#data-science-and-ml

1 messages · Page 138 of 1

past meteor
#

I haven't started the new role yet but I'm curious how big their maturity is on keeping costs down

lapis sequoia
#

pretty nice summary:

SIMD is the 'concept', SSE/AVX are implementations of the concept. All SIMD instruction sets are just that, a set of instructions that the CPU can execute on multiple data points. As long as the CPU supports executing the instructions, then it is feasible for multiple SIMD instruction sets to coexist, regardless of data size.

past meteor
#

In my current role people are just happy if it's running

#

My boss keeps paying absolutely insane cloud bills without knowing why they're so high

buoyant vine
#

I guess it depends what market you're in

#

I am in adtech so cost per user/site/operation has an big impact

past meteor
#

Right now I'm in R&D with many people that aren't really concerned about going to prod

#

The worst thing is that one guy decided to put all our services in 1 azure resource group

#

Drives me crazy

buoyant vine
#

How is the whole Azure AI experience? I could never understand their naming schemes or pricing of services

past meteor
#

To figure out the cost of a project you need to find/filter each individual resource and tally up their costs instead of just taking the entire RG

#

Hmm, in all honesty. We do all ML/AI stuff on prem

#

We have 2 Quadro cards, it's very comfy

buoyant vine
#

Fair enough

past meteor
#

The place I'm moving to does everything in databricks so that's that

#

I'm not sure how I feel about being bound to spakr for everything

#

I guess streamlining your stack saves engineering time

buoyant vine
#

Tbh imo, ML/AI wise, as long as what ever you are doing can go into ONNX, you're entire deploy setup and infra side is just a breeze

past meteor
#

I don't want to get philosophical but

#

The bigger issue is that the people working on the models are typically not engineers and have never heard of ONNX

#

But my sample size are just people I know irl of course

buoyant vine
#

So I agree, but I argue that it is not hard to introduce them to onnx. Especially if you're using something like PyTorch, really all it becomes is a "Hey just add this 10 lines of python" or you can do it via CI or some CLI tool for example

#

Sometimes a magic script is the best kind of script

past meteor
#

I see what you mean

#

I'm projecting a bit, most of my colleagues deliberately scoped things to never have to worry about prod

#

which is entirely possible in R&D, just call it something like "proof-of-value" and then it's done

buoyant vine
#

Yeah, I mean I think that is probably a good thing, but maybe the missing step there is "we create this model, send it to the blackbox in CI that does the training and then spits back results"

past meteor
#

Which is unironically what I was trying to make

buoyant vine
#

Effectively what I ended up doing for our web classifier model which I actually don't understand why it isn't a mainstream library, is a framework that lets multiple people work on seperate models within the repo, and then run them via just run_model("my.path.to.pthon_file:MyClass") and then everything else is handled for them, which also lets me keep the system generic and abstract enough to automatically convert the models, load datasets, etc...

#

Probably the best bit of time I invested in and 2k LOC was setting that up, because it just made life so much better

past meteor
#

What was your reasoning of not just having them give you the trained model

#

Is it something they couldn't do from their machines?

#

Or machines they have access to

buoyant vine
#

They depend on CI /remote servers for training since which multiple GPUs and 24GB VRAM

#

and they are all working on the same repo

#

the biggest issue is when you have multiple people working on different models to do the same thing in order to compare, is you end up with them writing lots of bits of random code all over the place and different entrypoints to train and test the models

past meteor
#

Specifically CI for model training is interesting because I can think of many times I'd want a new mode in a way that is decoupled to my commits

#

Like, new data arriving

#

But I suspect your use case is vastly different from the ones I have in mind

buoyant vine
#

this approach meant the entry points for training, testing and exporting all are the same across models, and the testing system didn't have to be specialized for each model

buoyant vine
#

the only painpoint is the system never really ended up setting up checkpoints

#

so if it failed right at the end you'd be 🫡 Waiting another 24 hours

lapis sequoia
#

pheeeww

covert cave
#

hi, from sklearn.preprocessing import MinMaxScaler,StandardScaler sc=StandardScaler()
X_train= sc.fit_transform(X_train)
X_test= sc.fit_transform(X_test)
for these codes I got this error: TypeError: Property names are only supported if all input properties have string names, but your input contains ['str', 'tuple'] as property name/column name types. If you want the property names to be stored and validated, you should convert them all to strings, for example using X.columns = X.columns.astype(str). Otherwise, you can remove the property/column names from your input data or convert them all to a non-string data type. do u know?how can I fix this error?

#

How do you suggest I do this?

past meteor
#

Put it in 3 backticks like so ``` your code ```

past meteor
# covert cave How do you suggest I do this?

Also, you can put python between the 3 backticks, like so:

```python
X_train= sc.fit_transform(X_train)
X_test= sc.fit_transform(X_test)
```

that gives you

X_train= sc.fit_transform(X_train)
X_test= sc.fit_transform(X_test)
covert cave
serene scaffold
#

!paste

arctic wedgeBOT
#
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.

past meteor
covert cave
#

I solved it now , It worked after disabling name verification like X_train= sc.fit_transform(X_train.values) . thanks a lott

errant bison
#

I'm trying to create an ai which helps in traffic congestion. How can i do that

serene scaffold
orchid forge
#

what does this "notions of dimensional modelling, ETL and the basics of data engineering go a long way" part means?

errant bison
spare forum
#

also there is different data architectures note that etl isn't the only way to go

past meteor
umbral delta
#

so i have a bunch of images that look like this, along with the same image with a red outline around the number, and a mask black/white image. i have an iot device thats takes centered pictures of the water meter centered / always in a similar place, how could i center the data in the dataset?

pallid badge
#

Hi, are there also Discord channels for scientific computing around or Slack communities? How does one find such things?

pallid badge
#

I import data, clean it, process it, and save it somewhere?

serene scaffold
#

@dreamy topaz "object oriented programming" means different things to different people. If you want to do machine learning in python, you should know how to make a class in python, and know how to use classes that other people have created.

nova matrix
#

I have a dataframe column with one trend where the null values lay between 2 known rows
and then sudden blocks of NaN values (50 rows)
i want to fil only the first trend of Nans
and not the second
how would I achieve this

serene scaffold
#

@nova matrix what method do you want to use to fill nans in the first trend

nova matrix
#

mean of the value before and value after

#

I tried different methods but my dataset has 50 million rows so I was wondering if I could do this with a package

serene scaffold
#

@nova matrix you can use isna and shift
I'm standing on a train or I'd show you

#

@nova matrix these are the key

scenic parcel
#

What's a nice solvable problem for LSTMs

nova matrix
serene scaffold
nova matrix
orchid forge
eager plume
#

I'm new to python world.
I wanna ask about data science's prerequisites, roadmap, curriculum.

Is the ML,DL and DA,Stats combines together to become Data Science?????

unkempt apex
orchid forge
#

I'm so fucking happy, finally I'm understanding how to do exploratory data properly, I'm literally getting it now.

#

The world ways of understanding is so fake and stupid, I came up with my way to understand which is crazy

#

My solution to solve a freaking problem is "don't ask people for help"

serene scaffold
wooden sail
orchid forge
#

THAT YOU DON'T QUESTION THE FREAKING DATA AT THE BEGINNING ITSELF

serene scaffold
#

What? Why are you shouting?

orchid forge
#

People keep saying this thing that first you understand your problem statement bla bla bla bullshit

orchid forge
serene scaffold
orchid forge
#

The first step is actually to accept the fact that even a data scientist/analyst can't help you because he/she would only tell you what everyone would tell you

wooden sail
#

you might consider that if everyone tells you something, including experts in the field, maybe you're wrong

orchid forge
#

I'm not questioning any different solution coming from a expert, it's just that now I have found "MY OWN" way to solve shit

#

That's all which is 100000% better and it's working for me

wooden sail
#

and are you well equipped to show and explain why the method works?

past meteor
#

Otherwise you're looking for a needle in a haystack. Look at it this way, how much harder would it be to do data analysis if I took a dataset with a meaning and renamed all the columns into A, B, C, D etc.

