✨│ai-help

[simple ore]

yeah, the encoded phonemes and pitch

during training it builds a link between phonemes and spectrogram, during inference it uses phonemes to build a specrogram, and it "fills the gaps"

encoder training masks parts of the sequence and the model has to generate something that matches the original

[glacial pollen]

I don't think your explanations will be any good for newbies in here

[simple ore]

so the process repeats with different parts of the sequence being masked every step until the model can generate the entire sequence on its own

yes, it is a bit advanced

kl almost always goes down and its contribution to training is rather small, so it can be ignored

[steel forge]

Suno to create a beat similar to the artist, "2000s hip hip chipmunk soul, Chicago rap"

Extract the vocals and instrumental with UVR

Infer the vocals with Kanye RVC model in Applio

Combine new vocals and instrumental in audacity. Do some tweaks to the vocal mix and you're done

Just using kanye as an example btw

[waxen jasper]

hey everyone, i have a question, i wanted to make a ai voice singing a music, but when i use rvc, the ai voice is also singing the instruments that are played on the song, so it makes it weird, how do i fix that ?

[low shard]

you have to separate the vocals and instrumentals

https://docs.ai-hub.wtf/rvc/resources/vocal-isolation/

it doesn't automatically do that unless you use aicovergen or aicovermaker or weights.gg

[knotty moth]

seems like just another troll

[brittle wing]

How do I fix this with Applio?

[simple ore]

dont run install as admin

get a compiled version and unzip it into C:\Applio

[brittle wing]

Yeah it installed fine without admin when you suggested, but then clicking run-applio I got this, I also didn't run admin on this.

[knotty moth]

try redownload completely while turning off firewall and defender smartscreen

[simple ore]

download 4.5gb zip

unzip to C:\Applio

not to some other weird folder

wait until unzip finishes

[brittle wing]

I can't seem to find the location of C:\Applio? or do I have to create 1?

[simple ore]

yes, of course

[brittle wing]

Okay, I will give it a try. Also does the compiled version usually takes a while to extract than others?

[simple ore]

if you use 7zip, not that long, if you use windows, it is a slowpoke

so about a minute

[crystal gull]

🤔 I want to have my ai talk in voices using tts, what would be the best setup for that? What app has api support? All local 🙂

[brittle wing]

Okay, seems like the complied version does not come with the install bat files?

[crystal gull]

You just run the run-applio no?

[simple ore]

that's the whole point

[brittle wing]

Okay, its working now. I had to wait a bit longer than expected.

thanks guys!

Whats better in Precision? fp16 or fp32?

what batch size should I apply with 30+ mins of datasets? I have an RTX 3090 24gb.

Pitch extraction algorithm, which one is better for singing?

and lastly, what Index Algorithm should I use?

[analog obsidian]

fp32

[analog obsidian]

16

[analog obsidian]

rmvpe

[analog obsidian]

choose "auto" in applio

[crude flame]

ive used bs 6 on 30 minutes of audio before

and that model is one of my best (in terms of most fire emojis)

[analog obsidian]

too low batch size in big datasets can cause the model to be stuck in suboptimal predictions, and also you're slowing the convergence too much

• noisy graphs

[crude flame]

it started ot-ing at 92 epochs

[brittle wing]

Thank you

[crude flame]

yeah iirc it was wild

[analog obsidian]

thats bad

models should converge at 200 epochs

[crude flame]

i dont have the logs anymore but it sounds fine so

[crude flame]

ive had a model ot at 68 epochs

[analog obsidian]

yes but u made it to be stuck in a bad local minima

so it just overfitted

[analog obsidian]

😭

[crude flame]

[brittle wing]

is it possible to still game while training? if so, lower the resolution?

[analog obsidian]

ideally u want them to converge at around 170-200 epochs, then you train until they start to overtrain

[crude flame]

https://discord.com/channels/1159260121998827560/1255616722480926860 this one

[lavish lintel]

Congratulations Razer by Weights!

Your Grotle is now level 25!

[crude flame]

was 20 min not 30 but yk

[analog obsidian]

yes reduce graphics and caps your fps to 60

[brittle wing]

Thank you

[analog obsidian]

it sounds good because the dataset is good but the model got stuck in a suboptimal place

too low batches causes model to be focus learning one specific thing rather than trying to learn more

[crude flame]

imagine being so good at making datasets that your model is bad

[brittle wing]

Out of curiosity, how many EPOCH to train with? 30+ mins datasets with added pretrained like KLM? batch size 16

[analog obsidian]

dont worry we learns from our errors, today i learned what segment size is thanks to my error xD

[analog obsidian]

set max epoch to 500 and watch tensorboard

if you notice g/total just goes up for over 1 hour, stop training

and select your lowest point in the mel graph before the g/total rising

[brittle wing]

can you provide some examples? if you want

[crude flame]

bs is like the only thing im confused with in rvc, ive heard so much misinfo that im confused on whats real 😭

[brittle wing]

Good example vs a bad example?

[analog obsidian]

past red circle = overtrained

[analog obsidian]

use between these: 4, 8, 16
16 for 12 minutes and above, and decrease it to 8 if the graph is too smooth

[brittle wing]

As long as its some what flatline, then anything after that stop training?

[crude flame]

thx 😭

[analog obsidian]

g/total should always go down, if you notice a rising trend like the image i sent for over 1 hour you stop the training

when the graph rises means the model is getting confused

so nothing useful there

that is the margen of error, so you want the less error possible aka the lowest point

rising is more errors

[brittle wing]

Thank you

[analog obsidian]

only use bs 4 for very small datasets like 5 minutes or below

[crude flame]

a while back (like mid ai hub 1 days) scr used bs 1 and said it actually made their model more accurate to the source with like a set of like 20 minutes or something

[analog obsidian]

bs 1 is pure noise

[crude flame]

i remember testing that and the model came out decent

i think i got lucky 😭

[analog obsidian]

you're just training the model with only noise at that point

[crude flame]

idk how it worked

the only downside i remember is it sounded wobbly

[analog obsidian]

same as your case, model found an suboptimal local minima

then learned from that

and only that

[crude flame]

now i know better 😭

[analog obsidian]

also didn't scr used to inference things that were in his dataset? because doing that is def not a good way to measure generalization

[crude flame]

he did a seen and unseen test

[analog obsidian]

every batch size works technically so like u can sure train a batch size 1 model but you're training pure noise

forcing the model to stay in one place

[frozen ledge]

-rvc

[azure marsh]

📚 Documentation

AI HUB Docs

https://docs.ai-hub.wtf

🍏 Applio Docs

https://docs.applio.org/

✨ More guides

• How to use RVC Mainline Colab by Cauthess
• AICoverGen Colab Guide by Eddy (Spanish Helper)
• Create a model with RVC disconnected (colab) by Angetyde

[frozen ledge]

-colab

[azure marsh]

📒 Google Colab Notebooks

• Applio, by IA Hispano Google Colab
• RVC Disconnected, by Kit Lemonfoot Google Colab
• RVC Mainline, by Hina Google Colab
• AICoverGen-WebUI, by Hina Google Colab
• AICoverGen-NoWebUI [English], by Ardha, fixed by Eddy, Hina and Gdr Google Colab
• AICoverGen-NoWebUI [Spanish], by Eddy, Hina and Gdr Google Colab
• UVR5 NO UI, by Eddy Google Colab
• UVR5 UI, by Eddy Google Colab
• Modified W-Okada's Voice Changer, Google Colab
• 🆕 FaceFusion UI, by Nick088 Google Colab
• 🆕 FaceFusion NO UI, by Nick088 Google Colab
• 🆕 EasyGUI, by Rejekts Google Colab

ℹ️ Note

While the Colab free plan provides up to 12 hours of daily usage, the GPU is typically available for only about 4 hours each day on average.

More commands: -audio, -hf, -local, -kaggle, -realtime, -rvc, /help

[analog obsidian]

tldr; sounds more accurate because its actually overfitted

so thats why the model did not improve much more than just a couple of epochs

[red cliff]

Is there a known reason as to why sometimes if you're throwing a full song size vocal into RVC and there's a long silence somewhere in it, when the vocal comes back in, there can be a lot of artifacts? I was gonna look into splitting the audio up programmatically (I think Applio implemented that?) in hopes that it would help. Maybe it would speed up the overall inference too

[brittle wing]

Should I Cache Dataset in GPU?

