#data-science-and-ml

1 messages ยท Page 126 of 1

river cape
#

I still didnt get it . What happens if the decison boundary goes through (0,0)

past meteor
#

idk it's hard to explain because it's quite intuitive if you don't overthink it ๐Ÿ˜…

#

still, what if all inputs are 0

#

what will you propagate?

river cape
past meteor
#

it'll output 0 and especially if you take regression and not classification examples theres 10000000s of examples where an input of 0 does not mean an output of 0

#

Say you pick relu

wooden sail
#

what you propagate is the gradient of the loss evaluated at the point, and some activation funcs do give 0 if the input is 0. i'm not sure either of those are good arguments to why a bias is used

#

probably better to just think of it as getting extra degrees of freedom. one way of reinterpretting the bias is to embed the original input in one dimension higher and then doing something like a rigid transformation or a shear in N+1 dimensions

#

it's the simplest way of getting a nonlinear effect in N dimensional space, too

#

those aside, the choice of threshold is the simplest explanation

past meteor
#

The non-math answer of "if your input is zero, should your output be zero" is also fine imo

#

Like, what's the mean net worth of a baby

wooden sail
#

should it though? classifier networks never do that. if your output is to be interpreted as probabilities of a categorical distribution, you don't get 0 as an output for an input of 0s

past meteor
#

stuff like that

#

If you have a softmax you can say "should each class be equiprobable for an input of 0"?

#

I mean you're definitely right though

wooden sail
#

but the question was "why" from the beginning

spring field
#

more > less ๐Ÿ˜

wooden sail
#

that's what we're discussing ๐Ÿ˜› "what" it is that it's doing

#

zestar just gave a good example where you might not want to get a uniform distribution out of a uniform input, but a linear transformation can't do that

#

you can turn Ax + b into a linear transformation if you like too, but only in a dimension of N+1 or higher

river cape
#

Could we say that bias is more of a starting point for our model?

spring field
#

not really no, it's usually initialized as 0

wooden sail
#

in some sense, sure

past meteor
#

I'd say yes

wooden sail
#

you remember all those + C's you get in differential equations and integral problems?

#

you find that + C from your initial and boundary conditions, or other constraints

#

same thing here

past meteor
#

Imo I'd say yes but it's also a wrong way to look at it

#

You have to stare to regression coefficients for this one

#

Like, econometrics style

#

then this stuff sinks in (from an applied pov at least)

river cape
#

Hmm i guess it's topic to research about

spring field
#

but you can make Mx do the same as Mx + b if you add another dimension to M

wooden sail
#

that'd be my preferred argument tbh

river cape
wooden sail
#

you can extend M into [M b] and x into [x 1]

past meteor
#

(that's also how affine transformations are implemented a lot)

wooden sail
#

that turns the affine transformation in N dimensions into a shear in N+1

past meteor
#

add a one

#

and multiply ๐Ÿš€

wooden sail
#

it doesn't, it addresses matiiss'

#

i'd imagine so if you use layers that don't map 0 to 0 or generally aren't monotone

#

simplest way of moving stuff around

unkempt apex
# 
import torch.nn as nn
import torch.nn.functional as F

class LaneDetectorCNN(nn.Module):

  def __init__(self):
    super(LaneDetectorCNN, self).__init__()

    # conv2d ( input, output, kernel, stride )
    self.conv1 = nn.Conv2d(3, 16, 3, 1)
    self.conv2 = nn.Conv2d(16, 32, 3, 1)
    self.conv3 = nn.Conv2d(32, 64, 3, 1)


    # so here output from conv3 which is 3D is getting converted into 1D
    self.fc1 = nn.Linear(64 * 14 * 14, 128)
    print(self.fc1.shape)
    self.fc2 = nn.Linear(128, 512)

  def forward(self, x):

    x = F.relu(self.conv1(x))
    x = F.max_pool2d(x, 2, 2)

    x = F.relu(self.conv2(x))
    x = F.max_pool2d(x, 2, 2)

    x = F.relu(self.conv3(x))
    x = F.max_pool2d(x, 2, 2)

    x = x.view(-1, 64*14*14)
    x = F.relu(self.fc1(x))

    x = self.fc2(x)

    return x

# -> for conv
# output_size = ( Input_size - Kernel_size + 2 * Padding / stride ) + 1
# -> for max pooling
# output_size = Input_size / stride


class Decoder(nn.Module):
    def __init__(self):
        super(Decoder, self).__init__()

        # Transposed convolutional layers for upsampling
        self.tconv1 = nn.ConvTranspose2d(64, 32, 3, stride=1)
        self.tconv2 = nn.ConvTranspose2d(132, 16, 3, stride=1)
        self.tconv3 = nn.ConvTranspose2d(16, 3, 3, stride=1)

    def forward(self, x):
        x = F.relu(self.tconv1(x))
        x = F.relu(self.tconv2(x))
        x = self.tconv3(x)

        return x
#

how it is now? correct or not?

#

but one thing to be noticed here is,

after feature extraction part in encoder!!, those features are passed to fully connected layers, so we have to flattened the input first then!!

#

@final kiln

I was thinking to download TuSimple dataset and then train my model, what do you think?

#

so it's fast process now!

rich moth
unkempt apex
#

shit!!

#

estimation is 16 hours by GPT

rich moth
#

Im using a 4090 its really slowing down around the 10 epoch. This is the first time Ive tried training anything

unkempt apex
#

the dataset TuSimple is 25 gb

#

so on 3060 it is approx.. 100 minutes

rich moth
#

I combine the power of VQVAE and CLIP for image reconstruction and aligment with textual descripts

#

Its really starting to take shape

#

Im going to add coming but it needs to run greyscale for a bit

#

Im training it on the flickr30k dataset

#

You aint buying it or what?

hollow escarp
#

Hi, I have object detection model which returns such resoult how can i extract bbox from such output shape?

image.png
small wedge
unkempt apex
#

how about above code?

# 
RuntimeError: shape '[1, 64, 2, 2]' is invalid for input of size 512```
#

okay sorry for that!!, I was just curious

unkempt apex
#

always follow experts!

#

come on let's focus on

#

problem now

#

I have created now decoder but

#

x_chw = encoder_output.view(64, 2, 2)

#

x_chw = encoder_output.reshape(1, 64, 2, 2)

#

I am not able to understand now this error

#

should we try -1 so that it will automatically do that stuff

#

again same error

#

ohhh

#

how can I download this paper?

#

what for this then?

final kiln
#

That's my improved error message

unkempt apex
#

so in encoder the output shape is 1, 512

#

so we convert this to do upsampling

#

hey it's vector

#

then we need to create tensor like matrix

#

where?

#

before decoder's conv

#

yeah

#

in decoder?

#

x = F.relu(self.fc1(x))

x = self.fc2(x)

first of of , why on second fc there is no relu function applied?

#

need to delete then

#

ahh, now it is getting worse, we are only swapping some codes not getting any output

#

we go from resnet to cnn

#

we need to focus now

#

ignore chat gpt, I am just asking some question to it

#

then how to reshape this

#

encoded output

#

also is there any need to add ReLU function in second fc??

#

do we need labelled data for that?

