data-science-and-ml

[pearl blaze]

Let me join it

https://www.freecodecamp.org/news/all-the-math-you-need-in-artificial-intelligence/

All mean majority part of ai/ml like 70 to 90 percent

And 1 more thing can i learn ai / ml and web scraping / flask / djanago, fastapi means stuff related to web

And is learning those 2 both does help me?

[proven pier]

Help you what? Learning anything with concrete uses can help you. People didn't design those tools just to be wastes of time and electricity

Can my car help me? Well, depends. Am I trying to drive somewhere, or am I trying to hike up a mountain? I dont think my car will help me hike a mountain

You're asking very vague questions. You should be able to figure out if something will provide use to you

[upbeat prism]

f = x * y + x / y

In pytorch, how do I get the gradients of the multiplication and division here? Preferably without defining them separately and also without using retain_grad. I'd like to use a hook. I know that works for nn:modules for some reason but for this is doesn't.

I just want all the intermediate gradients. 🙂

[serene scaffold]

Yes. You might need to define x and y with requires grad, though

[upbeat prism]

(Pdb) x = torch.tensor([[1.0, 2.0]])
(Pdb) y = torch.tensor([[3.0,4.0,5.0,6.0],[3.0,4.0,5.0,6.0]])
(Pdb) torch.matmul(x,y)
tensor([[ 9., 12., 15., 18.]])
(Pdb) torch.matmul(x.T,y)

How does pytorch know the representation of x? Could be a row vector, could be a column vecotr no? I assume they just look at the shapes and just transpose one of them?

[spare forum]

the way you defined it is not ambiguous, it's a row vector

 x= torch.tensor([[1.0],[2.0]])`

this is a column vector

[proven pier]

As gabiga said, they're not ambiguous. It might be ambiguous to you, but there's specific syntax. They would never auto transpose. You're responsible for dimensional compatibility, as it should be.

[final cobalt]

I have a simple question with a no doubt complicated answer

How do I detect a vanish or exploding gradient, why do they occur (not as a matter of math, but as a matter of architectural choices), and what do I do about them?

[upbeat prism]

ah actually true thank you

[upbeat prism]

the gradient vanishes because it gets smaller and smaller and at some point surpases what your computer can "display". it can only display numbers in a certain range, so at some point it's just basically -inf.

You can battle it by clipping your values.

[jaunty helm]

simple example: a single layer rnn, with just 1 weight w.
assume w is say 2, and we run it for 100 days. the value we get at the end is input * 2^100, and this huge value will make its way into our gradients, making them "explode"
assume w is say 0.5, run it for 100 days again. the value at the end is input * 0.5^100, which is tiny, and makes the gradients also tiny, thus "vanishing"

[lapis sequoia]

and single layer rnn is the same as 100 layer perceptron

sigmoid has vanishing gradient

relu not as much and generally shallower networks have it less

with batch norm or other types of norm you get less of it

or some fancy initialization like LSUV or orthogonal

[brave sand]

what is the minimum number of layers a CNN can have?

[unkempt apex]

what is minimum number of layers a NN can have? what do you think?

[brave sand]

2 right? input and output. so a CNN has 3 minimum because of the convolutionary layer?

[unkempt apex]

2 layers??
input , output?

then where is hidden layer?

how will your NN will learn about data then?

[brave sand]

so there is a minimum of 3 layers for the most simple NN?

[unkempt apex]

you can say that!

[brave sand]

well is it the same for a CNN?

[unkempt apex]

so a CNN is just -> convolutional layer + NN

[unkempt apex]

why conv layers?
to extract more features from data ( typically images )

and then this extracted features will gets converted into 1D tensor ( matrix )

and this tensor willl get feed into Input layer of NN

understood?

like this, here feature learning is CNN and classification is NN

[brave sand]

isn’t this more than 3 layers?

[unkempt apex]

the conv layers are separated from NN

so you can add more than 1 conv layers

[brave sand]

but the minimum is 3 right

[unkempt apex]

https://www.datacamp.com/tutorial/introduction-to-convolutional-neural-networks-cnns

read this

[unkempt apex]

for NN yes!

[final cobalt]

I'm trying to understand the relationship between problem complexity and gradient collapse

I've learned - by banging by head against the desk for a few days - that exploding or vanishing gradients are largely a symptom of too large a network applied to too simple a problem. Fair enough. I'm trying to understand the exact nature of why a more complex problem prevents gradient collapse within more complex networks. This is what ChatGTP said about the matter:

"More complex tasks provide richer and more structured information that helps deeper networks stabilize and perform well. Complex problems have multiple layers of abstraction, more diverse data, and rich error gradients, which allow the network to avoid issues like vanishing gradients or collapsing, especially when using architectures designed to support deep learning."

Would y'all say this is correct?

[severe shore]

hello, im wanting to get into ai, i have past experience with c++ and c# for software/game development with basic knowledge of python, im wondering what's the best way to start learning about machine/deep learning and what pre-requisites i need before i try learning about it, like what math, tools besides a python ide (pytorch, tensorflow, etc.) and whatever else i need to know

just looked through pinned mb xd but please let me know of anything else by all means

[rich moth]

its starting to look better, but now everything is in black and white. Ill see how more training goes.

[rich moth]

yes an input , convolutional and output layer

[rich moth]

Here's an example of what mine looks like ```class TokenImageCNN(nn.Module):
def init(self, embedding_dim, output_dim, h_prime, w_prime, c_prime):
super(TokenImageCNN, self).init()
self.embedding_dim = embedding_dim
self.h_prime = h_prime
self.w_prime = w_prime
self.c_prime = c_prime
self.output_dim = output_dim

    self.projection = nn.Linear(embedding_dim, h_prime * w_prime * c_prime)

    # Convolutional layers
    self.conv1 = nn.Conv2d(
        in_channels=c_prime, out_channels=64, kernel_size=3, padding=1
    )
    self.bn1 = nn.BatchNorm2d(64)
    self.pool1 = nn.MaxPool2d(kernel_size=2, stride=2)
    self.conv2 = nn.Conv2d(
        in_channels=64, out_channels=128, kernel_size=3, padding=1
    )
    self.bn2 = nn.BatchNorm2d(128)
    self.pool2 = nn.MaxPool2d(kernel_size=2, stride=2)
    self.conv3 = nn.Conv2d(
        in_channels=128, out_channels=256, kernel_size=3, padding=1
    )
    self.bn3 = nn.BatchNorm2d(256)
    self.pool3 = nn.MaxPool2d(kernel_size=2, stride=2)

    self.fc1 = nn.Linear(256 * (h_prime // 8) * (w_prime // 8), 1024)
    self.fc2 = nn.Linear(1024, output_dim)

def forward(self, token_embeddings):
    # Project token embeddings into image-like structure
    projected_embeddings = self.projection(token_embeddings)
    projected_embeddings = projected_embeddings.view(
        -1, self.c_prime, self.h_prime, self.w_prime
    )

    # Apply convolutional layers
    x = F.relu(self.bn1(self.conv1(projected_embeddings)))
    x = self.pool1(x)
    x = F.relu(self.bn2(self.conv2(x)))
    x = self.pool2(x)
    x = F.relu(self.bn3(self.conv3(x)))
    x = self.pool3(x)

    # Flatten and apply fully connected layers
    x = x.view(x.size(0), -1)
    x = F.relu(self.fc1(x))
    x = self.fc2(x)

    return x```

[pallid badge]

Hi everybody. Would somebody maybe know how to detect the horizontal lines, how many there are, and how much in y are the gaps. I ploted here np.sort(np.diff(sin^2(chi))) vs index. chi is discrete, delta chi = 10deg, 36 x. That reaveals the steps.

[left tartan]

You want to identify runs where the lag values are identical to current value?

[pallid badge]

I would like to know how many flat parts there are reliably.

Here: I split 360deg in 36 segments a 10deg=delta_chi. That should give me 9 flat sections

But within those flat segments there is some tolerance and the distance between each flat section separates them clearly.

[left tartan]

Yes, the trick is to label the data first for whether the value equals previous value, then you can group by the cumsum of every transition (not equals)

N other words, identify all the points where it's -not- a run. Then count those. If you just want a count, then you could count all of those where the gaps between them are larger than 1

If you want to work this out, you can open a help thread: start with code with an example dataframe and we can show you how

[pallid badge]

I don't understand. Sorry If I may: It gets harder, when I split 360deg in delta_chi=1deg steps. Then I would still get the flat segments , way more, but with a different spacing on the y-axis between them. I would like to identify this dynamically, no matter in how many segments I split my circle.

[left tartan]

Start a help thread with code for an example data set

[pallid badge]

Ok, I try this.

One last thing: I tried it with DBSCAN, it works , but still it is akward to find the right tolerance automatically

[pallid badge]

Done

[final cobalt]

    def loss(self, recon: torch.Tensor, batch: torch.Tensor, mu: torch.Tensor, lv: torch.Tensor):

        rc_loss = nn.functional.mse_loss(recon, batch, reduction='sum')
        kl_loss = torch.sum(1 + lv - mu.pow(2) - lv.exp()) * -0.5

        rc_mean = rc_loss.mean()
        kl_mean = kl_loss.mean()
        rc_std  = rc_loss.std()
        kl_std  = kl_loss.std()

        rc_loss = (rc_loss - rc_mean) / (rc_std + 1e-9)
        kl_loss = (kl_loss - kl_mean) / (kl_std + 1e-9)
        
        return rc_loss + kl_loss

I'm trying to find a simple way of normalizing the losses from different sources such that they are within more or less the same frame of reference

[desert oar]

If the losses aren't bounded in any meaningful way, subtract from each its sample mean, and divide each by its sample standard deviation

Standard data analysis technique: it rescaled everything to units of "standard deviations away from the mean"

oh, I see that's what you're doing, with a correction for 0 standard deviation

when std dev is 0 just bypass division and set loss to 0: there is only one value, so that value is the mean, so the result is just 0 after subtracting the mean

you're doing a kind of supervised autoencoder? trying to balance reconstruction loss with classification loss?

[final cobalt]

Yeah - mostly for practice at this point. As I understand it, adversarial "real or fake" classification can work wonders for image quality. I'm just practicing on MNIST digits right now but I'm going to scale up once I get all this figured out

As you say, the rescaling approach is probably simple and easy, but it's giving me trouble. Some kind of nan in rc_loss - and I can't even fathom where it's coming from

My weights are initialized with Xavier, and my layers are all pretty standard

    def __init__(self) -> None:
        super(VAE, self).__init__()

        self.dsize = nn.Upsample(scale_factor=0.5, mode='bilinear', align_corners=True)
        self.usize = nn.Upsample(scale_factor=2.0, mode='bilinear', align_corners=True)
        self.funct = nn.ReLU()
        self.noise = nn.Dropout(0.00)

        self.conv1 = nn.Conv2d(1,   16, kernel_size=3, stride=1, padding=1)
        self.norm1 = nn.BatchNorm2d(num_features=16)
        self.conv2 = nn.Conv2d(16,  64, kernel_size=3, stride=1, padding=1)
        self.norm2 = nn.BatchNorm2d(num_features=64)
        self.conv3 = nn.Conv2d(64, 256, kernel_size=3, stride=1, padding=1)
        self.norm3 = nn.BatchNorm2d(num_features=256)
        self.mu = nn.Linear(4096, 4096)
        self.lv = nn.Linear(4096, 4096)
        self.deco1 = nn.Conv2d(256, 64, kernel_size=3, stride=1, padding=1)
        self.deno1 = nn.BatchNorm2d(num_features=64)
        self.deco2 = nn.Conv2d(64,  16, kernel_size=3, stride=1, padding=1)
        self.deno2 = nn.BatchNorm2d(num_features=16)
        self.deco3 = nn.Conv2d(16,   1, kernel_size=3, stride=1, padding=1)

    def forward(self, x: torch.Tensor) -> torch.Tensor:

        x  = self.dsize(self.noise(self.funct(self.norm1(self.conv1(x)))))
        x  = self.dsize(self.noise(self.funct(self.norm2(self.conv2(x)))))
        x  = self.dsize(self.noise(self.funct(self.norm3(self.conv3(x)))))

        x  = x.view(x.size(0),  -1)
        mu = self.mu(x)
        lv = self.lv(x)
        x  = x.view(-1, 256, 4, 4)

        x  = self.noise(self.funct(self.deno1(self.deco1(self.usize(x)))))
        x  = self.noise(self.funct(self.deno2(self.deco2(self.usize(x)))))
        x  = self.deco3(self.usize(x))

        return torch.sigmoid(x), mu, lv

[desert oar]

Do you have a bug somewhere further up the pipeline?

I just recently fixed a bug caused by numerical instability leading to NaN that occurred ~20 lines of code above the actual error site

[final cobalt]

I'd like to run a possible normalization strategy by y'all. The above didn't really work even after fixing the error

MSE has a maximum possible value, and a discriminator can be made to output probabilities of real vs fake between 0 and 1. This should normalize the two values such that all I need to control is the weights as hyperparameters.

Sane?

[scenic parcel]

Does a 4090 make training ml models faster

[agile cobalt]

Doing multiple operations in parallel at the same time is faster than doing each of them sequentially one at a time, and GPUs allow for you to perform many operations which ml models require in parallel, so yes, if used correctly GPUs can be orders of magnitude faster than CPUs

That said, you have to setup and configure multiple things, and it will only make specific operations faster. It's not a magic plug-and-play device that will make everything faster, but it is effectively a prerequisite for training huge neural networks

[agile cobalt]

as for talking about "a 4090" in particular, depends on what you're using as a reference.

