✨│ai-help

[graceful obsidian]

cause, my fork was used already by 5~ people, me locally and on jupyterlab ( saturn cloud ) and no such issues occured

hmm

[leaden yacht]

That did not work

wait a minute

the error says it cant find the file

js as it did with dotenv

[graceful obsidian]

it says so whenever there's 1) path issues 2) naming issues 3) ffmpeg issues

you have ffmpeg added to path?

can you confirm "ffmpeg" working in cmd?

[leaden yacht]

wtf

[graceful obsidian]

as I thought

[leaden yacht]

I js looked into the env variables

[graceful obsidian]

ye?

[leaden yacht]

I was looking at path and ffmpeg wasnt there

[graceful obsidian]

https://phoenixnap.com/kb/ffmpeg-windows

follow this ^

trusted one

[leaden yacht]

though I recall downloading it as one of the first things when I got this laptop a few months ago

[leaden yacht]

dw I know how to do it lol

otherwise I would not be lookin at the env vars

[graceful obsidian]

phew, good

also pardon that I " simplify " stuff or so, habit from helping newbies ( which could at one point be reading the chat history. That's less issues for me later on

but ye, removed a bit of dc offset in ur audio and made it proper mono
( rvc / ffmpeg would most likely fuse the channels and that with dc differences could be, yea. better not

[leaden yacht]

dw about that

I have that too

but this was not an ffmpeg error yk

There are files not found in the python libs

so I think imma js remove all of my packages and reinstall your list

[graceful obsidian]

like I suggested, I'd purge all python stuff you have

get clean n fresh 3.10.6, my list and ffmpeg added properly

then it has to work 100%

maybe I should prepare venv method too in future hmmm

[leaden yacht]

already busy

I js did pip freeze into a requirements.txt file and then pip uninstall -r reqs.txt

should be done soon enough

installing the reqs now

[graceful obsidian]

🗿

[leaden yacht]

brb

[graceful obsidian]

alr

[leaden yacht]

back

its done

[graceful obsidian]

ye, check if it works now

[leaden yacht]

nah

but hold on

I forgot to restart after installing ffmpeg lol

The packages being reinstalled didnt work so its either ffmpeg or idk anymore

Ffmpeg is seen as a command now

[graceful obsidian]

that's hella sus

[leaden yacht]

finallyyyyyyyyyyyyyyyyyyyyyyyy

it was ffmpeg after all smh

[floral token]

Petition to get @graceful obsidian on the AI Hub payroll

[graceful obsidian]

lol

[leaden yacht]

weird tho cuz I already had the ffmpeg folder in my C:\ drive, but it got removed from env vars

[leaden yacht]

defi

hold on

Why aint my 4060 showing

[graceful obsidian]

huh 🤔

[floral token]

oh no

[graceful obsidian]

weird, would detect 1050 or tesla t4 no issues

[floral token]

update from me @graceful obsidian I installed requirements for your fork using miniconda instead of powershell/regular terminal and it seems to be installing everything as if it didn't exist before. I'll keep you posted.

[leaden yacht]

and how come T:\hifi_testing_new-main>py infer-web.py --port 7897 is the command starting it but the site opened on port 8000

[graceful obsidian]

8000's the port I use on saturn
so it's kinda hardcoded

but that's not the issue

as for gpu

I just noticed mainline changed the code

Gonna compare the configs ( was doing a fetch a week ago or so

maybe they screwed something up

[leaden yacht]

alr

[graceful obsidian]

and just now fixed

brb

[leaden yacht]

ping me when ur back

but whats the difference between using ur fork and mangio??

[graceful obsidian]

should be detecting the gpus

what the heck

[leaden yacht]

well it isn't

yk what

imma js try it

Ill put in 0

and try to see what it does

[graceful obsidian]

stability, bugs present in mangio, and those ^

also lack of bs auto syncing

[leaden yacht]

maybe its a text error

[graceful obsidian]

issues you guys have were never happening

neither on rvc devs serv

when I had my testers

aaaand I use it on my machine / saturn environment as well

so that's hella sus

[graceful obsidian]

that won't work

[leaden yacht]

should I use rmvpe or rmvpe_GPU

[rare goblet]

Ayo? @leaden yacht level 7 !!!

[graceful obsidian]

because 0_gt has them all sliced and processed properly

[leaden yacht]

oh

[graceful obsidian]

well, let's try another way to exclude few things

get this one
https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI/tree/main

they have other method for processing audio

[floral token]

since I'm having a similar issue should I try this fork?

[graceful obsidian]

that's the og rvc

actually

ye, try as well

we'll see if it's machine dependent or some specific cases don't like my fork

[leaden yacht]

but isnt that one outdated aswell?

[graceful obsidian]

nope, this is the mainline, it is the og maintained by devs

constantly

recent changes were only minor thingies and some overhauls but still

get the repo tho

not the package

[leaden yacht]

hold on

I actually did try the thing I said

I put in 0 for gpu and got this as output: ['infer/modules/train/extract/extract_f0_print.py', 'T:\hifi_testing_new-main/logs/Gregory-FNaF-SB', '8', 'rmvpe', '64']
infer/modules/train/extract_feature_print.py cpu 1 0 0 T:\hifi_testing_new-main/logs/Gregory-FNaF-SB v2 False
exp_dir: T:\hifi_testing_new-main/logs/Gregory-FNaF-SB
load model(s) from assets/hubert/hubert_base.pt
move model to cpu
all-feature-166
now-166,all-0,0_1.wav,(108, 768)
now-166,all-16,0_118.wav,(149, 768)
now-166,all-32,0_14.wav,(149, 768)
now-166,all-48,0_16.wav,(149, 768)
now-166,all-64,0_180.wav,(99, 768)
now-166,all-80,0_20.wav,(149, 768)
now-166,all-96,0_22.wav,(149, 768)
now-166,all-112,0_28.wav,(149, 768)
now-166,all-128,0_51.wav,(98, 768)
now-166,all-144,0_72.wav,(149, 768)
now-166,all-160,0_91.wav,(149, 768)
all-feature-done

[graceful obsidian]

aside that

[leaden yacht]

isn't that how its supposed to be?

