data-science-and-ml

[tender hearth]

High school thesis or uni?

[turbid field]

college uni thesis

sorry i dont have enough knowledge yet

so im trying to grasp what to except

[tender hearth]

did you propose it yet?

I don’t think it’s novel enough to be accepted

at least, based on my experience ha

[fast geode]

where can I talk bout web scraping?

[tender hearth]

research ko back in high school was using a face recognition model in a video compression algorithm to keep faces high quality

Can be used in video conferencing

You’d need to go much deeper than just “classifying types of vehicles”, since there’s hundreds of models that already do that online

[turbid field]

pinoy?

[tender hearth]

yeah

[turbid field]

nope hindi pa

[tender hearth]

English only in the server pls, it’s a rule

[turbid field]

ayon kasi recommendation samin ng civil prof

[tender hearth]

But we can take it to dms

[turbid field]

ay

why is it not novel enough

[turbid field]

^

[tender hearth]

Because it’s a very basic problem and very solved

But if you propose and they accept it, go lang

But if I were you I would try finding another problem or make it more specific

[turbid field]

even with raspberry pi implementation to count cars and classify them? also pala we will log it in web-based datebase e

[turbid field]

yaaa we will ask it pa

our drowsiness detection for car implementation got rejected tho

[turbid field]

will be use for traffic survery for civil eng

[tender hearth]

If you’re gonna use it as part of another research, then that’s fine

As long as it’s not the main part of your research

[tender hearth]

What was the premise for the rejection

Also which uni is this?

[turbid field]

too hard to get datasets daw

[tender hearth]

Okay, valid tbh

[turbid field]

may before researchers daw that coducted almost same concept but they gave up daw eh

[turbid field]

dlsu

but no main hehe

[tender hearth]

ohh, I was abt to get excited haha

main here

[turbid field]

AY WEH ASHUDHUASHUASHASDHASHDAH

[tender hearth]

What will it be about

[turbid field]

what yr?

electronics no

[tender hearth]

Secret

[tender hearth]

^

If you do decide to go forward with the vehicle classification, there’s lots of resources online for that

[turbid field]

civil eng stud in our uni manually count cars for research survey

idk anong tawag sa survey

but they were counting from 9am to 8pm?

so we did consult the research adviser, and told us to automate it

[tender hearth]

https://www.dlsu.edu.ph/wp-content/uploads/pdf/research/journals/jciea/vol-3-1/6galvez.pdf

Look at their metho

Then adapt it to your use case

[turbid field]

ohhh using mobilenet

thanks

[tender hearth]

Yepyep

[toxic holly]

yeah im pretty sure

im using libraries im using cv2 and mediapipe

i managed to get a real time handtracker that counts hwo many fingeres you have up

and then using that i wanna develop it to do sign language as wel

for the model im using pytorch

and thats it

wbu

[lapis sequoia]

Hello everyone

[regal galleon]

(Note: I am a newbie in this field)

K-Fold Cross-Validation, f1 score, and train-Test Split, statistical tests such as McNemar's.

Despite all this, I don't think it's exactly the right way to understand it. It comes out after you try it in a real-life problem. The model can perform very well. What about outliers? Values ​​that change over time? So I think the best way is to test it with different data. A test data that is statistically different.

Things like train test splits only increase the model's reliability. They can be misleading. I've experienced this.

In fact, increasing the number of data seems like the only real solution. 😅 👀

[glacial root]

right now for the number classifier i'm doing it from scratch (just numpy), but when i do the sign language one i'm thinking of doing a real time one, not sure if i'll do it from scratch or with cv libraries though

typically how do mle interviews run, as in what's the structure

[daring karma]

A friend and I are attempting to create an open source project that automatically identifies unattractive clauses in end user license agreements/privacy policies. I mostly come from a background of reverse engineering and security research and not data science, but I am very passionate about this.

I figure NLP is likely our best bet?

My understanding is that transformer models are somewhat unwieldy and kind of like a black box.

[serene scaffold]

NLP is "the subset of AI that deals with natural language", so saying "NLP is the best way to automatically extract certain clauses from a legal document" isn't really saying anything.

[serene scaffold]

most of AI these days is a "black box".

[daring karma]

I think what I mean to say is that techniques such as NER or text classification have more definite output shapes...?

[serene scaffold]

any program you can possibly think of to solve this problem will, by definition, be NLP.

I guess that's not your point of confusion.

[daring karma]

Right.

[serene scaffold]

can you tell me what you mean by "definite output shapes"?

[daring karma]

i.e. the model has a finite set of outputs, as opposed to the output being a stream of tokens with indeterminate structure and length

The ideal output might be spans of text with "low-level" details such as: [BINDING, ARBITRATION]

[main fox]

For classification tasks, yes. The model classifies data into categories it was trained on (zero shot is an exception).

NER is a type of classification where each word is classified into one of the entities specified in training

[daring karma]

My guess is that for this purpose, an NER model might be ideal. But instead of entities like "PERSON" or "ORG", we use low-level legal primitives, like "LIMITATION", "BINDING", "WAIVER", (more research is needed on my part for the ideal collection of these)

[main fox]

Sounds like it, yes. You could go the NER route or test specific prompting of an LLM.

[daring karma]

Is it possible to get an LLM to determinately emit structured output?

[serene scaffold]

if the interactive LLM is instruction-tuned, it will almost always adhere to instructions like "return the output as a JSON with this format"

[daring karma]

Interesting, I'll need to investigate how performant something like that is

[glacial root]

would a real time sign language identifier be a good resume project

[main fox]

Yes

[serene scaffold]

that would be incredibly ambitious.

[raven cloud]

Hi!
Stuck on not being able to get llama.cpp-python to use the gpu.

• Nvidia-SMI shows that cuda 12.5 is installed
• I installed doing : pip install llama-cpp-python --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cu125
• Here is the script im using:

from llama_cpp import Llama

llm = Llama(
      model_path="../../models/DeepSeek-R1-Distill-Llama-8B-Q4_K_M.gguf",
      n_gpu_layers=-1, # Uncomment to use GPU acceleration
      # seed=1337, # Uncomment to set a specific seed
      # n_ctx=2048, # Uncomment to increase the context window
)
output = llm(
      "Q: Name the planets in the solar system? A: ", # Prompt
      max_tokens=32, # Generate up to 32 tokens, set to None to generate up to the end of the context window
      stop=["Q:", "\n"], # Stop generating just before the model would generate a new question
      echo=True # Echo the prompt back in the output
) # Generate a completion, can also call create_completion
print(output)_

After running this keeps pooping up:

print_info: EOG token        = 128008 '<|eom_id|>'
print_info: EOG token        = 128009 '<|eot_id|>'
print_info: max token length = 256
load_tensors: layer   0 assigned to device CPU
load_tensors: layer   1 assigned to device CPU
load_tensors: layer   2 assigned to device CPU

The logs has literraly zero mentions of GPU (Crtl+F) does not show any results for GPU
Any help appreciated

[glacial root]

it would?

i thought the complexity of it is typical of resume projects

[serene scaffold]

uh, yes? are you talking about just drawing bounding boxes around signs, or translating whole communications into a written language?

[glacial root]

just signs

not whole communications

[serene scaffold]

what will the input be? images that are cropped around exactly one sign?

[glacial root]

no i mean if i had a camera in the application and it could classify what i'm "saying"

just single letters though

[serene scaffold]

you know that sign language isn't just letters, right?

[glacial root]

though maybe i could have a system where it records the letters to be able to "type" messages

[glacial root]

yeah but a lot of the full words are dynamic signs right

those i would have to figure out how to do

[serene scaffold]

and that would be incredibly ambitious.

[glacial root]

yeah that's why i thought of just letters

or maybe i should just try an autonomous drone (probably would be tougher but it's what i want to work with in the future)

[serene scaffold]

how would you distinguish between signs that are letters and signs that are not?

[glacial root]

well it would only pick up the sign if it's able to classify it into one of the letters right

[raven cloud]

Ok got it working but what gives ?

{'id': 'cmpl-b1422a3e-f391-4f18-813b-7d7466d574bc', 'object': 'text_completion', 'created': 1741033924, 'model': '../../models/DeepSeek-R1-Distill-Llama-8B-Q4_K_M.gguf', 'choices': [{'text': "Q: Name the planets in the solar system? A: 1. Mercury 2. Venus . . . Hmm, wait, that's it? Only two? No, that can't be right. I must", 'index': 0, 'logprobs': None, 'finish_reason': 'length'}], 'usage': {'prompt_tokens': 14, 'completion_tokens': 32, 'total_tokens': 46}}

32 max_tokens and the question prompt were also default

[serene trellis]

any tips to reduce state space for RL QLearn? currently doing a biclustering algo but since its for a recommender system i need to guarentee that there's at least an instance of each row in at least one bicluster

since if an user isnt in a bicluster i cant generate recommendations

[glacial root]

yo this might be a dumb question but i genuinely don't know, if i'm implementing a neural network from scratch and use np.gradient, does that still count as done from scratch?

[devout sail]

You decide

Is it for some school work?

[small wedge]

I don't think from scratch has a strong definition, but you probably are cutting out the hardest part of doing it from scratch by doing that

[serene scaffold]

In order to truly do something from scratch, you must create the universe

[glacial root]

nah

just to learn

[glacial root]

so for the sake of learning it's better not to use it

[devout sail]

Depends on how far you want to go then. Implementing the autograd yourself is common in school exercises

[bright thorn]

can somebody help me with a virtual enviornment issue via anaconda prompt. I am basically trying to import keras and tensorflow into my jupyter notebook but I keep getting errors that tensorflow keras cannot be found.

[hearty depot]

did u activate it?

[serene scaffold]

see #1346312369277636650 if you want to help this person.

[bright thorn]

yes I activated my virtual env, I conda installed tensorflow and had to do it with keras too but still doesnt work

im using a mix of chatgpt and stackoverflow and im going insane lmao

[hearty depot]

r u using vscode or like the jupyter itself?