If you compare it with using a CPU, the difference can be of orders of magnitude
If you compare it with a slightly weaker GPU of the same manufacturer, the different isn't going to be that huge
If you compare it with a stronger GPU, it'll be weaker

[scenic parcel]

Well then my hyperparameter optimization is taking like 5 minutes per trial so I dont even want to know how long it'd take without one. I'm really putting it to work though, it was super fast for small example problems

[agile cobalt]

make sure that you're actually using it - again, you must set it up for it to work at all, and some libraries don't benefit from it

[finite thicket]

Hey guys, I'm trying to get into ML and I've got a lot of questions. I'd say I'm very proficient with Python, but I'm a complete newbie when it comes to this stuff. Sorry if I sound a bit vague here and there, as the exact details of the project are confidential.

I'm trying to build something where I can input an image of a face (from a frame of a webcam) and get back a certain kind of text response. I've already scraped 90k pairs of image urls of faces and text from online. When I input an image of a person, I want the model to give me back a response based on the data that I've given it. As in, I want the model to give me a response in the same style/tone as the rest of the data I've given it.

I know I'm supposed to use RAG, which I found a few tutorials online for. But how do I go about this for images? Do I use an existing model to come up with a description of facial features for each image that I have, and then store it in a vector database as text? Or is there a way I could directly put these images inside a vector database? Then how do I connect my LLM (Ollama) to that database? Are there any other important details I'm missing?

I also have no clue where to start- how do I get my giant csv of data into a vector database? What vector database do I use? Any advice would be greatly appreciated

[rich moth]

You can do this with an advanced RAG using a multimodal retriever and a vector DB, I like elasticsearch. You can embedded images, text, tables, etc. You would just need and embedding model for images but you can process it

[finite thicket]

what exactly is a multimodal retriever? and wdym by embedded images and text?

sorry i'm new to all this stuff

[rich moth]

Im got the quantum embeddings going and boy did this thing slow down lol. Only 185 hrs for 1st epoch. Im only using a batch of 4 too. I threw in some temporal cross-modal fusion to spice it up. Some other things, think i over did it.

[rich moth]

check this out https://docs.haystack.deepset.ai/docs/choosing-a-document-store

Check out encoding also with like BERT models or sentence transformers for text and look for image models too,https://docs.haystack.deepset.ai/docs/retrievers

[finite thicket]

I don’t really know where to start, how do I start making embeddings and putting them in my vector db? And should I go the route of storing images or text descriptions of the image?

And how would I make it so that it retrieves text using the image, and not the other way around

[rich moth]

The compression rationm is 588x the input data lol

[rich moth]

I can help you, though Ive only used elasticsearch for vector db. But check out the models first I'd start there, check out BERT and sentence transformers for encoding embeddding data

[finite thicket]

Will do. It’s almost 1am rn but I’ll check it out tomorrow, ty 🙏

[rich moth]

no worries, later

[upbeat prism]

Hello,

I'm writing my own tensor subclass to implement some functionality for research. Basically I want to overwrite reduction behaviour e.g. the mean in the cross entropy loss or the summation in matrix multiplication.

I can achieve a lot of that with those tensor subclasses written in python but there's also e.g. Tensor::sum() on the C++ level which I think I might not be able to overwrite with a subclass.

Question 1: Can I overwrite Tensor::sum() on the C++ level with a tensor subclass?
Question 2: Can I overwrite it on a C++ level?

[wooden sail]

what do you want to rewrite these for?

[upbeat prism]

https://arxiv.org/abs/2307.03056

Imagine you have a node with 4 incoming edges. Backprop would sum it up but maybe you are curious about: Which one of those 4 edges has the highest or lowest gradient? Or how is the gradient distributed over those edges? stuff like that. So you overwrite the summation to e.g. only consider the max. value or whatever you wanna do.

[wooden sail]

a first, unrelated, observation is that this doesn't seem to be published anywhere with peer review, so proceed with caution

[upbeat prism]

it is I just gave you arxiv because free

[wooden sail]

but yeah this basically requires you to write your own differentiation engine or fish out the intermediate results from the standard ones

[wooden sail]

google scholar doesn't show it being published anywhere reputable, but maybe you have access to something private i somehow missed ACL, smh

you can modify the backend code and compile your own custom module based on that, yes

but i would stop and wonder if this is something you want to use long term and/or possibly make public, or if it's just to test some preliminary results for research

in the latter case you can sidestep all of this by building the custom functions directly on top of pytorch or jax and just relying on the JIT

[wooden sail]

this requires going kinda in depth into the packaging and distribution of the module you're modifying to see how it's linked to the backend

[upbeat prism]

Yes, as I said, most of it can be achieved with tensor subclasses. I'm just wondering if I can overwrite things like Tensor::sum() as well as in if someone uses the C++ API, it takes what I overwrote in the subclass as well.

The thing is that we basically hook into the dispatcher but that doesn't matter if I wanna overwrite something using the C++ API

[wooden sail]

yeah but you'd have to recompile and link the module

that would replace the functions everywhere in the module, which may or may not be what you want

[upbeat prism]

Yeah, still trying to figure out everything because there are so many moving parts. But I have a pretty good idea now I think.

[wooden sail]

this probably isn't something you really want to do though... these backends are usually BLAS-like

with a matrix multiplication function defined separately for almost any case you could think of. different precisions, real, complex, transposed, symmetric, hermitian, etc

so even modifying matrix multiplication might require modifying some 20 functions (depending on the specific backend)

[upbeat prism]

yes

[wooden sail]

might be worthwhile to see which backends are compatible with the module you're using, see which one is the least problematic to fiddle with, and go with that

you would also have to distribute your module along with the backend, so i guess some reading about licensing is in order

[upbeat prism]

yes I already did a lot of research into it, the question really was just about if pytorch provides way to overload methods like sum etc. 🙂

I'm fully aware that e.g. matmul gets dispatched to cuBLAS which is even closed source. I currently just overwrite it for a specific usecase and use a basic naive implementation. To make it go fast, you'd have to write your custom kernel. I'm completely aware of all that.

I just noticed yesterday that I might be able to save a lot of work if I might overwrite some of the basic reduction oeprations on the C++ level. I read the code for the autograd engine ages ago but never looked into how PyTorch is exntesible on the C++ level.

[gilded belfry]

Is it possible for the inpainting model to ignore certain parts of the image when making predictions?For example, I have a picture like this. While inpainting certain parts of this image, I want black pixels not to be used when making predictions. Is this possible?

this is LaMa inpainting model output:

[final cobalt]

What sort of diagnostics can I do upon a neural model in order to get a better idea of the health of the activation and gradient flow?

[oblique isle]

who have worked here on CTGAN ?

[shrewd geyser]

Hi guys, hope you are doing great!

I'm having a big question. I've been working on a machine learning project about Customer Segmentation (RFM analysis), after cleaning, feature engineering, analysing, and more, I can't find clusters in the data, like all the samples are divided in only one group... Any recommendation on what to do? 😅

Thanks in advance guys!

[serene scaffold]

Hello, remember to always ask your actual question. never ask if someone knows about the topic and wait.

[novel mango]

import pygrib
import eccodes
import tempfile
import os

ds = pygrib.open('Tools/merged.grib')

with tempfile.NamedTemporaryFile(delete=False) as temp_file:
    temp_file_name = temp_file.name

    for msg in ds:
        with open('Tools/merged.grib', 'rb') as in_grib:
            gid = eccodes.codes_grib_new_from_file(in_grib)
            eccodes.codes_write(gid, temp_file)
            eccodes.codes_release(gid)

    h123 = ds.select(step=120)

    for msg in h123:
        values = msg.values
        new_values = values / 2

        gid = eccodes.codes_new_from_message(msg.tostring())

        eccodes.codes_set(gid, 'step', 121)

        eccodes.codes_set_values(gid, new_values.flatten())

        eccodes.codes_write(gid, temp_file)
        eccodes.codes_release(gid)

with open('Tools/merged_modified.grib', 'wb') as outfile:
    with open(temp_file_name, 'rb') as infile:
        while (gid := eccodes.codes_grib_new_from_file(infile)):
            eccodes.codes_write(gid, outfile)
            eccodes.codes_release(gid)

os.remove(temp_file_name)

ds.close()

in the output, theres a new time1 dimension that is replacing time when time is the dim that takes the new modified values, i have no idea what to do, searched all over internet/Ai and everything and asked but got 0 answers anywhere

[lapis sequoia]

1

which is 1 convolutional layer

for example it can learn to apply some filter

[rich moth]

seems like its finally starting to work. I've never seen the images with so much noise. reminds me of the white noise from tv's back in the day.

[final cobalt]

What are you making?

[unkempt apex]

he is mad scientist!

[rich moth]

I’m tryinng to build this multimodal model that blends images and text in a way you don’t usually see. I’m mixing things like VQ, CLIP, BLIP, diffusion, and this token image CNN I built. The CNN takes the token embeddings from the text and projects them into an image-like structure, then I’m using attention mechanisms to fuse both the images and text at different levels. It’s kind of like merging both modalities in a more dynamic and layered way... thats hopefully the idea annyways 🙂

[unkempt apex]

and what's the usecase?

[tidal bough]

Perhaps you can include that entire black part into the inpainting mask, and then just don't use that part of the inpainted area?

[rich moth]

So the main idea behind it all is to make image generation from text more expressive and controllable., Im thinking for example you can throw in a complex prompt with multiple objects, attributes , etc to try to capture more of the semantic relationship between text and images. Hopefully you can even refine those images by tweaking the text, trying to iteratively improve what it creates.. Maybe even a type of image captioning system or visual question answering too. But who knows man, it's all still pretty experimental at this point.

[finite thicket]

do you think BERT is best for my use case?

i read through the article, and i have a general idea of how it all works, but I'm not really sure where or how to start working on it

i don't really know what steps and what order i need to take to develop a model like that using BERT

[final cobalt]

I'm trying to find a way to normalize different components of loss functions so they're easier to optimize, and, offer better control over which loss function (goal) should take precedence at a given time

I'm building a VAE system with a twist - multiple specialized encoders. To do this I'll need various loss functions. Reconstruction loss of course, strong coupled with adversarial loss to enhance image quality. Training individual encoders also requires contrastive learning loss, and ensuring there is no overlap between the encoders requires some dientanglement pressure

In the interest of sanity, I've settled on normalizing all losses to between 0 and 1 (then muliplying each by 100 to prevent gradient collapse)

I figure I can express MSE loss between 0 and 1 by dividing the loss by its maximum possible value (for an image of a given size). The various adversarial losses are sigmoided on output, so they're already between 0 and 1.

The only thing I'm a bit fuzzy on is normalizing contrastive loss - this measures the distance between two encodings. I'm thinking I can use the average of the cosine distance between the encodings and the ratio of their magnitudes. This should reflect differences in both direction and magnitude equally while also staying within 0 and 1

And yes, I know this is all a bit weird. Multioptimization of different loss components is an open question right now, and it makes things much more orderly in my mind to normalize everything, then worry about weights separately. I havn't forgotten that different components of the loss will need different emphases at different times.

[random nest]

Not enough information to help you. What kind of clustering algorithm are you using? Tell us more about your data - number of variables, total sample etc

[rich moth]

The plan to normalize the contrastive loss sounds solid. I pretty much follow the same route for normalizing from 0 - 1 For the reconstruction loss its mse based. I have dynamic weights also but ive been brainstorming how to optimize it for the clip loss, i was thinking maybe thresholds and possible epochs but i'd rather stay away from it. This is what I got for my diffusion loss. You gave me a few ideas though.

    max_mse_value = img_size[0] * img_size[1] * data_range ** 2
    mse_loss = F.mse_loss(x_recon, noise)
    normalized_mse_loss = mse_loss / max_mse_value
    return normalized_mse_loss * 100 ```

[rich moth]

Did you check out the vector databases? What were you doing exactly again? lol

[rich moth]

What about EMA for dynamic weights?

[serene grail]

Very cool project

[rich moth]

thanks dude!

[final cobalt]

EMA?

I was thinking of writing a recurrent attentive system with a delayed reward to control hyperparameters

Something which learns to recognize short and long term patterns in learning rates and changes in hyper parameters, and which has a selective attention based memory that only hangs on to what it thinks is important

Also, I'm glad to see I'm not the only one whose thought of normalizing learning. You even multiply by 100, which is what I was planning on doing!

[finite thicket]

#data-science-and-ml message

i did, but i don't really know what to use to generate embeddings and store them in a vector db

[rich moth]

Exponential moving average. Try the recurrent attenntive system, sound like a really good idea.

ive made some changes and my metrics are looking a lot better and the reconstructions are coming in clearer.

[rich moth]

hey what about a multi agent with dynamic weight assigment? You can have multi agents for like spatial, temporal, dense, etc , You can agents handling different aspects. Each agent focuses on a X features in your data, then the system to adjust how much each agent contributes.

[finite thicket]

kinda confused on how to start my project

[rich moth]

Remind me again of your goal. Sorry,.

nevermind i see it up above, im gonna read it agian

[rich moth]

I see, I would suggest a CLIP model. You can use it to embed the text and images into the vector space

infact, see if you can get the sentence-transformer clip model to work and tell tell me how you did it, lol

[finite thicket]

hmm okay

i looked up CLIP model and this is what i found "So, for example, an image of a dog and the sentence “an image of a dog” would end up having very similar embeddings and be close to each other in the vector space."

this isn't exactly what i'm looking for though, the text i'm saving with the image doesn't necessarily describe the image

the text kinda refers to the person's face, but its its own thing

essentially I want to use an image of someone to look up a similar image in the vector database, then use the text thats paired with it

even if the image and the text aren't directly related

[rich moth]

Then just use a sentence transformer for the text and something like facenet

just remember to match the output dims of the models

I mean theres ways to get around it, but it would be easier to make

[finite thicket]

sorry get around what?

and so i just need to store these embeddings in a vector database? in what structure do i store it? i'm new to vector databases as well

[scenic parcel]

I should do a kaggle comp

[zenith lark]

Hey everyone I'm new to open ai APIs and I was testing it last day, it started to say insufficient token add billing address, can you guys tell me how much will I have to spend I'm a student so the budget is tight and I wanna learn and play with the api

[finite thicket]

i dont really understand what this repo is, is just the code they used to train it?

how do i actually use this model

[finite thicket]

also what’s a sentence transformer

[gilded belfry]

I gave all pixels outside the polygon an intensity value of 150, but I don't know how to ensure that the model does not use these pixels when processing.

