data-science-and-ml

[serene scaffold]

what type is model and what is the full import statement for it?

[gleaming osprey]

model is of type Blip2ForConditionalGeneration. Here is the code that imports it: ```py
model = Blip2ForConditionalGeneration.from_pretrained(
"Salesforce/blip2-flan-t5-xl-coco", device_map="auto", low_cpu_mem_usage=True
)

[serene scaffold]

that's the code that defines model, but it's not an import statement. but I was able to find the documentation for that class without it.

[gleaming osprey]

oh I'm sorry

from transformers import Blip2Processor, Blip2ForConditionalGeneration, AutoModelForSeq2SeqLM, AutoTokenizer, BitsAndBytesConfig

[serene scaffold]

this person got a similar error: https://github.com/huggingface/transformers/issues/22563

[abstract wasp]

Im kinda lost on how I can go about in getting into either the space industry or healthcare w ml/ai. For health, I found some courses online by Stanford to get some certifications. Do you guys think I’m on the right path? I currently only have my bachelor’s in ml and 2 years of experience as a research assistant/data analyst.
Is anyone here in either industry?

[serene scaffold]

your best bet might be to get a masters.

@gleaming osprey try setting ignore_mismatched_sizes=False when you call Blip2ForConditionalGeneration.from_pretrained and see if you get a different error

also, the problem is arising when trainer tries to save the model.
there's also something about shard. I wonder if device_map="auto" is the problem. do you have the whole model on one device?

in either case, I have to run. I hope this was more helpful than nothing.

is that the standard I'm holding myself to now? wow.

[abstract wasp]

I would do that but it’s too expensive 🤡🤡😢

[molten elk]

Can anyone recommend a resource for self study? I can already code since I've been doing Web Dev for some years now (but in Java and React mostly)

[gleaming osprey]

Thanks for the help nonetheless! The issue is with Checkpointing and sharding and offloading some of the model (purpose of module map), how do i know?

1. I looked through the transformers source code for clues
2. If i disable checkpointing i get no errors

Thing is, i want checkpointing.

[weary timber]

i coded a neural network and when i train it he values come out as inf

weights and biases

what do i do

[wooden sail]

the first thing would be to play around with the learning rate. try making it smaller

[weary timber]

did that

[wooden sail]

did you make everything yourself from scratch or are you using something like pytorch? cuz another issue might be just using the wrong derivatives

[weary timber]

myself

from scratch

[wooden sail]

and you're sure the math is correct?

[weary timber]

if you want i can send the formulas i used but idk if you can understand it

i followed this video wait

https://www.youtube.com/watch?v=tIeHLnjs5U8&list=PLZHQObOWTQDNU6R1_67000Dx_ZCJB-3pi&index=4

[wooden sail]

i mean, the math and formulas are easy, the question is whether they are coded correctly 😛 have you tried some small scale toy scenarios that you can verify by doing the math on paper?

[weary timber]

what does small scale toy scenarios mean

[wooden sail]

like just one layer with few parameters

[fallow coyote]

currently self learning ML using Hands On Machine Learning but Im on Chapter 2 and Im not really enjoying it as I don't fully understand the functions I'm using when I go along with the guided projects. What 3 ML techniques should I focus for now as a beginner? I know linear regression is one of them but what other 2 should I work on?

[wooden sail]

something you can realistically do on a piece of paper

[weary timber]

did it on paint does that count

[wooden sail]

if you did the full math with numbers and used it to corroborate the results your code spits out

you can verify every number in your code vs calculations you make on paper

[weary timber]

the problem seems to be in my sigmoid deriv function

i tried without it and it works

not true values but no inf or nan

lemme send that function rq

def sigmoidDeriv(x):
return x*(1-x)

[wooden sail]

that's not the correct derivative of a logistic sigmoid, is it?

[weary timber]

i dont think so

[wooden sail]

shouldn't it be f(x) * (1 - f(x))? where f(x) is a sigmoid

[weary timber]

i copy pasted it from somewhere since i had more stuff to do

[weary timber]

yea bro im questioning that too lol

it should include sigmoid too

lemme fix that

@wooden sail sorry for the ping but i wanna ask something

how much should it take in an average pc to train with 60000 img dataset

[wooden sail]

depends on what size the images are, how big the network is, and how good your computer is

[weary timber]

mnist

[wooden sail]

60k isn't that many, anywhere from like 10 minutes to "it's not worth it" depending on if you have a gpu or not

[weary timber]

its an average pc

[weary timber]

oh ok

[wooden sail]

idk what "average pc" is

without a gpu on an old processor, probably several hours

[weary timber]

oh ok

[earnest hull]

can someone help me with a python project in data mining. plz dm me

[serene scaffold]

Hello, people won't want to dm you to figure out if the question is something they can help with. Please ask your whole question in this channel.

[earnest hull]

nvm i got it

[deep egret]

guys, what do you think about applyting ML/DL into cybersec?

[glad bane]

I need help. I am currently studying Multimodal RAG: Chat with Videos
https://www.deeplearning.ai/short-courses/multimodal-rag-chat-with-videos/
In the course, there is a use of bridgetower-large-itm-mlm-itc using predictionguard. When I want to try it on a local laptop, following all the examples in the course, I am currently working on the chapter L4_Multimodal Retrieval from Vector Stores. I am having trouble with bridgetower-large-itm-mlm-itc using predictionguard, which I do not have an API KEY for. So I searched for information on huggingface and found https://huggingface.co/BridgeTower/bridgetower-large-itm-mlm-itc. But the next problem I encountered is how do I make a function to solve this problem?

# helper function to compute the joint embedding of a prompt and a base64-encoded image through PredictionGuard
def bt_embedding_from_prediction_guard(prompt, base64_image):
    # get PredictionGuard client
    client = _getPredictionGuardClient()
    message = {"text": prompt,}
    if base64_image is not None and base64_image != "":
        if not isBase64(base64_image): 
            raise TypeError("image input must be in base64 encoding!")
        message['image'] = base64_image
    response = client.embeddings.create(
        model="bridgetower-large-itm-mlm-itc",
        input=[message]
    )
    return response['data'][0]['embedding']

Can you suggest how I should modify the function to successfully use bridgetower-large-itm-mlm-itc locally?

[surreal rock]

Q5. Write a program to filter count vowels in the below-given string.
string = "I want to become a data scientist"

[surreal rock]

..

[odd stratus]

!e

string="Text goes here"
count = sum(c in 'aeiou' for c in string.lower())

[arctic wedge]

:white_check_mark: Your 3.12 eval job has completed with return code 0.

12

[dreamy isle]

!e ```py
string="I want to become a data scientist"
count=sum(map(string.count, "aeiou"))
print(count)

ooo

[odd stratus]

forgor the +1

?

[dreamy isle]

no
i see what's wrong

!e ```py
string="I want to become a data scientist"
count=sum(map(string.lower().count, "aeiou"))
print(count)

[arctic wedge]

:white_check_mark: Your 3.12 eval job has completed with return code 0.

12

[odd stratus]

OH WAIT, nevermind i forgot to make it lower as well lmao

[odd stratus]

lmao, i thought it was just from off by one or smthn

[deep veldt]

im confused on how to use yolo for training, what do i write in the .yaml file??

[inland totem]

Guys I am trying to train a model to respond with a structured json response
specifically:

import json
from datasets import load_dataset
from transformers import (
    AutoTokenizer,
    AutoModelForSeq2SeqLM,
    Seq2SeqTrainer,
    Seq2SeqTrainingArguments,
    DataCollatorForSeq2Seq,
)

# Load Flan-T5 model and tokenizer
MODEL_NAME = "google/flan-t5-large"
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
model = AutoModelForSeq2SeqLM.from_pretrained(MODEL_NAME)

# Load dataset
dataset = load_dataset("json", data_files={"train": "train.jsonl", "validation": "validation.jsonl"})

# Preprocess the dataset
def preprocess_function(examples):
    inputs = examples["input"]
    outputs = examples["output"]
    
    # Convert output to a string if it's a JSON object or list
    outputs = [json.dumps(output, ensure_ascii=False) for output in outputs]
    
    # Tokenize inputs and outputs with padding and truncation
    model_inputs = tokenizer(inputs, truncation=True, max_length=512, padding="max_length")
    labels = tokenizer(outputs, truncation=True, max_length=512, padding="max_length")
    
