#data-science-and-ml

1 messages · Page 111 of 1

raw mortar
#

Rust has a dedicated ml framework, but I've yet to test it out

final kiln
#

has been a good experience

raw mortar
final kiln
#

nice

raw mortar
final kiln
#

yeah it's a fairly thin wrapper around torch

lapis sequoia
# final kiln yeah it's a fairly thin wrapper around torch

I managed to properly install pytorch with cuda into VS Code using pipenv pipenv install --index https://download.pytorch.org/whl/ "torch==2.2.1+cu121" "torchvision==0.17.1+cu121" "torchaudio==2.2.1+cu121". Was much harder than I expected.

P.S. python version "3.11.0", pipenv version "2023.12.1", GPU: NVIDIA GeForce RTX 3080Ti

final kiln
buoyant vine
#

Tbh @final kiln You might find the Burn optimization stuff quite interesting if you like the idea of the compiler benifits.

They are doing quite a bit of work atm around automatic kernel fusion and using Rust's RAII to optimize tensor memory patterns.

final kiln
#

.

sterile flare
#

hey anyone worked in R studio before ?

lapis sequoia
#

question: if my classification task involves an addition of a feature/column (e.g. confidence rating) in my training set vs my new dataset to predict on, will this throw my model off (inputs)?

lapis sequoia
desert oar
#

or are you saying that your training set has an extra feature that won't be present in the test set and future data? in that case you can't use it because you won't have it. it's not a matter of throwing the model off, it's a matter of not having the data to use

desert oar
dawn light
left tartan
#

I was unaware. Do you have something on topic to contribute?

#

!mute 1216171349786234963 1d chill, my friend.

arctic wedgeBOT
#

:incoming_envelope: :ok_hand: applied timeout to @lapis sequoia until <t:1710553523:f> (1 day).

broken arch
#

this server is ass im out

woeful sapphire
#

I’ve managed to reshape **__Table A __**into Table B, firstly with pd.melt and then pd.pivot. My follow up question is, how do I reorder the values in the Studies column from

[’adoption’, ‘eligible studies’, ‘studies not activated’, ‘total studies’,] to

[’adoption’, ‘eligible studies’, ‘total studies’, ‘studies not activated’]

I tried using

studiesDtype = pd.CategoricalDtype(['bls adoption','eligible studies','total studies','studies not activated'], ordered=True)
df1["studies"] = df1["studies"].astype(studiesDtype)

but this messed the table as all other countries are found in the regions they don’t belong to. See the table Error screenshot.

@desert oar

table_Before.png table_After.png table_Error.png
final kiln
#

today I'm gonna try to get some tooling going for this salad of languages and then study rust's type system in detail, I can't really do this without at least a good understanding how rust and cuda work, rn im very much a newbie in both

#

I think that building the binary automatically will have to happen over an aws instance, so there's that issue too

#

this is gonna force me to build and hold an image which is not for production, which is something I would like to avoid

#

I'm gonna try to map nvcc inside the container

#

installing it is a bit of nightmare because the prod image is using debian and it just is what it is

#

I'm gonna change the base image to pytorch's instead of prefect, the reason for using the prefect base image actually no longer exists, it was easier to maintain compatibility with their deploy system by using their image, but now I'm doing the deployments via actions only

#

so the hypothesis is that taking pytorch base image and installing the rest of the nvidia toolkit is not gonna be hard

spark nimbus
#

Running into an issue with pandas-on-pyspark where shifting a column with type timestamp will cause a mismatch error, saying it expected float but got timestamp. Repro: ```py
df = ps.from_pandas(pd.DataFrame({"date": pd.date_range('01-01-2000', '01-10-2000')})
df['date_prev'] = df['date'].shift()

frail hedge
#

Anyone know if Keras3 can work with older Tensorflow versions?

#

Or should I just try and get wsl2?

final kiln
#

I'm doing so many cursed things in rust and it's making me happy

final kiln
#

Oh no here comes malloc time

#

LETS GOOOOOOOO

#

I added two vectors on the GPU

#

With only 2 dimensions

#

I consider this an absolute W

naive forge
#

Could someone help me understand the inner workings of a 3 layered neural network during back propogation if we used cost function MAE and a sigmoid activation ( 1/(1+e^-x))? 🙏

#

like how to calculate the gradient 🙏

final kiln
#

derror/dw1 = derror/dlayer1 * dlayer1/dw1

#

Expand the last term as much as needed til you get to the last layer

naive forge
#

how would you get the first term tho 👀

final kiln
#

You have the analytic expression for any given layer

#

So you compute that, it will depend on the value of the current weights and on the input coming from the previous layer

naive forge
#

yea, but how to find the derivative of cost with respect to any layer

#

because cost is only defined with respect to last layer

final kiln
#

Cost is defined with respect to the entire function, which is a composition of layers

#

So you apply the chain rule

#

error = layer1(layer2(x))

#

Or maybe

E

#

Error(layer1(layer2(x)))

naive forge
#

what am i missing 😭

final kiln
#

No, because the output activation i oughta depend on output activation i - 1

naive forge
#

yea

final kiln
#

You can compute the derivative using that formula already, but you'll need to expand it when you apply the chain rule

naive forge
#

but it depends on all previous neurons in preious layer, how to find derror/d(some particular weight)

final kiln
#

If you go through the calculations for a simple composition of two linear functions

#

You'll see that you can calculate it as you have all the terms

naive forge
#

would you mind walking me through an example 🥲 🙏

final kiln
#

Gimme a sec

naive forge
#

😭 ty so much

final kiln
naive forge
#

ohhhhh

#

that makes sense, ty so much 🙏

final kiln
#

Good thing you ask, I actually need to start thinking about these for my new custom layer

desert oar
final kiln
#

Got tensors being shared over to the cpp code

#

After this I got to send them to CUDA and calculate stuff there

#

And after this I can get the cpp boiler plate for the forward and backwards pass

#

And then all I have to do is to code the calculations in cuda

mint mist
#

bro pls sum1 help me man i got no clue how to do this and i need the grade yert

Define positive_revenue_df as the subset of movies in df with revenue greater than zero.
Code is provided below that creates new instances of model objects. Replace all instances of df with positive_revenue_df, and run the given code.
Use this code to get started:







      positive_revenue_df = 

# Replace the dataframe in the following code, and run.
regression_outcome = df[regression_target]
classification_outcome = df[classification_target]
covariates = df[all_covariates]

# Reinstantiate all regression models and classifiers.
linear_regression = LinearRegression()
logistic_regression = LogisticRegression()
forest_regression = RandomForestRegressor(max_depth=4, random_state=0)
forest_classifier = RandomForestClassifier(max_depth=4, random_state=0)
linear_regression_scores = cross_val_score(linear_regression, covariates, regression_outcome, cv=10, scoring=correlation)
forest_regression_scores = cross_val_score(forest_regression, covariates, regression_outcome, cv=10, scoring=correlation)
logistic_regression_scores = cross_val_score(logistic_regression, covariates, classification_outcome, cv=10, scoring=accuracy)
forest_classification_scores = cross_val_score(forest_classifier, covariates, classification_outcome, cv=10, scoring=accuracy)
    
What is the mean of the 10 cross validation scores for random forest regression?
# 
It appears that the variables budget, popularity, runtime, vote_count, and revenue are all right-skewed. In Exercise 6, we will transform these variables to eliminate this skewness. Specifically, we will use the np.log10() method. Because some of these variable values are exactly 0, we will add a small positive value to each to ensure it is defined; this is necessary because log(0) is negative infinity.

Instructions
For each above-mentioned variable in df, transform value x into np.log10(1+x).

What is the new value of skew() for the covariate runtime? Please provide the answer to 3 decimal points.
#

this too :'(

left tartan
left tartan
mint mist
#

thanks!

final kiln
#

Aight, first torch tensor has been funelled to my kernel

#

Now it's time to figure out how to make a custom layer out of this

final kiln
#

Can't believe how smooth this is going

#

the buildup to this moment has been insanely long

2024-03-15-172747_1069x619_scrot.png
errant escarp
#

guys and also another question. What version of keras and tensorflow does GOOGLE TEACHABLE MACHINE USE? I HAVE CREATED A CODE THAT WORKS PERFECTLY BUT THE PRE-TRAINED GOOGLE MODEL ISN'T COMPATIBLE with the installed keras how can I fix this problem? I have downgraded keras and idk how to check the model info. I asked AI and it said that I could use h5py to manipulate the model .h5 file but it says it is not highly recommended because it can cause more problems. How can I solve this issue?

#

sorry for caps I'm pissed 😤

serene scaffold
#

No one is going to DM you. If they answer, it will be in this chat.

serene scaffold
errant escarp
errant escarp
#

not only image model

serene scaffold
#

Can you get to this menu and click the Tensorflow tab?

errant escarp
serene scaffold
#

She shows only the tensorflow.js tab

#

I suspect we'll see something informative on the tensorflow (not js) tab

#

Also I have to walk home from the coffee shop that I'm sitting in.

errant escarp
#

I have tried downgrading keras below 2.20 as I said because the model has a module named groups and keras >=2.20 doesn't use it anymore

lapis sequoia
#

Hi,is it worthy of learning the ML algorithms from scratch except just for knowing how they work and having an actual idea what they do ?

serene scaffold
lapis sequoia
#

So its been a while that I started learning ML through some courses I first learned to implement algorithms from sklearn etc now Im following a course in coursera and they implement the algorithms from scratch

#

because I know the pretty much do the same thing if i do it from scratch or import it

#

ill still be messing with the parameters

solar patrol
#

hey everyone, pandas-profiling is throwing attribute error to me

#

what am I supposed to do now? I have already checked the versions

teal lance
#

So I decided to instead use the fact that my strategy generates signals , use the signals to find where they happen the most and least with historical data 📊 this helps find patterns or days where I might have no edge and can avoid it before hand