Since my goal is to create an empty room, the floor needs to return to its original state. there should be no distortions

I separated all parts of the room into polygons ( floor, walls, ceiling).

In this way, the layout of the room will not be disturbed while inpainting

[tidal bough]

What I mean is: instead of having the model inpaint only the polygon, you can inpaint the polygon and the outside pixels, and then replace the outside pixels with the ones from the original image. (Though it's not very clear to me whether there'll be anything left, in such a case - do you only have 2 segments to the image?)

[gilded belfry]

i separated the image into 5 different polygons. With your method I have to combine all the images to complete the pixels outside the polygon. i think this does not solve the problem because other images also need this inpaint process. but thanks for your advice

When I give the pixels with a value of 150 the value "nan" in the mask, the shape is preserved, but the pattern of the ground is still not what I want.

[steep cypress]

Hello 👋 I'm training and autoencoder on signal data using LSTMs for anomaly detection.
For normalization I'm using sklearn.StandardScaler. For . fit(), should I only pass in the cleaned data without any deviating signals or the entire data?

[upbeat prism]

Can someone give me a nice neural network that's as simple as possible but does something? Can be made up but actually needs to learn whatever. Should only have like a dozen or so nodes in total.

[tidal bough]

The ones at https://playground.tensorflow.org/ come to mind

[quaint mulch]

https://iamtrask.github.io/2015/07/12/basic-python-network/

[desert oar]

That's a really cool toy/demo

[urban helm]

Okay thats actually impressive as hell.

[rich moth]

I let it run last night till this morning

that tensorflow playground is pretty wicked.

[pearl blaze]

Guys i wanna automate Arithmetic progression

Example if you wanna find n (number of terms)

But you have only last term and first term

And any random term

You have to make new formula

Another example, you wanna find d and you only know sum and and last term

Then New formula you have to use

And to make this python program complete you have to write logic for all 87 trillion

So your program can work in any possible states

And that possible states 87,178,334,440 that take million of years to implement in your program,
So the only solution you left with train ai for this Arithmetic progression

The ai will generate formulas on the spot instead of storing them all

My question : is simple Ai model can automate whole Arithmetic progression?

[final cobalt]

So, I've built a little VAE to reproduce MNIST digits. It's converged pretty well, it can replicate the general shape of the digit just fine

But the edges are a bit blurry. Of course there's a limit to how detailed a VAE can get - but I wonder if this is because the final layer has kernel size 3x3

Does this mean the pixels are being set multiple times as the deconvolution slides across the layer above? If so, is there some way to add a final layer with kernel size 1x1 such that every pixel is set only once?

Or, is that functionally equivalent to what's already happening?

[wooden sail]

i'm not sure i understand what you mean. are you given the terms a_1, a_N, and some a_n in between them, but not what n and N are?

[rich moth]

!paste

[arctic wedge]

Pasting large amounts of code

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

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

[pearl blaze]

Bro you know Arithmetic progression?

[wooden sail]

if you have n or N, this is pretty trivial. if you don't, then it can have either infinitely many sols or none, depending on how you choose your n and N

in either case, you don't need ML for this

[pearl blaze]

That's just a example

[rich moth]

I made a little something you can play around with. https://paste.pythondiscord.com/KGQQ

[rich moth]

im training something else so i can check it out right now though

[pearl blaze]

Is it work for all 87 trillion states

There can be 87 trillion formulas jn arithmetic progression

[rich moth]

lol no

[pearl blaze]

So what's this

I mean what this program capable of

[pearl blaze]

You should debug it !

[cedar tusk]

seems to me as a bruteforcing mechanism

u want to use ai to have a filterset for less iteration requirement?

[craggy agate]

I have got the YOLOV8 Nano model in the onnx format, how can I convert it to Tf lite?

[upbeat prism]

I need help designing a transformer model.

Given:

• Vocab: Number from 1 to 20.
• Input: Sequence of 10 numbers

Task: Does the first number on our list of 10 numbers show up again?

E.g. [1,2,3,4,5,6,7,8,9,1] -> Yes, [1,2,3,4,5,6,7,8,9,20] -> No

I want to overfit a model to it, such that I get 100% accuracy (assuming a balanced dataset).

I want the model to be as simple as possible but I can't figure out anything that actually trains. I'm currently trying this:

# Simple Transformer with an Attention Head for First Token Repeated Once Task
class SimpleTransformer(nn.Module):
    def __init__(self, vocab_size, hidden_size, num_heads, num_layers, ffn_size):
        super(SimpleTransformer, self).__init__()
        self.embedding = nn.Embedding(vocab_size, hidden_size)
        encoder_layer = nn.TransformerEncoderLayer(
            d_model=hidden_size,
            nhead=num_heads,
            dim_feedforward=ffn_size,
            activation="relu",
        )
        self.transformer = nn.TransformerEncoder(encoder_layer, num_layers=num_layers)
        self.fc = nn.Linear(hidden_size, 1)

    def forward(self, x):
        embedded = self.embedding(x)  # Embedding layer
        transformer_out = self.transformer(embedded)  # Transformer Encoder
        # We care only about the first token's output for the task
        first_token_output = transformer_out[
            :, 0, :
        ]  # Select the first token across the batch
        out = torch.sigmoid(
            self.fc(first_token_output)
        )  # Apply sigmoid to the first token
        return out

I generate 10k samples and use

vocab_size = 20
seq_length = 10
hidden_size = 128
num_heads = 1
num_layers = 2
ffn_size = 256
batch_size = 64
lr = 0.05
num_epochs = 100

Any LM expert here who might see something obviously wrong?

e.g. I'm actually unsure about the sigmoid.

[spring field]

can you share your metrics plots?

[upbeat prism]

I have no plots, it literally just stay more or less at the same loss/accuracy (which is 64%, a bit better tha nguessing)

[spring field]

for one, that learning rate seems pretty large, try with a way smaller one, like 0.001

[upbeat prism]

first time doing anything wit htransformers 😄

[upbeat prism]

tried 0.001, 0.005, 0.0001 etc. let me retry

feel free to suggest any metric to track

the whole code: https://bpa.st/LDUIK

[spring field]

the loss for one
also accuracy

[upbeat prism]

https://bpa.st/4O4EO loss & validation accuracy

[upbeat prism]

seems like increasing the data really helps a lot - curious. I tried the same with a veeeery basic bert model and there I needed waaaay less data. Now I'm at 50k

thanks

[rich moth]

what do you guys make of this?

[final cobalt]

Hey smart people!

https://hastebin.com/share/sosukokohu.ruby

I'm trying to build a super simple specific-use-case image sharpener. I'm going to try to remove the frames from and then outfill pokemon cards. I want to bring them all up to a consistent resolution first though

I have about 1000 google colab credits, or about 100 hours of Colab time. But, that time is money, and so

I want to make sure I've covered my bases before I sit down and start training

Any thoughts?

[clever current]

could you provide more context on what you're trying to do?

Or maybe I'm out of the loop. I can see why Slack has threads to keep everything together...

[clever current]

this really isn't my area of expertise, but here are some thoughts/questions: I can't tell where you're making sure your input images are all the same resolution first? Instead of using Google Colab, could you train and test locally? Super dumb question, do you have a test dataset so your model knows what success looks like? Or define some success criteria somewhere?

[glass jetty]

I'm having trouble with PyArrow+Parquet: I want to store data partitioned by a field a, which seems to work. When reading, however, it seems I can choose between streaming reads or filtering on the partitioned column, not both: ds.parquet_dataset doesn't support filtering by columns not present in the files (because of partitioning), pq.ParquetDataset does, but can only be read all-at-once into a pyarrow table or pandas df. What gives?

[rich moth]

Heres that ai powered symbolic regression system Ive been working on. Its up and running and im using it to train models on this dataset. Im using genetic programming and its utilizing polars. right now is doing the optuna parameter search.https://huggingface.co/datasets/yoshitomo-matsubara/srsd-feynman_easy

[rich moth]

Maybe you can use "pyarrow.dataset.Scanner" https://arrow.apache.org/docs/python/generated/pyarrow.dataset.Scanner.html#pyarrow-dataset-scanner

[rich moth]

ya check that out., maybe ds.Scanner.from_dataset I see that in there.

[glass jetty]

Hm, I'll take a look, thanks

[rich moth]

np. good luck

[glass jetty]

Ah, but that doesn't work with a ParquetDataset

Only with normal datasets, which already support the to_batches() method

[lapis sequoia]

?owner

[glass jetty]

(I figured it out: If you give the normal dataset the partitioning argument, then it can successfully filter on those columns.)

[rich moth]

!paste

[arctic wedge]

Pasting large amounts of code

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

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

[rich moth]

Im trying to design the preprocess data to be as broad and flexible as possible when dealing with datasets . You guys got any ideas or suggestions?
https://paste.pythondiscord.com/JRXA

probably should have went with pandas but i wanted to use polars

[jaunty helm]

datasets/models are different so there's no fit-all method

for example, you should not ordinal encode if you're using say linear regression

nulls in a ParkingLotArea feature might indicate that there's no parking lot, so you should fill with 0

etc etc

also

• if you pass train and test separately, the ordinally encoded values might not match
• if you concat train and test and pass the 1 df in, filling nulls with mean is leaking data from test, StandardScaler() is also leaking cause it sees test data, etc

for stuff like StandardScaler(), TruncatedSVD, etc. what you should do is .fit() (and .transform()) on train, and only .transform() on test

[final cobalt]

Not dumb questions at all!

To answer - I'm taking random 512x384 (4:3) crops from magic cards and marvel snap cards, adding some artifacts by doing an intentionally poor JPEG compression as well as a bit of noise injection, and then running that through the sharpener. I'm comparing the result against the original high-def cropping

[compact monolith]

If you find yourself looking for meaningful data to play with.... just start a data logging habit/hobby.
There are so many things i would like ML/AI to help with here, but I don't know where to start.

[final cobalt]

I'm working on a little project to enhance, remove the frames from, outfill, and then upscale pokemon cards. I want to start by sorting them all by frame type

Sadly, there are too many cards and too many types of frames to label them myself. Are there any straightforward ways of automated class discovery?

[scarlet anchor]

Hey, for training ML Models is it feasible to use AWS or amazon sagemaker or saturn cloud?

[serene scaffold]

training models is what sagemaker is for. idk what saturn cloud is.

[scarlet anchor]

okayy

whats the pricing?

[serene scaffold]

I don't know

[scarlet anchor]

ok

[quaint mulch]

Nice Viz.
This looks like timeseries.
Some easy things are finding correlations, forecasting, anomaly detection, imputation, nowcasting

[quaint mulch]

plot per iteration, instead of per epoch, if you are only showing 2 epoch

[pearl blaze]

Nice

[clever current]

Maybe you could limit your picture set to cards with all the same frame type? Start simple and train the model. I found this pokemon card database website and if you click Advanced Search, Format, Standard Legal, it appears to show cards with only a certain kind of frame. For the most part

https://www.pokemon.com/us/pokemon-tcg/pokemon-cards/4?cardName=&cardText=&evolvesFrom=&format=modified-legal&hitPointsMin=0&hitPointsMax=340&retreatCostMin=0&retreatCostMax=5&totalAttackCostMin=0&totalAttackCostMax=5&particularArtist=&advancedSubmit=

[clever current]

from a data analyst or business analyst perspective, you could try to answer questions like: what uses the most power in my home? how can I make my home more energy efficient?

[final cobalt]

This could be useful, but I've already downloaded all the cards.

Now that I'm properly rested (and I can see straight). Here's my project in a nutshell, and how I think I might be able to pull it off

I've downloaded 18000 pokemon cards. I want to sharpen them all to a standard resolution, remove the frames, outfill the background, and then upscale them to 1024x1024. This will be good practice with a bunch of tools, while hopefully being fairly simple

Unfortunately, the very first problem I'm coming up against (beyond sharpening, which should be fairly simple) is that there are like 30 different types of frame scattered throughout the dataset - not including variations by color

If I want to remove the frames, first I need to teach the system to recognize them. This would be simple enough if they were labeled, but they are not, sadly.

What I can do is create a bunch of pokemon cards using a card maker and use this to teach it to recognize all the frames it can. This should function fine for a large portion of the dataset, leaving only the outliers

Then, I figure I could make a bunch of cards using card makers from other games and, with a focus on generalization, try to train a system capable of one-shot classification

The only other option is a siames network trained on contrastive pairs of cards and augmentations of themselves. I think the idea here is that if it can learn to align the embeddings of a card and, say, it's flipped and slightly color shifted version of itself, the latent space starts to self organize. After seeing enough cards with very similar shapes, the orderliness of the latent space makes k-clustering a possibility

Any thoughts?

[desert oar]

some kind of image segmentation algorithm?

[final cobalt]

Hehe - see my wall of text above to know where I'm at with all this. Not sure if any of it is viable

[desert oar]

Yeah, I don't entirely follow all the details but I feel like you're thinking of some big complicated system instead of one step at a time

First thing is first: you want to extract the art from the image of the card, right? So just focus on doing that.

I agree this is a great project BTW. In part because it requires you to do exactly this kind of thing, pick it apart into sub projects and solve each piece one at a time.

[final cobalt]

Totally. Once I'm done I should have enough vocabulary and experience to tackle something super worthwhile - and, I'll get expanded art pokemon cards out of the deal

That said - I know I need to tackle this one step at a time. The first step, after detail enhancement, is learning to tell what's image and what isn't. If every card had the same frame then it would be simple. But there's a zillion of them - even the "base set" has undergone a bunch of modifications over the years

[final cobalt]

How about this.

