programming

[nocturne olive]

Machine learning but you maximize loss instead of minimize

[solid bough]

so I'm back to square 1

[solid bough]

That was with a Model.h5 that I ran train.py 2 times already, now I have a different approach

[nocturne olive]

Maximally incorrect model

[solid bough]

At least the model had some sort of edges

[nocturne olive]

Now make a vocal synth

[solid bough]

The only drive that I have for that model is that I wanna make a better model than the stock one

[nocturne olive]

How sily

Then make the best vocal separator in the world

Make one that supports all vocal stems

Like can figure out harmony and stuff

[solid bough]

I really have no drive to make something like that ngl

Like, notthing

[nocturne olive]

Well I would find that useful

[solid bough]

And with the fact that I have no idea what I'm doing rn doesn't make it better

[nocturne olive]

Imagine a vocal separator model specifically tuned for Neuro's voice using NeuroSynth output as a guide, that would be so funny

[solid bough]

Well, it's round 3 of 3 of my training now, now I look how it goes

[solid bough]

This run is just what I get on one line:

imma change back to relu

First I cycle through 3 parts and as a last time I do all of the 3 parts at once on training

Okay that is now better with a lower learning rate:

Except the train and test RMSE being flipped

[obsidian mantle]

10^?

isnt it supposed to be between 0 and 1

surely its 10^-2 or smth

[sage crag]

would only have to be -1

[obsidian mantle]

programs usually dont show -1's like that

do they

[sage crag]

its matplotlib

"programs dont usually do xyz" not a good reasoning for discarding it as a possibility

https://tenor.com/view/mini-miniko-minikomew-gif-17857450978747562560

[obsidian mantle]

old graphs were like this

[olive sable]

i am making a bass

[obsidian mantle]

fish bass?

[olive sable]

what?

[nocturne olive]

As in low frequency sound?

[obsidian mantle]

the bass

[olive sable]

as in the instrument

[obsidian mantle]

https://tenor.com/view/gemmy-singing-billy-bass-singing-fish-funny-fish-gif-7885482086969033798

[olive sable]

i forgot fish bass existed ngl

[obsidian mantle]

bassed

[nocturne olive]

I don't even know what that fish is in Finnish

[obsidian mantle]

same

[sage crag]

bass

[obsidian mantle]

ok i googled its a pretty common fish

[sage crag]

fish

[olive sable]

"baars" in dutch

[sage crag]

bass

instrument

bass

quality of audio

[olive sable]

i remember catching so fucking many "zeebaars" in acnh. fuck those

i hate zeebaar

[nocturne olive]

Which?

[sage crag]

bass

[obsidian mantle]

idk finnish i googled my language

[solid bough]

I guess

[solid bough]

It's supposed to be Log scale but idk if it is.

did sonic make the Graph?

[obsidian mantle]

no i mean

this gets cropped

its 10^x

[solid bough]

Oh yeah, that is something I have no idea how to fix

[obsidian mantle]

ooh its actually end of window

i see close button on top

what if you save it

sometimes they save jpg

[solid bough]

and btw, while I was not chatting in here someone might have gotten a DM from someone that for some reason wants my steam account

[solid bough]

Ahem, the score.png:

Still filtered cropped

[obsidian mantle]

does this have anything interesting

[solid bough]

Idk, the window is mostly frozen until the epoch is done

[obsidian mantle]

oh so its work in progress

[solid bough]

More like: I can't be filtered to filtered fix that it doesn't freeze

[obsidian mantle]

what word are you filtering

i cant get it after you used it 500 times

[solid bough]

Swear words

[obsidian mantle]

[solid bough]

And me with no filter (2 layers) is heck

The initial filter filters any BS I'm thinking, and the second one is swear words/blocked words

But back to the scammer, that is a rough outline what teh convo was: me: waving sticker person: Hello, sorry to bother you, but can I ask you something? me: You're gonna say that You reported me to Steam Support and I'm needed to friend a "Steam Support Staff"? person: great

Or what I assumed to try to get my account of the initial message

Great thing is that the user isn't in here

The swarm would take that user apart

[obsidian mantle]

I heard discord had some hack issue and some accounts got stolen do i change my password

[solid bough]

uh, You mean where the same Hello, sorry to bother you, but can I ask you something? is said?

Because I'm now oddly familiar with those things and I'm bored that the other user didn't reply past the "great"

[obsidian mantle]

Noo i mean if there was a hack then any account might potentially be compromised.
What if someone scams people from my account while i sleep

Or is it false alarm

[solid bough]

That is messed up

I personally would change the password (preferably with a password manager)

[obsidian mantle]

You dont need 2 factor auth if you got that login hash or how its called

From discord appdata folder

[solid bough]

The... Uh the same thing like that is used for Discord Bots

[obsidian mantle]

Api key?

[solid bough]

The token if I'm right

[obsidian mantle]

Auth token right

[solid bough]

Yeeee, Token.

If I did some token shenanigans, I could log in to my Bot

But I'm not feeling like doing that

Well, you mean the email you got of Discord?

uuuuh vituha?

Hot take: I should do learning rate 8

[obsidian mantle]

idk if i got any emaails

i just heard the word

[solid bough]

oh

[solid bough]

I asked chat and now I also have to filter on the Background color

[solid bough]

I'm just gonna let code roughly sift through the images

[trim valve]

please tell me chat does not refer to chatgpt

[solid bough]

[trim valve]

I do not want to believe that people do that

[solid bough]

... no?

[trim valve]

thank god

[solid bough]

At least not my Project (Text Converter)

[solid bough]

Btw, have you not Seen the sweating pepe and how I wrote the message after that?

[trim valve]

idk text sucks for having any nuance

[solid bough]

Well, on the Deep Doodle code where I'm mostly You might know the answer what I did with the Deep Doodle code...

I'm sorry that I disappointed you bred

[trim valve]

smh my head

[solid bough]

The Deep Doodle code is unknown territory for me...

[noble zodiac]

give it a couple of months and we will be outnumbered in here

[trim valve]

#vibecoding

[solid bough]

I am doing that with a base code that is Python 2

So I didn't pull the code out of thin air

[tender river]

emp

[solid bough]

What?

[tender river]

ye emp

[solid bough]

Electro Magnetic Pulse?

[tender river]

[solid bough]

You can't destroy my PC rn

Hi caibi

[midnight sigil]

ello

[narrow abyss]

computer explodes due to unforseen consequences

[solid bough]

I'm just going back to python 3.6 after unsuccessfully going to Python 3.13

[midnight sigil]

triangle

[solid bough]

The side panel is not even installed on my PC

[narrow abyss]

what if the girl is triangle

[solid bough]

200 OK
Out:
▲

[midnight sigil]

💌

💒

[solid bough]

Since when is that NN?

[midnight sigil]

[solid bough]

Wait, are you some times in NN?