[graceful obsidian]

because rmvpe vs rmvpe-gpu is using cpu

it will work regardless

[leaden yacht]

ohhh

rip

[graceful obsidian]

and hop is set for cpu rmvpe variant as it's more accurate

imma debug the stuff but first need you guys to test mainline rvc repo

and dl this:
https://drive.google.com/file/d/1xY4DM4D-dwqLIzpelZKfYD2CMYu305Ng/view?usp=sharing

it's the runtime that's coming prepackaged with packaged rvc releases ( which are outdated compared to repo itself

[leaden yacht]

this one isnt even working

[graceful obsidian]

ye, you need the runtime

[leaden yacht]

oh its not included?

[graceful obsidian]

you unpack the runtime folder to rvc's folder

it is, but in packages which again, are outdated

so I isolated it ( my fork uses that for realtime voice changer, hence why

[leaden yacht]

damn why did they make everything so complicated 💀

[graceful obsidian]

wish I knew ngl

but kalo, mangio and rest of the team didn't make it any easier

doing bazzinga mess with their releases

leaving it kinda broken and outdated with few hardcoded quirks

[leaden yacht]

I put the runtime folder (extracted) into the og rvc folder

[graceful obsidian]

yes, now it should work

[leaden yacht]

hope so

[graceful obsidian]

let's first see if processing works and if gpu's detected

if all turns out that way, might be I'll have to do another huge fork revision

💀 hope not

last one I had 2 or 3 months ago fffff, took me few days, porting certain things into new structures

[leaden yacht]

hope so, I js want this to work lol

[graceful obsidian]

btw, as another potential debug thing

where do you store rvc?
could it be disk / paths issues?

[leaden yacht]

T:\

There I js put the folder

[graceful obsidian]

and ffmpeg is on

[leaden yacht]

shouldnt be a problem as even mangio had no error's

[leaden yacht]

yup

[graceful obsidian]

hmmm

[leaden yacht]

all python dependencies are there too

[graceful obsidian]

ye

how's the runtime?

[leaden yacht]

Js moved it

starting now

Back to mangio it is ig

[graceful obsidian]

can you update the torch

tho, aside that

[leaden yacht]

alr

[graceful obsidian]

check processing itself

first

and as for torch

pip install torch torchvision torchaudio
but that later

[leaden yacht]

btw I installed the versions u gave in the reqs.txt

[graceful obsidian]

should work regardless

[leaden yacht]

already statisfied

torch is installed

[graceful obsidian]

a

well, how's processing, works?

[leaden yacht]

hold on ill use that command rq

wow it didn't have torchvision installed

isn't it supposed to do that automatically?

[graceful obsidian]

well

[leaden yacht]

restarting rvc

[graceful obsidian]

I've no clue how mangio was working then

[leaden yacht]

I mean I did js flush all of my packages

[graceful obsidian]

orrrr, they had quite a bit modified / older detection mechanic

oh, makes sense then

[leaden yacht]

and reinstalled them using your list

[graceful obsidian]

torch stuff should always be installed after python tinkering

[leaden yacht]

nope

no gpu detected

[graceful obsidian]

huh, guess you're bound to use mangio
if even mainline doesn't work

something's up with your machine or os should I say

[leaden yacht]

imma del the og rvc fork as there arent any differences in errors with urs and og rvc

[graceful obsidian]

you use win 11 by any chance?

[leaden yacht]

yeah

[graceful obsidian]

Orrrr that is the issue in some way

others with issues had win11 too

[leaden yacht]

probably that then

[graceful obsidian]

but can't confirm that really, just a speculation

[leaden yacht]

alr

[graceful obsidian]

we'll just uhh

we'll just remove auto syncing

[leaden yacht]

imma restart mangio and hope I dont get any dependencies errors

[graceful obsidian]

first check if mangio runs, if issues, use their reclist

[leaden yacht]

I will

already installing the reqs

[graceful obsidian]

then, after confirming it works, Imma have to remove syncing somewhat

[leaden yacht]

well mangio was even using another torch version

and did not require torchvision

Mangio does find my gpu tho

maybe look into the mangio code too

[graceful obsidian]

maybe will adopt their detection as a fallback option

processing works too?

[leaden yacht]

and if u do, make sure u get the right one from their github as there are 2 (one for only conversion and one for also training)

[leaden yacht]

I would

[graceful obsidian]

ye ik

I have all releases

stored

[leaden yacht]

Everything works in mangio

nothing went wrong

I trained js a few hours ago

the model isnt the best but it works so yeah

[graceful obsidian]

works now too?

[leaden yacht]

Imma restart the training and let it run for a few more hours

[leaden yacht]

it always did

You said the graph wasn't right in mangio tho

[graceful obsidian]

ye, that's the downside
auto syncing

but it's burried so deep in the code

last time I removed it, it took me the most of time

as kalo / others liked to scatter the code

throughout many files

[leaden yacht]

u said it was inaccurate

[graceful obsidian]

rather than have it unified

[graceful obsidian]

that's still correct

[leaden yacht]

wha

auto syncing?

[graceful obsidian]

because their math behind it is inaccurate

it can't be calculated 100% accurately

[leaden yacht]

hm

[graceful obsidian]

it will log at wrong steps, the tensorboard metrics

[leaden yacht]

any other forks that are better then?

[graceful obsidian]

giving you not accurate metrics

[leaden yacht]

that work?

[graceful obsidian]

they're built on top of what mainline is

some adopted mangio's feature like I did

some didn't

so if mainline doesn't work for you, quite a high chance they won't either

[leaden yacht]

true but even mangio is, some js changed it

[graceful obsidian]

the thing with mangio is

last version was made based on mainline going through transition

and general file structure, script separation / unification process

it's just a code mess inside

[leaden yacht]

hm

ill see whatever other things the guide has

[graceful obsidian]

@leaden yacht see?

even just the config ( fp16/32, gpu and so on detection )

is very different in structure

[leaden yacht]

true

but it works

and mainline does not

[graceful obsidian]

which is a huge sus

as in 90% of cases, esp on rvc devs

it's the opposite

mangio causes " no data " issues and such

ffmpeg, gpu

etc

that's why I find it hilarious this time 🗿

[leaden yacht]

I mean I could try applio, is that more accurate than mangio?