# 
    self.fc1 = nn.Linear(64 * 14 * 14, 128)
    self.fc2 = nn.Linear(128, 512)
```I need to revise this thing now
#

so from last layer of conv2d, we have 64 feature maps of size 3x3, with stride of 1
right?

so what is 64 * 14 * 14

#

what is 14 doing here?

# 
class Decoder(nn.Module):
    def __init__(self):
        super(Decoder, self).__init__()
        
        self.fc1 = nn.Linear(512, 128)
        self.fc2 = nn.Linear(128, 64 * 14 * 14)
        # Transposed convolutional layers for upsampling
        self.tconv1 = nn.ConvTranspose2d(64, 32, 3, stride=1)
        self.tconv2 = nn.ConvTranspose2d(132, 16, 3, stride=1)
        self.tconv3 = nn.ConvTranspose2d(16, 3, 3, stride=1)

    def forward(self, x):
        x = F.relu(self.tconv1(x))
        x = F.relu(self.tconv2(x))
        x = self.tconv3(x)

        return x
#

good?

#

yeah I am doing that now

#

x = x.view(-1, 641414)

what this line do?

#

converting 641414 into 1d plane?

#

I mean vector??

#

so how can I reverse that?

#

in encoder ( forward method)

# 


class Decoder(nn.Module):
    def __init__(self):
        super(Decoder, self).__init__()
        
        self.fc1 = nn.Linear(512, 128)
        self.fc2 = nn.Linear(128, 64 * 14 * 14)
        # Transposed convolutional layers for upsampling
        self.tconv1 = nn.ConvTranspose2d(64, 32, 3, stride=1)
        self.tconv2 = nn.ConvTranspose2d(132, 16, 3, stride=1)
        self.tconv3 = nn.ConvTranspose2d(16, 3, 3, stride=1)

    def forward(self, x):
        x = x.reshape(64, 14, 14)
        x = F.relu(self.fc1(x))
        x = F.relu(self.fc2(x))

        x = F.relu(self.tconv1(x))
        x = F.relu(self.tconv2(x))
        x = self.tconv3(x)

        return x
#

good?

#

yeah make sense now

#

run the whole thing and got same output

#

RuntimeError: shape '[64, 2, 2]' is invalid for input of size 512

#

torch.Size([1, 512])

from encoder's output

#

whole?

#

x_chw = encoder_output.reshape(64, 2, 2)

#

after getting encoder's output

#

heh?

#

lemme give yoou code first

give me paste bin

#

yeah now new error

#

RuntimeError: mat1 and mat2 shapes cannot be multiplied (256x2 and 512x128)

#

hmm lemme think

#

from where that matrix 256x2 came?

#

lemme give you full code then

#

yeah I am doing that

# 
encoder = LaneDetectorCNN()
encoder_output = encoder(input_image)
print(encoder_output.shape)

# so here we are converting simple 1d vector (output) into tensors
# so 32 channels tensors size of 4x4
# x_chw = encoder_output.reshape(64, 2, 2)

# now we will pass this to another t_conv to generate image


decoder = Decoder()
decoder_output = decoder(x_chw)
#

??, can't understand

#

shit typo

#

RuntimeError: Given transposed=1, weight of size [132, 16, 3, 3], expected input[1, 32, 16, 16] to have 132 channels, but got 32 channels instead

slender sentinel
#

Is Cs degree required to job in it or no ?

unkempt apex
unkempt apex
#

not for now!!

slender sentinel
#

I'm from buisness background should I get into it or no

unkempt apex
#

expected 132 channels but got 32 , whoaaa

unkempt apex
#

??

slender sentinel
#

Project management

#

To get into it can you tell me a roadmap

unkempt apex
unkempt apex
slender sentinel
#

to learn coding for eg which language should I focus on

hollow escarp
unkempt apex
hollow escarp
slender sentinel
#

Ok Bro

unkempt apex
#

hey @final kiln
GPT is still saying you don't need a fc layers in decoder if you are using transpose2d layer

small wedge
small wedge
arctic wedgeBOT
#
Rules

10. Do not copy and paste answers from ChatGPT or similar AI tools.

small wedge
#

It's not a reliable source of info, so much that it's banned to use it here for answering questions

unkempt apex
#

ahh, he is calling ambulance now!

hollow escarp
#

Thats the model which i used

unkempt apex
#

let's focus gentlemen!

#

[132, 16, 3, 3]
what is this?

hollow escarp
unkempt apex
#

3, 3 size of that matrix?

#

they are already I guess so?

# 
self.conv1 = nn.Conv2d(3, 16, 3, 1)
    self.conv2 = nn.Conv2d(16, 32, 3, 1)
    self.conv3 = nn.Conv2d(32, 64, 3, 1)


    # so here output from conv3 which is 3D is getting converted into 1D
    self.fc1 = nn.Linear(64 * 14 * 14, 128)
    self.fc2 = nn.Linear(128, 512)
#

this from encoder

hollow escarp
#

So i converted the model which had output shape of float32[1,5,8400] to model which has shape as on img below float32[1,1,80,80,65]

image.png image.png
unkempt apex
# 
        self.fc1 = nn.Linear(512, 128)
        self.fc2 = nn.Linear(128, 64 * 14 * 14)
        # Transposed convolutional layers for upsampling
        self.tconv1 = nn.ConvTranspose2d(64, 32, 3, stride=1)
        self.tconv2 = nn.ConvTranspose2d(132, 16, 3, stride=1)
        self.tconv3 = nn.ConvTranspose2d(16, 3, 3, stride=1)

and this from decoder

hollow escarp
#

I know it's a bit complicated but there are no other way to do it

hollow escarp
unkempt apex
#

congrats we created a CNN

#

torch.Size([1, 512])
torch.Size([3, 20, 20])

#

what is this now ,
3 channels, and 20x20

#

on what basis?

#

they are already 3x3

#

then which value? 5?

#

and on max_pool?

#

in forward?

#

it's 2

#

x = F.relu(self.conv1(x))
x = F.max_pool2d(x, 2, 2)

#

torch.Size([1, 512])
torch.Size([3, 26, 26])

#

this is not making sense to me

#

for reversing?
to convert feature_maps into a proper imag?

#

torch.Size([1, 512])
torch.Size([3, 71, 71])

good ?

#

torch.Size([1, 512])
torch.Size([3, 131, 131])

#

how much we want

#

128?

#

then we got 131

#

no I got 218

#

for kernel size of 39

#

let's keep that as it is

#

why now?
out input image was of 3,128,128

so why output for 131?

#

which loss?

#

to use?

#

cluster loss function?

#

hehe? where?

#

3 represents RGB?

#

why to 1?

#

ohh!! then it's what

#

ohhhhh

#

so how to be on that 1 shape

#

so it's fixed now!

#

can't we plot the output image??

then we will keep that loss and stuff

#

hey what about dataset?
do we need labelled dataset for this?
I am ready to manually label this ( only 10)

#

wait quick question

# 
import torch.nn as nn
import torch.nn.functional as F

class LaneDetectorCNN(nn.Module):

  def __init__(self):
    super(LaneDetectorCNN, self).__init__()

    # conv2d ( input, output, kernel, stride )
    self.conv1 = nn.Conv2d(3, 16, 3, 1)
    self.conv2 = nn.Conv2d(16, 32, 3, 1)
    self.conv3 = nn.Conv2d(32, 64, 3, 1)


    # so here output from conv3 which is 3D is getting converted into 1D
    self.fc1 = nn.Linear(64 * 14 * 14, 128)
    self.fc2 = nn.Linear(128, 512)

  def forward(self, x):

    x = F.relu(self.conv1(x))
    x = F.max_pool2d(x, 2, 2)

    x = F.relu(self.conv2(x))
    x = F.max_pool2d(x, 2, 2)

    x = F.relu(self.conv3(x))
    x = F.max_pool2d(x, 2, 2)

    x = x.view(-1, 64*14*14)
    x = F.relu(self.fc1(x))

    x = self.fc2(x)

    return x

# -> for conv
# output_size = ( Input_size - Kernel_size + 2 * Padding / stride ) + 1
# -> for max pooling
# output_size = Input_size / stride


class Decoder(nn.Module):
    def __init__(self):
        super(Decoder, self).__init__()
        
        self.fc1 = nn.Linear(512, 128)
        self.fc2 = nn.Linear(128, 64 * 14 * 14)
        # Transposed convolutional layers for upsampling
        self.tconv1 = nn.ConvTranspose2d(64, 32, 40, stride=1)
        self.tconv2 = nn.ConvTranspose2d(32, 16, 40, stride=1)
        self.tconv3 = nn.ConvTranspose2d(16, 3, 40, stride=1)

    def forward(self, x):
        x = F.relu(self.fc1(x))
        x = F.relu(self.fc2(x))
        x = x.reshape(64, 14, 14)

        x = F.relu(self.tconv1(x))
        x = F.relu(self.tconv2(x))
        x = self.tconv3(x)

        return x
#

should I apply that ReLU to that remaining layer?

# 
    x = self.fc2(x)

here we have not applied?

#

how many images should I label?

#

hey come on!

#

that was 25 gb

#

wait I found one with 262mb

#

what about comma2k19?

#

๐Ÿ˜‚

#

suggest fast dataset!!

#

they are in GB's

#

and I am still searching

#

it will take time to upload on gdrive then

#

or I should run the code on local?

#

okay comma2k19 then

#

just take a look at their labels!!

#

or we just need 2 lines for lanes?

#

EDA?

#

one dataset I found which has 2 seperate dir, one contains image and another it's labels

#

can I only download specific part of whole dataset?

#

I like curved lane dataset

#

how can I download this

#

okay got it

#

how can we combine labels and images??

#

how much will take to train??
it is total 40gb data

#

okay now going to sleep!!

rich moth
#

Wow after about 25 epochs its working well.

image.png
rich moth
#

Whoa...

reconstructions_epoch_1.png
#

Is this uncommon to get these kind of results from 30 epochs?

#

Dude this is wild.. it started with abstract shapes then refined that, all of sudden one epoch BOOM , now we have color.

#

What do you guys think?

rich moth
#

you are trying to call a integer instead of a function?

mental rampart
#

nvm....

mental rampart
#

im stupid

rich moth
#

What was it. No, you're not, I dont know you but using python probably doesnt make you stupid. Anyways, what was the issue, I wanna learn too ๐Ÿ˜›

mental rampart
#

wait my laptop giving me issues while im training the neural network
give a sec

rich moth
#

ok

rich moth
agile owl
#

so why does everyone scale standard deviations by the square root of time when they can convolve the probability mass function and actually do it right

image.png
#

I remember when I learned that convolving distributions yields their accumulation and I was like so I've been misled

rancid zealot
#

As a rule:

data science <=> doing things right

agile owl
#

so it seems, so it seems

bright sable
#

Could anyone recommend a library or something thats good for visualizing dataflow through the model? I would like to ideally see the gradients too (like an image where each pixel is a weight, alongside an image where each pixel is the derivative for the weights given the input and target output). Any suggestions? If nothing then anything that would be a good starting point to fork?

agile owl
#

isn't that what airflow is for

bright sable
#

Isnt airflow for distributing workload? Thats what I see in the pytorch docs anyway ยฏ_(ใƒ„)_/ยฏ

agile owl
#

It's a platform to "author schedule and monitor workflows"

#

monitoring involves visualization I'm pretty sure

bright sable
#

Well I can't find anything for it, just logs etc

agile owl
#

Datasets View sounds like what you were describing to me

bright sable
#

It does look like a general overview, but not really showing the data/weights/gradients as images

agile owl
#

What you described seems pretty intricate but I think you might be able to leverage the listeners to get that data and render visualizations yourself using matplotlib or something

rancid zealot
unkempt apex
rich moth
unkempt apex
#

Ohh my ....

#

How much GB of dataset?

rich moth
#

4.39 GB it looks like

unkempt apex
#

4 GB for 30k rows?

#

What is this ?? I couldn't understand

rich moth
unkempt apex
#

Is it relational database?

#

I mean I am not able to understand that 'rows' part?

rich moth
unkempt apex
#

Now make sense

spring field
rich moth
#

In includes other meta data, but thats all im using.

unkempt apex
#

Yeah I am on mobile , it lags on larger files

#

Then it will take 'days' to train on cpu

spring field
#

yeah, don't train on a CPU

unkempt apex
#

My frnd has 3060

rich moth
#

Heres where its at on the training currently.

image.png
spring field
#

what GPU do you have?

rich moth
#

msi 4090

spring field
#

wait wait wait, did you not split the dataset for training and evaluation?

rich moth
spring field
#

did you split it 50/50?

unkempt apex
#

Google colab only allow gpu for limited time

#

Do u know any alternative

spring field
#

paperspace

unkempt apex
#

Free??

spring field
#

no, but it's pretty cheap

unkempt apex
rich moth
#

Its not underfitting or verfitting and its obviously learning lol

spring field
#

cheapest paid GPUs are ~55 cents an hour
though occasionally there are free GPUs available as well
(if you pay for the 2nd tier subscription (1st/default is free) which is 8 bucks a month, then there are more GPUs to choose from, so more likely to have a free GPU available)
either way, it's not like you'd be constantly using them, you build the model, prep everything on the CPU and then when you're ready put it on cloud to train

spring field
rich moth
spring field
#

also may I suggest plotting those metrics

rich moth
rich moth
#

Im going bonkers trying to get that dataset split for train and test and validation? Does this look right?

image.png
#

Im confused why I cant use train or val only test.

spring field
rich moth
spring field
#

huh, did you specify train_size?

rich moth
#

so if change train to test, it works, but then i get this error.

#

KeyError 2

#

ok

spring field
#

did you read the docs for what split does?

#

I mean, it's optional in the first place

royal summit
#

Hey , I am a newbie in data scinece ,does any one knows where i have to begin in learning data science

rich moth
#

Ya it does look hella suss haha

spring field
#

it doesn't look sus... the package is simply using the interface that you have defined

#

it's basically a callback

spring field
rich moth
#

Hows it look fellas?

image.png
#

shit! i forgot the train size

#

Will that matter though?

#

Nevermind I read its calculated as a complement to test size, train size that is.

#

It definitely seems much more random now.

image.png
spring field
#

looks great

rich moth
#

Ok so I retreated to building a script try some brute force random hp search

#

I need a super computer , lol

dawn light
#

i created my own NN from scratch and it did just fine (on the MNIST data set)
One of the instructions from the guide asked us to take a look at what the neurons were learning (in particular the input to the first hidden layer), so i plotted the weights of each neuron of the hidden layer

The image from the guide is the one on the top, what i got is the one on the bottom
did i do this right? it seems like my NN's plot of its weights looks like whitenoise to me with a bit of patches

image.png
kind herald
#

Hello. SO i want to create an Ai that takes in a string of letters and outputs an output such as numbers with 2 decimal places. What might be a good idea to do this?

lapis sequoia
kind herald
unkempt apex
#

๐Ÿ˜‚ interesting

lapis sequoia
#

you dont need to create "AI" to make that

kind herald
#

What method

#

are you thinking about is what i mean

lapis sequoia
#

can u explain a little bit more what are you trying to create

kind herald
# lapis sequoia can u explain a little bit more what are you trying to create

I want to create soemthing that takes a scramble for a rubik's cube such as below:
" B' R D2 R' F2 R' F2 R B2 R2 D2 U2 R' D L D F R2 F2 D2 U' "
and outputs a time. that is within .5 of the time i actually got on the scramble a majority of the time. Like on that scramble i got 9.72. I want it to output something from 9.22-10.22

lapis sequoia
#

sorry bc my first language is not English but are you trying to predict the time it takes to solve a Rubik's Cube based on a scramble involves? bc im not understanding

unkempt apex
#

how to get kaggle api

#

I am searching that on profile

#

and still searching

unreal geyser
#

your text is not even a meaninful text to add embedding layer for RNN/LSTM , not simple numbers to feed regression models

#

might need to encode string letters to some numbers but then keeping the relation and meaning to data will be difficult , need to try a lot of techniques

twin relic
#

Hi, Can we do both SWE and data sci/Ml at the same time

#

I still at univ, I am not really sure about what path to really choose, tho we have to do a final sem thesis with data science/Ml .

agile cobalt
#

yes, there are a bunch of intersections like

  • model deployment (relatively high level and somewhat common)
    • integrating with existing applications, or creating new applications that benefit from them
    • to some degree scaling and overall maintenance
  • writing fast and efficient libraries for data science (low level, relatively few people working on it)
unreal geyser
#

depends on what do you mean by ML , if not research level , then it won't be hard to use Pretrained models via API or finetune them on your data,

#

no math needed unless data science or more like cleaning data and preparing proper data , but you have to decide where you want to focus more , ML is ever growing field with new good paper release on month levels

unkempt apex
#

btw I am downloading that TuSimple dataset 23 gb!!!

unkempt apex
urban canopy
#

For chat GPT programming I don't feel like I need the AI itself. If I could search the training data (which is has plenty of well-annotated examples) with an ordinary search function, it would be just as useful for me.

But the only way to use the training data is through the AI so I use the AI.

jaunty helm
#

and since the training data is a mixed bag (among other reasons), you really can't just trust what the AI gives you, always double check

urban canopy
#

I still couldn't 100% trust the training data either.

jaunty helm
# urban canopy You don't think that the search feature could narrow it down? I don't think the...

You don't think that the search feature could narrow it down?
not really? I can't imagine it at least, especially when you have like at least TBs of data right
just look at google nowadays

I don't think the training data has much in the way of horrible examples. The amount of hand-curation that goes into training these things is enormous
dunno how I'd check that honestly, but at the very least, the AIs these very curated datasets produce still often spit out crap to the point where we have rule10

#

it's also probably very unsustainable to just have this massive data stored all at once

urban canopy
# jaunty helm > You don't think that the search feature could narrow it down? not really? I ca...

I am curious how the AI narrows down training examples?

(thankfully I use the AI for the non-security-critical parts of my cloud codebase, and if it fails during testing no big deal ask it again).

If it is just as likely to give out crap, how is it more useful than searching the data and picking the best few matches to the keywords? If I could pay for API keys to query the data itself rather than download it (way to big and secret) that is OK.

jaunty helm
# urban canopy I am curious how the AI narrows down training examples? (thankfully I use the A...

I am curious how the AI narrows down training examples?
LLMs don't remember each specific example though, it just learns which words are more likely to come next given the previous N words, and right now it produces decent code with that mechanic

searching the data and picking the best few matches to the keywords?
cause that's not exactly easy? at what point does some data constitute a match with a keyword? be too lenient and you get too many results, be too strict and you get too few; there's also a lot of ways to say the same thing

urban canopy
#

"be too lenient and you get too many results, be too strict and you get too few".

A possible way around this is to set up a simple weight-vector and take the best-scoring matches.

"there's also a lot of ways to say the same thing"

This is one way I see LLM's as helping. I wonder how well word-vectors would work here?

It is still unclear to me that AI is much better than simpler techniques. Simply because I cannot access the training data without using AI. In this case "simpler" means that there is a small AI with few weights (word vectors, etc) that makes queries to the training data.

#

Rather than an AI trained on the whole thing.

jaunty helm
past meteor
#

I need to revive my blog ๐Ÿ˜ฉ

What content would you guys be more inclined to read? Longer, more informative and detailed or shorter, to the point but perhaps more shallow

#

Interested in the opinion of both experienced people and beginners

urban canopy
#

Any examples where the generalization helps you a meaningful amount?

jaunty helm
jaunty helm
urban canopy
#

Like AI art? Yes that would be harder to do with my "simple AI + database" idea.

If I wanted "simple AI" to make art, I would have to first:

  1. Query a thousand or so images from the training set that match my word vectors ("car", "blue", not "city", "shading mood", "non-cartoon style" etc). Weight images by how well they match (training data has captions).
  2. Splat pieces of images that match randomly (preserving origin + destination with some jitter). Splats may be a "multigrid" with splats at different scales?
  3. Develop a way to smooth the splats to varying extents.
  4. Develop a metric for how well the splats match each-other.
  5. Adjust smoothing as well as replace badly-fitted splats simulated annealing and/or gradient descent.

Sadly, I will never know how well this works until I can get a billion well-annotated images or so.

#

Maybe I will try this with mandelbrot set images that I can generate a few thousand pretty easily (but no prompt, just use the whol few thousand I generate).

twin relic
unkempt apex
#

and there is a catch
Artificial Super Intelligence

twin relic
unkempt apex
twin relic
urban canopy
#

There are some problems with big AI, most notably the loss of control. Python is a very complex tool but I feel still in control. Not so with GPT.

