data-science-and-ml

[pine escarp]

nice

im too broke to buy pro version though

[wet canyon]

this is how original and the clustered images look

but this is how it gets saved

[pine escarp]

it it me or that image is black

[wet canyon]

it is black yes

[unreal condor]

It has integrated terminal

[pine escarp]

do you get any error

[pine escarp]

oh jus like pycharm

[wet canyon]

i'm not entirely sure where it's going wrong, but i assume it has to do with the way it's getting resized again. or maybe it's an issue with cv2?

[wet canyon]

not when i try saving it with cv2. if i use matplotlib, it throws an error saying it's not the right data format or something like that, which led to me writing this line of code

clustered = np.clip(clustered / 255.0, 0, 1).astype(np.uint8)

[unreal condor]

Have you tried display the output before saving it

[wet canyon]

i have, yes. it displays it properly. it just isn't saving it properly

[unreal condor]

I think cv2.COLOR_RGB2BGR caused the problem, try different format ?

[pine escarp]

so it displays properly

but when you save it

its all black

totally unrelated but your pfp is scary man

[pine escarp]

i mean moto

@unreal condor have you tried any other IDEs that support notebooks

[unreal condor]

Too lazy to try

[fallow coyote]

You can just install the modules/libraries for jupyter notebooks on pycharm community edition. Thats what I did

[pine escarp]

one guy said that only pro version supports that

[wet canyon]

i've tried saving it without modifying clustered, i've tried it by dividing it by 255 and converting to int to get into the 1 to 255 threshold, i've also tried across cv2 and plt, nothing seems to be working so far 😵‍💫

[wet canyon]

i get that a lot lmao

[pine escarp]

well jus kidding

[fallow coyote]

I think thats for built in support. You can still install the modules for jupyter via pip (or for me, conda) and still access jupyter notebooks but you wont have any built in support from pycharm

[pine escarp]

do you mind if i try it on my notebook

[pine escarp]

ohh fair

do you have a pic of how it looks

[fallow coyote]

Even better, look up Pythan for Data Analysis by Wes Kinney and follow the instructions there

Gonna have to look online a bit to install conda on pycharms though

[unreal condor]

Then i probably don't know the problem behind it then

[pine escarp]

alright mate, thanks

[pine escarp]

it still works on conda?

@serene scaffold what shall we do

[serene scaffold]

I was on the train when we were talking and now I'm doing stuff

(I like trains)

[fallow coyote]

Yeah. Gonna have install conda from the website though. Pycharm professional, to my knowledge, provides it in built for you

[unreal condor]

Counter arguement: I hate train

[pine escarp]

same, im always blown away by how powerful their engines are

[pine escarp]

alright

[pine escarp]

no worrues

worries

what notebook do you use though

you said that you work ai company

[pine escarp]

whyyy

[unreal condor]

I have carsick

[pine escarp]

sucks

[pine escarp]

i want black theme

but with my current theme, i cant see the boxes where we write the code

[warm mortar]

Hello everybody,
Anyone an ESRGAN expert here

[spring field]

it isn't??? /s

[left tartan]

Just ask the question.

[spring field]

#python-discussion message

^

[toxic mortar]

from langchain_community.document_loaders import PyPDFLoader

loader = PyPDFLoader(
    file_path = "./example_data/layout-parser-paper.pdf",
    password = "my-pasword",
    extract_images = True,
    # headers = None
    # extraction_mode = "plain",
    # extraction_kwargs = None,
)

What does :

  extract_images = True,

Even do?

[fallow coyote]

Wait Anaconda is malware/unsafe? Ive been using conda for all my jupyter notebooks. Should I just switch to pip or another package managed?

[spring field]

no, lol, but pretty much, yes

it's rather invasive, that's all

[serene grail]

It's not dangerous or anything

^this is the reasoning

[left tartan]

There's also the scientific argument about reproducibility: that a conda forge recipe is reproducible (built from source) vs a pypi wheel that isn't necessarily

[unreal condor]

Change your theme then

[devout sail]

Hello guys , I'm new to data science and I've started with learning Python by Kaggle .

Any tips from you to follow will be appreciated ❤️

[serene scaffold]

> implying there's a paper where the reference code reproduces the findings

[unreal condor]

Learn basic ML concepts first like linear regression, logistic regression, loss function, gradient descent and maybe some traditional Algos like Naive Bayes, SVM, Decision Tree,... to know more about ML in general. After that you can start learning Neural Network/Deep Learning then maybe opt for a specialized field like computer vision or natural language processing and learn more about specialized NN structures like RNN, CNN, transformer, etc

[devout sail]

thank you so much ✨

[pine escarp]

I want pure black theme

Like that

But I can't see the box outline

[quaint rivet]

thanks

[pine escarp]

Bro

What do you use instead of anaconda

[past meteor]

Just virtual environments from the standard library can do the trick

[unreal condor]

How about less black theme ?

[lapis sequoia]

Deep learning and artificial neural networks are approaches used in machine learning to build computational models which learn from training examples.

from https://en.wikipedia.org/wiki/Neural_network_Gaussian_process

[lapis sequoia]

beautiful vis of GPs imho https://en.wikipedia.org/wiki/File:Infinitely_wide_neural_network.webm

[pine escarp]

That works, I tried it

But you know

I like the dark black theme

It looks kinda good to my eyes.

If only there was a way that I can edit the themes

[pine escarp]

Does python has a library that visualizes a neural network?

[lapis sequoia]

you can use tensorboard, netron, ..

[spare forum]

Torchviz too

[lapis sequoia]

^ probably that one is better

i wonder if the fact the NNs tend to a gaussian process over theta (params) maps to large neural networks optimising to different minima of the same quality

[dry field]

anyone pls look at #1277217265439342674

it just takes a second

[lapis sequoia]

are those magnets? just curious

[dry field]

this?

[lapis sequoia]

yeah

[dry field]

its an un-preprocessed synthetic image

I had a good result with YOLO OBB before

now i have nothing

[lapis sequoia]

if you have nothing it mustn't be the nn

[dry field]

I have the ground truth

masks + bboxes

[pine escarp]

Guys what kernel do I use for my jupyter notebook

[dry field]

ipykernel

[pine escarp]

In VS code

I switched to VS code

[dry field]

select an interpreter and that'll do

install python ext

[lapis sequoia]

what does 'nothing' mean tho?

[dry field]

ooo nothing means that the model is unable to learn

[pine escarp]

I did but it still asks me for a kernel, I used the ipykernel

[lapis sequoia]

i dont think you are describing the problem well with due respect

[dry field]

yea my bad - i misunderstood

the prediction is messed up

[lapis sequoia]

alr

doesnt vscode suggest you what install in the notifications? the bell icon bottom right

[dry field]

what's the architecture you can think of? @lapis sequoia

[pine escarp]

Lemme check, I didn't notice

[lapis sequoia]

yolo is fine for many tasks

[dry field]

😦 it doesn't work - the same issue

[lapis sequoia]

my point is that you should describe the problem with more detail in the help channel

[dry field]

you can specify the things u need

[lapis sequoia]

show the output, add an image with the current bboxes, link the model,...

[dry field]

yea i cant share the whole dataset

but gimme a moment

[lapis sequoia]

not that, just an image with the bounding boxes drawn

we don't know what is failing currently, nor how bad is failing.

[dry field]

sure gimme a moment

[lapis sequoia]

i assume you fine tuned it as well, with the classes you need, and your own dataset.

[dry field]

yep my dataset is COCO format

[lapis sequoia]

that's pretty good

[lapis sequoia]

you need to tweak so it discards the boxes that overlap

[lapis sequoia]

yolo is so neat

[pine escarp]

what should i select

[dry field]

that's why YOLO OBB worked

i dont think i can do that- many rods are inclined at sm angle

[lapis sequoia]

uhmm i dont remember off the top of my head, but try click on python envs, then try the other one XD

[dry field]

do you think if I implement RTDeTR with oIoU loss

[lapis sequoia]

i think this is about just the right type of parameters

for filtering, not for training

yolo should do fine

[dry field]

yea I think OBB type of models

woud work

[lapis sequoia]

does this help @pine escarp https://code.visualstudio.com/docs/datascience/jupyter-kernel-management

what do you expect though? @dry field that's pretty good result imho

but you may expect smth else

[dry field]

what you saw was the ground truth

they were just the bboxes

[lapis sequoia]

i meant u should show the prediction

[dry field]

there u go

[lapis sequoia]

i don't think that's the architecture

[dry field]

that's not YOLO but it behaved the same

[lapis sequoia]

my advice is try to train it for a very simple task, and check you are doing the process correctly

that's pretty good image to upload to sam

just for fun though

[dry field]

this is with mask R cnn

[pine escarp]

Yeah, does help, I'll jus use the python environments, cause for a existing jupyter server, i need to open a server and paste it's link in vscode which is a hassle

[dry field]

just select the python interpreter and vsc will do that for you

[lapis sequoia]

sounds good

[lapis sequoia]

that's great, right?

[dry field]

kinda'

but I trained it on 1/3rd of the dataset

now i have to do it altogether

[lapis sequoia]

nice

i assume training takes longer that yolo, the net is more complex iirc

[dry field]

its fine, I just want to get the job done 😢

[lapis sequoia]

imho you are doing great, just leave yolo if mask r cnn works

[dry field]

yep

thanks for the help!