[analog obsidian]

no

[brittle wing]

Thank you

[analog obsidian]

enable split audio in applio's inference

[red cliff]

@analog obsidian I'll give it a shot but any idea why the issue occurs in the first place?

[analog obsidian]

idk im not an applio dev 😭 also be sure to always have applio updated to the latest version, current one is 3.2.8 bug-fix

[brittle wing]

I can't seem to get this training to work?

[analog obsidian]

remove spaces from your model's name
preprocess and feature extract again but this time with a new name without spaces

[brittle wing]

Adding "_" to replace spaces work? or do I have to avoid anything like that?

Because I got an error again

[azure marsh]

Hey, Bizarre! Please use the command !howtoask to increase your chance of getting help by structuring your question in a way others can understand better. Also make sure you're asking in the right help channel:

• General RVC help: #✨│ai-help
• W-Okada / Realtime RVC: #🔍│help-w-okada
• AI image related: #🔍│help-ai-art

[analog obsidian]

hmm looks like its something else, im not an applio dev so i can't tell exactly whats going on here
you could try reinstalling the latest compiled version and try again

but better wait for a dev response

[brittle wing]

Yeah Im on the newest compiled version

[glacial pollen]

Which version did you get?

[brittle wing]

The newest 1 from github

ApplioV3.2.8-bugfix

complied version

[glacial pollen]

U used applio before?

if so, worked for ya?

[brittle wing]

Yeah from 3.2.6

[glacial pollen]

Not sure what's going on exactly with the mainline but, if you want, can recommend you my fork of it

a matter of 1 click install and running, is stable and has few nice things to help you train

[brittle wing]

Sure, whats your fork?

[glacial pollen]

Gimme a sec

[analog obsidian]

https://github.com/codename0og/codename-rvc-fork-3

[glacial pollen]

u fast 😩

was about to pack all into zip but guess it'll do

[analog obsidian]

yea because i was just installing it lmao 😭

[glacial pollen]

for now dl it that way

no zips atm

then run install and run fork bat

as always

[brittle wing]

Thanks, i will give Fork a try

[glacial pollen]

thx, lemme know of any potential issues
( should be stable tho

oh yea, and the new gimmicks you'll find at: Trainint tab, advanced settings and at the bottom:
@brittle wing

just in case

[brittle wing]

woaw thats new, whats the best option to pick?

[analog obsidian]

both

avg loss specifically is very good

[glacial pollen]

Basically, the warmup uhh

say you wanna train for 300 epochs ( approx )

then you could try 30 for warmup or 25

as for average.. recommend you first doing a test run for first few epochs and see how many steps you get per epoch

For example, if you have 40 steps per epoch, set the

to 10 or 12

can be even 8.
You get the point, it is some chunk of 1 epoch's total steps
not too much, not too little

it's to get an " avg " performance's metric from that epoch

[brittle wing]

Im training on 30+ mins of singing dataset on KLM 3 32k pretainer. I assume 500 epoch is enough

[glacial pollen]

yeaaa, you can always pause earlier in anything so no issues here

in that case, try 35 for warmup

( or if you wanna stick to 10% of total epochs rule, do that but I'd recommend 35 at first )

effects of warmup on rvc in general aren't well tested in field yet ( on actual pretrains, that is. )

[brittle wing]

Great, ill try

[glacial pollen]

Neat

[brittle wing]

@glacial pollen I got an error for installing, and then got an error for running Fork?

[glacial pollen]

Huh, this shouldn't happen

[analog obsidian]

is it normal that the rest of the graphs are missing?

[glacial pollen]

oh, you gotta type in " total "

in scalars to see em ( in filters tag )

it's normal

@analog obsidian

[analog obsidian]

oh lmao sorry im sleepy again, found them 😭

[glacial pollen]

Can you dump the full log ( from the very top to the bottom ) and send it in dm?

[brittle wing]

Sure, how do I dump the logs for you?

C:\Users\PC\Desktop\codename-rvc-fork-3-main\logs?

[glacial pollen]

just highlight all in the console, paste into notepad, save as txt and send me

also before that

what windows u running?

11?

either way, we can discuss it in more details in dm

[brittle wing]

Yeah im on 11

[glacial pollen]

update: issue fixed. Case's closed

[brittle wing]

@glacial pollen You recommend 35 on the "warm up phase" for 30+ mins of datasets on 500 epoch? What should I put in the "Frequency of avg running loss"?

[glacial pollen]

@brittle wing

pretty much you first gotta run a lil test for few epochs

we need to know steps you get per epoch

[brittle wing]

Should I lower my epoch 500 to something lower for now to test to get the results of warmup phase and frequency of avg running?

[glacial pollen]

you can run for 1 epoch really ( but do 2 )

[brittle wing]

Okay, then should I set the "warm up" + "frequency" on default settings for now?

[glacial pollen]

you can keep both at 0 for now

or whatever def value i left there, doesn't matter just yet

none of the options will affect steps per epoch you'd get

[brittle wing]

Okay, ill test it with 2 epoch. After its done, what should I look out for?

on tensorboard

[glacial pollen]

an example

you wanna check S value for your epoch

in this example you can see it's 25 steps per epoch

consecutively, 25, 50, 75 etc

Once you get what you have, we can think on what to try

[simple ore]

come on bro, you can't train using one 30 minute unsliced audio file

who does that?

[lean chasm]

where do i download the virtual audio cable?

[low shard]

youre following the written guides right?

[lavish lintel]

Congratulations Nick088 [ITA/ENG] by Weights!

Your Charizard is now level 59!

[low shard]

what’s ur pc gpu?

because the vac is already in the written guides, but im feeling youre following some outdated yt tut

[lean chasm]

AMD Radeon 780M

i already downloaded the cable btw

i just don't know how to use the okada with it

[brittle wing]

It’s sliced up into 1 audio file

Wav

[low shard]

eh usable i guess

[low shard]

-rt

[azure marsh]

Interaction has expired, use the command again for a new interaction.

💻 Local Realtime RVC

• Guide for deiteris' modified W-Okada's Voice Changer Client (Fork)
• Guide for W-Okada's RealTimeVoiceChangerClient
• How to use real time conversion with RVC (voice changer)
• RVC TTS

[low shard]

The 1st link is wokada deiteris fork, its better in performance

the 2nd is original wokada

@lean chasm i would highly suggest u to use the wokada deiteris fork instead of yt tut one

[lean chasm]

wait so i have to uninstall the wokada and get this one instead? (they are the same?)

[low shard]

where did you get your wokada? From a youtube tutorial?

Could u send the link of the guide you used or tell me where you got it from?

[lean chasm]

yeah

[low shard]

youtube tutorial one is old

All youtube tutorials are 1 year old

meaning you got an older version of the normal wokada

You should delete that one, and download the deiteris wokada fork

its way better in performance

you just have to read the guide

[lean chasm]

Download NVIDIA
Download AMD, INTEL and CPU

these are the two options, i got a graphic card too so does it count as nvidia

[low shard]

you got an amd gpu, so you should download that

you told me your gpu is amd radeon 780m

or do you got another nvidia gpu?

[lean chasm]

yeah

i got 2 gpu

so should i use the graphic card gpu instead or the amd?

[low shard]

whats ur nvidia gpu?

[lean chasm]

rtx 4070

[latent pumice]

I don't know if its here i should ask this, but i really need to know, Is there a way to use text to speech with RVC?

