's wrong with applio?

[unkempt sail]

like I said its training needs longer than RVC and I don’t really notice significant quality differences. A 45 mins datasets needs like 5 hours to train to 120ep while RVC needs like 1,5 hours.

People on free colab don’t have much time 😛 I wish I could train locally on my macM4pro but there seems to not be an option for it

[dawn lagoon]

the only no ui working option is Applio on Colab

you could use Applio Kaggle UI (Kaggle is another cloud computing platform, but gives 30 hours weekly of gpu compared to the max 4 hours of colab)

[dawn lagoon]

@karmic umbra may this be because of the fp32 improvement btw?

[dawn lagoon]

yeah i dont think theres mac training support locally, maybe Replay but i dont really know

[karmic umbra]

FP32 is slightly slower, but it all depends on the GPU

Not 4x slower

[dawn lagoon]

might just be the options they used on rvcdisconnected (old mangio) compared to applio then?

because its google colab

[karmic umbra]

but again we are talking about T4 which is way faster

so 5 hours sounds about right for FP32

Damn, I did not think about this

[unkempt sail]

I still wanna use RVC because applio training takes so long

[fleet pagoda]

have you read the above convo? the fp16 vs fp32 thing?

RVC disconnected as well as the og rvc use fp16 precision (enabled for RTX gpus and T4 in colab), and applio uses fp32

fp32 may be slower but doesn't suffer exploding gradients, as explained in:

tl;dr fp32 is generally more stable and less prone to artifacting (mainly caused by exploding gradients)

[unkempt sail]

yes but it doesnt change the fact that it takes longer.... and i was fine with RVC

if there just was a way to train on my mac M4pro 😦

[karmic umbra]

is it apple silicon (ARM)?

[unkempt sail]

@karmic umbra Yes

M4 Pro Chip

[karmic umbra]

there are some issues with math library on ARM, needs to use export OMP_NUM_THREADS=1 to run inference, but I would not try to run training

unless you want to melt it

also MPS is disabled in the latest build due to this nonsense.

[unkempt sail]

so in other words no way to train on mac m4pro? its weird because on github i've seen several people saying they do training on mac

[karmic umbra]

In the latest release it is disabled on mac

but it can be re-enabled

with OMP library limited to just one thread I dont think it will be productinve

[unkempt sail]

i think i have not enough knowledge for this because i don't understand a word

[fleet pagoda]

that's what you need to understand why "applio training takes so long" as you said

[karmic umbra]

To wrap it up. "Applio's training takes longer because it uses higher precision for a better training stability and to prevent the model from randomly blowing up. Unfortunately the free T4 on colab has a terrible performance at higher precision, much worse than less precise FP16".

[unkempt sail]

I put the batch size on applio to 15 now. Using colab on T4. Training is now faster (2 mins per epoch). Much better. It’s running since 1 hour and didnt error yet. Maybe I found the solution for my problem

[unkempt sail]

Or is it dumb to do that? Does it make a difference ?

[karmic umbra]

Do you have 50+ hours of audio to use batch 15?

[unkempt sail]

See, I had no clue that the batch has something to do with how much dataset I have 😂😂😂 it’s 29 mins Data

[karmic umbra]

4 to 8 then

[unkempt sail]

Shit.. but what happens when I keep going? It’s low quality model ?

[karmic umbra]

the results may end up bad.. or okay.. or good, who knows

you may get lucky

[unkempt sail]

It sounded not bad at 10ep

It’s actually a model of a live mic recording of a friend. So the quality is not the best anyways.