    # Add labels to model inputs
    model_inputs["labels"] = labels["input_ids"]
    return model_inputs

# Tokenize the dataset
tokenized_dataset = dataset.map(preprocess_function, batched=True)

# Data collator to handle dynamic padding
data_collator = DataCollatorForSeq2Seq(tokenizer=tokenizer, model=model)

# Training arguments
training_args = Seq2SeqTrainingArguments(
    output_dir="./flan-t5-greek",
    evaluation_strategy="epoch",
    learning_rate=5e-5,
    per_device_train_batch_size=1,  # Reduce batch size for memory efficiency
    per_device_eval_batch_size=1,
    num_train_epochs=3,
    weight_decay=0.01,
    gradient_accumulation_steps=4,  # Simulate larger batch size
    save_steps=5000,
    save_total_limit=2,
    predict_with_generate=True,
    fp16=True,  # Use mixed precision for speed
    logging_dir="./logs",
    logging_steps=500,
    report_to="none",
)

# Trainer
trainer = Seq2SeqTrainer(
    model=model,
    args=training_args,
    train_dataset=tokenized_dataset["train"],
    eval_dataset=tokenized_dataset["validation"],
    tokenizer=tokenizer,
    data_collator=data_collator,
)

# Train the model
trainer.train()

# Save the model and tokenizer
trainer.save_model("./flan-t5-greek")
tokenizer.save_pretrained("./flan-t5-greek")

the train.jsonl contains such data as:

{"input": "ΠΛΗΡΗΣ\nΚΥΡΙΟΣ (ΠΟΣΟΣΤΟ:100/100 )\nΟνοματεπώνυμο Φυσικού Προσώπου: REDACTED\nΟνοματεπώνυμο Πατέρα: REDACTED\nΟνοματεπώνυμο Μητέρας: REDACTED\nΗμερομηνία & Τόπος Γέννησης: 99/99/9999 Α.Φ.Μ.:99999999\nΔιεύθυνση: REDACTED, TK: 55555 REDACTED, Νομός: ΑΤΤΙΚΗΣ\nΑ.Δ.Τ.: Σ999999 99/99/9999 Α.Τ. REDACTED\nΑιτία/-ες Κτήσης: ΧΡΗΣΙΚΤΗΣΙΑ\n\n", "output": [{"full_name": "REDACTED", "father_name": "REDACTED", "mother_name": "REDACTED", "date_and_place_of_birth": "99/99/9999", "tax_id": "99999999", "address": "REDACTED", "postal_code": "99999", "location": "REDACTED", "municipality": "REDACTED", "adt": "Σ999999 99/99/9999 Α.Τ. REDACTED", "percentage_ownership_land": "100/100", "ownership_state": "ΠΛΗΡΗΣ"}]}```

tag me if anyone wants to help 😭

I have posted to https://discord.com/channels/267624335836053506/1310574647498117120, (I show the channel later)

[proper meteor]

3B1B!!!!

[weary timber]

3b1b + chatgpt for your questions

better than any free/paid course imo

[proper meteor]

can't disagree with ya mate

[agile cobalt]

ChatGPT is not exactly reliable though
there's a considerable risk of it hallucinating and giving you incorrect answers, so you must double check nearly everything it says

[proper meteor]

Yeah if you're already studying it's common to study from multiple sources

you should like read a book then like ask chatgpt for clarification

[weary timber]

if thats the case im cooked

[wild storm]

hey there guys

I'm looking for some basic information on training AI models

I have 5 days to make an inventory management and automatic reordering system, which takes in a bunch of data points such as:

• current stock amount
• rate of usage
• time to receipt of a new order of supplies
• reliability of specific suppliers (i.e. how long they take to actually deliver supplies versus their delivery estimation)
• historical data about the usage rate of specific items around certain times of year

and make a system that automatically decides how much of a stock to order, and when to order it, to ensure optimal storage costs and no shortages

it seems to me like it would be difficult to manually model a relationship between these variables in a way where I can write an algorithm for it

I'm not looking for anybody to solve my problem, I've just never messed around with AI before and I don't know up from down

If someone could point me in the general direction of the type of AI model that is optimal for this sort of work, I can learn everything myself from there

I just don't know where to start

[serene scaffold]

@wild storm thank you for explaining what you want to do. this is a great amount of detail for a beginner question.
have you heard of forecasting?

[wild storm]

making predictions about future data using the data you have?

[serene scaffold]

right.

[wild storm]

ah, so that's what I need to look into

thank you

[past meteor]

Start out with exponential smoothing or similar

Why? It's very very easy to explain to stakeholders and very easy to implement as well

Afterwards you can gradually add new variables and try different models

You can also get quite far with saying "demand yesterday will be the demand tomorrow"

5 days to come up with this is pretty much impossible though, at this point I'd focus on managing expectations lol

[main coyote]

Hello guys, can someone tell me a site where I can make my AI recognise open eyes/closed eyes by showing him a video/picture?

[shadow viper]

also, i dont know if the tutorial you follow will have everything but make sure you learn about MovingAverage

hello everyone
i'm working on this loan dataset i saw online, where the goal is to identify individuals who are most likely to repay loans while reducing the risk for financial institutions.

going through the data, i dont know how they arrived at Total_Income and its not the summation of ApplicantIncome and CoapplicantIncome.

when i checked the description of the features, they didnt include Total_Income.
could the Total_Income feature be there to throw me off or it means something im not seeing because i think its an error.
Thank you 🙏

[rose heath]

Hi all, I'm sorry if this is not the right sub for this question- but I'm looking for some good yt resources that explain python from scratch from a credit risk standpoint. Google search gave me a few courses but they are too expensive for me right now and yt vids I came across so far are not that helpful. Any suggestions would be appreciated as i have no experience in python or programming. However I have done FRM so I do have a basic level understanding of stats.

[shadow viper]

https://youtube.com/playlist?list=PL-osiE80TeTt2d9bfVyTiXJA-UTHn6WwU&si=oaXo1vh7t6zMp8dt

Corey is good. I didn't use Corey but the few I've learnt from him are alright. straight to the point, audible and visible enough.

[rose heath]

Thanks

[crimson raft]

Hey guys!
I'm doing my Master's in mechanical engineering with applied computer science, and I've got the basics of ML and AI from 2 courses I took during my masters, I also have basic programming skills in python.
Can someone please recommend resources ( Books, papers....etc) to learn PINNs and their programming? I need them for my thesis.
Thanks

[unkempt wigeon]

What have you used in the past?

[crimson raft]

I'm not following, what do you mean?

[unkempt wigeon]

Have you used pytorch?

Or tensor?

[wooden sail]

https://benmoseley.blog/my-research/so-what-is-a-physics-informed-neural-network/ this has a pretty clear explanation and some references

[serene scaffold]

@crimson raft wendigo is a beginner. if you've taken two ML courses, you probably know more than them.

I don't know anything about PINNs--sorry

[wooden sail]

the paper on SIRENs is also fairly clear IMO

https://arxiv.org/pdf/2006.09661 this one

those two were enough for me to set up a PINN for the 1D wave equation on a string

[weary timber]

can someone possibly explain this to me?

cause i cant understand shit from this

(the formula)

[serene scaffold]

do you know about derivatives in general? do you know the derivative of f(x) = x^2 + 3x?

[weary timber]

no :(

[serene scaffold]

okay. put everything you're trying to do with ML on the shelf and learn derivative calculus.

I know that's a big diversion, but it's critical.

[weary timber]

okay

[serene scaffold]

khan academy and 3b1b are your friend.

[weary timber]

yea im looking at 3b1b's channel rn

[serene scaffold]

you need to learn how to calculate the derivative of functions with one variable, and functions with multiple variables (but with respect to one variable at a time)

what you're looking at isn't as complicated as the notation makes it look, once you know what the notation means.

[crimson raft]

Thanks, appreciated.

[crimson raft]

Nice suggestion, but I'd rather recommend cal 1 & 2 Mit opencoourseware on YouTube

[crimson raft]

Nice.
Have you worked on anything else? Are you an engineer/ physicist?

[wooden sail]

with pinns or in general? i'm an engineer, one could say

my grad studies went in interesting directions, but engineering at the end of the day

[crimson raft]

Nice to meet a fellow engineer.
Since you worked on wave equation, I guess you are mechanical engineer.