Pick a card at random. For every other card, compute the MSE and select the 100 cards with lowest MSE. Have the system copy them to a working directory and manually inspect them. Remove anything not adhering to the most common card-type among them. The result should be a set of cards with the same frame. They become a cluster.

Repeat, this time omitting any cards already in a cluster.

Once every card has been clustered, repeat this process for clusters instead of individual cards.

[dry raft]

hey guys

for ViTs, do people usually use multiple transformer blocks or nah?

i may not be on, so reply to my message so i get the ping

[rich moth]

Anyone want to help me build this into a githib project? I feel like I got a solid foundation . I want to use GPU acceleration with rapids for polars but I was having issues with the conda install, i need to sit down and look at it again. Has anyone had success with it ?

[serene scaffold]

since you keep posting screenshots with lots of text instead of actual text that would help solve the errors you're having (I have asked you before to stop doing this), I'm going to start ignoring your questions.

[rich moth]

Alright man, seems really silly, but you're in charge.

[serene scaffold]

when you ask questions in a way that make people have to interview you in order to start helping you, you're wasting everyone's time, not just mine. if you're having issues with a conda install, the first step for the person who helps you would be to google the salient part of the terminal output from trying to install it. so if you want help with that question, you should post the whole output of running that conda install as text. Not a screenshot of some logging from an experiment.

[clever current]

Uh, I don't fully understand what you're saying since I haven't done work this complex. I did find this Youtube video about computer vision and cropping to the Region of Interest might be useful https://www.youtube.com/watch?v=kCyD0nfMwp0

[clever current]

I do wish this Discord had threads like slack, and I think the #1035199133436354600 channel auto-closing threads after an hour isn't that great, but your big screenshot doesn't make it the easiest for us to help you

I would also try googling the error message regarding the conda install, or try stack overflow. The problem you're running into sounds like something lots of others have run into too

[serene scaffold]

@clever current Discord does have channel threads like slack, in addition to forum channels (which are just bundles of threads). we don't allow non-mods to open channel threads.

[clever current]

Yes I know, but I see I can't start threads like you're saying

[serene scaffold]

but no matter where one asks their question--this channel, the help forum, SO--the asker needs to fully expose their question

oh sorry, you said this Discord. my mistake. I thought you were talking about Discord in general.

[clever current]

No problem! Agreed getting all the context is important

[compact monolith]

Nice. You hit a nail on the head. Reminded me of exactly where I want ML.
The power data.
Given nothing BUT the power data for a circuit, produce a list of "likely candidates" and their aggregate power consumption.

I can provide "hints" based on actual boiler plate device values, but ultimately it's a classifcation of the deltas while absorbing noise.
"Likely candidtes" means devices. Consider "An unnamed device consuming approximately 100W cycles for 5 minutes every hour. We suspect it is the same device, so we can label it."

I suck at ML though. I have the data, the CPU, 20yr SE exp, not near to 0 ML or AI.

[clever current]

Honestly my suggestions don't have to require ML. They're just basic analytics questions to me. You can just find max(power_usage) or make a nice graph of energy use over seasons

[compact monolith]

Already have that. My question is. Without having a power monitor on EVERY indivudual device, what can I learn from the aggregate deltas?

[clever current]

You don't need to use surgery tools when scissors will do. Also I'm curious your screenshot showed Celsius, are you based in Europe or something?

[compact monolith]

As a human, I can immediately spot the fridges.

ML should be able to also.

[clever current]

Ohhh it sounds like you want to take aggregate power consumption and try to de-aggregate it into individual devices, or at least make a best guess. I haven't really done a problem like that.

[compact monolith]

Exactly.

It's a dynamic feedback classification thing.

"thing" == not experienced in lingo.

My first attempt was to "hisogram" the deltas. quantise them to 10W intervals and ... no. it didn't work.

I off loaded data into PySpark and Java Spark, but just got lost.

the later did get me a lot more power over aggregation though.

[clever current]

You mean, histogram. And "the latter"

Sorry I don't think I can help you much with this, I feel like I've been spending too much time in this discord 😅 and I need to be applying to analytics jobs in my area (Texas) instead.

[compact monolith]

The system receives Watt values. If the last value was 1023W and the current is 1123W, then "someting that consumed 100W switched on".
Sorted right?
Unfortunately not. Noise.

This is why I went for "Quantise"

I got back to looking only at 500W changes and ... it still failed.

Obviously I have no idea what I'm doing.

[thorn pendant]

Hello every one, I hope you are fine, I'm 19, I started python 4 months ego and focused on django for about a month, I did some web applications pretty cool but I think of leavig all these to focus on AI/ML. Please I need some advice on where to start, good roadmap from beginner to senior and if I really made the right choice

[compact monolith]

Is your purpose to make salary or make engineering?

It matters because "AI" is a buzzword worth 90% of the value to companies.

They are wrong, but if you want in on the buzz and recuitment jazz.

[thorn pendant]

I will say, both to the first question. I am really passionate about programming in general and also the salary attracts me

[compact monolith]

Respect.

[thorn pendant]

do you have like a roadmap to guide me on this path?

[compact monolith]

My answer would be. If your salary and thus life doesn't depend on it....
Do whatever is interesting and productive.

If you are looking for a job. Different answer.

[harsh sun]

So I want to take an existing model for image recognition with camera vision, and add another class onto it. 1. whats the best way to go about that? 2. I have a dataset but how must I format each of the images for the model? I put paper towel below the subject in order to provide a contrast but idk how else I have to format it.

[thorn pendant]

can you be more explicite please

[compact monolith]

AI/ML etc are poorly understood by those who recruit, the get little help from those that know.

I am only trying to ask if you are looking for a job or have a more specific topic.

Let me put this another way>
https://youtube.com/watch?v=BKorP55Aqvg

Don't end up this guy.

[thorn pendant]

Actually I'm just exploring what I can and settle on something sweetable for me. I chosed AI ecause I have seen the exponential growth in its popularity and also because it somehow ties with what I like in programming(Logics). But please tell me more I feel so naive

[compact monolith]

If it's up to you, go big.

Go far.

No horizon.

[thorn pendant]

okay thanks and for the raodamp?

[desert oar]

MSE of what though?

You might be interested in hierarchical clustering... but I feel like I'm not understanding why this isn't just an image segmentation problem

[rich moth]

I was just wondering if anyone had any success with getting it installed.

[serene scaffold]

is the yes/no answer to that question literally all the information that you wanted? if not, you should bake the follow-up question into the original question.

"has anyone had success doing x? I'm trying to do x, and this is the issue I ran into: <code> <error message>"

just so everyone knows, @thorn pendant is also asking a similar question in #career-advice

[iron basalt]

ML is a basically a branch of mathematics, and so you will need to study the mathematics behind it. Since you already know some Python programming for web dev you should be ok on that end unless you want to be the person that implements all the ML libraries, in which case you will need to learn stuff like C++, CUDA (GPU programming in general), high performance parallel programming. The minimum required math that is often listed here is calculus (including multivariate), linear algebra, and probability/statistics.

Your web dev experience may also come in handy if you want some UI for what you are making and for working with databases if you have done that.

[desert oar]

I don't think you need to know any C++ or low level CUDA stuff unless you want to do ML eng specifically or research

Certainly not as a beginner... so much math and practical foundations to focus on

Not to mention basic data analysis skills and at least a bare minimum statistical intuition

[iron basalt]

Yes, only if you want to implement all those libraries everyone uses, it's an entire extra set of skills needed if you want to go down that path.

[desert oar]

Right, I just wouldn't include it on a roadmap for "AI/ML" in general

[iron basalt]

Some people really enjoy that, and so I bring it up as an option since it is its own unique job.

[desert oar]

Fair

IMO it's like putting low level networking protocols on a roadmap for web dev

Maybe useful in certain jobs but definitely not core

[iron basalt]

Yeah like how maybe you are into (backend) web dev, but really like databases so much so that you want to work on them in C/C++/Zig/etc. It's a different path that is still part of it all.

In the case of ML, you need the math either way.

[final cobalt]

https://tenor.com/view/waiting-comedy-ryan-reynolds-gross-sick-gif-3496531

[final cobalt]

I'm probably just misunderstanding. It is a segmentation problem - the problem is that I don't have the time (or the artistic skills) to segment even a small sample of every variation of every type of frame in every color

Which, in my very inexperienced opinion, means having it learn to do it on its own, and that means comparison of something against something else. That in turn means classification. Is there something I'm missing?

[desert oar]

Can you post an example? I saw you mentioned MTG above but then I thought you said Pokemon later on

And remind me, how many distinct frame types do you think there are?

And can they be grouped at all into similar categories? Like 3 different types that are all pretty similar

[fading wigeon]

Hey, I have a basic question about gradient descent. How does it find absolute minima? It seems like it would only be guaranteed to find a local minima. How do you ensure you find the global one?

[final cobalt]

Sadly, no. There are two dozen or so frame types, and they've undergone multiple minor alterations over the years. They also differ by color. In total, hundreds of different frame types

[unkempt wigeon]

Gradient = abs(min(output))

Loss = Gradient - output 

print(loss)

[agile cobalt]

it doesn't finds the global minima
(unless the gradient only has a single local minima, but that's not the case for ML)

if you train the same type of model on the same data with a different seed multiple times, it is entirely possible that it will reach a different minima and get stuck on a different loss each time

[desert oar]

Correct it is not a global minimization unless the problem is convex

That's part of why things like early stopping, gradient descent variants like "momentum", and parameter initialization can matter: you're not even necessarily trying to find a local minimum, because that might be overfitted to the training data

The actual objective function is generalization error, not loss on training set: but we have no way to compute that, so we use validation/test sets, cross-validation, resampling, etc to estimate

[desert oar]

Yes but how minor are the minor alterations? Can you show some examples? Are there any broad categories at all? Or are there hundreds of completely different frame types?

At the end of the day I do think you need to roll up your sleeves and manually annotate a couple hundred cards, just so you have a good baseline to test your algorithm. But I doubt you even need a CNN for a first pass at this

And regardless of whether you use a CNN or something simpler / more old-fashioned, yes you are going to have to go off-piste a bit and pursue some kind of "semi-supervised" approach, whether that's actually training a model on a mix of labeled and unlabeled data, or a multi-step thing

Fortunately, that kind of problem arises all the time in industry, and I think having experience solving that problem is tremendously valuable

You can also consider the active learning approach where you use a model to help you choose fruitful examples to label, instead of randomly sampling

[final cobalt]

Well, I've already toyed around with some possibilities for automated class discovery, and gamed out potential strategies for manual labelling. Fortunately, I think I've found a halfway decent solution

[desert oar]

But seriously, have you tried just scratching out some heuristics first? At minimum it can help you get a better idea of how a good model should work

[desert oar]

Good. What did you find?

[final cobalt]

Here's one example - EX versus EX mega

[desert oar]

Holy shit, those are a lot more advanced than what we had in the 90s

What would you even consider to be the frame boundary in those examples?

MTG would be much easier haha

[final cobalt]

And some examples of how the "basic" frame has changed - a few examples mind you, they change almost every expansion

[final cobalt]

Well exactly. Somehow I need to remove "everything but the art". Obviously it's a lost cost for some types of frame, but I can get most of them I'm sure

[desert oar]

You're not actually interested in classifying frame types, right? You just want to get the art extracted so you can upscale it?

Yeah, I wouldn't even bother trying to classify these

[final cobalt]

One helpful person suggested using the official Pokemon.com card database. I was initially skeptical because I doubted the search algorithm would cooperate - but it turns out I can search by set then card type then energy type. This will work pretty well for labelling I'd say

[desert oar]

Forget everything I said above, now my first instinct is to put together some kind of unsupervised vector embedding (autoencoder?), and then sample several dozen cards as evenly spaced as possible across the embedding space, for manual annotation

[final cobalt]

The images are much smaller, and that's both a blessing and a curse

[final cobalt]

SEE!? SEE!?

I'm NOT crazy!

[desert oar]

How does that help?

[desert oar]

I mean, these frames aren't even always "frames" -- forget it imo

[final cobalt]

Well - most cards are fairly simple. 80% or so I'd say have fairly straightforward layouts with a clear distinction between image and frame. I'm happy letting the other one's die off as casualties of war

[desert oar]

Ah, okay

In which case you might still want to focus on the frame, but with the specific characteristic that indeterminate or unusually complicated cards get rejected as such

But I would still avoid trying to classify all frame types

[final cobalt]

As for why I'm doing this at all - practice. I've got a big idea for a cool somthingorother, and I need practice. This project will force me to get my bearings on upscaling, autoencoding, segmentation, and perhaps even diffusion

[desert oar]

My initial instinct was that, and most cases, the art itself will be sharply different from the surrounding card background, so whatever you do should focus on finding that boundary

[final cobalt]

Yeah - some people mentioned canny and other simple algorithms, but I don't think that'd be quite right. And who knows! With the right architecture and maybe a pinch of meta-learning, the thing could work better than we'd think

For cards without frames, the ones "where its hard to tell where the frame begins and the image ends," that could be treated as a restoration/text-removal problem instead of a segmentation problem

Anyway - using this pokedex I can sort first by set - and cards have consistent shapes across sets - then by card type and then by color. These will result in folders with a few dozen cards each. A little manual inspection to fuse sets which didn;t change card layout and bam, labeled

[desert oar]

That's a very good idea, use whatever information you have available instead of starting from nothing

[final cobalt]

Well

The pokemon database has some decently robust anti-bot measures

Which means I'm going to need to crawl through it XD

[north adder]

Hello hope everyone is well. Im working on a feature in a application and saw that i can implement an AI Model. i want to compare Result with expected result and output if test is positive or negative. i never worked on projects so that would be my first project and i have a bit of knowledge in ML(Andrew NG Course). Can someone help tell me what type of problem is this called and what algorithm/s should i try. Thank you

[pearl blaze]

Can somebody tell me where we tell in this code that we need a steps of 5 on x axis and step of 20 on y axis ???

[jaunty helm]

are you looking for model evaluation?

