✨│ai-help

[glacial pollen]

And the average loss from my fork ( you check given epoch's performance )

and if it clicks for ya ( you can tell it does well ), you pick that one

[brave ermine]

my small brain and english already melted sadly

[glacial pollen]

Hmmm.. well, can always go live-stream in a call and show you what's what.
You can just listen if you wish

But either way~ that's that

[brave ermine]

if i understand these by myself it will be better i guess

[glacial pollen]

True

In that case wait

[brave ermine]

ill check tutorials and your messages

[glacial pollen]

Got some material prepared for such cases

A lil outdated but should still do, for the basics' sake

[brave ermine]

damn

[glacial pollen]

It ain't much man cmon

it's less than typical 2-3 min tiktok comments scroll

[brave ermine]

probably but i am totally new right now

[glacial pollen]

I tried to make it as noob friendly as it can be

even using some abstractions

[brave ermine]

there is so many graphics in tb

[glacial pollen]

well that's the reality

[glacial pollen]

or you can consider my rare as hell offer

just happens to have a good test subject rn ( training a model rn

[brave ermine]

maybe you can create a youtube video about this

[glacial pollen]

uhhh, I did have that in plans, ages ago

am lazy ngl

[brave ermine]

instead of teaching ppl 1by1

[glacial pollen]

yeah but sad reality is, majority of folks aren't even willing to read a bit of text

and I just feel like my effort would go down the drain lol

Anyways, like said before. That's all I can say or do atm. Rest is up to you buddy

[brave ermine]

thanks for help anyway ill figure out these by myself

this is just day 1

[glacial pollen]

Np and alrighty, best of luck ✨
yuh, take it slow in your own pace

( courage msg: Took me 1~ month to get handy with it, when I first started

so dw if you can't at first

[brave ermine]

i already created one model from my friends discord calls

is links allowed?

i covered a song with it maybe u can take a look

[glacial pollen]

#1159290752195633273
unless it's for the context's sake

or something

[crude flame]

-# not me taking 3 months just to download the tb

[glacial pollen]

A, if you want me to evaluate it / review, yea go ahead

[glacial pollen]

[brave ermine]

https://streamable.com/nllaxe

[glacial pollen]

Let's see

that's a 32k model right?

[brave ermine]

The sound is a bit muffled because it's from a discord conversation

[glacial pollen]

a

[brave ermine]

i didnt touch that thing

probably 32k

[glacial pollen]

seems like it, but discord itself can degrade audio quite a bit at times

Either way, it's quite good actually for your first model ( is it?

[brave ermine]

yeah i excepted for scream sounds but i got actually song

its my first model but i already used dozens of ready models

[glacial pollen]

Ahh, ye

https://www.youtube.com/watch?v=THxNdEsSGLA
Here's mine, first model I've made

but then, it was a prototype + my learning / test field model, effect of 2-3 months of work ( 40khz model, 10 mins )

shortly after, 48k training became available

[brave ermine]

what does do that 32k 40k thingies

[glacial pollen]

This is my current best ( and 2nd ever made oh irony ) model:
https://www.youtube.com/watch?v=2CW2Nyhtio8

Yea those are sample rates

32k = 32khz = 32000 hertz

there's 32, 40, 48 and soon hopefully 44.1khz ( Me, Noobies are working on new rvc pretrained models - those base models we all use to make our models ( the process's called finetuning btw )

[brave ermine]

i can feel the improve

[glacial pollen]

Yea, for a 10 min model, you can hardly find any that good sounding

[brave ermine]

so higher is better ot

or

[glacial pollen]

probs one of the best around on stock pretrains

[brave ermine]

what it depend

[glacial pollen]

It depends on what the sample rate of the audio you work with is

we have 44.1khz ( cd / standard music high quality ) and we have 48khz

[brave ermine]

what is sample rate 😭

[glacial pollen]

it's related to audio really

[brave ermine]

im clueless rn

[glacial pollen]

Means the amount of ( audio samples ) samples present in audio per second

higher the sample rate, higher the fidelity

but if we speak of 44.1 and 48, 90% of people can't tell a difference unless:

• they're working with audio
• are sensitive
• make music etc

[brave ermine]

how i can know a data is 44.1

or what rate

[glacial pollen]

basically, best way for newbies without some more advanced software, would be to use spek

[brave ermine]

its frequency of speak or smth

[glacial pollen]

It is this program:

You drop the audio in

and it shows you the:

[brave ermine]

my data was non stop talking cuz i cutted the blanks

[glacial pollen]

as you can see the vertical axis shows frequency range

measured in khz

( but given the nature of spectrograms, a form of representing the audio as you can see ) you wanna x2 the value

so, the audio file itself is 48khz since 24khz * 2 is 48

[brave ermine]

this looks to advanced

[glacial pollen]

then you can see where on avg it reaches

[brave ermine]

for me rn

[glacial pollen]

Yea, as mentioned before, take it within your pace

I'm mentioning it only because you asked

not that you should learn it rn or at all

[brave ermine]

so what u suggest howmany k i should pick

for default

[glacial pollen]

Should be matched with your audio's sr ( well, more or less. some deviance is allowed this or that way

[brave ermine]

what if not

idk my model matches or no

[glacial pollen]

then you can get sub-optimal performance or, plain bad performance

[brave ermine]

i need avarage and more performance i guess

for now i mean

[glacial pollen]

it's actually pretty simple tbf, you drop the audio and look for this range

in this case, it'd be around 24 or 20khz

so you can go 2 ways: 40k model or 48k model

no black magic here

[brave ermine]

fr

ill check next time

it looks simple

[glacial pollen]

yuh, it's free btw

Just one key thing to remember is the sample rate of spectrograms ( mel spectrograms to be exact )

( they can vary in colors and range, depending on what program you work with )

should be *2

and that's pretty much it

Here's the spek
https://www.spek.cc

[brave ermine]

how can i understand rainbow one

looks like random coloring book

[glacial pollen]

the rainbow one

is basically the same thing, just different color scheme I use

as I prefer it that way ( I can hear more details with varying " heat " in colors )

you can see the khz / hz axis on the right

and volume, in decibels ( db )

But that's iZotope RX program, advanced one for audio engineering / recovery / editing and so on

tbf, even audacity can preview spectrograms ( Audacity is probs the most know user-friendly ish program for audio editing and stuff ~ free as well )

[brave ermine]

i need understand these first

[glacial pollen]

In any case.. here's some terminologies you can check within your free time as long you feel like it:

• Sample rate in audio
• Mel spectrograms, what are they?
• Tensorboard and models' overtraining / undertraining ( overfitting / underfitting - same terms, can be used in-place )

yea, again, your own pace
I leave it all just in case if you wanted to recheck in future

[brave ermine]

you can sell lessons in online maybe

business

[glacial pollen]

well, I won't hide, my english has gotten pretty rusty ( spoken, that is

don't feel confident enough lol

[brave ermine]

u dont have to give lessons by live

u can sell videos i mean

[glacial pollen]

Oh yea, good point. But I don't quite like taking money for it

I was in your position, right, I was frustrated with lack of easy to take materials around

My goal in here, as long I feel like helping, is to make it as simple as possible yet without too overly-abstracted methodologies

and most importantly, not gate-keeping or pay-walling my knowledge, provided an individual is willing to give me and my teachings some time, just as I dedicate my time for them

[brave ermine]

anyway thanks for all and ur time

i gotta go

[glacial pollen]

Yea, np man. Best of luck and take care

[light pelican]

Hi, I’m getting this error when running TensorVENV.bat why ?

