#data-science-and-ml

1 messages Β· Page 125 of 1

unkempt apex
#

because other sites are working

#

local issue

past meteor
#

Okay, I was slower than expected but the RAG project is as good as done now

#

Spent too much time on the UI

#

I used sveltekit. It's what I know the best

#

This is what I went with. I put in mock data because it doxes me

redacted.png
#

Honestly. I'd never want to build a BIG frotend project

#

I don't pay enough attention to data flow

#

Yup. I can imagine. It's the same with my frontend code but I honestly don't care there

#

I think to do frontend well you need a checklist of how to manage state

lavish igloo
#

yo am i allowed to ask for help for my python script?

past meteor
#

I kind of mentally do it, I have some heuristics about which state solution goes where. I also try and keep as much out of the FE framework as possible

past meteor
lavish igloo
# past meteor definitely, ask away

so i dont have any background in coding ok. i have to make a scoreboard for project and its so annoying bc my teacher my program to have file handling

#

the script i have theres something wrong

past meteor
lapis sequoia
#

can any one help me my computer vision task?

past meteor
lapis sequoia
#

I am making a software aka web app to segment floor and apply pre-made texture. The texture will be in png formate and square shape tile.

Now by using segmentation model I segmented the floor, but now it time to apply texture in cm with dept in image that where i am stuck from many days the linear look of texture dosent look real so we need to do pov scale (near tile bigger further tile smaller) but here issue is of room direction again please sugggese me what I do in this

I tried prespective scaling but room direction is another issue i tried using inpainting but they never listen to prompt and never took my texture as parameter

keen comet
unkempt apex
#

how can we identify that out input image is in PIL or numpy ?

#

I mean mostly they are in PIL right? if we didn't convert into numpy!

#

hugging face is also not loading!!

but hey I didn't tried DDOS

keen comet
unkempt apex
#

he said possible reasons might be , I tried DDOS on that site, ( in simple visiting same IP from multiple PC on same location )

but I didn't do that

keen comet
past meteor
#

This can be good for UX but also bad. It depends. What I know is that this is for #web-development πŸ˜„

lapis sequoia
#

i have to calculate vanishing points?

past meteor
#

I didn't know openai has vector stores πŸ‘€

#

Means I could've been done with my project even faster

#

Yes but it's not just that

#

I could've ran a cron job on my raspberry pi to fill the vector store and then I could've done basically all the rest client side on gh pages without having a backend (if I used bring-your-own-key)

lapis sequoia
#

u mind if u dm u?

past meteor
#

At the expense of vendor lock-in

#

Rn I have a bunch of other concerns like hosting the backend and the ELT pipelines on my server (which isn't hard, I basically reuse the same CI/CD)

#

But it's still there

lapis sequoia
#

so look at real image and the texture my s/w applied I need to make that texture in tile shape like 30x30 cm repeating all over floor creating3d effect in image

image.png image.png
#

yes but some one else need in tile shape

#

i am stuck for 1 week

past meteor
#

Vendor lock in as in, right now I'm doing my embeddings and generation with OpenAI but in my code they're decoupled / behind an interface. I could switch the embedding models to ones that understand code better if I want

#

Or with a larger context window

lapis sequoia
#

like using prespective transform? function

past meteor
#

Ah, yeah like that

#

My CI isn't that advanced lol. I just SSH into my server and build the container

#

I'm not using specific cloud services

#

No, a VPS

#

It sounds messy but I pay a fixed fee per month to host all of my projects

lapis sequoia
#

u mean line look likt this and a point?

past meteor
#

It's very simple

unkempt apex
#

suggest me best vpn then for pc

#

free!!

lapis sequoia
#

in this the issue is if the floor area is less then floor also leak out of this 2 lines

past meteor
lapis sequoia
#

see this picture

image.png
lapis sequoia
past meteor
#

I wanted to do it old school and just "rent" a server full time

#

I wanted predictable pricing

#

But for GPUs I'd do what you're doing ofc

#

If you're hosting basic web apps you could get away with €40-50 per year

#

Especially since the things I rewrote in rust use so little resources

#

Your data stuff I just work on-prem

#

Or idk databricks

unkempt apex
# 
def load_image(image_path):
  image = cv2.imread(image_path)
  image_rgb = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
  return image_rgb


def preprocess_image(image):

  if not isinstance(image, Image.Image):
    image = Image.fromarray(image)

  input_tensor = preprocess(image).unsqueeze(0)
  return input_tensor


image_path = "/content/drive/MyDrive/lane_detection/original/002136-R.jpg"

image = load_image(image_path)
input_tensor = preprocess_image(image).to(device)

so now things are getting clear to me

#

but one thing is

after doing
preprocess_image(image), why we are doing .to(device)

#

are we using cpu/gpu for this function?

#

okay then it's fixed that .to(device) is use to do all operations on device which will be either cpu or gpu

unkempt apex
#

@final kiln

hey do you remember I asked about deeplabv3 model few days ago

#

now I have learned this, so now you can clear the parameters

#

because it is only predicting left truck

tranquil mist
#

Does anyone here work as a data analyst that ends up having to do things waaaaay above your pay grade ?
I’m working as a data analyst to support myself through uni and I started out just doing basic dynamic reports and answering business questions for the execs
But nowadays I find myself having to create forecasting models, clustering of clients to help my boss bring new products etc..
I feel like I’m learning a lot but I’m also doing a lot more than my colleagues which are doing simpler stuff

#

I’m the youngest one by about 10 yrs

unkempt apex
#

why you want to focus on money then?? , you are learning more as compare to others!

#

I asked you about deeplabv3 few days ago!!

#

I was having issues regarding output

#

see I have come this far

Screenshot_from_2024-06-10_17-21-52.png
#

now the question is why all objects are not segmenting?

unkempt apex
#

sry but where did I check only truck channel?

in my final mask output?

#

yup!!

#

I think there maybe 20 channels !!

#

but is this possible to combine all of this into one masked image

#

hmm? and how to do this?

#

changing in argmax?

#

ohh

#

then what ?
how can we get that output tensor?

#

wait that output variable has it

#

right?

#

ohh my bad sorry

#

yup I print that output variable and now got tensor

#

yeah clear now

# 
output_predictions = output_predictions.cpu().numpy()

do I Need to do this for 'output' variable as we have now removed 'output_prediction' variable completely?

#

ohh let me try then

#

hmm got it

#

yup got it !!

#

here all classes are different

#

so yeah we have from 0 to 20 classes here

#

no

#

it's
21, 512, 512

#

strange

#

but how did you find out that we don't need argmax because I read about it and it is usefull for getting indices of tensor

#

did not understand this!!

#

okay in code right okay

#

help(model) is professional info I guess

#

but nice thing

#

okay so that's why the only truck was segmenting

#

they all got segmented in traning phase , but when using argmax only one class is indexing ? right

#

inference? what's that

#

okay

#

yeah got it !! fundamentals

spring field
#

wait wait, but it does return the index of the largest value

unkempt apex
#

The argmax is one of the functions that PyTorch provides. By using the argmax function, we can get the maximum value of elements from the tensor as well as we can also get the indices of the tensor. In PyTorch, we have another function to return the second-highest elements from the input tensor.

spring field
#

why not?

#

but I'm certain it's the index, I use it for accuracy

#

lol

unkempt apex
#

I did this with for loop

Screenshot_from_2024-06-10_18-03-32.png
#

wtf did happen here

#

wait I used colormap by the way

#

I did this with claude's code!!

#

means?? class 6?

#

ohh heatmap

#

right?

#

no it loooks scary now

#

what do you think about this?

#

it's AI generated ignore

#

so what now??

#

I need to learn about thresholding then !!

#

claude gives more accurate code than gpt

#

yup sir always!!

#

what this called as first of all

all the heatmaps and all other maps

#

I use that for only plotting

#

it is color map!!

#

I guess

#

should I try for different image?

#

okay

#

IOU

#

this is becoming interesting now

#

challenging but need to do research

#

okay one question,
this loss function is used in last layer right?

#

yeah got it!!

the way we normalize the input tensor

#

so which loss function to use??

#

okay thanks for the time !! gentlemen

#

then what to do??

#

softmax activation sorry

#

go for the interview!!

rare plume
#

Can someone help me with RL? specifically either DQN and custom environments

unkempt apex
#

I mean elaborate more!

rare plume
#

I want to create a model that generates a tensor of input shape(X, Y, Z) for example 3, 3, 3

[['0', '0', '0'],
['0', '0', '0'],
['0', '0', '0']],
['0', '0', '0'],
['0', '0', '0'],
['0', '0', '0']],
['0', '0', '0'],
['0', '0', '0'],
['0', '0', '0']]]

into [
[7, 4, 12],
[2, 16, 9],
[10, 15, 5]
],
[
[14, 3, 11],
[6, 8, 1],
[17, 13, 7]
],
[
[8, 6, 3],
[11, 2, 16],
[4, 12, 15] something like this, then we validate the tensor based on some rules I've set (I have the function in place already) if the tensor generated is valid, we give it a score based my scoring function (higher is better)

#

I've googled and it seems DQN is the best option for this, but all the DQN tutorials I've seen include openai-gym games, but I want my custom environment here

unkempt apex
#

@final kiln

do you think I should try YOLO8?

becasue I tried using softmax but still not getting proper output

#

because I think it can't predict about lanes

#

wait lemme search that

#

how was interview??