#

I've seen domain experts make assumptions and hypotheses that were incorrect but that in and of itself is very important. It lets you find out other interesting things like maybe your data quality is bad, the measurements are incorrect, you made a coding error and if you can exclude all of this their assumptions may have been wrong, which lead to additional questions

orchid forge
past meteor
#

Wait

past meteor
orchid forge
#

Oh okay

#

I have a bad habit to not trust myself when it comes to having a solution in my head and then I would just simply ask people if I'm correct or not which has kinda made me feel like I don't understand data analysis, but idk today morning I woke up feeling something else like I was just studying and suddenly things start to make sense because I was legit studying. I tried to have lil confidence in my way of understanding a data and it kinda worked for the first time and I just wanna be this person everyday now.

#

Somebody who trusts herself

past meteor
#

I have a few tips about this later but before we go there, can you still try and explain with your own words what you think I meant with domain expert?

orchid forge
#

Someone who is dope in his/her work field (in their industry work)

#

I guess

#

Why?

past meteor
#

Okay we're on the same wavelength then

orchid forge
#

Oh okay idk the professional words but I do understand the professional words okay

#

Zester

past meteor
#

So, I've done data projects (only talking about work ones) in additive manufacturing, health, sociology, finance, ... I'm absolutely not an expert in any of those. The only way I could get those done is by asking experts where to look when I get my dataset and to validate my findings

#

It's not about not trusting yourself

#

Often times data professionals are an expert in working with data and they throw you in the deep end in places you have no clue about

orchid forge
#

Ik but I wanna have my own journey with data.

past meteor
#

But you want to do this for a living?

orchid forge
#

Yeah for a while not forever

past meteor
#

Honestly, the more involved datasets I've worked with were not ones where you could "have your own journey with data". Simply because it would be like having all of your columns be named A, B, C, D, ...

orchid forge
#

😂

past meteor
#

Then you need someone at your side. Because the subject matter is just the way it is

#
  • 9 times out of 10 I'd not know where to look and that's absolutely normal. All the things I'd find were obvious and that makes sense.
orchid forge
#

Ofc but I love to play a role of some "smart" person who is finding things which already exists but i just love that "wow i just discover something new" feeling

past meteor
#

For my current project I do find stuff that matters but when we meet the doctor we work with he always says "but could you look at X, Y and Z in conjunction with ..."

orchid forge
past meteor
#

And that's where the gold is

#

Well, we're not domain experts in this server

#

I'm talking specifically about getting assistance from someone working in the field

#

And how that shouldn't make you feel like you're not doing a good job

orchid forge
#

I'm just a girl who wants to be happy thinking she's smart even if that's like basic for you all expert people

orchid forge
past meteor
#

You're misunderstanding what I mean

orchid forge
#

I just have this server and Google

past meteor
#

My point is simple

#

No matter how smart you are, if you get an average dataset you'd find in real life you'd struggle because you're missing a lot of context. This applies to me, you and anyone else

orchid forge
orchid forge
past meteor
#

Then you go for different sources, news articles, books, papers, blogs, ...

orchid forge
past meteor
#

I always try and immerse myself (even if there's people I can ask questions to)

orchid forge
#

Oh that's nice

iron basalt
past meteor
#

So yeah, to circle back. I'd say for an EDA the critical thing is to have background knowledge of the problem you're trying to solve. When you do this for a living there will always be someone you can ask

orchid forge
#

I love this server

past meteor
#

For now all you can do is probably just read info about the topic online etc

orchid forge
orchid forge
past meteor
#

If you want to practice what worked for me is doing Kaggle competitions. Specifically those called "tabular playground". They're ML competitions but people always do an EDA, you can stop there if you want. The gist is, you make your own notebook, your own solution and then you read others

#

I'd always make it 100 % before looking at others

#

It's a mix of learning technical stuff "wow, is that how to make that kind of plot with Matplotlib?" and data analysis/ML intuitions "Oh that's how they looks at stuff, interesting conclusions they drew from ..."

iron basalt
# past meteor Honestly, the more involved datasets I've worked with were not ones where you co...

Reading this Wikipedia article simulates the experience of seeing all those new strange column names for a specific field of work: https://en.wikipedia.org/wiki/Glossary_of_baseball_terms

Glossary of baseball terms

This is an alphabetical list of selected unofficial and specialized terms, phrases, and other jargon used in baseball, along with their definitions, including illustrative examples for many entries.

serene grail
iron basalt
#

What you don't know what a "cement mixer" is?

orchid forge
past meteor
iron basalt
#

Most extreme in things like business and law.

#

(And math, what you don't know the "hairy ball theorem?")

past meteor
#

I like this one because even the explanation has references to things you wouldn't know. Torii Hunter, grand slam, walk-off homer, ...

#

(especially for us Europeans lol)

iron basalt
# past meteor

Yeah, like trying to understand a legal document in another language.

past meteor
#

When I did the job in (additive) manufacturing I spent ym first week on the shop floor with the technicians and reading ISO standards in my spare time

iron basalt
past meteor
#

Also like just seeing the production steps

#

if I saw "wire EDM, 2 seconds"

#

You have to see the production process to gauge the plausibility of that

#

I suppose you could also make a histogram/boxplot of all lengths and come to the same conclusion

#

This was some part time gig I did as a student. By far the worst data job I've ever done.

iron basalt
past meteor
#

I had others that were nice though

#

This one just had a terrible company culture. To do anything I needed authorization from Denver (-7h) and India (+7h) needed to execute iti

#

The "database" was Excel files on a network drive with a SQL view on top, queries could kill the entire plant

iron basalt
past meteor
#

Denver tells me I have a go end of day on the first day I ask or if they're busy on the next. By the time I can ask India it's their midnight

iron basalt
past meteor
#

Ah and the other thing that I didn't like was people couldn't agree on basic terminology. For instance, what "how much % of parts were without defects this month" meant

#

For one group it meant parts manufactured this month / parts returned this month while for the other group (... me) it's parts manufactured in month A / parts returned of month A (irrespective of when they were returned)

#

So instead of cool stuff you can imagine I spent a lot of time in meetings getting people to agree (and write down) definitions

iron basalt
orchid forge
#

Hmm

iron basalt
# orchid forge Hmm

(Not implying that the CIA is involved, just that it's a pretty bad thing for an organization to do if even the CIA recommends it as a method to disrupt)

orchid forge
#

I don't understand what are you trying to say

serene grail
#

I actually often argue about basic terminology...
Self-reflection time...

orchid forge
#

But I wanna be a CIA man

#

Like it's so cool

#

Haha @small wedge

#

Funny right

small wedge
#

bro wants to be a fed whyme

orchid forge
small wedge
#

you do my friend and so much more

#

but i digress I don't wanna fill ds/ai channel with nonsense

orchid forge
lapis sequoia
#

noticed that lscpu in linux will tell you whether your cpu has avx,avx2 etc support (if i understand correctly)

buoyant vine
#

And under flags it will tell you, the supported CPU flags

#

On most x86 hardware, you will have AVX2 and the generations that came before it, i.e. SSE4, AVX, SSE3, etc...

lapis sequoia
#

thanks, yes

buoyant vine
#

Only very new consumer hardware like AMD's Zen4 CPUs have AVX512 features.
Some Xeons Golds had AVX512 several years ago, but some problems they can have is it causes the chip to overhead and thermal throttle

lapis sequoia
#

interesting, i've got those on intel, not a new laptop

buoyant vine
#

AVX512?

lapis sequoia
#

i was noticing that colab has xeon processors

buoyant vine
#

Most ARM server chips will ship with NEON, notably Ampere, atlthough they don't mark it in their CPU info on VMs.
Some newer gen ARM chips like AWS Gravaton have SVE1 & 2 as well which provide more powerful SIMD operations on the cores

lapis sequoia
#

avx51f avx512dq avx512ifma avx512cd sha_ni avx512bw avx512vl avx512vbmi umip avx512_vbmi2 avx512_vnni avx512_bitalg avx512_vpopcntdq rdpid movdiri movdir64b fsrm avx512_vp2in...

buoyant vine
#

Fair enough, what chip is it?

lapis sequoia
#

i7

#

bought it second hand, 2 years ago

buoyant vine
#

I meant what generation

#

I'm guessing either Icelake or Skylake

lapis sequoia
#

11th

#

idk the codenames

buoyant vine
#

Intel just with the worst possible naming schemes on their chips 😔 And then AMD is now doing the same

lapis sequoia
#

just use the number

#

:-)

buoyant vine
#

Right it's Rocket lake, so post-icelake

lapis sequoia
#

post ice age

#

have u tried torch/rocm in amd gpus?