File "C:\Users\mathi\AppData\Local\Programs\Python\Python310\lib\site-packages\tensorboard_vendor\html5lib_trie_base.py", line 3, in <module>
from collections import Mapping
ImportError: cannot import name 'Mapping' from 'collections' (C:\Users\mathi\AppData\Local\Programs\Python\Python310\lib\collections_init_.py)
Keeping the command prompt open...
Appuyez sur une touche pour continuer...

[simple ore]

you should have just a regular tensorboard

not tensorboard_vendor, tbh

also since it is installed in global repo (under C:\users) you dont need a virtual environment activation (venv)

[dim jewel]

Hi guys, I have a question related to TensorBoard.
Does the lowest loss/g/total means the best quality?

[light pelican]

how ?

[simple ore]

pip uninstall tensorboard_vendor pip install tensorboard

[flint solar]

lowest loss/g/mel

[simple ore]

then you can just run it like tensorboard --logdir=X:\Applio-MRF\logs

[light pelican]

I’m still getting this error.
File "C:\Users\mathi\AppData\Local\Programs\Python\Python310\lib\site-packages\tensorboard_vendor\html5lib_trie_base.py", line 3, in <module>
from collections import Mapping
ImportError: cannot import name 'Mapping' from 'collections' (C:\Users\mathi\AppData\Local\Programs\Python\Python310\lib\collections_init_.py)

C:\Users\mathi>

[simple ore]

uninstall tensorboard_vendor

[glacial pollen]

#✨│ai-help message
#✨│ai-help message
and foremostly:
#✨│ai-help message + pics associated with the msg

Trusting solely g total loss isn't the most ideal if you aim for " best quality "

[dim jewel]

In your TensorBoard,
Would it make sense to fuse your 6k FM with 14.5k model to get the best of all?

[glacial pollen]

doesn't work like that sadly

sure, you'd reinforce traits good traits A but you'd also reinforce weak sides of each, traits B

You're kinda looking for a converged ish balance

Pretty much

[latent cypress]

can i use fp32 for model training on my rtx 4060 8gb? 👉 👈

[glacial pollen]

Uhh, you can try

but quite likely you won't be able to go crazy with batch size

on my 12 gig 3060 I can squeeze typically 12-13 bs

( I'll see what I can do further in my fork with memory optimizations

should re-add in-place and gradient checkpointing maybe... perhaps would help a tiny bit in terms of memory

[latent cypress]

i always use it on 8

[glacial pollen]

Well, then you should most likely be okay

[latent cypress]

is there a huge difference between fp16 and 32?

[dim jewel]

or you can start retraining at 6k for different result?
Is it an option, to find a balance, yes, but a balance favoring either FM or MEL, whatever you like more, instead of avarage?

[simple ore]

@latent cypress

so you can do batch 4 just fine

maybe 6

[glacial pollen]

Look man, I was having your mindset a year ago and I'll tell you right away
even if I used to say one can merge

it is no more a method I favor nor recommend

experience and tons of experiments have shown me it is def not the best way to go as well..

[dim jewel]

What would your advice then be?

[glacial pollen]

to focus on all rather than pick the models based on individual metrics

if one metric dominates in performance or is heavily undeperforming, something's bad going on

typically, after certain point, stuff should be somewhat 'balanced'

fm is your most important metric, this one takes a while to truly kick in

mel is pretty much related to mel spectrograms 'fidelity', you can think of it that way
but what influences it the most is how clean and uniform your dataset is

For example:

• if you mix sources in your dataset, mel could be having issues
• if the fidelity is initially bad, that on it's own also has it's implications as we're using pretrains that learned on, sure, not the cleanest but nevertheless, full ( frequency range wise ) audio

[latent cypress]

what about mri hifigan?

[glacial pollen]

not usable, needs pretrained models, same for refinegan ( not shown on the table

[latent cypress]

okay, thanks

[glacial pollen]

Tho, if you really wanna merge ckpts, I'd typically keep the ratio 70:30

where the base ( 70% of it ) is the FM one

you just " clear it up " with mel one

and one consideration: try to ( if possible ) have them both similar in terms of kl loss

else funky results can happen at certain inferences

[simple ore]

no difference

refinegan slightly more

[dim jewel]

#✨│ai-help message
It's the only thing I found about difference.

[simple ore]

fp16 without proper gradient clipping and nan parameter checks would ruin your model's training

sure it is fast, but is it fast to lose 10 hours of training while saving 30 seconds per epoch?

[dim jewel]

Thank you for your thoughts

[latent cypress]

thanks for explaining guys🙏

[flint solar]

@simple ore does applio use fp32 by default?

[glacial pollen]

afaik not

each time I had to switch from 16 to 32

[glacial pollen]

^

[simple ore]

no, you need to switch it in settings then restart the app

restart by killing it and starting again, not by opening a new browser tab

[latent cypress]

does your fork didn't have "sync graph" option?

[glacial pollen]

it's automatic

been that way for a while in applio generally

[latent cypress]

ah okay

[glacial pollen]

ye

ps

use repo dl not zip

[latent cypress]

i see your fork use fp32 by default

[glacial pollen]

Gotta push newsest rev1 zip

[glacial pollen]

yes

if you need fp16, just update the precision in settings

[latent cypress]

yes, im using the repo

[glacial pollen]

Alrighty

[dim jewel]

Does this message regarding what number in parameters of KL, MEL, FM you should aim for still good?
#✨│ai-help message

[glacial pollen]

Those were what I believed at a time ( half a year ago or more ) were right

but in reality, it's all per-case, individual

generally, lower the better

there's no ceiling or limits

[dim jewel]

😂 These time remarks about your past-self are killing me

[latent cypress]

i got this error when trying to warmup training

[glacial pollen]

Did you enable averaging

[latent cypress]

no

[glacial pollen]

then enable it and set it to 100

tho, just in case

100% sure you did not enable that box?

@latent cypress

[glacial pollen]

Alr, seems like I'll need to update the logic

in any case, for your specific situation where there's 600 steps per epoch, you can try 100

you'll get 6 loggings of average from a given epoch

[analog obsidian]

for training from scratch there is a huge difference and for that reason fp32 is always needed in that specific case
for finetuning (when we train using a pretrain) there's barely any audible difference at all/none

[glacial pollen]

as you're actually meant to use it

there's absolutely no penalty from that, in case you're worried ( and it does not affect the core training itself in any way

[latent cypress]

it worked now

[glacial pollen]

Ah ye, gonna patch it today then

seems like instead of disabling the mechanism ( if it's unchecked ) it's still activating and since averaging involves division

default value 0 cannot be divided naturally

[analog obsidian]

yea i also noticed that you cannot start the training if you leave that option unchecked

[glacial pollen]

Yea, overlooked it I suppose or rather, something unexceptional in args passing happened.
Gonna look into it in a bit

[latent cypress]

now my model is took 104 steps per epochs, i thought the training will be stopped when the warming up done (i set it to 2 epochs)

[glacial pollen]

wdym?

warmup epochs =/= training epochs

average steps =/= training steps

warmup should be 5-10% of your total chosen epochs for training
average steps ( n value ) should be anywhere from 10 to 35% of your total steps you get per epoch

pretty much

[latent cypress]

my bad, i can't understand english very well 😦

[glacial pollen]

Nah dw man

I'm here to help afteral

[brittle wing]

Where is that from?