[graceful obsidian]

it is more so, whether it auto syncs or doesn't

[leaden yacht]

it also causes no errors

[graceful obsidian]

in which case, we sync it manually

as it should be done

[leaden yacht]

wdym sync it

tensorboard?

[graceful obsidian]

in simple words

an epoch as corresponding steps

per each

those are known after having at least one epoch

.pth in weights

mangoo release would auto calculate and sync it

" estimating " the steps

but kalo was a hard shell

not listening to me

[floral token]

Okay I'm back. extracting that runtime to OG RVC fork

I'll lyk

[graceful obsidian]

when I told him rvc aside of 0_gt splitting

does internal further slicing too

which can't be calculated, at least yet, in my case

[leaden yacht]

no idea what 0_gt is

[graceful obsidian]

0_gt is model dir's folder for ground truth samples

sliced and processed .wavs
( obtained from preprocessing step / first step in the ui )

amd so is 16k dir, which are 16khz .wav files

[leaden yacht]

but should I or should I not switch to applio?

[graceful obsidian]

I dunno applio well, ain't sure if they adopted the auto sync or not

[leaden yacht]

is autosync better than manual?

[graceful obsidian]

no

[leaden yacht]

alr

[graceful obsidian]

like explained, it's inaccurate

your epochs could be having 10 steps per each

but auto sync could do it 9 or 11 maybe

or 12

that causes tensorboard files logging wrong metrics

from " nonexistent " points
as in, irrelevant as they don't match epochs
and in such important cases, where one epoch could be 2583steps ( say, idk, 300+ epoch )
it might contain wrong metrics / loss logged or stored

tl;dr, autosync is crap

[lean valve]

hey!

could someone send a colab link with the latest rvc update?

and if possible a tutorial of some kind

ty

😭

[leaden yacht]

but if I don't like the model and think I accidentaly DID get the wrong one, I can js go back one save ?

I mean it doesn't have to be so hard

and I can even install ROD to detect overtraining

[graceful obsidian]

well, it is not that

imagine you have 500 epochs

[leaden yacht]

mhm

[graceful obsidian]

good luck and have fun testing each

" blindly "

as metrics tensor gives are false and inaccurate, not reflecting the actual epoch

btw

[leaden yacht]

but are they that inaccurate that you cant even make up where smth happens?

[graceful obsidian]

just tested

fresh dl of my fork repo

works as well

[leaden yacht]

nahh

[graceful obsidian]

what I mean is, it's not the repo either

[leaden yacht]

Go back down a lil bit

to the gpu

[graceful obsidian]

that is normal for my case

as I have amd

gonna test one thing

imma spoof having nvidia gpu in

and see if it's detected

[leaden yacht]

alr

but at least for me, the model was actually doing pretty good so far and I stopped it too early, so Imma restart mangio

I tested the latest save and it worked pretty well

[floral token]

So I guess it's mangio for me then too. seems like we're having the same exact problem

[leaden yacht]

I mean it was not good, but wasnt done training yet which explains that

[floral token]

just installed og rvc added the runtime folder and nothing

[leaden yacht]

same

and code's fork also doesn't work for me

[graceful obsidian]

ye weird thing

even if my fork was the faulty one, mainline should work 100%

[leaden yacht]

and how tf did my mangio train without ffmpeg tho

[rare goblet]

Ayo? @leaden yacht level 8 !!!

[leaden yacht]

I mean I did only install it apparently when I was tryna fix the error with your fork

[graceful obsidian]

oh

wait a second

you have ffmpeg

[leaden yacht]

but how did it js train without ffmpeg

[graceful obsidian]

but do you have

[floral token]

funny part was my mainline was working. it had my gpu and everything. but it was taking 7 minutes/epoch

[rare goblet]

Ayo? @floral token level 5 !!!

[leaden yacht]

damn

[graceful obsidian]

this

this is a bridge

between python and ffmpeg

which is required all the time

[leaden yacht]

for me (mangio) it was going like an epoch if not 2 per minute

dude

code, that is not the problem

the problem was fixed

[floral token]

I do not have this

[leaden yacht]

the whole problem, was that my gpu wasnt shown

which is not related to ffmpeg is it?

[graceful obsidian]

I am talking about preprocessing step rn

[leaden yacht]

I had no problems with that

[graceful obsidian]

gpu is another thing

[leaden yacht]

I fixed that

[floral token]

nvm I have it I had a typo lol

[leaden yacht]

but I suggest looking into the mangio code, even if its messy it could have smth you could use

[graceful obsidian]

alr then, so it was just the ffmpeg installation thing
as for gpu

must be the way it detects em

[floral token]

I think mangio is the only one I have not installed at this point lol

[leaden yacht]

mangio is the 1st one I installed and the one that worked the best

[graceful obsidian]

mainline / my fork and others probs too detect it like so

mangio on the other hand

reads torch device under ID 0

you guys have iGPUs ?

what's the ID for your gpu in task manager

0 or 1

[leaden yacht]

Never had any errors with it, except for my previous laptop which was because the mangio code needed to be changed cuz otherwise it wouldnt support GTX 1650

[leaden yacht]

0

I can change it to 1 tho

Idk why but if unused I can eject my gpu which does nothing but change that

[graceful obsidian]

well, I have 2 potential ways of solving it

[leaden yacht]

💀

[graceful obsidian]

1 thing would be disabling igpu if you have one, just gpu ( bios

other is, I can port the detection from mangio

lol

[leaden yacht]

yeah uh I aint changing my BIOS for this

[graceful obsidian]

as I thought.
idk, maybe I will look into it later later or tomorrow

might port it

[leaden yacht]

alr

for now imma train using mangio

which has not let me down yet

I hope it doesnt in the future too lol

[graceful obsidian]

as long it works, that is

ye

[leaden yacht]

exactly

and if it works, don't even touch it

[graceful obsidian]

just lame for the log syncing but ig, just have fun testing the ckpts x' )

that can be fun too

in a way

lol

[leaden yacht]

I mean I can make a full cover of like 4 mins in about 10 secs

what if I js take 30 seconds

I mean that will be going fast and I have nothing to do with my time

[graceful obsidian]

true

[rare goblet]

Ayo? @graceful obsidian level 23 !!!