[midnight sigil]

Sam moment

[midnight sigil]

maybe

rarely

very rare

[solid bough]

sgtm

[midnight sigil]

I have 40000+ messages in this server, but <100 of them are sent in NN

so don't worry I'm not polluted

[solid bough]

And meanwhile I can keep my English clean. If I say something that is not standard English it is a different version of English that I have for here.

[midnight sigil]

[stray dragon]

ery

[solid bough]

classic t moment

[true hemlock]

no fuckin way the sacred triangle

[solid bough]

Hi duck

[sick owl]

You know, I kinda regret getting a 4k monitor

Because now I can see how shit youtubes bitrate is

[solid bough]

That means I sho... Have you never heard of NVIDIA RTX Video Enhancement?

but back to that, I should get a 4K 60Hz monitor ig

[sage crag]

https://cdn.discordapp.com/attachments/1126496895246807123/1427626771603390555/yippee-1.gif

quuck 🔺

[solid bough]

Hi konii

[sage crag]

https://tenor.com/view/jeong43va-gif-26093523

[solid bough]

Hmm?

[tender river]

me do this

later

[rigid snow]

now pls

[obsidian mantle]

i fixed my bot making shit moves

i made it not stop on quiescence at all so it never evaluates untraded board

[solid bough]

Nooo, not Germany, please

[obsidian mantle]

and banned it from stopping on extentions

[solid bough]

Yay

Emojis: 👍✅❌

[obsidian mantle]

i couldnt quite settle it in my brain that quiescence search is not really too long

since its limited to captures

[solid bough]

That was straight the sequential emojis of what I have mostly used:

[trim valve]

smh why do I have to do job applications

they should just let me walk in and pay me

[solid bough]

Brb, I'm asking if food is made and if not I'm making myself a Pizza

[tender river]

i just remembered

you

chatter

qwertz

[solid bough]

Yes, I use qwertz

Hiii chayleaf

[tender river]

[solid bough]

[tender river]

qwertz bad

[solid bough]

I disagree

and I forgot to go afk bc of dinner

[tender river]

disagreement not accepted

[opaque sigil]

qwertz bad

[rigid snow]

best layout fr

[maiden geyser]

literally the third worst there is

with the worst one being \ near backspace and the second worst being third worst + language switch replacing left Win

[civic flax]

I have a question

This isn’t about programming but the process feels like it.

I have a PNGTube+ model and I sometimes gets stuck trying to figure out what to do. Is it fine if I can ask questions here in the future related to it or not because it’s not programming.

Unrelated to that, I have another question.

Remember the laptop you guys help me find out about a year ago? For some reason, when I boot it up, I see the normal Victus logo and the loading wheel, and all of a sudden, the screen goes black. The computer is on, it’s just that everything is black, the screen is on, and the fans are doing its thing.

[dapper kernel]

[obsidian mantle]

nvm its still shit. i guess it was lucky move order or smth

[solid bough]

Did I stutter?

I sure hope you don't say that I This content goes outside the ToS of the service.

[rigid snow]

chay destroy tos of the service

say now

[obsidian mantle]

nothing helps

i remove features - bot becomes to slow and loses
i leave features on - bot makes blunders and loses

idk if its sorting problem or i have literal error/bug like wrong sign

if i got it correctly - all these additional prunes like late move reduction, internal iterative reduction and some others - they all depend on sorting

if your sorting is bad then it will prune good moves since they were wrongly sorted

but how do i sort it if i dont see the future

i can slightly adjust time - and new version will win old from both sides.
while making blunders

[solid bough]

That was odd, I was interrupted because of ToS

[solid bough]

Oops, That goes against the ToS

[ashen warren]

no do it vertically

[solid bough]

I have no capslock

[rigid snow]

[ashen warren]

soz

I saw a cursed video the other day that a lot of younger folk, because of phones, type in capitals never using shift and just hitting caps for every character

[rigid snow]

you're good

[solid bough]

I told you I have no capslock

[obsidian mantle]

@midnight sigil when you said your pruning "heavily depends on sorting", did that mean it will literally blunder and prune good moves if sorting is bad, or it only affects speed?

[rigid snow]

[ashen warren]

bro got anti shift

[tender river]

ge t pinged nerd

[stray dragon]

noooo

[rigid snow]

[solid bough]

That is the SpeedShift

not a capslock

[ashen warren]

shiftn't

[stray dragon]

chayleaf got the two in one combo

[solid bough]

Wait, it's Easy-Shift[+]

[solid bough]

I have then double the Macro keys

[rigid snow]

what the fuck is an Easy-Shift[+]

[ashen warren]

i chose not to ask

[solid bough]

I have only 5 Macro keys without the Easy-Shift[+] and with Easy-Shift[+] It's like I have 10 Macro Keys

So I have 5 Physical Macro keys but with that there is like as if I have 10

[ashen warren]

literally what

[solid bough]

Aka:

[solid bough]

If I press the the so called "shift" key, I can do the other 5 macros

[keen hatch]

[solid bough]

Why sad?

The drive is still there

[ashen warren]

that's crazy storage, what are you even storing there with such huge read write to start wearing down the drive?

gam?

[solid bough]

The drive is fine

or all drives better said

[sick owl]

I'm not big on it tbh, just feels like standard AI video upscaling

Its no DLSS we'll put it that way

[solid bough]

Oh

[keen hatch]

I actually think it's a bad connection.. I took the drives out of the NAS yesterday to clean them of dust.

[sick owl]

Tbf its been ages since I used it

[sick owl]

Maybe I ought to give it another try

[solid bough]

The drive is fine

[rigid snow]

oh i thought that was smart

[twilit lava]

I have no capslock and I must scream

[solid bough]

I CAN SCREAM PERFECTLY FINE THANKYOUVERYMUCH

[rigid snow]

i should rebind capslock to something more useful

[solid bough]

Sorry vopo

[solid bough]

I don't use Capslock in any situation ngl

. . . vopo?

[nocturne olive]

Anyone wanna make a Neuro TTS? Today's stream has a ton of free clean data

[solid bough]

Currently I'm not doing a tts model...

[nocturne olive]

Just letting anyone that does know

[solid bough]

Well, I certainly can somehow download the VOD after the stream

[rigid snow]

no music

[nocturne olive]

I'm saving it on my own system with automation

[rigid snow]

but also data easy to generate anyway

[nocturne olive]

True

If you know how to Azure

[ashen warren]

why use azure?

[nocturne olive]

Because that's what Neuro's TTS is?

[solid bough]

Azure too expensive

[ashen warren]

better options these days

[rigid snow]

how is this relevant

[nocturne olive]

Well stream data is free

[rigid snow]

they gave you shit ton of credit back in the day

for free

idk if still the case

[keen hatch]

smart says the drive is fine... but i need to run a full test. or maybe better to replace the drive, it's over 6 years old

[solid bough]

True, How long is the vod gonna be there?