[glacial pollen]

my fork

@latent cypress in any case, bug's fixed, gonna wait for 20th epoch to finish ( doing some test rn ) then ensure the fix work and uhh, ye, send you a patch

guess I forgot of one switch when moving my fork's base from rvc to applio, my bad ~

[flint solar]

it alr has 50

[glacial pollen]

if I get what you're trying to do, you want to keep only 50 there

and not use the mute files

[flint solar]

i wanted to add 50

[flint solar]

yes

[glacial pollen]

ye, then pretty much just keep in there:
[50],

[simple ore]

no need to do that

just add 50

it allows the small slices to be stacked

other buckets will be empty

[flint solar]

so i js delete the mute file?

or should i delete them from filelist.txt

[analog obsidian]

do this

dont delete the mute folder from logs

[simple ore]

delete them from the filelist

[glacial pollen]

not sure if you answered that the last time so gonna ask in case:
Doesn't it get re-added to the filelist again upon starting the training?

that behavior was present in rvc itself

as for applio ain't sure, haven't tested yet

[analog obsidian]

before removing the mute files i get collapses but after removing them from filelist.txt the collapses stops and the training goes well
this is how it looks without removing the mute files from filelist.txt

[glacial pollen]

oh, yea I know, we already do know that

I meant it more in a way where I haven't tested 'removing' the mutes in applio scenario

in stock rvc it'd readd the entries to the filelist

[analog obsidian]

ooo i see

[glacial pollen]

the mutes do play a role actually

if you have them gone and use silence in the training, well, weird shit happens

it's pretty much not the " best best " handler ish for silence

quite literally mute reference

And generally, given it's been mutes all this time, it is no more collapses

👀 yeyy

[analog obsidian]

[glacial pollen]

btw, fixed the bug

[analog obsidian]

[simple ore]

no, just when feature extraction ends

[glacial pollen]

a, alright then. That makes it even better

[flint solar]

so this method is to avoid collapses?

[glacial pollen]

Well, you can try but I feel like getting rid of mutes will have certain consequences when model encounters silent things / has own longer silences

you can actually be a good in-field tester 🤔 if you want, that is

[flint solar]

we'll see...

[latent cypress]

okay

[brittle wing]

Yeah, whenever I have time, my next project is to test it with your Fork, 8 batch as you recommended for 30+ mins.

Since you gave me those numbers for Warm up phase and average running loss to train with.

[glacial pollen]

actually, perhaps 15 to 25% of epoch's steps would be better

just divide your steps by 100 and * 15 or 25 and that's the n value to use

also, in any case, update's in a sec, doing last checks prior to pushing

[brittle wing]

Thank you

[dim jewel]

Are steps in TesorBoard supposed to be steps or epochs?

[glacial pollen]

steps

the logging itself is done 'per epoch', or should I say, per N steps that = 1 epoch

How about this @brittle wing

Does this look confusing to an average user?

or not really guys?

[dim jewel]

Okay, for some reason it shows epochs as steps for me, not sure why.

[glacial pollen]

wdym exactly

[flint solar]

check this

[brave matrix]

im trying to train a model and i need some help. is this a good spot for help?

[flint solar]

yes, what do u need help with

[brave matrix]

I keep getting this error when im trying to train. i click "process data" and it says
No such file or directory: '"C:\Users\BLANK\OneDrive\Desktop\RVC/preprocess.log'"

[flint solar]

why do u got a forward slash there

[brave matrix]

I dont know

[flint solar]

remove it

[brave matrix]

how

thats not the folder im inputting

the path is "C:\Users\BLANK\OneDrive\Desktop\RVC Train\Mangio-RVC-v23.7.0_INFER_TRAIN\Mangio-RVC-v23.7.0\MODELTEST[vocals]A.mp3"

[brittle wing]

Way better, so just divide your 1 or 2 N epoch by 100 then multiply 15% or 25%?

[brave matrix]

i enter that where it tells me to put my training folder

which this folder has only one file and its the voice

[glacial pollen]

Nah, it's actually simple

example: your epoch is 35 steps, your .pth file displays it

you do 35 / 100 * % ( and my recommended %s are 15 or 25

[flint solar]

can u screenshot

[brave matrix]

ss logs or the screen?

either way can i dm you? i dont want my name getting out (the "blank") or the voice im trying

[flint solar]

okay

[brittle wing]

Does lower or higher percent mean anything?

Lower the better or higher the better of percentage?

[glacial pollen]

not really, it quite literally means a percentage of your ( epoch's ) total steps

as for the size / number itself well

the concept is:
So, I think it's most common to balance these factors by reporting an averaged loss over N mini-batches, where N is large enough to smooth out the noise of individual batches but not so large that the model performance is not comparable between the first and last batches.

~ Some dude from whom I borrowed the idea

And for that, 15-25% range is typically okay

to not overburden the logging or cause any overly big mess

pretty much

[brittle wing]

If unsure which percent to pick, why not go for the medium of 20%?

[glacial pollen]

ofc, you can experiment, if you want more fine logging, you can naturally do that

heck, you can even log every single step

but imagine the amount of points to analyze, pointless

[glacial pollen]

that can be done too, it's just a general / broad recommendation I give

not enforced naturally

In fact, I encourage users to experiment to find what they like the best

afteral, it's just a diagnostic metric

and everyone might have own preferences or analytic capabilities 👀

[brittle wing]

Thanks for the update on Fork.

I appreciate it, thank you again!

[dim jewel]

it's my 295 epochs and it is exactly 295 step on the Horizontal Axis

[flint solar]

-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

[flint solar]

@brave matrix

[brave matrix]

much appreciated

[glacial pollen]

Np man

[flint solar]

https://docs.ai-hub.wtf/rvc/local/applio/

heres the guide for applio

[glacial pollen]

what batch size? what dataset size?

anything else I should know? ( what release you use of applio )

[dim jewel]

Maybe it's just my local colab thing. I didn't use applio.

[brave matrix]

Ill ping/dm if i need help. ill try to fend this on my own though. i dont wanna be a nuisance

[dim jewel]

Was using 16 batch size for 43 minutes of dataset.

[glacial pollen]

eeee, local colab?

are you using a notebook on pc locally?

if so, why

[dim jewel]

well, google colab, but local I mean community like

[glacial pollen]

a

well, something must be off with it then

what does it base on, applio? rvc?

if applio, which version ( same goes for rvc, the version

[flint solar]

sure ping if u need help

[dim jewel]

rvc v2

[glacial pollen]

well, weird to have it log every step regardless

the default behavior of rvc is to use 100 or was it 200 steps ( default )

and not 1 🤔

try some applio colabs maybe or something else, not sure what's currently recommended as I don't use colabs / cloud

but pretty sure there def are things to try out

[dim jewel]

Oh, are there any general rules between data size and what batch size to chose for it?

[glacial pollen]

broad topic really

for now, as I'm a lil too busy to give full explanations

tho ye, it's advanced and rather than ' going by rules ' , there are principles to get familiar with, which in result allows you to understand what is what and why

if you don't wanna really bother with it, you should choose 6 or 8 ( 8 is a better / more universal choice ) and you should mostly be fine

unless dealing with some rather big ( and diverse ) sets or super small sets

[dim jewel]

Well, in files it does show that 295 epoch is actually 15045 step.

[glacial pollen]

Yea that's what I mean, the behavior is not normal

Aside, it doesn't work in a way where user input or whatever is determining the actual real / file's steps

that's handled internally

What I meant by saying 100/200 is default, is for logging purposes

@analog obsidian @latent cypress
https://github.com/codename0og/codename-rvc-fork-3/releases/tag/v3.0.0-rev1

bugfix / update

[latent cypress]

many thanks

[glacial pollen]

: ) 🔥

[analog obsidian]

thanks

[glacial pollen]

Enjoy ~

[dim jewel]

Any thoughts on how to make FM fall?

[glacial pollen]

@graceful patio You too should know, in case

[glacial pollen]

uhhh, well.. It takes time

that's the main and most important metric

because generator learns from features ( FM )

Takes time, proper configuration and consistency in the dataset

[dim jewel]

Configuration and consistency?