[leaden yacht]

except more important stuff

but who cares about that

[graceful obsidian]

neat, level matching my age

kek

[leaden yacht]

I mean its only my mental health 💀

anyways imma go eat smth

[graceful obsidian]

🗿

[leaden yacht]

havent eaten today

[graceful obsidian]

alr, have a good meal then

[leaden yacht]

thx

[graceful obsidian]

and gluck with the model

[leaden yacht]

thanks

btw if u need me to test smth u can msg me

I'd be happy to

[graceful obsidian]

gotya

will do when needed

[hushed lion]

Does anyone know RVC Google Colab Notebook with a multi-inference function? It was removed from EasyGUI...

[proper shale]

Maybe Applio?

https://colab.research.google.com/github/iahispano/applio/blob/master/assets/Applio.ipynb

Dunno if this is restricted or not so

[hushed lion]

thanks anyway

[rare goblet]

Ayo? @hushed lion level 1 !!!

[floral token]

can I ask, why did you want to switch away from mangio if it was working for you? @leaden yacht

confirmed mangio is working. your answer is in their code somewhere @graceful obsidian

[leaden yacht]

Code said that the logging for tensorboard was inaccurate and not right

so

yeah

that

sorry for the late response btw

js got back

I don't think the training went well

[floral token]

all good, did that prevent you from finishing your model or just end result was bad?

[leaden yacht]

I left it on while I went to eat

wait

[proud elbow]

true for shorter dataset and higher batch size (less steps per epoch)

[leaden yacht]

do u mean the inaccurate things?

[proper shale]

ehhh doesn't seem like that

[leaden yacht]

Cuz I did not notice anything

[graceful obsidian]

lots of collapses

[proud elbow]

use 0.987 smoothing

[leaden yacht]

oh yeah

[floral token]

yeah the logging inaccurate error

[proper shale]

yeah increase the smoothing

[leaden yacht]

I forgot I had to put that on

[graceful obsidian]

post a ss with 0 smoothin

and disabled outliers off

[proper shale]

and that too

[graceful obsidian]

or ideally, just send me the tfevents file

[leaden yacht]

this?

[graceful obsidian]

need one without smoothing and with disabled " ignore outliers "

now, click the 3rd button

" frame "

from those 3 under the graph

to resize

[graceful obsidian]

now, one sec

general tendency, flatlining

[floral token]

at what point in training is this tensorboard? I still only am at the actual model training.

[snow rose]

How to continue training? beacuse i think i broke it XD

[graceful obsidian]

these are collapses

to be avoided

[proper shale]

after u start training

[leaden yacht]

that meansss??????

[proper shale]

just insert same settings, dont preprocess or feature extract again and it should work

[floral token]

another bat file to open? where do I see that UI @proper shale

[graceful obsidian]

avoid purple zone

green ones seem most promising

[leaden yacht]

Is this the point of overtraining then?

[graceful obsidian]

if it's for total G

[leaden yacht]

this was the g/total yeah

[proper shale]

@graceful obsidian can u send real quic the TensorBoard venv bat

[leaden yacht]

do u need the d/total too?

[proud elbow]

show the axis scale, if there are some less than 30-40%

[graceful obsidian]

take a look at this

notice the tendency switch

[graceful obsidian]

and magnitude decreasing

[proper shale]

try out the model later

[graceful obsidian]

your set has quite a bit of silence + you could try smaller batch_size

[leaden yacht]

I have no idea whatever magnitude or tendency means ;-;

Oh I need to remove the silent parts too? I thought it only picks up the parts with noise

[graceful obsidian]

have a look at each zone's " range "

more or less

[snow rose]

no it does not work

[leaden yacht]

yeah thats true

[graceful obsidian]

it would go down, with a curve ish trending

[proper shale]

ya might have messed up on something then

[graceful obsidian]

then up and kinda flatlining above the prev avg

[leaden yacht]

so I should remove the silence and then try again?

[graceful obsidian]

mainly batch_size and silence

also, those, again, aren't accurate, that's just rough approximation

[snow rose]

give me instruction from the base

[leaden yacht]

but am I not supposed to look at it with this on and smoothing 0.987?

[graceful obsidian]

they might not reflect your model's epochs

but "inbetweens"

[leaden yacht]

its 1 file

[graceful obsidian]

off

[snow rose]

beacuse i have D and G path saved

[leaden yacht]

and smoothing?

[graceful obsidian]

smoothing shouldn't be really used as a " thing for granted "

[leaden yacht]

The guide says it needs to be 0.987

[graceful obsidian]

that's for people who can't estimate the trending

[leaden yacht]

I mean I can't

[proper shale]

the instructions are literally just... insert same settings (name of model, batch size, etc) dont preprocess/feature extract again, train model

[leaden yacht]

I don't even understand half of the things you were saying about it tbh

[graceful obsidian]

it's not hard

take a look

just imagine a line going through those points

kinda an " avg "

[leaden yacht]

mhm

[graceful obsidian]

that's what smoothing does

in a way

[floral token]

dramatic drops = bad

[leaden yacht]

ohhh

[graceful obsidian]

squashing all together and averaging

[snow rose]

@proper shale

[leaden yacht]

so smoothing js gets the averages?

[floral token]

think of that line like a magnet

[graceful obsidian]

but if you just visualize it yourself vs using smoothing

you can see actual peaks

and dips

[leaden yacht]

isn't that useless? it hides a lot of the important info

[graceful obsidian]

familiarize yourself with this

that's just some examples

and each case is different but similarities and patterns, as you go on with training models, start to look familiar

[leaden yacht]

first ill js remove the silences alr?

[graceful obsidian]

it's not as straightforward sadly

[leaden yacht]

wdym

[graceful obsidian]

because, imagine you have 3 samples

[leaden yacht]

I can js cut out the silence using audacity

[graceful obsidian]

each sample is a sentence

you wanna minimize the silence-gaps between samples

not the actual words / phonemes

[leaden yacht]

wait

[proper shale]

wait if you have replaced the d/g pretrained path thing you shouldn't have done that

[leaden yacht]

so I should have more seperate files instead of one long one?

