How can I measure latency in an AI model?

[fluid cypress]

Also I'd like to know the graph the model follows as it develops and trains, and whether it would be a bell curve or a geometric graph, and if the latency worsens with overtraining.

[rocky dawn]

what kind of AI model are you talking about?

and what thing are the "bell curve" and the "latency" doing with the model?

!howtoask

[simple jetty]

❓ How to Ask for Help

✅ Before You Ask!

1. Check Docs & Guides: Your answer may already be in the AI Hub Docs or the https://discord.com/channels/1159260121998827560/1159513888199540817 channel.
2. Search the https://discord.com/channels/1159260121998827560/1192011222023950368 : Look for existing posts that solve your issue. Do not invade someone else's post.

📝 How to ask?

Tell your:

• Full GPU Name: (e.g., NVIDIA RTX 3060)
• Operating System: (e.g., Windows 11)
• Detailed Description: What were you trying to do and what went wrong?
• Tutorial Used: Link to the guide you were following.
• Screenshot: A picture of the full error message is very helpful.

🚫 Prohibited Topics (We Will NOT Help With These)

To maintain a lega, safe & ethical community, we will NOT provide help for:

• (E girl, as an example) catfishing/trolling, scamming, impersonation.
• NSFW/Porn.
• Any illegal activities.
  Requests for these topics will be ignored and may result in moderation action.

<:matsuripray:1159685390156967936> Community Expectations

• Be Polite & Patient: Our helpers are volunteers. You may ping the Helpers role once.
• English Only: Please keep all conversations in English.

[fluid cypress]

I'm not sure which model I'll be honest... I think RVC Crepe v2

[fluid cypress]

I mean, would the development of the voice model follow a bell graph or not. Like, would the quality/latency increase as the voice is overtrained or would it continue going down, but slower?

[rocky dawn]

rvc v2 with crepe feature extraction?

[rocky dawn]

wdym with the bell graph?

[fluid cypress]

It says "RVC Crepe v2 + persistent storage / autosave"

[fluid cypress]

I mean would the development of the voice model follow a bell graph or a geometric/logarithmic one

Because I've heard when overtraining a voice model, it makes it sound worse, but I'm not sure if it changes the latency

[rocky dawn]

please link the reference, since I don't recall that term in the rvc development we're having with the Applio dev team and the original by rvc boss

[fluid cypress]

Of course

[fluid cypress]

https://colab.research.google.com/drive/1TU-kkQWVf-PLO_hSa2QCMZS1XF5xVHqs?usp=sharing&authuser=3#scrollTo=2pShUKPuERDq

Does this clear the air a little more?

[rocky dawn]

that colab came from the Mangio RVC fork dev that has discontinued the development since 2 years ago

so it won't work anyway

[fluid cypress]

Whaat D:

Gosh this is most upsetting..

Well- anyhow, are there any easy to use models you'd recommend?

[rocky dawn]

still not understand what you mean the "bell graph"

the tensorboard graph isn't like so

the g/loss graph usually goes down and eventually saturate where there's no noticable improvement as you test the model checkpoints

[fluid cypress]

Okay perhaps its better if I showed you

So, if my understanding's correct, the ai voice model becomes worse after overtraining, correct?

So by logic, it'd follow a graph like this, right?

But I'm not sure if I'm correct in my assessment on whether the latency worsen or not when developing the AI

Or rather, the overtraining just becomes "useless" rather than hurt the model itself

[rocky dawn]

no, still not understand how the tensorboard graph could have bell shape like that

and how you'd notice overtraining point with it

[rocky dawn]

refer to this @fluid cypress

[fluid cypress]

So overtraining doesn't make it "worse" , but rather, just useless?

Because I think I can recall a guy some guy talking about how their model got worse after overtraining

But again, not sure

[rocky dawn]

aren't you actually looking for training (finetuning) a model?

[rocky dawn]

have you actually tried training instead of a mere theorycrafting with the "bell curve" crap?

[fluid cypress]

I trained one like, two years ago with that google collab one..

I should've probably explained my motive, sorry if I'm being a little difficult

I'm trying to run an experiment on

Well my research question is "What is the optimal training duration for an AI voice model to minimize response latency based on a mathematical model of performance in a constant environment with similar internet speed and background activity?"

[rocky dawn]

the RVC trainer should have tensorboard graph you could check while the training is in progress

[fluid cypress]

I see..

[fluid cypress]

And where can I find that?

[fluid cypress]

Also is it simple to use?

[rocky dawn]

you should consider the working RVC colab like this

https://docs.aihub.gg/rvc/cloud/applio-colab/

[rocky dawn]

I don't still quite understand your pointless babble without actual results of your trained models

[fluid cypress]

Right, well, nevermind that for now, could you tell me how I can measure latency?

[rocky dawn]

that's why you should try Applio colab I suggested

[rocky dawn]

(no, not the "measured latency" crap, but rather like the model testing result quality)

[fluid cypress]

Are there any resources that outline how I can properly use it? Maybe a tutorial?

[rocky dawn]

@fluid cypress

[fluid cypress]

Isn't that the program itself?