[glacial pollen]

well yea

batch_size for instance

consistency as in, the dataset's consistency

same source, quality, sample_rate, balance in terms of what is in there ( i.e.: styles )
( yes there are people who mix different sample rate files in 1 set )

[latent cypress]

never take a look on fm graph, is the graph normal?, all i know the graph should be fall

[glacial pollen]

It should eventually start falling

if despite having arbitrary amount of epochs it doesn't

means something's wrong but unless that's the case, one should not worry and keep on training

The averaging includes all metrics from generator ( total g, fm, mel, kl ) and disc ( total disc loss ) so any form of training dynamics-balance problems can be easily spotted

[dim jewel]

How much is arbitrary?

[graceful patio]

got it thank you , just implemented it, training seems to be working rn btw (well the feature extraction , kinda slow , but it looks ppromising )

just got to see how it goes

[glacial pollen]

The feature extraction speed isn't related to any of my changes really

in any case, it's on Applio's side

However I'll mention, my fork on default runs in FP32

So have that in mind

if you need to change the precision, it's in settings ( Recommend to restart the fork after switching it )

[graceful patio]

i literally only have one file from the feature extraction and its moving really slow maybe i should restart well do

got it

(in dataset i mean )

which shouldnt take long

to process

even in feature extraction

but let me check brb

[lean kelp]

Hey guys

I just downloaded the software and it sounds shit how do i tune it to sound good? Im currently on a 4060

[latent cypress]

tbh, i didn't expect the fp32 took that long to train with my 24 minutes datasets, it's like i train my previous model with 1 hours + datasets on fp16

[glacial pollen]

FP32 is naturally slower and more intensive on memory

it is Full precision

[lean kelp]

Somebody help me please :(

[glacial pollen]

not half precision like fp16, but that's too advanced and I ain't gonna go deep on this topic at this moment

[lean kelp]

is there a good youtube video explaining how to sort it out?

[simple ore]

this is not a channel for realtime voice changer

[lean kelp]

oh my bad

[low shard]

-> #🔍│help-w-okada

[unborn turret]

chica

estas?

[tropic mesa]

Hey has anyone found the best pitch extraction for gravelly/deep voices? For me they always come out very mechanical sounding.

[austere stream]

Hello, this model has overtrained after 15k, right? I'm not the best at detecting when a model is overtrained but I am trying to get better at it.

[simple ore]

looks like a pretty small size of the dataset, right?

[austere stream]

Correct, about 19 or so minutes

[simple ore]

yeah, about 15k then

[austere stream]

Thank you!

[inner birch]

Cant use applio on amd gpu

it cause that error: https://pastebin.com/gypTk5E1

[simple ore]

what gpu?

[inner birch]

rx 6600

[simple ore]

you either 1) did not do hip sdk install correctly

2. running out of system memory

for 6600 you gotta use hip sdk 5.7 and proper libraries

[inner birch]

i think is my memory running out

cuz it was working yesterday

[simple ore]

if you got 16GB RAM it is not enough

👉 https://discord.com/channels/1159260121998827560/1159290161683767298

[dire meadow]

thx

[glacial pollen]

Reminder on why it's worth considering the use average running loss + normal vs just normal:

you can see the disproportions in feedback
( training early on but that's juts a demontration ).

[floral lava]

i did this and i get this error: Traceback (most recent call last):
File "C:\Users\biric\Downloads\RVC1006Nvidia\RVC1006Nvidia\infer\lib\train\process_ckpt.py", line 66, in extract_small_model
ckpt = torch.load(path, map_location="cpu")
File "C:\Users\biric\Downloads\RVC1006Nvidia\RVC1006Nvidia\runtime\lib\site-packages\torch\serialization.py", line 791, in load
with _open_file_like(f, 'rb') as opened_file:
File "C:\Users\biric\Downloads\RVC1006Nvidia\RVC1006Nvidia\runtime\lib\site-packages\torch\serialization.py", line 271, in _open_file_like
return _open_file(name_or_buffer, mode)
File "C:\Users\biric\Downloads\RVC1006Nvidia\RVC1006Nvidia\runtime\lib\site-packages\torch\serialization.py", line 252, in init
super().init(open(name, mode))
OSError: [Errno 22] Invalid argument: '"C:\Users\biric\Downloads\RVC1006Nvidia\RVC1006Nvidia\logs\aktaneren\sess\G_630.pth"'

can anyone help me about this?

[azure marsh]

Hey, Feldas! 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

[floral lava]

i couldn't find the second .pth in weights folder because i didn't select the "To save a final small model in the 'weights' folder at each save point:" and i stopped the training middle of it. So i dont have the second pth Therefore i did the ckpt processing from the bottom line. now i have this error

[glacial pollen]

you won't get a 2nd epoch from G file

1 G/D pair " holds " only 1 epoch

If you stopped mid-way during 2nd epoch's generation, it is gone

[knotty moth]

consistency over the quantity, for example in this case, the former part has lower cutoff and might be from another source, so you'd better yeet it out

[glacial pollen]

that data does not exist as G ( and so, D file too ) did not get updated / overwritten

As for why you get the error in the first place, I'd need to see your settings / params you set. Send the screenshot of that section in the ui + provide training details

[floral lava]

actually i had like 50 epoch

[glacial pollen]

well, in that case you can't get it either

a

wait

you're doing it that way

well, show the screenshot of the ui

[floral lava]

it's turkish but you can understand it for sure

[glacial pollen]

what do these box say

upper and lower one

Been a while I've seen that section in stock rvc

[floral lava]

it easier to do like this i think

[glacial pollen]

uh, tho man

do not use " "

provide the path as it is

without quotes

Also, pick the save name ( mode info box leave empty )

[floral lava]

bro i tried everything

[glacial pollen]

Do it and show the log

[floral lava]

and now it solved

[glacial pollen]

a

[floral lava]

omg

[glacial pollen]

well, then there's that

[knotty moth]

I'd suggest these options, the first one so that it will save and overwrite G_2333333.pth and D_2333333.pth to save disk space

[floral lava]

i tried everything xdd

[glacial pollen]

enjoy I guess

[floral lava]

thank you

[glacial pollen]

np

[glacial pollen]

^
unless you do pretrains, which I doubt you do

[hallow thistle]

If you've solved everything by yourself, then you don't really need me to do that.

[glacial pollen]

save every small weight + only latest G/D weights

[hallow thistle]

https://cdn.discordapp.com/emojis/1230288925600514048.webp?size=48

[floral lava]

yeah this was my first time i wil do these settings ty

[glacial pollen]

Good luck~

[floral lava]

ty you guys too

[glacial pollen]

First and foremost, do you have a gpu?

star

@restive mural In case you don't know to check it, let me know

Well no offense, but it'd be appreciated if you stayed as an active participant of the convo if you're already asking for help or a guide

[glacial pollen]

well then too bad. Kinda tired of wasting my time on such cases

¯_(ツ)_/¯

[hallow thistle]

Bro ain't answering.

https://tenor.com/view/discord-users-esmbot-gif-26228284

[brittle wing]

This gif had me dying

[simple ore]

I'm in this picture and I dont like it.

[brittle wing]

does anyone know what configuration I should use on wokada? I have a ryzen 5600g, rx 6600xt 8gb, 32gb ram

[hallow thistle]

Cray.

[hallow thistle]

Go #🔍│help-w-okada for W-Okada. This channel is for RVC.