[graceful obsidian]

nono, 1 file is fine

but you gotta remembe that dialogue lines

are just fused many samples ( each dialogue part )

into 1

[proud elbow]

do audio labeling

[snow rose]

0_0 wait so how to conitnue training that beacuse someone telled me to do that

[leaden yacht]

whats that

[graceful obsidian]

example is

[proper shale]

bruhhhh

[graceful obsidian]

this is a fused big sample

[proper shale]

okay so

[graceful obsidian]

but in reality

it's many samples / dialogues

next to each other, naturally

[proper shale]

ig u gotta type the path of the pretrained g and d files now

[proud elbow]

Analyze > label audio

[snow rose]

WHAT

[proper shale]

i dont remember where they are specifically, but you have to start from the folder name

[proper shale]

yeah

[graceful obsidian]

@leaden yacht first tell me, what batch size you use

[leaden yacht]

wdym what batch size?

[snow rose]

smth like that?

[leaden yacht]

8the batch per gpu?

[proper shale]

should be like, pretrained_v2/something something

[proper shale]

no thats literally your model's g d files, you need the actual pretrains

[graceful obsidian]

yes

cause the thing is

[leaden yacht]

8

[graceful obsidian]

depending on the set duration

smaller batch size

causes noiser graphs

( lower one )

higher batch sizes, cause more " consistent " graphs

[leaden yacht]

I set it to 8 when I given the advice that I should

[proven pecan]

Did rvc 3.0 ever come out?

[graceful obsidian]

but it is not as simple as " more consistent = better "

[snow rose]

wait what

[graceful obsidian]

and neither is any batch_size " the best "

each dataset and voice needs own settings

[proper shale]

no

[graceful obsidian]

one model might have smaller set, one bigger
one voice might be more complex in patterns, one not

the thing with batch_size is, more you use, more examples you give of voice samples to AI

and it's just as with people' learning

if you learn too little? you won't learn much

[proven pecan]

What’s the best ai voice cloning software atm for live voice changing?

[graceful obsidian]

if you learn too much? you won't learn much

[proper shale]

see, the pretrains are the heckin soul of your model

ya need that shit there, in the path you wrongfully inserted your d/g file

[graceful obsidian]

it has to be just right per model / set / case

balanced

[snow rose]

so i just need to put D and G into assets folder in pretrainedv2?

[proud elbow]

u can try any power of 2 values, i.e. 4, 8, 16

[proper shale]

no, you need the actual pretrains

[leaden yacht]

but how do I fix the audio? You said there was too much silence or smth

[proper shale]

which are in pretrained v2

[proper shale]

https://rentry.co/VoiceChangerGuide

[graceful obsidian]

for now you can ignore it

For your sub 10 min in duration set

[proper shale]

what's your sample rate

[graceful obsidian]

try batch_size 14 or 16

[snow rose]

32k

[graceful obsidian]

if the graphs are too " flat " or do funky things, attempt 12

8 might be too little

[leaden yacht]

alright ill try

[proven pecan]

Isn’t that project discontinued?

[rare goblet]

Ayo? @proven pecan level 3 !!!

[proper shale]

assets/pretrained_v2/f0D32k.pth
assets/pretrained_v2/f0G32k.pth

[leaden yacht]

do I js remove the pth files from my weight folder and start training again with the right batch size?

[proper shale]

the voice changer no

just the guide

[proven pecan]

What was?

[proper shale]

but it has enough info and is pretty recent

[snow rose]

and what next?

[proper shale]

now just out them in their right places @snow rose

[proven pecan]

I tried that project about eight months ago and it was pretty poor in terms of results. Has it improved much since then do you know?

[proper shale]

well, it works okay

[proper shale]

ya got it

[proven pecan]

I need something that works really good

[snow rose]

and what next?

[proper shale]

well, try it out

[proper shale]

now just train ig

[snow rose]

how do i continue training my denji model

[graceful obsidian]

@leaden yacht here's kinda a visual

FIGURATIVE

on how increasing / decreasing affects the noisiness

in " distribution "

[leaden yacht]

but isn't it different for each dataset?

[graceful obsidian]

CAN but doesn't have to be that way per those values for your case

^

[graceful obsidian]

^

[graceful obsidian]

Correct

But that's so you can better visualize it

In ideal conditions

we wanna strike inbetween

• just stable enough but not too noisy
• one that goes rather stably down ( downward trending
• sometimes doing few trainings, say, at 8, 12, 16 can help you find what is close to your good batch
• from there, you wanna find inbetweens, could be 13, but be 10

[leaden yacht]

and idk if it puts more load onto the gpu, but will my gpu be able to do it with batch size 14 or?

[proven pecan]

Was rvc 3.0 cancelled? I was waiting for that

[leaden yacht]

ah thx

[proper shale]

well, no

[snow rose]

MJ

[proven pecan]

What happened to it?

[snow rose]

help me out

[proper shale]

i think rvc peeps r still working on it, they've found some limitations probably

[proven pecan]

It was just about done /ready to release like 8 months ago

[graceful obsidian]

^ @leaden yacht and here

[leaden yacht]

alright

[proper shale]

something to do with contentvec from what I'm aware

they wanna make it better ig

[leaden yacht]

^

[proper shale]

or its just that rvc boss moved on

[proper shale]

?

[leaden yacht]

dont know if its any difference but still

[graceful obsidian]

vram size?

[snow rose]

what do i do next

[graceful obsidian]

if you have 12/16 gigs

[leaden yacht]

8gb

[graceful obsidian]

you're good

oh

well

[proper shale]

just train the model now

if you've already got same settings and all

[snow rose]

from the beggining?

[graceful obsidian]

In that case, you can try with 16, 14 and 12, see which one works. If all failed, go for 10

[proper shale]

it'll resume

automatically

✨

[proud elbow]

it can go up to 8 batch

[leaden yacht]

Alr, should I start with 13?

[graceful obsidian]

for instance, 15 gig tesla t4 was able to go comfortably batch_size 16 on 40+ min sets

[snow rose]

how it will know that i want to resume denji

[graceful obsidian]

I'd try 14 or 16 Martin

[leaden yacht]

Alr

what will happen if its not enough?

[proper shale]

because it'll load the d/g files automatically

just do it, itll work

[graceful obsidian]

we'll tell by the graphs

[leaden yacht]

alr

[graceful obsidian]

or your gpu will refuse to go on with that batch

[leaden yacht]

this good?