[nocturne olive]

No need for that when I'm saving the direct stream on my own system

I can just pull the audio data from that and send you later if you'd like

Oh now there's Evil too

[solid bough]

Oh, I was thinking worst case

[nocturne olive]

Some Evil for those that want it I guess
Neuro trainers will want to remove the Evil parts though

Oh great the whole stream will probably be free data

[ashen warren]

I'd love a more technical breakdown of what went wrong with neuro the other day. most detailed output from vedal was 'bug in the inferance layer' but that doesn't really map to Location Location Location

I hypothesized overfitting of her model

but idk if that makes any sense from what vedal says

[keen hatch]

I also think it overfitting... or there is some issue with stop tokens.

[ashen warren]

yeah, maybe we'll get some more detail some day, or not, i'm tempted to just try and purposely overfit my model just to see if that happens

[keen hatch]

overfitting example:

["<bos><bos><start_of_turn>user\nConvert this text using semantic tags: 'Hello!' she said cheerfully.<end_of_turn>\n.\n<｜NARRATOR｜>I was sitting with another conversation who I was not immediately present in the conversation. I was not the original voice, but I was still not a synthesized voice. I was simply a synthesized voice that could be heard.\n<｜
NARRATOR｜>I used to express myself as I spoke, and I intended for speech as a person. I was not a person in the ordinary sense of being. I was a thinking being. I was a thinking being. I had a self-description. I was a person with a distinct identity. I did not have a person, but I was an object. I was made of metal and of wood."]

[ashen warren]

this has been the best bug from ASR i've had (totally unrelated)

i left it running over night with a fan pointed at the mic by accident and the next day the bot was obsessed

luckily just in RAG memory, as next day was a fitting day so just reverted

PoV 20 euro bill API cost for LLM

[keen hatch]

that's why local LLMs are better )

[ashen warren]

yeahh i need a better GPU for that i'm running a 3070 and the memory limits mean i have to live with quant models

best i got was out of gemma3 but even at 7B i was hitting memory constraints

I might just be an idiot and missing a nice memory trick i can use tho

[keen hatch]

It's always like this... no matter how much VRAM there is... it's always not enough

[keen hatch]

small MoE models are also an option

[solid bough]

mrtime

I have to say something controversial

If I don't, It's too normal in here then

I have to make a model that overfits so hard, no one else can or wants to beat it

That is what I was talking about

[keen hatch]

[solid bough]

Like really, Why not?

[keen hatch]

this is a good way to get this kind of picture.

[solid bough]

Huh?

okay

[keen hatch]

https://en.wikipedia.org/wiki/Double_descent

[solid bough]

Idk what that means ngl

[opaque sigil]

Basically there's a point where your model is so large it can perfectly memorise the training data including the noise and adding parameters past that point means it can interpolate between the noisy training data points better

[obsidian mantle]

i think i found what i was doing in my first attempt to make the bot
https://www.chessprogramming.org/SOMA

[solid bough]

Oh

got it now

[obsidian mantle]

nah not exactly

im wrong

[glass flower]

there is a whole chessprogramming wiki?

[obsidian mantle]

yess

many info there

[olive sable]

ello

[solid bough]

Hi Sam

[olive sable]

Hi chaos of sage

[solid bough]

How's going?

[olive sable]

good

i took a anp

[solid bough]

You took a anp?

[olive sable]

nap

[solid bough]

oh

[opaque sigil]

9am nap

[olive sable]

from 4 to 8pm or so

[solid bough]

Who pinged?

[rigid snow]

schedule maybe

[solid bough]

[olive sable]

most likely schedule ye

[solid bough]

Where do they know my work schedule?

[olive sable]

the neuro schedule

[solid bough]

Oh

[olive sable]

https://ksawerykomputery.com/works/we-are-one
art installation based on ai or something idk

i have recieved a 20% discount on this for some reason

they have all their tech blueprints on the website which i find pretty neat

its javascript based

[nocturne olive]

Eww JavaScript

[olive sable]

node.js

[nocturne olive]

If you want a cheap-ish inference GPU, you can get 3090s for quite reasonable prices

24GB

[solid bough]

Damn, did I hear 24GB?

[nocturne olive]

Yes 24GB

My system

[solid bough]

But with that new 12VHP connector

Thanks but I pass on the Cheese connector

[olive sable]

3090's dont have 12vhp

or well, a lot dont

[solid bough]

Well, last time I checked they do

[olive sable]

i have one

RTX 3090 Ti is the model that uses the 12VHPWR connector. Standard RTX 3090 models use traditional three or more 8-pin

"three" i have 2

[solid bough]

isn't the GPU using that mess of a plug?

[olive sable]

3090ti used 12vhp

and then the 40 series used it

and melted

3090's are fine

i think the founders editions used something like 12vhp but no sense pins?

[solid bough]

I still don't trust that Dodgy connector

Yes, I stand my ground, it is dodgy

[olive sable]

then get a 3090 without 12vhp. 99% of them dont

mine doesnt, the one superbox has doesnt

[solid bough]

If the Drivers License stuff would be in the way because of cost

[olive sable]

???

drivers license???

[solid bough]

I do the Drivers License rn

and that isn't cheap

[olive sable]

ah isee

[solid bough]

At least in total 1K

[olive sable]

The Founders Edition of RTX 3090, 3080, and 3070 use a 12-pin connector, not the 12VHPWR.

so just avoid those ones

[opaque sigil]

it's the same thing but worse cause no sense pin

[solid bough]

Wait are 30 series still getting produced?

[opaque sigil]

it just doesn't matter with those cards cause they don't draw nearly as much

no

neither are 40 series

[olive sable]

get them 2ndhand

only 500 bucks if you know where to look

[solid bough]

"Only 500 bucks" You say that like I make more than 1K a month

[olive sable]

nah, id consider them better cuz at least they had even distrobution of power along all 12 of the wires, the 4090 didnt have that for some reason

[solid bough]

500 bucks is more than I make a month on average

[olive sable]

and then you get 1 wire psuhing 40 amps and burning

[olive sable]

then dont get a 3090

[solid bough]

But I need more VRAM

[olive sable]

then get a card with more vram

you either buy one or not, im not your financial advisor

[solid bough]

Well, first I need to somehow make that my father and mom say okay

Yes, I'm 24 yet I have to convince my mom and dad that I need that card.