[winter garden]

https://github.com/emirthab/RVC-GUI

Anyone try this? trying to find an inference program for mac

[knotty moth]

nope

[knotty moth]

-gui

[azure marsh]

https://cdn.discordapp.com/attachments/1122285248844144733/1203460490475343953/caption.gif?ex=65d12cec&is=65beb7ec&hm=bd2fb8d010006dd7c6e3c1c67d3ae846fd1478e1a3124c544c31b43086fe54aa&

[low shard]

Super outdated

[low shard]

For Inference (use models) Mac, You can:

• Locally (runs on your pc so the speed depends on that, you will have to set it up with the guides, probably won't be able to train, make models):
  • Applio: A fork of RVC with some extra features like Applio TTS, kinda faster and simpler but same quality tho
  • Mainline: The original RVC
• Cloud (remote good pc, easier and faster than ur PC but it's limited):
  • Weights.gg: Partnered with AI Hub, lets u do them easily but u may be in a queue
  • Ilaria RVC Zero: fastest and simplest that you can get for free
  • Applio Colab: max 4 hours, not granted, of GPU

[modern anchor]

I got the path file and CKPT but no index, what do i do?

[serene crown]

my ai model can't do high notes (specifically above G5) without cracking like crazy

how do i solve this issue? does this have anything to do with the pretrain model or something?

[hallow thistle]

-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

[low shard]

are you training it yourself or found one online?

you need only the added index and .pth

[low shard]

It could be related to your dataset not having high notes

[hallow thistle]

A .ckpt file? No way, that file name extension is used with some Stable Diffusion model.

[low shard]

^

[hallow thistle]

The ckpt file doesn't seem to be used anywhere in RVC and W-Okada, unless you're refering to pth which is the primarily file extension for RVC voice model.

[low shard]

ckpt is a model checkpoint file

safetensors is a safer version basically, which is usally the standard now

RVC uses Pytorch models format, even if it would be better to switch to safetensors tbh

[flint solar]

-gui

[azure marsh]

https://cdn.discordapp.com/attachments/1122285248844144733/1203460490475343953/caption.gif?ex=65d12cec&is=65beb7ec&hm=bd2fb8d010006dd7c6e3c1c67d3ae846fd1478e1a3124c544c31b43086fe54aa&

[flint solar]

-svc

[azure marsh]

https://cdn.discordapp.com/attachments/1159822429137412177/1194008385113301003/svc.gif?ex=65c13ef6&is=65aec9f6&hm=ce1f57bc316793f641c80f9b359ba9db7d1312af6918b1aed9c8468c74035bac&

Malik by Weights GOT BANANA LOL LOL LOL

HEY YOU @flint solar YOU FUCKING GOT BANAN LMFAOOOOOOOOO
HEY YOU @flint solar YOU FUCKING GOT BANAN LMFAOOOOOOOOO
HEY YOU @flint solar YOU FUCKING GOT BANAN LMFAOOOOOOOOO

BRO GOT BANAN'D ONCE XDDDDDD

[hallow thistle]

Ckpt also used for certain LoRA models.

[knotty moth]

more likely a GPT sovits model, yet many ppl still trying to use it for voice changer

https://tenor.com/view/spongebob-punch-boxing-boxing-glove-punching-myself-gif-21654165

[simple ore]

singing pretrain like klm 4.3, pitching down the source a bit

[modern anchor]

Found it on this server's voice models channel

[low shard]

Send the link

[knotty moth]

you should see if it is tagged GPT sovits

[latent kettle]

how do i copy my link in weights 😓

[viscid moss]

Why do you have a gif of me?

[viscid moss]

u mean ur profile link?

[eager shuttle]

The problem I have is the prior setup with pytorch and all of that, why can't it just all be done already so it's easier

[glacial pollen]

There is no problem. You run 1 .bat for installation, one .bat for running

[frank blade]

Friends, I need help, a group of people and I are putting together a dataset to create a pretrain with Brazilian voices, but how do you train a pretrain for the RVC?

[glacial pollen]

it's as simple

[modern anchor]

Oh yeah it is nevermind thank you

[latent kettle]

Yes. Byw thank you. I got my answer 😊

[latent kettle]

Getting runtime error in appio on kaggle

What to do now. Please help

This is my data transfer out limit

[carmine siren]

is anyone here i am also getting same error

[latent kettle]

Do you got any solution ? If yes please tell me

[carmine siren]

not yet , but trying to get in touch with any tech support who can help me out with this mess

[low shard]

@glad zealot what was the fix you did for fixing this in mainline?

it's happening to other notebooks too rather than just mainline

prob should tell vidal to use hina's fix

@carmine siren @latent kettle do yall get this while trying to train?

[carmine siren]

yes, i don't know what going wrong its auomatically intrupted the trainning process with 'run time error'

[low shard]

It's an issue with the code of the notebook, the mainline notebook fixed it recently

[latent kettle]

I completed 200 epochs amd then decided to train it more on 300 epochs.

[low shard]

but i'm guessing that vidal's applio's notebook is affected by the issue

[latent kettle]

Then on 260 epochs this error occurred

[crystal gull]

"It's an issue with the code of the notebook, the mainline notebook fixed it recently"

[dim jewel]

Guys, let's say I save every 5 epochs. Is there a way for me to retrieve let's say 3rd epoch?

[crystal gull]

um no?

"if i count 1 5 10 15 20 did I say 3?"

[glacial pollen]

Nah, simply train every single one

and if memory is your constraint, do like I did 1 year ago

supervise the training and actively delete crappy ckpts to save the memory / move all epochs periodically to some other drive

[crystal gull]

memory? you mean storage?

[glacial pollen]

I use abstractions

[crystal gull]

abstractions??

[dim jewel]

Good idea

[glacial pollen]

You'd delete it once you're full, your storage obviously, it doesn't go permanent

[crystal gull]

jesus you're so wrong

but ok

[glacial pollen]

💀

[crystal gull]

"if i download a file and then delete it, it technically isn't storage anymore, it was memory :D" follows that same logic

[glacial pollen]

Memory and storage is used nowadays interchangeably, you're making it difficult by having an issue with simplifications

[crystal gull]

No the fuck it isn't

[glacial pollen]

By definition naturally memory refers to temporary held data

[crystal gull]

You're actually insane

[glacial pollen]

Use google

Go ahead

[crystal gull]

And you're supposedly an engineer

[glacial pollen]

Also learn some respect towards people with higher rank in here

[crystal gull]

Please learn to be smarter

[glacial pollen]

Because you're having an ass pain due to some terminology

[crystal gull]

Your rank does not correlate to respect

And regardless of respect, if you're wrong you're wrong.

[glacial pollen]

You're a member on this server and you willingly accepted to follow the rules

It is in your duty to uphold, essentially what you agreed to or either leave

[crystal gull]

"If your memory is your constraint supervise the training and actively delete ckpts to save memory (WHICH ARE FILES NOT CURRENTLY EVEN LOADED IN MEMORY) / move all epochs periodically to some other drive" Great so you're taking about storage, and you mean storage.

[crystal gull]

Hmm show me the rule that says I must respect and agree with every thing someone with a higher rank than me says

Regardless of them being wrong

[glacial pollen]

I said I used an abstraction / mental leap

And you're practically having a pain in the ass over terminology

[crystal gull]

You literally cannot

[glacial pollen]

This ain't your essay or school work

[crystal gull]

Memory is not storage

In the same sense of permanent storage

Such as a drive

You literally said "If your MEMORY is your constraint" REFERENCING, RANDOM ACCESS MEMORY. "Delete ckpts to save memory" which ARENT LOADED IN MEMORY

[glacial pollen]

Bro, you're literally making your position worse now. You're disregarding my statement on using an abstraction for the sake of simplifying the explanation

[crystal gull]

Your abstraction doesn't work when you referenced a pc component such as memory being a constraint..

You know you're wrong and you keep arguing because you can't admit you made a mistake? I'm out.