[bright thorn]

jupyter

[hearty depot]

can u try running conda list

[glacial root]

yo guys do i need a virtual environment if i have anaconda

cause anaconda already comes with all the needed packages, so what else would i need a virtual environment for

[serene scaffold]

my recommendation is to fully remove anaconda from your system and forget it exists.

[bright thorn]

I ran it

keras and tensorflow show up

[serene scaffold]

@bright thorn I recommend closing your thread if you're not going to post the whole error message and you're receiving help here instead.

[hearty depot]

🤓

[bright thorn]

my internet is being weird with jupyter so im waiting for the error message to load up so i can ss it

[serene scaffold]

you should always post text as text, not as screenshots. people often need to google parts of the error message, and it's unreasonable to ask them to retype it from a screenshot.

[bright thorn]

ok

[hearty depot]

wait when u say ur using jupyter r u using the desktop app or the browser one
that's odd tho

[bright thorn]

browser as in when I load it up, it appears on my google chrome

[hearty depot]

i'll say tho if u cant figure it out soon just make normal venv or use something like uv/poetry

[bright thorn]

wym north venv and also its for my class so my teacher is kinda insisting that i do it in a virtual env

[serene scaffold]

anyway, anaconda was created when intalling dependencies like numpy was nontrivial. but that's no longer the case. so in the vast majority of circumstances, anaconda just creates a barrier between data scientists and the rest of the python community, for absolutely no benefit.

I work for an AI company with over 10 thousand employees, and anaconda is banned at my company.

[bright thorn]

type shi man

[serene scaffold]

like, they literally deleted anaconda from everyone's laptop one day. (except it wasn't on my laptop)

[bright thorn]

nice

[hearty depot]

akshually conda isnt just limited to python 🤓

[bright thorn]

@serene scaffold did you go to a prestigeious school before getting your current job

im just curious bc im deciding on colleges rn

[serene scaffold]

I went to an average state school.

[bright thorn]

gotcha

I wanna go to a school that has a mix of good student life but also good education obviously but I just dont know how I should view it honestly

[serene scaffold]

if you're a hard-working and skillful person, and you go to an "average" university, it will be easier for you to get the special opportunities that are only available to the hardest-working and most skillful students.

[bright thorn]

fair

but I guess I just worry that even with hard work that i won't be able to benefit any special opportunies if i dont go to the right school

[hearty depot]

come to my alma mater, we good and in city environment

[bright thorn]

whats the alma mater

[hearty depot]

georgia tech

[bright thorn]

ok so, im chopped and i could not get in 💔

this year for college apps was so fucked

i could have prolly had a chance if i was born a diff year but like i barely got into schools and places where I got rejected/waitlisted I most likely would have had better chances

[serene scaffold]

that's a legitimate concern. but it's not a secret what the special opportunities are. if you know what you want to specialize in, you should just email the admissions counselors and ask them what opportunities there are for your preferred specialty.

[hearty depot]

u can always transfer
ngl tho if u still in hs and u lock in rn and grind coding
u will become mbappe when u graduate regardless of where u at

[bright thorn]

well I mean my concern roots from the fact i wanna do computer science and or cyber security

[bright thorn]

im a senior gang

ot

oh wait*

i didnt read the entire message mb

[hearty depot]

u good bro

one of mu biggest regrets was not grinding harder in covid cuz i had all the free time in the world to get skills uo

[bright thorn]

yea I understand experience / practice is what helps but rn Im kinda just full with my hands on my current class bc im doing learning development for ai models

[hearty depot]

if u learning developiment for ai models in hs bro u ahead of the curve fr

like i shit u not there were ppl in my graduating class that couldnt do i/o in any language u doing fine

[bright thorn]

idk I mean my teacher says shi like that but atp in this current time of computer science where imposter syndrome is common i just feel behind

[bright thorn]

i appiricaite that I guess I just doubted myself a lot in my capability due to the fact i didnt really get into my nearend reach schools

like umd was my dream school but i got rejected, which was crazy from the start since they only accepted 100 comp sci applicants but I was just too cocky id get in

[hearty depot]

u'll be fine bro, i spent the first two years of college just drinking and partying and ended up fine w multiple offers paying 200k+

[hearty depot]

ah u from md?

[bright thorn]

i appriciate that too, im trying to lay back a lot more and realize I have the grind mentality but i dont wanna drive myself into the ground

[bright thorn]

nah sadly i live in nj

[hearty depot]

u dont need to even grind hard bro, just be consistent and u will be fine

like main reason why ppl cant code or get job coming out of college is that theydont code on consistent basis

[bright thorn]

good point, I just like being ahead so I guess thats why i thought it was necessary

[hearty depot]

i mean u already r ahead, like im assuming u going to some place like rutgers but at least at georgia tech a ton of kids entered w/o ever coding a single line of code

[bright thorn]

true, its like scary hearing all the statistics but i never hear how the same mfs arent utilizing their degree so i just feel mislead with how the market current is

[hearty depot]

market isnt great but most ppl that didnt get job didnt really grind or apply themselves outside of classroom

[bright thorn]

i got into a diff branch of rutgers not main campus but Im considering west virginia u, penn state, loyola md, u delaware or rochester institute

[hearty depot]

u can just write rutgers university if u go there, if cost isnt an issue id prob go to penn state or rochester

[bright thorn]

but yea like i said this year every school practically had a crazy surplus of applicants which even lead to rutgers lowering their acceptence rate to like 30

[hearty depot]

u'll be fine bro, the school does matter to an extent im not gonna lie to u esp for stuff like getting vc funding or very early stage startups but tech recruiting is mostly meritocratic so if u r qualified u will get a decent job

[bright thorn]

well preference wise the branch of rutgers i got into isnt hella safe + not much aid, and rochester i just cant seem to udnerstand why i dont like it but why do u say that school

[hearty depot]

better engineering/cs program than the others except maybe penn state

like i wouldnt really consider places like west virginia and loyola md as like places that r super strong at cs or where resarch is priority

penn state altho its lowk a party school it gets a lot of funding for research and also a lot of kids from my hs that went there didnt really have trouble finding a decent job givrn they were dilligent

[serene scaffold]

UMD's CS department is very selective--I live in DC, and I only know two people who have graduated from it.

(Maryland DC )

but I also now have the same job as both of them

[hearty depot]

damn u from the dmv too?

[glacial root]

why

[serene scaffold]

#data-science-and-ml message

[toxic holly]

yeah im doing the real time sign language. from scratch would be pretty tough i cant lie so im gonna use cv libraries

my friend recently bought a pi and a camera module so we're gonna try make the sign language in real time work with the pi and stuff

[cobalt eagle]

Good day, I would be grateful for your help and advice. I tried to use YOLO models found on the internet to detect a dog in real time on the camera image. I failed because the dog and the surroundings have very similar colors. Is it a good idea to create my model and train it on photos of my dog ​​(Siberian husky)? I only care about the script detecting my dog ​​in the selected square. Is training the dog model on a set of 300+ photos of my dog ​​a better idea that might work? 🙂 I plan to use YOLO8l. Thank you for the advice! best regards

[chilly sleet]

umm any recommendation for which gemini-openai proxy to use

[pine ibex]

That is very general statement. It takes over 200 applications even when applying to the job to even get a job or interview. I know this since even after getting a 3.89 GPA for my Master's degree and research to back me up I had a lot of issues, and it does not help with current layoffs from government grants drying up for public and private.

[buoyant vine]

😅 That is equally a very general statement

[turbid field]

why is jetson orin nano so hard to find

almost sold out anywhere

[tender hearth]

Mahal naman niyan pre

pretty sure you can run yolo type models in real time on an rpi

[turbid field]

ay really ba?

rpi with hailo kit?

[tender hearth]

yeah I did it for my high school research

Well not exactly, but I found out you could while doing my research

[turbid field]

rpi 4?

[tender hearth]

no, just an rpi

[turbid field]

4?

what camera did u use pala?

usb type or module?

[tender hearth]

I couldn’t tell you, I didn’t do it, I just discovered you could while writing the RRL, because it was important for our research that you could run our program without expensive accelerator devices like the hailo kit and jetson nano etc

There’s an article out there from tensorflow (or maybe PyTorch) on this topic, im sure you can find it

[turbid field]

ohhhh

we might change it kasi

ocr for license plate, vehicle classfication, and car counter

dont have idea gaano ka heavy sa computation

3 in 1

[ruby jungle]

Hi guys,
I'm doing this pie chart but as you can see all names titles are not visible, how can i make to show all names using matplotlib?

[long locust]

Might need to change the orientation of the labels, or add a legend

Another option is to lump all of the very small ones into a single category

[main fox]

For that many categories, consider using a bar chart instead

[jaunty helm]

use a bar chart

[glacial root]

dang, banned is wild

well then i think i'll uninstall it and set up a venv

[glacial root]

wait how does it create a barrier though

[serene scaffold]

because it creates a different workflow for environment management for, in almost all circumstances, no added benefit.

[glacial root]

nice

[glacial root]

oh i see

[glacial root]

yeah i think i'm gonna use a framework as well after doing this one implementation from scratch just for learning purposes

[glacial root]

i don't know if anaconda was the problem, but after getting rid of it my issue with git lfs was fixed

i had git lfs installed and initialized but it just wouldn't work and continued to give the large file error

now it's fixed though

nevermind i got my hopes up too soon

i really thought

that push was taking an outrageously long time compared to normal so i thought it was working, but after waiting literally almost 10 minutes it finished loading and gave me the same large file detected error

unfortunate

[warm iron]

Hey everyone this is my first big project in AI:
https://github.com/sashaheadshot/12-lead-ECG-classification
just wanted to share it here. 🥲

[toxic holly]

yeah fairs

we managed to get it running on the pi

but my neural network did not work lmaoo

it was so inaccurate icl i think it was j randomly guessing

[hollow lily]

hi guys

i hope this is the right channel

when i create a .txt doc, enter code and rename .py, the doc open, black, then shut down. here is th code: from openai import OpenAI

Step 1: Initialize the OpenAI client

client = OpenAI(api_key="your-openai-api-key")

Step 2: Create a chat completion

response = client.chat.completions.create(
model="gpt-4", # Use "gpt-4" or "gpt-3.5-turbo"
messages=[
{"role": "user", "content": "write a haiku about AI"}
]
)

Step 3: Print the response

print(response.choices[0].message.content)

i've already import open ai with cmd

if someone can help...