That would be why, if I can do something with a smaller tool just as easily I would prefer to do so.

Stuff like removing bias and modifying the behavior becomes easier.

unkempt apex
#

also you can think about going alongside with web 3

twin relic
twin relic
unkempt apex
#

then you should take a look at web 3

#

maybe, where do you live??

twin relic
#

I just want to be good enough for a job and get out of this country ๐Ÿ˜

unkempt apex
twin relic
twin relic
urban canopy
#

Loss of control does not mean "terminator takes over everything."

It means that I cannot understand the tool well enough to know when and how to use it in novel situation.

Like trying to modding minecraft from byte code instead of well-documented source code.

It makes the tool less useful for me, and limits how I can use it (still very nice to have examples calling APIs that I forget).

unkempt apex
#

for each type of ML

twin relic
unkempt apex
#

you take interviews?
or you go for interview?

#

are you doing any job currently

#

yeah it!

urban canopy
#

I think I can understand NIST image recognition networks fairly well...

It's more that I wish I wasn't forced to use the AI just to access the training data!

Use the generic tool for well-established use cases (such as interview prep). But be able to use other, simplier, tools when more control is needed.

#

I built/trained a network in tensorflow and studied the visual cortex which is vaugly similar. It is not that hard, no where near GPT. I could do it if I put my mind to it, so could you.

small wedge
#

You might have an intuition on what the models are doing for simple tasks like that, but the idea that you know it well enough to figure out what it will output (i.e. not make it a black box) would be revolutionary. You'd be able to preform the proverbial AI brain surgery by hand

rich moth
#

Should I narrow down the hyperparamter search zeroing in out these results?

image.png
urban canopy
#

It would be fun to try! For a simple NIST network with a ~thousand weights.

small wedge
#

Agreed, at least for enabling precise control and alignment

#

Plus if we could make models into more of a white box where we actually know what they are doing we can extract the basic formulas from models and tune out any useless or redundant connections

urban canopy
#

This sounds like a good idea. I really want to "poke" into NIST digit NNs and get some idea of the weights.

Sadly, DevOps is not as much "know things at a deep level" as much as it is "guess and check how to get cloud component A to talk to B when auth C is passed. And remember obscure configuration settings that make it work but GPT doesn't realize."

As DevOps changes to a paradigm of workflow optimization (which will resemble highly automated bio labs in terms of skills needed), a shift back to deeper understanding may happen because more time is available for the core algorithms (and more algorythims will drive the automated workflows).

unreal geyser
#

but its not easy , at first to understand whats going on

#

if you already are familiar with backend then , try to get exposure of ML but stick to your strong point ,
backend devs actually get good salary

rich moth
#

So Im continuig with my hyperparameter search. I zeroed in on 256 dim and 768. Cya guys in a few hours,, gotta get some new tires, there goes a thousand bucks lol

unkempt apex
#

I waited 2 hours for that dataset to download

#

and now it failed on end point

#

shit happens!!

twin relic
#

I do enjoy backend stuffs , do you think it's best to learn python backend + Ml in that case ?

spring field
spring field
river cape
#

Hey guys

#

Lets say I have some non linear data

#

Like this

#

Now using a multi-layer perceptron , lets use two perceptrons.

#

How will those two perceptrons find the best line , because the data is non linear right?

iron basalt
river cape
iron basalt
#

(This does not actually decide how the weights are updated, just how some information (the gradient) is computed which can then be used to update the weights)

agile owl
#

I'd prefer the line

#

the variance is ridonculous

iron basalt
spring field
# river cape

I think you're forgetting about the sigmoid here, that's the non-linearity

agile owl
#

let's not miss the forest for the tree though sometimes the bias-variance tradeoff of neural nets and the time to tune it just loses to a line

#

(or logistic regression)

iron basalt
river cape
agile owl
#

the sigmoid is what makes you get something other than a line

#

the shape of the curve

#

it's a non-linear transformation

iron basalt
#

(You need some non-linearity, because linear with linear just gives you linear (or really in this case, affine))

#

(You can try this by modifying the desmos s(x) to identity (make it linear))

devout python
#

Hola people, I have been working quite extensively with pandas, but I run into one problem which frustrates me to no end, when I run a apply lambda function on a dataframe and it returns nothing, it doesnt return a dataframe no rows, but one that has shape 0,0. e.g.
df_output=df_input[output=df_input["col"].apply(lambda x: True if x in dict_test else False)]

#

if all rows are false the resulting dataframe is completely empty without columns

#

is there anyway around this?

serene scaffold
# 
df_output=df_input[output=df_input["col"].apply(lambda x: True if x in dict_test else False)]

This looks syntactically incorrect.
Also, what type is dict_test?

# 
df_output = df_input[df_input['col'].isin(dict_test)]

See if this works.

arctic wedgeBOT
#

:incoming_envelope: :ok_hand: applied timeout to @rigid void until <t:1718396032:f> (10 minutes) (reason: duplicates spam - sent 4 duplicate messages).

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

rich moth
#

Keeping the num embedding at 256 seems to be sweet spot, so does the embedding dim of 768 and 512. Which makes sense because they seem to align with the clip models embedding dim of 768 and its output if 512. Heres where I'm at so far, my score so far.

image.png
devout python
#

Dict is a standard dictionary

#

The code works but when empty its just annoying

rich moth
# river cape

i built an ensemble model that combines multiple LLMs to generate responses. It uses retrievers to fetch relevant information, processes the retrieved information to extract key insights, and generates responses using multiple LLMs guided by prompt nodes. A final selection step chooses the best response based on various criteria. The system utilizes Elasticsearch for storing conversation history and long-term memory

pipeline_diagram.png
#

The functionality is there , it just doesnt look very pretty. Heres a tiny output of the debug out, but when you ask it a query it will explore explore existing and new knowledge and dynamically in real time incorporate it that back into its though process . All this gets ranked and the top answered get saved back into its memory for other converations related to it.

Screenshot_2024-06-10_161139.png
#

I though instead of just training these things over and over again. Give them the ability and tools to learn new knowledge on its own accord, then save and reapply this new found knowledge. Heres an example how it searches and extract information from the web based on the query. Its more of a researchers tool then an conversational one.

Screenshot_2024-05-27_204532.png
#

Well not saved, that was a bad word, but encoded back into the elasticsearch index

serene scaffold
#

@rich moth nobody is going to read these big screenshots of text

rich moth
serene scaffold
# rich moth I dont blame you

What is your reason for posting them? It doesn't communicate anything other than "there exists on my computer a big blob of JSON data"

rich moth
serene scaffold
rich moth
unkempt apex
#

anyways!!!

rich moth
#

Its hard to find ways at articulating this stuff. I find pictures to easier sometimes, you know picture worth a 1000 words.

spring field
rich moth
#

Why not have open dialogue about it instead?

spring field
#

No, sorry, just couldn't resist the opportunity to make some light-hearted fun

rich moth
#

Its all good man, i could do better

spring field
#

Fair enough, the presentation could do with some actual visualizations or maybe just some simple formatting so that it's clear what's happening at least to some extent, yk, like at least pprint or something, that should immediately improve the presentation

serene scaffold
#

Pictures of text aren't worth more than the text they contain--they're just harder-to-read text.

rich moth
#

I trained and evaled it just like we talked about.

#

This is the first epoch. Something doesnt seem right

#

Seems good

spring field
#

"doesn't seem right"
"seems good"

so, which one is it? pg_rofl

rich moth
spring field
#

lol

rich moth
#

lol its overfitting after 10 epochs need to intergrate more datasets

#

I was hoping to automate this.

#

Can I share my current implemenation of doing this in python?

spring field
#

sure, why not

#

I found a tool for making these fun little diagrams

AE.png
agile owl
#

what is it

spring field
#

PowerPoint kekw

agile owl
#

powerpoint can be pretty good

serene scaffold
spring field
#

AE

agile cobalt
#

even in this channel it doesn't feels very obvious that AE means auto-encoder

spring field
#

well, tbf, there's the diagram for context ๐Ÿ˜„

#

(which Stel asked what it was, so I guess that's not helpful anyway...)

#

mmm, they generally seem more rotated when looking them up on the internet it seems pithink

The-structure-of-a-deep-autoencoder-with-encoder-decoder-and-the-code-bottleneck-in.png
spring field
frigid cove
#

How do I apply EarlyStopping from Pytorch Ignite into my Vision Transformer model?

frigid cove
#

Can I use EarlyStopping from Pytorch Ignite with the trainer from the transformers library?

past meteor
odd meteor
frigid cove
#

I'll have to look for the pytorch lighitng earlystopping. I have to train a ViT

past meteor
#

Coding early stopping each time is one of the reasons I moved to lightning

#

There's no reason not to use lightning (unless you're a beginner, then I think writing it out is good)

frigid cove
#

Oh jesus

#

Welp I'm a beginner x,d I still have to implement the ViT but I'll use a pretrained model from HuggingFace

odd meteor
dawn light
#

what does it mean if i got a training curve that looks like this (for ANN), did my model overfit or underfit or whatever?

image.png
past meteor
#

If this is some exercise you need to get through ASAP or you're just exploring, use lightning

odd meteor
frigid cove
#

The training function is implemented from scratch, but there's no early stopping there as the training epoch are very small

#

My dataset consists of 100K+ images in the training
I'll also use another dataset but with distinct labels later but my group is still dividing it

dawn light
#

this is for an ANN btw that's coded from scratch, so it's just an exercise

frigid cove
#

Could you give me some advice on how to implement it? I'm also thinking of the amount of training epochs I have to do since the dataset is quite big

#

With a batch_size of 32 the max training epochs would be around 3406 which is quite an insane number and based on what I've read it could make the model overfitt

frigid cove
odd meteor
dawn light
dawn light
frigid cove
odd meteor
frigid cove
#

Could you give further insight in how to apply EarlyStopping in the training function?

#

Like

#

Well I've read that you can use the loss

#

So if I give the loss from the nn Module could that be used as a parameter of the early stopping?

odd meteor
frigid cove
#

I'm sorry if I ask too much, I'm new too this and I'm a person who asks lots of questions

frigid cove
odd meteor
odd meteor
# frigid cove Could you give further insight in how to apply EarlyStopping in the training fun... 

import torch

class EarlyStopping:
    def __init__(self, patience=7, verbose=False, delta=0):
        self.patience = patience
        self.verbose = verbose
        self.counter = 0
        self.best_score = None
        self.early_stop = False
        self.val_loss_min = float('inf')
        self.delta = delta

    def __call__(self, val_loss, model):
        score = -val_loss

        if self.best_score is None:
            self.best_score = score
            self.save_checkpoint(val_loss, model)
        elif score < self.best_score + self.delta:
            self.counter += 1
            if self.verbose:
                print(f'EarlyStopping counter: {self.counter} out of {self.patience}')
            if self.counter >= self.patience:
                self.early_stop = True
        else:
            self.best_score = score
            self.save_checkpoint(val_loss, model)
            self.counter = 0

    def save_checkpoint(self, val_loss, model):
        '''Saves model when validation loss decrease.'''
        if self.verbose:
            print(f'Validation loss decreased ({self.val_loss_min:.6f} --> {val_loss:.6f}).  Saving model ...')
        torch.save(model.state_dict(), 'checkpoint.pt')
        self.val_loss_min = val_loss
#

Now, inside your training loop, you'd need to use an instance of this class to implement EarlyStopping

frigid cove
#

Thanks!

odd meteor
frigid cove
#

Mmmm I got a question, can I train my ViT with a 4060 in my laptop, given that the amount of images is 100K>?

spring field
past meteor
rich moth
frigid cove
rich moth
#

How do you know you've reached convergence?

image.png
spring field
#

just saying, but that's from google

nova matrix
#

Anyone familiar with NLP can guide me how I can extract stuff from a paragraph of text like Date, Time ,Total Cost

#

East Repair Inc.

1912 Harvest Lane
New York, NY 12210

BIN To Ship To
John Smith John Smith
2 Court Square 9787 Pineview Drive
New York, NY 12210 Cambridge, MA 12210
ary DESCRIPTION
1 Front and rear brake cables
2 New set of pedal arms

3 Labor Shes:

Terms & Conditions.
Paymentiis due within 15 days

Please make checks payable to: East Repair Inc.

RECEIPT

Receipt # us-001
Receipt Dato 11/02/2019
P.O# 2312/2019
Due Date 26/02/2019
UNIT PRICE AMOUNT
100.00 100.00
15.00 30.00
5.00 15.00
Subtolal 145.00
Sales Tax 6.25% 9.06
TOTAL $154.06

Smithy

#

for example from a paragraph like this

deep veldt
#

If i have two linear layers (or any) does it mean i have two hidden layers?

spring field
#

depends on if you want to consider the activations and various norms and regularisation as part of that layer or a separate layer

diagrams would likely just draw 2 hidden layers, I would consider there to be at least 4

odd meteor
hollow escarp
#

Hi i have following code which finds the closest boxes to camera:

def detect_closest_license_plate(
  session: onnxruntime.InferenceSession,
  image,
  img_width: int,
  img_height: int,
  logger: logging.Logger,
) -> ClosestPlate:
  predictions = get_predictions(session, image, img_width, img_height, logger)
  if len(predictions) == 0:
    return None

  camera_center = np.array([img_width // 2, img_height // 2])
  closest_plate = None
  closest_distance = float("inf")

  for license_plate in predictions:
    x1, y1, x2, y2, conf = license_plate
    plate_center = np.array([(x1 + x2) // 2, (y1 + y2) // 2])
    distance = np.linalg.norm(plate_center - camera_center)

    if distance < closest_distance:
      closest_plate = ClosestPlate.from_dict(
        {
          "bbox": (x1, y1, x2, y2),
          "confidence": conf,
          "plate_center": (plate_center[0], plate_center[1]),
          "distance_to_camera": distance,
        }
      )
      closest_distance = distance

  return closest_plate```

And im wondering how could i get distance from camera to plate_center
#

I want to make this function return result only if the object is some distance to my camera

#

Eg object crosed that line, so its like aproximently 2m from my camera

hollow escarp
proper crag
#

Is feature engineering is like mini model to raise the accuracy of the actual model?

serene scaffold
#

this is not a test. I just want to know what you know.

proper crag
serene scaffold
#

features are just properties of "things" that you can use for machine learning.
if you're trying to teach a model to take "a house" and predict its value, you would need features about each house. And the features would be the size of the house, and the number of rooms, and whether it's detatched or a town house, etc.

features are basically the columns in your table.

proper crag
#

Ty for the explanation... tried to watch few yt videos...most of em doesnt really explain ...like most of their explanation is kind of explanation which targeted to ML engineer that already know

serene scaffold
past meteor
#

@final kiln how was prefect? Would you use it for data pipelines or just orchestrating infra like you were doing?

#

I'm currently using Dagster and it's a very big PITA. I need to refactor it away. Airflow is an option or just good ol' cron

#

Observability is exactly what I want (or I'd use cron)

#

Rust for training models or Prefect in general?

#

I'd just use Airflow in the real world

#

it's a hobby project

#

most popular, robust solution in this space

#

I guess that's what I should do

#

Looking at the docs, this would be something I'd deploy myself

#

OSS products that also have a managed alternative typically have confusing docs

#

At least, ime

proper crag
serene scaffold
proper crag
past meteor
clever summit
#

Hello. I need help.

Source code: https://paste.pythondiscord.com/JICA
Error: ```py
Traceback (most recent call last):
File "d:\xampp\htdocs\dishub\AdminLTE-3.2.0\mesin_penghitung_2\sistem.py", line 27, in <module>
net = cv2.dnn.readNetFromDarknet(model_config,model_weights)
cv2.error: OpenCV(4.7.0) :-1: error: (-5:Bad argument) in function 'readNetFromDarknet'

Overload resolution failed:

  • Can't convert object to 'str' for 'cfgFile'
  • Can't parse 'bufferCfg'. Input argument doesn't provide sequence protocol
What does this error mean? What should I do?
agile owl
#

The rate of housing starts to the population in the United States continues to decline.

#

Never recovered to early 2000s levels after financial crisis

#

If that ratio continues to decline you would expect the price of houses to continue to go up for the same house in different time periods

hollow escarp
worthy sun
#

they provide the position of the plate center in the 2D image plane

#

not the actual 3D distance from the cam

worthy sun
#

if you have the camera's intrinsic parameters

#

I think he got offline

#

Not you, I'm talking about @hollow escarp

#

Like i said, it can be measurable directly from 2d coordinates

#

they provide the position of the plate center in the 2D image plane

#

not ACTUAL 3d distance

rich moth
#

After 25 epochs it has generalized on unseen data , pretty well. I want to introduce more datasets. But I ran out of disk space ๐Ÿ˜ฆ

reconstructions_epoch_4.png
spring field
#

yooo, that's great

rich moth
#

Not bad, huh? Thanks for your help

spring field
#

I think it's a VQ-VAE + CLIP

#

yes, that would be the core of it

#

I do wonder what CLIP brings to the VQ part though

#

unless it's a VAE + CLIP

#

^

rich moth
#

CLIP introduces multimodal aligment and semantic understand to the VQ-VAE model . Seems like its in enhancing the reconstruction process using text.

#

You were talking about VQ-VAE the other day which gave me an idea after I looked it up.

#

It's a multi-dimensional space where each point represents a possible combination of features from the two models.

clever summit
#

Hello. I need help.

Source code: https://paste.pythondiscord.com/JICA
Error: ```py
Traceback (most recent call last):
File "d:\xampp\htdocs\dishub\AdminLTE-3.2.0\mesin_penghitung_2\sistem.py", line 27, in <module>
net = cv2.dnn.readNetFromDarknet(model_config,model_weights)
cv2.error: OpenCV(4.7.0) :-1: error: (-5:Bad argument) in function 'readNetFromDarknet'

Overload resolution failed:

  • Can't convert object to 'str' for 'cfgFile'
  • Can't parse 'bufferCfg'. Input argument doesn't provide sequence protocol
What does this error mean? What should I do?
rich moth
#

yes

lapis sequoia
#

Hi,
Hope u are doing well,
I am working on a time series (wind speed)forecasting using CNN-LSTM-Attention model, the data is highly variable as shown in the figure attached, the model forecast generally follow the trend of data (red curve) but fails in details, i am using exogenous features too such as temperature, humidity and ect.
Please what can u suggest to improve the quality of forecasting ( I am getting a R2=0.6 and a MAPE=32%).

lmmmm.PNG
rich moth
#

thats a really good idea

#

keep em coming ๐Ÿ˜‚

#

Maybe your interested in working on it together then.

#

No worries . How is the jjob ssearch going?

#

Really? Im surprised. I know the tech industry has been having lots of layoffs. It wont be long till you land something

#

I'd hire you, if that makes you feel any better ๐Ÿ˜›

spring field
lapis sequoia
#

really?, i tried transformers too but my model performed it

#

if u have something to suggest so it's welcome

spring field
#

have you used transformers for timeseries?

#

maybe, but afaik, RNNs are more versatile

#

I don't remember :p

lapis sequoia
# 
    def __init__(self, attention_size):
        super(SelfAttention, self).__init__()
        self.attention_size = attention_size
        self.query_dense = layers.Dense(attention_size)
        self.key_dense = layers.Dense(attention_size)