[north adder]

yes

i want to know if possible which lagorithms should i read about that might help me solve this

[jaunty helm]

depends
if your model predicts some number like 12.3 then it's a regression task
if your model predicts from a set of possible answers, like it must be either 'apple' or 'orange', then it's a classification task
different tasks has different metrics you can score your predictions vs. real values against, which in turn will tell you how good your model is

[north adder]

it is a classification model with eithe rpositive or negative

it compares result with expected result and conclude if positive or negative

[jaunty helm]

then it's classification (only positive or negative), and a binary classification at that (only 2 possible outcomes)
simplest to understand is accuracy, or (number of tests the model got right) / (total number of tests)

[north adder]

yes thank you but do you know what algorithms should i research about?

[jaunty helm]

wdym by 'what algorithm'
what algorithm to do what? to do the predictions? to measure how good your model is?

[north adder]

do predictions

[jaunty helm]

then what you're looking for are the models, not how to evaluate them
what about looking at some simpler ones first like logistic regression?

[jaunty helm]

and, if you're already following a course, shouldn't it tell you what models are?

[north adder]

i was thinking of logostic regression but i am working with texts not numbers

[jaunty helm]

then the keyword you're probably looking for is NLP, or Natural Language Processing

[north adder]

i will check it out

thanks

[north adder]

is there anything else you advice me to read about?

[jaunty helm]

not really
what you said basically only tells me you're doing an NLP classification task

[north adder]

okay thanks

[desert oar]

in the definition of the lines you plotted, in the above cells

[desert oar]

there are many resources on "text classification". the technique is almost always the same: figure out a way to represent each document as a list of numbers, and put that into your logistic regression or whatever model

[pearl blaze]

where is the definition of lines in this code? and what is definition of lines?

plt.title("Bank robber got caught")
plt.xlabel("Time (in minutes)")
plt.ylabel("Distance (in km)")
ax.set_xlim([0, 40])
ax.set_ylim([0, 100 ])
plt.plot(time, robber_distance, c = 'red')
plt.plot(time, sheriff_distance, c= 'brown')
plt.axvline(x=30,color='green' , linestyle = '--')
plt.axhline(y=75,color='green' , linestyle = '--')
plt.show()```

[desert oar]

In ancient times we used techniques like tf-idf and feature hashing. Now we use a lower-dimensional representation called an "embedding" or "vector embedding", derived from another model, word2vec being a classic example

The fasttext library/tool is a great easy way to do text classification using word2vec-style embeddings

[desert oar]

they aren't defined here plt.plot plots pairs of x and y points. they happen to be connected with a line and they happen to lie on a straight line because that's how they were defined. so it looks like a smooth line, but it's actually individual "dots" connected by line segments. is this homework?

[pearl blaze]

no , i m learning it

[pearl blaze]

i want gap on y axis of 10 instead of 20 , how can i do that ?

[desert oar]

normally we define lines by their y intercept, not their x intercept. you need to do some algebra: find the y intercept such that y=0 when x=10

y = ax + b and you have some known constant a. plug in y=0 then solve for b, and then you can find b as a function of your desired x intercept

[pearl blaze]

i think that is x intercept

[desert oar]

corret

[pearl blaze]

then ?

[desert oar]

then what? then you modify how the line is constructed in the code, after you know a and b

[pearl blaze]

in my code where is a and y?

[dire island]

Good day, please does anyone know what to do?

[final cobalt]

Where is a good place to host a big dataset of images? (Around 50Gb)

[scarlet anchor]

I am tying to use Google translate API in my ML project. but its paid, are there any better alternatives?

https://cloud.google.com/translate/docs/reference/rest/?apix=true

[upbeat prism]

Task: Given a list of 10 numbers from 1-20, is the first token repeated once?

I want to solve it using a basic Transformer. I'm using PyTorch's transformer and I overfit it so I basically get 100% accuracy. I then look at the gradients using model.transformer.layers[0].self_attn.register_forward_hook(print_qkv_gradients)

but I get:

(Pdb) (query == key).all()
tensor(True)
(Pdb) (value == key).all()
tensor(True)

So the input gradients for query, key, value for all 10 "tokens" are the same. How can that be?

[serene scaffold]

For who?
Amazon S3 might be an option

[main fox]

When I created my datasets of images for a project previously, I had decided I wanted to use transfer learning and resized all the images to the size expected by the model. This greatly reduced the size of the dataset and allowed me to store it locally. Your experience may vary of course, but just something to consider.

[final cobalt]

Its more that I think it might be useful for others. Its a database of pokemon cards. I can resize them appropriately

Doesn't Kaggle and HuggingFace have dataset hosting?

[desert oar]

S3 or equivalent

[unkempt apex]

upload on kaggle, if you got high speed network

[winged turret]

Does anyone know how to web scrape a table off of kaggle? Been trying to parse out these html elements but can't figure out how to extract it properly.

[winged turret]

Yea I'm aware that I could just download it, but I'm doing a portfolio project where I want to show I know how to web scrape from a proper dataset as such as this: https://www.kaggle.com/datasets/muhammadroshaanriaz/e-commerce-trends-a-guide-to-leveraging-dataset. Been sorting through different classes and <div> elements, but no luck having the info for my dataframe.

[winged turret]

If anyone could help me solve my problem, I'd appreciate it!

[left tartan]

You're trying to scrape from the CSV data table that's shown on the page?

[winged turret]

yes

at least one of them

[left tartan]

That's going to be fairly hard, since that's dynamically loaded as you scroll through

Not a trivial project

[winged turret]

I see, how would you recommend doing so?

[left tartan]

Well, I wouldn't since there's APIs and other ways to retrieve the data. But, if you needed to scrape it, you'd have to spend some time inspecting the network view in dev tools and see what requests/responses are made and reproduce that.

[worn mountain]

does anyone know matlab's im2gray equivalent in python? I used pillow and I got different mean intensity values, for instance max mean intensity I got from python is 188.5 while matlab gave 189.4
same picture, same double precision

[worn mountain]

if i take mean intensities
python gives 79.7905864197531
matlab gives 79.685138421249505

[rich moth]

ok its ready to rock if anyone is interested in trying it. you can process a dataset from hugging face using the args and switch the models around depending on your resources. you'll need an elasticsearch server up and running though.
https://paste.pythondiscord.com/YTIA

[final cobalt]

Sick

Gimme an hour or two

And I'll have over 18000 pokemon card images downloaded AND categorized by series, set, type, subtype, and energy color - with associated tags if needed

[final cobalt]

I've almost got these cards sorted - but I've got a question

Should I sort them by color? The goal is to build segmentation masks which separate the frames from the art. Would breaking them apart by color make it easier or harder for the system to find what's frame and what isn't?

[tepid field]

Cool

[untold fable]

does sklearn use in industries

i mean does chat gpt use that

[jaunty helm]

probably no

chatgpt uses neural nets, and sklearn isn't focused on neural nets

it's more likely that nn focused libraries like tensorflow or pytorch are used
but because open ai's models are closed we won't know for sure

[odd meteor]

Depends on the company and your JD. So yeah, sklearn is used in industry.

[left tartan]

Separate from your question: if you're asking is 'sklearn' worth learning? The answer is yes, imo. It's a basic building block that anyone working with data should be familiar with. (And I'm curious if anyone disagrees)

[untold fable]

Like talking about the best example of machine learning like YouTube recommendations page

Does they use sklearn

[untold fable]

Or sklearn is just for basic understanding

[wooden sail]

sklearn more deals with "classical learning" instead of "deep learning"

that means it mostly uses classical optimization algorithms instead of deep neural networks (though it does have some neural network capabilities)

not every problem requires deep learning

[serene scaffold]

banned

[tawdry monolith]

Can someone suggest me free resource to learn math needed for ml

[serene scaffold]

check the pins

[main fox]

Anything you'd recommend for modern NLP?
Parts of speech tagging, NER, Q&A, summarization, etc

[serene scaffold]

no one cares about part-of-speech tagging anymore. POS tagging has only ever been a means to an end, and there are now better means to those ends that don't require it.

pretty much everything since 2022 has been about applying interactive LLMs in different ways.

[main fox]

Noted, thanks
Would you say the focus on LLMs is well placed, and if so do you have resources you'd recommend for developing and deploying LLM apps?

[serene scaffold]

if you're just developing and deploying them, you're not actually doing NLP. that's just pure software engineering, at that point.

procuring enough GPU compute would be an issue.

[main fox]

I wonder if my question is... misguided.
If someone asks how to learn more about ML, people typically recommend ISLR or Hands on ML by Geron.
These cover theory and applications.

I'm looking for NLP equivalents, in that vein.

[serene scaffold]

@main fox how well do you already understand concepts like neural networks, train/test, and regression

[main fox]

Decently well, can develop and deploy. Always more to learn of course.

[serene scaffold]

What do you mean by develop

Because once you get to deployment, the AI stage is over. That's only about software development

[main fox]

I understand what models to use to answer specific questions, limitations of the models, assumptions, why we do train/test, avoiding data leakage

[serene scaffold]

Okay, great

[main fox]

Were you gauging my understanding to offer a recommendation or for something else?

[serene scaffold]

I have a link to a YouTube series that I'll send you when I get home

I'm at the gym, not skipping leg day.

[main fox]

Thank you 🙂

Sounds good, enjoy your workout

[serene scaffold]

Enjoyment is for the weak
To progress is to suffer.

[main fox]

Destroy your body and recover stronger 💪 Get those machine legs that cause quakes with every step

[final cobalt]

Now that I've got my dataset

What is the standard approach to having a network learn on its own how to segment features from images?

[serene scaffold]

@main fox https://www.youtube.com/playlist?list=PLoROMvodv4rOSH4v6133s9LFPRHjEmbmJ

[serene scaffold]

are you sure you're asking how to segment features, or are you trying to segment entities? and you do know the difference between the two?

[main fox]

Thank you

[harsh sun]

So I want to take an existing model for image recognition with camera vision, and add another class onto it. 1. whats the best way to go about that? 2. I have a dataset but how must I format each of the images for the model? I put paper towel below the subject in order to provide a contrast but idk how else I have to format it.

[serene scaffold]

Does your dataset contain labeled instances of the classes that the model can already classify?

Does it have any unlabeled instances of those classes?

[harsh sun]

They are all unlabeled. However, theres only one new class.

And they are pertain to that single class.

I basically took a video of the subject, and extracted each frame to serve as training data.

[serene scaffold]

if you continue training it on only the new class, but you don't have any labeled instances of the other classes, you won't have a way to determine if its performance on those classes has degraded.

so I would get ahold of instances for every class that the model needs to be able to classify.

[harsh sun]

Well, I would use a pretrained model that has labels for the other classes. I would solely be adding a new one.

[serene scaffold]

I know. but training it on the new classes can degrade its performance on the old ones. (and it might not even learn the new one.)

also models do not have labels. instances do.

[harsh sun]

So I need to retrain the entire thing?

I might be messing up the terminology.

Instances are the training batches right?

[serene scaffold]

No, instances are the things that the model needs to be able to classify. images, in this case. a batch is just a batch of training instances.

but a batch is not an instance. a basket of eggs is not an egg.

you don't need to retrain from scratch, necessarily. but you still need labeled instances for all the classes that the model has to know.

the instances for the old classes might only go in the test partition of the data.

but it's probably a good idea to keep instances of the old classes in the training data, to reinforce its existing knowledge.

[harsh sun]

Hm I see. The goal of this project is to detect ID cards on people. So what if, instead, I use the pretrained model to detect people, and I have a separate model that solely detects the ID cards. Would that be a more accurate solution?

[serene scaffold]

that is, the model takes an image of a person who has an ID card (perhaps in their hand, perhaps displayed on their clothing), and the model indicates where the ID card is in the image?

[harsh sun]

Well wouldnt the second model just be trained on the id card itself (I put it against a white paper towel for contrast)? So then I can run the image and it will just check to see if it can see an ID card. The first model will confirm whether theres a person, and if both are there, then I get a true value.

[serene scaffold]

None of that goes without saying.

Do I understand correctly what you want your model to do, or do I not?

[harsh sun]

My message was reiterating the goal...

Because its not necessary as of now the specific location of the ID card

[serene scaffold]

you're not reiterating the goal, you're telling me about your implementation plan for the goal, but I'm still not certain I understand what the goal is.

[harsh sun]

The first model will confirm whether theres a person, and if both are there, then I get a true value.

That is the goal

[serene scaffold]

that's not a goal. that's an implementation plan.

what purpose do you ultimately need for this model to serve?

[harsh sun]

proof of concept

[serene scaffold]

proof of what concept?

[harsh sun]

checking if someone is wearing their ID

[serene scaffold]

okay, so the purpose of the model is, given an image of a person wearing an ID, to output whether the person is wearing their own ID? Or if they're wearing any ID?

[harsh sun]

Wearing any ID

I think it would be too difficult to detect whether its their ID

[serene scaffold]

okay, this is a concrete goal

[harsh sun]

Yes

[serene scaffold]

now the first model: what is it trained to do?

by the way, this isn't a classification task.

[harsh sun]

An arbitrary yolo pretrained model to detect a wide variety of classes.

[harsh sun]

Oh its not?

[serene scaffold]

at least, you're not adding an additional class to an existing classifier

I suppose "IS WEARING ID" and "IS NOT WEARING ID" is a binary classification task

but you're not going to get there by adding classes to an existing classifier

I'm not sure how you'd do this without a dataset of images, where the pixels of each ID badge in the image are indicated.

[harsh sun]

yeah I think maybe just a new model itself.

[harsh sun]

Wdym?

If I show it a ton of pictures of IDs, wouldnt it train the filters to detect patterns in the IDs themselves?