buoyant vine
#

not via torch

lapis sequoia
#

via tf?

buoyant vine
#

no just onnxruntime

#

we dont use TF, which is think is a general pattern in the industry

lapis sequoia
#

i see, for inference or training as well?

buoyant vine
#

TF is just not a very fun framework to try use in comerical aspects and then deploying

#

Mostly inference

lapis sequoia
#

interesting, i've used tf for training, onnxruntime for web deployments

buoyant vine
#

Can do training, but normally it is easier to use our PT framework so we just use nvidia

lapis sequoia
#

i looked at it today and ratio of stars to issue was high

#

sorry, low

buoyant vine
#

What on earth is ratio of starts to issues supposed to suggest lol

lapis sequoia
#

well, i'd expect that there aren't that many users, but there are many issues

#

if it works 4u, that's fine though...

buoyant vine
#

🤨

lapis sequoia
#

could also mean programmers aren't very concerned about fixing issues. it's nor far from tf's n of issues

buoyant vine
#

that is... 😅 Definitely not how you want to judge that lol

lapis sequoia
#

to me the response to user queries is quite a good proxy for success of a library

wooden sail
#

tf has both way more stars and fewer issues than numpy, and pytorch has way more issues and fewer stars than tf

lapis sequoia
#

imho it's a good library though

buoyant vine
#

They are both massive libraries

#

Commercially I would generally lean towards pytorch over TF though

#

Theres quite a few beginner tutorials for TF

#

but generally in our experience PyTorch's frameworks are just more intuitive and normally faster on things like train time and memory

lapis sequoia
#

on Keras metrics Pytorch is the slowest

#

but could be a limitation of keras, idk

buoyant vine
#

Keras metrics... With pytorch ?

lapis sequoia
#

i meant as a quick glance of a repository, low n of stars but high of issues does not seem great, was a minor observation..

#

Yes, why?

buoyant vine
#

why on earth would you use that xD

wooden sail
#

keras did split back out of tf

lapis sequoia
#

I'll let you guess..

buoyant vine
#

Keras is effectively built for TF

lapis sequoia
#

no

buoyant vine
#

It is

wooden sail
#

keras started out and is now again separate from tf

lapis sequoia
#

no, just read about it

buoyant vine
#

In reality, it is

lapis sequoia
#

Keras 3 is multibackend, and runs well with Jax and Torch as well

buoyant vine
#

Like, I hate to break it to you, but using Keras and PyTorch is just a weird choice

lapis sequoia
#

In fact, is in the road to support MLX

#

it's not because you have the flexibility of testing the speed of many models

#

in != frameworks

wooden sail
#

(also for completeness, jax and tf use the same backend)

buoyant vine
#

If you're using torch, it is infinitely easier to use something like PT lightning & friends than Keras

lapis sequoia
lapis sequoia
wooden sail
buoyant vine
#

yeah

#

I think it is a bit odd keras chose to abstract themselves out of just TF

#

ultimiately it is probably going to still be solely used with TF 😅 Because that is what all the tutorials will use

wooden sail
#

probably for a while, yeah. it did start out separately from tf though so maybe they can pull it off ok

lapis sequoia
#

it's actually nice conceptually

wooden sail
#

i don't remember when keras was acquired

lapis sequoia
#

it's an abstract symbolic layer

#

that specifies computations

wooden sail
#

looks like google acquired keras in 2017

lapis sequoia
#

keras has tutorials, they run on any backend

buoyant vine
#

and does some things like Multi-GPU and multi-machine training without a lot of the footguns

wooden sail
#

i have to admit i've never heard of pytorch lightning before

buoyant vine
#

I'd describe it as the thing you end up finding when you start training large scale models or multi-gpu models

#

because it manages all the checkpoints and device mounting for you

wooden sail
#

same here, i've yet to work with more than one a100

#

usually in the team "if it doesn't fit there, go back to your notebook and do better math"

lapis sequoia
#

you can test that in tf

#

create multiple logical gpus from 1 gpu

buoyant vine
#

I generally like it for that reason

#

but more so in the context of when working with teams

#

it just helps prevent people writing some cursed loops or unreadable blocks

lapis sequoia
buoyant vine
#

If you're in a team where everyone is good at Python and best practices, then it matters less imo, other than the multi-machine stuff

wooden sail
lapis sequoia
#

i admit i think pytorch will win to some extent

#

i just like fchollet general thought process i think

buoyant vine
#

From the work I have done, I think it has already largely won the comerical space

wooden sail
#

they work very differently, it actually makes sense to try out different backends for some problems

lapis sequoia
#

he is so smart lol

wooden sail
#

the way jax/xla and pytorch build the computational graph is very different, and this affects the speed and memory usage of some operations

buoyant vine
#

I think a lot of people start with TF and Keras doing tutorials and learning

lapis sequoia
#

maybe, i did the opposite, started with pytorch

buoyant vine
#

But seems to be the defactor for comercial projects where we go "okay we need a classifier or something" it is pytorch time

#

I am guessing because PyTorch seems to be the choice for academic related things

wooden sail
#

that is the case, though i do see (and push for) more jax lately

buoyant vine
#

Flash attention comes to mind with the recent LLM stuff

lapis sequoia
#

are you guys trying to help the world with your models

buoyant vine
#

no

wooden sail
#

no

buoyant vine
#

we are trying to make money

#

and in my case, classify half the internet

wooden sail
#

what i work on is either to publish papers or to make money, or both

jaunty helm
#

(guess I'll ask here as well)
sry to interrupt, but
If you were getting people, who are completely new to programming, into python (mainly for statistical analysis), would you recommend installing from python.org, or use something like anaconda, or even just colab?
personally I've not used anaconda much at all so idrk what's good/bad about it

lapis sequoia
#

do you recycle

buoyant vine
#

Personally I am not a huge jax fan, but I can understand why it exists

wooden sail
#

the biggest seller for me is really just that it looks like numpy, and most people already know that as background

buoyant vine
wooden sail
#

well, i say that, but the next thing right up there is the ease of writing the code to look just like the math on paper does

lapis sequoia
#

oh i think i've seen you in the math forums now

wooden sail
#

i would say conda is still my preferred way of using python, it brings most of the stuff i need bundled together. also requires no sudo perms so it's easier to use it on compute clusters with restricted perms

lapis sequoia
#

i think you explained a bit about the condition number to me

#

could be it wasnt you

wooden sail
#

that sounds like something i would mention, but i can't say i remember

lapis sequoia
#

i had no idea about it, it was quite useful

#

(it was years ago XD)

wooden sail
#

then there's no way i remember

buoyant vine
lapis sequoia
#

ig you use more Julia than Python for ML? Or maybe it's not the right question

wooden sail
#

i've been meaning to learn julia, but don't know it. i mainly use python. used to use matlab before

serene grail
#

I've only heard bad things about matlab

lapis sequoia
#

i see. it seems quite a beautiful language, at least from examples

wooden sail
lapis sequoia
buoyant vine
#

I think Julia is cool, but it suffers a bit from the issue of being very niche.