Cause i have a problem called "My pc is in my brother's room and i don't wanna wake him up" So i wonder if there is something that does the voice from RVC work with text to speech

[lean chasm]

i don't think there is

[latent pumice]

Welp, it was worth asking

[low shard]

Ah that's way better

You should get the Nvidia one then

[arctic willow]

Hello guys, I have some problem with the program, I have no sound

[low shard]

There are different Text To Speech (TTS) AIs:

GPT So Vits: RVC isn't as good as GPT So Vits for tts, but gpt so vits (few shot tts, which means needs just a lil training for models) can't use rvc models (and viceversa), and its only limited to: english, chinese & japanese, if you wanna check gpt so vits instead, read https://docs.ai-hub.wtf/tts/gpt-sovits/

Freemium 11labs: An easy way to do TTS is https://elevenlabs.io/, you can't use RVC model on this but its a mostly premium easy way for good quality TTS

FishSpeech: FishSpeech is a 0 shot (no explicit training needed) TTS, if you got a good pc you can use it locally else use their site

With RVC Models:

RVC is natively for Speech To Speech, but forks such as ilaria rvc mainline & applio have built in tts (using Microsoft Edge TTS to make a generated tts audio, which i suggest you to choose a tts model that is the same gender and language of the rvc model you wanna use, and then convert it with rvc)

If you wanna do tts locally with RVC Voice Models (if you got a good pc):

• You can get Applio in our docs
• While Ilaria RVC Mainline here (no guide as of right now)

If you don't got a good pc you can do tts with RVC Voice Models on cloud:

• Ilaria RVC Zero (Running on A100 GPU, free fasted rvc on cloud) and the guide

• Use Applio UI Colab (with google colab T4 free daily limit gpu)

• if you don't wanna use edge tts, you could try another tts ai from our tts index and use the output as an input in rvc

[low shard]

Which program? Elaborate

There's a lot of different tons of AIs

[arctic willow]

MMVC

[low shard]

If you mean the realtime voice changer for calls, Wokada, be sure to download the deiteris fork from the written guide and to not follow yt tuts

[low shard]

That's called Wokada as it's developer's name

This is the wrong channel

Use #🔍│help-w-okada

And be sure to not follow yt tuts

[arctic willow]

thanks

[low shard]

Yw

[lean chasm]

@low shard i downloaded the nvidia version and its runs on the web?? i thought it was a application

[low shard]

it runs on your pc gpu

it just got a web user interface

WebUIs (like Gradio & Streamlit) are used ALOT on almost every single AI Applications

They are way easier & faster to costumize/build for developers

And most importantly, it can be used on cloud (remote good pc), as many people like me don't got a pc good enough for AI
(A normal application program built with qt or tkinter wouldn't be possible to be shown on cloud)

Dw about it, uses your gpu

[lean chasm]

about the vb cable, i've done like the instruction but i can't use it in discord

i can use it on the web perfectly but discord is not recieving my audio

[torpid loom]

would yall say this is overtraining?

[simple ore]

go to the correct tab of the tensorboard (scalars)

and show all other related charts

[glacial pollen]

#1159290752195633273 is the place you want
( and don't spam-advertise please. )

[dim jewel]

What is the difference between AI covergen/Mangio/Applio/Mainline?

[glacial pollen]

Forks ( different takes on what rvc originally does )

• Mainline is typically more mentioned in context of original rvc
• Mangio was the first fork of rvc ( it's hella messy and outdated )
• Applio is kinda a successor of Mangio but maintained by different people / team.
  Packs the most features and can be considered more useful, modern and advanced than rvc

• covergen is probs only for covers and not training but I haven't used it so can't say for sure, I'd recommend to avoid it as it's most likely either old or too niche

[dim jewel]

Thank you for explainig it

[glacial pollen]

if you need a lil more functionality, recommend you my new fork ( applio based )
Other than that, Applio is your best bet

[wintry torrent]

My internet has been out for 26 hours but its finally back
i finished downloading the models and now when i run python src/webui.py it gives me this, should i be worried about anything

it opened the webui perfectly fine but will that affect anything

should i? and if so how do i

[glacial pollen]

Don't

[brittle wing]

Where should I stop training?

or should I kept it going?

[glacial pollen]

best course is to let it train for longer until you see dips or actual signs of overtraining

then you'd just pick an epoch from before the dip happens

( is why I recommend saving every single epoch during training )

[brittle wing]

Yeah I saved everyone 1 epoch

[glacial pollen]

oh ye, in that case, lemme do some example scenario for ya

It's just one of possible situations

naturally it doesn't ( and won't ) be like that 1:1

but you get an idea

[brittle wing]

Okay, in this case just keep it training? it's maxed out 500 epoch

[glacial pollen]

But then yea, it's a pretty meh scenario anyways because

[brittle wing]

its already finished lol

[glacial pollen]

The loggings you see

are like

Okay so, remember how I mentioned an epoch can have N steps ?

[brittle wing]

Right, i remember

[glacial pollen]

Now, the problem is, applio and rvc are logging in a manner where the actual logging point

references only the last step from a given epoch

so it's biased because

Lemme get an example pic

The green circle, is how it logs

So in reality, epoch could do an awful overal but the last logging ( last step ) could be " good "

or could be the exact opposite

Hence why I proposed average loss in my fork
``It's still not the most ideal approach

(( because in proper scenarios, training models has 1 extra phase in training where evaluation happens, where model's tested on unseen data and then scored appropriately ))

but def better than what it is rn``

[brittle wing]

oh gotcha

[glacial pollen]

Yea

So I'd take it with caution, the metrics themselves

best to follow what I mentioned before + just actively testing the models by inferencing
just follow your ears

[brittle wing]

next time, instead of 500 epoch maybe do 700 or 1000 epoch. since its still showing sighs of training?

[glacial pollen]

I'd say, a good approach is, double or triple your expected training ( total epochs ) time

that's how I do at least

👀

[brittle wing]

👍

[glacial pollen]

Cause restarting the training has it's drawbacks compared to just doing it all in a 1 go

[brittle wing]

Should I wipe out any sorts of that data from the system and do a complete restart?

[glacial pollen]

Tho, why u using covergen

[glacial pollen]

When restarting the training you want to not touch the previous files

essentially what matters ( files created during first training - wise ) is the:
G, D files and tfevents file + most recent epoch ( small weight: .pth model )

[wintry torrent]

I dont want to use UVR i just want to put in a link and have it do everything for me

[glacial pollen]

A link?

wdym

[wintry torrent]

like a youtube link

[glacial pollen]

oh, welp.

[wintry torrent]

it automatically extracts the voice from the songs

is there anything else like that

[glacial pollen]

Keep in mind what you use

is heavily outdated
just so you know

[wintry torrent]

i can tell it keeps asking me to update stuff

[glacial pollen]

No in general, it is outdated

structure wise

[wintry torrent]

if im going to use UVR whats the best webui to use with it

[brittle wing]

Thank you

[glacial pollen]

You don't have to use uvr really, mvsep does the job

[wintry torrent]

whats that

[glacial pollen]

it's an audio separation site

associated with " voice separation " discord - ish

[wintry torrent]

Is it free

[glacial pollen]

It is, if you register an account you can use it as you please with some limits ( file size / length wise ) + you can't use ensembling but that's not important

only drawback ( but imo not as much ) is the queue

Nothing crazy to call it bad tho. That's your best quality bet anyways

bs-roformer / mel-roformer models for separation ( which mvsep does use ) are beasts

My recommended flow of work is:

• 1. Get Applio ( or my fork if you intend to train in future )
• 2. yt-dlp is a nice tool that lets you dl yt audio in best quality the yt's servers provide
• 3. uploading the audio to mvsep to get your vocals and instru
• 4. Using applio for covers

[crude flame]

You can also use this google colab: https://colab.research.google.com/github/jarredou/Music-Source-Separation-Training-Colab-Inference/blob/main/Music_Source_Separation_Training_(Colab_Inference).ipynb

It has more models, some of which are better then the ones on mvsep.

Here is a google spreadsheet with scores on how the models perform ( higher is better): https://docs.google.com/spreadsheets/d/1pPEJpu4tZjTkjPh_F5YjtIyHq8v0SxLnBydfUBUNlbI/edit?usp=sharing

[glacial pollen]

Uuuuu, that's a nice one

[wintry torrent]

Is applio better than the rvc webui?

[wintry torrent]

I havent used a colab before but dont they cost money?

[glacial pollen]

Well, the core idea and principle is the same however, it's more modern and more optimized
aka, up-to-date

[crude flame]

you get around 4 hours free daily

[wintry torrent]

Besides, it looks way more complicated than mvsep

[wintry torrent]

Ill try it if mvsep doesnt work out

[wintry torrent]

Okay thanks

[crude flame]

its not hard at all

[wintry torrent]

Also, where do i get voice models

[glacial pollen]

Well, notebooks / colab is def harder to set if we speak of just drag-n-drop websites naturally

But sometimes it might be worth it
Esp in this case

[crude flame]

Here is a guide for it: https://docs.ai-hub.wtf/rvc/resources/vocal-isolation/#cloud-uvr

[wintry torrent]

Should i download the applio desktop app?