[lapis sequoia]

d be nice if you can show the result afterwards

[dry field]

sure

[lapis sequoia]

ur welcome

[dry field]

the task is to detect rods in SEM images

and my approach is to create a synthetic dataset

process it and use it on real time data

[lapis sequoia]

oh, it did look familiar to me

[dry field]

so this is an unpreprocessed image

[lapis sequoia]

those are crystals from some synthetic chemistry

[dry field]

hehe

zinc-oxide

[lapis sequoia]

nice, zinc forms neat structures

[dry field]

do you think if there can be a paper out of this work

if there are results

[lapis sequoia]

it forms cute hexagons as well,

[dry field]

true but our SEM produces very bad images smtimes

and im required to make a general purpose model

[lapis sequoia]

im not at research level to tell you, imho comparing different NNs performance for estimating and separating crystals is a valuable work

sp if you also estimate the size either with the diagonal or smth

[dry field]

tru but im gonna make a website

zno-explorer.vercel.app

[lapis sequoia]

(or even the distribution of sizes)

[dry field]

my website works on good images at a specific zoom level

good meaning- good contrast

[lapis sequoia]

i do know of people doing similar stuff wo NNs though, but still, i think is great stuff

yeah nice, u using,,,what was it

ONNX ?

[dry field]

it is running detectron on the backend

[lapis sequoia]

oh runs serverside

[dry field]

https://zno-explorer.vercel.app/ yep tunneled to ~my~ lab

sry its not my lab, its the uni's lab

[lapis sequoia]

nice, oh yeah i forgot this is 3d for a second

[dry field]

im only 19 to have one xD

[lapis sequoia]

so wait the height and width need to consider perspective somehow

[dry field]

yea there's a txt file has sm params - from the SEM

that has to be uploaded to wbsite too

[lapis sequoia]

i see

[dry field]

real height is actually hard to measure and i have abs no idea

[lapis sequoia]

u learning react for the plots?

[dry field]

no

i knew some react

[lapis sequoia]

oh it's plotly and js

[dry field]

yea

but yea it kinda; sucks for now

[lapis sequoia]

there are good react libraries for it, but if you dont need it thats fine

[dry field]

im focusing on training the detection

[dry field]

tru but yea will get back to it later

[lapis sequoia]

makes sense, that's the difficult part

gl !

[dry field]

thanks!

[pine escarp]

guys

https://cdn.discordapp.com/attachments/366673247892275221/1277225487365570590/UCiISP6Qk1RYJaxTAQAAYH3KtwJ6nidLowVF6KiOlUemAgAAwLqlluZeyNLoquHWAwAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA..png?ex=66cc646d&is=66cb12ed&hm=dda14a784b0d444137288810f37a0ec15282596e95816d883c334c3b0cb104e3&

what is this

[dry field]

thats a synthetic image

[lapis sequoia]

imho to make it for a paper it needs this minimum:

1. be reasonably accurate or measure some error (you do include confidence, maybe other metrics as well.)
2. let user select different nets
3. has a comparison of NN results for a specific set of images
4. have some parts open source.

jic it helps.

[dry field]

doesn't mean anything

[dry field]

1. there are people to measure
2. let's see if the task can be done with 1
3. alr tried YOLO OBB, detectron, SAM
4. sure the frontend can be opensource xD

[lapis sequoia]

the full thing has to be for publishing id say.

(i.e reproducible.)

[pine escarp]

doubt, what are you trying to achieve with that image

[dry field]

the focus is on the dataset- so it can be closed and be open for a few people with consent

[pine escarp]

ohhh

so the image doesnt mean anything?

[dry field]

yep

also the simulator can be opensource- which creates the whole scene

[buoyant vine]

🤨

[errant bison]

Where can i find events related to ai?

Like seminars or workshops for what is ai, etc

[proper crag]

is 4.6 GHz clock speed good for training locally?

also gonna deploy the model to a docker

then feed it a livestream data from my VM

ML model

[analog marsh]

guys is there tool or something to use so i can customize ai for my website
i want to ask user question based on them im gonna suggest stuff but i want to it using AI if that make sense

[lapis sequoia]

reading this if smone wants to join https://en.wikipedia.org/wiki/Bayesian_inference

pretty neat article

[unreal condor]

umm no

[jaunty helm]

like a cpu? if by model you mean a LLM, hell no
if you don't want it to train for your entire lifetime, you need a gpu to train a LLM

[proper crag]

ML model

SVR

bruh i didn do LLM

[jaunty helm]

depends on how much data do you have

iirc the time complexity for a SVM is like O(n^2 * d) or O(n^3 * d) where d is how many features you have

[lapis sequoia]

from google

Linear SVM has a complexity of O(n d), making it suitable for large datasets with a moderate number of features. Non-linear SVM with kernel tricks can have complexities of O(n^2 d) or O(n^3), which can be computationally expensive for large datasets.

but it's doing regression in his case, ie SVR

[jaunty helm]

don't think the time complexity differs between the classifier and regressor

anyways, if you have >= like a million data points then SVR won't run very well

in fact 100k probably already runs pretty slow

[pine escarp]

Most deep learning stuffs require GPU

I did CNN with my potato laptop

Took around 9 hours

😂

AND GUESS WHAT? I DIDN'T SAVE THE PERFORMANCE HISTORY ☠️

Now I have to train it again jus to get that training history

This time ill save it with pickle

But man

I don't wanna wait 9 hours

Is there like cloud or online thing

Where I can run it

[jaunty helm]

colab / kaggle gives you free gpu per week

[lapis sequoia]

why dont you do it by hand

[pine escarp]

Yeah noticed that

Are they fast

[jaunty helm]

faster than your cpu, that's for sure

[lapis sequoia]

you can save checkpoints to gdrive as well, jic it breaks/run out, and continue next day

[unreal condor]

much much faster

[lapis sequoia]

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

not a low goal this one:

"Divine Benevolence", or an Attempt to Prove That the Principal End of the Divine Providence and Government is the Happiness of His Creatures (1731)

not a minor fact id say that a lot of people these days are bayesian (and atheists), and this was one of his motivations (apparently)

Others speculate he was motivated to rebut David Hume's argument against believing in miracles on the evidence of testimony in An Enquiry Concerning Human Understanding.

[pine escarp]

I'll try it for sure then!

[pine escarp]

Enlighten me.

[lapis sequoia]

Bayes doesn't help if you take the wrong evidence, which ig is pretty obvious

In other words, bayesian thinkers may be too certain of being selecting the right evidence (especially in complex environments.)

[woven sundial]

which format is better .h5 or .pb?

because idk if i should use keras or not

[lapis sequoia]

.h5 in keras saves the weights onl, and .pb isnt really encouraged anymore from keras.
Use .keras format, if you need it for the web, convert to .onnx.
There are caveats, that's a short opinion.

Guide: https://www.tensorflow.org/guide/keras/serialization_and_saving @woven sundial

[woven sundial]

i want to make ai photo denoiser and idk which format should i use

[lapis sequoia]

.keras

[woven sundial]

alright thanks

[midnight orchid]

It’s doing everything but eat the food grr

[past bramble]

day 1 kaggle report:
started out with a deep learning course although i know it cuz it's been a long time i used those concepts

[lapis sequoia]

gl :-)

[sterile heath]

Have you considered the possibility of it not being hungry?

[odd meteor]

• Social media
• Eventbrite app
• Joining AI communities

[woven sundial]

so i can use model.load_weights('.h5') and load_model at once?

[whole pendant]

what is degree of freedom in t distribution ?

oh nvm i got it

[umbral surge]

Are there any convenient models that can output sequences of varying length? Like if you have trajectories where there is a varying amount of missing points at some place in the trajectory, can this be done? And in a way that makes use of the information available after the missing segment, i.e., not just taking the preceding points and making predictions without using the context of the points following the missing segment

[lapis sequoia]

I assume you are using tf>2.16 and keras 3.X
Load weights requires a definition of the architecture in the code; load model doesn't, it loads architecture + weights (model.)

If it's a more complex problem I'd add in in #1035199133436354600, with a detailed description of usage, versions, etc. I don't mind to help further (if i can.)

[real furnace]

Very complex stuff here man but its slowly coming together

New UI

[pine escarp]

What does it do?

[real furnace]

Well it plays roblox and since roblox isnt just one set game it is a community of millions of games I made it capable to determine the objective of the game based on anything presented on the screen and im planning to add even more

Eventually want to get a universal game player ai software and just have different models based on the games

I plan for this to be open-source and people can pick it apart and make it even better and maybe I should just make models for each independent game

@pine escarp

[pine escarp]

That's nice man.

Good luck with it.

[real furnace]

would help for space as its near 1GB already 😭

[pine escarp]

[real furnace]

Thanks I will definitely try my best

The idea is to just have my own army of ai gamers

[pine escarp]

Keep us updated man.

[pine escarp]

[real furnace]

Alot of money spent on just cloud stuff if I were to but.