[proper shale]

unless u screw up

[proud elbow]

it still runs but will get significant bottleneck

[snow rose]

all good?

[leaden yacht]

bottleneck?

[proud elbow]

performance

[proper shale]

id just change "saving small finished model to weights folder" to yes

and then u can do it

[proud elbow]

though not related to the results

[leaden yacht]

ill js try 12 first

[snow rose]

i just click train model yes?

[proper shale]

yeah

[snow rose]

error

[proud elbow]

it takes around 10 gb

[graceful obsidian]

it's dataset dependent

[proper shale]

okay, what's it say

[graceful obsidian]

10 gb for 10 min set at batch 12 doesn't quite scale like that

tesla t4 15 gig was fine with 16/18 for batch 45 mins + set

[proud elbow]

talking about vram, but not really related to the results

[graceful obsidian]

ye, I talk about it too

[snow rose]

"system cant find the specified path"

[graceful obsidian]

there's no saying batch_size 12 for 10 min set will take 10 gigs

[leaden yacht]

I have a 4060 8gig with about 9 mins set

[graceful obsidian]

it's not linear like that

[proper shale]

remove the spaces from "vocal spliter" folder name

and try again

[proud elbow]

colab T4 only takes 6 gb vram, yet local one takes 7~8 gb on batch 8

[graceful obsidian]

you forget about allocations

and the os itself having to reserve some for display

colab doesn't do that

[graceful obsidian]

for approx 10 min sets, typically you want above 12 batches

4-8 ranging is rather for smaller or more monotone sets I'd say

[leaden yacht]

Ill set it to 12 alr?

I mean I only have 8gigs of ram

[graceful obsidian]

you can try
but there's that thing with mangio crepe itself

if you go for that instead or rmvpe ( in case you do, ofc )

[thick swan]

yo anyone know if there any sites running still locally

or public

[leaden yacht]

I always use rmvpe

[graceful obsidian]

a, then no issues

[leaden yacht]

for conversion or training

both

[graceful obsidian]

Alrighty

[thick swan]

either both or conversion

i cant figure out how to run local on my mac anymore

[leaden yacht]

wait how do I reset tensorboard @graceful obsidian ?

[rare goblet]

Ayo? @snow rose level 6 !!!

[snow rose]

still

[leaden yacht]

Because I opened the tensor tab but the old graphs are there

[graceful obsidian]

did you open it up with a specific dir in location or per-file/path-to-file

[leaden yacht]

I js opened the tab

It was still open

[graceful obsidian]

use this

put it in any folder which would contain folders, and each ( named after your model for convenience ) has your models' tfevents files

[leaden yacht]

I already have tensorvenv

[rare goblet]

Ayo? @leaden yacht level 9 !!!

[pastel oak]

@proper shale any idea on how to determine batch size on ov2 pretrain roughly?

[graceful obsidian]

that's just an executor for tensor

not really installer or such

[leaden yacht]

I know

but i alr have tensorven.bat, which does exactly the same

[graceful obsidian]

works like so:

just in case

then whenever you delete the tfevents file / folder ( or add something new )

just refresh the tf page

and you can just deselect older ones to not show up

leaving only the most recent

[leaden yacht]

I know

But I forgot to change the name lol

anyways its training now

[snow rose]

@proper shale

[graceful obsidian]

Typesome

[snow rose]

me

[graceful obsidian]

Jeśli potrzebujesz asysty, pisz na dm

If you need assistance, write me a dm

[snow rose]

where are you from

[graceful obsidian]

you guessed it right

[leaden yacht]

but @graceful obsidian , what do I need to do the next time to improve the dataset?

[graceful obsidian]

well

these

newest and easiest workflow

my older methods are too complex

[leaden yacht]

and what was wrong with my current one?

[graceful obsidian]

well, silence and noise I think

had just a quick look really

But audio processing and such advanced stuff is something I don't teach people anymore, too complex

[leaden yacht]

yeah I figured out that last part already

a little too complex

[leaden yacht]

so I should have less silence?

[half cove]

just truncate

[graceful obsidian]

between dialogue lines / sentences, ye ( do not touch gaps between words in sentences or between phonemes )

that helps with collapses

[leaden yacht]

alr imma edit that then

[graceful obsidian]

25 to 40 ms is fine

between em

[leaden yacht]

damn thats short

I mean have u heard the timing in my set 💀

[graceful obsidian]

well yea

that's why, yea, I do it actively

as I build my sets

🗿

cause audacity's silence truncating doesn't care if it's between-dialogue gap or between-phoneme gap

sure, there's labels too butttt, I dunno it and haven't used it

[half cove]

labels are pretty good

its the thing i usually use

[graceful obsidian]

Thought of that, or whisper ye

but for now, my method works so, but will surely give it a shot eventually

[brittle wing]

Is that mode collapse?

[leaden yacht]

how does that work?

[brittle wing]

45k is it collapsing?

It's NOT overtraining If someone's gonna call it that No it's not my dataset is big to overtrain that early

[proud elbow]

"label sounds" does split into many files like RVC preprocess but without cutting middle of sentences, also it helps reducing collapses

[graceful obsidian]

yea I know, I meant that I dunno the ui of it / workflow of it in the program itself

Am well aware it's similar to whisper, just, via vol threshold and gap control

[lusty spruce]

guys i got an error when i wanna download a model

"JSONDecodeError Traceback (most recent call last)
<ipython-input-6-cf2b15191ac5> in <cell line: 31>()
31 if os.path.exists(config_path):
32 # File exists, proceed with creation of creds and client
---> 33 creds = Credentials.from_service_account_file(config_path, scopes=scope)
34 client = gspread.authorize(creds)
35 else:

5 frames
/usr/lib/python3.10/json/decoder.py in raw_decode(self, s, idx)
353 obj, end = self.scan_once(s, idx)
354 except StopIteration as err:
--> 355 raise JSONDecodeError("Expecting value", s, err.value) from None
356 return obj, end

JSONDecodeError: Expecting value: line 1 column 1 (char 0)"

this is said

[proud elbow]

outdated colab, surely

-colab

[azure marsh]

☁️ Google Colabs

• Applio, by IA Hispano Google Colab
• AICoverGen-WebUI, modded by Hina Google Colab
• AICoverGen-NoWebUI [English], by Ardha, fixed by Eddy, Hina and Gdr Google Colab
• AICoverGen-NoWebUI [Spanish], credits to Eddy, Hina and Gdr for translating and fixing Google Colab
• Ilaria RVC, by thestingerx Google Colab
• RVC Disconnected, by Kit Lemonfoot Google Colab
• easyGUI, by rejects Google Colab
• 🆕 UVR5 NO UI for Google Colab, by Eddy Google Colab

How to use?