#

15 is for person

#

wait I am searching

#

still finding which class belongs to which object

#

yup I also find out this

#

These weights were trained on a subset of COCO, using only the 20 categories that are present in the Pascal VOC dataset. Also available as DeepLabV3_ResNet101_Weights.DEFAULT.

#

whole dataset??

#

it's 4gb

#

hmm searching now

#

it's available on kaggle

#

and it was created in 2012

#

here they have specified 20 classses

#

and yeah there is not lane marks here only some car detection and birds

#

is this is the reason that only left truck was detecting?

rare plume
#

Anyone help?

unkempt apex
spring field
unkempt apex
#

what about YOLO8?

unkempt apex
# spring field ^

yeah he is the same guy!!

I don't even know about fundamentals of RL so sorry!

unkempt apex
#

I thought you have deal with that anyways!

rare plume
unkempt apex
#

reinforcement learning discord
search this on browser

#

and you will get that

#

and yeah dont take pressure!

rare plume
#

They are dead mate

unkempt apex
#

no

rare plume
#

Yes

unkempt apex
#

sry yes

#

I never gave look at there chats sry

#

shit! πŸ˜‚

rare plume
#

Obviously

unkempt apex
#

what about Learn AI Together?

#

you are there

#

so ask there!!

#

maybe first ask in general discussion

rare plume
#

I'll try

odd lodge
#

Best book on machine learning?

unkempt apex
odd lodge
unkempt apex
#

check pinned messages

#

on top bar

odd lodge
#

Can you help me

unkempt apex
#

always sir!

unkempt apex
odd lodge
#

Book in pinned msg

unkempt apex
#

okay are you begineer?

odd lodge
#

Yes

unkempt apex
#

what you know in coding?

#

and what about you interest in ML?
NLP?? , Self driving Cars?? , healthcare, finance??

#

or pure data science , handing with data?

#

if you don't know then it's okay

#

maybe start with

-> types of ML
-> algorithms in each type

odd lodge
#

I just want to make an ai

#

Like Jarvis, πŸ˜„

unkempt apex
#

then you need to read more articles!!

#

just read whatever you can read about AI and ML

jaunty helm
# odd lodge I just want to make an ai

if by "ai" you mean the stuff that's mega advertised these days, look into neural networks and pytorch (they call it "deep learning" but that's just a buzzword for big neural nets)
though it might be better to start off with more fundamental stuff, e.g. you'll at least be working with numpy, and sklearn has many "classical" (read: not neuralnet) algos that are also interesting

odd lodge
odd lodge
unkempt apex
#

yeah for data!!

unkempt apex
jaunty helm
# odd lodge Pandas important?

pandas is if you actually do like preprocessing, data cleaning, etc.
if you're just fed a good dataset, you can skip it (though this almost never happens, it's a dream scenario really)

odd lodge
jaunty helm
odd lodge
#

Hmm

unkempt apex
odd lodge
#

Then suggest me now, I just complete a playlist on basic python..... What should I do next?

unkempt apex
#

and why we use that!!

odd lodge
unkempt apex
jaunty helm
odd lodge
unkempt apex
odd lodge
#

Hands-On Machine Learning with Scikit-Learn, Keras, and TensorFlow: Concepts, Tools, and Techniques to Build Intelligent Systems,

#

I found this book trending in machine learning

spring field
#

TensorFlow is not particularly trending

#

it's very much on the decline in fact

odd lodge
spring field
#

I get it, I'm just saying that TensorFlow is on the decline, sooo

odd lodge
#

Ohh

odd lodge
spring field
#

check the pinned messages in this channel

odd lodge
spring field
unkempt apex
#

new things

Screenshot_from_2024-06-11_01-04-57.png
unkempt apex
#

now adjusting confidence levels
the more it goes like 0.9 and all it stop segmenting

and less it goes like 0.1 or 0.01 it tries to segment road as a car ( maybe because box )

so I think 0.128 is best I read about this in github issue post

#

he is able to do this!!

rich moth
#

Alright ! It works. ```import torch
import torch.nn as nn
import torchaudio
from scipy.io.wavfile import write
import re
import matplotlib.pyplot as plt

Device Configuration

DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")

Audio Configuration

SAMPLING_RATE = 22050
HOP_LENGTH = 256

Load Pre-Trained Models

tacotron2 = torch.hub.load('NVIDIA/DeepLearningExamples:torchhub', 'nvidia_tacotron2', model_math='fp32').to(DEVICE).eval()
waveglow = torch.hub.load('NVIDIA/DeepLearningExamples:torchhub', 'nvidia_waveglow', model_math='fp32').to(DEVICE).eval()
waveglow = waveglow.remove_weightnorm(waveglow)
utils = torch.hub.load('NVIDIA/DeepLearningExamples:torchhub', 'nvidia_tts_utils')

Text Normalization Function

def normalize_text(text):
text = text.lower()
text = re.sub(r"[^a-zA-Z0-9\s]", "", text)
text = re.sub(r"\s+", " ", text).strip()
return text

Save Mel-Spectrogram to an Image

def save_mel_spectrogram(mel, filename):
plt.imshow(mel, aspect='auto', origin='lower')
plt.colorbar()
plt.title("Mel Spectrogram")
plt.savefig(filename)
plt.close()

Text-to-Speech Synthesis Function

def text_to_speech(text):
text = normalize_text(text)
sequences, lengths = utils.prepare_input_sequence([text])
sequences, lengths = sequences.to(DEVICE), lengths.to(DEVICE)

with torch.no_grad():
    mel_outputs, _, _ = tacotron2.infer(sequences, lengths)
    
    mel_outputs = mel_outputs.squeeze(0).cpu().numpy()
    save_mel_spectrogram(mel_outputs, "mel_spectrogram.png")
    
    mel_outputs = torch.FloatTensor(mel_outputs).unsqueeze(0).to(DEVICE)
    audio = waveglow.infer(mel_outputs, sigma=0.8)
    
return audio

Save the Audio to a File

def save_audio(audio, filename="output.wav"):
audio_numpy = audio.cpu().numpy().squeeze()
write(filename, SAMPLING_RATE, audio_numpy)

Main Execution Block

if name == "main":
text = "Hello, how can I assist you today?"
audio = text_to_speech(text)
save_audio(audio, "output.wav")

tranquil mist
# unkempt apex then it's good sir!!

Oh yeah I’m not complaining about the learning part, it’s more so that I have way too many responsibilities and I feel like my academic performance is suffering because of that.
It sucks because I really need this job to pay for university 😦

#

Also being the only one with intimate knowledge of many of our datasets (by virtue of working with them a lot) means I’m usually the main guy higher ups go to for business questions
I guess I showed a little too much initiative 😦

keen comet
copper mirage
#

where can you guys find text corpus to train natural language models?

#

in txt file if you know

buoyant vine
#

to train it to do what

#

LLMs in general need a metric fuck ton of data to train from scratch

serene scaffold
copper mirage
#

For typing assistance, so like something like gmail’s autocomplete when writing emails, but really any large text corpus from a tv show script will be great too

serene scaffold
#

because any language model that you can make on your computer will suck

#

(compared to "large" language models)

buoyant vine
#

also for the record

#

you are probably better (for your local use case) to not use AI

#

it is a common thing to do algorithmically

#

something like Symspell can do compound aware completion, it just requires loading a dictionary, and the rest you can realistically do with a trie and set of common phrases and terms within emails

copper mirage
serene scaffold
copper mirage
rich moth
#

This is super crazy. It's evident my system posses some kind of transfer and recursive learning. It's learning and applying that knowledge dynamically in real-time. It then saves in a conversation database which is embedded and vectorized for later use.

image.png
buoyant vine
#

Ngl that screenshot of a wall of text generally means not much to most people

#

also from the text in that screen shot... It looks like it is largely regurgitating nonsense

marsh plover
#

How would I sort a multi-level dict like this by name?

#

a = {'abc': [{'name': 'fro'}, {'name': 'lucky'}], 'def': [{'name': 'Leyda'}]}

#

I know its already sorted, but in case it wasn't, how would I do that via sorted()

#

While keeping memory efficiency in mind. Imagine if this dictionary was 1GB for example.