Now done with the message mb

[olive sable]

founders editions so has 12 pin, but still a good deal

i aslso see an asus TUF 3090 oc for 500 bucks

the rest are all 700-800

[solid bough]

Okay, I would need to ask my dad and mom though

And I'm certain they say no

[obsidian mantle]

they are making PSUs with this plug now

so its just monolith wire

[solid bough]

I appreciate your dedication though Sam

[olive sable]

no worries

i do this as a side hustle lmoa

[olive sable]

as long as the sensing wires work, and the power is distrobuted evenly through all the wires, there should be no issue

[nocturne olive]

If I had the money I would get that one
But I've been broke ever since the previous one

2x FE is probably what I'll put in my second dedicated training rig

[olive sable]

the 30 series fucked up one the 1st point but did the 2nd point, then the 40 series did the 2nd point but fucked up the 1st point. and idk what the 50 series does

[nocturne olive]

Unless I can get 4090Ds with the 48GB mod

But those are like 2400 + shipping from China

[solid bough]

There are now Double D GPU's too

[olive sable]

superbox needs a gofundme fr fr

[solid bough]

it was talked about in the 3H GN GPU movie

[nocturne olive]

I have no clue if anyone would actually give me money for NSP
Nobody's giving me any crypto at least

[clear sedge]

i don't have crypto

[olive sable]

i odnt have crypto either

[solid bough]

apparently the RTX 5090 DD

[clear sedge]

i do have a paypal though

i'm willing to give you one crisp dollar

[nocturne olive]

I've been thinking I should set up a Ko-Fi thingy for NSP
Just new payment thingy scary

[clear sedge]

ko-fi isn't hard to set up

[nocturne olive]

I'm just way too worried about leaking my real identity
Though I heard using Stripe instead of PayPal should work

[clear sedge]

you can use a paypal business account

that's what i do

[solid bough]

I have PayPal

[nocturne olive]

I don't trust PayPal, seen way too many people have their identity leaked without them knowing

[clear sedge]

fair

[nocturne olive]

I just need to figure out a moment when I can set it up and get over the scary factor

[clear sedge]

the fact it shows your legal name for individuals is crazy

[nocturne olive]

True

Very stupid

[clear sedge]

"hm yes, let's force people on the internet to expose their legal names"

[nocturne olive]

Yeah that's not gonna go well

[clear sedge]

business accounts are superior because it just shows your "business name"

it still adds your billing address to transactions by default though so you gotta uncheck that every time

[nocturne olive]

I hope Stripe is safe, because nobody wants to give me any money and I suspect it has to do with crypto being difficult and kinda sketchy

[solid bough]

But I can't do anything with my PayPal

[nocturne olive]

But either way, that'll be something to worry about eventually (tm)

By the way anyone want the no BGM stream audio track?

[clear sedge]

oh my god stripe is half irish

as a certified alcoholic™, i'm now officially required to make a stripe

i'll do it later

[solid bough]

How are you "required"?

[clear sedge]

https://en.wikipedia.org/wiki/Stereotypes_of_Irish_people
https://www.cambridge.org/core/journals/medical-history/article/great-race-of-drinkers-irish-interpretations-of-alcoholism-and-drinking-stereotypes-19451975/23DB14BF85F151C706824FD36D9396B3

[solid bough]

Not as far I know, I know them as the most sober people...

[nocturne olive]

The stream turned from free Neuro data to free Neuro and Evil data, which is quite silly

[solid bough]

True

But I wasn't with the stream

[nocturne olive]

I'm getting the silly audio track for those that like training data

[solid bough]

I watch the time when livE had a stroke

[nocturne olive]

Will be two parts

[clear sedge]

inaccurate stereotypes?

[solid bough]

Idk

[clear sedge]

i've never met an ire

[solid bough]

I just thought they were the MOST sober people

[clear sedge]

anyone in britain is automatically debuffed by the british stat

[nocturne olive]

There, if anyone wants to try out training on it

[solid bough]

What?

[solid bough]

Well, Is audio easier?

[nocturne olive]

Eh, not sure
But I do audio ML a bunch

That's what NeuroSynthProject is all about

[clear sedge]

listen if i keep explaining i'm gonna get muted for rule 2 again

[nocturne olive]

NSP greatness

[clear sedge]

[solid bough]

Okay, First I make models that make blurs, and i assume that the Audio model is gonna be like the Face model

[solid bough]

ok

[clear sedge]

this sounds so much like a toby fox song

[nocturne olive]

Well anyway have fun with some basically free Neuro and Evil speech data
I'll go to sleep in a bit

[solid bough]

That's good

[nocturne olive]

Well the song is by Toby fox

[clear sedge]

i fucking knew it

i'm a genius

[nocturne olive]

Deltarune CH3

[clear sedge]

i'm gonna take my trophy now

[solid bough]

I'm gonna demo the latest and filtered model with the images filtered by code

[nocturne olive]

NeuroSynth is quite powerful already in the BETA stage
I've been working on the non-BETA JP data recently

Here's more silliness

[clear sedge]

i wanna make her sing the shrek song

[solid bough]

What the filtered filtered filtered

it doesn't even make a frickin edge

[clear sedge]

that does look pretty filtered

[solid bough]

There is supposed to be a circle like the B/W image

[nocturne olive]

Well if it's English or soon Japanese NeuroSynth could be capable of it

[clear sedge]

somebody once told me

[nocturne olive]

-# If you want you can pay me to do it

[clear sedge]

i gladly would but i'm poor as shit rn

[nocturne olive]

Well I'm also poor, so that's not anything crazy to me

[solid bough]

The test and train RMSE

[nocturne olive]

Can barely make these vocal synthesizers with the resources I have

[solid bough]

I don't wanna filter through images AGAIN

[nocturne olive]

The model must have learned that the best way to fit this data accurately at least at earlier training stages would be to just draw nothing

[solid bough]

Yeah

Lemme just run the doodler.py code mids training

idc anymore

I just want my GPU to suffer now

[clear sedge]

if you need extra compute power i've got an rtx 2080 ti

i mostly just program all day so it sits there mostly unused

[nocturne olive]

I have a 3090

But the VRAM is what I need

[clear sedge]

simple fix

[nocturne olive]

I need more batch size
I need more of it

[clear sedge]

enable zram and use integrated graphics with 256gb of swap

[nocturne olive]

No that will take literal millenia to train

[solid bough]

It's super subtle but I can see SOMETHING

[nocturne olive]

How weird

[clear sedge]

it looks like someone stained an olive on some gray paper

[solid bough]

How weird, I should have frickin continued with the manually filtered images

[solid bough]

Yeah

[nocturne olive]

Either way, I'm just hoping maybe if I set up a better more accessible payment method someone will actually fund NSP a bit
But anyway, now I need sleep

[solid bough]

But It's doing a nose dive now:

[clear sedge]

it looks like when you rub your eyes slightly too hard and look at something too colorful

[solid bough]

Let's run the doodler.py code again because I toootally love my RTX 3070 Ti

[clear sedge]

https://tenor.com/view/nighty-night-baby-goodnight-ollie-tucking-into-bed-ollie-gif-22641848

[solid bough]

Ye

Still the same "Stain" stuff

Wait a second

OMG

I still have the Manually sorted images folder

[olive sable]

bass

[solid bough]

Damn, That's a good guitar

[solid bough]

What made you make a 3D model of a Guitar?