🤗 Hugginface Spaces

• Applio, by IA Hispano Huggingface Spaces
• Ilaria RVC, by thestingerx Hugginface Spaces
• RVC-HFv2, by r3gm Huggingface Spaces
• AICoverGen, by r3gm Huggingface Spaces
• Advanced RVC Inference, by r3gm Huggingface Spaces
• RVC v2 Huggingface version, by Clebersla Huggingface Spaces

More commands: -audio, -local, -overtrain, -realtime, -rvc, -help

[lusty spruce]

so i just use another one?

[leaden yacht]

instead of having silence in between, and problems with that etc, should I js use all voice lines seperately?

all of those files?

[lusty spruce]

ty

[rare goblet]

Ayo? @lusty spruce level 1 !!!

[proper shale]

yeah idk tbh. it cant seem to find the gt wavs so you either screwed up on the model name or you'll have to restart from the beginning

[proper shale]

just test and see, but since we're talking shit below 10 mins 4-8 can do it ig

but it varies

[snow rose]

Codename tried to help me with this in dms but my brain cant handle this mess so i will start from beggining i think

[proper shale]

just remember next time to change no settings, ig

[fading summit]

does anyone know how to get your huggingface model link to work?

[jade shore]

Someone explain to me how I make a cover. I used to know but now i forgor

[rare goblet]

Ayo? @jade shore level 1 !!!

[leaden yacht]

Cloud (google collab) or local (hosted on your own hardware)?

[jade shore]

how do you use google colab

?

sorry im just dumb

[leaden yacht]

I don't use it myself, I have used it in the past but don't know about whats changed. What I can give you is the link to the official guide though

[leaden yacht]

no you're not, don't worry

[jade shore]

yes plz

[fading summit]

how do u make a huggingface link for your model?

[leaden yacht]

https://docs.aihub.wtf/rvc/cloud/ilaria-rvc/
https://docs.aihub.wtf/rvc/cloud/applio-colab/

Personally I recommend one of these 2, as I have used applio before and know it works well, and I have heard that ilaria works really good too from others.

[jade shore]

i remember using this: https://colab.research.google.com/drive/1Gj6UTf2gicndUW_tVheVhTXIIYpFTYc7?usp=sharing a while ago but im not sure it will work

[leaden yacht]

If you click on the links you will find the guide, with tutorial and full explanations about what things mean

If you don't understand something js send your problem here, and if you want you can ping me for it

[jade shore]

ok thanks so much

[leaden yacht]

I would js use the guide as those are frequently updated

[jade shore]

ok

[leaden yacht]

sticking with the guide makes it a lot easier

[jade shore]

@leaden yacht, I put in this link, https://huggingface.co/AstralBlade/Model/blob/main/FriskStoryOfUndertale.zip and it said that the download failed! What do i do?

[fading summit]

im still trying to figure out how to use huggingface cause i have a model just dont know what to do on huggigface

i made a model on weights.gg

and i just put in google drive

[leaden yacht]

sorry js seeing this now, where did you put it? can you send a screenshot?

[jade shore]

yes

[rare goblet]

Ayo? @jade shore level 2 !!!

[leaden yacht]

Does it also give an error msg with it?

[jade shore]

yes

[fading summit]

mine did the same

[leaden yacht]

I entered the link in my browser but it also didnt directly download

hold on imma check smth rq

Could you try this link: https://huggingface.co/AstralBlade/Model/resolve/main/FriskStoryOfUndertale.zip?download=true

[jade shore]

ok hold om

[leaden yacht]

At least in browsers it auto downloads then, hope it does the same for collab

[jade shore]

IT WORKED!!!

[leaden yacht]

Yooo

Great

So whenever you are having a problem downloading from huggingface, js take the huggingface link, ending in .zip and add "?download=true" to it

Which will make a direct download link

the link you used just sent you to the page of the file, where you can select what to do with it

[jade shore]

OH! I get it! thanks!

[leaden yacht]

yw

[jade shore]

https://media.discordapp.net/attachments/1194077895627374694/1217941586613370890/image0.gif

[fading summit]

so u just put the zips in the huggingface page

not the ptf

[leaden yacht]

pth?

nah you don't

you need both the .pth and index file

and put them in the same zip

[fading summit]

ok i did

[leaden yacht]

just upload that to huggingface

[fading summit]

what about the license?

i put unknown

[leaden yacht]

I don't know about that as I've never used huggingface before myself

I am making an account atm tho

[fading summit]

yeah itsvery confuzing

thats why google drive is easier

[leaden yacht]

True

but still once you get the hang of it you will probably find it a lot easier

question for the model makers, is the training doing well?

[low shard]

MIT

[low shard]

yep, let it train more

[leaden yacht]

so I js need to wait like a few hours and see if it has gotten a lowest point and starts going up again?

as in it starts overtraining then?

[fading summit]

does this work https://huggingface.co/mjfan1999/MichaelJacksonInvincibleEra2001V2/blob/main/MJInivincibleEra2001V2.zip?download=true

[full nest]

Why i can't do step2

[low shard]

yup, u need to see when it goes to the lowest point and then goes onlyy over that, that is training

btw you can also check https://rentry.org/RVC_making-models

[low shard]

!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.

[leaden yacht]

It isn't

at least not for me

https://huggingface.co/mjfan1999/MichaelJacksonInvincibleEra2001V2/resolve/main/MJInivincibleEra2001V2.zip?download=true

this does work tho

[fading summit]

im trying to put it in rvc and it says invalid link

[leaden yacht]

Idk why but the one u gave me is the same as I sent

[leaden yacht]

@fading summit if ur link doesn't work try to copy and paste this one

Yours doesn't work for me either

[low shard]

ur zip has a metatada.json which is not needed

also you need to put the added.index

[leaden yacht]

its not mine

[leaden yacht]

^

[fading summit]

whats added index

[rare goblet]

Ayo? @fading summit level 4 !!!

