programming

[next rover]

Self learning would be self trained

[nocturne olive]

No that is also not the same thing

Self trained implies trained by something made by itself, differnt from self learning which is learning at inference time

[next rover]

Alright I'm sorry. But that's the reason I'm trying to learn

[nocturne olive]

Whatever the case, the Neuro is trained on Twitch chat thing is false information spread by clueless idiots

And no current public architecture can do self-learning

At least for text models

[next rover]

Ok so then what resources do I need

[nocturne olive]

Depends what you want to learn

[next rover]

Well I have Mistral and python

[nocturne olive]

What kinds of models you want to work with, to what extent you want to train them, how from scratch do you want to do them

I work with vocal synthesizer models trained using open source tools

[next rover]

And the goal is to first create a bot that I can interact with locally to test and then put him in my own discord server

[nocturne olive]

Stupid easy, simply grab llama.cpp and any random language and then link the two

As long as you have enough GPUs

[next rover]

Will there be limitations down the line if I wanted to give it a voice or use it on twitch.

[rigid snow]

pitch should be fine regardless of how shit the output is considering the instrumental does not bleed into the separated vocal. if you use a bleedless model that should be totally good, idk how reverb affects things though. i'd dereverb with a separate model

[next rover]

I've got a GeForce RTX 3060 and an i9 12900

[nocturne olive]

12GB 3060?

[nocturne olive]

Mainly hardware

[rigid snow]

tts fucking sucks

to implement

[wide flicker]

Oh, I have something similar actually written in python and using gpt-3, except the llm itself isnt hosted on my machine

[nocturne olive]

I hate cloud models

[next rover]

Not sure from what I remember Nvidia GeForce RTX 3060 but it seems to be great. Last time I checked 4090 was the best one

[rigid snow]

most of the models are way too big for quality of output. like the most performant way to do custom voice tts is still a small model + rvc after all these years

[next rover]

I will need to upgrade my ram. I've got 16 GB ram

[nocturne olive]

The VRAM is what matters

[wide flicker]

It's not my first choice but for testing purposes it suits my needs; I figure once I have more tools and agents defined I can work on migrating to a local model

its very early stage rn

[nocturne olive]

Money though

[next rover]

Ah yeah vram. That's what the 12 GB is. I've heard about 24+

[nocturne olive]

Yeah do you have a 12GB 3060?

[next rover]

So I'm assuming I need to try and reach that

Yes

[nocturne olive]

Good then you can run 8B models just fine

[next rover]

Let me double check

[wide flicker]

i paid $10 sometime in the spring and havent run out yet so I figure im fine for now LMAO

[nocturne olive]

24GB is found on the 3090, which is what I recommend for inference

[obsidian mantle]

Why cant llm's train on the fly

[rigid snow]

i've given up on local llms it's just not worth it they're pretty bad

[next rover]

I'm apparently using Mistral 7B but I don't know what the 7B means

[nocturne olive]

I will never pat for LLMs

[wide flicker]

¯_(ツ)_/¯

[rigid snow]

7b parameters, model size

7 billion

[obsidian mantle]

many

[rigid snow]

as llms go that's not many

[nocturne olive]

How could they? They're statistical models

[rigid snow]

oh btw what about the titans arch what happened with that

[obsidian mantle]

Isnt that "memory" vedal made for neuro - technically learning?

[rigid snow]

no

[nocturne olive]

No

[nocturne olive]

7 billion parameters, at 8 bits per parameter it requires 7GB of VRAM for the model alone

[rigid snow]

it's like if you had no memory whatsoever and relied on a journal to "remember" things

[obsidian mantle]

Hmm

[rigid snow]

llm memory is what it is, a journal

[obsidian mantle]

Do they really need training then

Their speech is pretty much alright?

[rigid snow]

uhhhhh yeah? to understand language?

[obsidian mantle]

Is it about complexity of things they can discuss

[next rover]

Does the ti at the end tell you anything

[rigid snow]

at the very least

[next rover]

I'm trying to figure out how to know what GB it is

[rigid snow]

the bottleneck with ai inference is usually the memory bandwidth

and i don't think the ti has any improvements on that?

[stray dragon]

usually LLMs do pre-training and then fine-tuning, training too much further can lead to overtraining, and extra pre-training on a fine-tuned model is less stable for that than extra fine-tuning (pre-training is more ideal for on-the-fly, since you get the loss function by comparing output to input, instead of relying on a human or better LLM to produce an example response to compare to for fine-tuning)

also training is much more intensive than inference, so you'd notice slowdowns or simply be unable to run both at the same time unless you outsource the training to another computer and have it run in parallel with the inference model

[obsidian mantle]

So you need certain volume to just run it, and memory speed to train/run efficiently?

[stray dragon]

and any solution for that would be pretty hacked-together i feel

[next rover]

8GB

Damn not 12

[stray dragon]

you could do "sleep cycle training" or something where you have it experience things as it runs and then it stores that data to train on later during off-hours

[rigid snow]

idk what it is for training, but pretty sure the actual performance is what you need, but for running (inference) it's memory speed and amount is what you're after

[stray dragon]

it's a lot easier to not update the parameters much and to add a memory system or something though

[obsidian mantle]

Inference= running?

[stray dragon]

inference = producing responses

yes

[obsidian mantle]

Is 5080 good for ai

[stray dragon]

they crippled it with the shitty ass vram lol

16gb is alright

but i wouldn't want to spend that much on 16gb vram

[rigid snow]

"buy datacenter gpu pls we want $2000 margin"

[next rover]

So I'm definitely going to need a better graphics card if I ever want to run something like this is what I'm gathering at least to the extent of streaming

[stray dragon]

12gb is enough to start off

get more as you need more

[next rover]

Yeah but I got 8GB

[stray dragon]

ah

[next rover]

I was wrong

[stray dragon]

can work with smaller models i suppose

[next rover]

You suppose. Oh god I feel so inferior lol

[stray dragon]

7B is a common standard but there's smaller models out there that are still pretty good

[stray dragon]

buying big ass gpu is expensive, don't do it unless you're sure you need it

[rough bloom]

there are some cases where some continuous learning like this is done now
the big use case I'm aware of is basically continuous RL since you usually want your data for that to be on-policy (from the newest version of the model)
usually that happens during training before a model is released but there was recently an article by Cursor that they're applying it to their tab completion model too
(but AFAIK they all kinda "cheat" by using checkpoints)

[stray dragon]

ooh it's finally catching on

[rigid snow]

well you see i think that they're all shit below like 30b fuck local llms they're insanely stupid. and you can't exactly run anything larger on one consumer grade gpu at a sane quant

[next rover]

I use my computer for gaming and music production and I plan to start streaming. Was gonna go the v tuber route

[stray dragon]

and you want to make a local LLM based discord bot with a vtuber model as well? damn that's busy

good luck

[next rover]

I got some other ideas as well. But with this recent one. I want to get as far as I can with it

[rigid snow]

how is it busy lmao what

music + the bot

[next rover]

The discord one was just to test it and see how it would work

I figured testing something like this on discord would be beneficial to learning and seeing if I could do it

[stray dragon]

maybe gaming doesn't count but meh "i want to make local llm discord bot that is also a vtuber" is a pretty big commitment on its own i feel

[rigid snow]

yeah but music is a thing you can assume a person doesn't do professionally

[next rover]

I don't want to do exactly what vedal did

[stray dragon]

ah

i have zero experience with it

didn't know

[next rover]

More so a fun AI thing I could bring out from time to time

The majority would be my stuff

[rigid snow]

so one can just allocate less time to that

[next rover]

But it seemed like a really cool thing to have

[nocturne olive]

Oh you're on a 3060Ti? That means 8GB VRAM guaranteed and you're screwed

[rigid snow]

it's like all art, most people who do art don't have a career around art

[stray dragon]

yeah it sounds a lot less busy now lol

[next rover]

I'd say music is just as hard. I just know more about music but even still I wish I was better

[rigid snow]

i'm not saying it isn't hard

[next rover]

Yeah lol

[stray dragon]

thanks nshittia!