IMG_4197.jpg
teal lance
raw mortar
odd meteor
raw mortar
mint palm
#

I have an OOPs related bug I believe,
My model is this

class PadPrompter(nn.Module):
    def __init__(self, train_conf):
        super(PadPrompter, self).__init__()
        self.pad_up = nn.Parameter(torch.randn([3, pad_size, image_size]))
        self.pad_down = nn.Parameter(torch.randn([3, pad_size, image_size]))
        self.pad_left = nn.Parameter(torch.randn([3, image_size - pad_size*2, pad_size]))
        self.pad_right = nn.Parameter(torch.randn([3, image_size - pad_size*2, pad_size]))
        
        
        self.prompt = torch.cat([above 4 parameters], dim = ... )

     def forward(self, x):
        print("#1 CHECK inside forward", self.prompt)
        return x+self.prompt

else where in code, I am updating paramter of PadPrompter's instance: prompter

        torch.nn.utils.vector_to_parameters(w_l, prompter.parameters())
        
        for  param in prompter.parameters():
            print(f" #2 CHECK outside forward: {param[0][0]}")

after torch.nn.utils.vector_to_parameters paramters in #2 check , I am able to verify that paramters have indeed changed. But later on in forward again i see old parameters

kind wren
#

What is the best AI I can self host on a Ubuntu with an AMD graphics card? It seems like all the AI models only work on Nvidia graphics cards.

abstract wasp
#

which tensorflow version is stable? i reinstalled tf and reran an old code (that was working fine b4) but now its giving me some error: 2024-03-15 23:57:05.807478: W tensorflow/core/framework/local_rendezvous.cc:404] Local rendezvous is aborting with status: OUT_OF_RANGE: End of sequence
I get 0 trainable parameters too

odd meteor
raw mortar
lapis sequoia
#

what range do I use for segmentation mask
if I have 2 values do I say 0 and 1 or -1 and 1

lapis sequoia
#

it just creates Black image

#

because i have a tiny thing that it needs to segment the rest is black so it just learns to make everything black

#

does anybody know how to fix that

lapis sequoia
#

i figured it out by saying loss(preds, targets) + loss(preds*targets, targets) * scale so that white stuff affects it more and it fixed it

latent thistle
#

Hello!

I'm posting this to see if anyone is knowledgeable in algorithms and artificial intelligence. Currently working on an unbeatable AI project for the game of Gomoku, I have constraints that require the Minimax (or variant) algorithm to have a depth of 10, as well as a calculation time of less than 0.5 seconds. The project is in Python with the possibility of developing in C.

I'm asking for help to figure out how to optimize my code to achieve a depth of 10 in less than 0.5 seconds each time.

If anyone thinks they can help me and is knowledgeable on the subject, it would be great to exchange ideas! Thank you!

tired rose
#

Hello, does anyone know a good way of creating an AI chatbot with some pre-trained conversational capability as to not train it from scratch?

solar patrol
#

its still not working

odd meteor
solar patrol
solar patrol
raw mortar
#

or downgrade numba to something less than 0.59.0

solar patrol
#

I see isee

#

thanks everyone for helping

final kiln
#

as im finishing up this project, I've settled on what my next one is gonna be, a dedicated ML language

#

not like mojo, I'm gonna introduce new concepts and ideas

past meteor
#

Like a programming language?

final kiln
#

yeah

past meteor
#

What features are you thinking of?

final kiln
#

actually, not a lot yet, this is an evolution of my idea of tricking the compiler into doing math checks

#

but I like the challange and will be thinking where I can innovate

past meteor
#

Ah I thought you'd have gone down the differential programming route

final kiln
#

well yes autograd will be built in

past meteor
#

but that's a deep rabbit hole 🐰

final kiln
#

for now im just gonna collect the problems that exist, and see how a new language could potentially solve them

#

one could be standardization

#

like what onnx tries

past meteor
#

It could also be something that isn't particularly innovative but is fast and ergonomic

#

that would already fill a gap imo

final kiln
#

yeah im going for ml dev experience

#

imagine the compiler pointing out mistakes in the model

past meteor
#

I don't want to scare you but you should definitely consider that there many fantastic langs with garbage tooling

final kiln
#

they released the tooling with the language

past meteor
#

So if you're doing this I'd say consider from day 0 how you'll integrate with LSP, how you'll do testing, ...

final kiln
#

exactly

past meteor
#

that's the way

#

Looking forward to your ideas! Always feel free to ask me to bounce ideas, it's something I find really cool

final kiln
#

awesome I appreciate it, I'll definetly need it because I'm not very experienced overall, I'm even thinking of starting the project very slow so I give myself time to gain experience in the industry

#

im probably gonna have to build up my knowledge, do an interpreter first, then start getting into simple compilers

#

There's one thing that mojo got super right, which is that no one is gonna leave py so intereopability is a must

#

Way I'm thinking, is that modules get compiled to python compatible binaries, which you can import directly into py

#

Like how Cython does it

#

I don't like the optional superset route tho, I think the type system has to be obligatory

solar patrol
#

I used this code

#

and it is throwing me an error

#

this is what I am getting in return

#

it says, it is a javascript error

raw mortar
solar patrol
#

C:\Users\Abhi\anaconda3\Lib\site-packages\ydata_profiling\model\correlations.py:66: UserWarning: There was an attempt to calculate the auto correlation, but this failed.
To hide this warning, disable the calculation
(using df.profile_report(correlations={"auto": {"calculate": False}})
If this is problematic for your use case, please report this as an issue:
https://github.com/ydataai/ydata-profiling/issues
(include the error message: 'could not convert string to float: 'S'')
warnings.warn(
C:\Users\Abhi\anaconda3\Lib\site-packages\seaborn\matrix.py:260: FutureWarning: Format strings passed to MaskedConstant are ignored, but in future may error or produce different behavior
annotation = ("{:" + self.fmt + "}").format(val)
C:\Users\Abhi\anaconda3\Lib\site-packages\ydata_profiling\model\missing.py:78: UserWarning: There was an attempt to generate the Heatmap missing values diagrams, but this failed.
To hide this warning, disable the calculation
(using df.profile_report(missing_diagrams={"Heatmap": False})
If this is problematic for your use case, please report this as an issue:
https://github.com/ydataai/ydata-profiling/issues
(include the error message: 'could not convert string to float: '--'')
warnings.warn(

GitHub
Issues · ydataai/ydata-profiling

1 Line of code data quality profiling & exploratory data analysis for Pandas and Spark DataFrames. - Issues · ydataai/ydata-profiling

solar patrol
raw mortar
#

try opening the html file in the browser directly

raw mortar
solar patrol
#

yes

#

got it

#

I mistakenly passed the wrong variable

#

mom was right

#

I am an idiot

raw mortar
solar patrol
#

thanks for helping me tho

raw mortar
#

no worries

graceful ledge
#

looking for some pandas advice

#

I am looking to generate a comparison of 2 xlsx files, where sheets correspond to 2d arrays. I'm looking to identify the differences across these two files. I'm using a for loop to create a dictionary of dfs for each file, then comparing. Am I overcomplicating this? I run into tuple issues at times

#

the arrays, generally, are the same shape across file A and B

#

TIA

#

wondering if it makes more sense for me to define a function that creates a class and subclass for each sheet, column, etc. to do this comparison

#

case sensitivity and sorting causes me some issues in my previous attempts to build this program

mellow vector
graceful ledge
#

each sheet is a state, each df is month (columns) by year (rows) with gasoline price in the values of the matix

#

file A is gasoline, file B is deisel

mellow vector
#

a for loop is probably fine but why move the data to dictionaries?

#

my core python is rusty but i would probably start by looping through df.iloc using a range(len())

graceful ledge
#

if shapes are comparable, I figure i could just A/B test each key, subkey, and value within dictionary A and B

#

probably why I run into tuple issues

mellow vector
#

hmm im definitely not the best person to give advice but i'd be shocked if there wasn't a simple method using pandas to conduct a/b arithmetic

final kiln
#

A feature I'll want to include is CUDA blocks

cuda {
   /// Your CUDA code here 
}

Something like that. Seamless integration for GPU code and getting my ecosystem to swallow Nvidia's, so I can find ways to control the chaos surrounding it rn

open raven
#

A dataset in csv file.
30970 numeric value fields each line, comma delimited.
21259 lines in csv
Gets imported to memory (device local memory 32GB for everything - all processes and data) using pandas read_csv()
Shape of resulting data frame: (1523, 30970).
Is it where pandas reaches its own limits? Otherwise where the mere 1523 in memory buffer from?

versed pilot
graceful ledge
#

wouldnt importing the excel data into classes and subclasses be a more effective way to do this, I likely would like to reference this data in the future

#

for calculations, visualization, etc

#

maybe im way off base

thorny zealot
#

To do a master's degree in artificial intelligence, is it better to do 3 years of computer science before or 3 years of electronics ?

open raven
thorny zealot
open raven
#

no, I step in into AI, since June ‘23

weary monolith
thorny zealot
#

lol

open raven
open raven
#

A dataset in csv file.
30970 numeric value fields each line, comma delimited.
21259 lines in csv
Gets imported to memory (device local memory 32GB for everything - all processes and data) using pandas read_csv()
Shape of resulting data frame: (1523, 30970), while it’s dtypes int64
Is it where pandas reaches its own limits? Otherwise where the mere 1523 in memory buffer from?

No improvement on change of dtypes for resulting data frame from int64 to int32.

desert oar
#

check the settings. and do you have any blank lines?

#

also why on earth do you have such a "wide" csv? is that raw count-encoded text data?

#

pandas i don't think has any limit on data size other than RAM available

#

if that's sparse data you might want to consider loading it into a sparse array instead of a pandas data frame

#

i think pandas used to have a sparse frame but i think they got rid of it a long time ago

raw mortar
desert oar
#

it's not just a matter of reducing memory usage, operations on sparse arrays can sometimes be faster due to not having to loop over a ton of zeros

graceful ledge
stone vector
neon field
#

how can i measure the accuracy of a computer vision project

final kiln
neon field
#

This is a posture estimation model
The output is the landmarks x,y coordinaties with visibility

final kiln
#

So a good measure could be the eucledian distance between estimated x,y and true x,y

neon field
#

Okayy thankyouu

final kiln
lapis sequoia
#

hey I need advice , I want to rotate a 2D plot as if it were in 3D : along any axis - x,y,z , like we can do with 3D plots .How can I do that ? any suggestions ??

final kiln
lapis sequoia
final kiln
lapis sequoia
final kiln
#

Uhm

#

As always the most flexible tool for this kind of job is pyvista

#

You can probably have a rectangle mesh and overlay the plot

#

But there's a simpler approach even

#

Can probably be repurposed, you just give it a flat topological map and overlay the plot image

lapis sequoia
mellow vector
#

is there a meaningful difference between df.column.copy() and df["column"].copy()?