Or the webui

Oh nvm the app is alpha

[glacial pollen]

go for webui

[wintry torrent]

Does it have to be on my C drive

[glacial pollen]

Ideally, ye. Just to avoid any potential permission / path issues

C:\applio\applio's code n shit

ex. ^

[wintry torrent]

"C:\Applio" is fine right

[glacial pollen]

yup

[wintry torrent]

Okay

[wintry torrent]

whats Pinokio

[glacial pollen]

Pinokio?

[wintry torrent]

is it like stabilitymatrix?

[glacial pollen]

Not sure what you're referring to

[wintry torrent]

https://pinokio.computer/

[glacial pollen]

I mean, I don't know this thing / neither used it so can't say much + this chat ain't for that

[wintry torrent]

oh okay sorry

[glacial pollen]

Best to avoid any 3rd party abstractions or whatever like that

as we only provide support for what we recommend

👀

[wintry torrent]

👍

[simple ore]

stability matrix / pinokio are products made by companies trying to take a niche "we make it easier for you to do x", but completely failing to keep things up to date and breaking shit that is not supposed to break

dont use them unless you're a complete dummy

ipad-generation idiot for whom a computer = screen and who can't tell .exe from .pdf

[glacial pollen]

My respect towards you just now increased by 50%

[wintry torrent]

Fair enough but i do have to admit SM is pretty useful, it makes model and package downloading very quick and easy

[dim jewel]

Hi, I have a question.
Regarding index algorithms. As I understand, Fiass is default while KMeans used to decrease the files size for longer datasets. Is there a trade off for using or not using KMeans on longer datasets?

[glacial pollen]

Pretty much

tho I personally always go for faiss I guess

( ignore the subpoints 3 and 4 from the 2nd ss)

But then, imagine if the index was 1-2 gigs ( you never know what kind of datasets people would want to use ) yea
rip memory, rip efficiency

The biggest I've tried to use faiss on, so far and without any issues, would be 48~ mins of data
Anything bigger wasn't attempted by me

[glacial pollen]

I think it was like 200+ or sub 3xx mb

Don't remember, was half a year ago

[dim jewel]

Thank you for help again)

[glacial pollen]

yea it's alright, in fact I highly recommend gpt for explanation of more technical concepts

it's quite good ( and usually accurate ) in abstraction

[simple ore]

kmeans only kicks in after 200,000 samples in the dataset

[glacial pollen]

yooo wtf, I can hardly think of having 200k samples 💀

[wintry torrent]

Finally got applio to work

it has been downloading for 2 hours at 1mbps

I hate my country

Anyway, where do i get models

[glacial pollen]

rip on the dl speed. Feel ya tho, in 2009, in my small town, we'd have 64kbp/s ( some darn awful wireless tube based )

lol

Imagine the pain back then 💀

[wintry torrent]

Dang

[glacial pollen]

Either way, glad it works for ya now

[wintry torrent]

but i mean in 2009 everything was less than a few mbs

[glacial pollen]

Well true, ye

[wintry torrent]

Yeah but i do have a few questions

[glacial pollen]

go ahead

[wintry torrent]

Do i use fp16 or fp32 (3070 8gb)

for inference

[wintry torrent]

Also do you have any recommended plugins

[glacial pollen]

Well, almost all models are done in fp16 so

so I believe, there's no point doing it in fp32, not that it'd matter much anyway

you won't hear a difference in this scenario

[wintry torrent]

Okay

[glacial pollen]

As for pluggins, I don't really use them so can't say for sure

uvr's meh, we got standalone uvr and / or mvsep / colabs

Elevenlabs I don't use

Basically, no point adding em unless you use elevenlabs, got api and stuff ( I'd assume

[wintry torrent]

Ill stick with mvsep and if i have issues ill download uvr

What about voice models where can i get them

[crude flame]

You can search rvc ai voice models at:

• #1175430844685484042
• In #🔍│find-models , Do /find with @earnest musk
• https://weights.gg/ (login required)
• https://huggingface.co/models (but watch out cus in hugging face there arent only rvc ai voice models)
• https://voice-models.com/
• https://thevoicemodels.com/ (for Turkish Models, login required with discord and level 2 on their server)

if there isnt one, you can:

• #1159289738314919936
• #1191429836321849435
• make it yourself with our docs guides https://docs.ai-hub.wtf/essentials/how-to-make-voice-models/

[earnest musk]

:wave: @crude flame, How can I help?

Available Commands:
• @weights find <query> or /find <query> - Search for RVC Voice Models
• /create - Create an AI Cover
• /image - Generate an Image

[crude flame]

go away 😡

[wintry torrent]

Okay thanks

[crude flame]

yeahhh hush mode

[wintry torrent]

Is it normal for the download link to be that long

Also why does everyone have by weights after their name 😭

[crude flame]

meme

[analog obsidian]

wat, fp32 inference with fp32 models have an audible difference vs fp16 model + infeerence?

[wintry torrent]

ah okay i thought it was a staff thing

[glacial pollen]

I'd say it depends?
But tbf, most likely not

biggest thing between fp32 and fp16 is for training really

both stability wise and, well, 'max' potential you can squeeze out of the model ( quite likely given full precision and gradients' representation )

[wintry torrent]

Whats the site u recommended for downloading lossless audio from yt

[glacial pollen]

There's no such a thing as lossless audio from yt

All audio that goes on yt undergoes compression and other postprocessing

[wintry torrent]

Idk why i said lossless

i meant high quality

[glacial pollen]

Best you can do is use yt-dlp

arg -x

[wintry torrent]

Yeah that thanks

[glacial pollen]

usually u get .opus

( -x makes it fetch the best audio for a given video the server has )

[wintry torrent]

Also can you tell me what models to use because i have no idea what these do

[glacial pollen]

Now, I am a lil busy so can't respond rn

will be back in a bit

[analog obsidian]

ohh ok, thanks for having fp32 enabled by default in your fork, i was training a model and forgot to check the default precision 😭

[wintry torrent]

Okay

[crude flame]

https://docs.ai-hub.wtf/rvc/resources/vocal-isolation/#best-models

[lavish lintel]

Congratulations Razer by Weights!

Your Grotle is now level 26!

[glacial pollen]

HAHA, knew it was a good idea

😌

[glacial pollen]

Alr, I'm free now

in terms of that

Razer already sent you docks

[wintry torrent]

What output should i use

flac right

[glacial pollen]

But if you want to hear an opinion from me personally? I always go for bs-roformer

I go for wave
less conversions / encodings, the better for me. like it raw Do not shoot it

[wintry torrent]

Yeah thats what i picked

Okay

[wintry torrent]

Dumb question but how do i put the audio files back together after inference

Also why is the output 68mb

[glacial pollen]

Audacity can do if you're not keen with masterring or mixing or music production

[wintry torrent]

Its gonna take 40 years to download

[glacial pollen]

unless you are handy ( and have ) adobe audition

alternatively, fl studio or any other daw

[wintry torrent]

Is there no online alternative

[glacial pollen]

Because wave is uncompressed, weights it's portion

[wintry torrent]

I literally cannot download anymore things

Egypt really sucks internet wise

Everything wise actually

[glacial pollen]

I mean yea there surely are but think of it that way.. not only you have to upload that sub 70mb wave

then the instrumental

and lastly, dl it all

Better to get audacity or such and call it a day

[wintry torrent]

Whats the smallest option other than mp3

flac?