[green iron]

are there any docs for the 7tv api? cant seem to find anything useful on the website or the repo

[rigid snow]

the backend code is the docs. they don't have docs

[next rover]

I don't have a career around it but I'm still hoping. I just need to better prioritize my time with actively investing in things that will help me in the future like learning skills

[nocturne olive]

I do audio ML and vocal stuff, I make cool stuff

[next rover]

Hell yeah

[nocturne olive]

The cool stuff in question

[next rover]

I finished a song and tested out Hatsune Miku for the lyrics

It's ok but not the best

[nocturne olive]

This is a real Neuro vocal synthesizer by the way

[rigid snow]

btw if you want to list channels and emotes and that kind of thing your best bet is the graphql api the rest api they have is pretty bad

[nocturne olive]

If you're committed to running inference stuff, I suggest a 3090 as a good value inference GPU

[next rover]

I can show you some of my stuff I don't think it's as good honestly

[nocturne olive]

I personally just silly around, throw some stuff at NeuroSynth, throw that at a DAW and throw a mix on top, and cool stuff just magically comes out

[green iron]

thx. Btw would be thankful if there's an example somewhere of fetching emotes(metadata only) from an emote set fetching https://7tv.io/v3/emote-sets/Id works for me

[nocturne olive]

You have Vocaloid? Interesting

I only use free vocal synthesis tools

[next rover]

Oh I can't post audio in here

[nocturne olive]

Lack embed perms

[next rover]

It was a thought to use it but I really want to work on my own voice

[nocturne olive]

Just keep chatting to unlock

[next rover]

Ok

Can I dm you then

[nocturne olive]

Your own synthesizer?

[next rover]

No like me the human hi

[nocturne olive]

Oh

[next rover]

Lol

[nocturne olive]

I only use synthesizer voices

Like Kasane Teto, Neuro-sama, Evil Neuro

[next rover]

I think there fun. But I feel like I have to. Real talk. I feel like I have to be the best and if I ended up not trying to achieve being a great vocalist it would not satisfy me

But yeah I have two projects I can share with you

[nocturne olive]

When it comes to vocal synth usage, it's not that hard once you get it on machine learning banks, but way harder on sampled banks like Teto or Miku

[next rover]

One was a thing I trashed that I still really like and then the vocaloid song I made that was a lot of firsts for me

[nocturne olive]

Also because I am the only one with NeuroSynth access that can do good stuff with it I am automatically the best NeuroSynth artist

[next rover]

Yeah getting Miku English was a process

[rigid snow]

if you still wanna do gql here's what that would kind of look like. it's just that you have more control with gql over what you're requesting

[nocturne olive]

JP banks to EN is kinda hard
JP to Finnish is easier

[next rover]

I'll just dm them

But yeah that's basically my portfolio.

At least what I'm comfortable sharing. And isn't like 20 second musical idea lol

They seem so simple. But it's weeks of grinding. And hours in between moments of inspiration

I don't know what the average is on song writing. But it can take like hours to create 20 seconds of music

[dry charm]

gql is a whole another SQL

it's like NoSQL but it give me the headaches

[rigid snow]

what

[dry charm]

exactly

[tender river]

allowing me to read pdfs on an ereader

[midnight sigil]

I've started multiple projects and I never finish a single one of them

[stray dragon]

real and true

and factual

[slender timber]

Progaming

[midnight sigil]

Progaming real

[cosmic sphinx]

a truly good project is never finished
it's constantly evolving

[midnight sigil]

[slender timber]

Tried debugging and ended up with way more errors than when i started

[stray dragon]

that sticker is great

[olive sable]

Hi

[gritty dust]

A wise man once said

[gritty dust]

hi

[olive sable]

Happy birthday cheese

[bright quail]

I was wondering how the twins, Evil in this case, is able to recognize itself. I understand she uses a Clip-style model or a seq2seq multimodal like Florence2 behind the scenes, but I guess is like fine-tuning the model with images of them?

[midnight sigil]

he said the coffee was eliv's face

contextual clues

[bright quail]

you think so? A lot of times in art review they seem to know which on is which (sometimes)

[midnight sigil]

they do have a vision module

not sure about the implementation, but according to the old geoguessr videos I think it might be just some vision embeddings with attention

[shadow sinew]

I would guess there's some fine tuning to recognize themselves. They need it for art review.

[midnight sigil]

but in this case, evil doesn't really need a vision module to answer "I'm so cute"

'cause she have the context of that coffee has her face

[shadow sinew]

I don't think he told her before showing her. Let me recheck the clip.

[bright quail]

yeah, I think it was something along the lines of: "look who it is"

and evil responding 'its me!"

[shadow sinew]

https://www.youtube.com/watch?v=ZwTtyewjhgY&t=85s

[midnight sigil]

I check the VOD and I guess she didn't know

my bad

how do I have bad memory

I'm not even 20

[solid bough]

I tried np.float16 there:py x_train = np.load('x_data.npy').astype(np.float32) / 255.0 y_train = np.load('y_data.npy').astype(np.float32) / 255.0

Yes

It's alive

Ahah... Uh wait, am I supposed to be worried? And that looks like a cursed heartbeat

Btw

HIII caibi

[midnight sigil]

[midnight sigil]

just funny

[solid bough]

Okay

And I'm training the worst model with my GPU rn:

Dw, There is gonna be a down movement again.

Now I can experience the pain of trying to get that exact code that runs right now to run on CPU alone 💀

[solid bough]

Told ya:

[tight tinsel]

so to even change the built in dpi for my mouse i need to install a full app??

[solid bough]

I think it depends, But I have a LG mouse and last time I checked, positive.

[slender timber]

I don't understand my errors

I just edited a little and it somehow spawned 30+ errors

[solid bough]

That's coding for ya, Should I just trial and error the code that trains the model on my GPU to run on a Intel® Xeon™ E5-1650 V3?

[slender timber]

I'm just gonna redo everything

(Said no one ever)

[midnight sigil]

buh I can't do this?

🥀

python plz fix

[solid bough]

oh god

[solid bough]

You can do it, I believe in you!

[slender timber]

I have backups

[rough bloom]

bestValue := currentEval

[midnight sigil]

chat this is why you use a version control system

[midnight sigil]

when did my python turn into a makefile

[rough bloom]

always has been

[midnight sigil]

why can't they unite both two

[stray dragon]

weirdly smooth loss curve

ah it's being fit to the data

i see

[solid bough]

lr is 0.00008

[stray dragon]

the curve in picture 1 is different than in 2

[solid bough]

Yeah, that first one was an earlier one, That is right now the state:

[stray dragon]

very different actually, must have been a big spike around epoch 600

[midnight sigil]

the real loss graph belike

[solid bough]

No, that is literally the raw loss values with no processing.