Pinns just wanna get a heads up before I get involved 😅

[wooden sail]

so that's where it gets interesting

i did telecom for my bsc, communications and signal processing for masters, and doing ultrasound sigproc for phd

i only have tangential knowledge about the physics. just enough to build a not-so-wrong model for the approximate propagation of ultrasound in "well-behaved" solids

i'd say my bread and butter is more the estimation part

but pinns work well for inverse problems/estimation, so

if you close both eyes and wave your hands, sound kinda works like electromagnetism

[crimson raft]

Wow, that really is an interesting turn.
I love electromagnetism and communication, never had the time to dive deep into it

[harsh scroll]

does anyone know how to get tensorflow with gpu (cuda) to work ive tried literally everything. i have the right tensorflow cuda toolkit and cudnn versions installed and wsl and nothing recognizes my gpu

[weary timber]

i dont mind. at the end im learning the math still to understand ai

[unkempt wigeon]

What math do I need to learn to make a neural network sorry

[serene scaffold]

if I tell you, are you going to learn it, or are you going to suddenly go in a different direction again?

[unkempt wigeon]

I will learn it i Believe it maybe linear algebra

Or not linear I'm sorry

[harsh ferry]

How much is an array is important for data science student

[deep veldt]

how can i add more images to a cvat task?

[main citrus]

Please give me an example for losting data because of encoding

[weary timber]

a lot

[serene scaffold]

pretty much everything is represented as an array.

[stone saffron]

aka yes its very important lmao

[cedar sentinel]

Greetings everyone

I'm assigned a binary classification problem and I was wondering if someone can help in the modeling part.

If someone is interested helping me, please let me know 💗

[jaunty helm]

you can just say the problem you have

[cedar sentinel]

Oh ok thanks.

[jaunty helm]

wdym by model? what are you trying to model exactly?
e.g. are you thinking that in a fail, fail, fail, ..., fail, success sequence, the more times it fails consecutively, the more likely that the next time will be a success or something?

[cedar sentinel]

Yeah exactly

[jaunty helm]

also, isn't this just visit_id?

[cedar sentinel]

Yes

[jaunty helm]

then... just put it in as a feature?

if there is such a relationship, then the model will pick up on it

[jaunty helm]

e.g. if you have a tree, then the splits will contain something like

if visit_id > 2:
    ... move to a subtree with higher success rate overall

well, what is it?
I'm also not exactly sure what you mean by "in a node level"

[cedar sentinel]

But what if we can make it more clear

[jaunty helm]

so what you really want is a way to emphasize specific features to the model?
I'm not aware of such methods if they exist tbh, if anything the model should pick up on these relationships on their own if these relationships indicative enough

your visit_id already encodes the information of how many times a specific task failed to complete

well, given a task with visit_id == 2, can you ensure that you'll have access to the same task, but when it was visit_id == 1 and visit_id == 0?
not only for the training data btw, any future data you need to perform classification on as well

2nd thing to consider, if you know for sure that engineers in previous attempts always have lower engineer_lvl, then including any feature regarding this is just redundant
e.g. a column is_previous_engineer_lower_level will always be True 100% of the time if the above assumption holds, thus providing no valuable information

sure, the question is do you have it?
given a task with visit_id == n, will you have access to data that details what happened before that?

[cedar sentinel]

Yes if only this wasnt the very first attempt

[jaunty helm]

then what you're doing is a pretty nice approach already
include features from previous failed attempts

[cedar sentinel]

Nice talk btw.

[jaunty helm]

ig you might run into trouble where

| visit_id | prev_1_eng_lvl | prev_2_eng_lvl | ...
|        2 |              5 |              3 |
|        1 |              7 |            ??? |
```one way I could see it working is set it to `0`, and let the model pick up on the fact that `0` means there's no one

[cedar sentinel]

Yeah you can do a bunch of things. That;s another approach as well

[jaunty helm]

or if you really want to emphasize, include binary columns like

| visit_id | prev_1_eng_exists(binary) | prev_1_eng_lvl | prev_2_eng_exists(binary) | prev_2_eng_lvl | ...
```probably have to test what works best

[cedar sentinel]

😂

[jaunty helm]

sure

[cedar sentinel]

Food for thought

[twin timber]

whats the time complexity for lookup in csv by pandas?
lets say i have a column with unique values called "id", and i do a lookup like this dataframe[id]["unique_id_value"]

is it O(1)?

[cedar sentinel]

To be O(1) it should use a hashtable. Let me check

Its O(n)

With an index is O(1) :
dataframe.set_index("id", inplace=True))

[twin timber]

i read somewhere that with index, it is O(logn) due to B-trees, i maybe wrong though

[serene scaffold]

B-trees are for certain kinds of actual databases, whereas pandas is all in-memory.
if you do df.loc[row_index, column_index], that's O(1). you should never have expressions that look like df[ ][ ] in pandas.

[twin timber]

thank you for the insight, im actually building a search engine based on files, i've chosen a dataset of 1M+ docs

what could be the better format for storing files? json or csv? im gonna be dealing with roughly ~500K-1M doc files

[serene scaffold]

JSONL

which is slightly different from JSON. you'll want to look into it.

and if you need to be able to quickly look up subsets of the document set (the whole document set is a corpus), you'll need to put it in a search engine like elasticsearch.

[unkempt wigeon]

More than 150% important because if you use a regular list python it takes a while for it to take every object out of the list but if you assign it as an array it'll make it so that it's in a format that's more of an integer most likely sorry

[twin timber]

we have to actually build a search engine from scratch so i need to look for a format for fast retrievals and less memory consumption

[lapis sequoia]

can anyone help me with python ML?

[wooden sail]

not if you don't ask a concrete question

[lapis sequoia]

i have to make a program that get a sound of person talking and it needs with background vs foreground to find the seconds the person talks
and it wants me to create an ML from the begining no ready dataframes.

[serene scaffold]

that's not a DS/AI question anymore.

[serene scaffold]

by "ready dataframes" do you mean training data?

[earnest fjord]

Hello, What would be the best path to get big in the AI and make it big

I don't have the resources for college till yet

[serene scaffold]

get a PhD in computer science with a research focus in a niche area of AI.

[earnest fjord]

I want to build up my skills, learning and hopefully make something in AI, but I can't afford the have a education in US till yet

[serene scaffold]

you can start learning about AI whenever you want. but you almost certainly won't be able to have a career in AI unless you get at least a bachelors--and probably a masters--in a scientific field with focus on AI.

Are you interested to learn about AI anyway? Or do you need to be able to "get big and make it big"?

[crisp plover]

are the same any good books that teach creating AI using python? Geared towards a beginner

[serene scaffold]

!resources data science

[arctic wedge]

Resources

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

[serene scaffold]

also this: #data-science-and-ml message

[crisp plover]

Ty

[earnest fjord]

I want to be able to get into AI from a problem solving standpoint, i.e. learning and hopefully finding a problem which I could then be able to solve and contribute that way.
I am not saying he's good, but Sam Altman like didn't have a degree, but still runs OPenAI

[serene scaffold]

Sam Altman like didn't have a degree, but still runs OPenAI
you need connections and an insane amount of luck for that to work.

[earnest fjord]

So are there no practical path for something like me in the AI?

[serene scaffold]

if you can't get a job or funding through connections, and you can't get a degree, then there is not.

it might work out anyway because of luck, but if you don't get lucky, you will have wasted time that you could have spent preparing for a different career.

[earnest fjord]

Does this applies to AI or with other paths in prgraming as well?

[serene scaffold]

AI is the most degree-requiring path within programming. Web development is the least.

Also, Sam Altman was in university, but he dropped out to found a company with a classmate. Which meant that he needed to have an idea that he knew how to implement, and which he could convince investors to give him money to do, and then it had to actually work.

[earnest fjord]

Should I focus on my current skills like Web Dev and Game Dev and be able to save up enough money to afford education in AI eventually?

[serene scaffold]

and it looks like 90% of AI startup companies fail.

[serene scaffold]

I wouldn't plan that far out. No one knows what the AI landscape will look like by then.

[earnest fjord]

Thats what I am confused with, I want to do something in Game Dev, Web Dev, App Dev and also in AI, but don't have a practical path on which one to choose as the future

[serene scaffold]

game development has some of the worst working conditions in all of programming.
you can start practicing web development and app development at any time.

[earnest fjord]

Game Dev is something that I find fun and something as a hobby. I currently am a web developer.
But problem with web dev and apps is that, those are kind of saturated and I am not sure I would be able to actually contribute in them

And overall what should I aim in terms of carrer

[brazen sparrow]

if you want a specific job start building a portfolio for that job, if you have a passion the job will fix itself

[crisp plover]

That last book, “dive into deep learning”, does it explain concepts start to finish in python? If you’ve read it

Very new to this so looking for something with actual code that can help me build networks myself

[tawdry sundial]

How does regularisation make any sense

So you find the best line fit and then you decrease the slope for some bs reason

“My model doesnt perform well, ill randomly adjust the weight”

[calm thicket]

the idea behind regularization is preferring simpler models over complex ones in order to avoid overfitting

i'm not sure what you mean by "decrease the slope"

[tacit sorrel]

Looking for some thoughts on this

Ive been writing a few of different agentic rags I’ve seen online on Jupyter notebook and I kinda thought it would be neat to go straight from Jupyter to api server so I can give it to a buddy to test out. Like have him hit certain functions

Thought it was clever enough to share lol

[serene scaffold]

Notebooks are for human consumption. You'll have to copy the relevant functionality into a py file and wrap the functions with flask or something

[spring field]

If you find game dev fun, do it as a hobby instead of a career

[toxic palm]

hi, any one familiar with aws glue with terraform. need one help pls.