#

i'd google it but i wouldn't know how to word it

fallen steppe
#

Hello I'm searching for tensorflow for my raspberry pi 5 with python 3.11. I found a guthub page which has the .wh files and stuff but it's not matching for my pi. It says cp311 but the .wh file is of cp310. Can someone help me find the tf version for me please

#

Here's the mismatch

db276699-8531-4898-bf90-049a8b1aa620.jpg
final kiln
fallen steppe
#

Yee I can't find the wheels for python 3.11

#

The image shows it to be for 3.11 but the wheel file reads as cp3.10

final kiln
final kiln
fallen steppe
#

Omg thanks

#

Sure one min

void crescent
#

i dont really know whats wrong with my model architecture but on the first epoch im at 10% accuracy

#

with negative losses

fallen steppe
final kiln
void crescent
#

heres the thing

#

im using a multi class dataset

#

but

final kiln
#

5% is weight initialization, and the rest is model arch

void crescent
#

for some reason

#

the system thinks its single class

#

so it forces me to use BinaryCrossEntropy

#

im using tfds's Colorectal Histology

final kiln
void crescent
#

the loss is at -300 rn

final kiln
#

im gonna see if I cross check their checksum later today, the wheels might be altered

final kiln
void crescent
#

when i used categoricalcrossentropy

#

it said that the y_pred's shape is [None, 0]

#

so i can tuse it

#

my guess is that the preprocessing went wrong

fallen steppe
final kiln
#

that's odd, I think the output layer has not been properly setup then, or you're using the loss function wrong

void crescent
#

dis is the architecure

# 
                             InputLayer(input_shape=(IM_SIZE, IM_SIZE, 3)),

                              Conv2D(filters=6, kernel_size=3, strides=1, padding="valid", activation="relu"),
                              BatchNormalization(),
                              MaxPool2D(pool_size=2, strides=2),

                              Conv2D(filters=16, kernel_size=3, strides=1, padding="valid", activation="relu"),
                              BatchNormalization(),
                              MaxPool2D(pool_size=2, strides=2),

                              Flatten(),

                              Dense(100, activation="relu"),
                              BatchNormalization(),
                              Dense(10, activation="relu"),
                              BatchNormalization(),
                              Dense(1, activation="sigmoid")
#

im kinda sure the output layer is right

#

and theres a mistake in preprocessing

final kiln
void crescent
#

oh shoot

#

it should b 8

#

thers 8 classes

final kiln
# 
                             InputLayer(input_shape=(IM_SIZE, IM_SIZE, 3)),

                              Conv2D(filters=6, kernel_size=3, strides=1, padding="valid", activation="relu"),
                              BatchNormalization(),
                              MaxPool2D(pool_size=2, strides=2),

                              Conv2D(filters=16, kernel_size=3, strides=1, padding="valid", activation="relu"),
                              BatchNormalization(),
                              MaxPool2D(pool_size=2, strides=2),

                              Flatten(),

                              Dense(100, activation="relu"),
                              BatchNormalization(),
                              Dense(10, activation="relu"),
                              BatchNormalization(),
                              Dense(num_of_classes, activation="that exp() activation that I totally forget the name ._.")
void crescent
#

nope

#

Dense(8) didnt work, its incompatible appareantly

#

ValueError: Shapes (None, 1) and (None, 8) are incompatible

final kiln
#

who's complaining ?

#

you might stll be using binary cross entropy

#

softmax that's the name, omgosh

#

you might not need it tho and let it be the raw outputs, at least in pytorch the crossentropy loss accepts logits instead of probabilities

void crescent
#

in that case im back to square one

#

because if i try to use BinaryCrossEntropy with Dense(8) it gives a similar error

tidal bough
#

there's also a minor difference that df["colname"] = ... works but df.colname = ... doesn't work and shows a warning, but that's for assignment.

mellow vector
#

ty i was going through a quiz, the instructor used dot notation and it annoyed me but i couldn't remember why

#

seems like it might not be a great habit

final kiln
open raven
final kiln
#

I'm doing a final test today before diving deep into the CUDA implementation of the forwards and backwards pass.

I'm gonna code a linear layer in cuda and apply backwards propagation to it from rust. It is not clear to me how the torch Function API works, so I need to poke at it til I get it right

#

Depending on this step tho, I reckon this is all done in a couple days or so, thus marking the beginning of the end of the project

#

There was a lot more that I wanted to do. I'm still weighing if it's worth it to advance to the second phase I had planned. I'll think about it when I get there tho.

latent thistle
#

Hello ! I am currently working on a school project that requires me to create an AI for the game of Gomoku which can beat me with as few moves as possible. We are required to use the Minimax algorithm or a variant. For performance reasons, I want to use GPU computing. i am doing the project in Python and would like to know which lib i could use for GPU calculations that would be suitable for my project. thanks!

serene scaffold
latent thistle
serene scaffold
latent thistle
serene scaffold
lofty thorn
#

hi everyone

latent thistle
#

Thanks !
https://github.com/farinaleo/gomoku/tree/dev/ here is the repository for the project. The code isn't very clean yet. The important functions are located in ft_gomoku/grid, ft_gomoku/AI, and ft_gomoku/rules. If someone could advise me on how to improve the algorithm and performance

GitHub
GitHub - farinaleo/gomoku at dev

Contribute to farinaleo/gomoku development by creating an account on GitHub.

tidal bough
#

does it have to be in python? do you have an estimate how much faster position evaluation should be to reach your goal?

latent thistle
tidal bough
#

What I'm trying to say is that if your program only needs to be ~10-100x faster, then it's about time you should rewrite the slow parts in a compiled language, and that should give you enough of a speedup. Whereas if it needs to be a million times faster, then the problem is with your algorithm and you should focus on that. So it's a good idea to make an estimate of that - e.g. measure how long a turn takes depending on depth for depths 1-7 or so, and extrapolate to depth 10.

latent thistle
tidal bough
#

I was thinking of evaluating random positions and not the actual gameplay (so that it can be done automatically), but from this... that first long evaluation takes ~1s for depth 7 and ~10s for depth 9, so we should expect about ~30x the time for depth 10 compared to depth 7. And in your video for depth7 some positions take ~15s to evaluate. So we're looking at maybe 450s for a bad enough position for depth 10, whereas you need 0.5s always. So there might need to be some algorithm improvements.

latent thistle
tidal bough
arctic wedgeBOT
#

ft_gomoku/data_structure/GameStruct.py lines 83 to 84

def get_player_turn(self):
	return self.player_turn```
tidal bough
arctic wedgeBOT
#

ft_gomoku/data_structure/GameStruct.py lines 108 to 109

random_player = random.randint(1, 2)
self.player_turn = self.player_1 if random_player == 1 else self.player_2```
tidal bough
#

in the actual minimax or the grid impl, I don't anything obviously bad, though.

latent thistle
#

Thank you! I'll consider your advice and optimize my code. thanks

lapis sequoia
#

Do anyone partition code using single Linkedlist in python

final kiln
lapis sequoia
final kiln
lapis sequoia
lapis sequoia
final kiln
#

your descriptions are not very clear, so I'm throwing guesses

desert oar
lapis sequoia
desert oar
#

also @lapis sequoia you might want to clarify your question more. provide enough information up-fron that someone can answer the question without interviewing you.

lapis sequoia
final kiln
lapis sequoia
desert oar
# lapis sequoia Noone replying there

if you asked the question the way you asked it here, you can't reasonably expect a helpful reply. chatrooms are somewhat asynchronous. especially in a big server like this one, you need to approach asking questions like on a forum: state your full question, with sufficient supporting detail for someone to answer it intelligently.

#

actually that holds true on small servers as well

lapis sequoia
muted hollow
#

Hi guys, im gonna choose my major in my college tomorrow. I want to continue my path as ai engineer but i can only choose between software programming and information system. Which suit best for my career guys

serene scaffold
clear dove
#

Hi everyone,
I am in search of a course or content through which i can learn Statistics for implementing it in Python afterwards.
can anyone please suggest me the same??

hot bridge
#

Hi there, personally I've learned basic statistics from that guy: https://www.youtube.com/watch?v=zRUliXuwJCQ&list=PLZoTAELRMXVMhVyr3Ri9IQ-t5QPBtxzJO

YouTube
Krish Naik
How to Learn Statistics for Data Science As A Self Starter- Follow ...

Statistical methods are mainly useful to ensure that your data are interpreted correctly. And that apparent relationships are really “significant” or meaningful and it is not simply happen by chance. Actually, the statistical analysis helps to find meaning to the meaningless numbers.

Python AI Tech news : https://www.youtube.com/playlist?list=...

▶ Play video
raw mortar
#

Josh Starmer arguably makes better videos tbh

https://youtube.com/@statquest

YouTube
StatQuest with Josh Starmer

Statistics, Machine Learning and Data Science can sometimes seem like very scary topics, but since each technique is really just a combination of small and simple steps, they are actually quite simple. My goal with StatQuest is to break down the major methodologies into easy to understand pieces. That said, I don't dumb down the material. Instea...

low gust
#

hi guy I looking for solution / guide toward deep Q Learning for the CartPole problem have looked at some example still don't really know how to execute in code any suggestions will be great thx

versed pilot
#

In terms of a course focused on Python, I found "Python Statistics Essential Training" by Matt Harrison on Linked In Learning to cover the basics pretty well, with some useful python tips and pandas 2 updates. https://www.linkedin.com/learning/python-statistics-essential-training-19258005/being-a-python-statistics-mvp

LinkedIn
Being a Python statistics MVP - Python Video Tutorial | LinkedIn Le...