I think the only thing is just doing ai with 3d games is just a hassle because it needs to percieve everything as if it was a human which isnt that hard but it can be

and im sorts making it so it can interperet anything and everything to learn

maybe I just test out some simple games for it to play first but

ionk im striving for gold here

atm it is capturing gameplay but playing the game is where its getting tripped up with

[pine escarp]

dw, you can fix it

[real furnace]

and I could make life easier and just have it inject into the game but its something I dont want to rely on

as that would mean having to bypass anticheats and stuff

[small wedge]

there have been models with relative success on a variety of games (including 3d) that just take the screen display as input and shove it through a cnn, which doesn't require you to hook in but those are a lot more difficult to get good results from and require a lot more resources to train properly.

I always just make my own little games in unity then it's easy to hook in, you could make a tiny clone of whatever game you want it to play and not worry about fucking around with cheatengine or whatever you're using to monitor memory on existing games

[real furnace]

The thing is I’m trying to train it to be able to play ALL Roblox games the only factor being that they are Roblox games but they vary as normal games at this point with the quality they are at

Which makes the project a whole lot harder with the because it’s supposed to interpret multiple genres

But I have a idea that will make it possible

Creating different models for each genre/ game would make it be able to understand what’s happening it the game much easier

[lapis sequoia]

nice ai vibe to have in the background :-) https://www.youtube.com/watch?v=hTrIHjJRYVg

[unreal condor]

This is such an ambituos project, good luck lad.

[fiery bane]

Here are some events in no particular order
https://neurips.cc/
https://iclr.cc/
https://icml.cc/
https://cvpr.thecvf.com/
https://aaai.org/
https://iccv2023.thecvf.com/
https://2024.aclweb.org/

[fiery bane]

what are you doing?

[lapis sequoia]

introduction to DL without heavy maths https://developer.nvidia.com/blog/deep-learning-nutshell-history-training/

[whole pendant]

What's DL

[midnight orchid]

Training ai or smth if it hit itself it will punishment if it food it good but I think he thinks if he survives longer his points are gonna be higher (which is)

[lapis sequoia]

le cun's YT channel https://www.youtube.com/@YannLeCunPhD/videos

[pine escarp]

Deep Learning.

[fiery bane]

like RL?

DL RL?

[lapis sequoia]

would NNs learn to think trained with brain patterns

[small wedge]

there are like 50 pages of discussion that could happen on that question lol

[lapis sequoia]

k

have you read minsky, kurtzweil and those people?

[small wedge]

nope

do they have opinions on the topic?

[lapis sequoia]

nope

[small wedge]

who are they

[lapis sequoia]

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

Minsky also built, in 1951, the first randomly wired neural network learning machine, SNARC.

also invented the confocal microscope etc, but thats unrelated

he apparently attacker perceptrons (considered part of cause for AI winter.), i suspect why:

Minsky's book Perceptrons (written with Seymour Papert) attacked the work of Frank Rosenblatt, and became the foundational work in the analysis of artificial neural networks. The book is the center of a controversy in the history of AI, as some claim it to have had great importance in discouraging research of neural networks in the 1970s, and contributing to the so-called "AI winter".[27] He also founded several other AI models. His paper A framework for representing knowledge[28] created a new paradigm in knowledge representation. While his Perceptrons is now more a historical than practical book, the theory of frames is in wide use.[29] Minsky also wrote of the possibility that extraterrestrial life may think like humans, permitting communication

he may believed that NNs were only learning statistical patterns, but not deeper concepts.

[small wedge]

Interesting, I think that's sort of a distinction without a difference

statistical patterns can represent "deeper concepts"

[lapis sequoia]

(he is frozen now)

[small wedge]

LOL

[lapis sequoia]

XD

uhm...i point to what's described here:

[small wedge]

but to clarify when you say "if you train an nn on thoughts could it learn to think" what kind of model exactly are you talking about (like a seq-2-seq thing?)

and what does it mean to think in this context, does that imply it's sentient? or just that it can predict a sequence of thoughts as represented in brain scan data?

[lapis sequoia]

i meant if the signal to learn would be the appropriate to learn to learn, in a way

i think surface statistics is what they describe here https://arxiv.org/pdf/2012.05208

[small wedge]

I think I understand what you're saying

like they don't get meta statistics

[lapis sequoia]

give it an input => predict a neural pattern (parts of it)

the input is the same that was given to the human that produced the neural pattern.)

[small wedge]

but it does doesn't it, isn't that why when you transfer learn or fine tune an LLM you always train the last few layers, because the early layers contain the "syntactic understanding" or the first order statistucs and the deeper layers contain the "semantic" or more complex/meta understandings, training the first layers creates instability and causes massive changes in the later layers leading to catastrophic forgetting

[lapis sequoia]

or give the neural pattern => produce the output, or both ways

im unsure that an LLM wouldn't learn deep thinking though, if that were the case

they confuse really simple things

[small wedge]

I would agree with the notion of the paper you sent

[lapis sequoia]

there are hundreds of those examples

[small wedge]

modern ML architectures are nowhere near as complex as brains

thinking at the level we do may require a more complex approach

[lapis sequoia]

yeah, ig, just rambling anyways, for fun

[small wedge]

ofc ofc, it's all conjecture but it's fun to talk about

[lapis sequoia]

:-)

[small wedge]

last I looked into this I was reading about spatiotemporal architectures like liquid state machines and spiking nn's

[lapis sequoia]

nice, just know the latter

[small wedge]

that account for the change of time in their threshold functions more like biological neurons

[warped arrow]

I tried to do "import openai" and I put in the terminal "pip install openai" but in my code it is put that openai its not defined

[small wedge]

they are very similar, I think LSM is a type of spiking nn

or the other way around one, been a while since I read on them

[lapis sequoia]

i see. hinton mentions this

that account for the change of time in their threshold functions more like biological neurons
i think

[small wedge]

!dashm

[arctic wedge]

Install packages with `python -m pip`

When trying to install a package via pip, it's recommended to invoke pip as a module: python -m pip install your_package.

Why would we use python -m pip instead of pip?
Invoking pip as a module ensures you know which pip you're using. This is helpful if you have multiple Python versions. You always know which Python version you're installing packages to.

Note
The exact python command you invoke can vary. It may be python3 or py, ensure it's correct for your system.

[small wedge]

oh yeah I see the paper you sent mentions spiking nn by name lol

didn't even realize

but it's an interesting question about surface statistics, like even with humans it seems our "meta thoughts" i.e. language that describes situations and experiences has a huge impact on our ability to learn and conceptualize things

[warped arrow]

Traceback (most recent call last):
  File "c:\Users\ihab\Downloads\chatgpt1.py", line 17, in <module>
    réponse = interroger_chatgpt(question)
              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "c:\Users\ihab\Downloads\chatgpt1.py", line 7, in interroger_chatgpt
    response = openai.ChatCompletion.create(
               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\ihab\AppData\Local\Programs\Python\Python312\Lib\site-packages\openai\lib\_old_api.py", line 39, in __call__
    raise APIRemovedInV1(symbol=self._symbol)
openai.lib._old_api.APIRemovedInV1:

You tried to access openai.ChatCompletion, but this is no longer supported in openai>=1.0.0 - see the README at https://github.com/openai/openai-python for the API.

You can run `openai migrate` to automatically upgrade your codebase to use the 1.0.0 interface.

Alternatively, you can pin your installation to the old version, e.g. `pip install openai==0.28`

A detailed migration guide is available here: https://github.com/openai/openai-python/discussions/742 

Why this error ?

[small wedge]

it tells you

[small wedge]

read it

[warped arrow]

Okay i'm going to test "pip install openai==0.28"

[small wedge]

remember to use -m if you fixed it like that last time

[warm mortar]

Hey everyone,
Does anyone use ESRGAN on google collab???

[warped arrow]

Traceback (most recent call last): File "c:\Users\ihab\Downloads\chatgpt1.py", line 17, in <module> réponse = interroger_chatgpt(question) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "c:\Users\ihab\Downloads\chatgpt1.py", line 7, in interroger_chatgpt response = openai.ChatCompletion.create( ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "C:\Users\ihab\AppData\Local\Programs\Python\Python312\Lib\site-packages\openai\api_resources\chat_completion.py", line 25, in create return super().create(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "C:\Users\ihab\AppData\Local\Programs\Python\Python312\Lib\site-packages\openai\api_resources\abstract\engine_api_resource.py", line 153, in create response, _, api_key = requestor.request( ^^^^^^^^^^^^^^^^^^ File "C:\Users\ihab\AppData\Local\Programs\Python\Python312\Lib\site-packages\openai\api_requestor.py", line 298, in request resp, got_stream = self._interpret_response(result, stream) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "C:\Users\ihab\AppData\Local\Programs\Python\Python312\Lib\site-packages\openai\api_requestor.py", line 700, in _interpret_response self._interpret_response_line( File "C:\Users\ihab\AppData\Local\Programs\Python\Python312\Lib\site-packages\openai\api_requestor.py", line 765, in _interpret_response_line raise self.handle_error_response( openai.error.InvalidRequestError: The model `gpt-4` does not exist or you do not have access to it.

I can't read even with google translate

[small wedge]

The model gpt-4 does not exist or you do not have access to it.

are you sure you have access to gpt-4? are you paying a subscription?

[left tartan]

We've already explained that your code is incomplete, it has nothing to do with colab or Esrgan

[left tartan]

Do any other models work? Like 4o turbo? https://platform.openai.com/docs/models/gpt-4-turbo-and-gpt-4

[fiery bane]

I reccomend reading about free energy principle

[warm mortar]

So could you compelete it for me??