At work it is always a weigh up of:

  • Does it need good performance?
    • No -> Use Python
    • Yes -> Use Rust

It is hard to justify Julia just for the data and ML side of things and getting everyone in the team to learn another language

wooden sail
#

it jits everything by default and has faster run times. it's also industry standard in many applications. sure burns a hole through your pocket though

serene grail
buoyant vine
#

Julia as a language is great, and makes awesome use of LLVMjit

#

but I think it is very niche

#

that being said, wish we used it instead of Python at one of my old jobs

lapis sequoia
#

i've no knowledge about that, but i think it's precise numerically (for linear algebra operations)

buoyant vine
#

would have been amazing there

#

The hard thing is always getting devs to learn it though

#

unless all you do is numeric computing or ~adjacent stuff

lapis sequoia
#

yeah, i guess we are on a tangent on a tangent

wooden sail
#

but wrapping it as part of a bigger product, not so much

lapis sequoia
#

do you have favourite ml researcher?

buoyant vine
#

True, but with ML specifically I think it can also be hard to weight up VS PyTorch/TF/Onnx

lapis sequoia
#

(can be a list)

buoyant vine
#

there are a lot of devs that know python, and a lot that know Python and an ML framework

lapis sequoia
#

would be interesting to see node js hop in

buoyant vine
#

Also RIP my data engine, 32 cores and 64GB of memory and it still OOMs when aggregating 😔

image.png
lapis sequoia
#

do you guys open problems to freelancers anywhere?

lapis sequoia
buoyant vine
#

No I mean CPU

lapis sequoia
#

you do onnx inference in cpu?

buoyant vine
#

This isn't the inference server

#

this is the system that just logs the data being spat out by the inference cluster

#

the inference system itself is on a GPU machine(s)

lapis sequoia
#

i see

buoyant vine
#

Although ngl, I am kinda excited for the AI CPU chips

#

they have enough FLOPS processing power to make it realistically pretty cost affective to run on them instead of GPUs

agile cobalt
buoyant vine
serene grail
#

Oooh, working on a search engine must have been interesting

buoyant vine
#

Was good, I got a bit burnt out though

#

Learnt a relatively important life lesson that on those sorts of projects and systems, it really makes a difference to actively use the tool so you understand why things are done in certain ways and can see it from the end user's POV

#

otherwise it can seem a bit like a lot of stuff not quite lining up

lapis sequoia
#

basically compiles computation graph to machine code and has fusion operations which i hear are useful...

#

but idk what it means

wooden sail
#

the best part about jax, aside from looking just like numpy (and therefore looking very similar to matlab and julia as well), is that it exposes XLA's features directly

#

you can choose to grad, jit, and map however you like

lapis sequoia
#

As a part of the OpenXLA project, XLA is built collaboratively by industry-leading ML hardware and software companies, including Alibaba, Amazon Web Services, AMD, Apple, Arm, Google, Intel, Meta, and NVIDIA.

#

's got some backers

past meteor
#

I started out with TF, it was what we used in uni for neural nets (that and matlab)

lapis sequoia
#

i'm unsure that tensorflow uses XLA compilation by default, where is that information?

buoyant vine
#

Also 👏 AMD please give me a desktop chip of this

image.png
#

Otherwise I am going to be building a really cursed laptop server

lapis sequoia
#

yeah, so it's not the default since that's false by default

wooden sail
#

you probably have more tops on a usual server processor

#

ah, tf doesn't have jit enabled by default?

#

you can always try to cause a weird error and check the traceback

buoyant vine
wooden sail
#

jax spits out hundres of lines of xla calls whenever you look at it wrong

buoyant vine
#

If their measurements are in anyway close to what they say they are

#

that chip is several times faster than what you could do with a 64 core Epyc

wooden sail
#

i somehow strongly doubt that from a 28w processor

lapis sequoia
#

im checking

buoyant vine
#

In reality is a 54W chip, which I suspect is what they are actually doing to get those numbers

#

they are doing Max boost frequency to get those tops

wooden sail
#

the cache sizes would also have to be compared because that neuters the real life performance

#

time to fish up some benchmarks

buoyant vine
#

For reference a 16 AMD ryzen on zen 4 will do about 1 TOPS from my testing

#

although I think Openblas tends to struggle past 800 GFLOPs because of memory bandwidth, also unsure if OMP is pinning the cores

lapis sequoia
#

tf.config.optimizer.get_jit()
returns ''

wooden sail
#

some old numbers from intel xeons from 2022 say 419 tflops with float32, and upwards of 1500 with int8

buoyant vine
#

That sounds far too high

wooden sail
#

the h100 lists 26 to 3000 tflops depending on the data type

buoyant vine
#

a A10g GPU is rated at 250 TOPS

wooden sail
#

for which datatype?

buoyant vine
#

int8

wooden sail
#

i need to look up the a10g cuz i'Ve never worked with it

buoyant vine
#

It is a nice GPU tbh, it is what the AWS G5 instances use

#

I am using it here just because it was the first one I know that has the specs clearly listed on nvidia's page

#

Sometimes getting spec sheets with numbers is a pain 😔

#

I suspect AMD's numbers there are int8 operations

#

complete guess, but any sort of floating point op seems far too high

wooden sail
#

yeah i would also think so

#

let's see here

#

if techpowerup is to be trusted, an a10g

image.png
#

and an a100

image.png
buoyant vine
wooden sail
#

yeah

#

this is the a100 from nvidia's site

image.png
buoyant vine
#

that probably about lines up with what i'd expect

wooden sail
#

and the h100 with 3000 tflops on int8 seems reasonable then

buoyant vine
#

Maybe that 50 TOPS number isn't the most unrealistic thing in the world then

#

1/5th the compute of the A10g, but 1/3rd the TDP

wooden sail
#

indeed, if they disclose the data type 😛 otherwise this will become like another "nanometer" thing

buoyant vine
#

Yeah, well once they release I'm sure we'll see some more numbers

#

otherwise 😅 I am doing some shopping and we'll do a comparison

wooden sail
#

TOPS with 1bit ints

buoyant vine
#

Man that would really suck to find out xD

#

Although I am sure it could be utalised for some interesting bits of compute

#

like data filtering

#

We can actually probably work out

#

since I suspect, AMD's TOPS will be MS' TOPS

#

because of the whole AI PC shit

wooden sail
#

mhm

buoyant vine
#

Least scuffed MS webpage

image.png
lapis sequoia
buoyant vine
#

PepeHands Litterally everyone is just saying TOPS but not the datatype

#

god damn it MS!

wooden sail
#

what i did find in an arstechnica link is that current amd 7000 and 8000 chips offer "12 to 16 TOPS"

buoyant vine
#

that is what the AMD website lists

#

I think that is using the internal GPU

#

because they dont have a NPU

#

but do all have the Radeon graphics

#

so might be that what they are doing via their Ryzen AI™️ drivers, is effectively proxying the operations as graphics ops

wooden sail
#

mhm

#

this whole thing is a shitshow

buoyant vine
#

Idk why they are so focused on them being in laptops tho

#

Like... Guys do you not see if you're getting 50TOPS of int8 performance how big that potential is for the server and desktop space?

#

just for inference and numeric computing

#

Ignore the whole MS AI PC stuff that no one actually gives a shit about

wooden sail
#

they're already selling it that way on the server lines for 2 or 3 years though

buoyant vine
#

Qualcom are making ARM based laptops with excellent battery lives, but don't put that as their biggest feature

#

Everyone trying to push for AI but 99% of applications dont support it

buoyant vine
wooden sail
#

pretty sure yes, the xeon i mentioned from 2022 explicitly says neural something or another

buoyant vine
#

Like the best we have gotten really is better SIMD support and larger caches which mean more per-core performance

wooden sail
#

efficiency this, performance that, AI everywhere

buoyant vine
#

They dont actually have any dedicated hardware IIRC

lapis sequoia
#

but they do have a page detailing why and how at least.

buoyant vine
#

what Intel have done is taken their perf cores and efficiency cores from the desktop market, and put it on Xeons

#

with a bigger cache

#

Going off their 'fact sheet' it seems it is just the E-cores and P-cores

wooden sail
#

i still have to find out what exactly that accelerator is

#

ah, intel AMX they call it

#

dedicated matmul hardware on the P cores

#

for 3 gens already

lapis sequoia
#

they've got some new stuff for gpus also

wooden sail
#

the trend i've seen in the last 10 years is that the server scene is like "upstream" of desktop and laptop hardware, just like this

#

and the neat features trickle down... with arguable salesmanship like shoving AI down your throat

buoyant vine
#

thonk_line Reading their performance sheet tho

#

it just looks weird

#

wait nvm

#

In their graph they did (old)FP32 vs (new) BF16

#

and then in the next did (old)int8 vs (new)int8

lapis sequoia
worldly wagon
#

Is there any documentation on using arrow keys to move sliders in plotly?pithink I just wanna confirm this can only be done in dash not standard plotly GO

serene grail
#

I think a few years in the future "AI on chip" on desktop will actually be quite useful and the chip manufacturers don't want to fall behind the competition on that front

wooden sail
#

i think that's still a handful of generations away. the gap between AI tasks you can run on a desktop vs realistic tasks to be run on compute clusters is still comparable to the distance between heaven and hell

#

unless you put server grade hardware in your desktop, which requires you to shell out several tens of thousands of moneyz

buoyant vine
wooden sail
#

like 3 years i think

#

no, my bad, 2023?