[glacial pollen]

Nah I am not wrong. I literally said I used an abstraction / thought shortcut, if you can't accept that, this is only your own issue

[crystal gull]

"If you have a memory constraint, delete stuff off your ssd / hdd" 🤯 🤯 🤯

[glacial pollen]

Alr, here's the thing. a fucking ssd, is a flash memory 🙂

you're the fucking dumbass in here 💀

end of the debate.

[crystal gull]

BAHAHAHAH

LMAO

get out

leave

[glacial pollen]

This is your last warning

people of your sort with 0 respect and toxicity are flushed instantly

[crystal gull]

no one is going to assume you mean flash memory when you say memory ❤️

[glacial pollen]

Memory can be of any sorts my dude

You can clearly read the context

[crystal gull]

you lost your rocker because I called you wrong, and you went on a rant trying to prove you're right.

[glacial pollen]

or are you a roblox brainrot

[crystal gull]

you also don't know he has an ssd

🙂

[glacial pollen]

Nah, I have issues with people like you who can't read fucking context

[crystal gull]

You lost when you said "Memory and storage is used nowadays interchangeably"

[glacial pollen]

Acting like they ate up all the knowledge despite 0 shit contribution nor understanding on what they even work with

[crystal gull]

If I walk up to someone and say im having issues with my memory, their first thought is about random access memory

[crude flame]

im wondering, why would we be talking about ram?

[crystal gull]

Not flash memory

[glacial pollen]

Honestly, if you wanna compete or whatever for the sake of your ego, go ahead and learn machine learning basics

because all of what you've been talking about is terminology so far

[crystal gull]

hold on let me be you for a second

[crude flame]

💀

[glacial pollen]

Since practically, right now? at given moment? your knowledge is of 0 worth

for me

Actually, there's no point producing myself on brainrot kids. Bro watched an anime 'bout Russian girl and a school genius and thinks he's the smartest, 1 step ahead

[crystal gull]

the fuck

[glacial pollen]

Quit

[crystal gull]

glad we're bringing personal insults into this, shows you can't argue.

[glacial pollen]

Getting to your level of cognitive abilities

[crystal gull]

and IM the one not showing respect

[glacial pollen]

yea, I have no intents of respecting a person disrespecting me

[crystal gull]

LOL

[glacial pollen]

simple as that

it's either mutual or there's none

[crystal gull]

having a differing view of something which turns out to be factual, is disrespecting you.

we in a dictatorship with this one

[glacial pollen]

Nah, this is not factual, you're just an extremist when it comes to vocabulary or terminology

You can clearly see the tendency of impractical / mistaken interchangeable usage of words " storage " and " memory " especially when it comes to flash memory or phone's storage

I used abstractions of " temporary " in a sense the epochs get deleted shortly after, but you make it seem like the end of the world

starting a drama

You're in the wrong here

[crystal gull]

great, so lets say you did mean flash memory when you said memory. ❤️ you assumed he had an ssd

still haven't answered that question

[glacial pollen]

Ofc, what is so wrong to assume having an SSD in 2025~

Don't make it seem like " gosh, you assume my gender " 💀

that's a fucking clown fuckettry

[crystal gull]

what is so wrong to assume when someone says memory they are talking about random access memory?

who the hells first thought is an ssd when someone says memory

[glacial pollen]

Memory dosen't have to be ram, it can as well refer to brain's memory

it is about context

[crystal gull]

LMFAO THE CONTEXT IS HOME COMPUTERS

[glacial pollen]

You should understand it well being a slavic language user

As far as I know, In russian there is no separate direct term for mass storage / pc storage and memory in a brain / biological sense

[crystal gull]

My primary language is English, thanks though.

[glacial pollen]

Well then I shall let you know In Polish that's the way it is

so especially in my language, the terms are used interchangeably

yet, in this particular context, being the convo / my help for the dude, it was a thought shortcut

and you make a world-end's drama outta it

[crystal gull]

Welp sorry pal but you're speaking english. When you say "memory" the first thought is random access memory.

[glacial pollen]

I shall let you know that the behavior you're presenting with being overly fixated on terms is pretty unhealthy

Once again, this is fucking discord full of newbies who rarely even realize they have gpu, if I decide to use the words or terminology I deem right, that's what I do

[crystal gull]

I shall let you know that you were wrong whilst writing the sentence, and then backpedaled the only way you could. To protect yourself, why? dunno.

[glacial pollen]

and it is not up to you to correct me AND expect me to follow your rules as you plead

[crystal gull]

Holy shit wow you proved my point here

[glacial pollen]

There is no point to be proven, you just like attention and being the smartest guy

[crystal gull]

If they can barely realize they own a gpu, then how the hell are they going to assume you mean SSDS when you say MEMORY

[glacial pollen]

if you really wanna be smart? shut the fuck up, learn and prove your worth in here

[glacial pollen]

And following that, aim for engineer

before that, your worth is of 0 to me as you've displayed no respect towards my person

[crystal gull]

🗿 I don't care about cosmetic discord roles

[glacial pollen]

simple as that.

[crystal gull]

You're flexing a discord rank

Bah

Google memory and show me what appears

[glacial pollen]

Nah, cause while you're being a drama queen here

[crystal gull]

Google computer memory

👍

[glacial pollen]

I work on rvc v3 improvements

cheers pal

[crystal gull]

It's okay, we know you can't admit you were wrong.

[glacial pollen]

Go and search through this discord, I can quite confidently admit when I am wrong

we know you're a factual person already

[crystal gull]

thanks

we know i speak facts

glad to hear

what is new

[glacial pollen]

See how simple it is to have you satisfied? 🙂

wish your parents did a lil better job

honestly.

[crystal gull]

has nothing to do with satisfaction

you flexed a discord role

and then told me i have to respect you because of a DISCORD ROLE

sounds like you need satisfaction

[glacial pollen]

I ain't in a need of satisfaction, in fact, I'd be more grateful if shitheads like you weren't a part of the community

quite openly

[crystal gull]

just saying

[glacial pollen]

just so you know

[glacial pollen]

^

too bad on ya my man, time to explore the internet or xda a bit better

[crystal gull]

just saying

😭 Yes they are used interchangeably, but not when you're talking about fuckin home pcs and people who are unknowledgeable

[glacial pollen]

See, that's the problem with you, within majority of communities, the terms are used interchangeably, we are not talkin about the proper terminology

[crystal gull]

Literally no one here will say "yup you're talking about ssds" when you say MEMORY

ITS BASED OFF THE TOPIC OF THE CONVERSATION?

LOL

[glacial pollen]

Context

but people of your cognitive abilities seem to not be able to comprehend it

too many neurons fired in the same time 💀

[crude flame]

do you happen to be on the UV wave length?

[glacial pollen]

lmao

[crystal gull]

yes im all seeing

[glacial pollen]

honestly mandi, you remind me of that one ex that wasn't capable of getting any slightest bit of context or conversation happening at a given moment

any form of cognitive reasoning

[crystal gull]

you remind me of someone who can't have an argument or conversation without bringing up personal and non personal insults.