[leaden yacht]

idk why, but when I click his link it js doesn't automatically download like it should

do you have that too Nick?

[fading summit]

whats added.index?

[leaden yacht]

A file, it should be made when training

[fading summit]

i trained it on weights.gg

[low shard]

its like the file that controles the accent, if you dont have it, you can delete the metadata file and zip the model with only the pth, and upload it on google drive or hugging face. Just that without the added.index, it will prolly wont be good quality

[low shard]

wait what? they didnt give you an added.index?

[fading summit]

@low shardno

[low shard]

@willow lichen wait what? weights.gg doesnt train added.index ?

[leaden yacht]

Thats what that does? So basically the .pth is the actual voice and the index is the accent?

[fading summit]

https://www.weights.gg/models/clvgr3dnl001z959xq8w5o3h0

look for yourself

[leaden yacht]

You're right

there is no added index

[azure marsh]

RVC Guides (How to Make AI Cover)

Documentation

🇺🇸 English (main)

https://docs.aihub.wtf/

​

Translation by country

🇧🇷 Brasil (PT-BR)

https://docs.aihub.wtf/v/brazil

🇪🇸 Spain

https://docs.aihub.wtf/v/spain/

🇫🇷 France

https://docs.aihub.wtf/v/france

🇩🇪 Germany

https://docs.aihub.wtf/v/germany

🇯🇵 Japan

https://docs.aihub.wtf/v/japan

🇰🇷 Korea

https://docs.aihub.wtf/v/korea

🇵🇱 Poland

https://docs.aihub.wtf/v/poland

🇷🇺 Russia

https://docs.aihub.wtf/v/russia

🇮🇹 Italy

https://docs.aihub.wtf/v/italy

[low shard]

basically yea

[low shard]

you trained it or you js found it?

[fading summit]

@low shard i trained it myself

[leaden yacht]

damn

[low shard]

weird asf wtf?

lets see what river says, hes dev of weights.gg (dont spam ping him ofc lol), idrk about this

[leaden yacht]

but shouldn't it always give an index?

[fading summit]

@leaden yacht you would think so

[low shard]

technically, i honestly never trained on weights.gg, but it uses rvc which it gives you that yea

[fading summit]

i waited in a queue of 180 last night

[low shard]

Sorry but i dont really know about this

[leaden yacht]

You NEED an index file to use the model tho right?

[low shard]

could be a bug

[leaden yacht]

could be yeah

probably

[fading summit]

it works on google drive when i use it

[low shard]

well, not really, you can use it by just deleting the metadata file and zip the model with only the pth

[leaden yacht]

Wait u can?

[low shard]

its just suggested to have an added.index because it helps with accent

[leaden yacht]

I thought to start conversion you needed both a pth and index

never knew u could js use the pth

While the model is training on my laptop, I can still watch yt or smth right? I mean I have an iGPU

[low shard]

yea, whats ur gpu btw? also if you train on weights, its all online, it doesnt matter your pc specs bc it will use remote pc gpus

[leaden yacht]

Oh nah im training locally

As I showed tensorboard js some time ago

[leaden yacht]

I have AMD Radeon Graphics built in with the CPU, and an RTX 4060 with 8gb vram

but my laptop has a big screen so wanna watch yt while its busy training lol

[low shard]

oh yea my bad got you 2 confused lol

It seems to be still training good so yea u can do other things meanwhile

[leaden yacht]

greattt

[leaden yacht]

its going pretty good rn yeah

at first it was a lot more randomness when I had only 1 file with a lot of silenced parts, but I chose js to use all of the loose voice lines from the game directly

[fading summit]

maybe its because i dont have weights premium?

[leaden yacht]

like 160 files or smth, and now its actually going pretty well

[leaden yacht]

no probably not

I mean whats the point in having a free product that only gives half of what you wanted to train

[fading summit]

@leaden yacht im broke lol

[leaden yacht]

me too

[fading summit]

im just uploading it without the index

[leaden yacht]

you could do that yeah

Although it will probably sound worse without one

[fading summit]

maybe but i cant do nothing about it

it didnt give me one

[leaden yacht]

I mean you could also use another model

[rare goblet]

Ayo? @leaden yacht level 10 !!!

[leaden yacht]

You don't specifically have to train one yourself

unless you really want to

but its up to you

[fading summit]

yeah i did i made a better mj 2001 model

[leaden yacht]

nicee

[fading summit]

ive already uploaded it google drive

it works fine on google

it just doesnt have a index lol

[leaden yacht]

Question about the training, does the d/total matter? and what information would I get from it?

[fading summit]

well everyone elses links say resolve/main but mine doesnt

https://huggingface.co/mjfan1999/MJInivincibleEra2001V2/resolve/main/MJInivincibleEra2001V2.zip

that sould work

butt i cant upload it to voice models cause i dont have permission

[low shard]

it should be https://huggingface.co/mjfan1999/MJInivincibleEra2001V2/resolve/main/MJInivincibleEra2001V2.zip?download=true, you need to right click the download button next to the zip and do copy link

[leaden yacht]

Ok so you need to go to https://ptb.discord.com/channels/1159260121998827560/1159514067187277865

then use the /submit command and fill everything in

[low shard]

check https://rentry.org/RVC_making-models#lossgmel-lossgkl-lossgtotal-lossdtotal maybe

[low shard]

check https://docs.aihub.wtf/extra/model-maker-role/

[leaden yacht]

just to confirm, does that mean the pitch is getting better (its going down)?

[fading summit]

i got it

[leaden yacht]

Yeah I saw u submitted it

niceee

I think you need to wait now

[fading summit]

yep

[low shard]

yep

[leaden yacht]

is this overtraining?

[fading summit]

@leaden yachtive never payed attention to those things lol

[leaden yacht]

good to know

[leaden yacht]

the graphs?

[fading summit]

@leaden yacht yeah

[leaden yacht]

You didn't have to pay attention I think, I mean weight.gg is probably automated

[fading summit]

this whole ai thing is new to me

[leaden yacht]

Im doing it manually on my own laptop

[leaden yacht]

same