#

This doesnt work sorted(a.values(), key=itemgetter('name'))

dreamy isle
#

sorted() returns a list

#

oh wait yea

dreamy isle
marsh plover
#

yea

#

but i also want to flatten the lists

dreamy isle
#

you can make a dict comprehension that does that

marsh plover
#

into a single list

dreamy isle
#

"flatten" how?

marsh plover
#

there's two lists potentially in .values()

#

'abc' is a list of dicts and 'def' is also a list of dicts

#

flatten as in only return a single list

#

and sort all values regardless of key (abc, def)

dreamy isle
#

so one sorted list result from a multi-level dict?

marsh plover
#

yes

#

but without loading everything into memory (if possible) and using generators when possible

dreamy isle
#

sorted() would need to load everything into memory anyway, but..
i guess an itertools.chain of a comprehension of generators from a.values() would work

# 
sorted(itertools.chain.from_iterable(map(itemgetter('name'), list_) for list_ in a.values()))
marsh plover
#

how would I keep all the values in the dict not just 'name'

#

how would i sort by a specific key, lets say not name but another key?

dreamy isle
marsh plover
#

ty I think I got it?? return sorted( chain.from_iterable(set(map(itemgetter(*values), list_)) for list_ in price_comparison_map.values() ), key=lambda x: x[-1], reverse=reversed)

#

values = ('key1', 'key2', 'etc')

dreamy isle
#

does the key= actually work?
it seems like it's indexing a set

spring field
#

obviously reversed will work too, but it's just... well, for one it can be confusing, for two, it's semantically incorrect

marsh plover
#

is i t better to catch a datetime strptime conversion error or do some if checks

#

Is this a good practice?

      # Ensure we are in the correct date
      try:
        effective_date = datetime.strptime(row['Effective Date'], "%m/%d/%Y")
      except ValueError:
        continue
      else:
        if effective_date.year != year:
          continue```
#

The idea is to skip the given row in the CSV if we failed to parse the effective date

rich moth
rich moth
marsh plover
#

Is doing two sorted() calls to get top 10 increases and top 10 decreases better to do with two separate sorted() calls? Or is it faster to just call sorted once and then reverse it via [::-1] instead to get both?

rich moth
#

Yes, its generally faster. Some situations where reversing a single sort list much be the prefered method like very small list.

rich moth
#

Is this a type of scoring system?

rich moth
# marsh plover ty I think I got it?? ```return sorted( chain.from_iterable(set(map(item... 

# Function to sort a list of dictionaries based on the 'value' key and get top increases and decreases
def sort_by_value(data):
    sorted_data = sorted(data, key=lambda x: x['value'])
    
    # Get top 10 increases
    top_10_increases = sorted_data[-10:]
    
    # Get top 10 decreases
    top_10_decreases = sorted_data[:10]
    
    return top_10_increases, top_10_decreases

# Function to sort nested dictionaries by a specified key
def sort_nested_dicts(data, sort_key='name'):
    for key, name_dicts in data.items():
        data[key] = sorted(name_dicts, key=lambda x: x[sort_key])
    return data

# Function to flatten and sort a multi-level dictionary by a specified key
def flatten_and_sort(data, sort_key='name'):
    # Use itertools.chain to flatten the lists
    flattened = chain.from_iterable(data.values())
    # Sort the flattened list by the specified key
    sorted_flattened = sorted(flattened, key=lambda x: x[sort_key])
    return sorted_flattened

# Sample data for sorting by 'value'
data = [
    {'name': 'item1', 'value': 10},
    {'name': 'item2', 'value': 30},
    {'name': 'item3', 'value': 20},
    {'name': 'item4', 'value': 25},
    {'name': 'item5', 'value': 15},
    # Add more items as needed
]

# Sample nested dictionary for sorting by 'name'
a = {'abc': [{'name': 'fro'}, {'name': 'lucky'}], 'def': [{'name': 'Leyda'}]}

# Sort the data based on 'value' and get top 10 increases and decreases
top_10_increases, top_10_decreases = sort_by_value(data)

# Print the results
print("Top 10 Increases:", top_10_increases)
print("Top 10 Decreases:", top_10_decreases)

# Sort the nested dictionary based on 'name'
sorted_a = sort_nested_dicts(a)

# Print the sorted nested dictionary
print("Sorted Nested Dictionary:", sorted_a)

# Flatten and sort the multi-level dictionary by 'name'
sorted_flattened = flatten_and_sort(a)

# Print the flattened and sorted list
print("Flattened and Sorted List:", sorted_flattened)```
lapis sequoia
#

hey yall, im trying to setup chatgpt 3.5 turbo api in python and just getting errors out the ass, can't even get one good answer from following official docs and yt videos to get the damn thing to properly supply an answer based on user string input

#

ive even paid $5 for usage credits lol

lapis sequoia
mild grotto
#

Hey, I don't suppose anyone can help me implement the Esoteric Pull kernel in python/cupy

#

I understand the algorithm, but I'm a little at a loss on how to write the kernel

#

I need a data structure which is (X,Y, 9) and then for each X,Y, I want to touch specific neighbors

#

I'm a little at a loss

#

Actually, I'll just try to explain it to chatGPT πŸ˜„

#

Apparently I can do it like this

kernel_code = '''
extern "C" __global__
void esotericKernel(float* data, int X, int Y) {
    int x = blockIdx.x * blockDim.x + threadIdx.x;
    int y = blockIdx.y * blockDim.y + threadIdx.y;

    if (x >= X || y >= Y) {
        return;
    }

    float sum = 0.0;
    int count = 0;

    // Add (X,Y,0)
    sum += data[(x * Y + y) * 9];
...

That more or less solves my confusion

#

GPU programming seems so magical, but slowly peeling back the layers

lapis sequoia
#

I'm going to have chatgpt explain and provide "example" code for my own C assignments when i retake in the fall lol

mild grotto
#

Yeah I commonly use it in my hobby projects. It's pretty great except when you start trying to do something no one has done before

#

I was getting pretty frustrated for a few weeks that my cupy code was so slow. Moving to c++ kernel is doing the trick

toxic condor
#

u guys really need to use cuda in ai case? for me it's just a wrapper, only caring about the python level api

keen comet
#

id say

toxic condor
keen comet
mild grotto
#

I'm actually doing something that isn't AI, I'm just using the same tech stuff, so yeah you don't need to do what I'm doing

#

I'm implementing the Esoteric Pull algorithm for Lattice Boltzman Method fluid simulation on GPU

toxic condor
toxic condor
keen comet
unkempt apex
#

and now he is able to do lane detection

keen comet
toxic condor
high agate
#

I'm currently building a machine learning model for a classification task. Right now, I'm at the stage of imputing missing values. Below is the code I'm using to fill in missing values for categorical data:

# Categorical columns
cat_cols = ['Gender', 'Married', 'Dependents', 'Education',
           'Self_Employed', 'Credit_History', 'Property_Area']

# Impute categorical columns
cat_imputer = SimpleImputer(strategy='most_frequent')
cat_imputer.fit(train_df[cat_cols])

train_df[cat_cols] = cat_imputer.transform(train_df[cat_cols])
test_df[cat_cols] = cat_imputer.transform(test_df[cat_cols])

When I run the code above, I get this error:
TypeError: Argument 'a' is not recognized as numeric. Support for input that cannot be coerced to a numeric array was deprecated in SciPy 1.9.0 and removed in SciPy 1.11.0. Please consider 'np.unique'.

How can I fix this error?

high agate
#

Can you enlighten me what does that error means?

#

What confuses me is why I got an error when I try to impute categorical data

#

These are the packages I'm using:

import pandas as pd
import numpy as np
from sklearn.impute import SimpleImputer 
from sklearn.preprocessing import LabelEncoder, MinMaxScaler
#

what should I do?

#

How can I determine the suitable version of Scipy to handle this issue?

unkempt apex
#

nice? so quick or what?

I am thinking now that I will label my data using this method

#

because YOLO8 is segmenting cars and trucks properly\

unkempt apex
#

but I have seen some videos of some people that they are using YOLO8 to detect lanes also!!

#

but how?
have they finetunned that ? or they trained the same YOLO8 model for lane detection also?

#

what do you think?

high agate
#

But I don't understand what the sentence "Please consider np.unique" in that error means. Can you explain its meaning?

unkempt apex
#

and unique is unique

#

simple!

#

hey @final kiln
the above code is appropirate I guess becaause he is using CNN ( U-NET ) to train the dataset

and that other guy which I provided above this is just using ROI ( Region Of Interest ) and some mathematicals term to derive lanes?

so what do you think which method should I follow, I mean I also want to learn the architecture

#

it is in tensorflow what about pytorch

#

the main goal of mine is to create my own model, so that I can learn more things

#

that's why I am simply labelling my data first to train it

#

but to label I need some pre-trained models

#

yeah but my goal is not just segmenting images and lane detection so that I can use pre-trained model
I wanna make my own model!!

#

yup, but data ?
you mean labelled data with object segmented parts and lane and stuff?

#

like cars, trucks

#

so that my model can understand this things

#

but that data is not labelled right?

#

so it will be hard for me ( my model ) to understand things

#

yup using architecture and stuff

#

so you are telling me that , I shall build a model like them

#

and train on dataset

#

yeah

#

like how original image ( data ) is being transformed into new image, in which image the model is segmenting cars, trucks and lanes ?