[midnight sigil]

it can't be

[stray dragon]

probably more like

[solid bough]

I'm being serious

[stray dragon]

it literally isn't, the values at different points change between pictures

[solid bough]

Because that is while the trai... You mean that you sent?

[stray dragon]

wait i'm mistaken, the numbers in the graph don't change that much

[nocturne olive]

There's not a chance that is the real loss unless this is an exceptionally simple model
Loss just is not that clean

[stray dragon]

still heavily doubt it's the raw loss values, yeah

that's gotta be smoothed

you do not see smooth loss like that anywhere

[solid bough]

Well, learning rate is "0.00008" and that is pretty much it

[nocturne olive]

Yeah not even close to that smooth
NeuroSynth at its most stable is nothing like that it still varies like crazy

[solid bough]

That is where the data get's held:py #Training the model print("Training...") text="Done, Training the Model (initialisation)" log_info(text, logg) train_rmse = [] test_rmse = [] train_score=[] test_score=[]
And that is the plotting routine:py def plotscrores(scores, test_scores, fname, on_top=True): log_info("Plotting scores...", logg) plt.clf() ax = plt.gca() ax.yaxis.tick_right() ax.yaxis.set_ticks_position('both') plt.plot(scores) plt.plot(test_scores) plt.xlabel('Epoch') loc = 'upper right' if on_top else 'lower right' plt.legend(['Train', 'Test'], loc=loc) plt.grid(True) plt.draw() plt.pause(0.01) # <-- updates live plt.savefig(fname)
There is no such smoothing being done.

[solid bough]

Nya >:3

And that is the loop that writes to the variables:```py
for i in range(num_epochs):
print(f"Epoch {i} of {num_epochs}")
text=f"Epoch {i} of {num_epochs}"
log_info(text, logg)

print(x_train.shape, x_train.dtype)
print(y_train.shape, y_train.dtype)

model.fit(x_train, y_train, batch_size=batch_size, epochs=1)

#Training results
mse = model.evaluate(x_train_mini, y_train_mini, batch_size=8, verbose=0)
train_loss = model.evaluate(x_train, y_train, verbose=0)
train_rmse = np.sqrt(train_loss)   # if your loss is MSE


print("Train RMSE:", train_rmse)

#Test results
mse=model.evaluate(x_test, y_test, batch_size=batch_size, verbose=0)
test_loss = model.evaluate(x_train, y_train, verbose=0)
test_rmse = np.sqrt(test_loss)   # if your loss is MSE

train_score.append(train_rmse)
test_score.append(test_rmse)

print(f"Test RMSE:", test_rmse, "\n")

model.save('Model.h5')
print("The model is saved.")

plotscrores(train_score, test_score, 'Scores.png', True)```

So god for saken nowhere it is getting smoothed

[stray dragon]

"plotscrores" lol

[midnight sigil]

my AI is rage-baiting me to get in his territory

and kill me

[rough bloom]

your test loss is using the training dataset
also you're evaluating it twice which is unnecessary, but that's just efficiency

[midnight sigil]

test_loss = train_loss

[solid bough]

Lmao, I wondered why the hell both are the same

[solid bough]

Where was that, I literally can't find where that line is

It literally doesn't exist on my code:

[solid bough]

I didn't name the function, ok?

[midnight sigil]

MY BOT BEAT 1300ELO STOCKFISH

[solid bough]

Ah thanks for showing.

[stray dragon]

congrats wtf

[midnight sigil]

tbh it's not a fair fight

I limited stockfish to only search for 5000000 nodes

[stray dragon]

and you started on white

[keen hatch]

My model recognizes itself in pictures based on the features present (even without separate fine tuning of the vision part of the model)

[stray dragon]

but it's something

[midnight sigil]

let me try raising it up to 10,000,000 nodes and start on black

[solid bough]

The curve starting to get movin:

[midnight sigil]

*differentiate

[solid bough]

It was smooth before, and now it is getting a bit squigly.

[midnight sigil]

it's because you have two curves

[solid bough]

But that is before fixing that

what t showed

[midnight sigil]

I guess yea

probably not what he meant

[solid bough]

Well, t was right that it was using the train variable instead of test.

[stray dragon]

i would've never spotted that error if shuni hadn't pointed it out

[solid bough]

Yeah, but I appreciate the help to find the error, next run is gonna use the loop in that configuration:```py
for i in range(num_epochs):
print(f"Epoch {i} of {num_epochs}")
text=f"Epoch {i} of {num_epochs}"
log_info(text, logg)

print(x_train.shape, x_train.dtype)
print(y_train.shape, y_train.dtype)

model.fit(x_train, y_train, batch_size=batch_size, epochs=1)

#Training results
mse = model.evaluate(x_train_mini, y_train_mini, batch_size=8, verbose=0)
train_loss = model.evaluate(x_train, y_train, verbose=0)
train_rmse = np.sqrt(train_loss)   # if your loss is MSE


print("Train RMSE:", train_rmse)

#Test results
mse=model.evaluate(x_test, y_test, batch_size=batch_size, verbose=0)
test_loss = model.evaluate(x_test, y_test, verbose=0)
test_rmse = np.sqrt(test_loss)   # if your loss is MSE

train_score.append(train_rmse)
test_score.append(test_rmse)

print(f"Test RMSE:", test_rmse, "\n")

model.save('Model.h5')
print("The model is saved.")

plotscrores(train_score, test_score, 'Scores.png', True)```

So that should be the fixed code

[solid bough]

Thanks for pointing that out shuni.ex

[rough bloom]

oh, I think I know why the loss curve is so smooth now

it's not the actual training loss

[midnight sigil]

awww man yea no

null pruning actually doing something at the end game tho

[solid bough]

That are the Parameters btw:py #Training data settings num_epochs = 200000 batch_size = 10 valid_ratio = 25 lr = 0.00008 #default 0.0008

HOW IS THAT STILL GOING DOWN???

[stray dragon]

you used a lower learning rate than usual

[solid bough]

oh. right...

[stray dragon]

the model will converge much slower

[solid bough]

my bad

[stray dragon]

and will tend to reach a lower loss when it does converge

[solid bough]

Well, each epoch needs maybe tops 1s

Like around 0.9s each epoch

[rough bloom]

the hyperparameters aren't the issue
it's that your "training loss" is evaluated using the entire training dataset after each epoch
it's not the loss used for optimization

it's not spiky because it's a huge average, not a sample like usual

[solid bough]

Yeah... But in my defense that is what the original code did too

[stray dragon]

interesting to see the average go down so smoothly

[solid bough]

But the upside is that I can train it on any data

Any

(Ahem, images)

[rough bloom]

because there's (probably) a lot of data

[solid bough]

Well, let's say I use a learning rate of 2, that doesn't go well

[solid bough]

10 images

that is the datagen.py parameters:```py

--- Configurable parameters ---

num_images = 10 # Number of images in your folder
samples_per_images = 4 #default is 10
dots_per_images = 100 #default is 60
image_w = 144
image_h = 192
image_dir = "pictures"
num_channels = 3 # Must match the model input channels
num_samples = num_images * 2 * samples_per_images```

Aka the code that makes the data the train.py uses

[opaque sigil]

No wonder then

[solid bough]

😅

Hiii 😅

[stray dragon]

i was thinking "epochs make me think of looping over an entire dataset 1 time per epoch, but surely that can't be happening here"

[solid bough]

https://tenor.com/view/nervous-nilly-30rock-drink-shaking-nervous-gif-15622273

[stray dragon]

10 images in the dataset

i see now LOL

[rough bloom]

that doesn't make a difference for how you calculate the loss
if anything calculating the loss separately like this is limiting because it slows down training

[solid bough]

And that is low

Like EXTREMELY low

[rough bloom]

yeah, that is exactly what's happening

[stray dragon]

honestly, upping the learning rate is probably better in this case

[solid bough]

Well, that is why I wanna get that code running on a CPU

[stray dragon]

if you do too many epochs over the same training data, it converges worse

[solid bough]

Well, I just run it like right now for Kicks n giggles.