In this video, explore a compelling example of the topic in action and learn how to use this space to grab the browsing learner. The TOC and course description exist on the course's home page to communicate what's in the course and who it's for.

hexed dawn
orchid cargo
#

Hey, how i do the data priprocessing? I have my book but im still little dizzy about eda and data priprocessing. And the one im very confused is about Manual feature selection. How do i selection it?

jaunty helm
hexed dawn
#

manual feature selection is only a way to understand how feature selection works, but in applied AI, you will want to use tools that help you with feature selection

#

(hm, might be a bit too technical, idk if someone has a link to a good tutorial ?)

jaunty helm
harsh minnow
#

Hey guys, what are the best performing open-sourced LLM's that can be fine-tuned on specific domain data, I also need it to havebetter reasoning, will any open source LLM with good fine-tuning be able to reason equally to GPT-4 or Claude 2?

agile cobalt
#

fine-tuning doesn't really helps "reasoning" in a general sense, at least not unless you have an absurdly large amount of data

final kiln
agile cobalt
#

I have no idea whatsoever about what you data is like so it could be fall anywhere from "non-tuned gpt4 fails terribly but fine tuning gpt2 works perfectly" to "gpt4 works perfectly as is, open source (even fine tuned) fails terribly"

final kiln
#

Math is the most obvious one, but math would be hard regardless right

agile cobalt
final kiln
#

Oh you mean languages he don't know

#

Yeah that would be an example

harsh minnow
#

I am trying to train the model for a specific domain of data. Let's say I need to train the model on medical data and pass all the medical exams. I am sure GPT-4 cannot be great at this right?

#

And will the benchmarks of the fine-tuned model exceed GPT-4? At least on the medical exam benchmarks?

agile cobalt
#

if you are going to work with something as serious as medical data: Don't ask for advice in a random discord server, do proper research based on published papers

final kiln
#

Could be a school project

harsh minnow
#

Yes, its just an example I used.

#

To convey what I want to achieve for example.

final kiln
#

I recall having a conversation about this a couple months back

#

And they told me it can be hard to do that

#

Look into Lora

harsh minnow
#

How "hard"?

#

Is it a training data issue?

final kiln
#

No idea, haven't tried it myself

harsh minnow
#

Is it a training data issue/size of the data needed or the underlying model itself is not capable at other closed sourced performing models?

final kiln
#

Uhm, I think it's just easy to destroy the information since the model no longer has a signal coming from the huge dataset it was trained on

But this is just a guess

#

There is this technique where you freeze the entire model and instead train extra weights that get added on top

young saffron
#

hi
i have a lab evaluation in this week in which i have to perform a eda on a database and provide some plots of it can anyone help were can i learn data visualization it would be a great help
thank you

harsh minnow
#

Also I have previously I've tried QLora, I need to dive deeper to see with much bigger data sets.

final kiln
harsh minnow
#

BTW is there anyone in this groups who might have an answer for this?

#

May be you can tag them to my question.

final kiln
final kiln
harsh minnow
final kiln
#

And if no one has tried it yet, well it's a potential paper

harsh minnow
#

Yeah this is so true, probably will do more research.

abstract rune
#

i am a student and recently i learn this in uni

data -> linear regression -> y [Rd -> R]
data -> principal components [Rd -> Wk]

now i wonder can we do this ?

data -> principal components -> linear regression

I am wondering can these concepts be combined 😅

clear dove
clear dove
raw mortar
wooden sail
open raven
#

Hi, Have got a data set with 45 variables. Among all those few are categorical. One categorical is comprised of 133 categories. Majority of 133 categories qualify to be handled by ordinal encoder. However few ones among 133 categories have s good chance to carry special meaning: zero, any, a/n. One further has the chances to be on either of two sides: special meaning or ordinal category.
A question regarding 3 categories of special meaning - any approach patterns exist to apply in such a case?

500 records among 175 thousands have one of those 3 special intent.
Eventually a source which could help to make progress in this matter?

desert oar
open raven
#

No idea, no hints in data set documentation, unclear if data set authors will respond. My thoughts run towards n/a for “not available”, alternatively “not applicable”.

limber kiln
final kiln
#

lets goooooo

#

aight, final steps

#

that's not a full impl yet, it's just a test layer, I gotta replace the core cuda kernels with the actual attention mechanism

tough latch
#

help me where the problem is , whatever is missing in your opinion

long canopy
#

is there a library that abstracts model and API interaction, so that I can freely switch between OpenAI, Claude or local models with minimal programming?

final kiln
#

uhm not that I know of, how different are they ?

#

olama seems to use a very similar api to open ai

long canopy
final kiln
#

I think it's super risky to build a product on top of open ai or anthropic

fringe ledge
#

help

final kiln
fringe ledge
#

im doing it for a school project

#

its about stroke and there is a lot with stroke and little with

final kiln
#

What I'm saying is that you have waaay more true labels than false labels

fringe ledge
#

yes

#

i need to find a new one

final kiln
#

So your model likely figures that it's just easier to guess all trues and it will still be right most of the time

fringe ledge
#

can you recommend me something?

#

a dataframe

final kiln
fringe ledge
#

its the true label

#

the ones with the strokes

final kiln
fringe ledge
fringe ledge
final kiln
#

Uhm I guess knn could split these

final kiln
#

How are you encoding the data

#

Uhm yeah, you need a better dataset and idk if it's reasonable to expect a generalization from a simple table dataset like that

fringe ledge
#

i cant find a good data base

final kiln
#

I'm sure things like age are predictors of someone having a stroke, but it won't be a perfect thing in the end right

fringe ledge
final kiln
#

Just cuz you're 80 don't mean you'll have a stroke

fringe ledge
#

yeah

final kiln
#

Theres a ton other factors like how healthy the person was, were they smokers, did they quit if they were

#

Etc

#

I think the best you'll get is potentially an overfit

fringe ledge
#

can you maybe recomend me a database?

final kiln
#

Or a fit to a prob dist

fringe ledge
#

i will just do a new one

final kiln
#

You might also find useful information about how people are trying to solve this problem

fringe ledge
#

ok thanks

#

🙏

trim needle
#

I don't want to disturb you, but are there LLMs who think, that is, think properly about an action that they want to do?

final kiln
#

Which might not be easy to define in the first place

trim needle
#

Well, there should think about it properly, as an example with arguments etc.

fringe ledge
#

i just want to know if i need a new data base or no

final kiln
final kiln
fringe ledge
#

🦶

final kiln
#

If you adopt a pragmatic way of thinking, and define these concepts operationally, then LLMs can think

trim needle
#

I don't want to sound like a professional (I'm not one) but I think I've managed to modify a T5 model in a way that makes it thinking...

uncut parrot
#

Hello guys I am newbie and just starting out with python and I want to create my own AI model and train it on different things so that it can perform productive and time taking tasks
I just wanna what's the first step and is there any preferred ytuber from u guys
Any help is appreciated
Thank you

final kiln
#

They have good documentation, so does hugging face

uncut parrot
#

Sure I will look into it

final kiln
uncut parrot
#

I am also planning on making a discord chatbot which can answer some questions and all
And also a bot game like dank memer bot if you know but a little bit simpler
I just wanna know if python will help me do it or do I have to learn another programming language

#

Like it will use some AI I believe as it's a Chatbot but I'm not sure tho

trim needle
trim needle
final kiln
#

Ig I personally believe something "thinks" when they are somehow able to have the experience of language. Like when you think, you are experiencing it, like when you look at the sky and experience the color blue. Idk at which point a bunch of electrons moving about in a particular way "wake up", so ig I just don't know if LLMs think

final kiln
trim needle
#

So my T5 model has just created its own language and logic in which it thinks.

final kiln
#

At its best, It very clearly joins concepts it has never seen before and produces something novel

final kiln
trim needle
final kiln
trim needle
#

I don't understand exactly

final kiln
#

Can you get it to try to convince you that it is awake

#

I tried that w/ gpt but it either says no or it is not very convincing

trim needle
#

It is based on a very small data set, and in German, you can be happy if he explains why he lied.

#

*you can read it in this thoughts

iron basalt
# trim needle I don't want to disturb you, but are there LLMs who think, that is, think proper...

LLMs (of the giving type I assume you are talking about), have feedback loops, however, they often lack an internal loop and so to get them to """think""" step by step, you can often explicitly prompt them to do so ("please give me all the step by step"). Since each step is now smaller and part of a feedback loop they can do some limited reasoning. This falls under Cybernetics if you want to learn more on this aspect of AI.

final kiln
#

Might be pretty expensive to get them in a loop like that

#

Unless it's a small model

iron basalt
#

It is expensive, there are also tree versions (basically tree search like in a chess bot).

#

Which are really expensive.

#

The trick is to have cheap internal loops.

#

These loops can even have hard coded logic, for example, maybe one module calculates/simulates a specific differential equation / control system.

trim needle
#

No, my model can think by thinking about what it wants to generate as the final output. He does this in a language he has made himself, over which he has no influence during training.

final kiln
#

I think things will get super crazy when these loops get cheaper to do, especially with multi modal

iron basalt
#

(They do arbitrary symbol manipulation for algorithms)

#

There are also Neural Proof Machines, which are like a simplified NTM.

trim needle
#

Just forget it, my English is too bad to translate and you won't understand it anyway...

iron basalt
trim needle
#

Now I don't feel like working on it anymore ._.

iron basalt
#

Historically the winning method in AI/ML has always been to have some basic unit that you can parallelize and spam a bunch of, individual neurons are kind of like level 0, this is like level 1 (a bunch of recurrent networks combined).

final kiln
#

Pretty cool, do you do research or is it industry work ?

final kiln
iron basalt
iron basalt
#

It's why we get something that feels kind of smart, but not really.

final kiln
#

Either that or it just needs more GPU

iron basalt
#

Well, it's a scaling problem, in the same way of using tick marks vs a modern numeric system.

#

I can just add more ticks (GPUs), but that is linear.

#

Much better is a better numeric system.