#

I have a dataset which contains 1000 images ( both left lane and right)

#

you mean first create one or use pre-trained?

#

ohhhh got it!!
so I can take reference how other people have done that already and create/write my own

#

right?

#

but how??

#

okay then it's scratch process

#

got it

#

yeah

I was thinking I would take reference from other people how they builld and then same I will build my own

#

yup now make sense

#

and what about image segmentation like cars, trucks ,YOLO8 is doing that perfectly do you think I should spend time for creatiing my own for this task

or I should only do it for lane detection

#

Lane Detection!!

#

yeah

#

this guy litterally did this already!!

#

okay got it!!

#

I was just giving reference

#

yeah got it , but with only CNN it will be hard

#

then starting with reading paper good idea I guess?

#

yeah !! I have watched that

#

so general question is ,
CNN retrieves features from input images

#

then how it identify such features as lane markings on road?

#

yeah I watched yt video about this

#

when you started coding?

#

that's readigng and undeerstanding of whole architecture!

#

no it was different shape I guess

#

it's trapezoid

#

but then I need figure out how many number of max pooling layers, and CNN layers all that

main citrus
#

Why plotly is not working with ipywidgets?

void crescent
# 
y_pred = model.predict(tf.expand_dims(X_test, axis=-1))
y_pred

So im getting an error when I try to add tf.expand_dims to the X_test, and when I dont it gives me an error that Dense layers are incompatible with rank 1 tensors.

ValueError: Creating variables on a non-first call to a function decorated with tf.function.
#

nvm fixed it was a colab problem

left tartan
dreamy sorrel
#

for the final performance of each fold in cross validation, do you consider the average accuracy and loss or just accuracy and loss of the best epoch?

serene scaffold
past meteor
serene scaffold
#

imagine not agreeing with me

dreamy sorrel
past meteor
#

If you have enough data the average perf on the folds will be similar so take the mean. if you have very very little data you can do leave-one-out-CV and then you probably don't want to look at n-1 folds

misty shuttle
#

how do i start learning scikit learn?

past meteor
# misty shuttle how do i start learning scikit learn?

There's not a lot to learn in scikit learn πŸ˜„

You call .fit on models to train them .predict to ... make a prediction. Aside from that you use ColumnTransformer to apply preprocessing (OneHotEncoder, StandardScaler) to one or more columns. After you have all these you assemble multiple things with a Pipeline. If you're working on tabular data you should typically use HistGradientBoostingRegressor or the classifier version.

The hard part isn't the code (so, learning how to use scikit learn) but rather conceptual understanding.

misty shuttle
#

tysm!

lapis sequoia
# misty shuttle how do i start learning scikit learn?

I would say scikit learn is unending, just like say numpy and pandas.
It's important to know what you want to implement and know the necessary sklearn models you want to use.
Overtime you become familiar with it.
As for how the particular model works, you can take out time to look into them.
Trust me there are a thousand and one things that even experts ml scientists don't know in sklearn.

long robin
#

Im learning the basic machine algorithms in Scikit learn.

I have learnt numpy, pandas, matplotlib, and seaborn

Just finished with Linear Regression.

I saw some people on kaggle. They do EDA first. Then they fit the data to different regression model, then find out the r2 score. They pick up the one with the highest accuracy or score, and then they tune it's hyper parameter. If the score was 0.76 they would try to enhance it to somewhat 0.81 (for example)

Is that the actual and the usual way people do it????

lapis sequoia
#

The closer r2 score is to 1, the more accurate your model is.

long robin
lapis sequoia
long robin
lapis sequoia
abstract mica
#

I need help learning the finer optimizations of .h5 caching with python and processing

unkempt apex
#

hey @final kiln

how can I determine correct number of input/output channels for CNN?

#

wait there is a pdf regarding to this

spring field
keen comet
#

guys ive downloaded the llama3 model from meta but dead ass have no idea how to run it with python as a chatbot.. anyhelp?

keen comet
#

like man how tf do i run this shit 😭

image.png
keen comet
#

fr ran torchrun --nproc_per_node 1 example_chat_completion.py \ --ckpt_dir Meta-Llama-3-8B-Instruct/ \ --tokenizer_path Meta-Llama-3-8B-Instruct/tokenizer.model \ --max_seq_len 512 --max_batch_size 6 like acording to the docs and it says "failed to create process."

#

am i acoustic or something?

long robin
#

Some guys in my college are directly jumping over tensorflow

Im doing Scikit learn first... Will do TF later... Is that alright

unkempt apex
#

everything is alright if you have strong fundamentals!!

long robin
rich moth
keen comet
rich moth
#

Ahh

keen comet
# rich moth Ahh

i downloaded the model with the download.sh from meta and now im trynna run it with transformers cuz docs didnt work and im missing config.json so now i requests ascess to hugginface, hope it works

rich moth
#

I'll help show me what you got when you are initializing the model.

keen comet
unkempt apex
#

I ask the same question to gpt and now he develops whole CNN

keen comet
rich moth
# keen comet well first i tried <https://github.com/meta-llama/llama3> and now i am trying ...

from transformers import LlamaForCausalLM, LlamaTokenizer, LlamaConfig

Load the Meta-Llama model

llama_model_name = "Meta/Llama-3-8B-Instruct"
llama_model = LlamaForCausalLM.from_pretrained(llama_model_name)
llama_config = LlamaConfig.from_pretrained(llama_model_name)
llama_tokenizer = LlamaTokenizer.from_pretrained(llama_model_name, padding_side="left", config=llama_config)

tru to init it like that.

keen comet
# 
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "meta-llama/Meta-Llama-3-8B-Instruct"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

messages = [
    {"role": "system", "content": "You are a pirate chatbot who always responds in pirate speak!"},
    {"role": "user", "content": "Who are you?"},
]

input_ids = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt"
).to(model.device)

terminators = [
    tokenizer.eos_token_id,
    tokenizer.convert_tokens_to_ids("<|eot_id|>")
]

outputs = model.generate(
    input_ids,
    max_new_tokens=256,
    eos_token_id=terminators,
    do_sample=True,
    temperature=0.6,
    top_p=0.9,
)
response = outputs[0][input_ids.shape[-1]:]
print(tokenizer.decode(response, skip_special_tokens=True))
rich moth
#

At the top there.

#

at the imports and replace everything after = with what I wrote and add the config.

spring field
keen comet
#

sorry am confused

#

wait figured it out with ma 2 braincells

rich moth
spice jay
unkempt apex
unkempt apex
warm copper
#

so crazy

# 
encoder = OneHotEncoder(sparse_output=False)

y_train_re = np.array(y_train).reshape(-1, 1)
y_test_re = np.array(y_test).reshape(-1, 1)

y_train_encoded = encoder.fit_transform(y_train_re)
y_test_encoded = encoder.fit_transform(y_test_re)


model = Sequential()

model.add(Dense(64, input_dim=X_train.shape[1], activation='relu'))
model.add(Dense(32, activation='relu'))
model.add(Dense(3, activation='softmax'))


model.compile(loss='categorical_crossentropy', optimizer=Adam(learning_rate=0.01),
              metrics=['accuracy', 'Recall'])

history = model.fit(
    X_train,
    y_train_encoded,
    epochs=75,
    batch_size=32,
    validation_split=0.2,
    verbose=1
)

results = model.evaluate(X_test, y_test_encoded)
print("Test Loss, Test Accuracy, Test Recall:", results)
#

my first model with MLP

frigid cove
#

Should all images of Vision Transformer be of the same format?

keen comet
frigid cove
#

By format I mean either jpg or png, not size

keen comet
frigid cove
#

Some are png and others are jpg

#

Most are jpg

warm copper
#

I need a higher recall

keen comet
warm copper
#

Test Loss, Test Accuracy, Test Recall: [0.9047948122024536, 0.5376852750778198, 0.5376852750778198]

keen comet
warm copper
#

intrustion detection @keen comet

#

it suffers significantly from imbalanced labeling

#

So I started with ensemble methods

#

since they are more resiliant against imbalanced data

spring field
# frigid cove By format I mean either jpg or png, not size

no, it doesn't matter, just make sure that they are loaded with the same value range, I remember that loading pngs with plt.imshow returned an array of floats in range [0, 1], whereas with jpegs it was an array of integers in range [0, 255]
or something like that
just a heads up in case you run into some weird issues, but otherwise... ehhh, shouldn't really matter

#

also be careful of the channel count as pngs could have 4 channels

#

it's probably easier to just convert them though, yeah

#

unless you've got a ton of them

warm copper
#

hullooo

ashen axle
hidden meadow
hollow escarp
#

Does anyone ever done some object detection stuff on orange pi devices

errant musk
#

What has your experience been like so far, if you've used it?

hollow escarp
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)
    # suppose input image is [ 3, 128, 128 ]


    # 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, 2)

  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

model = LaneDetectorCNN()
#

so as per the code the output is of tensor with shape [1, 2], in simple they are 2 values!!