[glacial pollen]

there's quite a few of things better than mp3

aac, opus / ogg

then there's crappy mp3

as for lossless, there's flac

[wintry torrent]

They arent available on mvsep

[glacial pollen]

If it ain't going for training ( but you use it for idk, mixing or something ), you can use flac

[wintry torrent]

How big would a 3min file be if converted to flac

[glacial pollen]

matters of individual case and compression ratio

[wintry torrent]

Im doing all of this for fun not really going to train anything

[glacial pollen]

Typically

[wintry torrent]

or produce anything

[glacial pollen]

" People often favor FLAC because it takes up significantly less space on their devices. FLAC files can be up to 70% smaller than the same WAV file. "

[wintry torrent]

Yeah thats much better for me

[glacial pollen]

In that case, go for flac

no issues whatsoever

[wintry torrent]

My problem is not with storage its with downloading the file

Is audacity easy to use

The thing i liked about aicovergen is that it did all of that for me

It seperated and combined the voices byt itself

[glacial pollen]

I haven't ever used audacity for mixing so I can't give you any opinion on that

it's rather simple and generic so

if you get some basics, you should do well

but if alignment of 2 files is what you want ( no effects, compression and such - generally mixing )

i.e. Vocals and music

you just put em both in, align to the edge, export and call it a day

so

[wintry torrent]

Why is export audio grayed out on audacity

[glacial pollen]

I can do it just fine

maybe try ctrl+a

[wintry torrent]

Yeah i figured it out

just had to pause

[glacial pollen]

a

yeee

you can't do it midway in auda

[brittle wing]

What’s a proper way to make 48k datasets into 32k?

For testing*

[analog obsidian]

u can either resample the dataset to 32k using audacity/rx studio/whatever
or using an script that uses soxr_vhq which is technically better than the above ^
or just selecting 32k sample rate in applio and let it to resample for you in the preprocessing (by default the resample is done using soxr_hq)
realistically speaking most people cant hear the difference between all of these options, so do the one which is the easiest for you

[brittle wing]

Do you know of resampling from FL Studio with “Edison” plugin does the same thing?

[analog obsidian]

idk but even if the case you did not resample the dataset to 32k, applio is going to do it for you

[glacial pollen]

It's more so about the algorithm that does the resampling than a tool

SoX's having the best currently known algo

[brittle wing]

Thanks, I’ll check them out

[glacial pollen]

Ye, search up sox resampler
( the setting for quality would be 'vhq ' )

alternatively, use my fork as it has it in use

[brittle wing]

I’ll try both

[brittle wing]

They have a GitHub?

[glacial pollen]

https://github.com/chirlu/soxr

[brittle wing]

Awesome, thanks

[glacial pollen]

Actually nope, default isn't using soxr

it's using librosa's default resampling algo, whichever it is

[analog obsidian]

why is not using soxr hq? mainline does 😭

[glacial pollen]

it does? 🤔

As far as I know, my forks were the only ones using soxr, no mainline no applio

well.. in any case, it's a matter of adding

in: root/rvc/lib/utils.py

[glacial pollen]

oh, well, it's commented out

[analog obsidian]

o

no way

😭

[glacial pollen]

also, last time I checked it wasn't there 🤔

or maybe I remember it wrong.. either way ye

can be easily added ✨

[analog obsidian]

i correct myself:
mainline almost got soxr

[glacial pollen]

[sonic agate]

@glacial pollen hi i'm using the correct channel now

help

[glacial pollen]

oh, you picked the wrong one 👀

gimme few mins, pushing the latest changes to ver 3

( cause you got version 1, that one's rvc based, not applio )

[sonic agate]

ohno

lol

i just noticed

[glacial pollen]

shhh

[sonic agate]

?

[glacial pollen]

🤫 let em not know lol

[lean chasm]

can i get a help with the vb audio cable? i followed the instruction for nvidia gpu okada and finished the setting but somehow the cable is not working in discord

[latent kettle]

How do I get Tensor board in Applio

[glacial pollen]

Then copy the link console gives you and paste in the browser's address bar

[latent kettle]

Before starting training
.

[glacial pollen]

well not really

if you're training

paste that in ur model's folder

then like so ^

Paste in the path and done. Gonna open up in browser

[knotty moth]

oh wait, pinokio also copies fluxgym and there seems someone using it in #🔍│help-ai-art having an issue on not getting saved lora checkpoint

[brittle wing]

@glacial pollen I'm ready to try out Fork, I remembered you needed 2 epoch to figure out what to put in "Warm up Phase", and "Frequency running loss"? Is this what you meant?

[brittle wing]

Whenever you resample into 32k with Audacity, when pressing export, do you keep the vocals in mono or stereo?

[simple ore]

rvc uses mono

[brittle wing]

oh ok, best bet would to keep vocals in mono?

[knotty moth]

rvc preprocessing will always convert to mono anyway

[simple ore]

both for training and for inference all the audio converts to mono 16k

training does use a full sample rate files, but still mono

[brittle wing]

Thank you

When should I lower batch size? lower the better? I'm using RTX 3090 24gb, I'm using 16 batch size.

[uneven horizon]

how to run tensor board? i don't see any bat file named such in mangio folder

[simple ore]

if you have python installed as standalone

then pip install tensorboard

and after that you can use it from command line like tensorboard --logdir=X:\Applio\logs

[simple ore]

the og pretrain with 50 hours of audio was trained with batch 16

for a regular model that is a major overkill

and would only lead to shitty results

[brittle wing]

I'm using the new KLM 3 32k

[simple ore]

what's your dataset size?

[brittle wing]

30+ minutes

[simple ore]

I'm sure wit 3090 you can try 4, 6, or 8 and see which one gives the best result

[brittle wing]

Thanks, I'll try it out

[simple ore]

make one folder with a dataset for batch 4

[brittle wing]

I assume 4 would be stupid slow?

unless beefy gpu?

[knotty moth]

batch 8 fp32 occupies around 16 GB

[simple ore]

batch size may affect an overall training speed

one epoch would be the about the same regardless of the batch size

batch 4 - 500 steps x 1s, batch 8 - 250 steps x 2s, same thing

[brittle wing]

Interesting, I wonder what 24gb would be

on average

[knotty moth]

btw do you see difference (quality-wise and gradient stability thing) between batch 8 on single gpu and batch 4x2 on dual gpu?

[flint solar]

There shouldn’t be any difference

As long as the gradients are synced

[simple ore]

also two cards may not actually get 2x faster training because of the sync

[safe sparrow]

how can i run tensorboard in a program that didnt come with tensorboard

[pure gust]

probably a silly question, but i downloaded a model and there is a json file, should that be uploaded somewhere or leave it?

[knotty moth]

not really an issue

[simple ore]

weights.gg has models named as model.pth and model.index, so json is needed to tell what the f it is

[pure gust]

theres only a pth which i used, no .index

[safe sparrow]

the directory, does it have to be the folder that has the events.out.tfevents? or the one just before it?

[simple ore]

either

if you want to see all logs, or a specific model's log in case you have 100 models there

[safe sparrow]

thank you

[simple ore]

100 model logs gonna take a lot of time to load

[knotty moth]

there should be one or few .tfevents file in logs\yourmodel

[safe sparrow]

actually im training a beatrice v2 model, and it has a tensorboard support, but the loss_g is just a straight line so

nevermind, it just updates itself based on the checkpoints

[flint solar]

G loss graph isn’t ur main focus

[safe sparrow]

what is?

[flint solar]

It’s just the average of fm Mel and kl

When choosing the lowest point u will choose it from the loss/g/mel graph

[safe sparrow]

Thank you sir!

[uneven horizon]

Does longer datasets take more time than smaller ones to train for the same number of epochs?

[flint solar]

Depends on ur batch size, but generally speaking yes

[wintry torrent]

is weights.gg's inference fully free??

Like is there no credit system

[flint solar]

No

[wintry torrent]

Then whats the catch

How do they profit

[flint solar]

There is no catch

[hallow thistle]

You can do AI cover on Weights for free. But I'm not sure why you would inference the entire Weights on your PC.

[wintry torrent]

I meant ai covers yeah

[hallow thistle]

Unless you don't wanna wait for the very rare long number queue and in hurry, you can buy their premium.

[flint solar]

Dats how they profit @wintry torrent

[wintry torrent]

I havent waited a single second for queue

[flint solar]

Where do u live

[wintry torrent]

Egypt

[flint solar]

Time zone difference dats why

[hallow thistle]

https://cdn.discordapp.com/emojis/1301008287147364353.webp?size=48

[wintry torrent]

ah makes sense

[knotty moth]

absolutely for the same batch size

[knotty moth]

premium subscriptions, ||but there are also gamified reward system (similar to civitai's credit system but in quite different way)||

[uneven horizon]

What’s the best batch size for 4060 ti 16gb?