[left tartan]

No, we're here to help you learn. Not to write code for people.

[warm mortar]

Do you know someone??

[left tartan]

You could try to hire someone on fiverr

[tight wave]

I want to build an end to end ml project in python, my ml skills are intemediate, but want to expand my skill, what are some fun projects i can try?

[lapis sequoia]

thank you ! :]

do you think there is enough evidence to support the existence of a god @fiery bane ?

do you know about 'ideasthesia'? it's related
do you know francisco varela?

[warm mortar]

I don't use

[verbal oar]

I offer help in upwork, but this is more of mentoring than freelance jobs

where should I promote on social media like linkedin this is not effective

maybe also career discussion related I mean as channel not my offer
I can help with building project but not code for you

[fiery bane]

kaggle? AICrowd?

[fiery bane]

Like, depending on purpose, we demand different level of rigor.
With that being said, I'm a theist, so I guess that's a yes for all practical purposes.

[fiery bane]

I heard about ideasthesia, idk how it is related to free energy principle

[lapis sequoia]

does consciousness count within what the representations that minimise surprise are? (ideasthesia is one interesting way of those representations.)

[fiery bane]

francisco varela
not personally. I saw his wiki before. I still prefer Karl Friston's approach

[lapis sequoia]

(i think donald hoffman proposes smth like that.)

i see. he is dead though, died quite young

it seems he was studying the same problem

[fiery bane]

I think these are two different things. Like, I don't see why consciousness is necessarily related to self-organisation. I mean. It might be so, but I haven't seen a convincing argument as to why

[lapis sequoia]

btw i asked cuz your name seemed chilean, smhow

[fiery bane]

lol my name is Italian

but I am not

https://gunslingergirl.fandom.com/wiki/Beatrice

the name is literally from a manga lol

[lapis sequoia]

oh, not too close, though many italian moved to arg at least in the 1900s

[fiery bane]

https://gunslingergirl.fandom.com/wiki/Beatrice https://gunslingergirl.fandom.com/wiki/Bernardo

[lapis sequoia]

is that you writing?

[fiery bane]

No lol, it is just one of my favourite manga

[lapis sequoia]

u do sound like a cyborg but not completely unemotional

k, never heard that before

[fiery bane]

I hang out in #VC quite a bit, so you can literally listen to me

[lapis sequoia]

I don't stand voice chats, so wont hear it ig.

The name probably mutates to Beatriz in South America

btw, there exists Bernarda

I don't think Beatrice is the opposite

Beatriz is a Spanish, Galician and Portuguese female first name. It corresponds to the Latin name Beatrix and the English and Italian name Beatrice.

I meant from this part:

was originally introduced by Karl Friston as an explanation for embodied perception-action loops in neuroscience.

That consciousness could be what the brain creates as an approximation to a reality that we can't know by experience

My interp. is that this is what Donald Hoffman says, i may be wrong

Apparently Bayes wanted to develop a theory of why god exists, counting miracles as evidence, when he developed the rule.

He should count suffering as well, though.

[fiery bane]

I think there are many interpretation haha.
Not sure what exactly Donald said, not familiar with him, but I won't be surprised if lots of people pointed to it and say "consciousness".
I personally don't really like that approach, I still think that preception-action loops can exist without conscionsess.
I think Karl himself only hinted at that, and not go as far as making that claim.
I think the biggest issue with consciouness is that we haven't properly defined it. How can we properly talk about something and investigate its causes, if we don't have a good definition of it.

[lapis sequoia]

fair !

[past bramble]

day 2 kaggle report: completed deep learning course

[lapis sequoia]

nice

[past bramble]

it suggested that i could try building image generation model

that's possible only through neural network?

[agile cobalt]

just to make sure: can you clarify what exactly are you asking?
(... is only possible via ... || as opposed to ...)

[ocean pawn]

Currently trying to implement MLP without using framework, but it can't even fit a straight line and the loss is somehow increasing

[lapis sequoia]

u wanna share the code?

[ocean pawn]

I would love to, but it might be a lot. Thanks!

I'll see

[lapis sequoia]

looks like a relu

[ocean pawn]

Yes, a single relu layer

[lapis sequoia]

uhmm...did you print the weights?

i expect it's collapsing

[spring field]

your entire network is just a relu?

[ocean pawn]

Sorry, I mean a single relu hidden layer

with 10 neuron

!paste

[arctic wedge]

Pasting large amounts of code

If your code is too long to fit in a codeblock in Discord, you can paste your code here:
https://paste.pythondiscord.com/

After pasting your code, save it by clicking the Paste! button in the bottom left, or by pressing CTRL + S. After doing that, you will be navigated to the new paste's page. Copy the URL and post it here so others can see it.

[lapis sequoia]

yeah, so i'd bet most weights are 0, or you are facing the dead neurons (i.e the a(x) has negative x for most inputs.)

can you remove the relu, and leave the weights only?

(that's basically a linear regression.)

[ocean pawn]

https://paste.pythondiscord.com/HWGA

[spring field]

did you normalize the inputs?

[ocean pawn]

https://paste.pythondiscord.com/35CA

[ocean pawn]

I did not, but I tried, it didn't make a difference

[lapis sequoia]

you've improved quite fast

[spring field]

you need to normalize the inputs anyway

[ocean pawn]

Honestly, I got bored for a while and watched cartoon instead for the last week, just got back to it recently

[ocean pawn]

Ok!

[lapis sequoia]

which cartoon

[ocean pawn]

A lot, actually

1. spy x family seasson 2
2. Yona of the dawn
3. Apothecary Diaries
4. Digimon

yeahhhh, I took a long break

Anyway, I am back

[lapis sequoia]

nice

i wonder whether one could debug without reading the code first, doesn't look bad imho, at least the call method

[ocean pawn]

a single layer with a single neuron is still acting weird

[lapis sequoia]

if you comment out the layers, and use 1 dense the 5->1 does it improve?

[lapis sequoia]

k, then we can simplify the problem

[ocean pawn]

Same

[ocean pawn]

 self.layers = [
            Dense(1, 1, key=key3),
        ]

[lapis sequoia]

yeah, that's good

[ocean pawn]

Loss/Cost is going mad tho

[lapis sequoia]

are you normalising like it was suggested?

that's an important step

what is your LR?

all the parameters for the optimiser, and which optimiser?

[spring field]

is your learning rate 10?

[ocean pawn]

I tried Gradient Descent and Adam

[ocean pawn]

Yes, acutally might change that, but I tried 0.1 before and it still broke

[lapis sequoia]

yeah, that's likely it

[spring field]

try with 0.001

[lapis sequoia]

note that the update is getting crazy, because your are jumping from one side to the other of the parabola

(so the sign changes, and it oscillates)

[past bramble]

i thought image generation uses other algorithms or concepts, not neural network

[ocean pawn]

Same graph

But the cost is stable

[lapis sequoia]

can you show the loss plot?

[ocean pawn]

But the prediction seemes to be the same

[ocean pawn]

Wait

[ocean pawn]

I will try and normalize it

My old normalization code is broken for nn so might have to take a while

[lapis sequoia]

yeah, you've -3000 to 3000 right?

[spring field]

you can procedurally generate images, there are tons of approaches, you can also use neural networks, yeah, a simple one might be VQ-VAE or some other VAE architectures, then there are also GAN architectures

[lapis sequoia]

that can blow up the squared error, if that's the loss

[spring field]

did you normalize the inputs?

[lapis sequoia]

i think the plot shows it didnt

it goes from -3K to 3K

[spring field]

I mean, they could've scaled the values back to the originals

[ocean pawn]

I did that, but it's weird

This looks better

[ocean pawn]

Might be a plotting issue

[verbal oar]

try grid search or random search

[lapis sequoia]

are you training with the normalised values? the loss is too large

[ocean pawn]

Yes

Unless my normalizer is broken

[spring field]

if you could share the code...

[ocean pawn]

Sorry, I'll try

But I've seperated the code quite a lot

[spring field]

no need to apologize 😅

[ocean pawn]

!paste

[arctic wedge]

Pasting large amounts of code

If your code is too long to fit in a codeblock in Discord, you can paste your code here:
https://paste.pythondiscord.com/

After pasting your code, save it by clicking the Paste! button in the bottom left, or by pressing CTRL + S. After doing that, you will be navigated to the new paste's page. Copy the URL and post it here so others can see it.