[serene scaffold]

Hello, please remember to always ask your actual question. Never ask if someone knows about the topic of your unasked question.

[toxic palm]

pyspark_data_transform_code.py

import sys
from awsglue.transforms import *
from awsglue.utils import getResolvedOptions
from pyspark.context import SparkContext
from awsglue.context import GlueContext
from awsglue.job import Job

# Get job arguments, get datalist https://www.datablist.com/learn/csv/download-sample-csv-files
# args = getResolvedOptions(sys.argv, ["JOB_NAME"])

# Initialize SparkContext, GlueContext, and SparkSession
sc = SparkContext()
glueContext = GlueContext(sc)
spark = glueContext.spark_session

# Create Glue job
job = Job(glueContext)
job.init("pyspark_glue_job")

# Script generated for node AWS Glue Data Catalog
AWSGlueDataCatalog_node = glueContext.create_dynamic_frame.from_catalog(
    database="crawler_output_db",
    table_name="customers_csv",
    transformation_ctx="AWSGlueDataCatalog_node",
)

# Script generated for node Change Schema
ChangeSchema_node = ApplyMapping.apply(
    frame=AWSGlueDataCatalog_node,
    mappings=[
        ("index", "long", "index", "long"),
        ("organization id", "string", "organization id", "string"),
        ("name", "string", "name", "string"),
        ("website", "string", "website", "string"),
        ("country", "string", "country", "string"),
        ("description", "string", "description", "string"),
        ("founded", "long", "founded", "long"),
        ("industry", "string", "industry", "string"),
        ("number of employees", "long", "number of employees", "long"),
    ],
    transformation_ctx="ChangeSchema_node",
)

# Script generated for node Amazon S3
AmazonS3_node = glueContext.write_dynamic_frame.from_options(
    frame=ChangeSchema_node,
    connection_type="s3",
    format="csv",
    connection_options={
        "path" = "s3://${aws_s3_bucket.destination_data_heidi.id}",
        "partitionKeys": []
    },
    transformation_ctx="AmazonS3_node",
)

job.commit()

glue_job.tf

resource "aws_glue_job" "pyspark_script" {
  glue_version      = "4.0"                                         #optional
  max_retries       = 0                                             #optional
  name              = "pyspark_script"                              #required
  description       = "This script will transform the catalog data" #description
  role_arn          = aws_iam_role.glue_role.arn                    #required
  number_of_workers = 2                                             #optional, defaults to 5 if not set
  worker_type       = "G.1X"                                        #optional
  timeout           = "60"                                          #optional
  execution_class   = "FLEX"                                        #optional

  command {
    name            = "glueetl" #optional
    script_location = "s3://${aws_s3_bucket.pyspark_script_heidi.id}/pyspark_data_transform_code.py"
  }
  default_arguments = {
    "--class"                   = "GlueApp"
    "--enable-job-insights"     = "true"
    "--enable-auto-scaling"     = "false"
    "--enable-glue-datacatalog" = "true"
    "--job-language"            = "python"
    "--job-bookmark-option"     = "job-bookmark-disable"
    "--datalake-formats"        = "iceberg"
    "--conf"                    = "spark.sql.extensions=org.apache.iceberg.spark.extensions.IcebergSparkSessionExtensions  --conf spark.sql.catalog.glue_catalog=org.apache.iceberg.spark.SparkCatalog  --conf spark.sql.catalog.glue_catalog.warehouse=s3://tnt-erp-sql/ --conf spark.sql.catalog.glue_catalog.catalog-impl=org.apache.iceberg.aws.glue.GlueCatalog  --conf spark.sql.catalog.glue_catalog.io-impl=org.apache.iceberg.aws.s3.S3FileIO"

  }
}

pls ask me for any additional details / screen shots.
it's been 2 days on this.
Destination bucket is getting created where the script output should get stored. However the data is not getting created inside it.
On the debugging process, i found that input data bucket is getting properly & input data is loading into it properly.
And then the glue job is running on that input data.

[final jay]

Are there people that get into computational linguistics with a degree in linguistics? Or is a degree in compsci better? I'm new to IT so I don't know a whole lot about industry demands these days

[serene scaffold]

I am a computational linguist. I was a linguistics major but switched to computer science. One of my coworkers has a degree in linguistics and a minor in computer science.

Note that computer science is not IT.

[final jay]

Oh I see, what's the difference between the two?

[serene scaffold]

in principle, IT is just "information technology", but it's generally understood to refer to degrees in information technology and the occupations that those with that degree have, which are not scientific.

a degree in IT is going to be a business degree, and if you have a degree in IT, you'd be filling technical support roles. not developing software or AI.

[final jay]

Is it a good field for someone who likes programming and linguistics? I guess that question is subjective, but I guess the linguistics aspect of it might be different from what I would think?

[serene scaffold]

it's the only field in linguistics that pays well, for one thing.

when I was in high school, one of my teachers had a phd in linguistics, and I told her that I wanted to go into historical linguistics. and she said "you should go into computational linguistics. they make bank."

she actually wasn't my teacher. she was my friend's teacher.

but she taught me the most important thing I learned in high school, so I guess she was my teacher.

I didn't actually do what she said to do until like 7 years later.

[final jay]

I'd probably do linguistics with a minor in compsci. I really find language interesting, and that'd be my biggest focus. Do you know where I'd find good resources to get into it? Is open source contributing a good way to dip my toes into the field to see if I like it?

[serene scaffold]

I don't know of any open source projects that are approachable for beginners and which actually involve theoretical linguistics

linguistics with a minor in compsci
do it the other way around.

especially in the age of deep learning, your explicit awareness of theoretical linguistics won't be as helpful as it probably would have been 10 years ago.

I have this chart on a poster in the room where I work. if I didn't, I probably wouldn't think about pure linguistics most days.

[final jay]

It's a shame that I wasn't on the computational linguistics grind when I was 8 years old. But jokes aside I see what you mean

[serene scaffold]

my advice to all fetuses is to get a law degree and adopt an extreme political ideology so that they can get on the supreme court when they're born.

[tacit sorrel]

I think that’s what I’m alluding to, build some kernel commands that make a cell an api I can share. Totally bypass the flask thing

Thoughts on that?

[final jay]

Every tech guru ever

[serene scaffold]

I don't even want to think about how a notebook could be turned into a web app without switching back to normal py files. not everything needs to be a notebook.

whatever solution you come up with, you're not "bypassing flask". you're just replacing it with something that's "flask for notebooks".

[tacit sorrel]

Yeah the sales pitch sounds more like bypassing flask. The engineering hack would just hide the complexities of flask or a web framework from your notebook

[serene scaffold]

flask isn't that complicated.

I haven't used it in years, but from what I recall, it would be a challenge to make it less complicated.

[tacit sorrel]

It’s more like it’s a pain to do every time for every iteration of notebook I do.

I.e im on a dedicated AI team tweaking various versions of our RAG and agent workflows and I do NOT want to make flaks endpoints for every tweak/version etc. which I currently do.

I would rather have certain cells be my handler functions automatically

[serene scaffold]

you know when the US adopted a containment policy to stop the spread of communism? I'm like that with use cases for notebooks.

[tacit sorrel]

Big mr. I don’t like thinking outside the box I see 😉

[serene scaffold]

I use IPython every day and notebooks most days. But I see their limitations and don't try to expand their scope.

[toxic palm]

AWS glue job is producing output files like run-1732763023996-part-r-00000.
can we get them in .csv format?

[dense needle]

What was this out of curiosity? The recommendation of what to study?