[rough bloom]

usually you'd never evaluate the entire training dataset like this, it's insane
only works with small datasets (or a really really fast model I guess)

[solid bough]

What did I do now toast.? 😅

[opaque sigil]

You could add some randomness to the images to make more input data and force it to better learn the actual features idk

Like idk, shift them or sth

[stray dragon]

yeah, transformations or masks or mirroring or rotations or blurs

[solid bough]

I have no idea how I can make the Code actually use the 600 images that I can train it on GPU right now.

[obsidian mantle]

What do you have in your bot

[solid bough]

if I go higher on image amount, I need to drastically drop the other stuff.

Except if I run on CPU

[stray dragon]

almost tempted to work on bot but i added a bunch of code yesterday and now it crashes without printing anything to console so uhhhh i think i shall procrastinate that a little bit longer

[solid bough]

Because I repurposed temporarily one PC that has 64GB of total memory

but no GPU. (not even iGPU)

[stray dragon]

curious to see how the changes will do (i've added my first idea for a pruning system when i heard about this challenge) but not quite that curious just yet

[solid bough]

If I said what epoch it is doing right now, it would be out of date by like a dozen already

[rough bloom]

your training code is inefficient
improve it

[obsidian mantle]

meanwhile i struggle to copy lague's code into c++ and fail for 3 days straight

Im not winning this challenge thats for sure

[solid bough]

I... didn't... Ugh. I have to come clean I guess.

Chat was doing the conversion from Python 2 code to Python 3 code

Yes, boo me, but it runs.

[stray dragon]

oh so it's ANCIENT machine learning code

lmao

[rough bloom]

Python 2

but yeah that makes sense

[solid bough]

Yes: https://github.com/HackerPoet/deepdoodle

That is straight the original

[solid bough]

All I was able to do was the print changing and xrange, but the keras n stuff was something I was just not good on

[rough bloom]

so the original code was already a bit weird but it's much worse after the modifications

[solid bough]

Yeah, What do you expect of Chat?

[rough bloom]

it's probably not even that hard to fix

[olive sable]

chatgpt is decent enough at bugfixing, but its bad at generating complex code imo

[solid bough]

I had to go to tensorflow, and I myself didn't know that before Chat did that.

[solid bough]

In the past where I tried it, it was not going great.

Is it now overfitting?

[slender timber]

Right i found what's causing my errors

[olive sable]

if you give it the problematic code and the error, 9/10 times it will find the issue in my experience

[slender timber]

Accidentally deleted some letters

[solid bough]

Yeah, but to run the code on CPU with the model staying the same, no.

[rough bloom]

just don't evaluate the losses twice, only do it once like in the original code
also make sure you're always specifying the batch size, it's probably crashing with more images because it tries to process the entire dataset as a single batch or something

[solid bough]

I tried lately to train on CPU only that I have the upside of 64GB total Memory as my advantage

And the PC is then dedicated only for training the model

OOO, that looks not too bad, it's still mids Training though:

"mids draining" 🤣

[rough bloom]

can't really tell because there is no test loss
but since you only have 10 images you aren't going to fit the model to the distribution you want anyway

[solid bough]

Right... Imma just wait until it goes straight to the right, if that ever happens...

[nocturne olive]

You don't know if it's overfitting without eval loss

[solid bough]

Yeah

[solid bough]

but that isn't too bad for my standards

[rough bloom]

it will happen eventually, but by that time your model is already cooked

[nocturne olive]

What kind of model is that anyway?

[solid bough]

Deep Doodle: https://youtu.be/eEFlk6sSv88

That is straight of the maker

[nocturne olive]

Oh so it's just an image to lines model

[solid bough]

Lines to images, Yes.

And I really REALLY wanna beat that model

And I do have data

but it's grooling to sift through the THOUSANDS of images for QC

Idk the right word, Corrections are appreciated

[olive sable]

trying to make cubemaps work in vulkan

[solid bough]

Here is a hug.

[olive sable]

its mostly just refactoring existing code

[slender timber]

I finally found my errors

I wrote Value instead of value

[olive sable]

i have to make stuff that werent settigns before parameters and put them in a diffrent file

[obsidian mantle]

i'll be so down if i dont make anything good today

[solid bough]

Yikes 😬

[solid bough]

https://tenor.com/view/hug-gf-hug-tightly-hug-tight-hugging-hugs-gif-17493334962318329984

[olive sable]

its not too bad, just copy pasting a lot of stuff

[solid bough]

I end that now:

And yeah, the model was still mids training when I started doodler.py

That can't go wrong

Well, it is that kind of model

Ooo, Second line

Yay

[solid bough]

How do I run that on CPU only?

Oh

You're on Phone

Sorry shuni.ex

But that is straight the code with the changes where you saw the error

[rough bloom]

why do you want to train a CNN on CPU

[solid bough]

screams in Model.h5ValueError: Dimensions must be equal, but are 8 and 3 for '{{node mean_squared_error/SquaredDifference}} = SquaredDifference[T=DT_FLOAT](sequential/activation_7/Sigmoid, IteratorGetNext:1)' with input shapes: [20,8,192,144], [20,3,192,144].

[solid bough]

I'm going to go straight to the point, The PC has 64GB total of RAM:

So that PC has more than enough RAM to train the model

[nocturne olive]

I don't even see a GPU

[solid bough]

Yeah, because the PC has none

[nocturne olive]

If there is no GPU it should automatically run on CPU

[solid bough]

Nothing

[rough bloom]

you do not need 64 GB of RAM to train this model, it's tiny

[nocturne olive]

Tru

[solid bough]

I was planning on doing 600 images as a test and then sifting through more images to expand

[nocturne olive]

Tensorflow doesn't even have GPU acceleration on Windows

[solid bough]

I have no idea, lemme look in the logs

That is all I see on the logs...
07/10/2025 16:04:46 [INFO] Checking if the data is set up
07/10/2025 16:04:46 [INFO] Data is there.
07/10/2025 16:04:46 [INFO] Loading Data...
07/10/2025 16:04:52 [INFO] ('Loaded', '8000', ' Samples.')
07/10/2025 16:04:52 [INFO] Attaching more channels and splitting the data.
07/10/2025 16:04:54 [INFO] Channels attached and split.
07/10/2025 16:04:54 [INFO] Shuffling the data...
07/10/2025 16:04:55 [INFO] Data got Shuffled successfully.
07/10/2025 16:05:03 [INFO] Setting the image format.
07/10/2025 16:05:03 [INFO] Building the model from Cache or Scratch.
07/10/2025 16:05:03 [INFO] ILoaded pre_model.h5
07/10/2025 16:05:03 [INFO] set up the optimizer successfully.
07/10/2025 16:05:03 [INFO] compiled the optimizer.
07/10/2025 16:05:03 [INFO] made a Model image.
07/10/2025 16:05:03 [INFO] Done, Training the Model (initialisation)
07/10/2025 16:05:03 [INFO] Epoch 0 of 200000

At that line:py model.fit(x_train, y_train, batch_size=batch_size, epochs=1)

So pretty much on the bottom of the file

Well, it also throws:2025-10-07 16:12:52.287128: W tensorflow/stream_executor/platform/default/dso_loader.cc:64] Could not load dynamic library 'cudart64_110.dll'; dlerror: cudart64_110.dll not found

Also:2025-10-07 16:12:57.806102: W tensorflow/stream_executor/cuda/cuda_driver.cc:269] failed call to cuInit: UNKNOWN ERROR (303)

But I think I can ignore that

[rough bloom]

is that on the GPU or the CPU machine?