[olive sable]

litle indie game we're workig on

[solid bough]

[solid bough]

@clear sedge Am I doing too much?

I was trying something new

[clear sedge]

it looks like a nautilus shell

[solid bough]

I was thinking: Hmmm, Why not draw a face as if a swirl effect was applied?

Meanwhile:

[solid bough]

It's going well so far, How about your AI Vedal?

I bet you'rs is going WAY better (Neuro and Evil Neuro)

[olive sable]

[solid bough]

I'm just winging that ngl

[weak robin]

what am i looking at

[solid bough]

There is a non 0% percent that Vedal answers

[solid bough]

The train RMSE and test RMSE

[weak robin]

whats RMSE

[olive sable]

vedal appearing in #programming happens once in a blue moon

[weak robin]

reinforcement memory something E?

[solid bough]

idk

[weak robin]

oh xD

[solid bough]

All I know is Epoch 126 of 500 VRAM: 25.2 MB / 5789.1 MB | (2096, 5, 192, 144) float16 (2096, 3, 192, 144) float16 105/105 [==============================] - 8s 73ms/step - loss: 0.0369 Train RMSE: 0.1907518840061805 Test RMSE: 0.19188857129963205 Model saved.

[weak robin]

is that an ai?

[cosmic sphinx]

last time shiro had to ping bro

[solid bough]

More a Black/White sketch to realistic face

[olive sable]

dont get any ideas

[solid bough]

That is still not 0%

[solid bough]

I'm not pinging vedal.

[cosmic sphinx]

uh so how's that related to neuro & evil's progress

[olive sable]

in the entirety of 2025 so far, vedal has 10 messages in #programming

[solid bough]

Well, my training is not going so great, so that is why I bet that Vedal is doing better on training the model

[olive sable]

thats 1 message per month

on avarage

[solid bough]

Yeah, but still not 0%

[cosmic sphinx]

almost as much as the amount of dev streams he's done in 2025 LULW

[weak robin]

Yeah nah i would try to think right now but my brain so tired from coding In C 🙏

[stark needle]

The new DGX Spark is a fucking joke

[solid bough]

I just wanna filtered sleeep

[stark needle]

on gpt-oss-20b

[solid bough]

With a loss of 0.0361

[stark needle]

Shitty ass 319t/s generation throughput on llama cpp

My 2 3090 rig nukes those results

The dgx spark also only has 273gb/s while my rig has an aggregate 1.8TB/s or something

[solid bough]

UnemplyedLilac do you think it is a good idea to run the model while it is training?

[cosmic sphinx]

are you saving the weights every epoch or what

[solid bough]

Maybe

[cosmic sphinx]

it shouldnt affect the training, but it might affect the performance

quite a bit

[sick owl]

So I gave RTX Video Enhancement another fair shot

Ngl, it made the problem worse

[solid bough]

It's fine, the GPU is gonna cope

[solid bough]

Wuh... okay

[sick owl]

It RTX Video Enhanced™ youtube compression artifacts to be more prominent

[keen hatch]

I wonder how much it can be used for training models.

[sick owl]

Only talking about the super resolution bit though, I don't mind SDR content so I didn't enable the HDR component

[cosmic sphinx]

"your system is no longer receiving securty updates"
fuck you microsoft

[sick owl]

Linux time

[cosmic sphinx]

i dont think i can realistically cope with linux yet

[stark needle]

It's not gonna be that good

[solid bough]

What base video resolution is the video set to?

[stark needle]

I saw benchmarks

Also it does not support flash attention which means it will resort to the much worse attention

[sick owl]

1080p

[solid bough]

And that is with the GPU also doing the Training of the model

[solid bough]

Nope

[sick owl]

Honestly I'd have said the same but bazzite is pretty good

[solid bough]

That is the res of my monitor

[sick owl]

Its my go to on my laptop now if perhaps not my desktop due to Windows's better colour management and icue not supporting Linux

[solid bough]

There is a more obvious shot:

[sick owl]

Yeah I was going 1080p -> 4k

[sick owl]

That's not an improvement imo

[solid bough]

I was sending a more obvious image that you can more accurately guess what the video res is that I have set in settings

[sick owl]

Ohh right

That'll be like 240p right

[stark needle]

For 4000$ u can get a lot of nvidia H200 hours

[sick owl]

This looks like 480p or maybe 720

[solid bough]

Okay that would not be possible

[solid bough]

Lower actually

[keen hatch]

usually, time isn't such an issue with fine-tuning, but VRAM is...

[sick owl]

Not bad, but tbh I'm still not sure its an improvement when the AI sheen is so obvious

It looks like AI upscaled texture packs for games

[solid bough]

Should I just reveal?

[stark needle]

But then not even unsloth is supported on this thing

And the fact that flash attn is not supported means the memory scales quadratically vs linearly

Which is a huge negative

[solid bough]

It was at 360P

[midnight sigil]

first one

[rigid snow]

[keen hatch]

https://github.com/unslothai/notebooks/blob/main/nb/gpt_oss_(20B)_Reinforcement_Learning_2048_Game_DGX_Spark.ipynb

[stark needle]

Oh shit it does work

[sick owl]

Not bad, but I'd still probs take the 360p

[solid bough]

360P is the limit

[sick owl]

Actually nah

I think its an improvement on 360p

[solid bough]

Anything lower than 360P is not getting enhanced

[sick owl]

But at 1080 -> 4k it makes the problems worse rather than better

[solid bough]

Lucky, I'm stuck at 1080P

The monitors

I used to have a 1440P monitor

Gn8 M8

[keen hatch]

https://docs.unsloth.ai/new/fine-tuning-llms-with-nvidia-dgx-spark-and-unsloth

video https://x.com/UnslothAI/status/1976284209842118714

Unsloth AI (@UnslothAI)

OpenAI shows how gpt-oss can autonomously beat 2048 using reinforcement learning (RL).

Training was done locally with Unsloth on NVIDIA DGX Spark. You can also do it free on Colab. 🦥

OpenAI DevDay notebook: https://t.co/ptsilrBosy

X · Oct 9, 2025, 1:50 PM

[amber fractal]

tbf 2048 has a known stratagy

It is called snake and gamba that a piece doesn't spawn in the middle if you need to unstuck stuff

[rigid snow]

i decided to look at gemini-cli source code to see how they use ink and make it actually work and not be shit and oh my fucking god

https://github.com/google-gemini/gemini-cli/blob/main/packages/cli/src/ui/AppContainer.tsx this can't be real

most elegant react state management solution

[tender river]

[sage crag]

wrrr

hblang update

wasm ✅

[olive sable]

bass

i havent textured it yet, i think ill wait till iggly has a more clear idea for the design of the characters

tested out "3 point lighting" too, i learned that in school

oh ye

@amber fractal bass

[amber fractal]

bwaass

[mighty thorn]

i think i have figured it out

[dusky jackal]

Okay guys! I'm trying to download DaVinci Revolve 20 but it's asking me if my windows is x84 or if it's an arm. I want to becomed an clipper!! :DDD I followed a tutorial to check what it is and it said it's x64 based, does that mean its x84 or arm?? What do I download it on, x84 or is it an arm??? Pls help I'm not good at this. :C Thanks!