[serene scaffold]

Yes. Though I was being facetious

[tawdry sundial]

Regularisation like laso and rigid apply penalty on trainable parameters. I was referring to regularisation on linear regression, it adds penalty the slope

Loss: mse + alpha * |m|

Where m is the learnable parameter in y = mx + b

[calm thicket]

where did you get this formula? this is not correct

[lapis sequoia]

I'm new to python and data science. Can you guys suggest some good projects with source code

[tawdry sundial]

How so

What exactly js not correct

[calm thicket]

ah. misread. yeah, the idea is to penalize larger coefficients unless it is a lot better than a smaller coefficient

https://scikit-learn.org/stable/modules/linear_model.html#ridge-regression

[tawdry sundial]

its confusing

[terse frigate]

hello friends... i am a master's in AI but i am unable to land a job. i have been contemplating on doing a certificate course online but cannot decide which one to opt for ?? Shall i got for Google Cloud or AWS Cloud ??? Please guide me 😄

[quaint mulch]

You decrease the slope for good reasons.
You find the line of best fit on the data you have at hand.
A good assumption is that your data might not be perfect.
For example, there might be some variable there that shouldn't have any correlation, but because you have limited data and bad luck, the line of best fit shows some correlation when there is actually supposed to be none.
In other words, the gradient is supposed to be zero, but coz your dataset is not perfect, you got some gradient.
some regularization counteract this by decreating that graident

[tawdry sundial]

But how does it decrease some gradients by a lot and others by just a little?

I assume it depends on how much decreasing the gradient would result in increase in mse

Correct me if i am wrong

Now that i am thinking of it, it would make sense that it would cancel out the gradient which has no correlation with the label instead of the other way around

[jaunty helm]

not sure what you mean exactly
if you have a model that optimizes some function objective(), then adding regularization can be thought of as optimizing objective() + regularization()
and we can design the regularization() function to encourage specific behavior; e.g. L1 = sum(abs(weights)) encourages the absolute values of each weight to be small
not sure where the "random" is coming from

[toxic palm]

Below is the script executed as part of glue job

# ------------------------------------------
# Bioler plate code
import sys
from awsglue.transforms import *
from awsglue.utils import getResolvedOptions
from pyspark.context import SparkContext
from awsglue.context import GlueContext
from awsglue.job import Job

sc = SparkContext.getOrCreate()
glueContext = GlueContext(sc)
spark = glueContext.spark_session
job = Job(glueContext)
# ------------------------------------------
# Read data in CSV From AWS S3
s3_path = "s3://source-data-heidi/"
customer_table = glueContext.create_dynamic_frame_from_options(
    connection_type="s3",
    connection_options={"paths": [s3_path]},
    format="csv",
    format_options={"withHeader": True, "optimizePerformance": True},
)
# ------------------------------------------
# Write data in CSV into AWS S3
s3_path = "s3://destination-data-heidi/"
glueContext.write_dynamic_frame.from_options(
    frame=customer_table,
    connection_type="s3",
    connection_options={"paths": [s3_path]},
    format="csv",
    format_options={"separator": "|"},
)
# ------------------------------------------
# Display schema and first 5 records
customer_table.printSchema()
customer_table.toDF().show()
# ------------------------------------------

job is getting failed & when i verify the logs, it says

An error occurred while calling o93.getSink. No usable value for path. Note: This run was executed with Flex execution. Check the logs if run failed due to executor termination.

could not understand what is this error message trying to say.
any help pls.

[north adder]

hello i want to create a model that compares actual result with expected result and flag it as positive or negative. but problem is i cant find this type of dataset. can someone help me on what to do in such case

[tawdry sundial]

L1 = sum(abs(weights))

[cedar sentinel]

Hi, quick question. I'm given a dataset were 1 column is called
engineer_note: A summary note left by the engineer after this visit. The note
is already tokenized, so all words are replaced with numeric ids. Each word is
always associated with the same numeric id. For example, 106 might always
refer to the word ‘wire’.

How can I use such information to build external features even though I dont have access to the dictionary ? I can think of just 1 hot-encoding. A word2Vec has no meaning and also any mapping from LLMs to get the embeddings has no meaning cause the dictionary might be different.

[tawdry sundial]

this is random, you could perform any other operation

[tawdry sundial]

how do you know that you have to decrease gradients? maybe we need to increase the gradients?

[jaunty helm]

this is random, you could perform any other operation
sure, you can make up some other strange complicated function and use that as the regularization function and get another result
L1 and L2 have quite a bit more history though (see lasso and ridge, MAE and MSE), and have been studied more not only in ML but statistics
the reason they're chosen involves errors being normally distributed or something... I'm not well versed in stats enough to explain thoroughly

how do you know that you have to decrease gradients?
uhh... sorry? walking down the gradient would decrease the value of the function(the objective) being evaluated, which is what we want?

[jaunty helm]

ig as a not-very-rigorous reasoning, it's not hard to reason that if we penalize sum(abs(weights)), then the model would prefer a solution with smaller weights, which makes it more stable (e.g. similar inputs would have similar outputs), which ideally means it generalizes better

[deep veldt]

is there any way i can make ultralytics yolov11 to reuse the runs folder? i just have trained it and forgot to export, it took me very long

[tawdry sundial]

walking down the gradient would decrease the value of the function(the objective) being evaluated, which is what we want?
it decreases the output, why would you want to decrease the output.

If i have a data set, and i find the line of best fit (linear regression with mse or mae) why would i willingly go against my data and enforce it to decrease the output by penalizing the gradients? It seems absurd.

this is confusing me

[jaunty helm]

no, we're not decreasing the output, we're decreasing the objective function, which is a measure on how good / bad the model is (usually defined in terms of lower = better)

originally the objective function F is just objective(), e.g. mse. we can find a best fit for it, which is the model that minimizes F
then we add regularization, and now the new objective function F' is objective() + regularization(), e.g. (mse + l1), which we now again find a best fit, which is the model that minimizes F'
this new model that minimizes F' may not be one that minimizes F, but due to how regularization() was set up (say l1, which penalizes weights with big absolute values), we can expect the model to have better stability

we wouldn't need to do this if the original data is very good, such that the best model for it just so happens to be the best model in general
but that's usually not the case

[wooden sail]

the standard statistical formulations for L1 and L2 regularization of a least squares cost function is superficially along the lines of: the least squares cost function corresponds to "maximum likelihood estimation" under gaussian noise. L2 regularization comes in when the parameter you're looking for is also gaussian distributed and you now do "maximum a posteriori estimation". L1 shows up when the parameter is instead laplace distributed (the values are 0 with high probability), and again doing maximum a posteriori.

the deterministic formulation for L2 reg has to do with improving the condition number of (almost) singular matrices, and for L1 it has to do with the convex relaxation of L0 regularization so that the solution is sparse (explains the data with as few parameters as possible, or "in a simple way")

[cedar sentinel]

Hi, quick question. I'm given a dataset were 1 column is called
engineer_note: A summary note left by the engineer after this visit. The note
is already tokenized, so all words are replaced with numeric ids. Each word is
always associated with the same numeric id. For example, 106 might always
refer to the word ‘wire’.

How can I use such information to build external features even though I dont have access to the dictionary ? I can think of just 1 hot-encoding. A word2Vec has no meaning and also any mapping from LLMs to get the embeddings has no meaning cause the dictionary might be different.

[frail mauve]

I am new to AI, I want to build a application which reframes videos from 16:9 to 9:16 using dynamic cropping, subject detection. How do I do this? Any suggestions?

Ping me while replying

[quaint mulch]

Because, usually, in the design of your pipeline, you usually risk overfitting, by throwing in more parameters than neccesarry.
From experience, it is seems what work best is to make your model slightly more complex, and then fix it using regularisation.

also remember, we usually standardise the data 1st

like, the best way to understand this is from the perspective of the bias variance tradeoff

[wooden sail]

this depends only on whether you formulate your task as an optimization or a minimization problem

the gradient always tells you the direction in which, locally, the function increases the most

oops wrong reply

[wooden sail]

this was for you

one usually writes ML problems in terms of the "error", "cost", or "loss". these are all things you want to reduce to 0 if possible

so you want the negative of the gradient. otherwise you make the error bigger and your model becomes worse

[cedar sentinel]

Any idea in this NLP task ?

[dense needle]

I think part of the confusion may be that you are viewing the minimization process as “going against the data.”

[proven pier]

I have a question about Experience Replay and Reinforcement Learning. Is there some paradigm where you adjust all rewards of any given measured episode at the very end using some form of discount? I know the standard Bellman equation already does similar iterative math with adjusting Q-values. However, I'm thinking about an episode where rewards are only calculated on the final cycle. And the reward is either very positive/negative (depending on how the system evolved). Well, given Experience Replay, obviously rewards are not going to be present on any random sampling. So I wonder if I can post-process my experience when an episode ends and give some sort of discount factor to all the steps? Maybe this is a horrid approach.

[spring field]

First thing that came to mind similar to what you describe were the concept of fixed Q targets, however, a quick search brought up a concept of delayed rewards which might be worth checking out, though I'm not particularly familiar with it.