[ocean pawn]

https://paste.pythondiscord.com/6F6Q

https://paste.pythondiscord.com/A5DQ

# @jaxtyped(typechecker=typechecked)
@jit
def __normalizer(
    x: Float[ArrayLike, "..."],
    argnums: Optional[Iterable[int]] = None,
    *,
    x_train_mean,
    divisor,
) -> Float[Array, "..."]:
    if argnums is not None:
        argnums = jnp.array(argnums)
        divisor = jnp.take(divisor, argnums, axis=0)
        x_train_mean = jnp.take(x_train_mean, argnums, axis=0)
    return (x - x_train_mean) / divisor


# @jaxtyped(typechecker=typechecked)
def get_z_score_normalizer(
    training_data: Float[Array, "data_count feature_size"]
) -> tuple[NormalizerFunction, NormalizerFunction]:
    x_train_mean = jnp.mean(training_data, axis=0)
    x_train_std = jnp.std(training_data, axis=0)
    jax.debug.print("{} {}", x_train_mean, x_train_std)
    return jit(
        partial(__normalizer, x_train_mean=x_train_mean, divisor=x_train_std)
    ), jit(partial(__invert_normalizer, x_train_mean=x_train_mean, divisor=x_train_std))

[lapis sequoia]

why not to just generate between -1 and 1 for now?

key, x, y = generate_data(
    key,
    (1000,),
    -1000.0,
    1000.0,
    lambda x: 3 * x + 1.0,
    -20.0,
    20.0,
)

[spring field]

you need to normalize y as well

[ocean pawn]

Huh, for regression, I've only normalized x but never y, I'll see

[ocean pawn]

I have no idea what's going on

My only theory is that I might've implemented the vectorization incorrectly (the matmul) but I have tried a few arbitrary data and it seemed to be correct

The gradient calculation should be done by JAX and it is probably working as expected

The weirdest part about this is that I've implemented the same thing with Equinox before and it worked, it only break if I tried to make it from scratch, so I definitely did something wrong

self.layers = [
            Dense(1, 5, key=key3, activation=jax.nn.relu),
            Dense(5, 1, key=key4),
        ]

I am very much confused

[spring field]

the y doesn't appear to be normalized here?

[ocean pawn]

Yeah, still haven't done that, will do now

[lapis sequoia]

if you multiply by -1 it's not too bad XD

[ocean pawn]

[ocean pawn]

Normalized y (then inversed it for plotting)

I also did this in Equinox and it worked without normalization with the same architecture

[spring field]

do you have sigmoid as the last activation function btw?

[ocean pawn]

No, it should be linear

[lapis sequoia]

this isnt that important, but i'd use mean squared error for loss

[ocean pawn]

OH I tried both

Both of them is broken

[lapis sequoia]

you tried now?

[ocean pawn]

Originally it's MSE, I changed it to MAE but it's broken either way

With mean squared error

The loss curve looks better with MSE tho

[spring field]

it should be a sigmoid, your output needs to be normalized as well

although maybe it doesn't actually matter

[lapis sequoia]

you dont need a sigmoid for 1 layer and linear data

[ocean pawn]

Huh, I though sigmoid is only for binary classiflication?

[lapis sequoia]

not even for multiple layers, but it collapses to a linear prediction

in the exit layer, yes, in the middle layers, no

[ocean pawn]

You're better off using ReLu in hidden layer right?

[spring field]

it depends, language models usually use GeLU

VAEs might use Tanh

[lapis sequoia]

u could share a full colab right

[spring field]

LeakyReLU might be better than ReLU too

[lapis sequoia]

that we can test

[ocean pawn]

I can give you the github link

But I forgot to commit for a week (oops)

I'll update it now

[lapis sequoia]

i'd not share ever a personal link here, but it's your choice

[ocean pawn]

Why not? Just curious, isn't github account public, anyway?

[lapis sequoia]

yeah, but linked information scares me

but it could be a random account, or yours if you are fine w it

[ocean pawn]

Oh fair enough

[lapis sequoia]

if the matmuls are correct, and the input data is fine, then assuming that the initialisation of weight is correct, one place to look at is the gradient

alr, i might open a codespace

[ocean pawn]

Thanks!

There's quite a lot of irrelevant regression code, so you can ignore at least half of the code

[lapis sequoia]

oh, i ran out of codespaces data

cuz i left many opened

[ocean pawn]

Try google IDX I suppose

No gpu is a bit sad

[lapis sequoia]

never heard abt that

[ocean pawn]

Preety much codespace but google and Gemini

It's still vscode

[verbal oar]

how's your experimenting with kernel method in svm, this is very interesting topics hmm related with operator theory

[lapis sequoia]

in the code you showed us, did you write gradient descent, or is a jax built in?

[verbal oar]

I assume this creates hilbert space?

[ocean pawn]

So I've tried gradient descent
But I am using optax.Adam currently

[lapis sequoia]

seems similar, idk enough though

[verbal oar]

you also said about muddy waters like inner product spaces

[lapis sequoia]

like the space <phi(x), phi(x')> is means its equipped with inner product right

[lapis sequoia]

right, so i got lost with many files in the repo

[ocean pawn]

Sorry, just look at example/nn.py

[verbal oar]

yeah with this <> notation

[ocean pawn]

And ml/nn.py

[verbal oar]

ah these things related with norm

[lapis sequoia]

idx looks neat

[ocean pawn]

It is neat, they promised android studio too

(I hope PyCharm get added after android studio)

[verbal oar]

interesting but also I suppose challenging topic

generalization of vector space hmm so does it module in abstract algebraic sense?

[devout sail]

can't really understand what should I do here , any help please ?

[spring field]

what are you supposed to do here?

[devout sail]

the same question that I'm asking lol
it's supposed to do logging there based on using Booleans and conditionals statements,

[spring field]

it looks like the code is already written for you?

[devout sail]

oh really lol , how dumb am I 😭

[serene scaffold]

hello, this question is for data science questions. questions like this would be better suited to a thread in #1035199133436354600

[devout sail]

isn't this #data-science-and-ml ? and the whole category is called Topical chat/ Help . so why even there is a python-help thread rather than this category was only for chat ?

I mean no offense just I'm confused now.

[serene scaffold]

your question about finding the remainder after integer division is not about scientific programming or machine learning.

[devout sail]

but still it's about python's usage in data science overall , isn't it ?

[serene scaffold]

it is not.

[devout sail]

ok then.

[serene scaffold]

is there a reason you thought it was? are you taking a data science course?

[devout sail]

I'm walking through the python course on Kaggle so I can learn data science. but still I'm just a beginner

[serene scaffold]

sounds good. sounds like you're still in the "absolute basics of python" section of the course. you're always welcome to come back to this channel when you start using things like numpy or pandas.

[devout sail]

thank you.

[lapis sequoia]

@ocean pawn

idk jax so asked chat gpt to make sure new weight is used. rest are silly mods by me to find the error.
(to be clear ur code was correct just didnt use the new weights)
https://paste.pythondiscord.com/H5IA

[ocean pawn]

HUHHHHH

What changed

I AM STUPID

I FORGOT TO COMBINE THE NEW PARAMETER

THANKS!

model = combine(params, static)

this single line addded to 106 fixed it

I am annoyed at how simple this fix is

At least it work now

Imagine training new parameter and forgot to use them, definitely not me!

@lapis sequoia Thanks for your help, without you, I doubt I'll realise this

I guess the solution to a problem is always the easiest one

Thanks, anyway!

[violet gull]

why does ML use cuda instead of a compute shader?

[lapis sequoia]

my pleasure

[ocean pawn]

Thanks again!

Guess now's the time to implement Adam myself train MNIST dataset

[chrome ermine]

I was told that this question didn't fit in #algos-and-data-structs and to post here instead.

So, the question. I'm trying to get alpaca dataframes to play with backtesting.py. I've tried to trim and match the dataframes as best as I could get but I'm now stuck.

Link to the paste.
https://paste.pythondiscord.com/U4FA

[serene scaffold]

@chrome ermine can you do print(df.head().to_dict()) and put the resulting text in the paste bin, for each dataframe that pertains to this question? and make sure there's a comment to say which is which.

@chrome ermine ignore my previous message

you have data = get_stock_data. what type is get_stock_data? please ping me when you reply.

[chrome ermine]

It's the 2nd file. Alpaca api stock bar data.

[serene scaffold]

an object that represents a python file is a module.

    bt = Backtest(data, RsiOscillator, cash=10000)
         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\Me\AppData\Roaming\Python\Python312\site-packages\backtesting\backtesting.py", line 1043, in __init__
    raise TypeError("`data` must be a pandas.DataFrame with columns")
TypeError: `data` must be a pandas.DataFrame with columns

A module is not a dataframe.

[chrome ermine]

The ouput from the module is a dataframe. I did use an variable to make it usable. I'm assuming I did it wrong?

[serene scaffold]

modules don't have outputs. modules are just containers for all the variables (including classes and functions) that are defined in the module when it finishes running.

[chrome ermine]

Should I have used a 'return' or a file ouput?

[serene scaffold]

you can't return from a module.
what variable in get_stock_data is the one that you think of as the "output"?

[chrome ermine]

huh. you got a point. I should have used a function call at the end then. That way the modual call pass a the function instead of trying to use it as a variable...

[serene scaffold]

what is your answer to my question?

[chrome ermine]

def get_stock_data(params):

[serene scaffold]

to avoid confusion, I recommend that you not use the same name for a module and something in the module.

if you had a module named foo.py that contained a variable (which includes classes and functions) named bar, you would do import foo in another file, and then get that function as foo.bar

# foo.py
a = 5
b = 3

def bar(x):
    return x * 2

# baz.py -- THIS IS A DIFFERENT FILE
import foo
print(foo.bar(foo.a))
print(foo.b + 1)

@chrome ermine what do you think the output of this program is?

[chrome ermine]

10
4

[serene scaffold]

correct.
think of how you can correctly import and use variables that you defined in get_stock_data.py to fix the problem that data is not a DataFrame.

