How to know if my model collapsed?

[shell hill]

I made a new model using a 100% better dataset from my last model's dataset but the new model that I made sounds bad compared to my old model, I've a doubt that my new model collapsed while training but Idk how that works so I'm kinda confused rn, can anybody help me with that?

im training a model rn as well, not sure if its goin ok or it collapsed

[round blaze]

collapses normally go very low... this seems fine for the most part

suspicious but

idk

just don't pick lowest points

[shell hill]

I always pick the lowest points 💀

[round blaze]

yeah don't, in this case

actually it's not even the best method in general if you're picking from g/total, it's just the best generator point and not necessarily the best model

check kl and mel

lowest kl/mel points are basically the best versions of your model

at least in terms of being able to reproduce what you gave to the AI

[shell hill]

Ah okie, I got it but that still doesn't help if the model collapses?

[eager flume]

Ayo? @shell hill level 2 !!!

[round blaze]

if kl/mel still goes down after collapses it's still improving in quality

dunno if that's uh

possible

but if it still improves keep going

[shell hill]

So u mean if the model collapses and kl/Mel still goes down, it still gives u the model that u want

But like, is there any ways to 100% avoid model collapse or does it just depends?

[round blaze]

Yes

[round blaze]

Make sure your silences are trimmed correctly I guess. But it can happen randomly at times

Also if you use custom pretrains like Ov2

[shell hill]

I use og pretrain 😄

I use multiple audio files in my dataset and every file's duration is different like 1 sec, 10 sec, 30 sec and 1 min, I used labels in audacity and it automatically did that, but tbh idk if that's ok or not

[round blaze]

It can also be noise or just batch size screwing you over

The first one though I think is unlikely if you clean your stuff, and I'm assuming you do

[shell hill]

I use 8 batch size for a 9 min dataset and I train in kaggle mainline

And the dataset is pretty clean

There are some official stems too

I think i should just merge all the audio files into just one file and train

Coz as I said there are some 2/3 sec files too in my dataset, which can cause the problem too imo 😄

[round blaze]

Mm

Yeah try doing that

[shell hill]

Alright I think I got my answers, thanks for the help I really appreciate it 🙌🏻

[round blaze]

You're welcome

Lmk how the next training goes