[agile owl]

more generally the fortran code people use to optimize things only works in one direction so even if you want to maximize something you just negate it first

that's the real reason it's always minimizing lol

case in point: minimizing the negative log likelihood

it would seem to be more straightforward to just maximize the log likelihood but the convention is to minimize the negation because that's just the code we have available to use

and everyone uses it

[quaint mulch]

I think yoiu misunderstood his question lol. He is not talking about the backprop gradient, he is talking about the model weights, and calling it gradient, making anologies to linear models.

[wooden sail]

ah well i didn't read the whole convo. can't do anything if a technical term is used wrong :p

[deep veldt]

how do i make yolo auto export upon exit while it's training? i've tried with a try finally and model.export but got TypeError: 'IterableSimpleNamespace' object is not subscriptable

[tawdry sundial]

the regularization() makes no sense, it could be anything.

L1 regularization is alpha * |slope| (where slope refers to the trainable parameter in linear regression), why dont we use the opposite of that and make regularization() be - alpha * |slope|

[tawdry sundial]

so regularization attempts to make the model more complex to avoid overfitting? how does making the model more complex help with overfitting? it seems quite the opposite, it would make more sense to make a model more simple.
For example, if a neural network is overfitting the data, decreasing the number of neurons or layers could solve the issue instead of increasing it.

[tawdry sundial]

i should have clarified, i was referring to the slope in y = mx + b where m is the slope and the trainable parameter

not the loss function gradient

[tawdry sundial]

honestly, the entire concept of regularization is confusing me. I might have to skip this part for later, maybe ill understand it better when get to deal with bad dataset

[wooden sail]

what part of it troubles you?

[tawdry sundial]

understanding why we add + alpha * |slope| to the loss function instead of - alpha * |slope|

[wooden sail]

what do you want the regularization to do?

[toxic stream]

Hi, would someone mind helping me with an issue I've had with my neural network?

I've been stuck at 76% accuracy and realisitcally want to get it to 95% for my school project

[tawdry sundial]

the way i understand it, regularization makes the model less prone to overfitting

[tawdry sundial]

have you tried adding regularization?

[toxic stream]

I don't think so

[wooden sail]

i prefer interpretting it as giving the parameters "structure"

the question is what structure do you want to give to the parameters

[tawdry sundial]

how does it structure the parameters?

[wooden sail]

if you have a loss function that you're minimizing involving the equation a line, and you add alpha * abs(slope) to that loss, then clearly the overall loss function is larger if the slope is big

so now you have forced the model to limit itself to smaller values of slope

the structure here would be "use small slope values"

[tawdry sundial]

god damn it

so its like limiting decision trees

[wooden sail]

if you subtract alpha * abs(slope) from a cost you are minimizing, then obviously this is useless. the minimum overall cost is minus infinity, which you achieve by making the slope infinitely large

[tawdry sundial]

by adding max leaf and max sample

[wooden sail]

so the negative version is useless

i've never worked with decision trees so i can't comment

[toxic stream]

Would you mind helping me with that?

[haughty ledge]

I see

[tawdry sundial]

idk how i didnt think of that, its was just used as an example anyways

honestly this makes a lot of sense now

thanks

[wooden sail]

something super important to keep in mind: the minimum of a cost function and the minimizer of a cost function are completely different things

the minimum is the value the cost function reaches. the minimizer is the parameter needed to achieve that minimum

so always play with the two. that's how we concluded that + was the correct regularizer here

it's the same distinction between min and argmin

in ML one should technically always write argmin cuz we care about the params. the shorthand for this is min, with the minimizer variable as a subscript. so it's actually argmin

.latex e.g. [
\argmin_{\bm{x}} \Vert \bm{y} - \bm{Ax} \Vert_2^2
]

[strange elbow]

Failed to render input.

View Logs

[wooden sail]

sigh

[tawdry sundial]

the minimizer is the parameter needed to achieve that minimum
I assume you are refering to hyper parameters, correct me if i am wrong

[wooden sail]

you are wrong

[tawdry sundial]

why

[wooden sail]

hyperparameters are a different discussion still

the actual value of alpha here is a hyperparameter, for example

or how many layers you will use, or number of iterations in an iterative algorithm

the parameters are the slope m and offset b

[tawdry sundial]

got it

[tawdry sundial]

share more details, i would like to look at the issue and hopefully learn from it

[toxic stream]

Should I share my model code?

[tawdry sundial]

yes, if its a lot of code, put it in pastebin

https://pastebin.com/

[toxic stream]

https://pastebin.com/3cj5btDJ

[tawdry sundial]

1sec

[toxic stream]

np

[tawdry sundial]

i expected a simple neural network, this is a large cnn (for me at least). I dont think i can help you with improving the model

[toxic stream]

ok np, that's fine

[tawdry sundial]

do you mind sharing the training/testing data? I would like to tweak it and test somethings out

[toxic stream]

The folders are quite large so it may be better if I just link the dataset I downloaded

Since I used a python program to split the picture roughly in a 10:90 ratio of validation to training data

[tawdry sundial]

Awesome

[toxic stream]

I think it was this one: https://www.kaggle.com/code/utkarshsaxenadn/celebs-face-recognition-facenet/notebook#About-the-Dataset

[bleak glacier]

Hello did anyone here try running torch / tension with PyPim and can share the experience?

By PyPim I mean in memory processing package, not the networking one, as in utilise the RAM for computation instead of gpu / cpu

The link to repo

https://github.com/oleitersdorf/PyPIM

I've tried reading this article
https://arxiv.org/html/2308.14007v2

But it goes way in depth and my knowlage base isn't good enough to fully understand what are the benefits / preformance gains vs gpu computation

[deep veldt]

im trying to write a siamese network that uses yolo as its sub-networks how would i make one?

[quaint mulch]

Yes.
limiting the number max number of leaf and samples ARE regularisation methods as well.

[forest wadi]

guys, I want some help with reinforcement learning problem

[serene scaffold]

don't wait for someone to engage. describe the problem in enough detail that someone can jump in.

[forest wadi]

Sure,
The RL problem is defined as follows:

• The agent has one action to take from a continuous action space, the agent chooses a float between 0.2 and 0.7
• The environment is taking this action and calculating some values to end up with a target value "SPEED"
• The state is defined with 5 parameters, SPEED and 4 more that goes in the calculation of the SPEED
• After each step, only SPEED is affected by the action that the agent took, the other 4 parameters come from a database and change and the start of each episode.

The main idea of the problem is we want the agent to be able to find the optimal number that achieves 8 < SPEED < 9

The problem is that most of the time the agent tends to keep choosing 0.2 and 0.7 even thought this action is heavily penalized, and even when the agent convergence after 800k steps it doesn't converge to the optimal solution
I tried PPO and TD3 with different exploration parameters and different learning rates

[plush gorge]

Hello there I want to build a real time sign language translator having continuous video prediction I have tried Mediapipe with BiLstm the results are good on dataset but soon as I deploy it for real time footage the model sucks and it shows random predi toon
I have an dataset of 256 labels but currently I am just using 8 labels for initial testing
Thanks

[unkempt apex]

that's called overfitting

when model sucks for new data

try to flexible while training ( what does that mean now?)
use different types of data

[plush gorge]

I tried with generating different types of video augmenation but the results are not much better
Initially there was no testing accuracy so i figured it is overfitting but after augmentation the testing accuracy improved.
So is it still possible to have overfitting even though i have good testing accuracy?

And my dataset consists of complex signs not just static images for alphabets.
I want it to be somewhat usable in real world sceanrios.

[unkempt apex]

that's the sign of overfitting
when your testing loss is less on trained dataset
but at the same time model doesn't perform well for new type of data

[unkempt apex]

can you share what you are using for training
and for real world

[plush gorge]

Yeah i can,
Can I DM them?

[unkempt apex]

you should send it here

so that others can also help you

[plush gorge]

👍

This is an example for the training dataset
In the real time one i recorded my own videos containaing multiple signs for testing it in real time using opencv

[unkempt apex]

which model you are using?

thanks for sharing this btw!

[plush gorge]

First the video is fed to the mediapipe to extract the keypoints lh, rh and the pose (All total 258 keypoints)
Then after collecting this data from each video
I train the model using the below architecture in tensorflow

                Input(shape=input_shape),        
                
                # Bidirectional LSTM layers
                Bidirectional(LSTM(64, return_sequences=True)),
                Bidirectional(LSTM(128, return_sequences=True)),
                Bidirectional(LSTM(64, return_sequences=True)),
                
                Flatten(),
                
                # Fully connected layer : ANN
                Dense(128, activation='relu'),
                Dense(num_classes, activation='softmax')
        ])```

The data extracted from mediapipe is first stored in the form of an numpy array which later laoded to be trained by the rnn

[unkempt apex]

stop using GPT btw!