[glacial pollen]

You disrespecting me was personal

Because if this is your entry attitude, first display

it is personal

[crystal gull]

show me one personal insult

ive said

[glacial pollen]

I quite literally explained you what I deem as personal

[crystal gull]

you quite literally explained that the majority of the people in here barely realize they have a gpu, but for some reason want to assume that people are going to think of ssds and flash memory when you say memory. and not think of random access memory sticks

You disproved your own statement with something prior you said

[crystal gull]

like I said show me one personal insult ive said

[glacial pollen]

Because up to this point, all you've really displayed is:

1. You can't read context
2. You're a type of drama-queen you'd observe when it comes to " phrasing, terminology and grammar ", close to grammar-nazi levels
3. You're uncapable of connecting the topic of conversation with flash memory ( assumption from my side or not )
4. You do not respect people and have no issues breaking the rules

[simple ore]

some people just dont know when to stop 🙂

[crystal gull]

when have i talked bad about or insulted, disrespected. anything you like, or do, or have. outside of this conversation.

because i can call a few times you did 😂

[red kayak]

go to dms please @crystal gull @glacial pollen

[glacial pollen]

Nah litsa, this is already over the top, 5 or 6 fucking douchebags within 4 or 5 days tripping me upside down over some bs nonsensical crap.
I ain't gonna go around and tolerate the absence of moderation

Honestly, if you intend on muting or banning me, go ahead, I do what I do because I want to and I ain't obligated to volunteer if 0 respect is kept towards my person

not when people of that sort are around

[crystal gull]

sounds like a trend..

anyways, does anyone use anything other than applio? if so, what would you rate it. would you suggest it?

[analog obsidian]

original rvc got outdated after applio devs started to do code changes

you can use it if you want but the original rvc is no longer receiving updates

been 1 year since they've updated it

[red kayak]

i see who initiated the argument

@crystal gull you clearly started all this and poked at him, you shouldnt have done that. he wasnt even being hostile towards you anyway, whats your issue man?

[glacial pollen]

Like I get it all, but when a, pardon, dumbass calls me out like that for absolutely no reason

I trip my shit

This is the 4th or 5th time this week as I said

I am constantly short on sleep due to the work on rvc and quite nervous lately

[red kayak]

alright relax, no need for that anymore

i see whats going on

[glacial pollen]

I am just being honest and open, that's all.

[crystal gull]

Doesn't give someone an excuse to carry on and provide multiple personal insults towards me based outside of the conversation now does it? My end was simply an argument over the terminology used and how he could have better explained it. ❤️

And perhaps codename should work on not getting so angry over an argument no?

[glacial pollen]

Nevertheless, I am out of the convo.
Ain't taking more disrespect or provocations from kids with attention deficit.
Do as you please if you may. Need a break

[crystal gull]

And also Codename talks so often about rules and following them, but after told to stop. Continued the conversation in the same channel whilst I ignored :l

Just saying

[red kayak]

sure you bring up a good point however, you were still being so condescending and hostile towards him, perhaps maybe you should learn to mind your manners

[glacial pollen]

ps. An adult and comprehensive person is able to tell the no-cross line and effectively estimate one's limits

because people have different thresholds

it's a basic skill to learn and have

[red kayak]

you clearly pushed him, i dont think you can argue against that

[crystal gull]

And telling people that they HAVE to "respect" him purely because of his role? And that we essentially have to agree and cannot argue?
Him flexing his discord rank is a bit much lol

[glacial pollen]

No, it only mentioned respect towards the higher ups as the rules mention it

it wasn't " respect exclusively towards me ".

[red kayak]

you are missing the point here...

[crystal gull]

Never once have I insulted to this level

[red kayak]

seems like you want to overlook the fact that you initiated the fight

[glacial pollen]

this is a dm, if you push me forward, don't expect nice treatment

if your goal is to poke a frustrated person, that's what you get, being fully aware

And it'd be nice to not cut the context.

end of the debate.

[crystal gull]

oh well he called me "fucking stupid" at the top aswell

[low shard]

Idk what yall on but poor the people who come here for help and see the latest argument 😭

[analog obsidian]

😭

[red kayak]

you started the fight, does that make sense to you?

[glacial pollen]

yea, as I told you Mandi, if you have no respect towards me, I am not obligated to show any towards you

[quaint agate]

💀

[low shard]

from help-rvc to help-arguing it took a single word man 😭

[glacial pollen]

Don't expect kissing your cheek or pampering you if you're an ass towards me
not everyone agrees, and I get that, but that's how I work

[crystal gull]

seems like you want to overlook the fact that regardless of who initiated an argument, both parties willingly sent messages back and forth and participated in an argument...

end of story you literally told us to stop arguing, I did. Codename continued and made it a bigger deal in the public chat when it could've been over with.

[low shard]

U looking for help?

[crude flame]

This new Ai Hub By Weights Drama Situation Is Insane

[low shard]

😭

[red kayak]

alright relax just go now

[glacial pollen]

Alr, you handle that.

[red kayak]

mhm

[crystal gull]

Anyway like you told us, I'm done and I'm out. It's not my issue, I didn't wish to carry the conversation on here any longer after being told off to go to dms.

[low shard]

the drama, it's massive

https://tenor.com/view/low-taper-fade-ninja-massive-the-meme-is-still-massive-yeah-no-massive-gif-9995754776955392143

[dim jewel]

You do always have good story to tell about your year-ago self, huh? 😄

[glacial pollen]

habit, do not mind that

[modern surge]

-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

[glad zealot]

I think it was the canvas render? It needs to use argb instead of rgb in matproto lib or whatever that was called, then change the canvas size to dynamic instead of fixed for some reason

[low shard]

You should tell that to vidal too

[glad zealot]

The error on mainline only happens when training the model btw, doing everything else (process, feature extract, index) is fine

[grand pecan]

Guys is this overtrain? in g total it has been going kind of flat in g total for 2 hours, I used a 1 hour 30 minutes dataset
https://ibb.co/0YfxqJB (d total)
https://ibb.co/MDgYswZ ( g fm)
https://ibb.co/pL73YMB (g ki)
https://ibb.co/n8V21dF (g mel)
https://ibb.co/2hbh6vq (g total)

[analog obsidian]

yea looks like it stopped improving

[grand pecan]

alr, it stopped around here right? https://ibb.co/MZn2wGX

[analog obsidian]

yup, after that the model seems to be close to converge

not really overtrained

[grand pecan]

oh should I let it continue then?, I was aiming at 1000 epoch, or should I use it already?

[analog obsidian]

you can wait until it fully converges (graph becoming extremely flat) or stop the training now and hear the epochs and see if you like them

tho the improvement is very minimal

not audible

[grand pecan]

Ok thank you so much 👌

[analog obsidian]

no problem, in cases where my graph looks like yours i'd rather stop the training since is not really worth to wait it for fully converge

[hallow thistle]

I don't know if you were trolling around or like you just didn't know how to talk to people. But like please understand that not everyone here would understand you as you do with others. On Discord, the message you sent, the message anyone sent, can be misinterpreted sometimes, you know, it's the internet right? If you come here to get help on AI, please talk like what a normal person would do. And sometimes, some people aren't really in their mood to talk, so like please give them some time to relax. Ok?

Nah, I'm way too late to respond all of this, man. If I had time to talk to him, things might have gotten worse because of the way I typed. I'm done doing dramama since the previous AI Hub.

Sometimes, y'all gotta take a break from your screen and AI for a while to focus on your health. No offense, this is my recommendation. I've done this one time between after the beef with a kid and before the previous AI Hub server exploded.

[brittle wing]

can i get help from learning how to get a ai voice changer

[hallow thistle]

RVC the audio conversion or W-Okada the realtime voice conversion?

[brittle wing]

rvc

[hallow thistle]

-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

[brittle wing]

can we call so i can share screen

[hallow thistle]

Would you like to run RVC on your PC locally or on a cloud service?

[brittle wing]

locally

idk

[hallow thistle]

[brittle wing]

add meh

[hallow thistle]

What GPU does your PC have? Just wanna make sure it's newer than GTX 10xx series to be able to run RVC.

[brittle wing]

ima get in a call and share screen

1650 gtx

oh

is my gpu good?

[hallow thistle]

This is the minimum GPU you can use for RVC and any AI program related.

[brittle wing]

k

help meh then

[hallow thistle]

Applio is a recently developed fork program of RVC GUI. Download this one.

[brittle wing]

k

help me

idk how to do ts

ello?