[toxic holly]

does it give any error messages or just the doc is empty and shuts down?

[hollow lily]

black screen of cmd pop and deasapear

and it does that a lots of time also with other pi doc

[toxic holly]

I think the issue is with the api_key declaration

That's not the command i dont think

[hollow lily]

i follow the steps on openai website and they give me the k, i just replace it with "your-openai-api-key"

[toxic holly]

idk then

[hollow lily]

do you think their can be a link with azure cloud shell?? cuz i've seen some malicious stuff on my laptop since a bit of time idk

[hearty depot]

this situation doesnt not look ideal for pie chart imo

use diff formm of visualization cuz its not apt for it

[unkempt apex]

ahh good one, I think you can also write a blog on this so that it will more informative

[warm iron]

Thank you! I actually published a paper. Got accepted by iccae, will be online at the end of this month i think.

[unkempt apex]

ohh nice, share that as soon as its published will love to read that

[warm iron]

Sure thing!

[glacial root]

lol

literally just a perceptron at that point, no backpropagation

[glacial root]

does tqdm work in jupyter notebook

[serene scaffold]

Yes

[odd meteor]

Congrats

[main fox]

import tqdm.notebook.auto instead of just tqdm

not that just tqdm doesn't work, but it might give you a warning

[hearty depot]

getting inference fast enough for real time seems hard af from scratch or more effort than needed

[glacial root]

yeah that's why i'm only doing it for my first one which is just classifying with the mnist dataset

nothing in real time

[hearty depot]

oh ye thats p doable

[glacial root]

i get an error message saying no module named tqdm.notebook.auto

and it doesn't work when i import tqdm.notebook

it's probably fine though cause when i just import from tqdm everything works fine

nevermind the tqdm bar that is working is kinda janky, first time it seemed fine but now it prints each number seperately instead of replacing

[toxic holly]

haha yeah

currently retraining it tho!

on like 6000/220 000 but hopefully by morning it will be done

[glacial root]

holy

damn that's crazy

it takes that long to train?

[toxic holly]

well im definitley not an expert im pretty new to nn's

so icl my code is probably pretty inefficient

[glacial root]

my code is definitely inefficient asl lol

[toxic holly]

it does around 100 images/second

[glacial root]

it takes like 20-30 seconds to set up a perceptron with the mnist dataset

[toxic holly]

for a macbook i cant complain that mcuh

[glacial root]

and it's just 2 hidden layers

[toxic holly]

but yeah my friend overclocked his cpu with a 6700xt or something like that so we're gonna brute force it tmrw anyway

[toxic holly]

30 seconds!?

how?

[glacial root]

i honestly don't know

[toxic holly]

can i see ur code lol

[glacial root]

wait let me check again, i might be exaggerating a bit

[toxic holly]

hah okay

[glacial root]

ok yeah it was 20

[toxic holly]

damn mines like 5 im pretty sure

[glacial root]

here's my code by the way

[toxic holly]

what libs are you using

[glacial root]

numpy and pandas

pandas is only to parse the data though

nums = data.iloc[:50000, 0].to_numpy()
input_data = data.iloc[:50000, 1:].to_numpy()
output = np.zeros((50000, 10))
expected_output = np.zeros((50000, 10))
for i in range(50000):
    hidden_layer_one = hl_one_feed_forward(input_data[i])
    hidden_layer_two = hl_two_feed_forward(hidden_layer_one)
    single_output = output_feed_forward(hidden_layer_two)
    output[i] = single_output
    expected_output[i][nums[i]] = 1```

def hl_one_feed_forward(activation):
    weight = np.random.randint(low = -5, high = 5, size = (15, 784))
    bias = np.random.randint(low = -50, high = 50, size = (15,))
    neuron = (np.matmul(weight, activation) + bias).astype(np.float128)
    sigmoid_neuron = 1 / (1 + np.exp(-1 * neuron))
    return sigmoid_neuron```

[toxic holly]

oh wow thats more low-level than me to be fair

[glacial root]

and then i have a method for hidden layer two and the output layer as well, pretty much the same as this one

[toxic holly]

yeah fair enough

[glacial root]

yeah this is my first time working with neural networks so

starting with just basics

[toxic holly]

to be honest same lol

ive only done the number recognising one before and thats basically a tutorial

how did you chooose the number 50,000 btw?

[glacial root]

there's 59,999 training examples

[toxic holly]

really?

my code counted over 220,000

somehow

[glacial root]

and the book i'm going through says i should use 50k and then save the rest 10k for hyperparameter tuning later

i'm trying to do it with as little guidance as possible and only learning theory from this book

not gonna even peek at the tutorial part lol

[glacial root]

the one i'm doing is the numbers one

not the sign language one

[toxic holly]

yeah fair enough

[toxic holly]

haha i love it

[toxic holly]

ohhh sorry my bad

[glacial root]

nah all good

[toxic holly]

i did that one

[glacial root]

i'm not sure what's with the number though

[toxic holly]

and i made a little twist which you could try

[glacial root]

59,999

[toxic holly]

<60,000

lmao idk

[glacial root]

all it took was one more training example lol

[toxic holly]

haha

[glacial root]

oh what's the twist

[toxic holly]

basically i made a little gui and a really basic pen tool using pygame (dont judge) and then you can draw a number in a square it will resize it to 28 by 28 (for the 784 parameters) and then guess the number

it makes it much more fun to test

[glacial root]

oh dang that's super cool

i was deciding to make it into a webpage where i could input an image which would be resized to 28x28 and then it would guess

i honestly have never even heard of what pygame is lol

there's a lot of things i should know but don't

kinda scary how far behind i am

[toxic holly]

its just a library for making games lol

[toxic holly]

haha you're not dw

python has so many different libraries

[glacial root]

i struggle with git lol

[toxic holly]

for example ive been coding for probably over half my life and ive never used pandas before

[glacial root]

and setting up venvs

[toxic holly]

tbf who doesnt initially

[glacial root]

when it comes to doing stuff in ubuntu i always have claude open telling me what to do lol

[toxic holly]

haha

maybe the best way to create ai is to use it to create itself!?

[glacial root]

hopefully i'm not too cooked when i go to uni

[toxic holly]

how old are you if you dont mind me asking

or you can reply in dms if thats easier

[glacial root]

17

[toxic holly]

ah cool, UK?

[glacial root]

nah i'm from the US

UK would be nice though, seems like a more chill place lol

[toxic holly]

oh no im from uK

lol america seems so much cooler

[glacial root]

nah lol

actually i don't know what it's like after high school yet

but high schools here suck

[woeful grove]

suck how, are they like easier?

[glacial root]

that's part of it, but the bigger thing is that it wastes a lot of time

[woeful grove]

yeah but an education is an education, no?

do you have any other options as an american

[glacial root]

yeah i guess not

system just needs a lot of fixing

[woeful grove]

every system has its flaws

[glacial root]

true

not just that though, just kinda hard to find people who are serious

seems like everyone at school is always on some form of tomfoolery

[soft kite]

I am trying to make a chart similar to this but as an added bonus, I would like to modify the bars to modulate their width kind of like a violin plot. the data itself is the first and last time someone talked in a server and the modulating width would be probably a rolling average of chat activity month over month. I am trying to do it in matplotlib but I am reallly having a hard time making a violin chart with the kind of modulation that I want to begin with, let alone with multiple horizontal bars. @fleet latch for the cross channel ping

[main fox]

Try from tqdm.autonotebook import tqdm

[toxic holly]

same

[toxic holly]

thats true everywhere you go

although tbf in uk highschool you narrow down subjects a lot more earlier which is good i guess

[glacial root]

it still breaks up the progress bar

though no errors though

is this just how it is with jupyter notebook?

[glacial root]

oh that's cool

[silent helm]

Hello chat. I'm planning to learn Generative AI. I am a freshman undergraduate at a university.

I don't know if I should learn machine learning and Deep learning before starting Generative AI

The problem I have with machine learning and Deep learning, I don't know where to start. If i start at basics I always have "learn X before Y" and "learn Y before Z"

[serene scaffold]

Generative AI is an application of deep learning.

You should start with basics and work your way up

[prisma tide]

Of course first you need to learn the basics deep learning is basic before learning GAI

[hearty depot]

learn stats and linear b4 u even think about going deep into gen ai otherwise u gonna struggle

[fervent vale]

anyone professional in datascience

..?

[chilly sleet]

why'd you if it doesn't align with your work. i had to learn it cuz of school lol didn't thought it would be useful in near future

[naive axle]

anyone installed conda in WSL2 for tensorflow setup?

[glacial root]

you can pretty much learn what you need to along the way though right

like if you just get started on a basic neural network

that's what i'm doing, outside of just basic matrix and vector operations and partial derivates i knew nothing when i started

[compact merlin]

Hi everyone,
Has anyone managed to set up OAuth2 with Casdoor in Dify? I'd really appreciate it if you could share your experience or any guidance. Thanks in advance!

[split olive]

Is it not possible to run threads in streamlit?

[hearty depot]

Sure but I’d still dedicate time to learn it well

[fleet latch]

Send an inbox for assistance @compact merlin

[serene scaffold]

please don't try to divert help to DMs. if you want to help this person with aythentication, please go to #cybersecurity

[modern thorn]

Hi. I do not know anything about llm and ai development. Right now i am building a chess game to learn ai or llm in the future. My main goal to learn ai by building chess ai. Can you share some topics possibly i need to learn in this journey?