[chrome ermine]

yea, this is where I get lost... I guess it's jus back to to python course I guess.

[serene scaffold]

is stock_data not the dataframe that you want?

[chrome ermine]

stock_data was the attempt to match case. I know I went wildy wrong somewhere. I just didn't know where.

[serene scaffold]

how do you know you went wildly wrong? because of "TypeError: data must be a pandas.DataFrame with columns", or because of something else?

[chrome ermine]

It started with the error. And then little by little, I tried to match case but kept getting the same error. Then after everything lookeed almost exact, minus the hours, I got stumped and looked for outside guidence.

[serene scaffold]

the reason you're getting "TypeError: data must be a pandas.DataFrame with columns" has nothing to do with whether or not stock_data is correct.

the code that causes that type error never even looks at stock_data.

[chrome ermine]

I know that now. I called everything before stock_data cause stock_data was never called. You showed me that

[serene scaffold]

you don't call stock_data. stock data is not a function. it is a dataframe.

[chrome ermine]

It's a variable I know. but it was never "referenced"? Would that be the correct jaragpn? I'm still new

[serene scaffold]

"referenced" is right

import get_stock_data

data = get_stock_data

all this does is make data a variable that refers to the get_stock_data module. and modules don't have output. so data is just a container of variables, like foo in my earlier example.

[chrome ermine]

I get that now. Which I why when i run get_stock_data by it'self I get the correct data frame but not when i run back_test

[serene scaffold]

if you have get_stock_data, the module, how do you get the dataframe that you want to pass to Backtest?

[chrome ermine]

put stock_data into a function with the dataframe ecorrections and call the function in back_test

[serene scaffold]

that would work, but it's unnecessary extra steps

for one thing, delete the data = get_stock_data line.
then do this
print(get_stock_data.stock_data)

[chrome ermine]

It ptinted correctly

[serene scaffold]

yes

what does this tell you about how to reference stock_data in back_test.py?

[chrome ermine]

ok. So I just call the variable. data = get_stock_data.stock_data

[serene scaffold]

you do not "call variables". the word "call" has a specific meaning in programming.
data = get_stock_data.stock_data would cause data to be another reference to the stock_data dataframe that is defined in get_stock_data.py.

@chrome ermine do you understand?

[chrome ermine]

I get it

I'm just reusing references

[serene scaffold]

right

does your code work now?

[chrome ermine]

nop

*nope

I get a suite of new errors now

but the issue I had is working

so... yes?

[serene scaffold]

getting a new error is always cause for celebration

(ie, a new error following a previous error; the first error isn't fun)

(because usually, a new error means that you solved the original error. but it occasionally means that you've descended to an even lower level of inoperable depravity.)

[chrome ermine]

Progress is progress

[serene scaffold]

and inoperable depravity is depraved.

[chrome ermine]

But a cart with a broken wheel can still be moved with the right weight distribution

[serene scaffold]

I would just pick up the whole thing and carry it to prove what a man I am.

[chrome ermine]

lol

The original issua has been solved. i now have to get to bed. I have work in the morning. I'll deal with new issues after wrk. thanks for your enlightenment

[serene scaffold]

I changed my pfp to reflect my role as an enlightener

[main fox]

Same

[fiery bane]

Not really... https://en.wikipedia.org/wiki/Procedural_generation

[fiery bane]

digimon 1?

[past bramble]

What type of layers would I use for image GAN? The list is long so it'll be helpful to have it shortened. I'll be using TensorFlow

[small wedge]

https://www.tensorflow.org/tutorials/generative/dcgan there's an official tutorial on it

you don't necessarily even need convolution, it just helps to introduce translational invariance and gets GAN's a bit closer to VAE's like stable diffusion

[raw pasture]

Short Answer: Provide one reason why logistic regression is better than linear regression for modeling a binary target/outcome.

[eager sundial]

guys do you know any nn plotter such as alexlenail's that has better quality?

[remote stream]

can anyone help me

whenever i try to load a pretrained model from hugging face i am facing this error

OSError: Can't load tokenizer for 'superb/wav2vec2-base-superb-ks'. If you were trying to load it from 'https://huggingface.co/models', make sure you don't have a local directory with the same name. Otherwise, make sure 'superb/wav2vec2-base-superb-ks' is the correct path to a directory containing all relevant files for a Wav2Vec2CTCTokenizer tokenizer.

[lapis sequoia]

it makes the output interpretable as a probability; a linear regression is unbounded

[remote stream]

@lapis sequoia

[lapis sequoia]

but you could get similar results if you train, and use some cutoff value.

[remote stream]

can you help me with the problem

[lapis sequoia]

i happily would, but that looks hard, and ive never used hugging face

[remote stream]

😢

[lapis sequoia]

can only wish you luck at this time, sorry :-(

[remote stream]

@summer plover

can you help me

[tawdry gyro]

Can someone help with this:
https://stackoverflow.com/questions/78918404/how-to-fix-artists-getting-cut-out-when-saving-a-figure-in-matplotlib

[lapis sequoia]

pretty neat highlighter https://web-highlights.com/ (add on)

[pine escarp]

Gonna try it then.

[past bramble]

thanks a lot

[lapis sequoia]

How Bayesian Statistics finds degree of belief for a parameter of model
https://en.wikipedia.org/wiki/Bayesian_inference#Making_a_prediction

[pine escarp]

do you understand the math in it

[lapis sequoia]

just conceptually

[pine escarp]

the math seems easy but i dont understand it

[lapis sequoia]

the way i see it is first simplifying it:

1. remove all the formulas but the first one
2. remove the second multiplicatory term and the letter D

then, you have the equation of a line where y=P, x=c

in this example, c stands for the century, it's a continuous variable from 11 to 16 iirc

[pine escarp]

does the python bot support latex?

[serene scaffold]

.help latex

[strange elbow]

Command Help

**```
.latex <query>

*Renders the text in latex and sends the image.*

[serene scaffold]

sir lancebot rather than python, but yes

[pine escarp]

noted

[serene scaffold]

.latex
$$ \frac{\alpha}{\beta} $$

[pine escarp]

i wanna learn manim

it looks so cool

[serene scaffold]

idkwhat nanim is

[lapis sequoia]

math animation

ma for maths, anim for animations

[serene scaffold]

I see

[lapis sequoia]

used by 3b1b

[pine escarp]

have you watched 3blue 1brown videos in youtube

[lapis sequoia]

uve got a flashy profile pic

[serene scaffold]

🔵 🔵 🔵 🟤
yes

[lapis sequoia]

i bet stelercus means magician or smth?

[serene scaffold]

no. "stelercus" has no meaning outside being my name.

[lapis sequoia]

is that greek? never heard a similar word

[serene scaffold]

my current pfp represents my persona as the homosexuality pope (green).

[serene scaffold]

it's latin-inspired.
greek would be "stelerkos"

[lapis sequoia]

cool name

[serene scaffold]

ty bb

[pine escarp]

is my name cool too

[serene scaffold]

sure

[lapis sequoia]

yeah seems from dragon ballz

[pine escarp]

[pine escarp]

its a japanese name so its common that you can find that name in japanese animes

do you guys want to see the projects i have to do within 6 weeks

for my training course

idk if i can do it

but lets hope i do it

[lapis sequoia]

start by iris

[pine escarp]

iris as in iris in our eye?

[lapis sequoia]

i think it's a wines dataset, i may be very wrong

[pine escarp]

ohhh

its prob that one

[lapis sequoia]

anyways, i said cuz it seems the simplest task, maybe 1 too

u gonna be a blown up rich man by the end? \s

[main fox]

Iris is a flowers dataset

[pine escarp]

ohh

i chose advanced projects

cause i wanna get job offers

they told me they would offer me placements if i perform well]

[main fox]

Good luck!

[hollow lake]

Does anybody have suggestions on how i can learn the math behind machine learning?

So my goal is try to learn the math behind ANNs, CNNs, so that when i build models, I can learn how to optimize them better.

Any resources would be great

[pine escarp]

idk if its true, but im learning a lot in the process so i dont really mind if i dont get an offer

i know i have asked this before

but for data science

is poetry good

or venv is enough?

[agile cobalt]

venv is enough for most things

if it isn't somehow, forget poetry and go straight way to miniconda/anaconda

[main fox]

Docker images? Containers?
Nonsense, just pip install requirements.txt

[pine escarp]

fairrr

[ocean pawn]

Nope the app one, I am definitely younger than you, so Digimon App is what I watched when I was a child. In all fairness Digomon App Generation or whatever it's called is really good too

[vapid storm]

My friend's media organization is hosting a pioneering event on AI agents. is anyone interested?

[woven sundial]

is there any way to gpu accelerate my ai on windows?

[violet gull]

Yes

[timber trail]

Yeah just ask gpt or your card provider

[lapis sequoia]

There isn't for Keras + Tensorflow backend (or TF only.) unless you use WSL2. @woven sundial

[tidal bough]

well, technically you can also downgrade to a tensorflow version that supported windows

but yeah, mostly you're meant to use WSL.

[abstract wasp]

Hi can anyone explain to me or link some info on how to deploy models?

[agile cobalt]

which kind of models? deploy to where?

[abstract wasp]

Deep learning models, and where, Idk lol. Idk much about this so if you could give me general insight it would be great.