#

(Except it's worse than linear because communication overhead between GPUs)

final kiln
#

That does make sense, but ticks have a winning streak rn and the hardware will keep on getting better right

iron basalt
#

Yes, but if you plan for today, you will find yourself stranded in the future.

final kiln
#

I've started cuda programming btw, so Im biased now

iron basalt
#

Well, you will still need that.

#

There is a base level of compute needed, even with way better scaling.

final kiln
#

I do strongly believe that hardware dictates where stuff is gonna go. Like, if we want better models we need better hardware, even if that means something totally different like wetware or quantum chips

iron basalt
#

Yes, hardware is the most important factor.

#

Software is about making the least worst use of the hardware.

#

I don't think we need quantum or anything like that though. It's just the wrong architecture.

#

Von Neumann is not it for what we are trying to do.

#

(Brains are not Von Neumann)

#

GPUs just so happen by chance to work well with Deep Learning, kind of lucky (that they already exist / work well for it / mass produced for lower cost).

potent sky
iron basalt
#

(It's also a major thing that is missing causing stuff like hallucination, lack of explanation / reliable reasoning (which is what a lot of people really want out of the LLMs, but they can't provide (well/cheaply) in their current form))

wide wind
#

hello everybody, does anyone know if there is an ai that takes a code and score that based on the accuracy, a code corrector ai

burnt fiber
#

Hello guys, I am a first year computer eng. student trying to progress on the path of data scientist.

Do you have any websites or courses that will help me on this journey?
What do you think about the roadmaps on Roadmap.sh?

final kiln
#

Would probably beat the human brains computational power

#

But the pattern I observe is, the thing that prevails and will work, tends to be the thing that either packs more flops in a unit of volume, or the thing that somehow makes better use of the flops per volume

#

Regardless of architecture or any detail really, if you had asked me b4 getting into this stuff a couple years ago, I wouldn't have guessed that the GPU was the current holy grail of AI, it's easy to see in hindsight, but it's not obvious

true parrot
#

Does anyone here have experience with pygam and patsy? I'm new to using them and the documentation is breaking my head

desert oar
torpid flame
#

did you guys see the new devin ai 💀

#

i doubt it’ll replace software engineers but seeing how much ai has evolved it just doesn’t promise anything good for us

serene scaffold
spring field
serene scaffold
spring field
#

oh, is that a counterpoint for not promising anything good?

serene scaffold
iron basalt
#

(Inner thought loop concept (not a new idea, but we have better tools for it now))

#

Even just 1 inner/outer loop pair seems to make it way more effective, so the question is what happens when you have N of them and in a hierarchy... #data-science-and-ml message

untold dove
#

its just not as comprehensive as i would like tensorflow is my Personal favorite

patent roost
#

Hi everyone, currently dealing with the OCR with google vision. I found that the text recognition doesn't have a fixed pattern. Let's say, the document A, I got the "First Name" right above the user input in a form, and the OCR can detect it correctly. However, when I tried to use a similar format document B, it failed to detect the first name.

Is there any other way to detect a form with OCR? Or am I use it wrongly? btw, Im actually using nodejs for doing so, but i don't think it really matter tbh.

#

What I currently do is trying to extract the content I want from the detected text. For example, I will get the value of first name from a string - "First Name: ${firstname value here} Last Name"

long canopy
#

L40 spot GPUs are 70 cents an hour

#

why do people even buy GPUs for AI

#

federate your training and do one of these for inference, smh

final kiln
#

Wait what are the specs on L40

#

48gb, that's extremely cheap

final kiln
#

I bet if I perform a Fermi estimate for the number of parameters in the broca, I'd get more or less the number of parameters in gpt4

iron basalt
long canopy
#

at which point you only get charged for disk space

#

then kill the instance when you're off the computer

#

add a hook to your computer's boot so that it starts the instance paused

#

so it's actually way cheaper than 0.70 cents an hour

#

this whole thing is pure llmops

final kiln
long canopy
final kiln
long canopy
#

nice!

final kiln
#

I think Ohio and some of the Asian regions also have cheap GPU

long canopy
#

for training or inference? or both?

final kiln
#

For training and developing

#

Text classificStion doesn't really require more

long canopy
#

hm right I would also go with something cheaper if it is for training

final kiln
#

One thing I really wanna include in my language ecosystem is solutions for increasing accessibility to AI, like tools to provision cheap or even free GPU, but idk how feasible that could be

#

Most people don't have a GPU, or the means to buy one, and in most cases even if they do, it's not the most cost effective way to go about it

#

I do recall an idea I had a year back when LLMs started getting big

#

Which was to construct a block chain type of thing that would somehow use ML compute as proof of work

#

Perhaps I could have a feature where you could opt in to share a small percentage of your compute to other users of the language while yours is idle

#

Each person would get an equal percentage of the available compute

#

Don't know

long canopy
open raven
#

For categorical variable at what threshold of variable cardinality one should stop to think one-hot encoding to be good choice?
If such threshold exists, is it fixed rather than depend on other conditions, e,g. number of all variables, or number of all cardinal variables?

final kiln
final kiln
#

I think the closest I've seen is vast ai, where you can rent your GPU

#

Depending on the commission, it might actually make it worth it to buy one now that I think about it

supple inlet
#

Hello everyone, wheres the best place to find help with Cuda libaries and nvidia drivers. Ive got a tesla p40 and using ubuntu and cant seem to get pytorch to recognise my gpu. I think I have a miss match in drivers. Nvidia-smi works and picks up my gpu

final kiln
#

check that the version of cuda matches the version the pytorch you installed was built for

supple inlet
#

Ah right ill check that

final kiln
#

honestly I think im gonna write a small utility that checks peoples installation and tells them what to do, cuz these driver issues are so common

supple inlet
# final kiln run 'nvcc --version´

I vaguely remember doing this but ill double check. I've definitely done gcc --version. Ive been following nvidia docs and alexa gordic yt vid

final kiln
#

pytorch should totally do it tho, idk why they dont give feedback for it

final kiln
#

tho nvcc is not strictly needed

supple inlet
jaunty helm
#

Is there an elegant way to select from a pd.DataFrame all rows that contains np.nans in a specific subset of columns?
E.g. I have a DataFrame with the columns a b c d e, I want to select all rows that contain nans in b c or e

final kiln
#

this is the output coming from my layer on the forward pass, I gotta figure out how to get it through the gpu in an efficient manner

2024-03-19-140545_1434x529_scrot.png
#

c runs over the sequence of embeddings, k runs over the embedding dimension

#

when the dot product is ocurring between the same embedding, it's possible to simplify it cuz the p's commute and M is symmetric

#

I should be able to just place them in memory and calculate all the coefficients at the same time right ?

#

there's only gonna be one copy of p and M, and M will be a 1d array, so I'll just need to setup clever indexing and the right combination of blocks and threads

desert oar
#

note that np.nan is kind of just a hack placeholder for "null" values. you can have non-numeric data with over forms of "null" that are not np.nan. if you specifically want to check for floating-point NaN ("not a number") you can use np.isnan instead of pd.isna

#

!d pandas.isna

arctic wedgeBOT
#
pandas.isna 

pandas.isna(obj)```
Detect missing values for an array-like object.

This function takes a scalar or array-like object and indicates whether values are missing (`NaN` in numeric arrays, `None` or `NaN` in object arrays, `NaT` in datetimelike).
desert oar
#

!d numpy.isnan

arctic wedgeBOT
#
numpy.isnan 

numpy.isnan(x, /, out=None, *, where=True, casting='same_kind', order='K', dtype=None, subok=True[, signature]) = <ufunc 'isnan'>```
Test element-wise for NaN and return result as a boolean array.
desert oar
#

!d pandas.DataFrame.any

arctic wedgeBOT
#
pandas.DataFrame.any 

DataFrame.any(*, axis=0, bool_only=False, skipna=True, **kwargs)```
Return whether any element is True, potentially over an axis.

Returns False unless there is at least one element within a series or along a Dataframe axis that is True or equivalent (e.g. non-zero or non-empty).
jaunty helm
desert oar
jaunty helm
# desert oar good, that's what i would recommend in general

On a tangent, I want to make sure that this isn't leaking right(Kaggle Titanic)? I pass this to a FunctionTransformer

def get_titles(df: pd.DataFrame):
    df = df.copy()
    df['Title'] = df['Name'].str.extract(r".*?, (?P<Title>.*?)\.")
    df.loc[df['Title'].isin(['Mlle', 'Ms']), 'Title'] = 'Miss'

    majority = pd.Series(['Mr', 'Mrs', 'Miss', 'Master'])
    sub = ~df['Title'].isin(majority)
    df.loc[sub, 'Title'] = np.where(df.loc[sub, 'Sex'] == 'male', 'Mr', 'Mrs')
    return df
```If it is I can't tell how it's leaking
jaunty helm
void crescent
#

guys why does this happen?
``` ValueError: logits and labels must have the same shape, received ((None, 23, 23, 1) vs (None,)).

 the training dataset shape is  (110300, 50, 50, 3) (110300,) so why exactly does this happen?
final kiln
void crescent
#

the last layer is a Dense(1)

#

so what exactly is wrong there

final kiln
#

Same answer, post the code if youd like more detail

void crescent
# 

def create_model():
    model = Sequential()

    model.add(Conv2D(64, kernel_size=3, input_shape=(50, 50, 3), activation="relu"))
    model.add(BatchNormalization())
    model.add(Conv2D(64, kernel_size=3, activation="relu"))
    model.add(BatchNormalization())
    model.add(MaxPooling2D(pool_size=(2, 2)))

    # Dense layers
    model.add(Dense(512, activation="relu"))
    model.add(BatchNormalization())
    model.add(Dropout(0.5))
    model.add(Dense(256, activation="relu"))
    model.add(BatchNormalization())
    model.add(Dropout(0.5))

    model.add(Dense(1, activation="sigmoid"))

    # Compiling the model
    adam = Adam(learning_rate=0.0001)
    model.compile(loss="binary_crossentropy", optimizer=adam, metrics=["accuracy"])

    return model
final kiln
#

So the dense layers are acting on the last channel, which is why it's turning out 1

void crescent
#

so i have to make it 1-d again

#

or i can change it to Dense(2) right?