        self.value_dense = layers.Dense(attention_size)
    
    def call(self, inputs):
        # Compute queries, keys, and values
        queries = self.query_dense(inputs)
        keys = self.key_dense(inputs)
        values = self.value_dense(inputs)
 
        attention_scores = tf.matmul(queries, keys, transpose_b=True) / tf.sqrt(tf.cast(self.attention_size, tf.float32))
        attention_weights = tf.nn.softmax(attention_scores, axis=-1)
        attention_output = tf.matmul(attention_weights, values)
        
        return attention_output```
# 
input_shape = (n_hours, features)
# Define input layer
inputs = Input(shape=input_shape)
inputs=BatchNormalization()(inputs)
# First residual block
conv1 = Conv1D(filters=64, kernel_size=4, padding="same",kernel_initializer="random_normal")(inputs)
conv1 = MaxPooling1D(pool_size=2)(conv1)
shortcut = Conv1D(filters=64, kernel_size=4, padding="same",kernel_initializer="random_normal")(inputs)
shortcut = MaxPooling1D(pool_size=2)(shortcut)
residual = Add()([shortcut, conv1])

#LSTMM1=LSTM(80, activation="tanh",return_sequences=True,kernel_initializer="random_normal")(residual)
lstm_out = layers.LSTM(80, return_sequences=True)(residual)
attention = SelfAttention(attention_size=64)(lstm_out)
flatten = layers.Flatten()(attention)
# Dense layers
LSTMM1=Dense(128, activation="linear",kernel_regularizer=l2(0.001))(flatten)
outputs = Dense(y_trainlstm.shape[1]*y_trainlstm.shape[2],activation="relu")(LSTMM1)
outputs = Reshape((y_trainlstm.shape[1],y_trainlstm.shape[2]))(outputs)
modelCNNLSTM = keras.Model(inputs =inputs,outputs= outputs)
print(modelCNNLSTM.summary())
modelCNNLSTM.compile(optimizer=Nadam(learning_rate=0.0001), loss='huber')
mc = ModelCheckpoint('best_modelCNNLSTM1.keras', monitor='val_loss', mode='min', save_best_only=True)
early_stopping = EarlyStopping(monitor='val_loss', patience=5, restore_best_weights=True)
reduce_lr = ReduceLROnPlateau(monitor='loss', factor = 0.9, patience = 4, min_lr = 1e-7, verbose = 1)
history1=modelCNNLSTM.fit(x_trainlstm, y_trainlstm, epochs=50 , validation_split=0.2,batch_size=64,shuffle=True,callbacks=[early_stopping,mc,reduce_lr])```
#

this is my model

spring field
#

would have to ask zestar about timeseries stuffs

lapis sequoia
#

i did and this my transformers model:

def transformer_encoder(inputs, head_size, num_heads, ff_dim, dropout=0):
    x = MultiHeadAttention(key_dim=head_size, num_heads=num_heads, dropout=dropout)(inputs, inputs)
    x = LayerNormalization(epsilon=1e-6)(x)
    x = layers.Dropout(dropout)(x)
    res = x + inputs
    
    x = layers.Conv1D(filters=ff_dim, kernel_size=1, activation="relu")(res)
    x = layers.Conv1D(filters=inputs.shape[-1], kernel_size=1)(x)
    x = LayerNormalization(epsilon=1e-6)(x)
    return x + res

# Define the Transformer model
input_seq = Input(shape=input_shape)
x = transformer_encoder(input_seq, head_size=64, num_heads=2, ff_dim=32, dropout=0.1)
x = transformer_encoder(x, head_size=64, num_heads=2, ff_dim=32, dropout=0.1)

# Flatten and Dense layers
flatten = Flatten()(x)
output = Dense(1)(flatten)
lapis sequoia
#

it's the encoder part of a transformer, i tried TFT model too but same results

#

but when i apply a smooth to my data i get good results as shown here

kllll.PNG
#

increasing the model capacity can not lead to an overfit knowing that i am using early stop which prevent overfit

#

this is my train /val errors graph

mpl.PNG
#

do u see that, like my model isn't complex ,comparing to a time series model ?

#

u meant adding a new conv layer at the beggining ?

#

but i am putting them at the beginning ?
so u meant that adding a third one ?

#

so in that case, my model will be LSTM-Attention model, is it your purpose ?

#

this one

nvbnx.PNG
#

before smoothing the target variable

#

but this lead for a weak metric results

lapis sequoia
#

i see but it's really improved thing

#

in reality, i did, i applied outliers detection, PCA, i add more features, i calculated the correlation matrix, i calculated the importance of features using random forest, i used sliding window technique to transform data into sequences and finally normalize data

#

i think yeah, but if i focus only on noise i will get worste results not like when i smooth data

rich moth
#

So I used that trained model I made to try and manipulate a jpg. It understood the color red and applied it, but not exactly what I wanted ๐Ÿ˜›

I only trained this thing for about 13 hrs on one dataset, I should have used multiple oness, but I need to recover some disk space from my wsl2 ubuntu instance.

image.png
lapis sequoia
#

In fact, this data represents wind energy data, and my goal is to predict the data 24 steps ahead that will help me schedule the energies. so maybe noise can be important too, but if i see that enabling noice poses a challenge i will smooth data ofc but i have to try with

errant bison
#

Which are some pf the good field for research in ai

serene scaffold
errant bison
#

Of*

serene scaffold
#

@errant bison generative AI is popular right now. not sure how it is as a research area.

#

though genai requires a lot of advanced knowledge.

frigid cove
#

Can anybody tell me why the logs appears to be like this:

#

But not like this: I've been following various tutorials on HuggingFace and Youtube

Screenshot_2024-06-15_154655.png
#

I'm using a local dataset that is inside the project folder in my computer

rich moth
#

Whoa. I got a great idea. I can augment the captions from the orginal dataset with a sentance transformer, capturing more of the semantic meaning by expanding its vocabulary and understanding of the scence, just slighty different, but same. You can describe and image in multiple ways. It seems like a good concept ?

frigid cove
#

This is the path of the dataset: train_dataset = Path("..\datasets\ASL_train")
test_dataset = Path("..\datasets\ASL_test")

#

I have no idea why the trainer looks so different in his video from mine

spring field
#

bread_pensive struggling with RL once again

gilded belfry
#

i got such a result when obtaining point cloud data from the depth image. is this the result i should get or is there a problem?

Screenshot_from_2024-06-15_23-46-05.png Screenshot_from_2024-06-15_23-46-21.png
agile owl
#

any opinions on DLinear

errant bison
rich moth
#

Ok I changed things around a bit in my vqvae model. It integrates sentence transformers, manifold learning to control how the images change and attention mechanisms to focus on the most important words.

#

Now its suppose to learn all this directly on datasets by measuring the vectors between the images and text in the trarining process. The autoencoder then trains to reconstruct those different vectors and in doing so learns how to captures the direction for image manipulation. The attention weights are also learning durning tthis process too, it figures out what captions are more relevent of the a given image and assings higher weights to them when aggregating the projected differences .

rich moth
#

now im stuck

image.png
#

oh its the embeddings of t he orginal data i think..

odd meteor
# errant bison Which are some pf the good field for research in ai

All research areas are pretty good. NLP has been getting much attention for a couple of years now especially with LLM & PEFT being one of its driving force.

If you wanna work on research areas that are not over hyped like LLM but equally good, then I think you should look into Reinforcement Learning, Federated Learning, Privacy-Preserving ML, AI Alignment, and Neuroscience.

odd meteor
rich moth
#

Its working! Wow the training time takes around 40 mins now, it use to be like 7 minutes

image.png
#

Im gonna need some serious hardware lol

rich moth
#

Thats from the first epoch.

image.png image.png
lapis sequoia
#

wind power data or also wind speed data is known by its high variability, all researchers in this feind said that, concerning the noise, we can not consider it as a noice in reality because it represents a real data( those picks are real one and we can't also consider them as outliers )

rich moth
#

Look how well its learning.

image.png
serene scaffold
#

Can you post text as actual text from now on?

rich moth
#

I apologize thanks for point that out. You got. Epoch [2/5], Train Loss: 0.0004, Val Loss: 0.0017, Train PSNR: 15.4822, Val PSNR: 16.1355, Train SSIM: 0.2424, Val SSIM: 0.2738

rich moth
#

I figure I'd update you guys a bit less often, im just really excited. But heres the results from the first run. It's important to note this version was without sentance transformers incorporated. Epoch [3/5], Train Loss: 0.0004, Val Loss: 0.0017, Train PSNR: 15.6918, Val PSNR: 15.6039, Train SSIM: 0.2612, Val SSIM: 0.2672 Training: 100%|โ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆ| 1551/1551 [34:24<00:00, 1.33s/it] Evaluation: 100%|โ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆ| 388/388 [03:58<00:00, 1.63it/s] Epoch [4/5], Train Loss: 0.0004, Val Loss: 0.0017, Train PSNR: 15.9803, Val PSNR: 16.6987, Train SSIM: 0.2814, Val SSIM: 0.3348 Training: 100%|โ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆ| 1551/1551 [35:23<00:00, 1.37s/it] Evaluation: 100%|โ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆ| 388/388 [03:42<00:00, 1.74it/s] Epoch [5/5], Train Loss: 0.0004, Val Loss: 0.0017, Train PSNR: 16.2615, Val PSNR: 16.8271, Train SSIM: 0.2983, Val SSIM: 0.3014

serene scaffold
rich moth
#

Honestly, thank you everyone for letting me share with you. I have a feeling im probably a bit annoying and being to honest here, but I dont have friends that are into stuff. And it

#

hard to find people to engage with I guess on my level. Im not quite as advanced as you guys obviously.

spring field
#

btw, you could add an accuracy metric as well, unless one of those abbreviations already is that

#

hmmm, does accuracy make sense for image reconstruction?

rich moth
#

I think you suggested it to me..

spring field
#

no yeah, fair enough, accuracy is probably gonna be pretty low and thus not a particularly great metric, yeah

rich moth
#

I trained it another two epochs, but I need to take a break, appreciate you people. Heres my results from that Training: 100%|โ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆ| 1551/1551 [34:05<00:00, 1.32s/it] Evaluation: 100%|โ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆ| 388/388 [03:55<00:00, 1.65it/s] Epoch [1/2], Train Loss: 0.0004, Val Loss: 0.0017, Train PSNR: 16.4321, Val PSNR: 17.1083, Train SSIM: 0.3143, Val SSIM: 0.3549 Training: 100%|โ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆ| 1551/1551 [35:45<00:00, 1.38s/it] Evaluation: 100%|โ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆ| 388/388 [03:51<00:00, 1.68it/s] Epoch [2/2], Train Loss: 0.0004, Val Loss: 0.0016, Train PSNR: 16.9217, Val PSNR: 17.3436, Train SSIM: 0.3533, Val SSIM: 0.3787
Talk to you guys later

high agate
#

Here's the code I used to created the confusion matrix:

cm = confusion_matrix(y_true = test_labels, y_pred=y_preds)
plt.figure(figsize=(15,15))
sns.heatmap(cm/len(class_names), annot=True,cbar=False, fmt='.2f',cmap='Accent_r')

plt.xlabel('Prediction')
plt.ylabel('Actual')
plt.title('Confusion Matrix');
4qYZEVAAAAAElFTkSuQmCC.png
#

Why the output displayed like that?

clever summit
#

Hello. I need help.

Source code: ```py
target=320
model_config=Path(file).parent.resolve()/'yolov3.cfg'
model_weights=Path(file).parent.resolve()/'yolov3.weights'

net = cv2.dnn.readNetFromDarknet(model_config,model_weights)
net.setPreferableBackend(cv2.dnn.DNN_BACKEND_OPENCV)
net.setPreferableBackend(cv2.dnn.DNN_TARGET_CPU)

Error: ```py
Traceback (most recent call last):
  File "d:\xampp\htdocs\dishub\AdminLTE-3.2.0\mesin_penghitung_2\sistem.py", line 27, in <module>
    net = cv2.dnn.readNetFromDarknet(model_config,model_weights)
cv2.error: OpenCV(4.7.0) :-1: error: (-5:Bad argument) in function 'readNetFromDarknet'
> Overload resolution failed:
>  - Can't convert object to 'str' for 'cfgFile'
>  - Can't parse 'bufferCfg'. Input argument doesn't provide sequence protocol

What's happening here? What does this error mean? How can I fix it?

spring field
clever summit
#

Okay, but what I don't understand is that if I remove the Path function, it shouldn't have had any problem, but instead:

image.png
spring field
#

str(model_config), str(model_weights)

#

when they are pathlib.Paths that is

clever summit
#

Aight. Thanks!

clever summit
#

Hello again. This time, I have a big question about directory path.

#

So, when I opened the exact folder in VSC where all my codes are written, my code ran well.

image.png image.png
#

But, once I open few files backwards, my code throws error instead.

image.png image.png
#

The thing is, my code is still the same, I have not changed it yet, but just because I open the 'wrong' folder in VSC (which technically contains the right folder), it gives a huge difference towards the result. Why?

river cape
#

HI guys what is use of the dummy variable trap?

past meteor
#

Aight, the rag thing is deployed

#

It doxes me very hard so I can't share ๐Ÿ˜ญ

#

Basic but functional UI

image.png
#

Honestly, I overengineered mine

#

Running a dagster instance for the pipelines, doing the UI in svelte, backend in FastAPI, pgVector DB, etc.

#

If I did it with streamlit/gradio + chromaDB I'd have been done in 4 days maximum

#

no

#

โ‚ฌ5-6 p/m server

#

That runs all of my other projects

#

When I finish the "host it yourself" section I think anyone could roll with it ๐Ÿ˜„

#

If you're applying a RAG like this may actually help

#

recruiters can use it to see what you've built

#

but that's also interesting ofc

#

Like I told you, the thing that interested people the most (tech and non-tech) was the stupid GenAI photobooth I made with Django for $largeMusticFestival in <1 week ๐Ÿฅด

#

automatically writing cover letters ๐Ÿ˜‚

river cape
#

What does this mean

image.png
past meteor
#

OpenAI

#

I have 4 gb ram

#

For ALL of my services lol

#

I have a spend cap ofc

# 
        current_spend = await self.retrieval_service.get_current_spend(date.today())
        if current_spend >= self.max_spend:
            raise MaximumSpendError()
#

My spend cap is brutal tbh

#

It can cost me max โ‚ฌ1.5/day

#

Which is 547 / y

#

nah I'm fine. I'll solve this with a clever feature ๐Ÿ‘€

#

"Apply for elevated spend" which sends me an email. I log into a management UI, generate a JWT and send it back to the person's contact details (+ delete them afterwards)

#

Yeah stuff like that can work for sure

#

I'd start with ollama for testing and then go to rust

#

Yeah, safetensors

#

or onnx

#

How efficient is it?

buoyant vine
#

it is relatively efficient

#

If I needed as much perf as possible I probably wouldnt use it th

#

not because of the inference but the API overhead and data wise isn't the best

#

yeah that is probably fine

past meteor
#

My issue is partly learning new APIs

#

When a threadpool and a basic endpoint could work

#

but that's a me problem

buoyant vine
#

Ollama's api is pretty simple though really

#

like for generative text it is pretty solid

#

The biggest issue i've had deploying models as an API before though (and this is less of a ollama thing and just a general thing) is load balancers hate them

#

your latency tends to be high and under higher load the load balancers end up opening thousands and thousands of connections ๐Ÿ˜”

past meteor
#

oof

#

Do a RAG, more impressive

#

Even though it's not that much harder

valid falcon
#

where is help about python?

errant bison
past meteor
valid falcon
#

ty

past meteor
#

Nice, lmk know how it goes

#

Yeah or actually

#

Why don't you just use AWS bedrock?

#

Last time I did a RAG before this I used Bedrock + Lambda

#

if you go with bedrock + lambda you don't need those

odd meteor
# errant bison Thanks a lot and what domain would u prefer, health care, education etc

I currently work on low-resource NLP, however, my domain of interest does not matter here.

It still boils down to you. I get that it can sometimes be
hard to narrow down / figure out a specific area of research interest.

I think you should make your decision based on what you find more impactful, fascinating, or very-confused-and-unclear-about-yet-curious-enough-to-find -out.

lapis sequoia
#

Hi guys, could anybody point me in the right direction? I want to train an agent from real-time frame capture. Which windows tools could help with this?

unkempt apex
lapis sequoia
#

Could OpenCV help with this?

unkempt apex
#

elaborate more about problem!

left tartan
agile owl
#

RIP

icy tapir
#

Hi, i have conducted a survey and in need of generating multi-level cross-tables like image shown below. I generated this using SPSS custom table function (CTABLES). I don't know if there is a similar Python package that does this. Any ideas ? Thanks

past meteor
#

at least use chromadb then lol

serene scaffold
icy tapir
#

raw data is .csv
i used SPSS for CTABLES function

serene scaffold
past meteor
#

I suffer from this too but making a kanban or similar on gh projects makes me feel a lot more accountable

#

I write down each idea and assign it to milestones

misty topaz
#

does anyone have any source or references about Q-learning? im kinda confused on how the algorithm works on python and whether to go with monte carlo or temporal difference

past meteor
#

Also, go with TD methods over monte carlo

#

So, Q learning, SARSA and so on

#

GH projects works really well for me and is low effort enough to consistently use

#

For this project I just make all ideas issues and assign them to a milestone or none (if I don't know what to do with it yet)

#

and then never do work that isn't linked to a milestone (at least, in spirit). Right now I'm in the "stabilising" milestone for my current project so no new features will be made

misty topaz
agile owl
#

ah doesn't get much better than this

image.png
lapis sequoia
#

the problem that i am the first using this data, so there is no previous work on it, so i can't compare

past meteor
#

I think I get what you mean

uncut orchid
#

Hi all! I am a surgeon in the USA (first-year resident) who is taking a year to publish some projects related mainly to biostatistical analysis / basic comparisons.

#

Would anyone have a good online guide / github guide for my usecases?

#

I'm new to coding but figured python would be a good language to focus on.

#

I mostly do basic descriptive stats, univariate or multivariate comparisons, survival analysis etc..

unkempt apex
#

that's nice to hear!!, well done SIR!!

rich moth
#

I'm running into an issue with the integration of sentence transformer. Im trying to expand the text_features tensor to get it to match the dimensions of the image features. You can see their shapes in the code below. But I get a error when expanding text features to image features, its expecting a 4 dimensional tensor. Im trying to project them to the same shape , how do I I get the tensor dimensions to match up?

text_features shape: torch.Size([2, 512])
image_features shape: torch.Size([2, 768])
Training:   0%|                                                                                                                                                                          | 0/1551 [00:00<?, ?it/s]
Traceback (most recent call last):
  File "/home/plunder/testrun01.py", line 347, in <module>
    main()
  File "/home/plunder/testrun01.py", line 328, in main
    train_loss, train_psnr, train_ssim = train(model, train_dataloader, optimizer, clip_model, clip_processor, sentence_transformer_model, manifold_autoencoder_optimizer, win_size, data_range)
                                         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/plunder/testrun01.py", line 224, in train
    output_data, vq_loss = model(images, augmented_captions, clip_model, clip_processor, sentence_transformer_model)
                           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/plunder/miniconda3/envs/qusar/lib/python3.11/site-packages/torch/nn/modules/module.py", line 1518, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/plunder/miniconda3/envs/qusar/lib/python3.11/site-packages/torch/nn/modules/module.py", line 1527, in _call_impl
    return forward_call(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/plunder/testrun01.py", line 151, in forward
    text_features = text_features.expand(image_features.size())
                    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
RuntimeError: expand(torch.cuda.FloatTensor{[2, 512, 1, 1]}, size=[2, 768]): the number of sizes provided (2) must be greater or equal to the number of dimensions in the tensor (4)```
wooden sail
#

rank usually means something else, but yeah

#

rank 4, order 4, 4-way array, tetradic/4-adic, among others

#

though rank usually relates either to the higher order SVD or to the canonical polyadic decomposition

unkempt apex
#

but how can a non-tech can join board?

#

is he joins for AI safety?

#

I was litterally amazed when Illya left!!

rich moth
#

Does anyone have any suggestions? I can post my VQVAEwithmanifold if anyones interested. ```Encoder output shape: torch.Size([16, 768, 56, 56])
Quantized shape: torch.Size([16, 768, 56, 56])
Difference shape: torch.Size([512])
Reshaped Difference shape: torch.Size([512])
Input shape to ManifoldAutoencoder: torch.Size([1, 512])