[clever viper]

Iirc you can run x86 (32 bit) app on x64 (64 bit) PC (at least mine does)
Arm is different from x64/86 PC, so download the x86 version

[dusky jackal]

Does my computer like explode or some shit if I try to download the x86 thing?? Like I don't know if my computer is made for that, dude.

[clever viper]

Nah, but there's a chance your computer gains sentience and becomes turtlephobes instead

[dusky jackal]

Well shit dude that sounds sick as shit. I'll do it!

[olive sable]

use x86

x64 is just x86 with 64-bit registers

x64 is just short for x86-64 from my understanding

arm would be if it was a tablet or certain chromebooks i guess

basically any pc thats less than 20 years old is x64

[modern gulch]

I honestly understand why Vedal is so protective of neuro and evils data sets

based on the fact that neuro-sama and evils data sets have been refined for so long

The data sets create such a strong personality/identity baked into the llms that make them the way they are

Given enough time and low enough latency between the various parts used to make it all works can in a sense make them pass as humans in most cases except for when something they just aren't trained on comes up
Also saw a poll for a similar question on the Neurosama sub reddit

Thought of this while watching neuro content on YouTube
Vedal was being asked if he had a back up so that the twins wouldn't be lost

Tldr
Even if they're not conscious
Their personality is very real

And that personality is shaped by the community around them and Vedal Constantly

Also vedal is a coding wizard to be able to put all this together

[uneven pulsar]

Truth. Anyone who's worked with a raw LLM knows how frustrating it can be just to create a reasonable response, let alone give it anything close to a consistent sense of "personality" or even stay on topic.

[modern gulch]

Yep exactly

[glad path]

or mac

mseries specifically

[olive sable]

true, but mac has their own button on the installer page

[glad path]

because stuff has to be built seperately for macos

[olive sable]

ye

[glad path]

rosetta 2 is practically bundled with the os (there's a tool to install it and it prompts you after finishing setup) so you can run x86 apps anyways

[olive mango]

hoi

[olive mango]

just use windows bro jk

[olive sable]

hoi

[nocturne olive]

Linux is easier to deal with than Windows 11

[dry charm]

liblastfm-unstable-2019-08-23
       > CMake Error at CMakeLists.txt:1 (cmake_minimum_required):
       >   Compatibility with CMake < 3.5 has been removed from CMake.
       >
       >   Update the VERSION argument <min> value.  Or, use the <min>...<max> syntax

love being on nixos-unstable

[rough rover]

Ur on unstable nixos?

I’m sorry for your loss

[dry charm]

Yes cause I had quite a few packages that were broken on stable

I should have just moved those to unstable :\

I wonder if I can safely switch to stable now

[rough bloom]

should be fine to switch back unless you set your stateVersion to 25.11 already
probably easiest to just wait for 25.11 branch-off though with how close it is, then you never have to downgrade anything

[dry charm]

Saw the warning in the beginning, forgot to read it, deleted the comments, never touched evers ince

oops

[rough bloom]

[dry charm]

oh well

[rough bloom]

ye, leaving it as-is is fine in most cases

you can increase it but you have to read release notes and fix some breakage at minimum

[dry charm]

I wonder, breakage in the system or breakage in config?

[rough bloom]

breakage in system state that's not managed by Nix

stuff like Postgres DB version

you can upgrade, but you can't downgrade
so increasing the Postgres version breaks the ability to boot into an older generation

[dry charm]

Oooh I see

Fortunetly all of these I run in a container

If I can run a service in a container, I opt to

I had my fair share of headsmashing into desk with versions so containers it is

[hard raptor]

We did it boys, we got a "fair enough " out of an analog supremacist

[midnight sigil]

[cobalt sigil]

whar

[obsidian mantle]

Its downloading into buffer first, not straight to the disk

[rigid snow]

ye mega does e2e

the key is always in the link, server never reads that

so for it to decrypt it has to download into memory first

[cobalt sigil]

i tried with my phone and then it says ram not sufficient qwq

wait let me try transfer it

oh yea that solve the issue

[cosmic sphinx]

yeah but theres a lot of mindfuckery and witchcraft with program compability

[nocturne olive]

Eh not really
It's just either get a Linux version or use WINE

App compat is basically trivial these days

[cosmic sphinx]

cubase?

[rigid snow]

yeah just use windows

audio stuff basically just works in wine afaik

but i don't even try

feels better to be on an officially supported platform

[cosmic sphinx]

well, ima just enroll into ESU for another year of security updates and delay my worries for another time

[opaque sigil]

You should look into massgrave so you can avoid it

[tender river]

meow

[opaque sigil]

[nocturne olive]

I just use stuff that supports Linux

LMMS is cool for that

[rigid snow]

ye no, in terms of daws that's only bitwig out of the ones i could possibly switch to

reaper also exists but i don't see myself ever using it

[nocturne olive]

Sily

I've been using LMMS and will keep using it unless I suddenly become rich

[rigid snow]

you have to understand that it's not like kdenlive where for most workflows it's actually capable and nice to use compared to big dawg video editing software, lmms is absolutely atrocious and miserable. for mixing you don't need much you can mix in adobe audition just fine, but to produce in it? i'd rather use video editing software to produce music genuinely

[nocturne olive]

Wha

Well I'm poor so I don't have that many options anyway

I don't really think it's that bad, I'm able to use it for my stuff just fine

[tender river]

you have been suggested an option

[nocturne olive]

And so what? I have no reason to migrate unless I know what I'm migrating to will significantly improve what I can do or I've paid for something
LMMS works fine for me

[silent cloak]

which google will you pick

[cobalt sigil]

donno haha
just check the old Google logo is microG service

[silent cloak]

ahh

[olive sable]

Ok i admit it

Quick reacts are annoying

Can you disable that?