but what are those 2 values now?

unkempt apex
#

hey @final kiln need some help , I am confused here!!

sick tulip
#

Suggest some good books to learn DS

fallow coyote
coral field
#

So for two of my input features (out of ~30), the skew resides around 70-80 after appling log/boxcox transformation, would it be reasonable to remove the features from the dataset?

spring field
hollow escarp
hollow escarp
spring field
#

I don't even know what the N stands for, lol

#

wait, it's probably Neural

midnight harbor
#

I need help in choosing a good embedding model,

Documents are medical related and searching those documents can be like sentence search or just keywords so

What embedding model Opensource would be great for this?

past bramble
#

if I have trained a model with TensorFlow on datasets of single eye images, how can i further use it to detect and track eyes on casual images of a person?

serene scaffold
past bramble
#

given a cropped picture of a single eye, it tells if its closed or pointed at right/ left/center direction

serene scaffold
past bramble
serene scaffold
#

do basically the same thing you already did, but with images that include "everything", not just cropped to the eye.

past bramble
#

i need to simply track eye in any kind of image, be it a person taking a selfie or having breakfast or in a scenery. I don't think training on such images can help detect eyes

serene scaffold
agile cobalt
#
    1. can we not do Black Mirror level creepy stuff?..
    1. in general your model can only do what it's been trained for
serene scaffold
#

if you have a model that has only ever been trained on images of eyes, everything that isn't an image of an eye--and only an eye--is meaningless to the model.

past bramble
#

how exactly are facial recognition models built then?

serene scaffold
#

if the model has only ever seen closely cropped images of eyes, there's nothing in the model that could be construed as knowledge of what eyes are, or where they exist in relationship to anything else in the world (like faces)

tidal bough
#

in theory you could use some image segmentation model that can detect eyes on an image, and feed the cropped eyes into your model, but it'd probably be both simpler and more effective to train your model on real images.

serene scaffold
past bramble
#

how does that work?

unkempt apex
#

lemme clear some confusion !!

so currently that CNN has 2 neurons at last layer okay!!, so the thing here is how we are using that neurons, because we just want that lane lines pixels to be detected right?, so I was confused that , are only 2 neurons are sufficient for this task ? because when I ran that code the output values was both negative e.g [-0.9xx, -0.8xx ]
so what this values are trying to say to us??

#

yeah I search about how to visualize output of CNN , but it was ambigious for me!

#

what??, converting those 2 last layer's neurons into 512?

#

yeah got it now!!

#

but hey, will it break the structure of CNN, because I was spending some valuable time understanding maths behind max_poooling and conv operations

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

it's interesting

#

wait lemme change those values and see the results firsts!! curious now....

#

what are hyperparameters here?

#

for example?

#

new output is interesting also ( no it's not it just tensor or vector )

#

so what we can do with that now??

#

torch.Size([1, 512])

#

can I ask how?

#

should I convert that into numpy as see as image , for fun!!

#

but is our approach correct??

and why did you choose 512 neurons?
because out input image is of size 128 so why big number?

#

okay interesting give me some time to read this !!

#

because I converted that into 128

# 
transform = transforms.Compose([
    transforms.ToPILImage(), # no need of this if image is already in PIL
    transforms.Resize((128, 128)),
    transforms.ToTensor(),
    transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
])
#

look here

#

yeah it's clear now

#

but 512 will expand those?

#

anyways lemme try atleast , and yeah how can this transform.Normalize value are only 0.5

#

A transposed convolutional layer, on the other hand, is usually carried out for upsampling i.e. to generate an output feature map that has a spatial dimension greater than that of the input feature map. Just like the standard convolutional layer,

that's why are we using 512 neurons??

#

so what do you think? 256 or come back to 128

#

yeah! but for now??

#

because input is 128

#

okay back to reading now!

#

-> torch.Size([1, 3, 128, 128])

let it be for now !!

#

wait how this text got bigger size??

#

because of -> ???

# 
def trans_conv(X, K):
    h, w = K.shape
    Y = torch.zeros((X.shape[0] + h - 1, X.shape[1] + w - 1))
    for i in range(X.shape[0]):
        for j in range(X.shape[1]):
            Y[i: i + h, j: j + w] += X[i, j] * K
    return Y
#

X is input tensor , for me it's output of CNN

#

what about K ( kernel tensor )

#

how can I get that

#

wait , kernel values are out weight values right??

#

for each layer, for example here we defined 3 conv2d layers so 3 kernel values

#

so we have 3 kernel values now??

#

right?

#

??

#

we have 3 kernel here? as per 3 conv2d layeres

# 
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 which one?

#

I think I am also getting confused

#

is there any article of docs regarding this?

# 
torch.nn.ConvTranspose2d(in_channels, out_channels, kernel_size, stride=1, padding=0, output_padding=0, groups=1, bias=True, dilation=1, padding_mode='zeros', device=None, dtype=None)
#

lemme clear then now!

#

the in_channels are output of our current cnn right?

#

in_channels (int) – Number of channels in the input image

#

channels are what layers??

#

no it's not

#

yup reading about that

#

input.shape = [3, 128, 128]

#

so here 3 channels, 128x128 width and height

#

so our output has shape of this torch.Size([1, 512])

so what does this mean?

#

1 channel with 512 values?

#

yup

#

so now question is how can we convert this?

#

in what ??

#

hmm,
[1, 512, 512] ??

#

but that's not possible

#

ohh hooo

#

we want to create tensor right to represent it as a image

#

and we have only 512 values!! ( vector )

#

so what will be size, lemme do calculation

#

1 is fixed in .shape() because it's channel

#

what about width and height

#

22*22 ??

#

23*23

#

exact shape not found !
ohh god

#

come on give me a hint atleast

#

is channel should be 1 only??

#

is it fixed?

#

bro I got different shapes

#

16 channels with 4, 8

#

32 channels with 4, 4

#

hey but we need only one right?

#

so how can I reshape this now?

#

yeah reading docs now

#

a = torch.reshape(output, ())

which value to give in this tuple?

#

and store in which variable, the name should be relevant?

#

okay what's next now, we got tensor now !! yeah!!

#

why batch is -1 here?

#

-1 represents flat vector values right?

#

and now I am getting confuse with those short names

#

x_bchw as x_batch weights

#

x_b1d what??

#

okay got it

#

now what's next?, encoder or what?

#

as per images

#

so it's batch size

#

yeah so D is out model's output
and B is .....

#

yeah!

#

I am now veery much confuse but anyways new things are always like this

#

πŸ˜‚

#

output.reshape(32, 4, 4)

so suggest a name for this now and then we will move to encoder part

#

explain it also please!!

#

and x ??

#

yeah I know thaty

#

okay now encoder right!

#

so what is out output is being now??

#

what it represents to us?

#

x_chw

#

it's tensor !

#

but for what

#

now encoder how can I implement that with out x_chw?

#

quick question!!
x_chw = output.reshape(32, 4, 4)

does this 32 channels represent as masks?
I mean the same masks which acts as features extractors from original image?

#

so is it we got 32 feature_masks?

#

I am wrong I guess?, because we have just converting 4, 4 into 32 channels!!

#

yeah so 4, 4 rectangles ( feature masks )

#

and they are 32

#

how can I implement this?

#

in pytorch

#

heh??, but on what size?

# 
X, K = X.reshape(1, 1, 2, 2), K.reshape(1, 1, 2, 2)
tconv = nn.ConvTranspose2d(1, 1, kernel_size=2, bias=False)
tconv.weight.data = K
tconv(X)```
#

we still need K value..

#

so from where we can get this value?

#

K value

#

we need this to pass into t_conv

#

here X is out x_chw

#

right?

#

hmm, got it

#

yeah

#

what about kernel size , we used 2x2 in conv2d !!

#

okay and what about input and output

output is fixes now, because we are using 32 channels right?

#

and in out input image the channels are 3

#

why?

but we reshaped into 32 right?

#

so we got different 32 feature_mask

#

okay so only reason we converted them to convert vector into tensor

#

what about stride shall we keep it as 1?

#

yeah sort of

#

but those were not images, those were feature masks

#

okay!

#

stride = ??

#

explain latent space in simpler words!! please

#

okay whatever!!, let's not focus onto that

#

m = nn.ConvTranspose2d(3, 1, 2, stride = 1)

#

what about this?

#

series? but why? and of what size ?

#

okay okay got it

#

we need final image as 128x128

#

so I need some layers probably again 3 to convert this into 128

#

hehehe

#

is decoder needed for out task?

#

now I am gonna confuse you so be ready

# 
tconv1 = nn.ConvTranspose2d(3, 1, 2, stride = 1)```
#

here we define fist tconv

#

with 3 input and 1 output!!

#

wait litterally 3 input channels are goint to convert into 1

#

where the heck maths gone?

#

1 output channel becaus we convert 32 of "x_chw" into 1 because it make sense

#

so how can I define input and output channels of tconv2

#

because out goal is 128

#

yeah I am keeping this on second tab

#

it's number of input channels in image

#

so for tcon1 it's just our raw original image which has 3

#

3, 128, 128

#

ohh for conv1 ??