```Traceback (most recent call last):
  File "/home/plunder/FUCKYA4.py", line 372, in <module>
    main()
  File "/home/plunder/FUCKYA4.py", line 353, in main
    train_loss, train_psnr, train_ssim = train(model, train_dataloader, optimizer, clip_model, clip_processor, sentence_transformer_model, manifold_autoencoder_optimizer, win_size, data_range)
                                         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/plunder/FUCKYA4.py", line 240, in train
    output_data, vq_loss = model(images, augmented_captions, clip_model, clip_processor, sentence_transformer_model)
                           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/plunder/miniconda3/envs/qusar/lib/python3.11/site-packages/torch/nn/modules/module.py", line 1518, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/plunder/miniconda3/envs/qusar/lib/python3.11/site-packages/torch/nn/modules/module.py", line 1527, in _call_impl
    return forward_call(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/plunder/FUCKYA4.py", line 158, in forward
    _, manifold_difference = self.manifold_autoencoder(difference.unsqueeze(0))  # Add batch dimension
                             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/plunder/miniconda3/envs/qusar/lib/python3.11/site-packages/torch/nn/modules/module.py", line 1518, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/plunder/miniconda3/envs/qusar/lib/python3.11/site-packages/torch/nn/modules/module.py", line 1527, in _call_impl
    return forward_call(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/plunder/FUCKYA4.py", line 32, in forward
    z = self.encoder(x)  # Latent representation (manifold)
        ^^^^^^^^^^^^^^^
  File "/home/plunder/miniconda3/envs/qusar/lib/python3.11/site-packages/torch/nn/modules/module.py", line 1518, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/plunder/miniconda3/envs/qusar/lib/python3.11/site-packages/torch/nn/modules/module.py", line 1527, in _call_impl
    return forward_call(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/plunder/miniconda3/envs/qusar/lib/python3.11/site-packages/torch/nn/modules/container.py", line 215, in forward
    input = module(input)
            ^^^^^^^^^^^^^
  File "/home/plunder/miniconda3/envs/qusar/lib/python3.11/site-packages/torch/nn/modules/module.py", line 1518, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/plunder/miniconda3/envs/qusar/lib/python3.11/site-packages/torch/nn/modules/module.py", line 1527, in _call_impl
    return forward_call(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/plunder/miniconda3/envs/qusar/lib/python3.11/site-packages/torch/nn/modules/linear.py", line 114, in forward
    return F.linear(input, self.weight, self.bias)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
RuntimeError: mat1 and mat2 shapes cannot be multiplied (1x512 and 768x512)```
buoyant vine
#