final kiln
#

No, you need to shape the output of the first part of your model into shape (batch, wtv number needed here)

#

The number needed will be 46

#

Uhm, potentially, idk what's the value of the last dim, since it has been projected to 1

void crescent
#

its 3

#

because r,g,b

final kiln
#

Not necessarily because the conv layers can add more channels

void crescent
#

in that case a Flatten() should work

wooden sail
final kiln
#

It's the tensor product of two unit vectors

wooden sail
#

unit as in unit norm or canonical basis vectors?

final kiln
#

Orthonormal, they're just there to make it a matrix

wooden sail
#

but generic orthonormal, not sparse orthonormal, right?

final kiln
#

I'm not brushed up on the terminology, ei dot ej is the Kronecker delta

wooden sail
#

i ask because you can consider multiplication by a rank 1 matrix as an orthogonal projection, so you should be able to replace the whole thing with a handful of clever dot products

wooden sail
#

so, canonical basis vectors

final kiln
#

I can leave it implicit by hiding the sum and the product

wooden sail
#

the tensor product of two canonical basis vectors is a matrix that is zero everywhere except for 1 element

#

and multiplying a matrix by another one from the right takes a linear combination of its columns

#

so any matrix times (ej tensor ei) yields a matrix that is zero everywhere except for one column

final kiln
wooden sail
#

all of the math you have there falls under the category of "nice on paper, terrible if you code it like that" :p

final kiln
#

I'm still gonna give it some more thought cuz there's symmetry going on

wooden sail
#

even before symmetry, almost everything is immediately zero

#

without any matrix multiplication at all, you can build the result by just indexing the matrix on the left

final kiln
#

Im very confused tbh, maybe I have some misconception, but the factor multiplying the tensor product is the (c, c') element of a 2d matrix

#

Stuff you learn in physics class tends to be very hand wavy so I wouldn't be surprised I've done a couple crimes against maths to get to the answer

wooden sail
#

i mean, none of it is wrong

#

just that almost everything is zero

#

no need to explicitly compute a bunch of zeros

final kiln
#

I'm not gonna compute the tensor product, it's for visual guidance only

wooden sail
#

ok, that's what i meant. you'd wanna directly index into another matrix and just assign the value there

final kiln
#

Yeah that's what will happen, the tensor products help me keep track of the indexes, ideally I wouldn't need them but I'm rusty

#

Neither the sums

#

I never did understand covariance and contravariance, so I don't do upper indexes >.>

wooden sail
#

😛

final kiln
#

I'm almost 90% sure no one does and everyone just memorizes the index rules

wooden sail
#

the easy rule is "row vector covector, column vector vector"

#

you can do einstein notation only with subindices if you're very explicit though

final kiln
#

Uhm, I do try that, but then I see sometimes matrices are upper down, other times down down, and then I'm just confused. I know it has something to do with the dual space of vector spaces, where linear functions live, but then I'm studying that stuff instead of getting stuff done, so I kind amove on

wooden sail
#

if you stick to the golden rule of "all vectors are column vectors", then row vectors form the dual space

#

then the row vector covector idea works 😛

final kiln
#

But then dual space is also a vector space

wooden sail
#

that's exactly the case

final kiln
#

It's confusing

wooden sail
#

why? that's a true statement

final kiln
#

Because, who's the dual of the dual ?

wooden sail
#

in finite dimensional vector spaces, that's isomorphic to the original vector space

#

and the map between them is the transpose

final kiln
#

So, like, theres no difference between a vector and a linear map, so the distinction doesn't look useful to me

wooden sail
#

in finite dimensional spaces, exactly so

#

except that if you arbitrarily pick one of them to be "vectors", the other is always covectors

#

this is super handy

final kiln
#

But in physics there's a huge difference between them

wooden sail
#

for example, it means gradients are vectors and jacobians are linear functionals, which is exactly how one treats it in differential geometry

#

and it makes taylor expansions easy to interpret: we take differential forms of a function

final kiln
#

One represents a real quantity, and thus transforms against the coordinate system, the other is a feature of the coordinate system and thus transforms in accordance to it

wooden sail
#

same here

#

and in both cases, the vectors and covectors form separate vector spaces

#

the dual space is exactly the vector space of linear functionals

#

in physics too

#

it's a separate vector space and you can assign it the meaning you like

final kiln
#

Okay, but here what decided what is a co vector was where I decided to put a coordinate system

#

Cuz if I put a coordinate system in the space of functions, then the other one has the other meaning

wooden sail
#

you can always apply a change of basis, what's wrong with that?

#

yeah, it just depends which one you want to study

#

that's just as true in physics as elsewhere

final kiln
#

In physics it does not

wooden sail
#

certainly so

final kiln
#

Some things are fundamental

#

You can't make them transform with the coordinate system

#

Whereas here is up to our choice

wooden sail
#

you can

#

whether it has any useful meaning is a separate matter

final kiln
#

That's a fair take actually

wooden sail
#

you usually don't do it because it's not immediately useful, not because it's not possible

#

"hey, check out this ???? i got? lmao"

final kiln
#

So in the case of the metric tensor, you get a down down indices right

#

Cuz it's a bilinear function

#

No, up up

#

I forget

wooden sail
#

i don't know off the top of my head

#

down down i think

final kiln
#

yeah they're doing down down in gr

#

oh and then it comes this stuff

#

it's a complete rabbit hole once you click on the "one-form"

wooden sail
#

that's where i was going when i said to treat the jacobian as a linear functional

final kiln
#

we call dx a little change in x, and you guys took it and built a giant castle of abstraction

#

I still get triggered cuz one of the professors in the math department decided that differential forms, manifolds and topology were appropriate for real analysis of 2nd year physics and engineering students, decided to invent his own notations, use PT instead of english nomenclature and allow nothing other than his own notes to be studied

viral escarp
#

Hey there! I'm fine-tuning and training both a YOLO model and an SSD model for object detection. I'm training them on a high-end GPU. However, I need to use these models on a CPU-only device. So, my question is: Is it okay to train on a high-end GPU, export the weights, and later use them with a CPU?

mild grotto
#

Hey all, I've got a cube laid out with each of it's 6 faces. Each step I'm adding some color and doing a gaussian blur (see gif). Because I'm doing each face separately, it causes discontinuities along the cube edges. I'm looking for speed/correctness suggestions I can try

cubeheat1.gif
wooden sail
#

for the edges it's easy: you can consider copies of the sides of the cube

#

the corners are tricky and there's no unique way of treating them

mild grotto
#

I saw an example saying you can use an Adjacency matrix (sparse). But it strikes me as probably going to be slow

wooden sail
#

should be fast if you implement it as a sparse matrix

#

defining the adjacency matrix for the pixels will let you do a graph convolution, which should be equivalent to considering copies of the faces as i suggested. i'm not sure what it'll do about the corners though

mild grotto
#

Ok I will try

#

seems like I can get it to blur using adjacency matrix on this test image at about 26 Blurs Per Second

image.png
#

My other image is bigger and can do ~53 blurs per second. Any optimization tips?

#

Maybe define dtype?

wooden sail
#

i'm a little surprised, the laplacian should be both symmetric and sparse

mild grotto
#

I can share the sample code I'm working from

wooden sail
#

you can also factor out the laplacian polynomial into a single matrix and multiply that by x

mild grotto
#

And the adjacency matrix generated like this:

adjacency = lil_matrix((num_pixels, num_pixels))
    # Define adjacency based on spatial proximity (e.g., 4-connectivity)
    for i in range(height):
        for j in range(width):
            node = i * width + j
            adjacency[node, (node - width)%num_pixels] = 1  # Connect to pixel above
            adjacency[node, (node + width)%num_pixels] = 1  # Connect to pixel below
            adjacency[node, i*width+ (j - 1)%width] = 1  # Connect to pixel on the left
            adjacency[node, i*width+ (j + 1)%width] = 1  # Connect to pixel on the right
wooden sail
mild grotto
wooden sail
#

ah that's what that is. lemme read how it works, one second

supple inlet
wooden sail
#

"slow matrix vector product"

supple inlet
mild grotto
wooden sail
#

try one of the ones it recommends there, csr or csc. i usually use coo. try them out and see what's easy to work with

mild grotto
#

I guess COO? (not sure what it is)

wooden sail
#

coo is coordinate array. it stores the matrix as tuples (row, column, value)

mild grotto
#

Ok I'll try a few things

wooden sail
#

the other flavors do things like storing indices of rows or columns with nonzero elements, then store those vectors completely. more efficient if the nonzero vectors are dense

#

i think your adjacency matrix might be banded and/or toeplitz too, so another option is to use fourier transforms. maybe try plotting the matrix as an image (i'm not super sure tbh, those corners are nasty)

mild grotto
#

Yeah the purpose of the cube mapping is to make a representation of a sphere that is optimized for matrix math (it's called "Cubed Sphere" and is used in weather simulations, afaik).

#

but it has disadvantage of corners

wooden sail
#

i like what you're working on 😌

mild grotto
#

the literature I read said that the corners should cause around only a 1% error rate 🙂

wooden sail
#

yeah probably not a huge issue, but there's no unique, perfect way of treating them

#

the adjacencies there are inherently 3d and you're applying a 2d map, any projection method you use is kinda wrong

#

graph methods are well motivated and understood though, so that's as good a method as any

mild grotto
#

yup, I think you just do a tiebreaker because yeah exactly, when crossing a corner boundary, there are 2 viable overlapping positions

final kiln
odd meteor
#

Is anyone attending ICLR or NAACL conference this year? We could plan a mini Python Discord dinner 😃

final kiln
#

The Vienna one I might depending on the date, not sure if I'll be travelling outside the continent this year tho, so the mexico one is unlikely for me

#

I think if I do visit the Americas I'll go for some conference in San Francisco

#

But don't know if this year will be possible

mild grotto
#

so it's like 12x faster

#

thanks a lot

wooden sail
#

nice

#

show a nice gif with the new result if you can

hollow root
#

Hey guys did anyone work on parsing SEC filings i need some help please

odd meteor
#

One thing I think people might be missing with the Gemma Open Models release is the section announcing grants given to academics to enable them push the boundaries of LLM research.