#

shit again wrong

#

then for what

#

1, 512

#

which we convert into 32, 4, 4

#

okay

#

make sense now!

#

arite ? what's this now?

#

πŸ˜‚

#

okay now next?

#

we pass this shape to tconv1

#

right?

#

we are decoding now!

#

are you still calulating?
let's do parallelly!!

#

GPT wrote a code for this !!

# 
self.tconv1 = nn.ConvTranspose2d(32, 16, 3, stride=2, padding=1, output_padding=1)
        self.tconv2 = nn.ConvTranspose2d(16, 8, 3, stride=2, padding=1, output_padding=1)
        self.tconv3 = nn.ConvTranspose2d(8, 3, 3, stride=2, padding=1, output_padding=1)
#

wtf did he wrote?

#

converting 32 into 16 into 8

#

and our goal is to get 128

rich moth
# unkempt apex ```py self.tconv1 = nn.ConvTranspose2d(32, 16, 3, stride=2, padding=1, output_pa... 
import torch.nn as nn

class Decoder(nn.Module):
    def __init__(self):
        super(Decoder, self).__init__()
        
        # Assuming the encoder output is a vector of size 512
        self.fc = nn.Linear(512, 32 * 4 * 4)  # Fully connected layer to reshape the vector

        # Transposed convolutional layers to upsample the data
        self.tconv1 = nn.ConvTranspose2d(32, 16, 3, stride=2, padding=1, output_padding=1)
        self.tconv2 = nn.ConvTranspose2d(16, 8, 3, stride=2, padding=1, output_padding=1)
        self.tconv3 = nn.ConvTranspose2d(8, 3, 3, stride=2, padding=1, output_padding=1)

    def forward(self, x):
        # Reshape the vector back to a 4D tensor
        x = self.fc(x)
        x = x.view(-1, 32, 4, 4)  # Reshape to (batch_size, 32, 4, 4)
        
        # Pass through transposed convolutional layers
        x = self.tconv1(x)
        x = self.tconv2(x)
        x = self.tconv3(x)
        return x

# Example usage:
# Initialize the decoder
decoder = Decoder()

# Example input (compressed representation from encoder)
input_vector = torch.randn(1, 512)  # Batch size of 1

# Decode the input vector
output_image = decoder(input_vector)
print(output_image.shape)  # Should print torch.Size([1, 3, 128, 128])
#

Nope its wrong it prints 1 3 32 32

unkempt apex
#

for reference only

unkempt apex
rich moth
#

They are good but he need to refine them, I often point out mistakes to it.

unkempt apex
#

atleast tconv are good

#

I got it now,
our original's image was of 3 channels

which out encoder converts into 32 to create feature_masks

now out task being decoder is to reverse this thing

#

yeajh!! I got it now!

#

now shape is out problem

#

lemme try some values

rich moth
#

What was it?

unkempt apex
#

wait lemme try atleast!!

#

you can if you!!

rich moth
#

It works! ```import torch
import torch.nn as nn

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

    # Assuming the encoder output is a vector of size 512
    self.fc = nn.Linear(512, 32 * 4 * 4)  # Fully connected layer to reshape the vector

    # Transposed convolutional layers to upsample the data
    self.tconv1 = nn.ConvTranspose2d(32, 16, 4, stride=2, padding=1, output_padding=0)
    self.tconv2 = nn.ConvTranspose2d(16, 8, 4, stride=2, padding=1, output_padding=0)
    self.tconv3 = nn.ConvTranspose2d(8, 4, 4, stride=2, padding=1, output_padding=0)
    self.tconv4 = nn.ConvTranspose2d(4, 3, 4, stride=2, padding=1, output_padding=0)
    self.tconv5 = nn.ConvTranspose2d(3, 3, 4, stride=2, padding=1, output_padding=0)  # Final layer to reach 128x128

def forward(self, x):
    # Reshape the vector back to a 4D tensor
    x = self.fc(x)
    x = x.view(-1, 32, 4, 4)  # Reshape to (batch_size, 32, 4, 4)
    
    # Pass through transposed convolutional layers
    x = self.tconv1(x)
    x = self.tconv2(x)
    x = self.tconv3(x)
    x = self.tconv4(x)
    x = self.tconv5(x)  # Adding the final layer
    return x

Example usage:

Initialize the decoder

decoder = Decoder()

Example input (compressed representation from encoder)

input_vector = torch.randn(1, 512) # Batch size of 1

Decode the input vector

output_image = decoder(input_vector)
print(output_image.shape) # Should print torch.Size([1, 3, 128, 128])


How to talk it through
unkempt apex
#

please enable that py!! in

#

code

rich moth
#

I guess it needed another laye

unkempt apex
#

in decoder?

rich moth
#

layer, geez my typing is lazy today :\

unkempt apex
#

lemme understand this code first

spring field
unkempt apex
#

self.fc = nn.Linear(512, 32 * 4 * 4) # Fully connected layer to reshape the vector

why you add this

#

we can done this in outside of class

# 
model = LaneDetectorCNN()
output = model(input_image)
print(output.shape)

# so here we are converting simple 1d vector (output) into tensors
# so 32 channels tensors size of 4x4
x_chw = output.reshape(32, 4, 4)
```like this
rich moth
#

If you omit the fully connected layer, the 512-dimensional vector cannot be directly reshaped into a 4-dimensional tensor because the dimensions would not align correctly. Transposed convolutional layers cannot operate on a 1-dimensional vector; they need 4-dimensional input.

spring field
#

well, a linear layer would also learn stuff

rich moth
#

That's what Im learning.

unkempt apex
#

interesting

rich moth
#

AI is great for alot of things, its been an awesome tutor. It might not always be right, but you paste entire documents in it and figure shit out together.

unkempt apex
#

then you will misunderstood

#

and in lot of time, it becomes habit to just copy and paste from that GPT

#

but eventually it not works here!

rich moth
#

I agree its a double side blade, it has pros and cons

unkempt apex
#

that's why I ask the same GPT to explain same code line by line and then I again read some artciles arount it

spring field
#

why would you even consider using cgpt for this pg_shocked

unkempt apex
rich moth
#

Nah, its a tool like anything else. If pride keeps you from using it, youll get left in the dust eventually.

unkempt apex
#

why this reaction?

spring field
#

cuz the pride month

rich moth
#

lol nice

iron basalt
#

Pride mentioned.

unkempt apex
#

pride month? new to me

#

need to search

spring field
unkempt apex
#

ohh it's nice btw!

#

we are getting distracted focus on code now

spring field
unkempt apex
#

πŸ˜‚

#

now I am confusing which lines to add in my existing decoder code , need to read now

#

so what do you think , shall I keep that as it is ( outside class )

or create a layer for just reshaping?

spring field
#

I concur

unkempt apex
#

in decoder?

#

I don't care if our output image is good with lane markings

#

which will eventually take some time to re-adjust our hyper parameters

rich moth
unkempt apex
# rich moth

shit you ask for this also!! but quite nice output

rich moth
#

lol

#

I dont fuck around, I got all 3

unkempt apex
#

but yeah he is right

rich moth
#

Let me ask

unkempt apex
rich moth
#

lol

unkempt apex
#

ask here lol!πŸ˜‚

spring field
#

latent spaces and stuff, he he

unkempt apex
#

self.tconv4 = nn.ConvTranspose2d(4, 3, 4, stride=2, padding=1, output_padding=0)
self.tconv5 = nn.ConvTranspose2d(3, 3, 4, stride=2, padding=1, output_padding=0)

can you explain about this/

spring field
#

it's basically upscaling

unkempt apex
#

we go from 16 to 8

#

and then 8 to 4

#

okay so we need in multiplies in 2?

#

yeah I use only 3

#

but here he uses 4th one

#

which converts 4 into 3

#

so should I use 4 in decoder??

#

in encoder
we did
3 -> 16
16 -> 32
32 -> 64

#

so what about decoder now?

#

we have in 32 shape already

#

so
32 -> 16
16 -> 8

#

yeah that's what I am thinking

#

heh?

#

so again (32, 4, 4) ??

#

then?

#

and with 2x2

rich moth
#

Im trying to understand this too.

unkempt apex
#

yeah becaue then it produce 256

#

torch.Size([1, 512])

#

yeah it's before

#

hey?

# 
model = LaneDetectorCNN()
output = model(input_image)
print(output.shape)

look at this

#

because it's vector which our last layer gives us

#

how can I print that now?

# 

  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
```you mean here?
#

that's the catch!

rich moth
#

print the shape of x?

#

oh

#

Oh i see

spring field
#

(also x_bchw)

#

(why is it so tiny?)

rich moth
#

sorry couldnt help myself

spring field
#

is X referring to the visual structure there?