[serene scaffold]

Not really. Images where the only entity in the image is the ID are going to be very different from images where the ID is one of many entities

And it would be nontrivial to bridge that gap

[harsh sun]

Wait, I've never heard of this being an issue. In every guide I've ever seen for object detection, you train the model on images of the object itself and then it can detect it with camera vision.

https://www.youtube.com/watch?v=-ZyFYniGUsw AFAIK they just trained on the object

[iron basalt]

Scikit in general is heavily used for all kinds of random small script stuff by scientists and engineers.

Its goal is kind of to be a Matlab substitute.

[serene grail]

That's interesting, I didn't know that
I don't hear much about scikit on this channel, I suppose most people are interested in deep learning now
So scikit would have stuff like K Nearest Neighbors then? I'm very new to machine learning

[iron basalt]

https://scikit-learn.org/stable/index.html

https://scipy.org/

[serene grail]

Thank you!

[final cobalt]

I asked this earlier but then I went shopping

I've collected a huge database of pokemon cards and I've sorted them by frame type

What's the standard approach to teaching an AI to segment the frames of the cards?

An autoencoder with contrastive learning?

[blazing urchin]

Hello,
I am working on a tabular data set that contains various inputs ranging from time/date/time of day/day of week through meteorological inputs like temp and whether or not there was snow on the ground, to various other numerical inputs pertaining to speed of vehicles and their weight etc... what I want to do is to run various algorithms on that data and then have the output feed a probability model. I am doing this as a hobby project on RC car racing.... I am not a data scientist but I can write some code in Python...what sort of problems I may encounter with this approach?

[upbeat prism]

I wanna do a very basic classification task using transformer i.e. a transformer encoder. Does anyone know of a very simple model that is pretrained?

[desert oar]

That sounds like a pretty standard task to me. The only problems you will encounter are the usual problems with doing statistics and machine learning, which could (and do) fill a textbook

[upbeat prism]

I mean you could take your data, random "AI model" and just trial and error or you could take a more machine learning route first and do a lot of analysis on your data. Like which features correlate to which etc. Here's a nice walk through (completely different data but concepts you coul still do) https://www.kaggle.com/code/ilialar/california-housing-analysis-and-preciction

[desert oar]

Oh, I misunderstood... yeah definitely don't try a bunch of different models just to see what happens

[blazing urchin]

Thanks for that!

[blazing urchin]

Do you know which text book, or can you recommend one? LoL... I am looking for more indepth explanations than what I can find in youtube lectures.

[grand breach]

how do i scrape data from semrush or other seo for ctr prediction problem ? by how i mean which specific url should I look for the data that I want ?

[river cape]

Hi guys

I was working with vgg16 architecture and while experimenting , I got to know we can freeze the layers of the architecture

Now my question , as we know that vgg16 is a 16 layer CNN model (13 conv and 3 dense) , when I freeze the first 6 layers , which layers will be freezed , will onty be the conv layer or both conv and pooling layer

[sterile jacinth]

anyone is here

[serene scaffold]

Always ask your actual question. If you ask your question clearly, people will answer it when they look at the channel.

[final cobalt]

Alternatively - make a patently incorrect statement!

Nothing gets you correct information faster than steadfestedly claiming something false XD

[lapis sequoia]

pooling is not trainable

[river cape]

so the first 6 conv layers will be freezed?

[lapis sequoia]

yes, probably

most frameworks let you list or filter trainable layers as well.

[versed pilot]

I think SciPy is more of a matlab substitute, general purpose science and engineering. SKlearn is only for machine learning. If you want to do a t-test or do a Fourier transform etc. you go to SciPy, not sklearn.

[desert oar]

Numpy + Scipy + Matplotlib is the Matlab substitute. The "scikit" ecosystem as far as I know was intended to consist of collections high quality libraries on top of that, but only scikit-learn became a household name

There's also scikit-image

[iron basalt]

Yeah, Scikit, which includes all of these things, although they renamed some things, scipy is what it's all built on, which is a wrapper for a bunch of common (mostly) numeric algorithms.

There is also the scikit build system, which sometimes shows up in stuff like robotics and other random projects that are using it for some reason.

Numpy is the defacto standard library for transferring dense numeric data/arrays between libraries and so it shows up in all of these.

Python's built in array type could be used, but Numpy just also has a bunch of other useful stuff so it's used instead.

And matplotlib for visuals.

(Although I don't really like Matplotlib, if you use it for basic things only, it's pretty much the easiest most direct way to plot things)

[trim oxide]

Hey all, I am working on a Project to solve PDEs with NNs and rn im trying out different initializations for a basic NN. However, it seems like all of PyTorchs inits are worse than just the plain default. Im using ADAM to optimize and Relu Squared as activation if that is relevent. I was just wondering if this is normal or if I may be initializing it wrong. Perhaps one might be better with a different combo of optimizer and activation?

[final cobalt]

I've got a question for y'all

What's a good way to monetize AI quickly and honestly

[buoyant vine]

Sell the shovels

[iron basalt]

You find a demand, and fulfill it.

[final cobalt]

https://tenor.com/view/otter-shock-shocked-otter-utterly-shocking-huh-gif-26697050

[agile cobalt]

what did you expect?

if it was that simple, we would just do it ourselves instead of giving away the idea for free in here lol

I guess that you can try building a RAG system using internal documents

[umbral charm]

So i have this very messy graph

without drawing the lines between each point we get this

So my question is how do i just plot the outline so therefore it looks like just a bimodial distribtion?

like this!

[final cobalt]

There's nothing wrong with polling for ideas from peers 😛

Besides - a lot of people have tangential knowledge they'd rather not apply themselves for whatever reason, but might be germane to the question

[severe hare]

There are half a dozen of repoes just like this one:
https://github.com/Marker-Inc-Korea/AutoRAG

Who wants to help me write a NN that can beat anyone in Magick the Gathering?

Nobody?

[agile cobalt]

many organizations that could benefit from it don't have the know-how to set it up

[final cobalt]

This is a very, very substantial task

You'll need a rules engine - that's a tall order in itself, even if you disregard the graphics

Then you'll need some way to teach it about a deck's strategy - mill vs self mill vs token spam vs voltron, etc

Those strategies will need to degrade into medium and short term goals, and then it'll need to relate card effects to those gaosl

AND it'll have to balance pursueing it's own goals against sabotaging the opponent's board state - except in cases where it can instead take advantage of the opponent's board state

You don't need one AI, you need many

[faint quail]

now this is epic (red is true bounding box and green is prediected)

[final cobalt]

https://tenor.com/view/we-bare-bears-ice-bear-pick-up-lift-up-gif-6549807532845206879

Ice Bear

BTW, are you building an auto targetter? Cool

[untold fable]

now i learn pipeling in sklearn

[untold fable]

which algorithm you are using

@faint quail would you share the code

[final cobalt]

Twas a good day sorting pokemon cards

[scenic parcel]

It turns out ml takes time for even the best hardware. Doing hyperparameter optimization and after 2 hours only 60 trials done. Maybe the 5090 coming out soon could be worth, but idk maybe dual 3090's would be better

[wooden sail]

in fairness, the 4090 is not the "best hardware" if you're doing AI

you'd wanna be able to parallelize the hyperparam tuning over several tasks running either on a gpu with a huge amount of vmem like an A100 or better, or over several gpus like a 4090 or better (ideally several A100's)

[limpid oak]

Hello, I have photographs which are collected from users where they are expected to capture it using mobile camera, but some of them them upload photo from print or other display device eg. taking photo from computer scrin.
How can we approach such issue, please guide

[scenic parcel]

Ok so vmem is the spec to go for, not vram? Because yeah as you can see Im not even using all of my memory

[wooden sail]

i meant vram, my bad

24 gigs is better than other consumer graphics cards, but not a huge amount either

the a100 series from a few years ago has 80 gigs of vram, and there should be newer models out

[iron basalt]

https://docs.nvidia.com/dgx/dgx-station-a100-user-guide/intro-to-station-a100.html

(4 interconnected A100s)

(320 GB VRAM)

[wooden sail]

i guess the h100 series uses the more recent architecture, but still has 80 gigs of vram. you usually find pools of several a100s or h100s in a compute cluster, like what squiggle showed

iirc the one at my uni has 12 a100s in the compute cluster

yeah there seem to be 2 of these bad bois https://www.gigabyte.com/Enterprise/GPU-Server/G492-ZD0-rev-100

[scenic parcel]

They let you use it if you ask?

[wooden sail]

at least here, you fill out a form and get a signature from a professor to verify you actually need to use it

then you're given an account and can submit jobs as you like using a scheduler like slurm

[iron basalt]

An interesting perspective is that the fastest super computer in the year 2000 cost $110 million (and was very large), and the Nvidia 3090 cost about $1500, but can do 3x more operations per second.

[scenic parcel]

It seems like I'm not maxing my vram the same way I do when I run an llm though, so I wonder if I just need more tflops or something

[iron basalt]

It's best to run a profiler to check.

(Nsight is part of the CUDA toolkit)

[scenic parcel]

Though I think moore's law died due to quantum tunneling so now our components just increase in size and heat instead

[iron basalt]

The next phase will most likely require different architectures and maybe a different medium (than silicon). And different algorithms more suited for those architectures (probably not backpropagation).

We will also have an in-between state, where you see stuff more like NPU that is being shoved into things. Where there are components with special purposes / architectures that the CPU cores direct to do whatever (like how the cores on a FPGA just orchestrate things).

[iron basalt]

The types of algorithms being used in practice are heavily driven by the available hardware design, and that is a feedback loop since they design for the popular software trends.

(Stuck in local minimum)

[serene grail]

That's actually very cool

[thorny lintel]

hey guys, anyone knows how to access gated hf models using tokens using inference APIs ?

[wooden sail]

i misread that as wanting to automatically generate tokens and requests maliciously, oops

[strong notch]

Is someone looking for a project on AI in healthcare? If yes, please DM me.

[thorny lintel]

oops!

gracias

[thorny lintel]

hi, I already have the access to the model, but the error comes like this:
already tried making multiple accs and diff access tokens

[upbeat prism]

So I run some PyTorch code that at some point dispatches aten.tanh_backward.default. Now I'm wondering what exactly that implements. Does someone know how I can figure it out? Do I just go an read the source? Does anyone know what the .default means?

Also is there something like a convention that _backward might be the derivative for the autodiff engine or something?

[untold fable]

Machine Learning A-Z: AI, Python & R + ChatGPT Prize [2024] hows this course

[serene scaffold]

you don't need to learn R.
I'm immediately suspicious of any intro machine learning course that mentions ChatGPT.

[untold fable]

i mean should it worth it

[serene scaffold]

if you have to pay for it, no.

[jaunty helm]

I agree with Stelercus, the course name containing ChatGPT does not give me confidence about its quality

[untold fable]

https://www.udemy.com/course/machinelearning/?couponCode=ST15MT100124A#reviews

this is link of my corse

[unkempt apex]

lol you can learn that for free

just search about it and read

[woven shadow]

Is anyone here very familiar with autograd?

i.e. implementation level familiarity?

How does torch.autograd perform df(x)/dx for extremely simple functions, like f(x) = e^x? Or f(x) = ln(x), which cannot be easily computed without using finite differential techniques?

Or is it acceptable to use such techniques with these very basic functions?

[wooden sail]

the derivatives of basic functions are stored/hardcoded

[woven shadow]

by basic functions, where is the line drawn?

[wooden sail]

you'd have to check the source code for pytorch's autograd to know exactly

[woven shadow]

Also, for univariate functions like I mentioned, how exactly are they stored on the DAG?

i.e. to exemplify?

if you have something like this expression: a + b in the DAG, the gradient stored is 1

For a/b, the gradient is 1/b

and for a * b the gradient is b

All these are wrt to a

[wooden sail]

with respect to a, sure

[woven shadow]

For univariate functions, how is it done?

[wooden sail]

exactly as you did now

[woven shadow]

for something like ln(a), how would I compute the grad?

[wooden sail]

you store ahead of time that the derivative of the ln is dx/x

same as you did just now with addition and multiplication. for "simple" operations, you hardcode the derivative

[woven shadow]

yeah, I understand

[wooden sail]

you trace the graph in the opposite direction and check the derivative of each node as you go along

[woven shadow]

But for illustration purposes, can you show me how it would be done wrt to a?
the classical derivative is 1/x, how would I express this wrt a?

[woven shadow]

i.e. would the grad of ln(a) be 1/a, for example?

[wooden sail]

yes

[woven shadow]

or would for sin(a) be cos(a)

[wooden sail]

just let x = a. the letter makes no difference

[woven shadow]

Ah thank you so much
I really appreciate this

[fallow coyote]

any good books or tutorials in how to use matplotlib and seaborn?

[rich moth]

i had a cool idea for a model i wanted to build. I call it Tabtransformer, pretty lame, I know. But it combines tabular data and texxt data . I'm using bert to handle the text transformer block for categorical features in the tabular data. training is off to a good start Starting self-supervised pre-training with BYOL... Pretraining Epoch 2/10: 9%|███████████▋ | 340/3750 [01:15<12:24, 4.58batch/s, loss=4.52e-6]

[faint quail]

Its trained using the yolov2 architecture and algorithm, but the optimizer is adam, all the code is written from scratch, using cupy and occasionally tensorflow. You probably wouldn't want the code since it'd be faster just to use PyTorch and train it like that

[desert oar]

Matplotlib has a lot of good docs and tutorials now on its official docs site

[iron basalt]

Enough until you can make more or less anything you want via composition with them.

[untold fable]

no i want that code if you wont mind

[faint quail]

it's a neural network so there's not much code to actually share besides the code for training it, which again I made from scratch using my own custom library and it'd be faster to just train your own model using yolov5 or yolov8 using pytorch

[faint quail]

brother what behemoth of a model are you training

[untold fable]

you use tensorflow to train it

and now you want to use pytorch to make it faster

[faint quail]

it's my own custom library but yes I use some tensorflow functions for the convolutional operations

heres an XoR solver using it (not including other files)

https://paste.pythondiscord.com/EWZA

[merry basalt]

gang

I am using numpy, and why can't it broadcast together (500,1) with (1,5000)???