[hallow thistle]

I don't usually join in a DM conversion as to avoid spam, but anyways.

[brittle wing]

ok

can we call so i can get help?

or nah

[hallow thistle]

First of all, download the complied one of Applio. https://github.com/IAHispano/Applio/releases

[brittle wing]

k

wut i do next\

nvm i see

[brittle wing]

wut do i do now

[hallow thistle]

Is it finished downloading Applio?

[brittle wing]

yessir

[hallow thistle]

Extract the zip to your directory like C:\Applio or D:\Applio.

[brittle wing]

k

[hallow thistle]

https://img.huffingtonpost.com/asset/62e29d62220000e9944fa3df.jpeg

[brittle wing]

DONE

[hallow thistle]

Double click on run-applio.bat.

[brittle wing]

did it

but do nothing?

[hallow thistle]

What? Did it show up any CMD? If not, then try run it with CMD and screenshot it.

[brittle wing]

For real

[hallow thistle]

https://tenor.com/view/kanye-west-stare-serious-gif-15710427

[brittle wing]

But you did the opposite regardless, this is just excuses

nvm it work

[hallow thistle]

It looks like you didn't extract Applio zip into C:\ or D:\ but instead a Downloads folder, which can cause Applio to be unable to run. It's best to extract Applio to somewhere like this.

[brittle wing]

Link?

its telling me to put voice model

wut i do now

[hallow thistle]

Now this part of putting the voice model, this can be tricky as since all pth and index files are stored in \Applio\logs unlike any other RVC GUI.

[brittle wing]

k

[hallow thistle]

Well, I don't know which person started it first. But if they wrong, they wrong.

[brittle wing]

do i just download a voice model

and thats all?

[hallow thistle]

Find a voice model at #1175430844685484042 or search up on Weights.gg.

[brittle wing]

there

i got one

what now

[hallow thistle]

Extract pth and index files from voice model zips into \Applio\logs, click refresh on the GUI part and the list for voice model will show up.

[crystal gull]

Clearly you're blind, also quit bringing up old things already dealt with kthx ❤️

[brittle wing]

Bro my eyesight is above 100%

What are you talking about

You're being disrespectful

Again

[crystal gull]

Stop speaking to me or ill ping the mod for continuing drama 👋

[brittle wing]

You're the one who enabled it.

[crystal gull]

Stop speaking to me or ill ping the mod for continuing drama 👋

[brittle wing]

i give up tbh to hard

And you keep threatening and calling me blind
Got it.

Yebye-

[brittle wing]

Link to the fork?

jk

[hallow thistle]

This was like 96% of finishing up Applio, and you gave up on it? Alright, might try again later next day, so good luck.

[brittle wing]

but can we get in a call pls

it more easier for me to share screen

then text

[hallow thistle]

That bro's expecting anyone to give him attention by fighting him.

Wouldn't be surprised if he's just a kiddo on Discord.

[brittle wing]

who me?

[hallow thistle]

The other guy.

[brittle wing]

o

[brittle wing]

Me or...?

u

[hallow thistle]

The other guy you talked to.

[brittle wing]

Idk if they're a guy or girl but nvm it's just arguments over nothing

[hallow thistle]

[brittle wing]

Mindless meaningless stuff

I think

There are far more important things in life, childish stuff

[hallow thistle]

An unserious argument an X/Twitter user would do.

[brittle wing]

I don't use Twitter

[quick jungle]

You mean 𝕏??

[hallow thistle]

https://tenor.com/view/ai-hub-aihub-ai-hub-discord-aihub-discord-aihub-mod-gif-18349908146504765143

[crystal gull]

@red kayak

[brittle wing]

I still call it Twitter...

[analog obsidian]

stop bringing this back

[hallow thistle]

Nah bro, I'm not trying to aggravate you at all. I'm just here to give my basic knowledge of installing AI, not a doing a drama.

[brittle wing]

How can I make ai vocals sound more realistic

[analog obsidian]

train a dataset of 30 minutes and above

[hallow thistle]

Why would you think I brought up all of this? I was just just too late to the chat and knew nothing about it.

[brittle wing]

I know that, I mean after inferencing when mixing the instrumental with the ai vocals.

[analog obsidian]

oh for covers, proper eq and compressors should do the trick

and for the inference audio it should be clean, no reverb and no harmonies

[brittle wing]

How do I regulate a preset

[brittle wing]

I kw that too

[analog obsidian]

i would recommend watching a youtube tutorial on how to mix vocals

[brittle wing]

How to set it up correctly I mean

[hallow thistle]

So like please understand that even if my typing can be hard to understand for certain people, I would never do such a drama at all.

[analog obsidian]

ohh no idea

[brittle wing]

How much reverb, how much eq and stuff

[analog obsidian]

that varies depending on the song

[hallow thistle]

[analog obsidian]

this is out of a rvc related topic but more of a fl studio/DAW topic

[brittle wing]

Is there a channel

[analog obsidian]

for vocal mixing? not here, sorry

[hallow thistle]

If it's about training a model, go to #📑│making-models.

[brittle wing]

Uh it's okay.

I already know how to train & resume.

[brittle wing]

i need help weights

i need help

[lost sail]

-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

[latent kettle]

So using mainline is good for now ?

[modest hare]

help

[azure marsh]

✍ Suggestions

• Search for it in AI HUB Docs or Applio Docs. You will probably find your answer there 📚
• Ask for help in #🔍│help-w-okada if it's related to real time voice changing but make sure to read #1297207135469305866 first
• Ask for help in #✨│ai-help for general help, but use the command !howtoask first to learn how to structure your question properly and increase your chances of getting a reply
• Last but not least, ask for help in #🔍│help-ai-art if it's related to AI images.

[analog vault]

I’ve been messing around with regular version w-okada for a bit, after a lot of researching and testing i got it to sound good, but not good to make me want to use it in friends discords. i’ve seen youtubers use it that had much better results, what am i doing wrong? or am i not getting high quality voice models?

is seed-vc real-time changer better quality and worth downloading? i can barely find any info on it and only one vid in english demonstrating it. or is deiteries fork worth my time trying out?

also i have 4080 so hardware not a issue

[potent tendon]

Where is the RVC generating channel

the one where we make ai music

[hallow thistle]

This is #✨│ai-help which is a channel for RVC. For W-Okada, go to #🔍│help-w-okada.

[hallow thistle]

RVC generating channel? A channel in this server or a YouTube channel?

[potent tendon]

This server

[hallow thistle]

There doesn't seem to be a RVC generating channel in this server. Unless you're referring to #🤖│bots, which is where you use bots to make AI cover.

[potent tendon]

There used to be one where

I can use a command

choose voice model and put a YouTube url

and it makes it

[low shard]

#🤖│bots

[hallow thistle]

Yes, the #🤖│bots is where you use bot commands.

[low shard]

You can use /create

To make an AI cover here

[potent tendon]

oh I forgot

[low shard]

With the Weights bot

[potent tendon]

I haven’t made one in a while

[hallow thistle]

[low shard]

Yw

[hallow thistle]

A channel for "RVC generating channel" could refer to a now-deleted channel in AI Hub.

[low shard]

Never heard of one named like that

Maybe astralabs

[hallow thistle]

Yeah, I assume if it used to be there long before I joined this server. Making another channel for that would be such a waste for this server, as #🤖│bots is one and only channel you can use bot for anything.

[latent kettle]

My loss/g/fm is increasing. Everything else loss/g/total, loss/g/mel, loss/g/kl etc are going down. What should I do

[simple ore]

this is the case where averaging the loss window may provide a better outlook

the normal logging happens at the end of the epoch, so the FM numbers are prone to wild fluctuations depends on what the model pick as the very last random sample