Honestly all the AI pushing stuff is what is going to push me to linux tbh

#

Im not a big linux person outside of deploying servers, but god damn I just do not give a fuck about copilot of all this AI shit

#

get it off my software that I pay for

#

At least make it an optional feature

buoyant vine
#

only issue I have, is linux still suffers so heavily from the "tech people only" mentality

#

it has got a lot better recently mostly because of steam

#

but you still need to effectively compile everything from scratch

#

and you can't get most non-techy people to do that

#

No my experience has been in general, there will always be somethings that you end up having to compile from scratch

#

or require dependencies that require building from scratch

#

or have more complicated install requirements that require some level of technical knowledge

#

Probably top of my list would be OpenSSL related things, i.e. some packages that require openSSL but can't find it

#

yeah

#

Although I have personally had issues before trying to install even a simple deb file

#

Website list it as the correct version for the OS? Check!
Checksums match? Check!
Try to install via CLI? ERROR! This file isn't a deb

#

Yeah, and that is realistically the biggest problem

#

the errors and issues that arise normally assume technical knowledge

#

and that just doesnt work with normal users

#

Kinda feels like you need an OS ontop of linux especially designed for non-techy users

#

Chrome books are pretty solid... If it wasn't so google heavy LOL

rich moth
#

I fixed it! Training: 1%|โ–ˆโ–ˆโ–Ž | 22/1551 [00:38<43:20, 1.70s/it]

Now we are training with sentence transformers as well.

#

Dude that was a serious pain in the ass, let me tell you

#

I also increased the latent space from 128 to 256, I feel thats the sweet spot. 128 just wasnt capturing enough I feel.

rich moth
#

Its not just a normal sentance transformer though, its a clip one.