#

i thought it was older

buoyant vine
#

All the AMX stuff and Xeon 6th gen stuff is 2024 from what i can see

wooden sail
#

amx was introduced in 2020, but only implemented in 2023 with xeon 4 (According to google)

#

but that means they were selling the AI on cpu idea for a while already

buoyant vine
#

I wonder if that is because of all the chip and arch issues intel were having previously

#

I remember they were having some serious issues with their new arch in the chip making processes

wooden sail
#

like self-destructing chips :x

buoyant vine
#

Mostly a desktop thing

#

But one thing is going to be interesting is looking at the price of the 6th gens

#

Going off their press release sheet, they make it seem like the 6th gens are the ones with actually good performance with AMX

#

the rest are just AVX512 VNNI

#

so what is the price of that chip

#

and how much does it weight up against the Epycs?

#

128 physical cores

#

fucking monster of a CPU

#

So the top line 6th gen Xeon is 128 P- cores and 288 E cores

#

so it is no slouch either

#

Will be interesting to see what the TDP and cost of that chip is

#

Competition is going to be 2x 9754S @ ~$11k a piece

left tartan
coral field
#

what are good statistics to measure the strength and direction of association between a continuous variable and multi-class categorical output?

proven inlet
#

Console:

Epoch 1/10 - Accuracy: 19.78%
Epoch 2/10 - Accuracy: 15.88%
#

it goes down 🙄

proven inlet
still parrot
#

huh

proven inlet
still parrot
#

idk I just started learning deep learning and all I had is errors Be glad you made it this far.

proven inlet
#

the code im using

still parrot
#

I know

proven inlet
#

but idk where error

proven inlet
still parrot
still parrot
proven inlet
proven inlet
still parrot
# proven inlet

by the way can't you do something like back feeding the info like telling what the answer should be and what it could be also?

proven inlet
still parrot
#

ohh

proven inlet
#

Which checks the result and compares with true one

#

if its right good 👍 if not change the weights and bias

#

By delta

#

derivative of sigmoid or ReLU

#

depends on what you are using

odd meteor
# coral field what are good statistics to measure the strength and direction of association be...

You can compute the mutual information score. Unlike computing the correlation coefficient, I think one of the drawbacks of mutual information score is in its result's explanability.

For a more detailed analysis, you'll need a statistical test. The kind of statistical test you need depends on the kind of categorical feature you're dealing with.

Do you have have an ordinal or nominal categorical feature?

Since you've already mentioned that the categorical feature is not dichotomous, we can rule out Point Biseral Test.

still parrot
#

I think you should just test it more with more data.

proven inlet
still parrot
#

ohh my

proven inlet
#

and it's just basically number recognition

still parrot
#

ohh

proven inlet
#

all of them has 100 hand written numbers

coral field
still parrot
#

there's a 25 hour long video teaching it

proven inlet
proven inlet
still parrot
#

pytorch but there's a video by another creator about doing it be scratch.

proven inlet
still parrot
#

ok

#

https://www.youtube.com/watch?v=Wo5dMEP_BbI&list=PLQVvvaa0QuDcjD5BAw2DxE6OF2tius3V3 https://www.youtube.com/watch?v=w8yWXqWQYmU https://www.youtube.com/watch?v=cAkMcPfY_Ns

YouTube
sentdex
Neural Networks from Scratch - P.1 Intro and Neuron Code

Building neural networks from scratch in Python introduction.

Neural Networks from Scratch book: https://nnfs.io

Playlist for this series: https://www.youtube.com/playlist?list=PLQVvvaa0QuDcjD5BAw2DxE6OF2tius3V3

Python 3 basics: https://pythonprogramming.net/introduction-learn-python-3-tutorials/
Intermediate Python (w/ OOP): https://pythonpr...

▶ Play video
YouTube
Samson Zhang
Building a neural network FROM SCRATCH (no Tensorflow/Pytorch, just...

Kaggle notebook with all the code: https://www.kaggle.com/wwsalmon/simple-mnist-nn-from-scratch-numpy-no-tf-keras

Blog article with more/clearer math explanation: https://www.samsonzhang.com/2020/11/24/understanding-the-math-behind-neural-networks-by-building-one-from-scratch-no-tf-keras-just-numpy.html

▶ Play video
YouTube
Green Code
I Built a Neural Network from Scratch

Don't click this: https://tinyurl.com/bde5k7d5

💚 Link to Code: https://www.patreon.com/greencode

How I Learned This: https://nnfs.io/ (by the awesome @sentdex )

I'm not an AI expert by any means, I probably have made some mistakes. So I apologise in advance :)

Also, I only used PyTorch to test the forward pass. Apart from that, everything el...

▶ Play video
#

https://www.youtube.com/watch?v=OGxgnH8y2NM&list=PLQVvvaa0QuDfKTOs3Keq_kaG2P55YRn5v

YouTube
sentdex
Practical Machine Learning Tutorial with Python Intro p.1

The objective of this course is to give you a holistic understanding of machine learning, covering theory, application, and inner workings of supervised, unsupervised, and deep learning algorithms.

In this series, we'll be covering linear regression, K Nearest Neighbors, Support Vector Machines (SVM), flat clustering, hierarchical clustering, a...

▶ Play video
proven inlet
#

Oh i watched the last 2 videos already but thanks ill watch 2 other 🙂

proven inlet
#

Thanks again 🙂

still parrot
#

you can also get a copy of his code

odd meteor
# proven inlet could you help me please? 🙂

In that case you need a non-parametric test.

You can use Kendal Tau or Spearman-Rank correlation instead of Pearson correlation.

Another alternative is, to carry out a Chi-Square test of independence to determine the association between the two variables.

You can even use ANOVA as well to test for differences in means of the continuous variable across the different categories.

still parrot
proven inlet
proven inlet
odd meteor
proven inlet
odd meteor
# coral field ordinal, the categories represent various "levels"

In that case you need a non-parametric test.

You can use Kendal Tau or Spearman-Rank correlation instead of Pearson correlation.

Another alternative is, to carry out a Chi-Square test of independence to determine the association between the two variables.

You can even use ANOVA as well to test for differences in means of the continuous variable across the different categories.

summer bolt
#

can anyone help me create a countdown bot that stays in a vc 24/7 and countdowns like this: 3 minutes till countdown
2 minutes til countdown
1 minute til countdown
30 seconds till countsown
10
9
8
7
6
5
4
3
2
1
Go
Please help me with this

odd meteor
proven inlet
summer bolt
still parrot
# summer bolt can anyone help me create a countdown bot that stays in a vc 24/7 and countdowns...

import time

def countdown(t):
while t:
mins, secs = divmod(t, 60)
timer = '{:02d}:{:02d}'.format(mins, secs)
print(timer, "minutes till countdown")
time.sleep(60)
t -= 1

print("1 minute till countdown")
time.sleep(30)
print("30 seconds till countdown")
time.sleep(20)
print("10 seconds till countdown")

for i in range(10, 0, -1):
    print(i)
    time.sleep(1)

print("Go!")

Start countdown for 3 minutes

countdown(3)

summer bolt
proven inlet
summer bolt
#

wrong channel xd

proven inlet
#

İts ok lol

odd meteor
# proven inlet This

@proven inlet I'm currently not free at the moment but I will take a look at it once I'm a bit free.

proven inlet
lapis sequoia
#

i ran it with 1 hidden layer

#

Loading Dataset...
Loaded.
Training Epoch: 0
Training Epoch 0 finished with %18.366666666666667 accuracy
Training Epoch: 1
Training Epoch 1 finished with %31.266666666666666 accuracy
Training Epoch: 2
Training Epoch 2 finished with %34.766666666666666 accuracy
Training Epoch: 3
Training Epoch 3 finished with %37.4 accuracy
Training Epoch: 4

#

seems fine, what's the issue @proven inlet

proven inlet
lapis sequoia
#

i don't remember if the remaining steps of backprop are correct, but if you are using MSE the derivative is correct

#

i'm showing you it does not with 1 hidden layer at least

proven inlet
#

is the problem using more than 1 hidden layer

proven inlet
#

most of them are 0

lapis sequoia
#

i commented some stuff

#

and changed the delta, and learning rate, i can't test much more now, just run it and check

proven inlet
proven inlet
lapis sequoia
#

does seem to increase with extra hidden layers as well.

proven inlet
#

How??