[glacial root]

yo guys don't tell me the actual best way as i need to figure it out on my own, but if i'm setting up a neural network from scratch, once i have set up the perceptron and am setting up the gradient vector, should i get a np array of every single weight matrix and bias vector

50k samples and 2 hidden layers before the output layer so both of these np arrays would be of size 150k which is kinda wild

[hearty depot]

are u not using autograd, also u shouldnt have that much arrays

[serene scaffold]

they're intentionally not using autograd

[hearty depot]

i mean its troll af to not implement backprop using autograd imo

[lilac lichen]

why not saying purpose of search?

[lilac lichen]

is it really needed? as i know conda can install linux packages for windows without problems

[toxic holly]

yeah tbf if there's no need for it then there's no need for it lol

[chilly sleet]

Wise words 💀🙏🏻

[glacial root]

isn't it better to implement my own gradient method though

to better understand how it works

[hearty depot]

i mean autograd is teh algo that is used for most modern deep learning and will be always relevabt

[glacial root]

i mean just for now though

after this implementation i'm probably gonna be using frameworks for everything

[hearty depot]

well autograd isnt a framework tho its just an extension of chain rule and viewing the gradiens in respect to the computation graph of loss

[glacial root]

yeah but that's all of backpropagation in one method right

which would defeat the purpose of implementing it from "scratch"

[hearty depot]

i'd disagree, and in terms of like theory it allows for a greater degree of complexity and understanding modern deep learning

[glacial root]

!d numpy.gradient

[arctic wedge]

numpy.gradient

numpy.gradient(f, *varargs, axis=None, edge_order=1)```
Return the gradient of an N-dimensional array.

The gradient is computed using second order accurate central differences in the interior points and either first or second order accurate one-sides (forward or backwards) differences at the boundaries. The returned gradient hence has the same shape as the input array.

[glacial root]

a wise individual here says to not use conda

better to set up a venv and install packages using pip

[naive axle]

but from tensorflow 2.11 only wsl2 supporting gpu version

does venv can have two cuda versions

[brisk meteor]

I think I wrote it right?

[hearty depot]

U could always just build from source if that becomes an issue

[weary timber]

i tried to code a image captioner but it seems to overfit, like it outputs the same sentence for all inputs, the output sentence is a gramatically correct but compeletely out of context sentence, i tried everything but i couldnt get to fix it someone help me pls

code : https://paste.pythondiscord.com/IW2A

btw when i train it it first starts with gibberish, then transforms to the output sentence that i get all the time

[glacial root]

wait yo how would this even work

cause how would autograd know what variables to take the partial derivatives with respect to

cause there's chain rule

[hearty depot]

u do it base off the loss node, basically the idea is given the loss function u take the reversed topological sorting and use that to determine the order to find gradients
this repo covers it p well in terms of code imo:
https://github.com/karpathy/micrograd

[glacial root]

topological sorting?

damn i haven't gotten that far in dsa

that's a part of dsa right

also the approach i mentioned earlier, would it work?

while it would use some pretty large arrays, the method itself probably isn't too hard since everything is gonna be plugged into one of two derivatives (chain rule of derivative with respect to either weight or bias)

[cursive badge]

this is my training function: ```# Create function to fit and score models
def fit_and_score(models, x, x_val, y, y_val):

np.random.seed(69)
model_scores = {}
for name, model in models.items():

    model.fit(x, x_val)
    y_predict_proba = model.predict_proba(y)

return roc_auc_score(y_val, y_predict_proba)```

I'm getting a value error when both x & y, and x_h1n1, y_h1n1 have the same shape

nvm i'm dumb

[cursive badge]

need to pass y_predict_proba[:,1] into roc_auc_score

[weary timber]

pls help me with this im desperate

[serene scaffold]

Can you explain exactly what you imagine what will happen when you "connect AI to visual studio"?

If you're asking how to use a code completion assistant--which is very incredibly more specific than "AI"--that would be a question for #editors-ides

[glacial root]

yo i'm sorry but i sometimes find it funny when people say stuff like that

like "connect an ai with visual studio"

i know they probably are just unaware and have yet to learn more but

my bad for this 💀

i really shouldn't laugh, there's a ton of stuff i don't know either

[serene scaffold]

The way I said all of that probably came off meaner than I intended. But I legitimately didn't know what they were asking at first, and I'm irritated by how I'm increasingly expected to know people are talking about generative AI when they just say "AI"

I haven't heard of the kind of thing that you're talking about.

[glacial root]

nah i get that, it's definitely irritating, but also that's why i find it funny

unless someone's asking something like that to me directly, then yeah it irritates me too

i think the dude genuinely just didn't know how vast "AI" is, just like the majority of the population

the term has become more of a marketing buzzword than a field of technology

but with the way they were speaking, i presume he was just referring to copilot lol

[past meteor]

I gave a guest lecture in an MBA recently and I noticed all of the students believed all of AI was GenAI. I think that’s not a battle we can win

[serene scaffold]

Did you ask them what they thought self driving cars were?

[past meteor]

I did actually

[serene scaffold]

Also gratz on being a guest lecturer
But MBA 🤮

[past meteor]

I asked them what they believed AI was, all of them said gen ai related stuff and then I showed all interactions I had with ml/ai that morning, most of them obvs not gen ai

My colleague has a Tesla so I brought it up haha

[serene scaffold]

And does he regret buying it given present circumstances?

[past meteor]

I haven’t asked but all my management has anarcho-capitalist vibes so probably not 🫠

[iron basalt]

Welcome to the club where some of us are still on the classic definition of "rational agents," which at this point is long forgotten.

[past meteor]

That’s a fight hardly worth fighting haha

[iron basalt]

At some point a bunch of people will be on "only generative AI is AI," and will be left behind too.

Maybe it will cycle (like fashion).

[cedar tusk]

so when will we be able to create companies using ai?

[glacial root]

what did they say

[glacial root]

for real

i mean mba i guess is chill to an extent, depending on where you're going into

but bba 🤮

mba is just an extension of that but the degree can the useful for some cases, i would rather just get a masters in my field of interest though

most business majors don't know anything but yapping

[long lion]

i need personal help on a project if interested email me [e-mail]

[arctic wedge]

Click here to see this code in our pastebin.

[glacial root]

you know what battle we can win though? making companies believe they can replace business "professions" with "AI"

[plain leaf]

yo chat, Can I like scrap data with Twitter API (free-tier), for like reading the tweets under specific "target words", and use the data to fine-tune a llm? and then if a user inputs a similar "target word" the model should fetch like last 1 week's tweets with the keywords and provide insights based on that, is it possible?

[serene scaffold]

what do you want the LLM to be able to do more effectively after the fine-tuning? Be as specific as you ever possibly can, with examples.

if you just want the LLM to be able to give you a summary of a set of tweets, fine-tuning it on that tweets won't help you.

[plain leaf]

Let's say new movie releases, I want the llm to go read like past 1 or 2 days tweet on "movie title", and say it is worth watching in theatre or not

[serene scaffold]

you should not do any fine-tuning for this.
you just need to have all the tweets that the LLM can consider in a database that you query when the user asks a question, and then pass those tweets to the LLM along with the user's prompt.

[plain leaf]

so we like give a live-feel to it? like past 1 or 2 days with APIs? I mean if we storing in database, it'll not be constantly updating weekly with new tweets right.

[serene scaffold]

you periodically update the database. you do not give the LLM a live feed. you only give anything to the LLM when the user prompts it.

[plain leaf]

periodically update in the sense like manually update on database?

[serene scaffold]

whenever there's more tweets that you want the LLM to be able to use, you put them in the database.

[plain leaf]

lemme give you another examples (just trying to understand if its possible to do), Let's say a new product releases, and I make a model, which tells what to improve based on tweets from users on a 1 week timeline, like if the tweets say like the user is experiencing lags, the model should suggest like "you need to fix server lag to boost user retention" for this I need to fine-tune a llm right?

[plain leaf]

and it'll be cool, if the model can retrive tweets, in a given time frame, after the user enters the product name or company name, like "Zoom" or something like that, just curious if it's possible to make such models

[serene scaffold]

the thing that an LLM does is generate text that comes after initial text. For interactive LLMs, the thing that the user says is the initial text.

If you want it to start generating answers to questions differently than it does currently, you need to train it on examples of user interactions where it answers those kinds of questions. you need to have tons and tons of training data to do this.

[plain leaf]

YES, that's what I mean

[serene scaffold]

to be clear: you need so many tons of training data to do this, that it's almost always better to embed more information in the prompt.

[eager horizon]

usually how many viewmodels dose an app have?

[plain leaf]

Let's say I'm able to train such model. can I do the Twitter posts info retrieval in real time?

[serene scaffold]

can you be more specific? viewmodels? app?

[eager horizon]

viewmodels ie the thing after repositories

[serene scaffold]

if the shadow president hasn't shut off the twitter API, you can add new tweets to your database in real time.

[eager horizon]

and app application

[serene scaffold]

I still don't know what you're talking about. is there a library you're using?

[obsidian plume]

hi so i am a beginner to all this and i thought of making a simple chapter wise question paper classifier it takes in a pdf made out of images of pyq of a particular subject and classifies it int chapter wise on providing the list of chapters in that chapter easy right! NO it isnt.

While trying to improve it, I encountered several challenges:

1. My initial prototype worked because the physics paper was text-based without diagrams, but...
2. The past year question (PYQ) papers are only available as scanned images.
3. Converting these images to text is complex because they contain diagrams, chemical structures, and scientific notation that regular OCR cannot process.
4. I considered using MathPix, but it wouldn't be effective for chemistry papers.
5. Another challenge is handling bilingual content—questions appear in both Hindi and English. Even if I separate the English text, processing mathematical notation in the options remains problematic.

This is just the data preparation stage. After that:

1. The text needs to be processed through an LLM for classification.
2. Scientific notation needs conversion from MathPix's LaTeX format to readable text.
3. Finally, the text needs formatting and PDF conversion.