[solid bough]

CPU

[rough bloom]

then ye it's fine

though you really should just train it with the GPU

[solid bough]

If there was enough VRAM 💀

[rough bloom]

how much do you have

[nocturne olive]

How do you not have enough VRAM for such a tiny model?

[solid bough]

Only 8GB

[rough bloom]

that's enough

a model this tiny will need practically nothing
more VRAM just means higher batch size but practically anything will work

[solid bough]

I wanted to train for a test 600 images, and then I wanted to sift through more images for qc to have like 1200

[nocturne olive]

Sounds tiny

[solid bough]

The vast, Vast majority of the images don't look where the camera was

But what trips me up is still:ValueError: Dimensions must be equal, but are 8 and 3 for '{{node mean_squared_error/SquaredDifference}} = SquaredDifference[T=DT_FLOAT](sequential/activation_7/Sigmoid, IteratorGetNext:1)' with input shapes: [20,8,192,144], [20,3,192,144].

[solid bough]

But I can't do 600 images because apparently I go OOM when doing that

[nocturne olive]

Wha

[solid bough]

Lemme just

[nocturne olive]

You're definitely doing something wrong then

Are you trying to load every single image into memory all at once?

[rough bloom]

yeah, very likely attempting a batch size of "the entire dataset"

[solid bough]

do the default settings

[rigid snow]

batch size = dataset size

[nocturne olive]

Also how are you using the GPU in the first place? Tensorflow has no GPU accel on Windows

[solid bough]

. . . 600!?

[rigid snow]

what

[solid bough]

I have 600 images

in total

[rigid snow]

don't make the batch size 600

[solid bough]

Yeah, I just set default stuff now

Except image amount, That is the total images number I have now

That... What the

huh

WHAT IS HAPPENING!?

It friggin runs now?

[solid bough]

Okay, What the filtered is batch size supposed to mean?

[rough bloom]

still trying to allocate way too much
at least it runs now

[solid bough]

YEA, BUT HOW???

Last time I tried it ALWAYS failed to run

[rigid snow]

basically it tells the training code how many samples from the dataset to load at the same time

[solid bough]

So a lower batch size would lower the memory needed, but something is... The file is there in that chat 🤦‍♂️

[rigid snow]

So a lower batch size would lower the memory needed
correct

[solid bough]

That is the file

[solid bough]

All I can say now is: WTF

What the, I can't

That is not train.py

that is of doodler.py

Wait, it's of doodler.py

[solid bough]

It was unrelated to the training

[languid escarp]

How to submit a Library of Ladev issue? It doesn't handle dashes, like bun-believable. https://libraryofladev.com/search?text=bun believable

[solid bough]

I know why it was working

[rigid snow]

https://discord.com/channels/574720535888396288/1337595628607242282

[midnight sigil]

I think the search function of my chess bot is good enough, for now. I have some issue on evaluation now

[solid bough]

It was using the old x_data.npy and y_data.npy

[midnight sigil]

PeSTO's PVTs doesn't really perform well on my model

surely shiro won't notice if I just copy the piece value tables from Boychesser

[solid bough]

Well, with the x_data.npy and y_data.npy corrected, I get:

I had to delete the files that it is what I set.

[obsidian mantle]

ok i think i made negamax with alpha beta with nothing else and it works

when i remove beta cutoff it slows down so i guess its all right

with quiescence it became super slow and super bad

it is literally the same as normal search but shorter wtf

is this correct

i dont understand where i put - and where i dont put -

im so dumb should have become a janitor

i think i got it

[obsidian mantle]

incorrect

i didnt make move here so i pass alpha and beta and everything unchanged

i'll do it really explicit this time and only then shrink

because it all gets messy

[solid bough]

What now? Is it correct or incorrect?

[obsidian mantle]

no its incorrect

we go deeper with a minus if we made a move

here i just pass it to different search mechanism with same alpha, beta, and sign

[solid bough]

Oh, well, after I’m done making my food, imma loook what I need to do that I can run my code on the CPU

[obsidian mantle]

this is correct (i think)

[warped frost]

I heard ass

Nvm pass

[obsidian mantle]

see how alpha and beta have no minus and passed on their places

[solid bough]

Hi

[warped frost]

Bye chat

[solid bough]

Why???

[warped frost]

No ass

[solid bough]

Huh???

Idk what you mean

Like, did you mean the word?

[obsidian mantle]

ooh i forgot about sorting.. thats why it looks slower than when i walked this path last time

[rigid snow]

qualcomm acquiring arduino

interesting

[solid bough]

mlntcandy

[rigid snow]

chaosminecraft

[solid bough]

That is what happens if I do 600 images with default settings of the GitHub

It just doesn't have the needed memory

[rigid snow]

i don't know what you got going on but what should fit into the memory during training is the model and the batch

[solid bough]

The training data

[rigid snow]

if it doesn't have enough memory but still runs at some point that must mean the batch size is low enough to allow that to happen. are the samples even in size? if oom doesn't happen consistently then samples not being the same size would be an explanation

[solid bough]

The time where I said "How?" it was with the data having only 10 images

[rigid snow]

what are you training even

object recognition? image generation?

[solid bough]

Something like that: https://youtu.be/eEFlk6sSv88

[rigid snow]

i see

[solid bough]

I wanted to make a better model

[rigid snow]

i don't know about that specifically

but 10 should be low enough? probably?

[solid bough]

That is actually too low compared to what the creator used, apparently the creator used "It was 1200 plus mirrors"

[rigid snow]

weird idk then

[solid bough]

That is straight the message inside the "" part

Yes, I indeed had the balls to contact the maker

Assuming the GPU of the maker having 8GB when it was trained, I wonder how it was done

[solid bough]

And that is the training data size when 600 images:

[obsidian mantle]

does anyone have any idea what "ply" could mean

[solid bough]

in total 1,85GB

[obsidian mantle]

its lague's code

[solid bough]

Wait, What are you trying to do?

[obsidian mantle]

im making chess bot

im so bad i have to look at other people's code

[solid bough]

Oh, I thought you were doing 3D stuff

[olive sable]

players?

[obsidian mantle]

noo

no its somethin search-related

[solid bough]

My bad vituha

[obsidian mantle]

ply moves ply tree idk

[solid bough]

Sorry

[rough bloom]

plays?