[knotty moth]

refer to my comment above

[uneven horizon]

So lesser batch size equals to lesser consumption of vram?

Also which is better higher batch size or lesser for overall training?

[knotty moth]

mostly between 4 or 8, and fp16 (as default choice for RTX gpus) theoretically halves the vram usage of fp32

the difference is just that fp32 may offer little better quality and gradient stability but also slower as well

[simple ore]

fp32 - better stability, less wild gradients (i've seen 30k+ with fp16)

1hr set fp32 batch 8

fp16 halves the vram usage used by the model / discriminators

but that would be something on top of ~4-5GB it takes anyway

so in this case it would be ~7-7.5GB instead of 9

[uneven horizon]

I’m new to this but what is fp32 cause i don’t see any such options in mangio

[simple ore]

kindly delete mangio and install Applio

[uneven horizon]

Can you link it?

[simple ore]

https://huggingface.co/IAHispano/Applio/blob/main/Compiled/Windows/ApplioV3.2.8-bugfix.zip

mangio is oudated and should not be used

[uneven horizon]

Thanks, i’ll check this out once my current training session gets over.

[simple ore]

in applio the default is fp16 (it is okay for finetuning), you can switch it to fp32 in settings, kill the terminal window and restart the app after

as for the batch size, it really depends on the size of the data set.. batch 40 may work with 100hr+ set, but it is excessive for 1hr set. Same as batch 4 may be okay for 10 min set, but not applicable for 10hr+

what's good for finding a tick in a matchbox is not good to find a bowling bowl in a potato field

[uneven horizon]

what do you recommend for 30-40mins to an hour datasets?

[simple ore]

=4 and <=8 generally

[knotty moth]

5k grad/g is the worst of my old models having made using default fp16 so far

[signal bloom]

Does anyone know what the most popular tts people are using?

currently using sapi5

[analog obsidian]

i like the stability fp32 provides but god it takes ages to cook a model with it 😭

[simple ore]

i'd rather have a good model

[knotty moth]

edge tts (free) and elevenlabs (freemium)

[signal bloom]

by freemium you mean like rate limited?

[knotty moth]

yep perhaps

[signal bloom]

got it. thank you vm

[knotty moth]

btw with truncate method and your custom slicer script, it barely causes negative kl but there might be some upward spikes in fm & mel

[simple ore]

or you can install something locally

f5-tts, fish speech, xtts

first two may require some finetuning

also depends on a language

[simple ore]

there should not be negative kl ever

buut somehow it happens during training from scratch with weird models

[analog obsidian]

what it did not caused negative kl for me

[knotty moth]

it was there with the old labeling method and rvc's default slicer

[analog obsidian]

almost 12 hours of training no issues for me

[simple ore]

my attempt of "cement" with refinegan did not go right

[knotty moth]

(almost 0)

[analog obsidian]

last time i faced negative kl was training some very damaged dataset

[simple ore]

I think the formula is messed up or the values calculated by the encoders

it should not be possible to have a negative, and here here we are

[analog obsidian]

rvc is not wokada #🔍│help-w-okada use this channel instead

[tawny nexus]

my bad thanks

[flint solar]

I’ve never had negative kl

😂

[knotty moth]

🤔

[analog obsidian]

i dont even know what causes negative kl

[analog obsidian]

lol

[simple ore]

    kl += 0.5 * ((z_p - m_p) ** 2) * torch.exp(-2.0 * logs_p)

    kl = torch.sum(kl * z_mask)
    loss = kl / torch.sum(z_mask)```

[knotty moth]

though literally no collapses with mute files removed from filelist

[analog obsidian]

must be dataset related then because mine was normal

[simple ore]

i've seen logs_q being way too high that causes that

[analog obsidian]

that using noobies method

[uneven horizon]

Applio training showing 65-85% cpu usage with 80-100% gpu usage consuming about 5.3 gb vram. Is this normal?

[flint solar]

"A small numerical error or negative log probabilities"

[analog obsidian]

aka weird implementation of kl

[knotty moth]

I'll try back to the old labeling method later

[analog obsidian]

is your dataset a bit compressed?

last time i got negative kl was in a compressed dataset

not saying compression causes that

[flint solar]

Is there a better method than labeling?

[analog obsidian]

noobies method

0.5 sec slices

[flint solar]

What IS the noobies method

Ohh

[analog obsidian]

truncate silence + script that slices the dataset in 0.5 seconds of chunks

[knotty moth]

no negative kl in this case, only some "upward spikes" in g perhaps

[flint solar]

Send the script

Ima try this later today

[analog obsidian]

is your cpu amd? last time there was someone with a similar issue

#🔊│ai-development message

[simple ore]

I'm using AMD GPU and because of that some stuff gets offloaded to CPU, but even there it is only 75% tops

[simple ore]

dont comment out hipass filter

use it

[analog obsidian]

is it good?

[uneven horizon]

Yes, any fix?

[analog obsidian]

i dont know honestly, i have an intel cpu

and my cpu usage is fine

[simple ore]

are you using 'cache dataset in GPU' checkbox ?

and I wonder if 'Resizeable BAR' affect it as well

[knotty moth]

I though hi-pass filter can be done by your own before

[uneven horizon]

No, it’s unchecked

[analog obsidian]

oh so is for removing dc-offset

i got lucky my dataset didnt had that

[simple ore]

there's likely an issue how the samples are being moved from regular memory to gpu and back, it may be using some CPU resources

you can try enabling that option and check the task manager's performance tab, as long as the shared memory is not used, you're good

and that should lower the CPU%... hopefully

[alpine valve]

anyone here with decent prompting experience, i need some quick help🙏

[analog obsidian]

@simple ore chunk_len=5.0, overlap_len=0.5 is this good? <

[simple ore]

to use 0.5s slices you need to remove mute files from filelist.txt and to add 50 to the batch size in train.py

[uneven horizon]

Shared GPU memory is around 0.9

[simple ore]

as long as dedicated vram is not at max, it is fine

[uneven horizon]

There’s plenty left

[simple ore]

i've shown the code used in rvc

they've implemented something weird instead

[knotty moth]

other distance functions:

[latent kettle]

Is there any applio hugging face space is available?

[distant turtle]

-colab

[azure marsh]

📒 Google Colab Notebooks

• Applio, by IA Hispano Google Colab
• RVC Disconnected, by Kit Lemonfoot Google Colab
• RVC Mainline, by Hina Google Colab
• AICoverGen-WebUI, by Hina Google Colab
• AICoverGen-NoWebUI [English], by Ardha, fixed by Eddy, Hina and Gdr Google Colab
• AICoverGen-NoWebUI [Spanish], by Eddy, Hina and Gdr Google Colab
• UVR5 NO UI, by Eddy Google Colab
• UVR5 UI, by Eddy Google Colab
• Modified W-Okada's Voice Changer, Google Colab
• 🆕 FaceFusion UI, by Nick088 Google Colab
• 🆕 FaceFusion NO UI, by Nick088 Google Colab
• 🆕 EasyGUI, by Rejekts Google Colab

ℹ️ Note

While the Colab free plan provides up to 12 hours of daily usage, the GPU is typically available for only about 4 hours each day on average.

More commands: -audio, -hf, -local, -kaggle, -realtime, -rvc, /help

[flint solar]

Gotcha

[simple ore]

setting environment variable like this seems to lower AMD CPU use significantly

need to close the app and restart it in a new window

[uneven horizon]

How do i use custom pretrained models in applio?

[simple ore]

check custom box, select custom G and D files

[uneven horizon]

Where do i put the files cause i don’t see the select option

[simple ore]

X:\Applio\rvc\models\pretraineds\pretraineds_custom

[glacial pollen]

well, technically just 1 to see. but ye

in your case you can set avg value to 23 or 30
( and warmup is up to you but I'd stick withhin min - 5% of total epochs and max range - 10% of total epochs )

[low shard]

• Applio ZeroGPU (Unofficial, some features may not work)
• Applio (cpu, official)

[latent kettle]

Also uvr hugging face space?