#

it does not for me

#

lemme try on colab

lapis sequoia
#

i said i modified a few things

#

in any case, im unsure if it'd get very far, since the surface can have many pockets

proven inlet
#

Uh you are using ai libraries

lapis sequoia
#

i ran a similar model with keras to compare

proven inlet
#

i need to get it working without ai libraries

#

thats my goal

lapis sequoia
#

the ai library is completely irrelevant, you can remove the cell.

proven inlet
#

no its not irrelevant you load x & y data with ai libraries

lapis sequoia
#

how do you want me to test the code otherwise?

#

that data is the mnist dataset, a dataset of handwritten digits.

proven inlet
#

i have it as local file in my project dir

#

but it doesn't work

lapis sequoia
#

well, maybe you didn't normalise the data

#

i can't guess the part in your laptop

proven inlet
#

it's just like that

image.png
lapis sequoia
#

(x_train.astype("float32") / 255)

#

i.e normalising

proven inlet
#

img_array = np.array(image).flatten() / 255

#

i do this

lapis sequoia
#

idk what is the issue with your data, i can't access it, but the network seems not totally wrong

proven inlet
#

The data has no issue

#

it's all correct

lapis sequoia
#

ok, maybe someone else can help

proven inlet
#

okay thanks

#

:)

#

@lapis sequoia This is what happens 🙏

image.png
lapis sequoia
#

try it, please, this settings:

# 
input_size = 28*28
hidden_layers = [10,10]
output_size = 10
print("Loading Dataset...")
images, labels = load_dataset()
labels = onehot(labels, output_size)

print("Loaded.")
nn = Brain(input_size, hidden_layers, output_size)
nn.train(x_test, y_test, 10, .2)
proven inlet
lapis sequoia
#

what happens with hidden_layers=[] ?

proven inlet
#

they're being added to self.layers which is list of Layer objects

lapis sequoia
#

no, if you use it empty

proven inlet
#

it would not work

lapis sequoia
#

why not?

proven inlet
#

inputs will be connected to output without anything

#

it didnt work

image.png
lapis sequoia
#

then it must be the input data-labels pairs the problem

proven inlet
lapis sequoia
#

you can load it into a grid of images and paste the labels

#

with the image

#

just for a subset

proven inlet
#

i do that like this:

def load_dataset():
    images = []
    labels = []

    for label in range(10):
        path = os.path.join("Dataset", str(label))
        for filename in os.listdir(path):
            if filename.endswith(".png"):
                image_path = os.path.join(path, filename)

                image = Image.open(image_path).convert("L")
                img_array = np.array(image).flatten() / 255
                images.append(img_array)
                labels.append(label)

    return np.array(images), np.array(labels)
lapis sequoia
#

imho there isn't, just errors.reverse() is not used

#

to me it seems correct, but we won't know until you plot image-label pair

proven inlet
lapis sequoia
#

plot it and check

proven inlet
#

input is image, expected_output is label. atleast it should be,

proven inlet
#

seems fine

image.png
#

onehot format

lapis sequoia
#

no, it's not what i am saying

#

plot the images with the label, like the image of 0 with the label, then you know it's matching

proven inlet
lapis sequoia
#

and the image isn't wrong etc

proven inlet
#

its just in array format tho

#

between 1 and 0

#

wait lemme append their path instead of their pixels so we can check

lapis sequoia
#

like this

proven inlet
#

i changed load_data function to put img path instead of pixel array and it gave me these results:

image.png
#

it seems to work

#

because its 0

#

0*

lapis sequoia
#

also, if you have [] you still have 10 neurons with activations, and corresponding weights

#

so with empty list, it should do something

proven inlet
#

empty list to which param

lapis sequoia
#

hidden layers

proven inlet
#

hiddenlayers didn't work for empty layers

proven inlet
lapis sequoia
#

i've tried it without trouble, i'm just saying that it should have been a simple test to try first

proven inlet
#

i already tried it without hiddenlayers

#

it does not work

lapis sequoia
#

could it be possible that you are only loading one number?

#

i can't guess if you don't plot random data points.

proven inlet
#

i've tested that before running ai

lapis sequoia
#

can you show a screenshot of the images in 0\1

#

and 1\1

proven inlet
#

Alr

#

0/1.png

1.png
lapis sequoia
#

and how many are u using? in total

proven inlet
#

1000 images total

#

Alr

arctic wedgeBOT
#

:incoming_envelope: :ok_hand: applied timeout to @proven inlet until <t:1722376265:f> (10 minutes) (reason: attachments spam - sent 10 attachments).

The <@&831776746206265384> have been alerted for review.

lapis sequoia
#

just one screenshot of the images in miniature

#

should fit 20 in one image

atomic tide
#

!unmute 1018096765225938985

arctic wedgeBOT
#

:incoming_envelope: :ok_hand: pardoned infraction timeout for @proven inlet.

proven inlet
#

Thx

proven inlet
#

Dataset/0

image.png
lapis sequoia
#

it's not shuffled @proven inlet , maybe try that

#

you are passing first all 0s, then all 1s, ...

proven inlet
#

is that critical problem for ai ?

lapis sequoia
#

yeah, you are forcing the net to a local pocket probably

#

and it can't get out of it

proven inlet
#

HMMM

#

That makes sense

#

Lemme try that

lapis sequoia
#

mine are shuffled ;-)

proven inlet
#

good point

#

but the ai tries all of the numbers

#

per epoch

lapis sequoia
#

but you update the weights as you go don't you

proven inlet
#

Yep

lapis sequoia
#

so try, you can create a random array of 1000 numbers and pick from there

#

and then labels[i], image[i]

proven inlet
#

thats better i think

lapis sequoia
#

as long as you shuffle them in the same order yes

proven inlet
#

Yep

lapis sequoia
#

otherwise they don't match

proven inlet
#

Yeah ima try it rn

#

it's literally unlearning 😭

image.png
# 
permutation = np.random.permutation(len(images))
images = images[permutation]
labels = labels[permutation]
lapis sequoia
#

chollet likes amd apparently

Screenshot_2024-07-30_230147.png
lapis sequoia
proven inlet
#

lr was 0.002

#

changing it to 0.2

lapis sequoia
#

and hidden_layers=[]

proven inlet
#

that was absolutely random tho but %20 crazy

image.png
#

😳 😳 😳

image.png
lapis sequoia
#

yup

proven inlet
#

i think

#

it starts to work

lapis sequoia
#

been sayin

proven inlet
#

hiddenlayers were already empty

#

for a long time

lapis sequoia
#

alright, gotta sleep

proven inlet
#

learning rate changed the game here i guess

proven inlet
lapis sequoia
#

yes, shuffling as well

#

ur welcome

proven inlet
violet gull
#

rl batch size how much?

small wedge
violet gull
#

how know if need more?

small wedge
#

if your model struggles to converge

#

it's a hard variable to isolate especially in rl where there might not always be a clear metric for convergence

#

might be safer to start high like 64 and work your way down

violet gull
#

im at 512

small wedge
#

ideally it should be the minimum number of samples required to get a consistent estimation of the true gradient

#

if your samples are balanced well then that's probably super overkill, otherwise it's probably fine you just lose a bit of training speed from going so high

violet gull
#

thanks you

eager plume
#

Data science is vast field.

8 am confused in Ai/ML/Ds/Da.
That what to choose first.

serene scaffold
spare forum
#

DS means "using data technique to solve problem" kinda, it's large, it can be ml ai etc... But it is not necessary, DA is analytics the name speak for itself you do analytics to track business things (kpi), DL is a subset of ML which is a subset of AI

lapis sequoia
#

"Understanding Mathematics behind floating-point precisions"

#

wait...

#

results are right though...but..

serene grail
#

32 bits is 8x4 right

#

maybe a typo

lapis sequoia
#

yes..!article reads nicely though, ig it's just a typo

lapis sequoia
#

Recently Microsoft released a 1-bit LLM variant namely BitNet b1.58 which uses ternary {-1, 0, 1} for every single parameter. Surprisingly it matches the FP16 or BF16 precision transformer model.

serene grail
lapis sequoia
#

it seems just to be about the values that the weights can take

#

you can see that on the first matrix

#

(it may be a lot more complex)

#

paper ends with:

Recent work like Groq5 has demonstrated promising results and great potential for building specific hardware (e.g., LPUs) for LLMs. Going one step further, we envision and call for actions to design new hardware and system specifically optimized for 1-bit LLMs, given the new computation paradigm enabled in BitNet [...