#

so, with that latent embedding, how exactly do you train word embeddings with it? using sth like cosine similarity as the loss function? like how do you bring the word vector closer to the image vector?

rich moth
#

cosine similarity, dot product, etc

spring field
#

like, do you embed the words and images in the same space?

#

I get that, but how do you tie in labels with that?

rich moth
#

there are different encoders and decoders models that use different similariys

#

Wouldnt those want to match the hidden embedding layers?

spring field
#

or prompts, yeah, like what vq-vae does (realizes that I could just go and find the paper on those...)

#

lol

rich moth
#

I was asking AI bout our chat and it wrote the whole dame code what you were asking. It alinged the image and text using a triplet loss function and ensuring silar embeddings

#

I ran it seems to work

image.png
#

Honeslty! There have been numerious "AH, HA" moments where I knew it was wrong and explained it, sure its not right, but these conversations is really how I pick up a lot. Critical thinking is key, i get what you mean.

#

Right, but understand what its spewing out is important otherwise were gonna have generations of braindead people lol

#

copilot is no bueno. Claude Opus. You should give that one a shot., or Just clade in general.

spring field
#

Deep Metric Learning, but like, VQ-VAE apparently indexes the embedding table, it seems I'll have to reread it a couple times till I understand it, lol, but yeah, it's not exactly triplet loss

#

well, yeah, but VQ-VAE tries to combine an image embedding and token embedding

#

I think... I'm still reading the dam paper, lol

#

wasn't that your plan with that new approach to transformers ducky_sus

#

SOTA go brrrt

rich moth
#

Look at this

image.png
#

This describe my entire system I build, I automatically load, preprocess and indexed datasets directly from hugging face into the ESDB, everything is embedded in a 768 dim, so I use models from hugging face that use the same dim.

#

Heres the diagram of it

pipeline_diagram.png
spring field
# rich moth Look at this

interesting, this is actually what I considered at the very first when I mentioned embedding images and labels, I just knew that VQ-VAE doesn't do that

#

so you can try embedding an image and a token in the same latent space and bring them closer and such

rich moth
#

Nice! I build a Conversation DB, and Two types of memory for summarys of conversations and and transcripts from other models, it gets ranked by scores, and the top answers get embedded back ito the conversation index.

#

I use cosline similarys on the memory to extract knowledge its learned from querys.

#

Combine with its ability to seach online and extracts information the same way, so its a type of transfer learning, but I swear the more I play with it the more it feels recursive learning too.

#

Heres what I mean.

image.png
#

I will put it up on github in a couple days, maybe you guys take a look at it

#

That would awesome. I would need a hand with that though, I made a front end and backend for it, just really simple but it worked locally.

#

great idea! let me go research this ill be back

hollow escarp
#

Hi, im having an issue when converting my onnx model to rknn format. I need to make my model produce smaller output, how i can do this?

image.png
#

Thats my model

image.png
rich moth
#

You guys gave me an idea from earlier. I combined VQ-VAE and CLIP to make a tool for learning and leveraging discrete latent representations of images that are aligned with textual descriptions.

I a full model ready to train on it. I think I got the learning rate dailed in after playing with it for a bit.

image.png
#

Ill share it with you guys if you want.

rich moth
#

I gotta take a break. I got this far but I got fix the padding issue.

image.png
#

So it uses (VQVAE) with a CLIP model to work with both images and text. The VQVAE compresses images into compact representations and then reconstructs them, while the CLIP model ensures that these image features align well with corresponding text descriptions. During training, it optimizes a loss that includes image reconstruction error, quantization loss, and a CLIP loss measuring image-text similarity. I got it loaded with the flickr30k dataset.

keen comet
rich moth
keen comet
spring field
keen comet
spring field
#

what's silly?

rich moth
#

Decreasing loss durning training is a good sign.

#

Usually the first epoch from my understanding should be taken with a grain of salt.

keen comet
rich moth
#

For my understand but please correct me if im wrong

keen comet
#

I don't think it's normal for my computer to random restart from nowhere when put in sleeping mode

rich moth
#

Maybe a device is waking it up

keen comet
spring field
rich moth
#

The model is train to learn off the data, theres nothing to learn first until it process a few epochs.

spring field
spring field
rich moth
#

Not sure yet,, I have it setup for test and validation but I need to fix the padding error.

#

Technically your right but im talking about measuring it learning progress.

vernal valve
#

would it be better to fine-tune an LLM with my own data (dynamic data that can change every year or so) or just do RAG on it?

#

it's a customer support AI

spring field
#

from my limited knowledge of this from an API-usage standpoint, it seems that fine-tuning is quite an expensive process actually and you say it's dynamic data... that likely means you'd have to retune the model every year, that's certainly not cost effective the way I see it. I'd suggest RAG

vernal valve
rancid zealot
#

i have the following einsum:
np.einsum('nsd,jd->njs', x, y, out = z)

d = 10, n = 1500, s = 100+, j = 200.

It's slow, because it's single-threaded. is there a multi-threaded matrix operation i could do that would accomplish the same thing?

#

i know matmul is multithreaded, but it seems difficult to accomplish that operation with matmul.

rich moth
spring field
# rancid zealot i have the following einsum: ```np.einsum('nsd,jd->njs', x, y, out = z)``` `d =...

so, you can do this

import numpy as np


x = np.arange(1500 * 100 * 10).reshape((1500, 100, 10))
y = np.arange(200 * 10).reshape((200, 10))

e_sum = np.einsum('nsd,jd->njs', x, y)

dot_transpose = (x @ y.T).transpose(0, 2, 1)  # <- this

assert e_sum.shape == dot_transpose.shape
assert np.array_equal(e_sum, dot_transpose)
print(e_sum.shape, dot_transpose.shape)

but it didn't really seem that much faster

#

nvm, it is faster

#

!e

from timeit import timeit

import numpy as np

N = 10

# reduced to 150 because otherwise this would take forever, lol
x = np.arange(150 * 100 * 10).reshape((150, 100, 10))
y = np.arange(200 * 10).reshape((200, 10))

print(timeit("""
np.einsum('nsd,jd->njs', x, y)
""", number=N, globals=globals()))

print(timeit("""
(x @ y.T).transpose(0, 2, 1)
""", number=N, globals=globals()))
arctic wedgeBOT
rancid zealot
spring field
rancid zealot
#

this is faster than my original solution, so i'm absolutely implementing it.

spring field
#

apparently np.dot has an out option

#

that doesn't change anything though

frigid cove
#

This is a beginner question:
Can I fine tune a ViT model one time with one dataset, and fine-tune it the ohter time with another dataset that has different labels?

unkempt apex
#

My internet was gone , that's why I went offline , now need to go college then will do this!!

spring field
# frigid cove This is a beginner question: Can I fine tune a ViT model one time with one datas...

well, you have to combine the labels (like, if they are different then label with index 1 from the first dataset is gonna be with index 1, but label with index 1 from the other dataset can't also be index 1 in the embedding table (unless they're the same label)), otherwise you'll mess up the embedding table, does that make sense?
as for the images, it probably doesn't matter that much, though you probably are gonna need to match their sizes if they're not the same

frigid cove
spring field
#

oh wait, you're not gonna embed the labels anyway... this is for classification, right? does same gestures also mean that the labels are identical? cuz like obviously different gestures could mean the same thing... hmmm
I think to be safe, it'd be better to just concatenate the labels, so basically, when you train both datasets your output is the concatenated array of labels from both datasets, for example

ds1_labels = [1, 2, 3]
ds2_labels = [1, 2, 3, 4]
combined = [1, 2, 3, 1, 2, 3, 4]

so when you train them, your output is of the combined size, that's the final output after softmax, it's 7 values in this example, the first 3 are for the 1st dataset and the next 4 are for the 2nd dataset, but otherwise you proceed as normal, you just need to know that when you're actually predicting and you get the max value from that array and it's the 2nd dataset, you need to subtract the number of labels in the first dataset from that max index to get the index for your actual label
for example, you're training the second dataset and this is your output for an image
[0.0, 0.0, 0.1, 0.1, 0.7, 0.0, 0.1]
the max index here is 4, but since it's the combined labels, you need to subtract 3, so you get 1 as the index for the label in the 2nd dataset

now, a concern of mine here is what to do if it predicts the label from the other dataset... now, one option I guess would be to assign a much greater loss to such cases

honestly, this is a bit of a tough one (btw, do take what I say with a bit of a grain of salt, not an expert yet 😁)

#

I mean, tbf, this does kinda make sense, it's basically the equivalent of merging the two datasets into one

rich moth
#

I finally got it training. I did 20 epochs might take a bit but we can test the results

image.png
rich moth
#

OMG! Its working better than I though lol

image.png
#

so image encoder of the VQVAE first extracts features from the input image. These image features are then concatenated with the text embedding obtained from CLIP, forming a joint representation. This joint representation is fed into the VQVAE's quantization layer and subsequently used for image reconstruction.