So at this point i feel like it isnt worth it as the outcome isnt that much valuable if we compare the effort its taking so should i quit this projects and how to build something useful that actually is worth it?

also im new to all this just learned python last year im ok with pandas and numpy too dont know dbms and flask will work on it after exam.

[plain leaf]

wait so does it automatically stoes tweets at set intervals, or does it only update when user requests for the "target word", grabbing new tweets with that term? (and thanks for being patient with my endless questions about this tweet-y business! 😅 you're a legend⚡ )

[serene scaffold]

up to you.

[plain leaf]

thankusomuchhh

[weary timber]

pls

[obsidian plume]

can someone pls help me

[serene scaffold]

nobody can help you until they know what the question is

[obsidian plume]

@serene scaffold here it is

[serene scaffold]

I'm not sure how to help with that, but hopefully someone else does

[obsidian plume]

@serene scaffold yea thanks btw are you the person who made the pokemon card game neural network?

[serene scaffold]

no

[obsidian plume]

@serene scaffold oh ok

[pseudo tulip]

Hi there

I'm new here

[lapis sequoia]

this is what chatgpt made 😭

[austere prawn]

4 knobs, let's start with 2 values each:
A: 1, 2
B: 64, 256
C: 0, 1
D: 8, 1024

So, 16 input combinations in total, each correspond to a scalar output value. I want to see what affect changing D from 8 to 1024 has. How to visualize? My intention was to pair up each input combination that has identical ABC inputs, compare the output of those two, and plot the resulting 8 values. Is there any plot that can help me do this grouping? I have the data in a pandas frame currently and just started using seaborne today. Catplot seems interesting, but can't figure how to utilize it for this. Any input?

[gray slate]

That's not bad for a model tha fundamentally does not understand geometry at all

[round rapids]

yeah ai is going to take our jobs

[gray slate]

It's going to take most jobs because most jobs are bullshit jobs. And that means it'll drive down wages - it'll hollow out the middle class completely

[lapis sequoia]

hold on lemme check the version as well

[gray slate]

Train it on the outputs of everyone and it'll be better than 50% of the people

[lapis sequoia]

GPT-4o

[wooden hill]

Ah.This place looks fun

[spice ravine]

which AI is best for math and physics

[glacial root]

these kinds of messages need to be restricted

as in they gotta be against the rules of this channel

[serene scaffold]

Can you explain what you think "AI" is?

[glacial root]

when doing backpropagation for a simple feed forward neural network, i should set up a vector such that each term is the partial derivative of cost with respect to either a weight or bias, then once i have the vector, i should add the terms to their corresponding weights and biases and then check if it increases the cost or decreases it, then if it increases i subtract it instead and if it decreases then i go with adding it right (in either case i would multiply the vector by a learning rate that i choose depending on how far the vector is from 0)

honestly i'm kinda lost with backpropagation, not even the goat 3blue1brown could save me

and if i'm having trouble understanding 3blue1brown then i'm definitely fried

[plain leaf]

yo chat, where to find metal scrap image data, as in like the image should have minimal rust in it, which should be recyclable, i want to know how or to where to find source (paid or free)

[small wedge]

what you calculate is the positive gradient, it will always tell you the direct to go to increase the cost. We then multiply that by -1 to get the negative gradient which tells us which way to go to decrease the cost. No need for conditionally increasing or decreasing.

[glacial root]

so then what's the learning rate

also yeah i realized how messed up my approach was earlier lol

now i realized though, i need to calculate the gradient by backpropagating to get the partial derivative with respect to each weight and bias

[small wedge]

how much you multiply the gradient by when you subtract it from the weights/biases

[glacial root]

and that learning rate is like the step we're taking right

[small wedge]

yes, it's one of the factors included in how big a step you take

[glacial root]

whatever it starts out as, from there we decrease it as we get closer to our minimum

i'm still failing to understand how to actually set up backpropagation

[small wedge]

if you're just learning about implementing backprop don't worry about decaying the learning rate, just consider it a constant step size

[glacial root]

yeah the learning rate part is pretty straight forward anyways, but my goal is to implement a full neural network

a pretty basic one though

it's for the mnist dataset

so essentially when we calculate the partial derivatives (find derivative and plug in the relevant values from what we have), that gives us the gradient right

but our gradient is what has the steepest slope

[glacial root]

i don't see how we get the steepest slope from this everytime

[small wedge]

a single partial derivative will tell you how the function changes when you tune a single parameter, if your partial is positive you know you need to go down to decrease the cost and vice versa. The gradient is just what we call all of those partials together in a tensor (or set of tensors)

[glacial root]

i see, so it's not necessarily the direction that has the steepest slope

[small wedge]

it the slope of the cost function for a single point over every trainable parameter in the model

[glacial root]

so pretty much what we do is we compute that gradient, then check to see if it increases or decreases our cost when we add it right

[small wedge]

the positive gradient (what you calculate using the chain rule) tells you the direction you need to go to increase the cost

so you just subtract those parameters instead and decrease the cost

no need to consider what the gradient actually says

[glacial root]

oh so that gradient will add to the cost

no matter what

and so we always just multiply it by the learning rate and then subtract it from our vector of weights and biases

[weak delta]

Whats a good book to master python from knowing basics
i heard about Fluent Python by luciano is that a good one or is there better
Like to write efficient good code + basically know inside out of python

[glacial root]

i'm thinking i should store all the weights and biases into one vector, and then have the gradient be in the same format so that they correspond to one another and i can just directly subtract, that's a good idea right?

[glacial root]

i think the best way to master python after you learn the basics is to just practice

no book will replace doing practice problems and projects

[grand breach]

should I fine tune CLIP on dataset or instead use cross attention to improve accuracy on image captioning task ?

[small wedge]

optimizationally yes, as far as making it easier to work with that might add a bit of extra struggle in the implementation

[pseudo lagoon]

Hello guys I am new to ai can anyone recommend any sources from where I can start learning

[dense needle]

The gradient is largest when applied to a vector in the same direction as the gradient

Or well the gradient is what it is but you don’t have to move in that direction

But the function changes by the most if you do

(So instead you move the opposite direction)

[obsidian plume]

first make sure if yore good in python learn pandas and numpy and then take courses on langchain on deeplearning.ai that what i did

[dense needle]

Also just to clarify because don’t think I explicitly said it— the gradient is the direction with the steepest slope

[cedar sentinel]

Greetings everyone. I have a very simple question yet so tricky. In the attached figure I created an end to end workflow of a model selection and evaluation. In principal, there's nothing wrong with it but the first split (to train/test sets) introduces bias although I bootstrap the test set in the final evaluation module. My question is: How can I do it robust ? If I repeat the exact same workflow 1000 times, I'll get a new "best configuration" and that's not what I want. I work with tabular data of 835 samples for a binary classification task with 80 features and imbalanced of 65-35 %.

Someone could argue that the 1st split (train/test) basically corresponds to reality and whatever the result is this should be reported.
Others could argue that the 1st split (train/test) might lead to poor scoring for the best configuration hence split was done once and sets might not be ideal.

[olive obsidian]

Hi, I'm trying to understand the use of numpy a bit better (in relation to ai/ml) and I often see something like: img = np.random.randint(<someting>); img.transpose((2, 0, 1)); and I'm trying to understand what the transpose() call really does.

From my understanding, in this case, it converts the pixels which are stored in memory like [r0,c0, r0,c1 ....]` (e.g. <r>=row, <c>=column, pixels are stored packed per row), into a planar format where the R, G and B values are stored as separate planes. Is that correct?

[wooden sail]

simply put, it's reordering the axes of your data. using your example, if the data is originally stored as [r,g,b], then img.transpose((2,0,1)) makes it so that it is now [b,r,g]

whenever possible, this is done only by changing the stride and not by rearranging the entries in memory

[olive obsidian]

Thanks! So the data is already planar? Instead of packed?

[wooden sail]

idk what you mean with "as separate planes"

since numpy falls back on c to handle memory and what not, what is done is that the full required memory (rows * cols * size of each element) is allocated as a contiguous chunk. you then have two options: store the data in row major, or column major order. numpy uses row major by default

maybe this helps

[olive obsidian]

Ok I understand, I'll try to explain better.

Normally when you load/decode an image like a png or jpg, the pixels are stored in memory as r0,g0,b0,r1,g1,b1,r2,etc. However from what I've learned so far, the AI/ML worls uses pixels where they are stored/extracted, like: r0,r1,r2,r...for red, and b0,b1,b2 etc.

[wooden sail]

how you index them and how they are stored in memory are two separate things, btw

you usually don't need to care about how they are stored in memory

[olive obsidian]

In my case it is important how they are stored in memory because I'm loading them in C++ and feeding them into TensorRT.

[wooden sail]

ah in c++ you can choose this yourself

the first order you showed would be column major

[olive obsidian]

When I feed packed R0,G0,B0,R1 instead of planar R0,R1,R2, G0,G1, etc. the inference results are completely wrong

[wooden sail]

why do you call it packed

if you look at the image i shared, it's exactly what i showed there

just reading the data in one order or the other

[olive obsidian]

I call them packed/planar and not column/row major because I talk about how the bytes are stored in memory

[wooden sail]

i mean, row and column major refers exactly to how the bytes are stored in memory

that's precisely what the words are used for

[olive obsidian]

ok

[olive obsidian]

I also tried to focus on how the red, green and blue values are stored in memory. But ... from your first answer I think that they are already separated/planar.

and the .transpose() only flips the order like you said

[wooden sail]

transpose does not change the memory layout

it changes the stride only

the two things are completely separate. you can keep your memory layout as is and just change how you map the memory to a matrix

[olive obsidian]

Ah ofc. The reason why I was asking, is because I'm curious why they use transpose here:

def convert_onnx(net, path_module, output, opset=11, simplify=False):
    assert isinstance(net, torch.nn.Module)
    img = np.random.randint(0, 255, size=(112, 112, 3), dtype=np.int32)
    img = img.astype(np.float)
    img = (img / 255. - 0.5) / 0.5  # torch style norm
    img = img.transpose((2, 0, 1))
    img = torch.from_numpy(img).unsqueeze(0).float()

[wooden sail]

because of what the math is doing

[olive obsidian]

My guess was, that they want to show how to reprocess the input data.