[low shard]

wrong help channel, well we don't have one for that

but replied to u in #🧬│ai-chat message

[low shard]

• https://docs.ai-hub.wtf/rvc/resources/vocal-isolation/#uvr-zero-gpu
• https://huggingface.co/spaces/r3gm/Audio_separator

the first one is made by @viscid moss , i think he updates his uvr much

the space should too ig

[viscid moss]

Well.. the HF space, not yet. I'm waiting to add the last missing model, to make a big release

Yesterday, 17 new models were added to audio-separator, which is the core of UVR5 UI. But there is one more that needs a workaround to work.

[low shard]

goodluck

[signal bloom]

any recomendations for generating more natural sounding audio using edge tts. Looks like it doesn't support SSML

[latent kettle]

@simple ore can you please tell me how to see tensor board correctly?

[simple ore]

run it, go to scalars tab

[latent kettle]

Then ?

There are too many graphs 📊

Some are going up some are going down

What to do ?

[simple ore]

https://docs.applio.org/applio/getting-started/tensorboard

[latent kettle]

Thank you 😊.

[ionic canopy]

So, out of curiosity, can yelling be apart of a dataset?
Like let's say, Eren Jaegers yelling mixed with his talking
Do I just need to put both together in their own group?
Like all the talking lines first, then the yelling?

[latent kettle]

I want to stop training on current epoch (65) how do I stop it in applio ?

[simple ore]

if you're saving every 10 epochs, then stopping right would get you epoch 60 model

if you chose to only save the final model, then it wont be saved until the very end

[latent kettle]

Oh no..

It was just hiding. Now I see it.

So I have to wait for at least 100

[simple ore]

do you have .pth files other than D/G in your model's folder?

[latent kettle]

Do over training detector works in applio ?

[latent kettle]

Yes I have one. Mymodelname_50e_5400s.pth

[ionic canopy]

Do you think you could answer the question I had?

Please

[simple ore]

do not include yelling in one file with normal speech

otherwise the normal spech would normalize to nothing

[loud condor]

how to update applio?

[simple ore]

download new version, unzip to new folder, move audios and models over, delete old folder

[loud condor]

thanks

[dim jewel]

Hey guys, I have a question. In theory, If a model has an accent, can more epochs decrease it?

[fast phoenix]

oop

yeah im very beginner at this

[brittle wing]

10 hours in, keep training?

[crisp void]

Anyone help me resume training?

[azure marsh]

Hey, TwoOne! Please use the command !howtoask to increase your chance of getting help by structuring your question in a way others can understand better. Also make sure you're asking in the right help channel:

• General RVC help: #✨│ai-help
• W-Okada / Realtime RVC: #🔍│help-w-okada
• AI image related: #🔍│help-ai-art

[crisp void]

!howtoask

[patent trellis]

How To Troubleshoot

__**GIVE CONTEXT.**__ 📝

• Don't simply mention your issue, like "my rvc is not working".
• Describe the step you are on, what you're trying to do, the RVC you're using, a screenshot, etc.
• The more context, the better.

__**BE POLITE.**__ <:matsuripray:1159685390156967936>

• Don't be desperate. You can ping a Helper, but if they ignore, they aren't available/don't know the answer.
• It's okay if you're frustrated, but don't take it into this server.
• Don't DM without prior consent.

__**BE PRODUCTIVE.**__ 🤝

• Don't ask for every little instruction. Put your own effort & test things by yourself.
• Don't ask to ask.
• Check if your answer is a Google search away/on our guides website.

If you've changed your mind about your message, feel free to delete/edit it.

[pseudo dagger]

It was here that people were working in a text to speech ai? cant find a good one that uses rvc models

[brittle wing]

I believe there’s 1 that is integrated in Applio

[idle stag]

So my applio has been working perfectly fine till right now "RuntimeError: cuDNN error: CUDNN_STATUS_NOT_SUPPORTED. This error may appear if you passed in a non-contiguous input." I saw someone saying to use the split audio option but that didn't work I ran the "install.bat" file again but that didn't help either what should I do?

Fixed it I was just running out of memory

[crisp void]

The model has completed 300 epochs, and I would like to extend it to 500 epochs. How can I continue training from epoch 300 without starting over?

Could someone teach me?

[simple ore]

what about other charts?

you're squeezing all the juices out of it

[simple ore]

depends on the application you're using

[brittle wing]

They look similar

[simple ore]

(personally I'd stop at 30k lol)

[pseudo dagger]

How do i add rvc voices to aplio in pinokio?

[brittle wing]

Alguien que hable español que me ayude a tener un buen modulador con voz de chica que no parezca taaaan robotica? porfavor

[glacial pollen]

I don't think there's anyone using pinokio so ( afaik

we don't quite officially support it

[pseudo dagger]

Ok im using it, i had a lot of problems installing so a just pinokio'ed it lol, so how its done in the aplio normally? it needs to be in a specific folder?

[glacial pollen]

I mean, you could have just ask for some support really

In applio, you'd place em in /logs/<model's folder>

within the folder, along with the index

[pseudo dagger]

My aplio logs has only one folder called mute

[pseudo dagger]

I always have a lot of throuble when it comes to A.I.s and i dont want do disturb anyone, just ask when im really in trouble kkkk

[glacial pollen]

Well then that's that. but keep in mind you're having a higher chances of support if you use tools that are known / used here

naturally

you see, if 90% of people use X1, hardly anyone will want to dive on their own into X2 to help 1 person

and as most of us or me, are against crappy automations ( pardon my lang ) the chances are even smaller

Imo Pinokio is for lazy people who aren't willing to learn a bit to do things right

That's a lil scary considering such people intend to work with artificial intelligence

it never was meant to be easy or easily accessible with no effort put into it tbf

That's like asking for troubles in 5-10 years

[pseudo dagger]

Thats me lol, im using it to make some memes as the site i was using got shot down, dont want to go a whole programming lesson just to make some memes

[glacial pollen]

ig, if reading up a bit of text that'd literally take ( at worst ) 10 mins is what you call programming lesson

then I suppose, you shouldn't be using AI for memes

¯_(ツ)_/¯

it's 2 clicks man, 2 clicks

1 .bat for installation, 1 for running

lol

And yet here you are asking about fixing N problem on an unknown site or a service
Taking you more time than it'd if you read a bit of instructions

I hope you get the point

[pseudo dagger]

Well thats news last time i tried it, needed to instal a lot of things write some texts in the prompt of command and deal with a lot of problems as they showed up

[glacial pollen]

Well, never hurts to ask

or to check the repository man

[glacial pollen]

That's what the help channels are for

So you know how to deal with them and maybe even help other users, if at one point you felt like it

it's a basic skill anyone should have, problem solving

Man, I'm kinda scared what's gonna happen to this new generation

Can't imagine ( no offense ) such people operating atom/nuclear-powered facilities in 20 years

[crude flame]

dw those same kids are going to be running our gov in like 30 years

[glacial pollen]

welp, good thing I'm not from US

nah, half joke

[crude flame]

still effects you

[glacial pollen]

Either way, Soryu
If at one point you changed your mind and actually decided to give applio a go, let us know
Always open for support as long you need it

[pseudo dagger]

Man if am using something no one here uses isn't a good thing actually? i can solve problems of other people that you guys cant help

[glacial pollen]

I suppose? not that I'd expect much of support towards tools not associated with the server anyways

[pseudo dagger]

Also i finded the solution, works the same as the normal aplio just needed to drop the file in the download section of the interface

[glacial pollen]

so if you wanna volunteer, go ahead

[glacial pollen]

well, in normal applio you'd do it manually unless you're lazy

same as it always was with rvc

But congrats on figuring it out ✨

[pseudo dagger]

Can do this too there actually, but i just droped the file no created a new folder for it

[brittle wing]

alguien español?

[low shard]

Please speak English here as it's in the server rules

[simple ore]

step 1) download the compiled applio version off Huggingface
step 2) unzip to C:\Applio
step 3) unzip your model into C:\Applio\logs\yormodelname folder
step 4) use run-applio.bat to start it

how hard is what? Why do you need pinokio for that?

[hallow thistle]

Please speak English, or speak Spanish at #🌍│español

[hallow thistle]

Isn't what it called a skill issue?