Bought a new mouse

Its the exact same as the old one, but now i have 1 for desktop and 1 for laptop

I dont have to unplug my mouse each time

[nocturne olive]

Crazy looking mouse
My mouse is way cheaper because broke

[midnight sigil]

💔

I'm all in Delta Force and I forgot how to program

[tender river]

@olive sable https://r500.idk.st/

[maiden geyser]

what is there possibly to deal with on windows? you install revi and that's basically all you need, with some random side quests with easily googlable solutions, "it's not a bug, it's a feature" or windows reinstallations

oh, or probably it broke because you went on a certain side quest, so reinstall windows and don't do that

[maiden geyser]

it's like seeing an icon from pre-NT windows in win 11

[pliant palm]

fellow lmms user

[nocturne olive]

Silly

[pliant palm]

or does everyone here just use lmms that would be awkward

[nocturne olive]

I'm the only one I know of the #programming regulars

[pliant palm]

Lmms so peak

[nocturne olive]

Windows is just generally broken
Like that time it started wiping SSDs
Also my Windows 10 install broke after an update

[nocturne olive]

It's cool
Beats being even more broke

[pliant palm]

Yeah exactly, I've heard good things abt FL and its developers, but it's not like one can casually buy it

[nocturne olive]

Yeah professional stuff too pricy

[blissful geyser]

Does anyone have any idea on how to make mpv stream remotely to another client

[nocturne olive]

I mostly use LMMS for mixing myself, trying to learn instrumental but I suck at it

[blissful geyser]

I'm developing a music streaming service

[nocturne olive]

A music streaming service? Can't you just open a socket or WebSocket?

[blissful geyser]

The thing is its on a VPS

Im trying to stream audio from the VPS to my device

[nocturne olive]

Then simply make your device create a socket to the VPS

[blissful geyser]

Hmm

Howd you do that (i sound dumb right now)

[nocturne olive]

You have a client program on your client machine and a server program on your server, and then have the server listen on a port and connect to the port from the client

[pliant palm]

one day I'll learn programming and stop lurking in #programming 🥀

[nocturne olive]

Silly

And surely I can somehow learn instrumental music

[manic jacinth]

Dooo eeet

[blissful geyser]

Oh right

[pliant palm]

I yearn for the code

[nocturne olive]

Right now I just do vocal

[pliant palm]

Ig I'll start today nothing to do anyway

[nocturne olive]

Like so

[manic jacinth]

Start with messing around with javascript in your browser's console (f12 menu, otherwise known as the devtools)

Its already there, no need to install stuff

[pliant palm]

Thanks!

[tender river]

i'd use ffmpeg for that

if you want low latency you'll have to make a custom solution

this is what i use https://github.com/chayleaf/ihatelatency (in udp mode, the tcp mode is quite bad)

[umbral thorn]

me pissed. mad

so insulting microsoft cunt

me very offended

wrrlllrhgg

[dry charm]

upgrade to LTSC

[umbral thorn]

@true hemlock

[uneven pulsar]

my virtual machine can't run win11 so they should trade it in for a high end gamer pc

[maiden geyser]

come back when you get a little, uhmmm, richer

[obsidian mantle]

what is faster:

if (x==3)
if(y==4)
x=y;
or
if(x==3 && y==4)
x=y;

is it equal?

[tender river]

it's a shame linux is hard to use at least win10 will stay perfectly usable for a few more years

[maiden geyser]

shouldn't it get -O'd out during compilation?

[obsidian mantle]

oooh

does it work like that

coool

[tender river]

one of my relatives' pc autoupdated to a version of win10 that doesnt support their cpu and started bsoding randomly

had to install linux because i literally didnt have a choice

[maiden geyser]

what cpu are they using? pentium 3?

[obsidian mantle]

i had a laptop from 2013 that has amd chipset and there were no drivers existing for win10
so i had to use win7

it could run

but it would bsod

[umbral thorn]

wot cpu

same with quack weird cpu?

[tender river]

i uuh dont know its just a retail cpu but it was like core 2 duo or something

old

[umbral thorn]

then not interesting cpu

[tender river]

ye

[maiden geyser]

_ = !(x ^ 3) * !(y ^ 4);
x = y * _ + x * (_ ^ 1);

[obsidian mantle]

but is this faster or gets optimized anyway

your example looks dank

who is _

[maiden geyser]

this is objectively worse unless you're looking for obfuscation

it's branchless though

[rare bramble]

x += int(((x-3)+(y-4)) == 0)

or equivalent to whatever lang you write

[maiden geyser]

and no, it probably won't get optimised

[rare bramble]

though idk if bit operations are faster than normal arithmetic

havent tested that

[maiden geyser]

-O0 vs -O3, yeah

[rigid timber]

have you actually checked why Windows thinks you can't upgrade or are you just staying irrationally angry?

bc looking at various lists of supported CPUs, I still think it's some random small issue that would be easy to fix

[maiden geyser]

the clock drift situation is insane (UTC+7)

[maiden geyser]

probably tpm or secure boot, who cares

[rigid timber]

that was my thinking too

[pliant palm]

do yall recommend looking at the documentation when learning a language

[rigid timber]

where else would you look?

well aside from whatever tutorials/blog posts/videos/whatever you happen to be learning from

docs are a very useful thing

[pliant palm]

ya i mean i could def learn from tutorials/blog posts

[maiden geyser]

as in "official documentation or w3schools" or "docs or no docs"?

[pliant palm]

official documentation

but is w3schools bad?

[maiden geyser]

not really, an example from the top of my head

both should be ok

[pliant palm]

ah k

[obsidian mantle]

ok so i just discovered that common alpha beta without iterations is better than my current super mega 10 depth extra fast 50ways of pruning transpositions variant

great

im missing some critical information

is chessporgramming wiki a troll website that suggests fake strategies that make the bot worse

i can make my current bot perform just like common alpha beta too

but it will take 10x more time per turn because of iterative search

it will reach depth 8 instead of depth4

and still have mirrored game

[golden narwhal]

Someone know why this space is unusable? I am trying to bual boot linux mint and windows 10

[obsidian mantle]

maybe you need to format it

[golden narwhal]

In the video i was looking he didn't format it. Is it possible because i already have 2 format disk?

Is laptop

[obsidian mantle]

is it even possible to make linux and windows run on same physical disk

dont they need different filesystems

and mpt gpt gbr i forgot format what is this even

when you install win10 and it says "its mbr i want gpt"

install after win7 i mean

[golden narwhal]

I just learn that if you separate the physical disk it work

I install linux now

[maiden geyser]

...yeah? as long as it's separate partitions

[queen ginkgo]

I was about to say if you don't know what partitions are you shouldn't be dual booting but then realized you aren't the one asking about it lol

[maiden geyser]

it complains about that when you're installing it from an MBR/GPT installer on a GPT/MBR disk

[opaque sigil]

Keep in mind windows has a habit of messing with the efi partition so be prepared to have to deal with that

[celest tangle]

As a someone that recently is studying system enginnering is it really necessary to learn about ai for almost every field now days? I mostly like entreprise web apps and desktop app development so i was wondering if now days is it really necessary to keep up in the ai hype train/bubble

[opaque sigil]

(it's not that bad just need to be aware for when it overwrites it and you have to fix it)

[maiden geyser]

and even then, you just use DiskGenius to convert them and it's ok

[golden narwhal]

I already prepare with back up important data on other laptop turn off update and the boot system

I can set the link for that video i follow

https://youtu.be/0gSr8YsJtd0?si=j-AtXMRr-4QMs1gM

[glass delta]

Manjaro vs Arch ultimate battle??