[obsidian mantle]

is it some slang

[olive sable]

oh its chess

[obsidian mantle]

plays.. iidk

yess

its ches

[olive sable]

you said league so i thought league of legends

[rigid snow]

league

[olive sable]

oh

[obsidian mantle]

i meant Lague its the name of some guy

[rough bloom]

Lague != league

[olive sable]

you didnt say league, i just cant read

[rigid snow]

lague the guy who made the chess ai videos

[olive sable]

ohh

sebastian

[rigid snow]

come on he makes fucking gamedev videos you should know him by name

[olive sable]

i barely know my own name

[rigid snow]

based

[olive sable]

you're setting the bar too high

soemtimes i hear a ping and then i look to discord and there's nothing there

discord has given me undiagnosed schizophrenia

[solid bough]

Then it is a reaction, I know that feeling.

[trim valve]

just turn off pings ezpz

[solid bough]

Now I need to start thinking which is: sdlkhjgklsdjhfghjklwsehjgrkewhl

[trim valve]

or more importantly just never get any pings

even easier

[olive sable]

i only have them on to mentions

[trim valve]

i accidentally turned off direct mention notifcations outside of dms

so i miss a lot

😭

[olive sable]

im 61st in the server ranking, its too late for me to never get any pings

i do enjoy reading them tho

unless its a ping about bugs

[solid bough]

Did someone say buks?

[trim valve]

@olive sable 🐛

[olive sable]

[trim valve]

my desk is once again a massive mess

😔

[solid bough]

What did I do now?

Oh that

[warped frost]

You know what else is massive?

[trim valve]

why can't this reverse engineering project be super easy so I can finish it and focus on real work like uni

[warped frost]

https://tenor.com/view/ninja-any-haircut-recommendations-low-taper-fade-you-know-what-else-is-massive-gif-3708438262570242561

[solid bough]

Nah, You. 🤣

You were not thinking of...

Uh

I can't say, not PG

[olive sable]

i have 2 options, either i make a new commandpool in the framemanager
or i have to pipe an existing one from main.cpp into the subfile, then into a function, and form there into a function called by that one

[warped frost]

Low taper fade

[solid bough]

-# totally wasn't thinking something else...
Oh, I am not familiar with that stuff, My bad...

[olive sable]

there really was no need to spell it out

[solid bough]

What do you mean spelling? I wasn't spelling.

[olive sable]

bruh

nevermind

[solid bough]

Spelling is when someone turns the word "Picture" to "P i c t u r e"

aka each letter on a word

[warped frost]

He meant magic

Like casting spells

[solid bough]

That made more questions than it answered.

actually, it didn't even answer...

[olive sable]

https://en.wiktionary.org/wiki/spell_it_out

[warped frost]

[solid bough]

Anyways, tensorflow really doesn't like running on CPU

[solid bough]

What line of that message?

[stray dragon]

damn he just had it spelled out to him

[rigid snow]

ericdoa my goat

[solid bough]

With the context I have I still don't understand

I'm just gonna work that the code can friggin work on CPU now 🤦‍♂️

[olive sable]

there are 2 lines there. you really cant figure out i meant the first one?

[solid bough]

No, I go in multiple paths while thinking what was meant, and some times I just happen to go the wrong one

[rigid snow]

https://google.com/search?q=imagine if ninja got a low taper fade

[olive sable]





thanks discord

[solid bough]

Oh great, lemme pull up my phone rq

[solid bough]

What emoji is that even?

[olive sable]

none

it failed to embed or something idk

[solid bough]

All I see:

[olive sable]

yes

that is what it says

because the emote failed

[solid bough]

Oh, on my iPhone it is just blank

[olive sable]

object replacement character (U+FFFC)

[solid bough]

WHAT THE

not your message

I reduce to 4 ima... Nevermind

[olive sable]

why do you even bother to send the message if you type nevermind in the same line?

[rigid snow]

erasing more effort probably idk

[solid bough]

Because of the "WHAT THE"

and that

But mainly for context that I was wondering how the code now works on the CPU only machine but that it then still didn't work out.

Sreams in binary streamTraceback (most recent call last): File "c:/Users/Sheep/Documents/temp/train.py", line 327, in <module> plotscrores(train_score, test_score, 'Scores.png', True) File "c:/Users/Sheep/Documents/temp/train.py", line 52, in plotscrores plt.plot(test_scores) File "C:\Users\Sheep\Documents\temp\.venv\lib\site-packages\matplotlib\pyplot.py", line 2842, in plot **({"data": data} if data is not None else {}), **kwargs) File "C:\Users\Sheep\Documents\temp\.venv\lib\site-packages\matplotlib\axes\_axes.py", line 1743, in plot lines = [*self._get_lines(*args, data=data, **kwargs)] File "C:\Users\Sheep\Documents\temp\.venv\lib\site-packages\matplotlib\axes\_base.py", line 273, in __call__ yield from self._plot_args(this, kwargs) File "C:\Users\Sheep\Documents\temp\.venv\lib\site-packages\matplotlib\axes\_base.py", line 419, in _plot_args for j in range(max(ncx, ncy))] File "C:\Users\Sheep\Documents\temp\.venv\lib\site-packages\matplotlib\axes\_base.py", line 419, in <listcomp> for j in range(max(ncx, ncy))] ZeroDivisionError: integer division or modulo by zero

Oh

Train RMSE: 0.28846472713957755
Test RMSE:  []```

[obsidian mantle]

this does not alter "depth" variable does it?

[opaque sigil]

no

[obsidian mantle]

instead of mating it repeats last turn move which is illegal

here it doesnt pass depth==max_depth for some reason

even though its supposed to be highest level turn

[opaque sigil]

are you even incrementing the depth anywhere

[obsidian mantle]

im setting it to 4 before calling starting search and decrement it when going down

can i get fen from cute chess

hmm

[opaque sigil]

iirc there was a way to have it log to a file and dump all states

idk specifics though

[obsidian mantle]

function looks like this, it detects mate but doesnt pass depth==maxdepth check and doesnt record the move

it is called with depth=4 (depth=maxdepth)

alpha is -9999999

[opaque sigil]

i guess it never gets to a checkmate

[obsidian mantle]

i dont get it

log prints board before move

wtf is this

in pictured board queen makes e7d7 (last move)

oh it bugged

nvm

so it detects bestmove e7d7, does it
enemy king moves
it detects mate but doesnt record the move so it makes last turn move

it correctly sees the board before move

it probably detected this mate 5-6 turns ago

wtffff

just do it man

mate him

[stray dragon]

?!

[obsidian mantle]

i know how to do it.
i'll print more info when turn count is close

to this mate

oh i think

i got mixed up with bot names and running wrong bot

[solid bough]

Hi shuni.ex

welcome back

[rough bloom]

@solid bough Colab notebook for your doodle thingy
don't even attempt to train on CPU, even inference sucks ass

[solid bough]

I really don't want Google thinking I'm doing anything but Training

[opaque sigil]

hey, running it on the cpu can be fast
all it takes is a smol useless model
and lots of pain manually vectorising

[solid bough]

I have a history where I used to run a render client on collab...

[solid bough]

If it even makes one

Welp, small image amount it is 💀

[rough bloom]

that's the thing with CNNs, they don't have many parameters so they're small but they still need a ton of compute

[opaque sigil]

i forgot how to implement CNNs ngl, wasn't that a 5 level deep loop or sth

well, for the shitty cpu version ig

ain't nobody got time to use mpi

[solid bough]

And more like a pain in my poyo to run code where it runs on GPU btw

[solid bough]

Btw, how am I supposed to open that file?