[agile cobalt]

you can deploy nearly anything to pretty much any hosting major provider, it can get expensive if you don't plan properly though

some random examples:

• https://aws.amazon.com/pytorch/
• https://onnxruntime.ai/docs/tutorials/azureml.html
• https://huggingface.co/inference-endpoints/dedicated

[abstract wasp]

Is it easy to deploy? Or is it a complicated process?

[agile cobalt]

for most models it should be nearly trivial, assuming you're willing to pay the cost for convenience

if your model depends on custom low-level code, non-python dependencies or other unusual things it could be a little harder, but if you can get it working in Linux it'll probably work

[lapis sequoia]

side note: for small models you can normally run them on the client

[agile cobalt]

side note: in some cases you might want not to (even if you could) in order to keep the weights secret

but yeah, running in the device (aka inference at the edge) is also an option for some applications

[lapis sequoia]

that's true

[past bramble]

day 3 kaggle report:
created handwritten digit recognition from mnist and made a pygame to play around with it

[unkempt apex]

share that game video if possible

[spring field]

Good job

P.S. you used pygame-ce, right?

[past bramble]

nope, just pygame, is that different?

[past bramble]

I'm not a pygame expert, it's just a grid with a button to check number

[unkempt apex]

its community edition

[past bramble]

are the features different

[unkempt apex]

https://www.deusinmachina.net/p/what-is-happening-in-the-pygame-community

[woven sundial]

how about exporting app to exe? its still imposible?

[magic bane]

https://hastie.su.domains/Papers/ESLII.pdf

Ensure you know multivariate calculus and statistics at a decent level first. Also consider learning some matrix theory.

[lapis sequoia]

i don't know that

[drowsy ice]

RL Course by David Silver do you guys think its revelant

Like what direction should I go to learn reinforcement learning?

[ebon torrent]

yes, you're absolutely right. I tried all the methods that exist in this universe

[ebon torrent]

Firstly, could you tell me what you think about the course? I believe it's relevant because it could be beneficial.

[woven sundial]

in pixinsight, photo proccessing software is .dll with tensorflow which allows to use gpu. so maybe could i use it to accelerate instead of pip library?

[past bramble]

wait so those context length limits on text models, they're the size of first input layer? 128k context length meaning 128k input neurons?

[small wedge]

Yes

[past bramble]

took me long enough to figure out

[serene grail]

Cool, I had no idea

[past bramble]

I realised this on a walk to my lunch

[small wedge]

Indeed, that is why models lose context once they output enough information

[past bramble]

how do they limit output tokens, is there a feature in nn to limit how many output nodes we need?

[small wedge]

Say you have 5 input neurons and you give the model ['how','far','is','the',''] as an input (just imagine it's encoded into a vector). The model will output 'drive' as the next token. To continue getting input we will now feed the original vector + the models output so the input becomes ['how','far','is','the','drive']. The model outputs 'to', and your next input becomes ['far','is','the','drive','to']. And you can see context on the original message is lost since there aren't enough inputs to represent the whole output and original input together

[small wedge]

Models output 1 token at a time

You run this process iteratively until you have all the tokens you want

[serene grail]

That makes so much sense

[past bramble]

is that possible in neural network to output one node at a time, or make a similar system?

[small wedge]

Sure it's possible but we don't do that, we output a vector representing the probability of each token in the vocabulary being the next output, using the softmax activation function

Then the one with the highest value is the choice. There is also a temperature variable that is often introduced that effects how distant the probabilities can be from each other, to introduce ambiguity in the output. These ambiguous output choices let us use a random decision to make the models output "more creative" and less deterministic while still being more coherent than an equally distributed random choice.

[past bramble]

interesting how it's so simple and I'm learning it now

[small wedge]

Yeah, and the fact that there is such simplicity leaves it open for a ton of innovation in these techniques. With the mind blowing results we've had so far on decades old techniques it'll be fun to watch where research goes over our lifetimes.

[lapis sequoia]

i wonder whether llms understand fallacies wo training say like this one https://www.fallacyfiles.org/amphibol.html

[small wedge]

They don't understand anything, they are likely word guessers

[lapis sequoia]

One evening I killed a kangaroo in my suit. I dont know how he got into my suit.

That's quite a whimsical and surreal statement! It sounds like you're playing with language and humor. The idea of a kangaroo somehow being in your suit and then being killed by you (inadvertently, perhaps) is both absurd and amusing.
If you want, I could help you develop this into a short story or a joke. What direction would you like to take it in?

i meant if the reply would be coherent

[small wedge]

Would just depend on how much of that was in its dataset I would think

And with how much of the internet and rlhf modern models have gone through I'd expect it to be fairly common

[serene grail]

There's a YouTube channel "AI explained" and the guy has made his own personal questions to test and benchmark LLMs (so there's no leakage), humans score 96% and the best models score 6-12% or something like that

[past bramble]

What are filters in Conv2D layers? I have used them without much of an idea how they work

[lapis sequoia]

prisms of numbers

[serene grail]

One example of a question he gave is somewhat similar to this but a bit more complex and phrased as a math question (but it's not actually about math)

So most LLMs he tested just treat it as a math question

[lapis sequoia]

wouldnt it be nice to see how much a human that never experienced a fallacy would take to understand, vs an LLM

read understand as 'model'

sutskever says something loosely related, to train an ai without references to consciousness, then describe it and see the reply

[small wedge]

A filter is the part of the convolutional layer that contains the weights

The number of filters you use is a hyperparameter, like the number of neurons on a Dense layer

[lapis sequoia]

actually, grown up humans cant understand language at all if they werent exposed in childhood

[small wedge]

The issue is we can't test that, the model just has to have a convincing output, not a logical understanding of the content we feed it

Which is exactly what it's done

[serene grail]

It was something like "John wants to buy this many chicken nuggets and he has this much money. Which size of whole chicken nugget boxes can he buy, given that he currently is in a coma and cannot purchase anything? Choose an answer: A) 4 B) 8 C) 24 D) 0"
and the models tend to just ignore that John is in a coma (presumably because the "math question" structure of the prompt means that the likely answer is just going to be math-related)

[past bramble]

can I consider it as number of neurons?

[lapis sequoia]

Given that John is in a coma and therefore unable to make any purchases, the correct answer is:
D) 0

[small wedge]

Ehhh depends on for what purpose but you can treat them similarly

[woven sundial]

can i use already built dll tensorflow library in python?

[past bramble]

what's the difference in Dense and Convolutation layers? Conv uses kernal operations on the input?

[lapis sequoia]

the answer i like the most is that LLMs do program fetching, but not program synthesis.

so they fail at novel tasks

[past bramble]

I know they're used for images, I'm curious on their differences

[serene grail]

That also makes sense to me

[lapis sequoia]

they populate templates (that's how abstract interpolation feels like), but cant make new templates

imagine overcoming that limitation

[small wedge]

Yes, each of your models filters are applied as a dot product in the convolution. The filters learn weights to extract important data out of the image, and also introduce translational invariance (i.e. a cat in the top left is still a cat in the bottom right.)

[jaunty helm]

they're still bad at generalizing / finding underlying principles honestly
asking them to do math on bigger numbers is more likely to fail

maybe grokking that supposedly takes 10x more training compute will fix it

[lapis sequoia]

agree on 2), not on 1)

imho 1) is what they are best at, but might depend on many variables

[small wedge]

The intuition behind convolution is that it has an "anchor datum" in this case a pixel in the center of the kernel, and it gives you lots of information about the neighbors of that pixel, and each new one as it strides across the image

Which is useful to learn patterns from and often reduces the dimensionality of the input

[lapis sequoia]

like: can they summarise? find intentionality on a text? (replying to purplys)

imho there are also enough samples for generalising in multiple contexts

now, can they solve the transposition (caesar) cipher, i think there is consensus they didnt in general

so im saying, it's a fragmented answer; and may depend on who knows what. (idk if it's an instrinsic or data limitation)

[jaunty helm]

I don't think I'm disagreeing with they can summarize / find intentionality?
I guess was more thinking stuff like yes, caesar cipher as you mentioned, or conway's, etc

[lapis sequoia]

imho 2 pieces we are coming to are @jaunty helm :
1) are they sample inefficient and why? (generalise from small dataset + compress information in general way), and
2) can they become good at novelty (do program synthesis.) ?

[past bramble]

Are Conv2dTranspose quite different in how they work from this? Do they just rotate the image?

[small wedge]

The need for transposed convolutions generally arise from the desire to use a transformation going in the opposite direction of a normal convolution, i.e., from something that has the shape of the output of some convolution to something that has the shape of its input while maintaining a connectivity pattern that is compatible with said convolution.

From the keras docs

[past bramble]

Im reading that page and still confused, I would need an example of application to understand

[lapis sequoia]

AC = D
AtD = C

tho not all matrices are invertible.

[serene grail]

t is transpose here?

[small wedge]

So seems like it could be used for taking data out of a bottleneck for example if you were using convolution to reduce the dimensionality, you could use a transpose convolution to increase the dimensionality back in the same way, which maintaining the invariance and signal data that you get from convolution

[lapis sequoia]

yes

[small wedge]

https://medium.com/@marsxiang/convolutions-transposed-and-deconvolution-6430c358a5b6

In a tranposed convolution, instead of the input being larger than the output, the output is larger. An easy way to think of it is to picture the input being padded until the corner kernel can just barely reach the corner of the input.