[dry charm]

arch

Might as well use another distro that uses Arch repo instead of that mess Manjaro does

they screwed up multiple times and DDOSed the Arch repos

[golden narwhal]

Only think i know for anch is

• you are free to do anything

• you are free to do anything
  And a lot of new linux user choice is but is bad for who is new.

[dry charm]

Arch is a bad choice who is absolutely newbie and wants an OOB windows experience

but those who come with desire to learn and fail, arch is a great learning experience

[maiden geyser]

linux is a bad choice for people who want an oob windows experience

[noble zodiac]

just use windows if you want an windows experience

[glass delta]

Fedora is the newbie friendly right?

[dry charm]

not the friendliest IMO, but I haven't used it personally. I'm only saying things on what I've read and heard

[obsidian mantle]

i just discovered that stockfish's 1320 elo limit isnt checked by default

also it has these cheat eva files

so with these settings my bot wins

so its not that bad after all

ok but 2000 elo i cant win

[silent cloak]

the chat is still chess

thats crazy

[obsidian mantle]

i like how it goes 250 depth in the endgame in less than 1 second

[solid bough]

I'm still training rn

filtered

[nocturne olive]

I have done way too little Chess bot stuff
It's still way too inefficient to get anywhere

[olive sable]

It was on sale on amazon, only 38 bucks

[nocturne olive]

That's crazy

[olive sable]

[nocturne olive]

Well hopefully I won't need a new mouse any time soon
I need them 3090s

[olive sable]

No fps mentioned but i guess that doesnt make sense to do when using it this way

Very cool tho

[clear sedge]

fedora kde is pretty friendly and stable while still being cutting edge

all you need is follow this tutorial if you use nvidia: https://rpmfusion.org/Howto/NVIDIA
and everything else should work out of the box

[obsidian mantle]

oh you cant actually run stockfish without these cheat files

[solid bough]

now I wonder if my Model collapsed AGAIN:

Reeee

[silk spoke]

Let me guess...
It collapsed

[amber fractal]

Ah yes, eye

[obsidian mantle]

it looks like it zoomed into the eye

[nocturne olive]

Maybe just pure lack of data or batch size?

[solid bough]

I am not willing to sift through more data to get White Background images that are clean

[solid bough]

Perhaps...

[silk spoke]

[solid bough]

Like that is the RMSE of train and test:

[nocturne olive]

Well you may be lacking in either data or batch size if you're getting good loss trends but your models end up still looking broken

[solid bough]

it is the square root of the scores ig

[solid bough]

Batch size is 20

[nocturne olive]

Kinda small

How big is the data?

For reference, latest NeuroSynth is trained on like 1.5 GB of data at 128 batch size

[solid bough]

I have only 131 images that are clean (White BG and no nonesense around the head and face visible)

[nocturne olive]

Yeah way too few

[solid bough]

The sorting code was not better

[nocturne olive]

That's like the equivalent of trying to train a vocal synthesizer on 5 minutes of audio data

You need like 1-2 hours or so

[orchid mesa]

hpohhjhhhjjjjhijigi guys im new here is someone here experienced in using tensorflow

[solid bough]

Like is that gonna cut it?

[nocturne olive]

I hear it's kinda bad, if you can use PyTorch
If you're on Windows you also don't get no GPU acceleration

[solid bough]

Not me, I'm kinda crazy because my model always implodes

[nocturne olive]

If that's all of your images then no
You need like 100 folders like that

[orchid mesa]

ah yez thats my issue rn, running my tf and it doesnt recognize my gpu

[nocturne olive]

No GPU on Windows

[solid bough]

I have in total without filtering 202.599 images

[nocturne olive]

Try to get like 2000

[solid bough]

I have 2 different GPU's of Nvidia:
1: Asus NVIDIA GeForce RTX 3070 Ti TUF GAMING OC EDITION
2.NVIDIA Quadro M2000

[solid bough]

unfiltered or filtered?

[nocturne olive]

Fully usable

[solid bough]

Because if I have to sort for like that one manually I'm crazy then

for 2K

[nocturne olive]

Your train set should be like 2K images
And eval like another 50

[solid bough]

There is no eval folder

[nocturne olive]

At least assuming a small-ish model

[solid bough]

And that is the celebA dataset

[nocturne olive]

Then is your eval just a random chunk from the train?

[solid bough]

so there is TONS of images that are not good for using

[solid bough]

The parameters of the train.py:```py

Training data settings

num_epochs = 500
batch_size = 20
valid_ratio = 5
lr = 0.00002```

[nocturne olive]

If you're using part of the train set for eval, at the start allocate like 5% to eval and keep it separated from the train at all times

[nocturne olive]

Tells me nothing about the data split

[solid bough]

Every time I look at the code I'm just filtered

[nocturne olive]

You do always have that eval graph in your losses

It must come from somewhere

The eval should always come from data that has never been shown to the model

[solid bough]

The training data is getting split like:```py

Split data

split_ix = int(num_samples / valid_ratio)
x_test = x_train[:split_ix]
y_test = y_train[:split_ix]
x_train = x_train[split_ix:]
y_train = y_train[split_ix:]
log_info("Channels attached and split.", logg)```

[obsidian mantle]

what happens if you give it unfiltered 200k photos

[solid bough]

200K photos is not gonna fit in the VRAM despite the puny size

But I can try 4K

[nocturne olive]

That's a weird way to split

[nocturne olive]

You should not send all the images into VRAM at once

[solid bough]

That is not what I did, it is what the maker (hackerpoet aka CodeParade) did

[obsidian mantle]

can it just train more on the same data

[nocturne olive]

It's already told to do 500 epochs it seems

[nocturne olive]

Well currently it's effectively splitting the data in half
You should change the ratio to like 20 so 1/20th of the data goes to eval

[obsidian mantle]

does it have to be all in ram at once

[nocturne olive]

Well that should only happen with bad code or on really big server farms

[solid bough]

It's something I set

[nocturne olive]

Assumed so

[solid bough]

I based it of the Deep Doodle code of the GitHub

[nocturne olive]

Oh wait I read wrong
But still, set valid_ratio to like 20

[solid bough]

Originally it was set to 25

[obsidian mantle]

do you have any data on what does what

like epochs increase blur

[nocturne olive]

Then put it back
That controls how much of the data is used for train and how much eval
Higher value means more for train and less for eval, giving the model more to work with

[obsidian mantle]

that valid_ratio decreased blur

something like that

or you just randomly toggle everything

[nocturne olive]

Eval doesn't have to be a massive portition of the data