I am simply adding the two vectors together

broadcasting rules should cover (a,1) + (1,b), so WTF is happening?

[wooden sail]

looks like you have 500 and not 500,1

[merry basalt]

ah yeah

thanks for the help

[finite thicket]

Traceback (most recent call last):
  File "D:\Projects\sync\get-dissed\get-dissed-prototyping\pixtral_test.py", line 17, in <module>
    llm = LLM(model = model_name, tokenizer_mode="mistral", trust_remote_code=True)
          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "D:\Projects\sync\get-dissed\get-dissed-prototyping\.venv\Lib\site-packages\vllm\entrypoints\llm.py", line 214, in __init__
    self.llm_engine = LLMEngine.from_engine_args(
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "D:\Projects\sync\get-dissed\get-dissed-prototyping\.venv\Lib\site-packages\vllm\engine\llm_engine.py", line 564, in from_engine_args
    engine = cls(
             ^^^^
  File "D:\Projects\sync\get-dissed\get-dissed-prototyping\.venv\Lib\site-packages\vllm\engine\llm_engine.py", line 325, in __init__
    self.model_executor = executor_class(
                          ^^^^^^^^^^^^^^^
  File "D:\Projects\sync\get-dissed\get-dissed-prototyping\.venv\Lib\site-packages\vllm\executor\executor_base.py", line 47, in __init__
    self._init_executor()
  File "D:\Projects\sync\get-dissed\get-dissed-prototyping\.venv\Lib\site-packages\vllm\executor\gpu_executor.py", line 38, in _init_executor
    self.driver_worker = self._create_worker()
                         ^^^^^^^^^^^^^^^^^^^^^
  File "D:\Projects\sync\get-dissed\get-dissed-prototyping\.venv\Lib\site-packages\vllm\executor\gpu_executor.py", line 105, in _create_worker
    return create_worker(**self._get_create_worker_kwargs(
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "D:\Projects\sync\get-dissed\get-dissed-prototyping\.venv\Lib\site-packages\vllm\executor\gpu_executor.py", line 24, in create_worker
    wrapper.init_worker(**kwargs)
  File "D:\Projects\sync\get-dissed\get-dissed-prototyping\.venv\Lib\site-packages\vllm\worker\worker_base.py", line 449, in init_worker
    self.worker = worker_class(*args, **kwargs)
                  ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "D:\Projects\sync\get-dissed\get-dissed-prototyping\.venv\Lib\site-packages\vllm\worker\worker.py", line 99, in __init__
    self.model_runner: GPUModelRunnerBase = ModelRunnerClass(
                                            ^^^^^^^^^^^^^^^^^
  File "D:\Projects\sync\get-dissed\get-dissed-prototyping\.venv\Lib\site-packages\vllm\worker\model_runner.py", line 977, in __init__
    self.attn_backend = get_attn_backend(
                        ^^^^^^^^^^^^^^^^^
  File "D:\Projects\sync\get-dissed\get-dissed-prototyping\.venv\Lib\site-packages\vllm\attention\selector.py", line 117, in get_attn_backend
    from vllm.attention.backends.xformers import (  # noqa: F401
  File "D:\Projects\sync\get-dissed\get-dissed-prototyping\.venv\Lib\site-packages\vllm\attention\backends\xformers.py", line 6, in <module>
    from xformers import ops as xops
ModuleNotFoundError: No module named 'xformers'

anyone know why this is happening? getting this error when trying to initialize vllm.LLM. i just ran pip install vllm mistral_common in my venv like normal.

and to confirm, i've ran nvidia-smi in my terminal and can confirm i have CUDA 12.5

[finite thicket]

ping me if you respond pls, thanks

[woven shadow]

Is this mistral_common you've installed a new or old version?

It's not a CUDA problem, these missing module problems are more likely to be a version issue, your version might be too old or too new @finite thicket

[finite thicket]

theyre freshly installed

just created a new venv and ran pip install vllm mistral_common

[unkempt apex]

https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/5993

[livid gate]

Is anyone here using Kedro by any chance?
I'd be really interested in your experiences with it and whether you'd recommend it

[lapis sequoia]

https://www.reddit.com/r/MachineLearning/comments/gmwfmf/discussion_are_metaphors_a_form_of_transfer/

[desert oar]

That's the most Reddit thread title imaginable

[scenic parcel]

Data is 2774 rows by 70 rows, trying the TFT model from darts

[finite thicket]

This talks about command line arguments, doesn’t apply to my python program

[unkempt wigeon]

I have a question has anyone tried making a neural network that acts similarly to them mannerisms like talking if people have already made a neural network that could talk or ones that have a habit of trying to make sure that everything's 100% perfect etc my apologies

[left tartan]

Have you used GPT?

[unkempt wigeon]

Yeah I have I'm curious because if you give up data on you like a video file will it pick up your specific characteristics like how you speak your mannerisms do you do something when you hear specific sound would it pick up on that in the video recording and if it were to generate and video have its own it probably would have the same reaction similar to the person

[unkempt wigeon]

Yes

[left tartan]

GPT incorporates many mannerisms, as you'll see from its responses, in an attempt to make it seem more "human".

[unkempt wigeon]

The reason why I'm interested is because well they can learn from a lot but if you give it a video on just one person like audio diaries could it learn to act like that person and if it learns some more data would it become more like the person or when it have its own personality similar to the person but not quite

[unkempt wigeon]

I know but that's programmed in I want to know if anyone who's made something similar to chat gbt too and came at a video on a specific person didn't pick up any it's the person's mannerisms how they talk etc

[faint quail]

but why

[scenic parcel]

45 of the rows are positive, negative, and neutral binary values for news articles written about 15 different companies

If I wanted to do embeddings instead, which seems like it'll be better, I'd probably need a lot more columns

[faint quail]

are u like analyzing news articles for predicting stocks?

[scenic parcel]

Yup

Never done this before though so results are terrible. So I'm considering embeddings

[faint quail]

dang thats a nice idea

good luck

[dark slate]

Has anyone worked with PyTorch?

It's just I am not sure I am using the autograd method properly

[rich moth]

Alright I made another version of the vqvae . It should be able to handle visual, text and tabular so Im gonna start training it on this dataset I found for Pokémon. https://github.com/HybridShivam/Pokemon?tab=readme-ov-file

[serene scaffold]

Hello, many of us have used pytorch. Try being more specific.

[dark slate]

I am trying to use the torch optimizers instead of specifying the gradients

However, I am not sure if I am doing it correct

I can try showing how I did it

If I am allowed to show it on this chat

[serene scaffold]

you can show code in any chat on this server.

!code

[arctic wedge]

Formatting code on Discord

Here's how to format Python code on Discord:

```py
print('Hello world!')
```

These are backticks, not quotes. Check this out if you can't find the backtick key.

For long code samples, you can use our pastebin.

[dark slate]

So, normally I have code with regular gradients on python of such:

a = 0.4
# Gradient
def sf(x, y):
    return (a**2)*2*x + a*y, 2*y + a*x
eta = 0.5
# Number of steps
n_iter = 4

r = 1.  # we will plot the function over x, y in [-r, r]

# Define starting point in the upper right corner of plot
xi = 0.7*r  
yi = 0.4*r
p_x = [xi] 
p_y = [yi] 

for _ in range(n_iter):
    dx, dy = df(xi, yi)  # computing the gradient
    xi -= eta * dx  
    yi -= eta * dy  
    p_x.append(xi) 
    p_y.append(yi)  

And the way I wrote using torch:

xi = 0.9 * r  
yi = 0.8 * r


x = torch.tensor(xi, requires_grad=True)
y = torch.tensor(yi, requires_grad=True)


p_x = [x.item()]
p_y = [y.item()]


for i in range(n_iter):

    z = f(x, y)
    
 
    z.backward()
    
  
    with torch.no_grad():
        x -= eta * x.grad
        y -= eta * y.grad
    

    p_x.append(x.item())
    p_y.append(y.item())
    
    # Zero the gradients for the next iteration
    x.grad.zero_()
    y.grad.zero_()

Is this a fair approach?

[serene scaffold]

you usually pick the optimizer you want to use and just step that optimizer.

[dark slate]

which means?

[serene scaffold]

the optimizer is an object that adjusts the weights of a model according to a certain procedure and in terms of a certain loss

https://pytorch.org/docs/stable/optim.html

[dark slate]

is it in relation to:

for i in range(n_iter):
    # Compute the function value
    z = f(x, y)
    

    z.backward()
    
    # Updating x and y using the gradients
    with torch.no_grad():
        x -= eta * x.grad
        y -= eta * y.grad
    

    p_x.append(x.item())
    p_y.append(y.item())
    
    # Zero the gradients for the next iteration
    x.grad.zero_()
    y.grad.zero_()

instead of optimizing this way manually

I should use torch.optim ??

[unkempt wigeon]

Is a histogram optional?

[serene scaffold]

There isn't some universal context in which histograms are used, so you'll need to be more specific

[unkempt wigeon]

Numerical histogram

https://www.w3schools.com/python/python_ml_data_distribution.asp

[unkempt wigeon]

Is this ok?

[unkempt wigeon]

Data distribution

[rich moth]

😂 Im using the csv file from kaggle for the text and tabular data and images nicely correspond to it.

I wanted todo it for MTG but the mtgdata I found wouldnt download

[unkempt wigeon]

If I was trying to use a girl Network make one do I truly need to have any type of histogram?

[final cobalt]

I wanna game something out with y'all

I want to build a neural style transfer tool - and in the long run, a pose and character transfer tool also

I want to try building a multi-encoder VAE and train one encoder to recognize style using a contrastive approach - feed in pairs of images by the same artist or from the same cartoon, have the system compare latents

The problem is that if one encoder is recognizing style, the other encoder would need to be capturing content and only content. Somehow I need to enforce disentanglement of the two systems. CGTP, my main sounding board and trust rubber ducky, says the only true way to enforce disentanglement is to give each a loss function designed to make it capture what I want it to capture

But I want the "content" encoder to pick up on everything the style encoder doesn't pick up on

Is there any kind of loss term you can think of to say "hey, your encoding is capturing information already captured by this other guy?"

[white tartan]

I am pursuing bachelors in Mathematics and computing can i make future in AI development with this domain?

[final cobalt]

In theory, you can pursue a future in AI dev without any education at all

Homebrew, baby

That said, you'll need two to four terms of calculus, one term a minimum of linear algebra, and some statistics

As well as a strong foundation in computing principles

[dark slate]

Btw, does anyone know how normalization works on data? Because I was curious whether it would affect the classification performance on regression models such as Logistic Regression?

Or is that dependent on its penalty?

[serene scaffold]

"normalization" can mean different things and the kind of normalization you should do depends on the data and the model

[white tartan]

My branch have advance mathematics and basic computing

like 60% maths 40% cs

[serene scaffold]

people without degrees almost never succeed in getting AI-related jobs.

[final cobalt]

I'm sure you're right - but for the sake of diligence, might I ask in what way you're qualified to make that statement?

Just need to be clear you're not just armchair pontificating

[dark slate]

Input normalization: Centering alone will not solve the problem. So, a solution is to ensure that all input variables have the same scale. One measure of scale is the standard deviation. Input normalization transforms the inputs so that each input variable has unit standard deviation

[serene scaffold]

I work in AI professionally in the US. I do not know anyone who has a job in AI who does not have at least a bachelors. Most of my coworkers have at least a masters, and many have PhDs.
My department also insta-turbo-rejects any applicant who does not have at least a bachelors.

[dark slate]

That’s the kind of normalization I’m talking about

[white tartan]

so can you tell is bachelors in mathematics and computing will support me someway?

[serene scaffold]

if you take AI-related courses as part of that degree, that would be ideal. you'd have a much easier time if you got an AI-focused masters after completing that first degree.

[final cobalt]

Hehe, so, super qualified

Well lemme ask you then, because I'd like to get into the AI biz myself

[white tartan]

yeah i am also starting learning AI sideway making projects and i am also planning to do masters in AI after bachelors as i could do that after mathematics and computing domain

[final cobalt]

I'm most of the way through my comp sci bachelor - I've got the calc and linear algebra out of the way, and my coding skills are...

Let's just say "enough," at least compared to my classmates

What else do I need?

[white tartan]

The problem was that i was not able to get AI DS bachelors as my rank was 30k out of 1.5 million students and for AI DS i needed 15k rank🥲

[serene scaffold]

it will be easier for me to answer your questions if you ask the whole question in one message.
if you want to get an AI job with only a bachelors, you need to max out all the AI-related experience you can get from that degree. take every AI course that undergrads can take, get involved in a lab. things like that. if you can't do all of that, you should expect to need to get a masters.

[final cobalt]

Well, I want a masters any way

I'll ask a more thorough question later - I gotta head off to that place

By which I mean school

[serene scaffold]

Stay in school, don't do drugs, and hail satan.

[white tartan]

can you accept my friend request in case i lose you

[final cobalt]

Be gay, do crimes

[serene scaffold]

I'm here all the time.

[white tartan]

Fine

[serene scaffold]

No. Being gay is good. doing crimes is not.

[quaint mulch]

if I already max out all the degree I can get, what's next?

[serene scaffold]

if you've done all the ML-related stuff that you can do during a bachelors degree, and you're done with that degree or almost done, you should start applying to jobs

[quaint mulch]

So I guess, it is just straight back to the numbers game, but hopefully, with better odds this time?

[serene scaffold]

if you're applying and it's not going well, you might need to apply to grad schools

[quaint mulch]

I already maxed out all the degrees, including grad schools

[serene scaffold]

so you have a masters degree? in what?

[quaint mulch]

my master was in teaching, but my PhD is in CS

[serene scaffold]

how much AI-related stuff did you do during your phd?

[quaint mulch]

I think I'm talking about industry jobs in particular, not academia

[quaint mulch]

100%

[serene scaffold]

okay, so you should have options. if you feel like your job hunt isn't going well, go to #career-advice and show your anonymized resume and describe your job hunting strategy.