If you're interested in this, try to apply before April 17! All academics (universities + non-profits / independent labs) are applicable!

https://docs.google.com/forms/d/e/1FAIpQLSe0grG6mRFW6dNF3Rb1h_YvKqUp2GaXiglZBgA2Os5iTLWlcg/viewform

If you're an independent researcher and in need of an "affiliation" to apply, you can get in touch with me or just DM ML Collective on twitter.

20240319_214801.jpg
Google Docs
Accelerate your research with Gemma, Google's open model

Google is providing academic access to Gemma, a family of lightweight, state-of-the art open models built from the same research and technology that we used to create the Gemini models. Up to USD $500k in Google cloud credits (USD $5K/award) will be collectively awarded through Google cloud credits to selected applicants for novel academic resea...

nimble cobalt
#

good evening any idea on what libraries should I use for some data visualizatioN? Im working a project that takes input from the user and then displays data on graphs and all sorts

agile owl
# 
Acumen Pharmaceuticals, Inc. is a clinical-stage biopharmaceutical company developing a novel therapeutic that targets toxic soluble amyloid beta oligomers for the treatment of Alzheimer’s disease. Management will participate in a fireside chat as part of the Stifel 2024 CNS Days on Tuesday, March 19, 2024, at 12:00 p.m. ET. The live webcast may be accessed under the Investors tab on www.acumenpharm.com.

Acumen Pharmaceuticals Presents Sabirnetug (ACU193) Fluid Biomarker and Target Engagement Analyses from Phase 1 INTERCEPT-AD Study in Early Alzheimer’s at the AD/PD™ 2024 Annual Meeting. Key findings include improvements in biomarkers associated with AD pathology, including significantly lowered CSF neurogranin and pTau levels. Company on track to initiate Phase 2 trial evaluating sabirnetUG in the first half of 2024 and Phase 1 subcutaneous study in mid-2024.

Acumen Pharmaceuticals to present Sabirnetug (ACU193) Fluid Biomarker and Target Engagement Analyses from Phase 1 INTERCEPT-AD Study in Early Alzheimer’s at the AD/PD™ 2024 Annual Meeting. ACU193 is the first humanized monoclonal antibody to demonstrate selective target engagement of AβOs, a soluble and highly toxic form of A β that accumulates early in AD and triggers neurodegeneration.

I am actually pleasantly surprised with the quality of the text summarizations I got using an abstractive model

desert oar
serene scaffold
#

@scenic parcel you asked about assembly? No.

scenic parcel
#

doing data science in assembly.. that would be something

serene scaffold
mild grotto
#

I'm getting a weird effect. This is using a dot product on an adjacency matrix while a particle moves adding white color. The dot product acts as a blur filter (or in this case, just spreading the color out). But when it wraps around to the original starting location, it no longer shows up.

output.gif
#

I'm still adding color, but it doesn't seem to continue drawing

#

Oh I bet it's actually just a really high value

#

Nope, it's 0

wooden sail
#

that would be a multiple scattering problem

#

try a series expansion

#

rn you do Ax (matrix vector product) per time step

#

try (A + A^2) x

mild grotto
#

Mmm, I'm not sure I understand

wooden sail
#

you get each time step with a convolution, right? or how are you doing it

mild grotto
#

self.data = self.adjacency.dot(self.data.flatten()).reshape(self.data.shape)

wooden sail
#

yeah, that's a convolution

mild grotto
#

Ok so how would I change it to A+A^2?

wooden sail
#

self.adjacency is A

agile owl
mild grotto
#

so... (adjacency*adjacency+adjacency).dot(self.data.flatten())?

wooden sail
#

dot, not *

mild grotto
#

ok

wooden sail
#

this is just a rectngle, not a cube, or?

mild grotto
#

it's a cube, but i'm just testing in the horizontal direction to make sure I have everything set up right

#

which I'm pretty sure I don't

wooden sail
#

looks like the particle wraps to the wrong location, so many the adjacency matrix is wrong

#

how are you moving the particle

mild grotto
#

the particle wraps correctly actually, but it uses different code

#

it tracks which face it's on, and when it leaves a face, I apply a rotation and translation matrix

# 
self.sides[1].east_transform=[[1,0,-res],
                                  [0,1,  0],
                                  [0,0,  1]]

and

if(self.position[0]>self.face.resolution):
        matrix=self.face.east_transform
        self.velocity = np.dot(matrix,self.velocity)
        self.face=self.face.east
mild grotto
iron basalt
mild grotto
#

The bug is that it reaches a point where the particles are still moving but everything is zero

output.gif
wooden sail
#

this is too complex for me to see if the psrticles move correctly

mild grotto
#

Ok I'll turn off the blurring

wooden sail
#

and just put 1 or 2 particles

mild grotto
#

this should explain

output.gif
#

You may need to click "open in browser" to see it clearly

#

I think about it like you are wrapping a ribbon around a present. The ribbon just goes straight, and when it turns a corner, it maintains it's direction on the next surface.

v4-460px-Tie-a-Ribbon-Around-a-Box-Step-11-Version-2.png
#

Here I try filling the entire matrix with the value 0.005. You can see it quickly fills with a really high value, and then all goes to zero

output.gif
wooden sail
mild grotto
wooden sail
#

all right

mild grotto
#

yup. As you can see it lines up perfectly

wooden sail
mild grotto
#

yup

# 
for i in range(len(self.particles)):
      p=self.particles[i]
      p.step()
      p_2 = (int(p.position[1]),int(p.position[0]))
      self.data[p.face.ID,p_2[0],p_2[1]]=255
wooden sail
#

that should all be correct. (btw that's equivalent to the series expansion i was talking about. at each iter, you do A dot (A dot (old data) + new data), which if you evaluate recursively is a neumann series)

#

then i would wonder if your adjacency matrix is correct

mild grotto
#

so, here is the basic part of it:

self.adjacency = lil_matrix((num_pixels,num_pixels)) # sparse, only few connections
    for i in range(res): #width
      for j in range(res): #height
        location = i*res+j
        for g in range(6):
          pixel = g*grid_size+location
          mods = [(0,1),(0,-1)]#(1,0),(-1,0)] #4 directions
          for mod in mods:
            new_i=(i+mod[0])
            new_j=(j+mod[1])
            new_loc=new_i*res+new_j
            if(new_i>=0 and new_i<res and new_j>=0 and new_j<res):
              new_p=g*grid_size+new_loc
              self.adjacency[pixel, new_p] = 1

I didn't finish the cross-boundary parts because I assume there is still a bug

wooden sail
#

i'm not able to check that atm cuz i just woke up and am not lucid, but the cross-boundary parts are exactly the issue 😛

mild grotto
#

Isn't the whole thing the problem though? Even if the particle stays still, the whole picture turns black

#

it creates 1 wave, and then stops

wooden sail
#

meaning the boundaries are wrong

#

it's working correctly across some boundaries, but it's also not wrapping up-down for example

#

it seems to be wrapping only across a few of the squares

mild grotto
#

Yes, that's true. But I'll isolate the problem more specifically

#

Thanks for your help so far

wooden sail
#

the easiest scenario is to properly turn a rectangle into either a torus or a cylinder

#

to completely wrap around one axis (or both)

#

rn it seems there's one horizontal boundary where this does not happen correctly for the cylinder case

mild grotto
#

I've made the central face wrap with itself.
The particle is stationary.
You can see the problem is that the wave starts for ~10 frames and then no longer propagates

output.gif
wooden sail
#

yeah, the left and up boundaries are off

mild grotto
#

so the problem really seems to be: why is the particle no longer able to propogate after about 10 frames? It's as if it's consumed the food, and there is no food left

wooden sail
#

well, all of them actually

#

the edges are off

mild grotto
#

No the point is just randomly placed, sorry it's hard to tell

#

I'll center it

wooden sail
#

i know that's the case, but your matrix is applying a convolution based on your adjacency matrix for the cube

#

your matrix is wrong

mild grotto
wooden sail
#

it's missing connections at the boundaries

mild grotto
#

Nope, it's a torus now

wooden sail
#

certainly doesn't seem to be the case 😛

mild grotto
#

the problem isn't the boundaries, it's the center

wooden sail
#

are you putting the particle back in?

mild grotto
#

The particle is stationary in the center of the wave

wooden sail
mild grotto
#

every frame, it adds 255 to that location

wooden sail
#

all right

#

because for me, the same procedure works

2-scatter.gif
#

i'd have to see more of the code tbh

wooden sail
mild grotto
# 

    for i in range(1):
      x,y=np.random.randint(0, res, size=(2))#TODO: sample  sphere
      self.particles.append(Particle(self.sides[1],x,y))
wooden sail
#

ok but you're not using that here

wooden sail
mild grotto
# 

  def step(self):
    self.data = self.adjacency.dot(self.data.flatten()).reshape(self.data.shape)
    for i in range(len(self.particles)):
      p=self.particles[i]
      p.step()
      p_2 = (int(p.position[1]),int(p.position[0]))
      self.data[p.face.ID,p_2[0],p_2[1]]=1#np.random.randint(0,2550)
wooden sail
#

then for this case it's just the one particle fixed in one place?

mild grotto
#

yup

#

so we can ignore the step function, as it doesn't move

wooden sail
mild grotto
#

(and we can ignore the possibility that x/y are swapped, because x==y)

#

p.step moves the particle

#

but we set the speed to 0, so it does nothing

wooden sail
#

the particle list, the update of the data, or the adjacency matrix

mild grotto
#

yup, both the number of particles, and location of the particles are constant (just checked)

#

it's not the custom code, it's something about the numpy specific code

wooden sail
#

are you using int as your dtype?

mild grotto
#

float16

#

just tried float64, no difference

wooden sail
#

idk, this all screams to me that the point is not being placed correctly, only at the first iteration

mild grotto
#

I am testing by checking the max value, and I can see (after the wave ends) that the max value is exactly at 384x384 (middle of the center box as each box is 256x256)

#

it's just not... spreading for some reason

wooden sail
#

that'd mean you're not multiplying by the matrix

#

i think you have two similarly named vectors somewhere and are updating the wrong one, maybe?