[wooden sail]

not because of the memory layout

from the exact code you shared, it's not even needed. you could just directly create the array with the shape 3, 112, 112. but later on when the data is not randomly generated, it can make a difference

[olive obsidian]

yeah that's why I thought they were using transpose() to clarify something ... like how the preprocessed input for inference should look like

[wooden sail]

sure

[olive obsidian]

so that you can make sure that the memory layout in e.g. C/C++ is correct

[wooden sail]

still has nothing to do with memory layout

[olive obsidian]

not in python maybe, but in C you have to make sure the input data is preprocessed in such a way that the pixels in memory are as expected.

[wooden sail]

also no

you're mixing up mathematical objects with programming objects

you can store the entries of a matrix however you like in memory as long as you apply the math to them in the correct way

[olive obsidian]

If "no" please tell me why remapping a loaded image into separate R,G,B planes gives me correct results when I feed it into a face detection model and that it doesn't work when I keep the pixels as is?

[olive obsidian]

yes of course 🙂

but it's about what the model expects

[wooden sail]

because you are treating the two things as if they were the same

it's because of how you wrote your code

[olive obsidian]

yes ofc 🙂

[wooden sail]

you made the two things the same, when they don't have to be

you could keep the memory layout the same and change the stride instead

[olive obsidian]

sorry, maybe I should have shared a bit more context

the thing is that I'm working with a .onnx model which I converted into a tensorrt engine. I use this model to learn about ai/ml and as a test I'm creating a face detection app.

so I load this model and want to feed images for inference into it.

what I found difficult is to figure out how to correctly preprocess the image data before feeding them into (any) model that expects images.

most repositories seem to share a python script which converts a pytorch model to onnx and I thought maybe those scripts share/explain how the data is supposed to look as they all look similar.

[wooden sail]

those scripts focus just on the "shape" of the data, e.g. that it's an m x n matrix

[olive obsidian]

ok thanks for explaining

[limber belfry]

Ill give you a code that is basically a tester for your model when i get home

[glacial root]

but if that's the case wouldn't we always be going in the direction with the steepest slope

or is the vector we find not always exactly the gradient

[gray slate]

The loss landscape is a lumpy mess, filled with hills and valleys. Part of shaping the architecture and training data is to smooth this out so we can "roll the ball down the hill" - descend the gradient - without getting stuck in a pool that isn't at the bottom (a local minima, rather than a global one)

Part of this is "feature engineering" and part of it is model design, the architecture

[left sapphire]

with open('smallbatch.jsonl','r') as f:
    data = [json.loads(line) for line in f]
    df = pd.json_normalize(data)

df.head()

Hi, pandas / json question. Does anyone know why the json_normalize() function is failing to flatten after it gets a few levels down? (response.body.choices)

[mental ether]

any idea on rocm support on wsl vs nvidia setup to train a deep learning model ?

has 2 setup where one is old laptop with rtx 1060, another one with amd 7900 xtx. but from what i have been reading online, wsl setup on window with amd gpu is annoying ??

[lilac lichen]

hi guys. I am deciding about ml hardware, but pytorch and tensorflow requirements are not obvious. i am seeking for cheap variants and so cannot be sure about specific card being capable of running frameworks due to card being obsolete. how to check compatability properly?

[glacial root]

i see

so model design is the setup of our multilayer perceptron right

as in how many layers we have and how many nodes in each layer

[gray slate]

Yeah but also the fact that you use the perceptron pattern, which parts are plumbed into each other, the flow of the information as gradient descent adjusts the loss landscape to shape to fit the overall function to the training data

[glacial root]

right now i'm just worried about getting the concepts down and trying a basic implementation from scratch, once i get backpropagation/gradient descent set up and have the algorithm which gives me a set of weights and biases that will give me a relative minimum cost, then i will look into all the ways i can improve the accuracy through hyperparameter tuning and improved model design

[gray slate]

I highly recommend "The Ancient Secrets of Computer Vision" for a deep primer on feature engineering and the sort of cool hacks that are used to build vision models. It's long, but it's a full course by the guy who wrote YOLO:
https://www.youtube.com/watch?v=8jXIAWg_yHU&list=PLjMXczUzEYcHvw5YYSU92WrY8IwhTuq7p&index=2

[glacial root]

thank you, i'll look into it

[gray slate]

Something a bit shorter and a bit more direct:
https://www.youtube.com/watch?v=TkwXa7Cvfr8

[tidal garnet]

Does locally training any LLMs or any ml models (like say diffusion models) makes any sense?
Or is cloud the way to go?
Asking cause I wanna buy a new gpu. I am not a pro but much better than a total beginner. Would keep this in mind.
how much would i be able to do with say rtx 4070 (thats the max i can go). Or I should just stick to colab/kaggle?

[merry ridge]

I don't really train that many models so I'm not a very good source of an opinion here: I own a 3080 and I think it has the same amount of VRAM and more cores? It felt like a waste of time to me.

[serene scaffold]

The amount of GPU power you might need varies wildly depending on what you're trying to do. Your decision about which GPU you get for personal use should only consider how you'd use it for gaming

If you can also use it for ML, yay

[tidal garnet]

I see. and i dont think and gpus are any good for that right?

[serene scaffold]

Good for what?

[tidal garnet]

oh i meant AMD gpus.
it autocorrected 😅

[serene scaffold]

Yeah, you'll want to stick to Nvidia.

[tacit sorrel]

anyone here been working with this mcp library? https://github.com/modelcontextprotocol/python-sdk?tab=readme-ov-file

I legit cannot get even the quickstart to work and I want to make sure its not just me

[iron basalt]

AMD works, but not all libraries support it, or support it well. It's the best option for price, but at the tradeoff of ecosystem / support.

Considering how difficult it is to get any Nvidia GPU at this point, it may be your only option.

(Other than using the cloud)

The key thing you would be looking for in a GPU is tensor core count, and VRAM (at least 16 GB, but you always want more).

Also not all generations are equal, they may have the same count, but tensor cores have improved a lot recently.

[tidal garnet]

at least 16?
I was more in the range of a used 3080 or maybe new 4070tbh.
would they not be okay even for learning and like a proof of concept thing?
ig it won't be worth it considering the model training part for gpu at all then haha?

[iron basalt]

You can use them for learning machine learning on more simple tasks. Just don't expect too much, they will be toy projects.

[tidal garnet]

I see. and if I go with cloud for getting similar experience, is that expensive?

[iron basalt]

That also depends on how much you are expecting. If your expectation is anything in the ballpark of ChatGPT or self driving cars, then I regret to inform you that this has gone outside the range of what an individual (usually (are you a billionaire?)) can afford.

But it will probably be cheaper and get you further. However, this depends a bit, since there is a curve here. If you have enough funding to buy enough of them and basically become your own mini cloud it can be worth it again.

This also depends on if the cloud provider you pick is currently willing to lose (go negative) to attract users, in which case it would be a good deal for you and better than doing it yourself (and many cloud providers are currently doing this, waiting until everyone is locked into their services to increase the price).

Overall, probably go with cloud. It will probably be cheaper and they have the biggest, most modern GPUs (that are not even consumer GPUs, they are better than those).

[tidal garnet]

na na lol. it's just personal level.
for doing to understand on my own.
and perhaps building something by fine tuning existing LLMs specially open source ones

[iron basalt]

If it's just for learning on your own, then a used 3080 is fine.

Main issue would be VRAM, since it's only 10GB.

(DDR6X)

[tidal garnet]

I see.. how much is good?
cause it only gets 12gb with 4070

[iron basalt]

For reference, the 5090 is 32GB (DDR7). And the 3080 has about 272 tensor cores (older generation cores) while the 5090 has 680 modern ones.

[iron basalt]

The 4070 is bluntly a scam.

For the price.

[tidal garnet]

I see.

then 3080 used would be the max

[iron basalt]

Nvidia likes to mess with their consumers a lot, you have to navigate their tricks.

[merry ridge]

If you are a university student, one option is to just ask your professors if they have compute power laying around and maybe turn it into a summer research project

[iron basalt]

Also the 4070 is not what you want, it's targeted at gamers.

Less tensor cores. It just has more VRAM for textures and models.

[merry ridge]

My department has a fair bit of resources that are mostly a polite request from a good student away from being made available to them

[iron basalt]

If your university has like a Nvidia workstation that is great (mutliple GPUs all interlinked).

(4 V100s interlinked, each 32GB VRAM, and 640 tensor cores)

[tidal garnet]

nah. I just graduated lol. been thinking to dive into ML again.

[vital vessel]

Hey, anybody online? I am self taught and don't know much about Python. I had to use a chromebook and webbrowser programming things like CodeHS to learn 😭 I just got an actual computer and I want to make JARVIS.

[serene scaffold]

What do you mean by "JARVIS"? Because the Jarvis in the marvel films is more advanced than the most advanced AI that currently exists. By a lot.

[vital vessel]

Yeah, I just mean I want something with a similar concept

I know it's not gonna be nearly as good

I want something that I can talk to and get a verbal response. I would (preferably) like to have it voice activated but that's a big what if

Basically Glorified ChatGPT

I think it's called machine learning? Idk. I want it to be able to remember some stuff if that makes sense

[serene scaffold]

Glorified ChatGPT. You won't be able to create an interactive language model that's anywhere near as good as ChatGPT. Those models train for months with terabytes of training data on enterprise systems.

[vital vessel]

Again, just using these as examples lol

[serene scaffold]

You're currently trying to operate at level 100. You need to bring it way down to level 1.