[solemn shell]

cant do an interference on applio

i followed the steps for amd gpu

the applio opened but is in a infinite loading to do an interference and are not using my cpu or gpu

[glacial pollen]

u sure you didn't accidentally click on or have the console in focus?

[solemn shell]

the terminal is not on focus

its says: Compiling in progress. Please wait...

after I try to interference

[uneven horizon]

If i’m using custom models in applio do i need check custom in embedder model tab?

[simple ore]

please go back to the install guide and read what it says at the very end

[simple ore]

only if you have a model that requires a custom embedder, I doubt there are any in the wild

[uneven horizon]

So just leave it at contentvec?

[uneven horizon]

Will this let me use the pretrained custom models?

[simple ore]

contentvec is the default feature extractor

pretty much all models use that

[knotty moth]

models using contentvec are compatible with original rvc

[glacial pollen]

rvc's hubert is in fact contentvec 500

so yes, they are

well, or should I say " contentvec's 500class model "

https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI/issues/2121

[uneven horizon]

What is Hop Length and what does it do exactly?

[glacial pollen]

@uneven horizon
simplifying / abstracting it without going into too much details:

[brittle wing]

uhh, I left my PC on and activated sleep/hibernate mode on my PC and its still training?? When I went to work and thought nothing of it.

possible that it can still train during hibernate mode? My PC is chill and not hot or anything.

I'm 1005 epoch in lol

Here's 21 hours of training, when is it over training and should be stopped?

[latent kettle]

Loss D ?

[brittle wing]

Here ya go

[latent kettle]

https://docs.applio.org/applio/getting-started/tensorboard

Read it

[clear mesa]

hey, I was curious, if you add more audio into the dataset, do you have to start training the model from scratch or can you just continue and enhance the existing model?

[latent kettle]

How do I resume Training on Applio

[brittle wing]

I hope I understood it, near after 16k is over training?

[latent kettle]

See all parameters G total loss d total loss kel ml

[knotty moth]

batch size 40 on short dataset? 💀

[brittle wing]

No, it's on Batch 6 of 30+ mins of Datasets.

[latent kettle]

Really

[brittle wing]

Yeah

rtx 3090

[knotty moth]

also you have resumed training with different batch size than before

[latent kettle]

I think there are no symbols of overfitting ?

[brittle wing]

Did I have to remove the logs from previous trainings? I believed I started a new one

[latent kettle]

But how

[brittle wing]

Meaning, it can still be trained?

I left 1500 epoch on it

[latent kettle]

Can you tell Me how do i resume training

On applio

@knotty moth

[brittle wing]

I wonder too

[flint solar]

U need ur logs folder

Use the same model name, and sample rate dont preprocess, dont extract features

Use same batch size and click train

[latent kettle]

Okay

[flint solar]

On rvc disconnected u need to type 23333 in the g/d number cell

I believe

[latent kettle]

Ohh. Thank you. But I'm training on applio

[flint solar]

Bet

[brave ermine]

is there any good tutorial to train ai model

[low shard]

what's ur pc gpu

[brave ermine]

4050 laptop gpu

[low shard]

How much vram

[brave ermine]

6

[low shard]

Not really the best

You can try either Local or Cloud

Local:

• Applio: A fork of RVC with some extra features like Applio TTS, kinda faster and simpler but same quality tho
• Mainline: The original RVC

You can train RVC models on cloud (remote good pc):

1. Prepare the Dataset
2. Setup RVC:
   Choose a cloud way to use RVC,

• Google Colabs (4 hours of daily gpu for free, not much hours, but easy to use):
  • Applio (ui)
  • Mainline (UI)
  • RVCDISCONNECTED (no ui)
• Kaggles (a bit harder to use and needs phone number but gives 30 hours weekly of better gpus):
  • Mainline (UI)
  • Applio by Vidal (UI)
  • Applio by Shirou (UI, no guide as of right now)
• Lightning.ai (Kinda hard, needs login, no issue with web uis or anything, but only free 15 credits monthly):
  • Applio (UI)
  • Mainline (UI, No guide as of right now)

Google Colab = Easier but risk of getting disconnected
Kaggle = Harder but way more gpu time

3. Be sure to know about the tensorboard

If you are looking for the easiest way and for free, is using https://weights.gg which ofc uses RVC

But I think Cloud would be the best

[brave ermine]

thanks ill take a look

[latent kettle]

What is Mel reformer?

[sonic agate]

a vocal separation model

[neon rover]

Voice cuts off… what to do?

[sonic agate]

where?

[neon rover]

on discord

[low shard]

wrong channel

-> #🔍│help-w-okada

[low shard]

yw

[glacial pollen]

you can't quite tell it tbf but

around 16k it's more or less where you should stop as past that the performance as you can see is regressing

if you want more accurate results, get my fork with averaged metrics

[hallow thistle]

A converted audio you inferenced on RVC can be cut off to silence abruptly if using a bad RVC model.

[glacial pollen]

as then you can see avg performance throughout epochs themselves too

so you can "more or less" see how an epoch performed ( still, on it's own data but better that than what stock logging on it's own is )

ofc, normal losses are still there too

but you can already see the differences

Stock behavior of logging is to log given epoch's last step's performance where averaging does log the loss over n steps ( of your choice ) within that epoch

reason I mention it is because stock loggings are hella inaccurate. Example;
Imagine your epoch is 67 steps, the logging takes place on step 67, that one could be great metrics wise but 80% of the steps in that epoch display rather mediocre or bad performance. You get the point

Naturally, having a proper evaluation phase during training would be the most ideal, where aside of training and own-losses, losses based on how model performs on unseen data ( evaluation set ) is also measured. That'd showcase the model's generalization. yet, we don't have that ( at least yet )

[glacial pollen]

ps. batch size of 6 might not be the most ideal option here ( esp for 30 mins), I'd highly recommend trying out 8, it's more balanced and since 8 is a number that is a power of 2, the performance of training is somewhat better as parallelism in a sense is in your favor

[brave ermine]

i succesfully trained and tested my voice model it worked well i used 200 epoch and 5 minutes data but i wonder howmany epochs and howmuch data length is ideal ? im looking for any tips for newbies

[glacial pollen]

You'd basically want to use tensorboard

[glacial pollen]

Looks like so

[brave ermine]

i looked over that but i didnt understand anything

[glacial pollen]

once you learn to evaluate what's going on with ur model on graphs, you can def improve ur models' quality

[glacial pollen]

yeah I can help if you're willing to read a bit and dive into it
( worth it tho

[brave ermine]

im not aiming to be a professional but i wish for better

im just using this for trolling my friends not business

[glacial pollen]

it's not really what professionals do

it is just what's used in all machine learning cases ( well, most, there's also keras stuff

cause " I'll train for N epochs as I think its's good " was not and won't ever be a rule to follow sadly

[brave ermine]

thats why we save bunch of epochs

i see

i mean checkpoints

[glacial pollen]

I mean yea, saving every single epoch and testing em is an option, but a pain in the ass tbf

[analog obsidian]

is a bit confusing at the start but honestly a couple of reading and you'll get it pretty fast

[glacial pollen]

you see.. if you don't wanna go that far into metrics, you can just follow simple rules~

1. Lower = better.
2. if it keeps on rising and keeps that tendency for a while = bad

it

is a nobrainer once you dedicate like 15 mins of your time into understanding it ( even basics will do, and you're already well prepared for most ml trainings), my dude

Imo reading basic graphs is a basic skill most of us should have in 5-10 years

[brave ermine]

ill check few tutorials i guess

[crude flame]

the tensor board is actually easy after a 10 ish minute read, trust me i actually used to be exactly like you and think that the tb was only for professionals

[brave ermine]

or ill learn by trial and error

[glacial pollen]

I mean, I can simplify it all

[analog obsidian]

should take you a couple of minutes to understand them, don't worry, does not require years of machine learning knowledge to understand them

[glacial pollen]

read this up

but in a short;
if you get my fork, evaluation of your models gets pretty easy

you get to more or less see how that one epoch does, in terms of performance

[brave ermine]

imma try my best

[glacial pollen]

Then you'd have like, two steps.
Normal graphs ( hypothetical scenario )