[opaque sigil]

Starting with TensorFlow 2.10, Windows CPU-builds for x86/x64 processors are built, maintained, tested and released by a third party: Intel
interesting

[rough bloom]

convolutions
one small kernel applied to every position of the input
(idk I never really dealt with CNNs except now)

[rough bloom]

upload to Colab

or another Jupyter instance but you very likely do not have one, so just use Colab

[solid bough]

Wait a second

I increased the batch_size thing, it ran faster

Yeah

Holy

[opaque sigil]

i guess you go over all rows, then all columns, then the same again and over all layers

so 5 level

nice

though

[solid bough]

Huh:

[rough bloom]

sounds about right

[opaque sigil]

not n^5 at least, more like zn^2k^2

[solid bough]

imma be quiet for idk how long

[trim valve]

there's a funny thing you can do wiht fourier transforms and convolutions iirc

its pretty neat imo

though I don't 100% remember what it is

[opaque sigil]

there's a lot of really cool things you can do with a good ol fft

[trim valve]

iirc its basically "instead of doing convolutions normally take the fft of the kernel and the fft of the image, multiply, reverse fft, poof convolution applied"

[opaque sigil]

that checks out i guess

since you're basically multiplying 2 polynomials

and you can use fft for that

[obsidian mantle]

i cant see what is wrong

is it because my last turn gets beta-cutoff

but why does it affect the winning bot and not the losing one

no, there are 35 moves and 10 cutoffs

it only checks one move and it gets beta cut-off

wtf though

so beta cut-off kills final search

but i dont use this high number anywhere than in calling initial search.. how can "evaluation" get this value

because opponent's moves are so bad he returns -999999999 which turns into 999999999

i just ban beta cutoff on max layer then

it works

ok tomorrow i add transposition + iterative search
and it beats stockfish ez ( )

[rigid timber]

9999 ELO tomorrow surely

[olive sable]

one of the joys of being an older brother is looking at your siblings math homework and thinking "which dumb fuck wrote these questions?"

[gritty dust]

Lol

[olive sable]

they had a x and y or z, but didnt specify if "and" or "or" took priority.

so i jsut assumed programmer style "and" took priority, and it was wrong

i need to smack that teacher

[hoary drum]

https://igorlinks.vercel.app

do u like my website

[rigid timber]

broken on mobile, so no

[clear sedge]

my predicament:

[clear sedge]

false advertising

[olive sable]

name: "Don't touch my links"
i guess i wont then

[obsidian mantle]

im gonna touch them

[clear sedge]

i have a whole 9 clicks per second

beat that nerds

[olive sable]

it is indeed broken on mobile

[clear sedge]

mobile is inferior anyway

[gritty dust]

Sam, I got a raspberry pi 16gb andddd radio equipment so... robot arm with AI vision remotely controlled will be peakkk

[olive sable]

if i use my other hand to hold the mosue i can do 9cps

[hoary drum]

nobody liking gayballs?

i like gayballs

why dont you guys like gayballs

[olive sable]

excuse me

nah

excuse you

[trim turtle]

huh

[hoary drum]

https://gayballs.vercel.app

i love gayballs

its the favorite out of my websites

[clear sedge]

holy shit i can play with them

[olive sable]

i dont like the website title

[obsidian mantle]

oh shit it supports japanese

[gritty dust]

Wtf I am not clicking that

[drowsy jungle]

I know nothing of coding aside from a little knowledge from watching Doug Doug code for a few years, mostly how to get AIs to cooperate. I have this idea for a game that I have been wanting to make. What is an beginner friendly coding language for game design. It needs to be simple enough that I can teach myself. I have chat gpt to help but I mainly want to use it as an assistant to learning like generating questions and problems to fix.

[obsidian mantle]

unity + they have their own ai helper

embedded into ide

they use c# for programming and there is a ton of guides

[sullen sierra]

interesting, so full utf8 support

[rigid snow]

“gayballs” is crazy work

[versed orchid]

although some characters are a bit cropped

[rigid snow]

mf WHAT

[hoary drum]

i aint putting each and every dot for every letter in every language

i automated it

[olive sable]

gamedesign gamedev is mostly unity ye, godot if you feel contrarian or unreal if you hate both yourself and your gpu works aswell, altho its harder

unity is the most widely used so a lot of resources on it

[versed orchid]

of course. and it's not like I complain - just an observation

why I can't attach pics btw?

[hoary drum]

idk

[tender river]

i've seen people use php for scripting but why laravel

[rigid snow]

this has 4k stars btw

[tender river]

php can be faster than python for some use cases (not a huge bar to clear i know)

[sullen sierra]

I once used pure bash for a game pathfinding backend, so, why not?

[rigid snow]

it's probably a replacement for ink but with laravel instead of react because ew react?? i am guessing here

ink being the thing that things like claude code and codex and gemini cli use

[rare bramble]

am I missing something or is the whole repo combined just some 200 lines of code?

[versed orchid]

check what it says in the readme

it's a micro-framework - it's just an addition to Laravel

[rigid snow]

these are pretty cool

[gritty dust]

whaaa

[rigid snow]

they started giving awards out for surpassing 10b, 100b and 1t tokens

on their api

[gritty dust]

ooo damn that's cool

[rigid snow]

to be clear i AM NOT openrouter

pic from twitter

[versed orchid]

Sure...

[obsidian mantle]

how much do 1t tokens cost

[opaque sigil]

tree fiddy

[vague shell]

varies based on the model, also if that's input or output tokens

for GPT5-Mini, 1T output tokens would be 2 million dollars + however much the input tokens cost

Open Router essentially resells it though, it is a connector to multiple API providers with a single unified API

[rigid snow]

the cheapest one would be gpt-5 nano on which approximately $5k assuming all tokens are input and cached

but probably like $40k if actual usage, but keep in mind 5 nano is crazy cheap

$2m is more realistic

[rigid snow]

all companies in the 1t club resell it (or you'll go bankrupt lol), other 1t ones i saw are from notion and cognition

(notion ai and devin)

[vague shell]

(nano is also crazy bad though, it only works for very easy tasks like summarizing text in the context or bad direct translations)

[rigid snow]

yeah

[versed orchid]

[rigid snow]

explain why this doesn't say bajillion dollars then

[versed orchid]

I can't attach pics still...

Sadge

[rigid snow]

chat more

[versed orchid]

SpAmMiNg??

[rigid snow]

no it's on cooldown

they have to be spread out

[versed orchid]

on cooldown? Like for a fixed time or fixed messages?

[rigid snow]

a message awards you 10-20 points but then a cooldown begins during which messages won't award points and you have to reach 1k i think?

[versed orchid]

Understood, thanks

Well... I just started chatting more actively a few days ago since finally I have some spare time after work and university

It'll take some time

[hoary drum]

oh shut up gayballs is better

[olive sable]

[merry plank]

https://tenor.com/view/kitten-keybo-lap-gif-19489640

me right now

installing RVC UI fully locally with code

[nocturne olive]

Applio?