[plush gorge]

if you want to understand the dataset more thoruogly i can send the sources of the datset

NO I use it rarely

when i have no option

[unkempt apex]

the code says it all

[plush gorge]

The base was originaly from there because i had no idea first how this worked then i modified it accordingly

[unkempt apex]

can u share , what features the mediapipe is extracting
before converting it to np

[plush gorge]

Sure

First i extarct the kepoints in form 3 numpy arrays first one being the data for poseture then the left hand and right hand and then concatenate it all
If a detetction is missed the array send 0 instaed of leaving it empty

def extract_keypoints(results) -> np.array:        
            pose = np.array([[res.x, res.y, res.z, res.visibility] for res in results.pose_landmarks.landmark]).flatten() if results.pose_landmarks else np.zeros(33*4)
            lh = np.array([[res.x, res.y, res.z] for res in results.left_hand_landmarks.landmark]).flatten() if results.left_hand_landmarks else np.zeros(21*3)
            rh = np.array([[res.x, res.y, res.z] for res in results.right_hand_landmarks.landmark]).flatten() if results.right_hand_landmarks else np.zeros(21*3)
            
            return np.concatenate([pose, lh, rh])    ```

I basically extarct all the x,y and z coordinate

[unkempt apex]

always add "py" after back ticks

do you have any video on test data?

how model is deriving features

[plush gorge]

Do you mean how the model is tracking ?

[unkempt apex]

this is your data extraction part which is fine
but I need what type of model you are using

[unkempt apex]

yup on test data

[plush gorge]

Ok i will record the tracking and send it over

This is the recording for tracking on real time data

[plush gorge]

Do you mean mediapipe's model ?
Then it is mp.solutions.hollistic

What is your recommendation for the solution of my problem

[unkempt apex]

tf happen to recording

[unkempt apex]

after extracting data from video and converting it to np.array , then on what model you are training it?

[plush gorge]

BiLSTM

model = Sequential([        
                Input(shape=input_shape),        
                
                # Bidirectional LSTM layers
                Bidirectional(LSTM(64, return_sequences=True)),
                Bidirectional(LSTM(128, return_sequences=True)),
                Bidirectional(LSTM(64, return_sequences=True)),
                
                Flatten(),
                
                # Fully connected layer : ANN
                Dense(128, activation='relu'),
                Dense(num_classes, activation='softmax')
        ])```

[unkempt apex]

ohh my bad

why LSTM?

https://stackoverflow.com/questions/65879627/lstm-for-video-input

read second answer from here

[plush gorge]

Becuase of its ability to relate with older data in NN so my though process was in videos model should be able to make sense with its older frames so and LSTM would be good

[unkempt apex]

update this first of all

can't see video here

[plush gorge]

I reuploaded it

You can try downloading if it isn't showing preveiw here

[plush gorge]

So basically i have to ditch Mediapipe here and just directly feed the video in this architecture?

[unkempt apex]

yeah!

CNN are capable to directly extract features! but they need lot of data to be more accurate

[plush gorge]

That's the problem here i don't exactly have a lot of data available with augmentation on average I have 60 videos here for a single label

[unkempt apex]

wait!!

this approach is also good

just confused why it is overfitting

[plush gorge]

That's what I am trying to figure out for past 4 days

[unkempt apex]

okay so for now!!

lets do some testing
-> keep the same model you have trained
-> don't test on your own video ( your PC camera )
-> just download random video from online and test

let's see what it does

[unkempt apex]

it's okay don't worry

[plush gorge]

Ok i will try the random videos
But by random vedios you are saying videos with same label set or any random signs

[unkempt apex]

you must clarify the signs which you are trained
if that random video has that signs then its okay

[plush gorge]

Nope its the same

I found 2 videos

[unkempt apex]

can u share how the model is responding to that videos

[plush gorge]

Sure

[deep veldt]

How can i make a siamese model?

[past meteor]

The idea is you take a standard neural net, output N neurons in the final layer

It takes two or three inputs when predicting

And then you compute some distance metric for your loss

[weary timber]

whats the best machine learning framework?

or library idk

[serene scaffold]

what are you going to do with that information if I give it to you? because you shouldn't try to learn machine learning in terms of a framework/library.

[weary timber]

after learning how the machine learning work im not gonna do a big project with numpy

[serene scaffold]

there's more to machine learning than neural networks, but you should use pytorch if you want something more practical, or JAX if you want less abstraction. don't use tensorflow as it's dying.

[weary timber]

i typed wrong i was gonna say machine learning

[serene scaffold]

you did say machine learning; did you mean to say neural networks?

[weary timber]

i edited it 😭

[serene scaffold]

when you do your big project, what is the model that you create intended to do?

[deep veldt]

do i need 2 subnet or just 1?

[cloud cosmos]

Hello

I need help with tensorflow I posted my question on the python-help page!

[weary timber]

i dont have anything in mind

[past meteor]

one, that's the tricky part when trying to understand siamese nets

Basically: 1 net, send in the first input, get the embedding, send in the second input, get the embedding and then calculate the similarity metric between both

[peak drift]

Hey there! I'm trying to create the following fourier transform:

f0 is given, f is known and sBB(t) is known. I can use np.ftt.ftt(sBB) to get the fourier transform of it, but I need it moved ( (f-f0) and (f+f0) )

[rugged mist]

responded in #1312458387832705035

[unkempt apex]

what was for second test video?

the model predicts accurately for first test video I guess

[plush gorge]

The first one was quiet which gives sad as a prediction

and second vedios was for sad

[unkempt apex]

for how many epochs you trained?

[plush gorge]

15 epochs or else it would start overfitting

I used early stopping to prevent that

[unkempt apex]

how much time it took? for 15

[plush gorge]

I have an gpu so it hardly takes a minute

And its just the extracted coordinated instead of images so yeah it takes very less time to train

[unkempt apex]

okay so from all this things
-> the data points ( np arrays ) which are extracted either lacks the info for the model

-> of the model size needs to be changed

haven't you look for some papers on this?

which approach are they following

[cerulean storm]

hi.. kinda off topic.. but can anyone with pika pro generate one dissolve effect for me? I really need it >(

[vital scarab]

Hi, Im pretty new to AI and Im currently trying to train a vanilla DQN with a prioritized replay memory buffer for a pretty simple task.
Its a small 2d football game and im doing a multicurricular approch where I start off very simple and then later add in more complex task with harder to get rewards. For the first phase I just wanted the ai to learn to walk towards the ball and pick it up, for this im giving a negative reward of -0.2 for every step and a positive reward of 10 for touching the ball. When the agent touches the ball I also end the episode. Now Ive been training my ai for around 200k training steps with a minibatch size of 24 and a learning rate of 0.025 and I literally see 0 progress at all. The AI is pretty much still walking around randomly. Now I know that vanilla dq learning is pretty slow but I feel like at this point there has to be something wrong with something in my code or approach since there is truly no visible progress at all.
This is the link to the pastebin of the code, if anyone with more knowledge and expertise could take a look at it I would appreciate it a lot!
https://paste.pythondiscord.com/APLQ

[karmic prism]

Who here works as an AI developer?

[plush gorge]

I have and they all are following somewhat similar architecture but none of them mentioned about testing on live feed even the people who published the dataset recommended an dataset but that didn't work that well if you want I can send there recommended architecture too

Some have even tried Gru and lstm

I could try them but I doubt it will have any difference

The thing of np arrays the problem could be with tracking
As if the frame dozens track the hand I pass an zero array instead of coordinates

[plush gorge]

And my another concern/question is can models learn this quickly

[unkempt apex]

it depends on data you passed

e.g -> raw images which gets passed to CNN takes lot of time

where are textual data learn quickly

[plush gorge]

ok so what could be tthe problem here

I tried GRU here results are better but not much better

[brave stream]

I'm trying to use a GAN model to try and generate spectrograms. I'm stuck on this issue where it seems like the generated images are limited by their resolution. The original shape of the sample images were 128 x 128, and while the output images are not that size (they're much larger), the image is limited to 128 by 128 rectangles (i went through the trouble of counting). I'm not sure why this is or how to counteract this.

I wasn't sure if this is really classified as the checkerboard pattern that I've encountered before or something else.