[odd meteor]

My friend's company is hiring at the moment. Your profile looks like a good match. You can give it a shot, who knows something positive might come out of it.

https://www.arcee.ai/careers

[fallen gate]

oh cool!

could you send over your thesis?

I am curious

[unkempt wigeon]

Does anyone understand now I'm trying to ask about needing to visualize the data using a graph or could I just not use a graph

The website that I'm using for machine learning for you ask and they are using something like this to work on the model to show data being normal and showing the data random or in between two numbers sorry

https://www.w3schools.com/python/python_ml_data_distribution.asp

[left tartan]

What is your question?

[unkempt wigeon]

Do I need to use a graph in general for a network to work?

[final cobalt]

Just polling again in case anyone has any bright ideas. Is there a way to force one neural network (encoder) to only capture features not captured by a second encoder in a generalizable way? The only method I know of is mutual information loss, with is massively expensive (computationally)

[left tartan]

I'm not sure I understand what you mean. Your first question was about "visualizing the data using a graph" but you're asking about "for a network to work". What do you mean by "for a network to work" and what does that have to do with visualizing the data?

[unkempt wigeon]

Do I have to have that line of code added in to network for it to work I know it's a stupid question I'm just curious my apologies

[final cobalt]

Anyone here ever use Mutual Information loss?

[odd meteor]

I've used it a few times

[tulip wyvern]

I’m having a lot of trouble. I want to use lightgbm to predict the name of a Pokémon given other features but lightgbm can’t predict categorical variables. But I also don’t want to ordinally encode the names of the Pokémon because there shouldn’t be an ordinal relationship between the Pokémon. What should I do??

[desert oar]

Who says it can't?

[charred egret]

https://lightgbm.readthedocs.io/en/latest/pythonapi/lightgbm.LGBMClassifier.html pretty sure all you have to do is encode your classes. If you have categorical features, you might need to specify the columns

[tulip wyvern]

Does this also work for the prediction feature being categorical?

[tulip wyvern]

When I tried to put in my prediction column as y it said it can’t read strings but then when I tried to put in my y as one hot encoded columns it said it couldn’t take the dimension

I tried googling the issue but I got nothing so I know I’m definitely just missing some key element

[charred egret]

Yeah that’s why it’s a classifier. All you need to do is label encode the classes. You can experiment and see.

[tulip wyvern]

Is there a better way to encode other than ordinal if the values have no ordered relation ?

[charred egret]

It won’t treat them as ordinal

0 to whatever number

[tulip wyvern]

Wait really??

[charred egret]

afaik

They have a special class for the classifier so yeah it shouldn’t treat it as ordinal. I don’t see why they would code it that way.

[tulip wyvern]

But like if I labeled charizard as 1, venasaur as 2, and pikafhu as 3, wouldn’t pikachu be closer to venasaur than it is to charizard ?

[tulip wyvern]

Oh shoot

[tulip wyvern]

https://scikit-learn.org/dev/modules/generated/sklearn.preprocessing.LabelEncoder.html

this?

Even though it just labels it 0, 1, … n (number of classes) it won’t make an ordered relationship where 2 is closer to 1 than it is to zero?

[charred egret]

Yeah that’d be an easy way. But if it was me i would use the pokemon’s pokedex number

[tulip wyvern]

would that solve my issue of the labels having a false relationship ?

Sorry I’m just really tripping out about that one small thing

[charred egret]

Try it. The classifier class should be aware that it’s not ordinal.

[tulip wyvern]

It can do that??

[charred egret]

if you’re using regressor then maybe it’s a concern

Try it and evaluate your model

If you look at their code examples they all use 0 to n-1 for classifier

[tulip wyvern]

Okay thank you I didn’t know these models could decipher that the relationship isn’t ordinal

That’s actually really helpful tysm I’m also prob gonna use pokedex number that was good idea

[rich moth]

I found the images also correspond the the numbers in the dataset , which is nice.

[tulip wyvern]

I’m sorry I don’t understand

[rich moth]

This dataset I found aligns with the images in the order of the numbers. https://www.kaggle.com/datasets/jaidalmotra/pokemon-dataset?select=Pokemon.csv

[desert oar]

The "classifier" model ignores order, the numbers are arbitrary

But the "regressor" model will interpret numbers as numbers with an ordinal relationship

[tulip wyvern]

O Im not using that dataset i downloaded my data from Pokémon showdown API

[tulip wyvern]

Omg dude that actually makes so much more sense 🤦‍♀️🤦‍♀️

Is this also the case for the features?

Do I have to one hot encoded the features as well? I assume so right

Like the input features

[desert oar]

No, completely unrelated. Classification is about the label only

lightgbm specifically has a way to state which features are categorical via its Dataset API, I don't know if it's doable through the sklearn interface

[charred egret]

You don’t have to one hot encode it but you have to specify the features that are categorical https://lightgbm.readthedocs.io/en/latest/Advanced-Topics.html#categorical-feature-support

[desert oar]

One hot encoding in a tree based model can be questionable

[tulip wyvern]

Oh dang so it automatically encodes it for me

[tulip wyvern]

Ya that’s sick tysm

[tulip wyvern]

Tree based model?

Sorry I’m new to all this

[desert oar]

Yes, you should read how lightgbm works

[charred egret]

If you’re using logistic regression then yeah you should probably one hot encode but lightgbm is a more complex model that can learn more complex shapes so it’s better in that regard

[desert oar]

Sort of. You can't put in strings from what I remember -- they need to be numbers, but you tell lightgbm that the numbers represent categories

[tulip wyvern]

Okay I will read about gradient boosting cuz I think lightgbm is just a more optimized version

[tulip wyvern]

Okay thank you very much

[final cobalt]

https://drive.google.com/drive/folders/1xFEbAzhbMc4WrrkG5BI_pJRmeZqCJaNy?usp=sharing

All magic card images

        # uid : a randomly generated 12-digit hex string
        # src : the actual location of the image on the internet
        # via : the webpage on the website which hosts the image
        # hid : the card's name
        # utf : the datetime the image was downloaded
        # bid : the id of the 'batch' of images downloaded in a burst
        # license : Rights Reserved WotC
        # attribution : the artist's name
        # category : N/A
        # subcategory : N/A
        # filename : '{uid}.png'
        # description : N/A
        # tags : N/A

        cursor.execute('DROP TABLE IF EXISTS Images')
        cursor.execute('''CREATE TABLE IF NOT EXISTS Images (
                uid         TEXT PRIMARY KEY,
                src         TEXT NOT NULL,
                via         TEXT NOT NULL,
                hid         TEXT NOT NULL,
                bid         TEXT NOT NULL,
                utf         TEXT NOT NULL,
                license     TEXT NOT NULL,
                attribution TEXT NOT NULL,
                category    TEXT NOT NULL,
                subcategory TEXT NOT NULL,
                filename    TEXT NOT NULL,
                description TEXT NOT NULL,
                tags        TEXT NOT NULL
            )
        ''')

[past meteor]

what would you do instead?

[unkempt wigeon]

May I ask a question I'm specific ways in a way could work like a 2d array for images of 3D array for three dimensional objects etc is there a paper that goes into detail?

[desert oar]

Integers & use the lightgbm option to declare which features are categorical

[past meteor]

Yeah, agreed. Catboost has the same option

If you forget setting them as categorical with lightgbm or catboost it becomes ugly though

[rich moth]

Tricky look at this dataset. It's 5 gigs of magic cards in the h5 format. This structure fits right into my model. https://osf.io/kq9js/

[final cobalt]

Cool! How many cards?

Is it just card art, or fan art also?

[velvet oak]

I wanted to learn machine learning, could someone please guide me towards some best resources to learn ML from?

[unkempt apex]

okay logical question now

I am trying to train VAE
but I have 5 input images for corresponding output image
didn't get that? my bad

2 directories -> sketch and photo
sketch contains images which will get trained model and
photo is like labelled images from which loss will get calculated

but for one photo image I have 5 sketch images

so can I train 5 train images for 1 photo images on VAE?

[unkempt wigeon]

May I ask her thought experiment

What's the possibility of a neural network quote unquote ingesting a virus and then the virus gets transcribed and then the neural network becomes viral in nature

[serene scaffold]

0

[unkempt wigeon]

I'm sorry

[unkempt wigeon]

But how can you be sure I'm at work only knows what date is coming in you can't differentiate certain types of data groups what if it infects an image and then parts of the code or it was generated to go into images and change specific parts which gives the neural network instructions as it's going through training data it imparts some of the qualities of both the virus and the network itself becoming an adoptable virus

[serene scaffold]

when you learn more about neural networks, you will understand that a neural network in itself cannot turn into executable, self-replicating code.

[unkempt wigeon]

You never know it's just meant mathematics and the same formulation Annette goes awry I can be reproduced to go right again if it seems the right mathematical formula puts it into a viral like cell it's own computer virus to to find a way of infecting other neural networks in essence around Network becomes viral itself and watching neural networks is more of natural science you can never be 100% confident in what they can really learn

[quaint mulch]

https://github.com/aprbw/ArianDLPrimer

[quaint mulch]

Why not just do it 5 times?

[unkempt apex]

nah, it will be too hectic

[unkempt wigeon]

If you have an optimizer and 3 Nerons that have the learned pattern it could reprogram via a built in optimizer

[charred egret]

Neural networks don’t execute code at least not on their own. To them it’s just data they will use to optimise to whatever the targets are. NNs aren’t some machine that suddenly knows how to execute x86 on the host machine just because it ingested bad data

Now if the Deep Learning framework you’re using has some CVEs then that’s a different issue

[unkempt wigeon]

What I mean is the virus somehow gets onto your terminal finds any active material and checks to see if it might be a neural network and if it isn't go on to the next file then the next file and then the next file until it finds a neural network and it would be a library consisting of its own data and maybe two you're on the race based off of that data which would be 'injected' into three neurons changing the neurons to learn oh that data that we thought was bad okay it's good even if it's a deep learning Network it can change many different things depending time what process it's going through

[charred egret]

Do you know what neurons contain? Just numbers. At worst you’d have a model that’s bad at predicting

It’d be doing matrix multiplication on wrong numbers

[unkempt wigeon]

Still enough bad inputs can turn into severely worse consequences for such a network

[charred egret]

It really doesn’t work like that. That’d be saying there’s a certain combination of a matrix multiplication that’d somehow mess up your computer. Or even

float64 * float64

that’d mess up your device

[unkempt wigeon]

Would it be possible for a virus to recognize it's is about to learn and replace the data files my apologies

[final cobalt]

Lemme run something past y'all

I need to make a diffusion model to outfill pokemon card art - to expand the art. To do this, I need 512x512 images (target size) which mimic the art style I'm after. But the only images in the right art style are the cards themselves - and the art from those cards is too small.

Could I train a diffusion model on 256x256 images taken from the cards and then use tiled diffusion to create high resolution, non-artificially upscaled 512x512 images?

And then train the outfilling model on those?

[rich moth]

I got the MTG data loaded but the training times, sheeesh Training Epoch 1/20: 1%|▏ | 12/1539 [09:50<22:27:24, 52.94s/it, Loss=2.1, PSNR=9.69, SSIM=0.142]

[final cobalt]

I was meaning to ask - how many cards are in that dataset you showed me?

[rich moth]

Its a good question. I haven't been able to figure that out. It says default cards, so I need to see maybe what that entails.

[final cobalt]

What are your needs?

Do you just need the images? Or the card text also?

[rich moth]

im using both right, i wanted to experiment see how they can relate both types .

[final cobalt]

relate pictures to card text/types?

Just thinking out loud, but I feel like your best bet would be to scrape wither https://scryfall.com/ or https://gatherer.wizards.com/Pages/Default.aspx for the actual cards and then just crop images from any cards whose frames make it easy. With multi-processing and and asynchronous requests you could probably download the entire database in a few hours

And have the cropping done in about as much time

[rich moth]

Good idea, ill check it out. let ya know

[unkempt wigeon]

What is the best use to make a 5 dimensional array for possible looking at a different timeline

[left tartan]

Wdym? How to make a 5 dimensional array?

[unkempt wigeon]

What I mean is making up five dimensional space for a neural network to understand concepts that the human brain can't through deep learning and simulations

[left tartan]

So you're trying to understand how to create a nn?

[unkempt wigeon]

Yes

[left tartan]

Maybe start with. https://iamtrask.github.io/2015/07/12/basic-python-network/

[unkempt wigeon]

Thank you

[rich moth]

You guys see Nvidia new open source multimodal? https://huggingface.co/nvidia/NVLM-D-72B
I found a quantized version but you still need 48 gigs of VRAM.

[final cobalt]

Hey smart people!

Here's my plan: I need a diffusion model capable of outfilling the card art from pokemon cards - I want to expand them from approximately 500x300 to 512x512. To make this model I need art in an appropriate style that's 512x512 or so, right? The only art I have, though, is the card art itself.

Two options - the first being just upscaling extracts the from the card art

The second, much more complicated but much more fun, if to train a smaller diffusion model on 256x256 random crops from the card art and used tiled diffusion to create 512x512 images in similar styles.

Both options will inevitably introduct artifacts into the dataset. One is simple, one is complex but also a "diffusion model with training wheels" and thus good practice

Thoughts?

[untold fable]

where to learn statics for data science

[lapis sequoia]

Helu can anyone provide the resource for nlp?

[rich moth]

I scarped the content again and the size is a little over 8 gigs, its 3 gigs bigger than the other dataset i found I also cropped the images. Want me to put it on my google drive?

[final cobalt]

I appreciate the offer, but when I need to I'll just scrape scryfall myself

That way I can be sure

[rich moth]

generate_converted_dataset(
    out_img_type='border_crop',
    out_bulk_type='default_cards',
    save_root='./data/converted/',
    out_obs_size=(224, 160),
    convert_speed_test=True,
)```

That made it really easy.