[merry plank]

😕

[nocturne olive]

Other RVC GUIs suck

[merry plank]

wha

[merry plank]

oh

[nocturne olive]

https://github.com/IAHispano/Applio

Drop whatever you're doing and get this

[merry plank]

oh it is just a zip file

well that makes things easy

[nocturne olive]

Yeah Applio is real easy, on Windows anyway

[merry plank]

thanks

[nocturne olive]

What are you gonna use it for by the way?

If you're planning to use it for Neuro and Evil I have good voice models for you

[merry plank]

train my own voice maybe becouse I hate singing

[nocturne olive]

That works, Applio has an easy training mode

[merry plank]

also maybe neuro and evil too

[nocturne olive]

Then you want these

They're deprecated in favor of NeuroSynth and EvilSynth, but they're the best of the best when it comes to Neuro and Evil RVC

[merry plank]

cool

[nocturne olive]

Remember to credit if you post anything using them

[merry plank]

yeah I will for sure do that

a cool nightmode too

[merry plank]

where do I put the voice models

idk what folder

[nocturne olive]

You use the Download tab and drop the files

[merry plank]

[nocturne olive]

You extract the model files, then go to the download tab and individually drop each PTH and INDEX file

[merry plank]

oh ok

[nocturne olive]

That will let Applio put the files in the correct locations

[quiet finch]

Did the Minecraft part of Neuro use mineflayer?

Or a part of it

[merry plank]

I do need to remove background from a song before putting it in right...

what ever I will find out after it generates

yeah

[merry plank]

[quiet finch]

Try UVR

[merry plank]

yeah

[merry plank]

I like the acapella of evil singing life

but yeah better if I seperated background music

[nocturne olive]

It's recommended to use vocal synthesizer output as a base

[merry plank]

😕

[nocturne olive]

See

[merry plank]

the what

[nocturne olive]

Something like SynthV Solaria

[merry plank]

oh

[nocturne olive]

It has way better quality

NeuroSynth and EvilSynth are a thing because SynthV is expensive so we want a free solution

And RVC is just limiting

[merry plank]

oh ok gotcha

[rigid snow]

or sing it yourself

[merry plank]

isn't syth more robotic though

[quiet finch]

True

[worldly plank]

Who made that?

[quiet finch]

lol

[nocturne olive]

Technically

[nocturne olive]

Nah Machine learning synthesizers are really good these days

See

[nocturne olive]

Me and Wispers, though I did the work on the mix
CHXI's instrumental was also used

[quiet finch]

Yeah，old vocoder is……

[rigid snow]

oh and also you get backing vocals in the vocal track as well as the lead vocal mixed together when stem splitting and rvc won't handle those well

[nocturne olive]

Yeah true

Backings are the death of RVC

[merry plank]

isn't neuro and evil karaoke RVC?

[rigid snow]

v1 is

[nocturne olive]

No 100% not

Oh yeah V1 and V2 probably

[rigid snow]

beyond that voice synth

[nocturne olive]

But V3 no chance

[rigid snow]

i don't think v2 is rvc

[merry plank]

oh ok

[nocturne olive]

I think V2 is RVC or RVC on top of a basic synthesizer

[rigid snow]

maybe the latter? or maybe i'm just wrong i'm not confident enough

[nocturne olive]

But either way V3 is 100% not RVC

V3 is certainly a native synthesizer

Just like NeuroSynth

[merry plank]

so neuro and evil singing is pretty much vocaloid

[nocturne olive]

NeuroSynth is cool

[nocturne olive]

Eh not really, closer to SynthV AI

[quiet finch]

But it's impossible to train a model by self

[nocturne olive]

Clueless

[quiet finch]

It must cost much

[nocturne olive]

Literally what NeuroSynth is

[nocturne olive]

NeuroSynth is done on a single 3090

Which I have at home

Very cheap compared to an LLM

You should not try to assume things about vocal synthesizers if you don't know what you're talking about

[rigid snow]

it wasn’t even a 3090 at first

[nocturne olive]

Yeah at first a 4070Ti which has half the VRAM

[merry plank]

I have 4060 I think

[nocturne olive]

Should be fine enough for inference

The 8GB of VRAM will limit you in training though

Does that really matter when the end result is cool?

You don't do audio ML so don't go assuming things someone like me that does do audio ML can correct

[rigid snow]

superbox gatekeeping voice synth training pipeline

[nocturne olive]

Yes I have been doing so for a while now

[rigid snow]

genuine question why

[merry plank]

I still think RVC can be more realistic sounding

[nocturne olive]

Something like ENUNU?

[patent shard]

Not default models
Like FT2 bleedless or big beta 6x or something not old

[nocturne olive]

RVC is way more limiting

NeuroSynth can already do stuff RVC can't really do

[merry plank]

yeah like the original songs?

[nocturne olive]

Yeah vocal synth can do easy original songs

Silly

[quiet finch]

oh

right

I will keep in silent

[rigid snow]

with good enough input rvc will still sound better than neurosynth 90% of the time though

[obsidian mantle]

attempted to add transpositions and iterative search, now it misses easy mates, and is super slow
uhh at least nothing seems to be broken

[rigid snow]

i don’t have access to neurosynth so i’m not sure if it’s the model or the tuning that isn’t good enough

[nocturne olive]

With good enough tuning NeuroSynth will always sound better than RVC once trained on real data

[rigid snow]

maybe it already is better but no one can actually use it that’s able to make it do better

we’ll never know

[nocturne olive]

Well we're cooking up the first organic data model, NeuroSynth-BETA-JP

Data is at 2/46

[quiet finch]

Oh amazing!

[rigid snow]

skipping rvc pretraining or no?

[quiet finch]

Will it be open-sourced

or just a demo

[nocturne olive]

Yeah no more RVC

[rigid snow]

[nocturne olive]

All directly from Neuro data

[obsidian mantle]

what does "organic" mean here

[nocturne olive]

Not synthetic as in not RVC

[quiet finch]

data

like train data?

[obsidian mantle]

oh

[nocturne olive]

NeuroSynth will be available to the public once we have English organic

[obsidian mantle]

i thought its
organic datamodel
and not
organicdata model

[rigid snow]

base neurosynth is trained on vocals made to sound like neuro with rvc (synthetic data), so not that

[nocturne olive]

We're just abandoning RVC for good and moving on to better things

[nocturne olive]

Yeah, and some of them are not that good for Neuro

[rigid snow]

[nocturne olive]

NeuroSynth-BETA-1 was very British

This might be BETA-1

[quiet finch]

RVC is limited by the quality of the base model.

and it's uncontrollable.

Is this the reason？

[nocturne olive]

RVC causes generational loss so it's better to not have RVC at all for NeuroSynth training

The generational loss is significant enough that NeuroSynth-BETA models up to 3.2 always miss any note higher than G5

[rigid snow]

tbh mostly just sounds like sloppy tuning

[nocturne olive]

True I would tune that completely differently now that I know what NeuroSynth actually wants

But the model itself is also very bad

I forgot I did this

I think this was for Arisu

I forget which tuning type this has

My guess is old tuning as new tuning should begin at Raise up your bat

[merry plank]

https://tenor.com/view/cat-jam-cat-jam-head-bounce-kitty-gif-17946989

[nocturne olive]

I think this predates new tuning by a week

[merry plank]

ugh I have to use audacity to combine vocals to instrumentals

https://tenor.com/view/audacity-the-audacity-discord-caption-discord-meme-meme-gif-16372139078567264810