mild grotto
#

I have it set up to do dot product first, then add the new data (so that I can see the unblurred point

mild grotto
orchid forge
#

Need some coding help with a data analysis related code

#

Is anyone interested to help

raw mortar
mild grotto
#

I forgot to normalize the dotproduct

# 
      self.data = self.adjacency.dot(self.data.flatten()).reshape(self.data.shape)/4``` that divide by 4 is kind of important
wooden sail
#

aha, there we go

#

i was also just in the middle of replicating your code, too

#

with a 1/4 factor and on a torus, we get this

result_flat.gif
#

that's no fun, the animation isn't showing

mild grotto
#

I can open in browser to see it

wooden sail
#

i guess you can kinda see that the rings start getting more intense

lofty thorn
#

where can i learn tensorflow

#

I can't find any

lapis sequoia
#

im gonna throw this message out there in hopes it falls on the suitable reader: im working with a SVM classifier model and im having trouble reaching past 80%+ accuracy score for my classification reports. could i consult with someone knowledgeable for advice?

wooden sail
# mild grotto ``` self.data = self.adjacency.dot(self.data.flatten()).reshape(self.data....

here's a way of making the kernel larger. you probably notices that the current one blurs in a sort of diamond pattern, since it only considers the immediately adjacent pixels. you could consider the diagonal pixels, or you can square the adjacency matrix to see pixels 2 hops away. i also scaled it by a little bit more than 1 so the animations turn out nicer (they still blow up, just slower)

# 
import numpy as np

N_0 = 100
N = N_0*N_0 #after flattening
torus_adj = np.zeros((N, N))
for p in range(N):
    col, row = divmod(p, N_0)
    indices = [
        (row - 1)%N_0 + N_0*col, 
        row + ((col - 1)%N_0)*N_0,
        row + ((col + 1)%N_0)*N_0, 
        (row + 1)%N_0 + N_0*col
        ]
    torus_adj[indices, p] = 1/4

filter = np.eye(N) + torus_adj + torus_adj@torus_adj
scale = np.sum(filter[0,:])
filter /= 0.97*scale
mild grotto
wooden sail
#

you need both

mild grotto
wooden sail
#

without modulo you get the boundaries wrong

#

(and squaring you also get the boundaries wrong btw)

#

just a little less wrong because the first step is computed with modulo

mild grotto
#

What would cause the boundaries to be wrong? Is it because it will double-count across the corner boundary?

wooden sail
#

because you won't get the wrap-around effect at all

#

what we need is that the bottom of each column is connected to the top, and similarly for the edges of the rows

#

actually, maybe if the adjacency is computed correctly, we do get the correct result after squaring. smh

#

i'm super rusty with graphs

mild grotto
#

looks like it's working

output.gif
#

Some of the medium gray blur appears to "vibrate" so I might need to do some more tests, but I'm going to call this a success

wooden sail
#

awesome

mild grotto
#

close up looks like it's doing a moire pattern

image.png
wooden sail
#

i get those too if i use the proper graph laplacian, but not the way i wrote it above

mild grotto
wooden sail
#

right, that way i don't get a pattern, it's just flat

#

the proper graph laplacian would be 4 I - A, where I is an identity mat and A is the adjacency matrix with just 1s (you'd have to normalize after)

#

and it looks like squaring the adj matrix works as intended. once we reach the edges, the pattern grows faster as we'd expect and the shape is no longer circular

#

(the identity mat is also only needed for stationary sources)

mild grotto
wooden sail
#

without parentheses

#

4 I is the incidence matrix (number of edges per vertex) for this graph

mild grotto
#

Wouldn't that make mostly high values where I don't have adjacncy?

#

oh wait, identity matrix, not ones matrix

wooden sail
#

this is already exactly what you're doing for the stationary particle

#

you apply the adj matrix scaled by 1/4, then put the particle back where it was with amplitude 1

#

so you're doing (4 I + A)/4

mild grotto
#

Ok so 4*np.eye(res*res*6)-A

wooden sail
#

the graph laplacian canonically has the opposite sign for the adjacency matrix, giving you a checkered pattern

mild grotto
#

oh, interesting

#

Oh, so this requires that I allocate 1 terrabyte of memory

#

can I just multiply by -1?

# 

      self.data = -1*self.adjacency.dot(self.data.flatten()).reshape(self.data.shape)/4
wooden sail
#

wait wait, i'm still playing around with it

mild grotto
#

looks like multiplying by -1 isn't good enough, but creates interesting patterns

output.gif
wooden sail
#

yeah

#

i think the issue is really just considering the 4 immediately adjacent pixels

#

using adj + adj^2 should do, without the negative sign

#

also, i used np.eye because i'm doing a small scale simulation, but you should use the sparse version of the identity or keep inserting the value by index manually as you were doing before

#

otherwise you get a dense matrix

mild grotto
wooden sail
#

by whatever the sum of one row of adj + adj^2 is

#

here i used adj + adj^2 with the normalization i just mentioned, and instead of an identity matrix i added in the source scaled by a cosine to get some ripples

# 
filter = torus_adj + torus_adj@torus_adj
scale = np.sum(filter[0,:])
filter /= 1*scale
mild grotto
#

?
self.adjacency =4*identity(res*res*6,format='csr')-self.adjacency

output.gif
wooden sail
#

trippy lmao

mild grotto
lofty thorn
#

I have to complete this project..

Can i know what are the prerequisites to complete this project
" 1.Analyze facial expressions in video sequences as time-series data using recurrent neural networks (RNNs) or long short-term memory (LSTM) networks. "

wooden sail
mild grotto
deep bough
#

Hii I kinda need help,
How can I extracting Words from Audio to Text without Auto-Correction

#

like if someone is prounancing 'klass' speech recognition convert it to 'class'

#

I want to stop that

mild grotto
#

Ok gonna call it done for the night, thx

output.gif
deep bough
#

@wooden sail 🙂 can you please help

mild grotto
deep bough
#

as everyone has there own accent example
Indains usually prounance 'Dha' not 'The' (Im not stereotyping anyone)
so if I try convert there speech to text I want it as it is.

mild grotto
# deep bough as everyone has there own accent example Indains usually prounance 'Dha' not 'T...

https://en.wikipedia.org/wiki/Phoneme
You want this

Phoneme

In phonology and linguistics, a phoneme () is a set of phones that can distinguish one word from another in a particular language.
For example, in most dialects of English, with the notable exception of the West Midlands and the north-west of England, the sound patterns (sin) and (sing) are two separate words that are distinguished by the subs...

#

You are trying to write phonemes using roman letters. You should use linguistic notation instead. That's what speech recognition should do

#

then, it's up to you if you convert the phonemes to words or not

#

And then you'll need training data. Probably for whatever dialect you're trying to record. For example, Mandarin needs different training examples for [má] [mǎ] [mà] [mâ]

wooden sail
mild grotto
#

and this is about as far as I can help as well. I've never done it

deep bough
#

basically I'm trying to make program for pronunciation checker and I made a logic where I'll convert someone's speech to text and will check if he said it right or not.

deep bough
#

I'll check

wooden sail
# 
import numpy as np
from scipy.sparse._coo import coo_matrix as coo_type
from scipy.sparse import coo_matrix as coo

N_0 = 50
N = N_0*N_0 #after flattening

rows = []
cols = []
values = []

for p in range(N):
    col, row = divmod(p, N_0)
    indices = [
        (row - 1)%N_0 + N_0*col, 
        row + ((col - 1)%N_0)*N_0,
        row + ((col + 1)%N_0)*N_0, 
        (row + 1)%N_0 + N_0*col
        ]
    cols.extend([p]*4)
    rows.extend(indices)
    values.extend([0.25]*4)

torus_adj = coo((values, (rows, cols)), shape=(N, N))

filter = torus_adj + torus_adj@torus_adj
scale = np.sum(filter[0,:])
filter /= 1*scale

if type(filter) != coo_type:
    filter = filter.tocoo()
#

the values could just as easily be 1 or anything else, since the "filter" is scaled afterwards by normalizing the rows

#

this is fun, what a huge nerdsnipe

final kiln
#

i drank to kool aid on index notation, it's pretty handy

2024-03-20-105428_1744x682_scrot.png
#

M for metric, E for expansion, C for contraction, same letters sum up to same value (c, c', c'' sum up to Nc)

#

im gonna calculate the gradients for this to see if there's optimization that can be done there, but my current plan is to focus on the term inside the softmax

#

no advantage, I'd just be doing what pytorch does already, which is to propagate the gradients, I'm gonna focus on the Mpp thing

abstract rune
#

Guys
I had 2 uni courses this semester

  • MLF (Machine Learning Foundations) - it was focused on maths [linear algebra etc]
  • MLT (Machine Learning Techinqes) - it was focused on application [algorithsm like PCA, Regression etc]

I feel like I have some holes in my understanding of them, I was thinking to read a book and make some notes so that I understand the whole thing clear and better

Can you suggest something it

wooden sail
#

if you wanna look at the math part with some small amount of applications, gilbert strang's book on linear algebra is great

final kiln
#

it's a general form tho, so I might be able to find something more efficient to compute and use that as my F

left tartan
final kiln
#

like, might as well just have more parameters in the layer than to spend time calculating this

2024-03-20-120956_1465x588_scrot.png
#

i dont really need F tho, I need f and g, so maybe I can construct them

abstract rune
abstract rune
abstract rune
final kiln
final kiln
abstract rune
abstract rune
final kiln
wooden sail
#

i guess they were asking about the image with the corollary you shared

final kiln
#

oh

abstract rune
final kiln
#

just a random arxiv paper I found online

abstract rune
#

a bit advance stuff for me

wooden sail
#

always double-check what you find on arxiv, a lot of the content has never been published nor reviewed

final kiln
#

I know, but I won't be using it anyway

#

served as a good steping stone to get to what I wanted

river cape
#

Hey guys , so for SVR,
We can predict like right,
y_pred = sc2.inverse_transform(regressor.predict(sc1.transform([[6.5]])).reshape(-1,1))

Whats the use of reshape(-1,1)?