#

Well I guess there is no "normal" one, I meant for justal text, not images.

lapis sequoia
#

Guys I am stuck.
https://pytorch.org/tutorials/intermediate/reinforcement_ppo.html
This is my Pipfile:

[[source]]
url = "https://pypi.org/simple"
verify_ssl = true
name = "pypi"

[[source]]
url = "https://download.pytorch.org/whl/"
verify_ssl = true
name = "downloadpytorch"

[packages]
torch = {version = "==2.2.1", index = "downloadpytorch"}
torchvision = {version = "==0.17.1", index = "downloadpytorch"}
torchaudio = {version = "==2.2.1", index = "downloadpytorch"}
ipykernel = "==6.29.3"
matplotlib = "==3.8.2"
pandas = "==2.2.1"
seaborn = "==0.13.2"
opencv-contrib-python = "==4.10.0.82"
mss = "==9.0.1"
tensordict = "==0.3.0"
torchrl = "==0.3.0"
tqdm = "==4.66.4"

[dev-packages]

[requires]
python_version = "3.11"

Why does this line give from tutorial give error?

base_env = GymEnv("InvertedDoublePendulum-v4", device=device)
fallow coyote
#

Is grokking machine learning a good book to go through for getting into ML?

serene scaffold
rich moth
# 
Traceback (most recent call last):
  File "/home/plunder/FUCKYA7.py", line 380, in <module>
    main()
  File "/home/plunder/FUCKYA7.py", line 361, in main
    train_loss, train_psnr, train_ssim = train(model, train_dataloader, optimizer, clip_model, clip_processor, sentence_transformer_model, manifold_autoencoder_optimizer, win_size, data_range)
                                         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/plunder/FUCKYA7.py", line 248, in train
    output_data, vq_loss = model(images, augmented_captions, clip_model, clip_processor, sentence_transformer_model)
                           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/plunder/miniconda3/envs/qusar/lib/python3.11/site-packages/torch/nn/modules/module.py", line 1518, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/plunder/miniconda3/envs/qusar/lib/python3.11/site-packages/torch/nn/modules/module.py", line 1527, in _call_impl
    return forward_call(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/plunder/FUCKYA7.py", line 174, in forward
    aggregated_difference = aggregated_difference.view(batch_size, -1, 1, 1)
                            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
RuntimeError: shape '[11, -1, 1, 1]' is invalid for input of size 256```

Noo!
#

Anyone have any insight?

serene scaffold
#

Like, if you have a tensor of shape (4, 2), that has 8 elements. So you can't reshape it to (2, 3) because that would only be six elements. Every element needs somewhere to go

#

The -1 in that error message represents "whatever number would make this work". If you tried to reshape a tensor with 12 elements to (3, -1, 2), the -1 would get solved to 2, because 3 times 2 times 2 is 12

#

In your case, there's no integer that makes it work

#

Does that make any sense, @rich moth ?

rich moth
rich moth
violet gull
#

in reinforcement learning is there an alternate to completely random, random discovery. I dont want my agent taking completely random actions but i want it to explore

rich moth
#

I also had the wrong latent space of 256, i changed it to 768 and things to be training again. fingers crossed thank you

stuck kiln
#

How do you quickly learn statistics? I know NO statistics, but I've been told that CS majors are expected to know some. I have summer off, so how can I learn statistics quickly before next school year. Is Khan Academy good enough?
<#databases message>

spring field
#

Khan Academy is probably good, yes

spring field
rich moth
# 
Evaluation: 100%|โ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆ| 388/388 [05:10<00:00,  1.25it/s]
Epoch [1/5], Train Loss: 0.0001, Val Loss: 0.0001, Train PSNR: 14.2016, Val PSNR: 15.3402, Train SSIM: 0.1445, Val SSIM: 0.2102
Traceback (most recent call last):
  File "/home/plunder/OK.py", line 364, in <module>
    main()
  File "/home/plunder/OK.py", line 357, in main
    reconstructed_images, _ = model(sample_images, [""], clip_model, clip_processor, sentence_transformer_model)
                              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/plunder/miniconda3/envs/qusar/lib/python3.11/site-packages/torch/nn/modules/module.py", line 1518, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/plunder/miniconda3/envs/qusar/lib/python3.11/site-packages/torch/nn/modules/module.py", line 1527, in _call_impl
    return forward_call(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/plunder/OK.py", line 138, in forward
    tokens = sentence_transformer_model.tokenize(caption)
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/plunder/miniconda3/envs/qusar/lib/python3.11/site-packages/sentence_transformers/SentenceTransformer.py", line 319, in tokenize
    return self._first_module().tokenize(texts)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/plunder/miniconda3/envs/qusar/lib/python3.11/site-packages/sentence_transformers/models/CLIPModel.py", line 71, in tokenize
    inputs = self.processor(text=texts_values, images=images, return_tensors="pt", padding=True)
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/plunder/miniconda3/envs/qusar/lib/python3.11/site-packages/transformers/models/clip/processing_clip.py", line 98, in __call__
    raise ValueError("You have to specify either text or images. Both cannot be none.")
ValueError: You have to specify either text or images. Both cannot be none.```

Dang  I'll have to look into this more
#

Oh I see ,, I think its from the visual metric.. Doees that look right to you guys? It cant display the text output and the image?

rich moth
#

What do you guys think?

#

I might have create a dummy for the text for now. Hopefully this works.

#

Is it me or does that train loss and val seem suss?

serene scaffold
#

@rich moth can you put all of OK.py in the paste bin?

#

!paste

arctic wedgeBOT
#
Pasting large amounts of code

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

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

serene scaffold
#

the code that you put in the paste bin has to be the exact code that caused the error. If the line numbers don't match up, it's useless.

frigid cove
#

Can anybody tell me why am I getting this error despite the fact that I've already installed accelerate and ran all the cells again in Collab?

Screenshot_2024-06-16_220820.png
#

This is the error. Even though I already have accelerate installed in this notebook

Screenshot_2024-06-16_220830.png
upper crag
#

@frigid cove I think this is a package dependency error

frigid cove
#

Two things:

  1. Is this shit overfitting? The loss seems to reduce too fast (Gave it 5 epochs)
  2. Why is it stuck at 7/12970 despite that the loss is reducing
Screenshot_2024-06-16_224022.png
rich moth
# serene scaffold its actions will become less random over time.

I was able to bypass it for now passing an empty tensor to the visuals for the captions the model produces, though I would like it add it, the model is finally progressing to the next epoch. Training: 100%|โ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆ| 1551/1551 [40:19<00:00, 1.56s/it] Evaluation: 100%|โ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆ| 388/388 [05:26<00:00, 1.19it/s] Epoch [1/5], Train Loss: 0.0000, Val Loss: 0.0001, Train PSNR: 14.2622, Val PSNR: 14.6331, Train SSIM: 0.1447, Val SSIM: 0.0673 Training: 9%|โ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–Ž | 139/1551 [03:51<39:48, 1.69s/it]

#

Heres the visual.

reconstructions_epoch_0.png
#

Seems really dark, see how it plays out, might need mess with the parameters at bit.

coral field
#

for my regression model, while my outputs that don't require performing any y-transformations recieved permutation importances of: [8.489870495152069e-06, 2.8473137634987127e-05, 3.601992801842216e-05, 8.994490795259866e-05, 0.00014122290689252414, 0.00016634087419826262, 0.00018651374166957848, 0.0002985011378653167, 0.0003217476455222398, 0.0003648501915516668, 0.0005390584133917069, 0.0005510453761112769, 0.0005579521130572043, 0.0006692111046020154, 0.0007342147839425814, 0.0009495427176870068, 0.0010499860369553222, 0.0011171318800323078, 0.0014396640246170038, 0.0014551954296537551, 0.0016762460340624529, 0.002773101506971252, 0.0028425825210629868, 0.0029898274331477967, 0.003008392228598817, 0.0034062715902328537, 0.0036089692723145993, 0.005192301826945134, 0.005426078843544332, 0.008128894172522275, 0.00891206066296612, 0.012957150932567437, 0.01876308316741349, 0.0194401900313785, 0.019999771452021975, 0.024811335893464116, 0.027843860140545785, 0.046640726175167076, 0.06081653164310454] , why did the output that did require box-cox have permutation importances of these: ````[-24226294443.848663, -12645536262.108978, -8895211489.339962, -6409687232.491282, -6088093136.653414, -938402035.502533, 0.0, 0.0, 0.0, 0.000164794921875, 16940292.432662964, 95529664.47017364, 146881268.19153443, 232146763.68365327, 619801830.0394211, 772500704.3520828, 778769318.7313598, 881445589.4466995, 1019127781.606067, 1587857749.527005, 4347401829.157445, 8176495248.835803, 8916961444.561548, 10201667386.495699, 22008501374.868847, 26985242275.276268, 29101324791.81298, 29154799119.23103, 33261882890.59846, 35941789613.710495, 64786094845.843445, 109258007129.88943, 109541725905.22734, 148586347865.06268, 323553642632.7811, 389284519697.52673, 488638808637.0902, 613591860565.783, 1136668394142.4473]

#

for reference, before the output used box-cox, the feature importances were closer to the first example

rich moth
coral field
rich moth
#

it can handle zero and negatives value . Box-cox is designed for postive i thought.

#

just a thought

coral field
#

yeah all of my values are positive

#

the values above arent model predictions, they're feature importances

rich moth
#

oh gotcha

coral field
#

yeah...

#

like I really dont know why the permutation importances are that wack

#

i also tried log transformation but to little avail

coral field
#

but if i dont use BC transformationg for my "y", even though the permutation importances look more "normal", the accuracy is also higher

rich moth
# 
Evaluation: 100%|โ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆ| 388/388 [05:26<00:00,  1.19it/s]
Epoch [1/5], Train Loss: 0.0000, Val Loss: 0.0001, Train PSNR: 14.2622, Val PSNR: 14.6331, Train SSIM: 0.1447, Val SSIM: 0.0673
Training: 100%|โ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆ| 1551/1551 [41:31<00:00,  1.61s/it]
Evaluation: 100%|โ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆ| 388/388 [05:11<00:00,  1.24it/s]
Epoch [2/5], Train Loss: 0.0000, Val Loss: 0.0001, Train PSNR: 15.8758, Val PSNR: 17.7024, Train SSIM: 0.2390, Val SSIM: 0.3333
Training: 100%|โ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆ| 1551/1551 [40:28<00:00,  1.57s/it]
Evaluation: 100%|โ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆ| 388/388 [05:17<00:00,  1.22it/s]
Epoch [3/5], Train Loss: 0.0000, Val Loss: 0.0001, Train PSNR: 16.5871, Val PSNR: 16.8975, Train SSIM: 0.3003, Val SSIM: 0.3589
Training: 100%|โ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆ| 1551/1551 [41:02<00:00,  1.59s/it]
Evaluation: 100%|โ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆ| 388/388 [05:41<00:00,  1.14it/s]
Epoch [4/5], Train Loss: 0.0000, Val Loss: 0.0001, Train PSNR: 16.6891, Val PSNR: 17.2012, Train SSIM: 0.3141, Val SSIM: 0.3452
Training:  11%|โ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆ                                                                                                                                              | 175/1551 [04:53<37:03,  1.62s/it]```

Seems ok so far, the pictures are really coming to shape. Not so dark now. Im gonna have to experiment with different clip and sentence transformer models.  Roberta seems like a good choice.  I could a better clip model also, need to free up room though lol.
buoyant vine
#

Yeah I think you probably want some plotting going on

#

Loss of 0.0000 is a bit sus though when the SSIM is changing

#

Also not sure to what extent you are using the sentence transformer for, but if you're not looking to fine tune the model parameters on Roberta, I'd recommend using intfloat' s variety of pre-trained models for mapping text into vector space

lapis sequoia
#

Is there any way possible that given a audio file which will be passed through OpenAI whisper that we can extract the phonemes out of that file?

serene scaffold
lapis sequoia
spring field
#

I don't think f-strings have to do anything with safety pithink

#

is it even possible to prevent prompt injections?

#

well, any object you pass there is gonna get converted to a string, sure, but I'm not sure how that really helps

#

ah

spring field
#

lmao, is this how you gonna automate job applications pg_rofl

#

lol

#

(wait, like actually though?)

buoyant vine
#

It would be cool, although ngl I would be sweating at the idea of it hallucinating or just straight up fucking up

#

"Ah yes Lisan is indeed a terrorist attempting to sacrifice people to the duck god so they can enhance their crypto bitcoin prector ai"

viral bobcat
#

so i wanna go into AI but i dont know how, can enybody give me something to start

agile cobalt
coral field
#

I'm currently look at my regression problem's feature importances through Sklearn's "permutation_importance" method. Because my output data is skewed, I'm using a Box-Cox transformation on my output to help reduce errors. However, when I plot the "permutation_importance" values of the Box-Coxed values, I get the attached image's plot, whose values range from -24226294443.848663 to 1136668394142.4473. However, if I do not perform y-transformation, my values appear far more reasonable, from 1.978362396543032e-07 to 0.049242178178920196. While I know the "permutation_importance" method returns relative, not absolute, values, why is there such a large discrepancy in between the transformed and non-transformed? And would the feature importance data even be valid if I am taking the importances on the transformed outputs? Any help is greatly appreciated.