[vital vessel]

fair enough

I had something that was working but it had limited commands

and I had to push a button before it would start listening

[serene scaffold]

A beginner project you could do is make a spam detector

Or something that decides if a picture is of a dog or a cat

[vital vessel]

Also, what can I use for these things? I have VSCode installed and I think I have the python extensions for it installed too, but idk how to tell if they work or not

[serene scaffold]

Vscode is just a code editor. You can use whatever code editor you want.

It has no bearing on how your code will run.

Just don't install anaconda

[vital vessel]

hahaha, why?

[serene scaffold]

They've been gaslighting ml people into thinking it's still relevant

[vital vessel]

oh

[serene scaffold]

I'm signing off now.

Don't do anything bad

[vital vessel]

okay lol

I'll ask around and try to learn some more

thanks for the help

[serene scaffold]

I'll probably be back in like nine hours

[vital vessel]

alright

[glacial root]

you seem to be in the same boat as me

trust me, we are no where close to recreating jarvis lol

not just us, mankind as a whole

what i'm doing right now and what i think you should do as well, is implementing a basic feed forward neural network using just numpy

and pandas but only to parse the data and get it into a numpy array, nothing else

before that probably learn some basics of linear algebra and multivar calc though

and then the rest you could probably learn as you go

[gloomy swan]

If anyone is looking to get some open-source experience with AI Engineering and LangChain, I have a repo which has some issues to tackle, probably more intermediate than beginner-level stuff.
https://github.com/GGyll/condo_gpt

In the readme there is a video that explains the project in case you like visual explanations 🙂

[brisk bay]

does anyone know a fix for PIL UnidentifiedImageError?

basically a lots of images in my dataset are corrupted and I can find a way to ignore corrupted images in dataset during testing loop

[past meteor]

Did you join this server just to advertise your project? 😅

[odd meteor]

Isn't this what exception handling can take care of? Or perhaps, it might be better to find all corrupted images and remove them from your data.

[ornate latch]

Guys anyone interested in doing some deep learning projects with me?
I am learning so yeah if anyone would like to join in guiding the workings of DL architectures & other stuff.

[gloomy swan]

I just joined and I posted it, but not really tbh.

[thorny zealot]

How much data science should you learn before learning machine learning?

[woeful escarp]

Hello, I am starting in ML, I would like to work in a project to improve, send me DM

[glacial root]

i think the question should be how much math should you learn before ml

[lament warren]

Hi, I'm working on a small project to approximate polynomial function using tensorflow.

I use 30000 datapoints from x = [-1, 1] and use the polynomial to create the corresponding y dataset

I then create a model using this general format

model = tf.keras.Sequential()

model.add(tf.keras.Input(shape=(1,)))

model.add(Dense(12, activation= 'relu'))

model.add(Dense(1, activation='relu'))

This is just as an example

I train it at 20 epochs with batch size 12 with validation data (data is split in 30% training, 20% validation and 50% testing)

It takes about a minute to train the model, is this normal?

[opaque condor]

https://paste.pythondiscord.com/AQRA
Is this convelution net ok I do see an issue but I don't know how to solve it

[hearty depot]

this isn't a convnet

[opaque condor]

?

[hearty depot]

i dont see any convolution layers, this looks like a standard mlp

[opaque condor]

I was trying to get it to work

[glacial root]

yo do you think it's worth it to work on an implementation from scratch for each of the most important neural networks

like convolutional and recurrent

also by the way i got an implementation of the gradient function working, i'm still not sure how the numpy method for gradient works

cause it doesn't give us the individual changes for each weight and bias that we need to make

[hearty depot]

Don’t use numpy gradient method u got write ur own for ur operators in tensor class

[glacial root]

tensor class?

also yeah i wrote my own

[glacial root]

UnboundLocalError: cannot access local variable 'hl2_to_output_weight' where it is not associated with a value

does this mean out of bounds

[serene scaffold]

remember to always always show the code also.
are you getting this in a lambda or what?

[viral wadi]

I need help

whenever i use import cv2 or whatever i am importing it says module not found

[serene scaffold]

you can have more than one python on your computer. if you're sure you installed something, and you're getting an import error, you probably installed it to a different python.

[viral wadi]

how am fix

[serene scaffold]

import sys
print(sys.executable)

add that to your code to see which python your code is using

[viral wadi]

@serene scaffold

[serene scaffold]

please always copy and paste text directly--don't ask people to read screenshots of text

[viral wadi]

oh sorry

[serene scaffold]

what did you do to install cv2?

[viral wadi]

wym

im new to python

[serene scaffold]

you get an import error when you try to import cv2, right?

[viral wadi]

btw

yh

[serene scaffold]

what did you do to install cv2?

[viral wadi]

i do not understand

i did import cv2

[serene scaffold]

you have to install cv2. it doesn't come with python.

[serene scaffold]

please give the text in this screenshot. I need it for something.

[glacial root]

no i'm not using a lambda, is it fine if i sent the notebook file

[viral wadi]

then how do i run it on someone elses computer without them knowing

🥀

[serene scaffold]

No, it needs to be flat text (not a json, which is what a notebook is).
you can do python -m jupyter nbconvert --to script --stdout your_notebook.ipynb

[serene scaffold]

without them knowing that it's running??

[viral wadi]

😊

they can know whats running idc

but i need to test my webcam and part of my script is screwed

[serene scaffold]

please do this if you want to continue.

[viral wadi]

ok but can i edit out the nane of my pc

bc its embarrassing

[serene scaffold]

with your keyboard

[viral wadi]

ok

what the debug said or what i typed in

[serene scaffold]

this

[viral wadi]

ok

[serene scaffold]

@glacial root still there?

[glacial root]

yeah the command wasn't working so i just exported it as a python file

[arctic wedge]

Click here to see this code in our pastebin.

[glacial root]

should i click the check

[serene scaffold]

It said please

[glacial root]

or is this fine

[serene scaffold]

which line causes the error?

[hearty depot]

are u using

venv?

[viral wadi]

visual studio

[glacial root]

one sec i'll send a screenshot of the entire error

[viral wadi]

PS C:\Users\pcname\Downloads> & 'c:\Program Files\Python313\python.exe' 'c:\Users\pcname.vscode\extensions\ms-python.debugpy-2025.4.0-win32-x64\bundled\libs\debugpy\launcher' '59458' '--' 'C:\Users\pcname\Downloads\import sys.py'

[serene scaffold]

Please do not post screenshots of text.

[hearty depot]

u can still use venv in visual studio

can u do pip list?

in terminal?

[viral wadi]

yh

[glacial root]

oh then i'll copy and paste it

UnboundLocalError Traceback (most recent call last)
Cell In[11], line 8
6 if array_max > max:
7 max = array_max
----> 8 gradient_descent(cost_gradient, max)

Cell In[9], line 3, in gradient_descent(cost_gradient, max)
1 def gradient_descent(cost_gradient, max):
2 if max > 1:
----> 3 gradient_descent_step(cost_gradient, 0.01)
4 elif max > 0.0001:
5 gradient_descent_step(cost_gradient, 0.001)

Cell In[8], line 2, in gradient_descent_step(gradient_vector, learning_rate)
1 def gradient_descent_step(gradient_vector, learning_rate):
----> 2 hl2_to_output_weight = hl2_to_output_weight - (learning_rate * gradient_vector[0])
3 output_bias = output_bias - (learning_rate * gradient_vector[1])
4 hl1_to_hl2_weight = hl1_to_hl2_weight - (learning_rate * gradient_vector[2])

UnboundLocalError: cannot access local variable 'hl2_to_output_weight' where it is not associated with a value

[serene scaffold]

@viral wadi okay, do this command.

'c:\Program Files\Python313\python.exe' -m pip install opencv-python

[viral wadi]

ok

didnt work

invalid syntax

wtf

[hearty depot]

also @viral wadi
i recommedn using wsl in the future it will make life easier

[serene scaffold]

you run it in the powershell terminal. not as python code.

[viral wadi]

ok good

[glacial root]

also can you tell me just enough to give me a slight hint

[serene scaffold]

I think that's too advanced for them for the moment

[serene scaffold]

I'm working on it.

[viral wadi]

The filename, directory name, or volume label syntax is incorrect.

[glacial root]

sorry this is what i meant

the way i said it earlier may have come off as a bit rude

[viral wadi]

WTF

[serene scaffold]

@glacial root

def gradient_descent_step(gradient_vector, learning_rate):
    hl2_to_output_weight = hl2_to_output_weight - (learning_rate * gradient_vector[0])
    output_bias = output_bias - (learning_rate * gradient_vector[1])
    hl1_to_hl2_weight = hl1_to_hl2_weight - (learning_rate * gradient_vector[2])
    hl2_bias = hl2_bias - (learning_rate * gradient_vector[3])
    input_to_hl1_weight = input_to_hl1_weight - (learning_rate * gradient_vector[4])
    hl1_bias = hl1_bias - (learning_rate * gradient_vector[5])

It looks like you're trying to change global variables in a function, but that's not how it works.

[viral wadi]

@serene scaffold

its not working

[serene scaffold]

I need to know what you did and the whole error message.

[glacial root]

oh so even if they're global variables i have to put them as arguments

[viral wadi]

the error is "The filename, directory name, or volume label syntax is incorrect."

[glacial root]

dude

[serene scaffold]

do you know the difference between variables and objects?

[glacial root]

you should probably send him the whole thing

[opaque condor]

I followed the tutorial what can I say

[viral wadi]

i closed it so ill redo it

[glacial root]

yeah

but there are no objects here

[serene scaffold]

@viral wadi if you run code and it causes an error, I need to see the whole code and the whole error.
if you run a powershell command and it causes an error, I need to see the command and the whole error.
always send both parts in the same message.

[viral wadi]

C:\Windows\System32>-m pip install opencv-python
'-m' is not recognized as an internal or external command,
operable program or batch file.