#

so i guess that's the underlying idea / goal

serene grail
#

Interesting, so it's simplifying the values in the matrices for simpler calculations and that (in theory) gives the same results for less compute?
From my skimming that's what I got

lapis sequoia
#

exactly

#

this happens in quantisation (converting weights from FP to Int) as well, but up to a byte (int8 numbers.)

serene grail
#

very interesting

lapis sequoia
#

idk why it's called bit though, probably just being ignorant

serene grail
#

8 bits is one byte right?

lapis sequoia
#

yes

#

i've just seen this device called 'friend' has anyone seen it?

serene grail
#

I haven't

lapis sequoia
#

if they would be able to run a 1 bit llm in small devices, it'd be a big money thing

#

(that device must use wifi)

serene grail
#

hmm I don't know if AI is good enough yet for a device like that

lapis sequoia
#

maybe, i do enjoy talking to llms a lot. but never buy trendy tech

#

just mentioned cuz of the 1 bit llm

serene grail
#

oh yeah, if LLMs get much smaller that would be very good for this market

lapis sequoia
#

do you find dull talking to llms?

serene grail
#

not really, I'm just kind of too lazy to open up a tab in a browser and use an LLM lol
and I don't have a real use case for it besides having fun

lapis sequoia
#

i think you can 'hey google' with Gemini

#

but yeah, not that useful

serene grail
#

I like the idea of something like AI dungeon/novel AI more than a chatbot, they basically let you write a story together with an LLM. that can be more fun for me

lapis sequoia
#

nice, haven't heard ab it

serene grail
#

that's what got me interested in AI in the first place, although I didn't take it seriously until recently
AI dungeon used to use GPT2 in the beginning because that was the most advanced model

lapis sequoia
#

just realising log(3) is 1.58, the number in the llm

#

not sure what it means

#

basically having -1,1,0 reduces matrix muls to additions

serene grail
#

there are 3 options for the numbers, -1, 0, and 1, maybe it has something to do with that

lapis sequoia
#

yeah, the measure of the entropy/information of each bit i think, but still idk

#

so for the alphabet is log(26), more possibilities, more entropy (i mean, not for the alphabet, cause it's not random in english text.)

tranquil ledge
#

hello has anyone worked with TABLEAU DESKTOP before for data analysis ?

past meteor
cinder elk
#

hey, I need some help, does anyone know how can I count the number of screws if they're overlapping or touching each other in this image, I tried dilation but didn't work

IMG-20240731-WA0000.jpg
lapis sequoia
#

i'd guess SAM could be useful there

agile cobalt
#

maybe something like https://sites.google.com/view/f-vlm/home ?
might be overkill, not sure

F-VLM

Abstract
We present F-VLM, a simple open-vocabulary object detection method built uponFrozenVision andLanguageModels. F-VLM simplifies the current multi-stage training pipeline by eliminating the need for knowledge distillation or detection-tailored pretraining. Surprisingly, we observe that a

cinder elk
#

I've been asked to do this by classical image segmentation methods and not use any AI💀💀
that's why I'm scratching my head over this problem

serene grail
#

Do you need the exact number of screws or is there an acceptable error margin?

cinder elk
#

95% accuracy it says

whole stone
#

are the pics always at the same distance and using same sized screws?

unkempt apex
cinder elk
cinder elk
unkempt apex
#

then opencv will be best ( wait,.. is there any other options?)

spare forum
whole stone
#

i would try opening instead of dilation

spare forum
past meteor
#

If you're going to only use classical computer vision (no deep nets) you'll need an entire pipeline of steps

whole stone
#

also your class looks super cool, idk what you mean with it sucks

unkempt apex
#

if you change the original image and maybe apply some filter , with identifying contour for each pixel , it is possible

past meteor
#

The ones that are facing up could probably be found with a hough transform (circle detection)

#

The ones that are flat ... maybe with template matching

#

I'd try it iteratively, first use basic template matching to see how many "hits" you have

#

the nuts are super simple, that's something you can 100 % do with template matching and post processing

lapis sequoia
#

i know of people doing it with ROI algorithms

past meteor
lapis sequoia
#

it can not be that hard..(not saying it's easy either.)

agile cobalt
past meteor
#

yes

#

Or rather, there's extensions of the basic template matching that can handle rotation and scale

#

Like SIFT

cinder elk
past meteor
#

If you know the right words you'll find an example pretty quickly

cinder elk
#

wow, thanks @past meteor

serene grail
#

woah this is cool, I should look into cv

past meteor
cinder elk
#

I had only been trying to do this just by OTSU and contouring. Definitely learning new things here

past meteor
#

And we're at the right sentence

#

"How to find all objects in the scene with SIFT"

past meteor
#

okay, I really got nerdsniped by this one

#

Maybe just template matching but rotating your template and changing the scale can help

#

So basically, doing half of SIFTs algo

cinder elk
#

at that point I might use sift as well, I'll sit to code once again after my dinner let's see how this pans out

cinder elk
left tartan
past meteor
#

Or template + hough transform like I mentioned

#

Hard to say what'll work without trying it 😄

spare forum
violet gull
#

hi friends, I ran this for several hours https://github.com/pytorch/examples/blob/main/reinforcement_learning/reinforce.py and i am wondering why it does not look as expected. its made by pytorch so i would expect it to converge on such a simple model but it clearly hasnot converged. The red is the running average and the blue are individual episode scores

image.png
GitHub
examples/reinforcement_learning/reinforce.py at main · pytorch/exam...

A set of examples around pytorch in Vision, Text, Reinforcement Learning, etc. - pytorch/examples

agile cobalt
#

it sounds like it's been broken for a while?
that example hasn't been updated since 2022 too, and there is an issue claiming it does not works at all https://github.com/pytorch/examples/issues/1213

  • edit; could be that this guy is trying to run an older version of gym or maybe gym vs Gymnasium

just guessing though, I haven't tried running it myself though

violet gull
violet gull
agile cobalt
violet gull
#

its not convergent

agile cobalt
#

...........oh

lapis sequoia
#

do you guys know how to get to formula (1) ? from the previous one

Screenshot_2024-07-31_211004.png
#

it's from this paper https://arxiv.org/abs/1412.0233

arXiv.org
The Loss Surfaces of Multilayer Networks

We study the connection between the highly non-convex loss function of a simple model of the fully-connected feed-forward neural network and the Hamiltonian of the spherical spin-glass model under the assumptions of: i) variable independence, ii) redundancy in network parametrization, and iii) uniformity. These assumptions enable us to explain t...

wooden sail
#

it's explained in the text just underneath

#

you'll have to take out a pen and paper and try out their construction with 2 layers to see it

lapis sequoia
#

it has to be derivable algebraically i assume

#

i did test it, but that's not my doubt

wooden sail
#

what's the question?

lapis sequoia
#

how to get from formula on top, to the one in the bottom, algebraically

#

or if it's got a name or smth

#

for example, sigma isn't there

wooden sail
#

matrix multiplication can be defined as a double sum over the individual elements of the vectors and matrices involved

#

that's about it

lapis sequoia
#

that's not just mmul, you've got an activation

wooden sail
#

they replace the relus with a binary matrix A

lapis sequoia
#

that seems an approximation

wooden sail
#

why?

lapis sequoia
#

sigma doesn't just output 1 or 0

wooden sail
#

it does if you use a relu

lapis sequoia
#

oh, it

wooden sail
#

if you take an input x, a relu outputs either x or 0

lapis sequoia
#

it's not s sigmoid

wooden sail
#

the paper says everything it's using

lapis sequoia
#

i still don't think it's obvious the multiplicatory of weights

wooden sail
lapis sequoia
#

which is why i asked if there was some derivation

wooden sail
#

they explain it explicitly in words right under the equation

lapis sequoia
#

we are talking ab different things

wooden sail
#

everything you've asked is written there

lapis sequoia
#

you don't just build the formula reading the paragraph, do you?

wooden sail
#

yes you do

lapis sequoia
#

ok, sorry im not like you, so i asked to try to understand

wooden sail
#

using the paragraph and the definition of matrix multiplication, plus the info they gave above (using only dense layers and relu)

#

that's why i suggested you do it explicitly for 2 layers on paper

lapis sequoia
#

i already did that

wooden sail
#

did you really?

lapis sequoia
#

yes, why?

wooden sail
#

if you shows pics of your work on paper, others will be able to help more easily

#

since you'll have more concise questions about the parts that seem to be off

lapis sequoia
#

ill ask elsewhere, thanks, you must feel good now.

brave yew
#

guys... has anyone here worked with sentiment analysis pipeline of transformers library? for some reason my longer inputs are taking up to 10x less time than my one line inputs

#

i am using my gpu to process... atleast that is what i have specified in my pipeline method

unkempt wigeon
#

I I can use some help I'm trying to learn no networks unfortunately I need to hear somebody saying it for me to truly learn because I always lose my place with reading and sorry

violet gull
unkempt wigeon
#

I tried and I still don't understand it I'm sorry

thorn flame
#

Hi! I'm trying to build a recommendation system as an ML noob, is there some kind of good resource that could help?