EP_CANCER_Importance.png
unkempt apex
#

can I ask for RL?

#

it's complicated error though

spring field
#

just ask, no harm in just directly asking your question and rather hoping that someone answers than waiting for permission to ask, then ask, then wait
you can occasionally (not too frequently) also bump your question a bit for more attention I guess, but just ask it

P.S. RL is kinda not simple though, yeah

unkempt apex
#

because I have created custom environment with help of gym, for my Pong game

int that environment everything is executing fine
but if ball touches one of the wall edges ( window of env) then it's shape changes and now I am not able to draw things

#

here is the window
that rectangle is striker, and that red dot is out ball ( I don't know how to convert it into circle)

Screenshot_from_2024-06-17_20-30-19.png
#

so it works fine initially, the striker is also anticipating it's position

#

but when that ball hits wall, it should get reflected, I mean it's velocity should get reversed but I think it's shape is also getting converted

#

should I upload code?

long canopy
#

anyone had any success with LLM-automated documentation generation?

long thunder
small wedge
unkempt apex
#

here we go!

#

I think ball is getting out of bound of the screen

#

I think I need to apply circle equation formula to represent accurate circle in that numpy array, but that's not requred now

toxic palm
#

For https://datalemur.com/questions/sql-bloomberg-stock-min-max-1
pls let me know, where the below solution is getting deviating from the problem.

SELECT ticker,
TO_CHAR(date, 'Mon-YYYY') highest_mth,
MAX(open) highest_open,
MIN(low) lowest_mth,
MIN(open) lowest_open
FROM stock_prices
GROUP BY ticker, TO_CHAR(date, 'Mon-YYYY')
Bloomberg SQL Interview Question | DataLemur

Bloomberg SQL Interview Question: Write a query to retrieve the highest and lowest open prices for each FAANG stock by month over the years.

rich moth
mild herald
#

I'm not sure if this is the right place to post this, but I am trying to program something that makes a simple query engine using phi3, the most important part is that this is running on Google Colab. The thing keeps alternating between two error messages, ReadTimeout: timed out and ConnectError: [Errno 99] Cannot assign requested address. If anyone has any ideas that would be great.
My code is ```py

Define system prompt and query wrapper prompt

system_prompt = "You are an instructor teaching people driving lessons about the rules of the road. Your goal is to answer questions as accurately as possible based on the instructions and context provided. Make sure to reference the document and explain how you got your answer"
query_wrapper_prompt = PromptTemplate("{query_str}")

Initialize the Llama model

llm = Ollama(
model="phi3",

#Changes how much it's allowed to generate
context_window=320,
max_new_tokens=100,

generate_kwargs={"do_sample": True},
# Give it the prompts from before
system_prompt=system_prompt,
query_wrapper_prompt=query_wrapper_prompt,

tokenizer_kwargs={"max_length": 100},

)

Set the LLM and other settings

Settings.llm = llm
Settings.chunk_size = 512

Create a vector store index from the documents

index = VectorStoreIndex.from_documents(documents)

Create a query engine from the index

query_engine = index.as_query_engine(include_text=False, response_mode="tree_summarize")

Define the predict function

def predict(input_text):
# Querys the engine from input
response = query_engine.query(input_text)
return str(response)

prediction = predict("What are signals used for?")
print(prediction)```
The data is a drivers manual btw

buoyant kite
lapis sequoia
#

yeah so i needed to use the api

#

maybe ill apply for card tomorrow lol

#

thanks ill take a look

rich moth
#

I added chatgpt-2 to introduce caption generation as well as the image generation. Lets see how this works out, the training just started.

#

Last night i set it up to train 15 epochs before bed, I woke up to the system rebooted. I was super bummed.

unkempt apex
#

why rebooted?

mild herald
#

Anyone have experience with google colab?

unkempt apex
#

everyone here has experience with that!

dusky cargo
#

looking into deep learning a lil bit. theres the idea of using a cost function, and then doing gradient descent to find optimal, weight and biases to tune. but the cost function just spits out a single number, how do you find a line from that on which to descend on?

rich moth
coral field
#

If I performed transformation to my x values to reduce skew, can I still use sklearn's "permutation_importances" to get feature importances? What about y- transformed data?

mild herald
mild herald
unkempt apex
mild herald
#

I've done immense amounts of googling for the exact same issues, this keeps happening over the past week

rich moth
#

Maybe related to timingout from the server.

#

let me see if I can run it locally. one sec

rich moth
#

Doesnt it have a requested timeout parameter?

#

I cant load up my jupyter server, that shut down did something nasty to my file system in my wsl2 ubuntu instance. I cant write anything.

odd meteor
mild herald
#

I'll give it a shot

rich moth
# mild herald I seem to keep getting errors whenever I try to run my ollama code, I put all th...

I found this online. https://onexception.dev/news/1081606/run-ollama-in-google-colab

Running Ollama in Google Colab: A Comprehensive Guide

This article provides a step-by-step guide on how to run Ollama, a powerful AI platform, on Google Colab, a free cloud-based Jupyter notebook environment. Learn how to set up your environment, install necessary packages, and configure your Ollama instance for optimal performance. Whether you're a seasoned AI developer or just getting started, th...

odd meteor
# mild herald I'll give it a shot

If for some reason you don't wanna run it locally, another alternative would be, using CrewAI to create an AI Agents, then plugging it to Groq so that the Agents could leverage top tier LLMs.

Groq doesn't have Phi3 though but they do have Llama3 8B and Llama3 70B, Mixtral, Whisper, and Gemma

And the beauty of it is, it cost $0.00 (they have a paid plan of course, however their freemium plan still packs a lot)

You shoukd check it out.

https://console.groq.com/docs/models
https://crewai.com

GroqCloud

Experience the fastest inference in the world

honest sierra
#

Hello! how can i run python in a super computer? i have a function f(n) that has a large time complexity and i can only perform values up to 30 in a reasonable amount of time in my computer. how can i get more values? would colab's gpu work for this?

agile cobalt
#

GPUs are not magic, they can perform operations in parallel, but would be useless if your function cannot run efficiently in parallel - and even if it can in theory, it can take some effort to make it actually run in parallel

#

realistically you should reduce the time complexity of it by optimizing it

honest sierra
#

Damn thats kinda hard

#

Also would I need to modify it in order to run it in parallel

#

Thank you anyway!

acoustic skiff
#

I have nominal categorical training data (several categories) and a continuous numeric target, what statistic test do I use to see whats the best features to add to my model? When I google this question I see stuff like: https://pythonfordatascienceorg.wordpress.com/chi-square-python/#chi-square but this is for when they're both categories I think, not sure what to do when my target is float

gilded dagger
#

Hi , Guys
Does anyone work rn in the programming field. i just Wana ask em some questions about the career and work . I'm fresh graduate ai programmer.

serene scaffold
rich moth
gilded dagger
rich moth
#

hey

Epoch [2/5], Train Loss: 0.0000, Val Loss: 0.0001, Train PSNR: 16.1015, Val PSNR: 17.9194, Train SSIM: 0.2466, Val SSIM: 0.3354
Epoch [3/5], Train Loss: 0.0000, Val Loss: 0.0001, Train PSNR: 16.7126, Val PSNR: 17.9312, Train SSIM: 0.3012, Val SSIM: 0.3463```

I totally forgot to create a visual plot for these metrics. I'll get those up so you dont have to deal with this  .
But here's the thing: the PSNR and SSIM metrics are improving, and the reconstructed images are looking better with each epoch. However, the train and validation losses are still super low (0.0000 and 0.0001). It's got me wondering if the gradient slope might not be measured correctly.
Is it possible that the model is actually learning well, even with these low loss values? Or could there be an issue with how the gradients are being calculated?
spring field
rich moth
# spring field what loss function are you using?

so I use is a combination of reconstruction loss, vector quantization loss, and CLIP loss. The reconstruction loss is the MSE between the original and reconstructed images. The vector quantization loss includes the MSE between the quantized vectors and the encoder outputs with commitment cost, the CLIP loss measures the similarity between image and text features from the CLIP model

spring field
rich moth
#

I use total_loss = recon_loss + vq_loss + clip_loss

spring field
#

if you're gonna plot them, might I suggest plotting each of them separately and the combined one as well? that might provide more insight into why it's so low
are you using cosine similarity? because if you are, you should invert it and normalize between 0 and 2 basically

rich moth
rich moth
spring field
#

that's great

#

can't wait to see the plots ๐Ÿ˜

timid urchin
#

Hi, I am trying to run a CNN model and i keep getting this error, does anyone have an idea?

Screenshot_2024-06-18_at_9.08.38_AM.png
wooden sail
#

it's missing a dimension (which might not be a problem), but you fed in something of size 154 instead of 145

timid urchin
wooden sail
#

yes

timid urchin
wooden sail
#

where did you create the training and testing data?

wooden sail
#

so you need to make X_test have the same size in shape[1] and shape[2] as X_train does

timid urchin
#

is this what you are referring to

Screenshot_2024-06-18_at_10.08.28_AM.png
wooden sail
#

and then the 3rd line, where you tell the model "this is the only shape you will accept"

#

so naturally X_test needs to have that shape

timid urchin
zealous shore
#

i need help

#

anyone here good with tensorflow on python

gaunt granite
#

Guys how much math should I know to start off with PyTorch ??

agile cobalt
#

uh, given that just removing "nerf" from that sentence goes from fun toy to seriously dangerous autonomous weapon, I feel like this is not the sort of thing you should be asking online

past meteor
#

Nice, you're going fast

#

How are you doing it? What stack?

spring field
#

regarding*

#

does it not end properly?

image.png
rich moth
# spring field regarding*

Frontend looks clean, this is what I need todo. I used flask and react for my frontend/backend but yours looks good. What did you end up using? Does look like its getting cut off at the end is this another bug?

#

@final kiln

past meteor
#

bedrock ๐Ÿ”ฅ

# 
    async def enforce_spend_limit(self):
        current_spend = await self.retrieval_service.get_current_spend(self.date)
        if current_spend >= self.max_spend:
            raise MaximumSpendError()
#

I do it on the full service ๐Ÿคก

rich moth
#

there you go

#

Ok so I kinda the captions working, but not really lol . Wasnt what I was expecting. But maybe because its the first epoch ill see how it goes.

image.png
#

Are you using an API like chatgpt for this?

#

You should check out mine sometime. It definently fun to play around with. Its an ensemble learning setup i think my biggest model is a 3B blenderbot distil. But yours looks good buddy

#

Or just check it out I can make it private on github and give you a link, you can run it or check it out give me feedback, whatever really.

#

Oh my graph is done now!

image.png
#

kinda lol

#

wow remember how they use to be really dark? Now it looks a bit like a bright negative. This is the first epoch

image.png
deep sleet
#

is there a benifit for mapping a series in panda with another series instead of using a dictionary

agile cobalt
deep sleet
#

Ohhh

#

Thx!

violet gull
#

If a RL agent achieves a maximum score on certain life cycle what would cause it to only hit it rarely instead of hitting it consistently. It proved it is capable?

#

Ex. Max score = 48 but average = 36

small wedge
violet gull
#

I really want this to work so I can scale it up

small wedge
#

I think I'd recommend running a test with a reasonably small number of agents and seeing exactly what each score was at the end of a cycle with average 36

violet gull
#

Wym

small wedge
#

we know that the random chance used for exploration is going to cause it to skew

violet gull
#

By the end the random chance is pretty much 0 and never triggers

#

Already verified that

small wedge
#

okay I'd still wanna see the actual scores that are being averaged

#

so you could see if it's a few outliers or if the models are consistently not preforming well

violet gull
#

Iโ€™ll rerun it later and ping you with the data