C:\Windows\System32>

[glacial root]

i'm not gonna lie i kind of forgot about objects, it's been a while since i've worked with java and in python i haven't really worked with classes

but i see the issue now

[serene scaffold]

you're missing the first half.

'c:\Program Files\Python313\python.exe' -m pip install opencv-python

[hearty depot]

i'd prob learn the math behind ml first then blindly following tutorial
it'll help u more in the long run too

[viral wadi]

I tried that too

C:\Windows\System32>'c:\Program Files\Python313\python.exe' -m pip install opencv-python
The filename, directory name, or volume label syntax is incorrect.

[serene scaffold]

try making the C upper-case

[viral wadi]

ok

C:\Windows\System32>'C:\Program Files\Python313\python.exe' -m pip install opencv-python
The filename, directory name, or volume label syntax is incorrect.

C:\Windows\System32>

@serene scaffold

[serene scaffold]

download and install this. https://cmder.app/

[viral wadi]

i no no wanna get ratted

ok then

[serene scaffold]

idk what you mean by ratted

most developers use a bash console, not powershell.

[viral wadi]

ok

[serene scaffold]

I don't know enough about powershell to keep helping you.

[hearty depot]

@viral wadi if u just need to script and dont want to deal w windows environemnt other option is to juse use colab

[serene scaffold]

everything in python is an object. there is no escape.

[delicate apex]

RAT malware - remote access trojan

[hearty depot]

🤓 blud about to spill some generational yap about pyobjects

[serene scaffold]

# creates a numpy array object and assigns it to that variable
hl2_to_output_weight = np.random.randint(-5, 5, (10, 15))

# creates a new object and overwrites the original variable, but only for the current functions
hl2_to_output_weight = hl2_to_output_weight - (learning_rate * gradient_vector[0])

[viral wadi]

"how to rat someones pc and delete system 32😊"

@serene scaffold

[serene scaffold]

idk what could have caused that. you could try using google colab instead.

[serene scaffold]

this is for you, @glacial root. sorry for not making that clear.

[viral wadi]

i got it

i had to do the easiest thing ever

[serene scaffold]

great.
open cmder and type which python

[viral wadi]

is it supposed to be transparent

[serene scaffold]

doesn't matter

[glacial root]

oh wait yeah i forgot about that

[viral wadi]

λ which python
/c/Program Files/Python313/python

[glacial root]

i see

[viral wadi]

stelercus what happen

[glacial root]

so should i just not use a function

[viral wadi]

λ which python
/c/Program Files/Python313/python

@serene scaffold

[glacial root]

dude

chill

give him time lol

[viral wadi]

NO

[serene scaffold]

YES

[viral wadi]

ok

[glacial root]

even functions are objects right

[serene scaffold]

okay, do python -m pip install opencv-python

[glacial root]

i find that wild

[opaque condor]

I know somebody already I commented but I still understand the equations it's just my brain is it comprehending what's the error

https://paste.pythondiscord.com/AQRA

[glacial root]

i had a stroke reading this

[viral wadi]

please be quiet mechanical fox

[glacial root]

bro what

you are not the only human being in this chat

[serene scaffold]

you're not more entitled to help than anyone else

[viral wadi]

oh

[glacial root]

...

[viral wadi]

when will you help me

stelercus

[glacial root]

bro...

[serene scaffold]

this

[glacial root]

if you don't mind me asking

how old are you

@viral wadi

[viral wadi]

ty

and i do mind you asking

where can i go to learn python

does duolingo have it

[serene scaffold]

that's for human languages. but it's not even good for that.

[viral wadi]

:C

[serene scaffold]

@viral wadi in cmder, do python -m pip install opencv-python

[viral wadi]

i did

[serene scaffold]

okay, what were the last few lines of the output?

[viral wadi]

Successfully installed numpy-2.2.3 opencv-python-4.11.0.86

[notice] A new release of pip is available: 24.3.1 -> 25.0.1
[notice] To update, run: python.exe -m pip install --upgrade pip

[serene scaffold]

okay. where in your computer is the python program you want to run?

[viral wadi]

uhhhh wdym

like you mean a text py file

[glacial root]

oh sorry

but the reason i was asking

was that i just want to let you know

you gotta be more patient

and you gotta understand that others are people as well

[serene scaffold]

yes

whatever folder it's in, you need to use the cd command to go to it

[viral wadi]

ok ill make one

downloads

[serene scaffold]

so it might look like cd C:\Program Files\my_stuff\example
and then once you're there, you do python the_code.py

where the_code.py is your python file.

I'm going to have dinner, so I'll be back later.

[viral wadi]

ok bye

its confusing what do i swap the mystuff and example with

[glacial root]

condition = True
while(condition):
    cost_gradient = gradient_cost()
    max = 0
    for parameter_array in cost_gradient:
        array_max = np.max(np.absolute(parameter_array))
        if array_max > max:
            max = array_max
    if max > 1:
        hl2_to_output_weight = hl2_to_output_weight - (0.01 * cost_gradient[0])
        output_bias = output_bias - (0.01 * cost_gradient[1])
        hl1_to_hl2_weight = hl1_to_hl2_weight - (0.01 * cost_gradient[2])
        hl2_bias = hl2_bias - (0.01 * cost_gradient[3])
        input_to_hl1_weight = input_to_hl1_weight - (0.01 * cost_gradient[4])
        hl1_bias = hl1_bias - (0.01 * cost_gradient[5])
    elif max > 0.0001:
        hl2_to_output_weight = hl2_to_output_weight - (0.001 * cost_gradient[0])
        output_bias = output_bias - (0.001 * cost_gradient[1])
        hl1_to_hl2_weight = hl1_to_hl2_weight - (0.001 * cost_gradient[2])
        hl2_bias = hl2_bias - (0.001 * cost_gradient[3])
        input_to_hl1_weight = input_to_hl1_weight - (0.001 * cost_gradient[4])
        hl1_bias = hl1_bias - (0.001 * cost_gradient[5])
    elif max > 0.00000001:
        hl2_to_output_weight = hl2_to_output_weight - (0.0001 * cost_gradient[0])
        output_bias = output_bias - (0.0001 * cost_gradient[1])
        hl1_to_hl2_weight = hl1_to_hl2_weight - (0.0001 * cost_gradient[2])
        hl2_bias = hl2_bias - (0.0001 * cost_gradient[3])
        input_to_hl1_weight = input_to_hl1_weight - (0.0001 * cost_gradient[4])
        hl1_bias = hl1_bias - (0.0001 * cost_gradient[5])
    else:
        condition = False```

@serene scaffold this is my function now

how long can i usually expect this to take to run

so far it's been like at least 2 minutes and my computer sounds like a fan lol

[serene scaffold]

it really depends. you usually want to write it in such a way that it periodically tells you the current loss.

[glacial root]

wouldn't that be even more computationally heavy though

cause it would have to recalculate the loss every time right

though that takes like 5 seconds

[serene scaffold]

you're always calculating the loss. you just need to print it sometimes

[glacial root]

oh i didn't have it calculating the cost every time

but now i changed that and am having it print every time

[serene scaffold]

printing it every time will be annoying for the human

[hearty depot]

why not just using logging module so u can have that when u r debugging

[glacial root]

it's just a for loop though

one sec i'll send the code

oh wait it goes over the limit

i'll use that github thing to send code, i forgot what it's called

[glacial root]

what's logging module

also i'm doing this on jupyter notebook

[serene scaffold]

I'm winding down for the day, so I probably won't be answering questions in the immediate future.

[hearty depot]

https://docs.python.org/3/library/logging.html
it's kinda like print() but u can have it trigger based on certain conditions like whether u r debugging or for some some other reason

[glacial root]

i see

oh wait i remember what it's called now

github gist

https://gist.github.com/AnishM101/ec5adbef1fd67d2feb684402e54a5d1e

this is the code for it

only the first element in the cost vector has been going down

and very slowly, except for one time when it just shot down and then went back up to relatively where it was before

[hearty depot]

is there a reason why u arent summing up mse?

[glacial root]

i am

i'm taking an average though

average across all training examples

also perhaps the reason why it's so slow is cause i'm using a cpu

[hearty depot]

u arent summing up the components tho, u should get a scalar

[glacial root]

do you guys usually pay for a gpu

[glacial root]

oh that's what you mean

it doesn't really make too much of a difference though right

[hearty depot]

it does tho

ur supposed to have the loss for mse as a scalar and not doing his impacts

backprop

[glacial root]

so it should be done before i take partial derivatives?

[hearty depot]

yes, u should get the partial after u get mse
in fact a lot of modern ml liubraries dont let u call backwards() unless the tensor is a a scalar

[glacial root]

i see

[glacial root]

is this also why it's so slow

[hearty depot]

Run it under cProfile to benchmark bottlenecks

[glacial root]

i'm sorry but i have no idea what that means

[past meteor]

it's a a profiler, a way to figure out what the bottleneck in your code is

[iron basalt]

You are using Python (without numpy), also use a smaller batch size or just SGD.

[glacial root]

i have no idea what a profiler or a bottleneck are

[glacial root]

i'm using numpy

setting up a neural network without numpy would be diabolical

[iron basalt]

Also you don't really compute these things in separate passes always. It's like if you add 2 to every value in an array and then multiply it by 3. You could do this in a single loop. But what you are doing is like having a loop that adds 2 to each value, and then another after that multiplies them by 3.

[glacial root]

and yeah i cut down the batch size from 50k to 5k training examples

[glacial root]

i see

[iron basalt]

See if you can combine loops into one.

Do things all at once when you have the info right there.

Note that this separate loops stuff still happens when using numpy, just it's hidden.

[glacial root]

yeah that's what i did

[iron basalt]

And it's why it's slower than doing it in plain C still.

[glacial root]

like with setting up the layers and the cost function

[glacial root]

i did all of that in one loop

same for backpropagation wherever i could

[glacial root]

i see

makes sense

[iron basalt]

The real issue is going to be large loops in Python.

Even though numpy is not optimal it's so much faster than pure Python that it does not really matter.