#

Wow.. look at the difference in loss.

image.png
#

holy shit lol

spring field
#

loss on its own doesn't convey much of a story, you should add some additional metrics and some sample visualisations and stuff, yk, so that you can somewhat see what the network is actually doing

rich moth
frigid cove
#

i didnt understand quite well the merging of the labels. the 4 would represent the label that is not from the other dataset, right?

#

i mean the thing is that it has the same amount of labels in both dataset, but what it's different is the type of sign language

#

mmmm

#

so the amount of labels would be 48?

#

jesus fucking christ is that feasible stiill

spring field
#

wdym?

frigid cove
#

i mean those are too many labels

spring field
#

not really, no

frigid cove
#

what?

spring field
#

why would it be too many?

frigid cove
#

what do you mean? so it's not feasible to do this? like, actually train the thing?

frigid cove
spring field
#

if you have enough samples for each class I don't see the issue

#

like language models have vocabularies that are in the 10s of thousands of tokens and stuff

#

so anyway, basically my idea boils down to essentially merging the two datasets and adding those prefixes to differentiate the datasets
but you can train one dataset first and then the other one, you just have to then merge the labels of both datasets

#

by merging here I just mean concatenation

#

like it's easier to visualize it as just merging the two datasets completely, adding those prefixes to the labels, so now you have like all the labels you could possibly predict, but then you first train with only one dataset and all of those combined labels and afterwards you train with the other dataset and again all the combined labels

#

now, as I understand you just want to do simple classification? so basically you give it an image of a gesture and it tells you what the gesture is, right?
I mean, you said you want it to identify the differences, but I assume it just means that you want the model to be able to differentiate different gestures itself

rancid zealot
#

_<

spring field
#

huh

#

what exactly were you doing?

#

cuz that is not an issue with memory

rancid zealot
#

the operation i specified before, with the dimensions i'm working with.

#

should become a 1566 by 240 by 200 output array

spring field
#

okay, that's a different shape, can you provide all the specs again? the shapes and what you're doing and like the code that causes this error and such

rancid zealot
#

np.einsum('nsd,jd->njs', X, Y, out = Z)

with n=1566, j =200,s=240,d=10. (actual dimensions now; i was guessing before).

#

You had suggested (x @ y.T).transpose(0, 2, 1)

#

i tried implementing with np.matmul, which @ calls (i think), and it gave that error. so i tried @ directly, and it still gave that error.

#

so weird.

#

er..to be clear, einsum works but is slow. i was hoping to replace with a more performant option.

spring field
rancid zealot
#

uh, the error message i posted.

spring field
#

cuz it works just fine for me

#

!e

import numpy as np

x = np.arange(1566 * 240 * 10).reshape((1566, 240, 10))
y = np.arange(200 * 10).reshape((200, 10))

e_sum = np.einsum('nsd,jd->njs', x, y)

dot_transpose = (x @ y.T).transpose(0, 2, 1)  # <- this

assert e_sum.shape == dot_transpose.shape
assert np.array_equal(e_sum, dot_transpose)
print(e_sum.shape, dot_transpose.shape)
arctic wedgeBOT
# spring field !e ```py import numpy as np x = np.arange(1566 * 240 * 10).reshape((1566, 240, ...

:x: Your 3.12 eval job has completed with return code 1.

001 | Traceback (most recent call last):
002 |   File "/home/main.py", line 6, in <module>
003 |     e_sum = np.einsum('nsd,jd->njs', x, y)
004 |             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
005 |   File "/snekbox/user_base/lib/python3.12/site-packages/numpy/core/einsumfunc.py", line 1371, in einsum
006 |     return c_einsum(*operands, **kwargs)
007 |            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
008 | numpy.core._exceptions._ArrayMemoryError: Unable to allocate 573. MiB for an array with shape (1566, 200, 240) and data type int64
spring field
#

good one... well, anyway, as you can see, that's a memory error

#

you're not getting a memory error, so it's some other thing

rancid zealot
#

to be clear, it worked fine for you? (not the eval bot)

spring field
#

yes

rancid zealot
#

weird.

#

maybe it's my version of numpy?

#

i'm using an older version for compatibility with tensorflow.

spring field
rancid zealot
#

hmm

#

yes. then that opens up another question.

#

my issue is my implementation, i guess.

#

wow, damn it numpy

#

type(X) returns np.matrix

#

f-ing nonsense.

#

putting in np.asarray(X) in place of X makes it work

rich moth
#

try float32 instead of 64

rancid zealot
#

wish python actually did proper typechecking. all my functions are properly typed, but the only good that does is helping linting.

spring field
rancid zealot
#

data is now coerced to an array on initialization of the model

rich moth
#

Use like pycharm iin visual studio code.

#

what i use.

rich moth
rancid zealot
#

does np.ndarray.reshape cost time?

rich moth
spring field
spring field
rich moth
#

Ok ill work on, thanks

spring field
#

cuz for all I know, it could've been terrible initial values or some stuff like that, or the accuracy might be only say 60% despite the loss
ofc, it's a good sign that the loss is dropping, but it doesn't really convey much on its own

rich moth
#

roger makes sense, we need more data!

long robin
#

Im learning all the traditional ML algorithms in Scikit learn.

We do standardization of the X_train and X_test in the start, before training.

Once the model is trained, once we have tuned the hyper parameters to get the best result and error metrics, when we finally go for training the final model on the entire dataset, do we need to fit the final model with the scaled X too?

And if I am deploying the final model, then do I need to deploy the scaler object too?

hollow escarp
simple pagoda
#

Will market still be demanding AI ML engineers after 20 years

spring field
# simple pagoda Will market still be demanding AI ML engineers after 20 years

20 years? I mean, if there will still be a market... or humans for that matter left, maybe, it's really hard to say, 20 years is quite a long time
suppose the cycle continues as it has been going thus far and assuming we don't all go extinct, there will likely be demand for AI engineers, though it may be different AI, not necessarily ML

rich moth
#

Compared to the first image its really coming into shape

spring field
#

looks fantastic

rich moth
#

thank you, im going hit the sack. ill let you guys know how it did it the morning, thanks for the advice

hollow escarp
rare fog
#

Which career?

lament moat
#

!rule ad

arctic wedgeBOT
#
Rules

6. Do not post unapproved advertising.

wooden sail
#

@past meteor what's the name of that book for ML + python that you often recommend? a friend is asking for a suggestion and that's the first that came to mind

real kayak
#

where can I find some assistance on ai training task?

serene scaffold
long robin
#

I am a good maths student. I am learning all the traditional ML algorithms and their practical implementation with sklearn (I have learnt numpy, pandas, matplotlib, seaborn). Even if I understand the underlying maths properly, at some points I am so confused or burnt out. I'm starting to doubt myself.

pithink ~ Just wanna code real awesome nueral networks in tensorflow one day and build do projects... but now i doubt myself so much

long robin
#

The way I learn it
Hmm... πŸ€”
I have a course of Jose portilla....
I learn stuff from there...
See code on kaggle
I also refer to the documentation...

That's how I learn... And I do try to implement that into real world dataset too

Is there any other or better way I could learn the ml stuff??? Plz tell πŸ˜“

#

Okieee

#

Alright πŸ‘
One last question...πŸ˜“πŸ˜…

A few guys from my college are directly getting into tensorflow and neural networks....

They just know the theoretical and mathematical part of all the traditional machine learning algorithms... Is it possible to go into tensorflow just by knowing that much knowledge, without even implementing stuff with sklearn?

wooden sail
long robin
past meteor
lapis sequoia
#

Hello where do I learn Neural networks? like youtube or textbooks? recommend me some pls.
I have very basic understanding of NNs and some common ML algos (Thats just my current level knowledge)

past meteor
spring field
#

!res
I don't exactly feel qualified handing out such advice, but if you want to go into software development I guess you can start by learning a language, like Python, see the resources in the linked page
for career-related questions you can also go to #career-advice

arctic wedgeBOT
#
Resources

The Resources page on our website contains a list of hand-selected learning resources that we regularly recommend to both beginners and experts.

lapis sequoia
#

thank you

#

!!

long robin
#

But don't u need to learn those generic ml algorithms...first ...?

#

Really? ☠️☠️☠️

#

U sure?

#

Okk
☠️ Thank for the info

#

Im new in this field... That's I don't have that much knowledge ☠️☠️

#

Ik I sound so stupid but yeah

#

Yeah
It learnkng Scikit learn was not a time waste ...ik that... But still I could have started with tensorflow

spring field
#

(btw, instead of tf, go with pytorch)

long robin
spring field
#

by what?

#

ah, sure, at some point that will happen probably, yeah, le cycle of software or whatever, lol

past meteor
#

It'd be hard for that godlike API to exist in Python but there's a lot of possibilities in dependently typed languages

#

Or ones with very strong compile time programming

river cape
#

Does anyone as to why do we use a bias term in perceptron?

past meteor
#

otherwise your decision boundary would have to go through the origin (0,0)

#

Let's say you're measuring happiness. On an average day you're 6/10 happy. Certain factors make you happier (eating your favourite food) and others make you unhappy. The bias here is 6