The visual under that line is helpful in intuiting it

[serene grail]

[past bramble]

thanks a lot those graphics helped to understand

[small wedge]

Np

[past bramble]

dumb question, if we normalize the dataset, don't we need to use same normalization on any data we'll use the model to predict on?

[jaunty helm]

yes

[past bramble]

so is it better to use batch norm inside the nn to avoid that

[small wedge]

There are some methods of normalization that don't effect the data and some that do

Dropout for example we don't use during inference

Batch norm we do

Batch norm is generally preferred, combining them can cause issues but can be useful with special types of dropout

[jaunty helm]

sure, you can put a batchnorm layer right after the input layer
though do remember that batchnorm isn't free, and calculating std/mean once is probably faster

[lapis sequoia]

chat gpt answer to this ^ prompt is quite interesting (hopefully the most frequent one.)

[past bramble]

what are channels in input shape?

[small wedge]

An extra dimension of the input

Like r g b color channels

You get a separate 28x28 matrix (channel) for each color on a 28x28 rgb image

[pine escarp]

Can be grayscale too right?

[small wedge]

Greyscale uses 1 channel

[lapis sequoia]

answer 1

[pine escarp]

Oh yea.

[lapis sequoia]

answer 2

last part was excluded, it says:

Reinforcement Learning and Evolutionary Methods: In some approaches to program synthesis, reinforcement learning and evolutionary algorithms are used to explore the search space more effectively. These methods can adapt and become good at handling novel situations by continuously improving their strategies based on feedback.

[small wedge]

Based gpt

Non gradient based evolution strategies are goated

No worries of local minima or saddle points

[past bramble]

for rgba, it would be 4 channels?
how are they helpful if the inputs are 1d numbers for each pixel?
I might be taking it wrong, i made mnist model yesterday where each pixel was a number b/w 0-255

[lapis sequoia]

yeah, but have their own set of complex issues as well...

ppl suggest combining those

[small wedge]

The represent the intensity of each color at each pixel. It's just a way to store more information easily.

Say pixel x,y has a red of 240, a green of 220 and a blue or 250 with an alpha of 255, it will be an off white color

You divided by 255 to normalize the data so instead of 0-255 it will range from 0-1. This prevents the model from making bad assumptions about the data, like thinking a value of 255 is worth 255x more than a value of 1 for example.

[past bramble]

so they define array size of each color pixel? for grayscale [...], rgb [..., ..., ...]

[small wedge]

Yes

[past bramble]

For a image generator trained on dataset of specific images, I wouldn't need any input. Is it possible to create a model without inputs?

I have seen ones that create noise for inputs, I want to make one without noise

[agile cobalt]

the output would be always the same if it had no inputs, which is typically not very interesting?

[woven sundial]

WARNING:absl:Compiled the loaded model, but the compiled metrics have yet to be built. model.compile_metrics will be empty until you train or evaluate the model.??

what that means

[small wedge]

And also not possible to train at all

[past bramble]

well that's not intended

I'll start one soon, kaggle had some flowers dataset

[slate raven]

I need the code that does the following (chatGPT cannot really do it):

• I have a mapping (126 data points) from 2d pixel coordiantes to 3d world coordinate
• I need code to find the extrinsic matrix transformation (I have already the intrinsic)

[fiery bane]

https://kidger.site/thoughts/just-know-stuff/ here is the list of things to learn. There is a section for maths

[fiery bane]

It is possible to trian

[fiery bane]

so, each datapoint is a 2x3 matrix?

[drowsy ice]

Lot like other courses seems overtly complicated

[slate raven]

it's a dictionnary where the key is a tuple (x and y coord) and the value is a tuple (x, y and z coord)

[past bramble]

For the image generator model I'm creating,

I feed random noise as inputs, and the real images as labels, is that all?

is the concept of separate discriminator and generator models needed for this? Or is my approach appropriate?

[verbal oar]

and now you make distribution close and close to given distribution

[eager plume]

Hi,
To understand DS and ML practically, which tools or libs should I learn/use.

[serene scaffold]

do not try to learn DS and ML in terms of libraries
do not try to learn DS and ML in terms of libraries
do not try to learn DS and ML in terms of libraries

[fiery bane]

numpy, pandas, matplotlib, altair, sk-learn

keras
tf, tf2++
pytorch, pytorch lightning
jax

[serene scaffold]

do not try to learn DS and ML in terms of libraries

[fiery bane]

everyone learns differently. If someone wants to start from libraries, and somehow managed to get all the math and concepts by reading the pure code and docs alone, all power to them.

Let them try, until they are get what they can get out of library alone, and then let math and CS fill in the rest afterwards. I don't think it realy matters where people start.

[eager plume]

College Curriculum has DS and ML subjects, Theory stuff.

Also they are using R language, But I am Python Geek

[serene scaffold]

this channel sees a lot of people who go off and try to "learn pandas", and then they come back and say that they've "learned (all of?) pandas", and someone tells them to "learn sklearn", and they never actually get anywhere.

[fiery bane]

well, if they want to get somewhere, they can always come back and we can point all the other apsect they are missing.

[eager plume]

@fiery bane

Can u both tell me how much Maths(like Linear Algebra, Statistics, such broader fileds) need to master to learn

1. DS
2. ML (Basic DL too)

[fiery bane]

depends on your definition of "master"
my witty answer is: this much math https://micromasters.mit.edu/ds/ is a good approximation of a master degree

In another sense, IDK? No one on earth has "mastered" ML, especially DL, that's why it is an ongoing topic of research

[serene scaffold]

for data science and several kinds of machine learning, the most important things are probability and statistics.
when you get into things like neural networks, linear algebra and calculus (especially derivative calculus) become important.
it's also helpful to understand graph theory, as that's used to model a lot of problems.

[fiery bane]

In another sense, this list is good enough for a PhD, so if you consider a PhD freshgrad a master, then there you go https://kidger.site/thoughts/just-know-stuff/

[uneven stirrup]

https://gamengen.github.io/ entire game (doom) conjured by Ai

theres no game engine here, nothing managing internal logic or variables, just input, linear algebra, and frames on the screen

[gilded belfry]

Can anyone suggest an algorithm that will separate the parts of different colors in this picture with straight lines? 4 pieces, none of the pieces have a completely certain intensity value.

[fiery bane]

k means?

[verbal oar]

which llm book do you recommend?
first can be theoretical, second can be practical

i know about deep learning and NLP but not llm

[mild dirge]

See the image as a large number of 3d (RGB) points. Cluster these (with kmeans if you know there will be 4 colors, otherwise check other clustering algos). Then every pixel is assigned to a cluster. @gilded belfry

[past bramble]

day 4 kaggle report: unsuccess trying to create image generating model

I'll try tomorrow with something simpler like mnist

[unkempt apex]

Keep Grinding!

[verbal oar]

hmm maybe start at first with VAE then GAN

[past bramble]

what's vae?

I'll try finding videos for this, I learn better that way

[gilded belfry]

yes i thought the same and k means works well

my problem solved

[verbal oar]

variational autoencoder

mit have lecture on deep learning VAE

[abstract wasp]

Do you guys think a masters in AI/ML is worth it? I just graduated w a bachelor’s in ML and I have 2yrs of research experience. I just started my job search some days ago. I feel like I still need to learn a lot in ML, I really don’t feel prepared to get the ML job so I’ve been applying to data analyst/science jobs rn.

[serene scaffold]

I feel like I still need to learn a lot in ML
there's so much to learn that this feeling never goes away.
you probably should have started your job search a few months ago. but give it a few months and see what your callback rate is per application. track what positions you're applying for in a spreadsheet.

do you mind telling me what university you got the degree from?

[abstract wasp]

i graduated from ucsd (cog sci - ml and neural computation), tbh they taught us the basics. w my job search i noticed they ask for aws and they didnt teach us that, we didnt get to learn in a lot depth as i wish they wouldve. im grinding through a gen ai book rn cuz we honestly missed out on a lot

[left tartan]

The good thing about having to learn AWS is: you can't possibly learn all of it, and neither will anyone who ever interviews you. So, you just need to know enough to be useful.

[pine escarp]

I'm doing B.Tech Computer Science & Engineering (Data Science). In my place (maybe in your place too ) they don't teach you the skills we need, for example pandas tabluea power bi etc. We have to study it on our own. Get certifications.

They jus teach theory part related to computer science.

Maybe some programming too!

I'm not sure how is the syllabus in foreign countries.

But I'm planning to do my M.Tech in America or UK.

[abstract wasp]

yeah they didnt teach us tableau either so i learned it by myself ;-;

[pine escarp]

Tableau is actually easy to learn.

[abstract wasp]

im going to start the aws cloud practitioner program soon too, im using the free course on the aws website and hopefully get the certificate soon, i was thinking of doing the the ml engineering path:
https://d1.awsstatic.com/training-and-certification/docs/AWS_certification_paths.pdf

[abstract wasp]

yeah thankfully

[abstract wasp]

yeah ik ;-;

[left tartan]

For many jobs, certs just don't matter. If it's what motivates you to learn something, great, but you can prob pickup what you need much more easily

The other part: can you truly learn it without having a project/job/something that reinforces it.