#

I believe the process is the same for all types, just training data could be different

#

I'm given 1hr30mins to solve this. For an ML noob, is this feasible?

buoyant vine
#

Eh kind of

#

maybe not for a complete noob tho

#

A basic recommendation system, you want some vectors which represent user interests

#

and the easiest method is to mean pool those vectors of the users lists of interests

#

to get an average vector that hopefully encapsulates the context of all rolled up vectors

#

and then KNN search over some dataset of all the vectors to find similar things

thorn flame
#

I understand I need to convert the input datasets to pandas dataframes.

buoyant vine
#

Ehhhhhhh it kind of depends

thorn flame
#

I also understand I could use the surprise package for the algorithm and model evaluation (I could be wrong)

buoyant vine
#

easiest honestly is spit out a bunch of keywords for what ever content

#

and then feed it into some LLM encoder to generate the emebeddings that can be used for KNN

#

Have a look at the sentence transformers library for that

mental rampart
#

hey , i was having some doubt regarding tensor conversion from text in pytorch and nlp

#

if my tokens are words, how should i convert them to tensor?

thorn flame
thorn flame
#

Also by transformers lib, do you mean hugging face's?

thorn flame
thorn flame
thorn flame
#

I believe I could get the cosine similarity after getting the embeddings with sentence transformers

buoyant vine
#

Yes

#

I am about to go to sleep, but what you want

#

Is say a user wants some recommendations relating to star wars, you would encode that text/keywords into an embedding with sentence transformers

#

And then do knn search for close/similar results in your database

#

Where the keywords / records in your database have also been encoded with the same sentence transformer model

#

If you want recommendation by something like, "given a user's watch history recommend some other shows the user might like" then you take the average of all the embeddings the user has watched and use that averaged embedding to do the knn search

#

Or if you want "user watched X video, recommend them some other similar videos" you'd take the embedding of video X (be that generated from keywords or what not) and use that for the knn

#

Basically your whole goal is to get a index of various embeddings for the dataset of content you want to be able to select and return

#

Then it is just a case of generating a query embeddings to suite what sort of application you want

#

Make some sense?

thorn flame
#

It's basically an ecommerce platform

#

To provide products for users based on purchase history and whatnot

#

And browser activity

#

Doesn't knn search compare with preferences of other users??

buoyant vine
#

No?

#

It is completely arbitrary

#

All it does is calculate the distance between two points in a graph effectively

#

A vector is like a set of coordinates or a postcode

thorn flame
#

Which lib are you suggesting I use for that?

#

scikit-learn ??

buoyant vine
#

When you do KNN search, you effectively are asking "what are the closest other data points available to me from this point?"

thorn flame
#

Cos the KNN class from surprise lib works with some kind of Reader format

buoyant vine
#

I would suggest Sentence transformers to generate the embeddings, and PyNNDescent as the index

thorn flame
#

Index how?

buoyant vine
#

You give it a bunch of embeddings, generated by what ever content you have

#

And it will make a index that can be searched, i.e. you give it a query vector/embedding, it gives you the top K back

thorn flame
#

I dig

#

It basically implements its own knn algo I guess

buoyant vine
#

Quickly* rather than brute forcing checking every point

#

I wouldn't worry about the algo ATM.

#

It doesn't really matter to your use case

#

Just that it is faster than brute force

#

And provides a convenient way of going "hey here are all my embeddings, give me the closest points to X"

#

Start with something basic, i.e. don't worry about taking in the user history or what ever

thorn flame
#

so what I'm imagining is having to pass a user with necessary context to my function, and it returns a list of products based on my products dataset

buoyant vine
#

Just take some input text to begin with, encode it then do the knn search

#

Once you have that it should start to make more sense

#

And then you can start looking at doing it off of the user history

thorn flame
#

Do I need a db for this?

buoyant vine
#

No

#

You can do it in memory, doesn't really matter

thorn flame
#

Everything is stored in memory?

#

I see

buoyant vine
#

Unless your dataset is huge, but I suspect that is not really a issue here

thorn flame
#

I'm wondering if a possible result can return multiple similarities

#

Instead of just one

buoyant vine
#

That is what the K is

thorn flame
#

I see

#

Makes sense now :)

buoyant vine
#

so say with nndescent you can tell it "take the top 10" etc...

thorn flame
#

That's dope

#

Thanks. I'll run with this

#

Would be my first actual intro to ML if I'm successful with the engine heh

cloud cosmos
#

Hello

#

Im trying to make a text-speech ai and I don't know where to start exactly any tips?

brave yew
brave yew
cloud cosmos
quasi crag
#

Would be my furst actual im trying to make twxt spech to im succesfull with the engine 🏋️

cinder elk
raw tree
#

Hey - quick question - can you/what is the sanest way to load safetensors into keras ?

raw tree
#

wut O.O

brave yew
raw tree
#

If so, absolutlely no clue

brave yew
raw tree
stoic topaz
#

hi, im in high school and have a free choice data analysis project (using powerbi)

any interesting dataset/analysis ideas?

woeful sorrel
#

Hellllo friends

#

Man how to read a CSV file in pandas

cinder elk
#

can I use feature matching algorithms like sift/surf to find matches that are close enough but not perfect matches?

#

reference

image0.png
toxic mortar
#

Code:

!pip install -q transformers==4.31.0
from transformers import DistilBertTokenizerFast
from transformers import TFDistilBertForSequenceClassification
  sentiment_model = TFDistilBertForSequenceClassification.from_pretrained('distilbert-base-uncased',num_labels=2)

Error:

tokenizer_config.json:   0%|          | 0.00/48.0 [00:00<?, ?B/s]
vocab.txt:   0%|          | 0.00/232k [00:00<?, ?B/s]
tokenizer.json:   0%|          | 0.00/466k [00:00<?, ?B/s]
config.json:   0%|          | 0.00/483 [00:00<?, ?B/s]
model_training/kinesiologie_tape_new/sentiment/tf_model.h5 exists on GCP = False
Sentiment analysis stage (1/2) - est. time is 2 minutes
model.safetensors:   0%|          | 0.00/268M [00:00<?, ?B/s]
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-31-7a684e7281c2> in <cell line: 2>()
     30   ))
     31 
---> 32   sentiment_model = TFDistilBertForSequenceClassification.from_pretrained('distilbert-base-uncased',num_labels=2)
     33   optimizer = tf.keras.optimizers.Adam(learning_rate=5e-5, epsilon=1e-08)
     34   losss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)

2 frames
/usr/local/lib/python3.10/dist-packages/transformers/modeling_tf_utils.py in build(self, input_shape)
   1129     def build(self, input_shape=None):
   1130         call_context = get_call_context_function()
-> 1131         if self.built or call_context().in_call:
   1132             self.built = True
   1133         else:

TypeError: 'NoneType' object is not callable
#

Any help?

cerulean violet
#

Hello I am trying to make a small AI model which is for my bot used for moderation,any idea where to start?

cinder elk
#

what are your intended goals?
does it need to process images/videos or just for text?
if text only look into nltk for sentiment analysis and keywords.
Or you can use langchain

tranquil ledge
#

Hello does anyone worked with Tableau desktop before

#

i want to create this table