[finite night]

Does anyone have experience with FSDP+QLoRA? Looking for help on this problem. I have spent 10+ hours and still can't come up with a solution 😢 .
https://discord.com/channels/267624335836053506/1312775104664571995

[severe knoll]

Can anyone tell me how to get started with data science/ml, i know python and basics of algebra and probability. any tips for what i should learn next etc...

[swift furnace]

CS50 Artificial Intelligence

[proper meteor]

but i love using tf with keras

[weary timber]

https://www.youtube.com/watch?v=tIeHLnjs5U8
can someone (if possible) explain to me how backpropagating the error works based on this video?

[serene scaffold]

Can you describe in your own words what the error is? This isn't a test. I need to know what you currently know

[weary timber]

cost

(current - desired)^2 as i know

[serene scaffold]

Can you describe what this value represents?

[weary timber]

hmmmm

how bad the neural network is

higher = stupider

[serene scaffold]

One moment

@weary timber I'm walking home and then I'll help you

[weary timber]

👍🏿

[serene scaffold]

@weary timber do you understand how a neural network is like a function? when you put a value into the neural network, the result is a function of that input and all the weights of the network.

[weary timber]

i know that the neural network is a function but i cant really understand what you said at tthe end

[serene scaffold]

the weights of the function are an inherent part of the neural network, right?
but the instance that you put into the network (such as to train it) is not

[weary timber]

yep

sorry for the late replies i translate some words to understand well

[serene scaffold]

that's alright.
the whole point of training the network is to change the weights until you minimize the loss. right?

[weary timber]

yep

and for that as i know you calculate how sensitive the cost function is to the weights and biases so you can change the values of them correctly

aka derivatives

[serene scaffold]

so when you write (current - desired)^2, you're hiding a key consideration: current is the result of a function

suppose your neural network is f, and the input is x, and the weights are w. then the result is f(x; w). do you understand how I wrote that?

[weary timber]

what is the ; for

[serene scaffold]

to indicate that w is an inherent part of f, but x is not.

[weary timber]

oh ok makes sense

[serene scaffold]

so it's not really (current - desired)^2, it's (f(x; w) - desired)^2

see?

[weary timber]

yes

[serene scaffold]

you can also say loss(x, desired) = (f(x; w) - desired)^2

the next thing. w isn't a single value. each weight is its own value

when you look at a neural network diagram, each line between nodes is its own weight

still with me?

[weary timber]

yeye i dont wanna interrupt

[serene scaffold]

I want to make sure you understand what I'm saying at each step

if you don't understand, then it's my job to find a different way to explain it

[vestal mirage]

how can I teach ai to play games?

[weary timber]

i understand it to this point

[weary timber]

yes

[vestal mirage]

I said how

[weary timber]

sorry

[vestal mirage]

need help

[weary timber]

you should start from scratch and learn how ai's work for that

[serene scaffold]

@weary timber this is where the calculus starts: when you do back propagation, you're calculating the derivative of (f(x; w) - desired)^2 with respect to each individual weight in w

[vestal mirage]

how can I learn it

any links?

[serene scaffold]

@vestal mirage I'm in the middle of an in-depth explanation for memo; I'll try to help you next.

[serene scaffold]

do you understand this?

[vestal mirage]

sure thanks

[weary timber]

yes

[serene scaffold]

and then you multiply the result by the learning rate, and subtract that much from the individual weight.

[weary timber]

ye i can do that part

the part i struggle is when it comes to going to other layers

from there you will calc the error (cost ) for each layer as you go

[serene scaffold]

you don't need to think of backpropogation as an iterative process

the neural network is ultimately one big function

when you write it out as one big function, individual weight variables might appear multiple times, but that's okay

you just take the derivative of the big function with respect to that weight

@wooden sail what I'm saying is accurate, right?

[weary timber]

so i dont calculate the cost for each layer and just use the output layers cost for all weights

?

[serene scaffold]

well, keep in mind that the whole point of this is that the computer is doing it for you. but mathematically, you don't have to think of it as something that cascades through the layers.

[weary timber]

what

[serene scaffold]

which part are you confused about?

[weary timber]

all the message

[serene scaffold]

when you train a neural network, you don't have to do any math, because the computer is doing it. right?

[weary timber]

i mean i write it one time and thats all so yeah

[serene scaffold]

@weary timber you want to understand how the math for backpropogation works. specifically, how it "moves" between the layers. right?

[weary timber]

YES EXACTLY

[vestal mirage]

@serene scaffold can you help me now?

[serene scaffold]

do you know the chain rule from calculus?

[serene scaffold]

I'm not done helping memo.

[weary timber]

yes

[serene scaffold]

@weary timber the reason it's called "backpropogating" is that when you calculate the derivatives for the weights in one-layer, you can re-use those calculations to calculate the derivative of the layer before it.

well, the error more so than the derivatives

@vestal mirage you want to start small with this kind of thing. what game do you want your AI to play? connect four?

[vestal mirage]

Like snake game.

wait let me change my username xd

[serene scaffold]

@vestal mirage look into reinforcement learning

[vestal mirage]

but I dont want to make it working for my own games

I want to make auto play for other games like google snake game

[weary timber]

one last thing. can you say this sentence again simpler my english isnt englishing

[vestal mirage]

I know Turkish @weary timber

[weary timber]

youre turk or?

vaay

[serene scaffold]

please speak Turkish in your DMs

[weary timber]

sorry sorry

[serene scaffold]

it's okay

[weary timber]

so i use the calculations i did to find the cost at output layer for the layers before that?

[serene scaffold]

yes

[weary timber]

ohhhhhhhhhhhhh tysm

[serene scaffold]

you are welcome 💚

[vestal mirage]

any links

[serene scaffold]

No; I've never done reinforcement learning. sorry

[unkempt apex]

I can

Why RL but?

[vestal mirage]

im trying to learn it

i want to make bots

[unkempt apex]

Read about basic RL

DQN, Q-network

Agents and all that stuff

Then play with openai gym library

RL needs time to learn remeber this

[dapper peak]

Hi guys, what are the best online tools that you guys have used to learn the python skills
For data science

[iron basalt]

https://www.deeplearningbook.org/contents/mlp.html#pf30

It's the DP solution to the problem.

(When you are doing math on paper, substituting in the values for variables, you are always doing a table-filling (DP) algorithm by having written down the resolved values on paper (except not in a table, maybe just written all over the place) and reusing that over and over in other places (you could not do this, but it would take forever by hand and so people tend to do DP naturally))

[steady crystal]

I find myself needing to store a lot of (30G+) multiline text, as well as a variable-length header describing the data, so I need a dataset format that is streamable and performant. Which combination of dataset libraries/formats are the best for this?

[spring field]

it's behind http, but oh well: https://incompleteideas.net/book/RLbook2020.pdf

[steady crystal]

JSON Lines seem perfect for this except, I need to store which classes and there and such

[olive hinge]

Why are Encoder-Decoder neural networks stacked with two LSTM layers that are independent of each other

[wooden sail]

yep

[serene scaffold]

Ty
I wonder if I'd understand it at this point without you and @past meteor
Certainly no thanks to my deep learning "instructor"

[past meteor]

Controversial but I think maybe they should just teach backprop with auto differentiation in mind from the get go

[vestal mirage]

not found

[past meteor]

Every time people want to do neural nets from scratch they’re “hard coding” the derivatives etc. instead of making something small like micrograd

https://github.com/karpathy/micrograd

[spring field]

as mentioned it's not an "https site", so, perhaps, your browser just doesn't want to let you view the resource

but also take a look at the pinned messages in this channel

[iron basalt]

Understanding compute DAGs (dependency graphs) in general can make understanding a lot of stuff in math easier (and also often results in the optimal algorithm (DP)). I agree that this should be covered.

[plucky iron]

Can anyone solve my doubt related toCopy-Move Image Forgery Detection gaussian model?

Here this code, after resizing images and feeding them to the model with combined masks, it gives the error: ValueError: Arguments target and output must have the same shape. Received: target.shape=(1, 256, 256), output.shape=(1, 2048, 2048) for the below code:

# Assuming the input images are of shape (256, 256, 3)
# and the masks are of shape (256, 256)

# Combining the masks for training
def combine_masks(masks_copy, masks_paste, masks_forged):
    print(np.array(masks_copy).shape)
    print(np.array(masks_paste).shape)
    print(np.array(masks_forged).shape)

    # resizing
    masks_copy = cv2.resize(masks_copy, (256, 256))
    masks_paste = cv2.resize(masks_paste, (256, 256))
    masks_forged = cv2.resize(masks_forged, (256, 256))

    # checking shapes
    print(np.array(masks_copy).shape)
    print(np.array(masks_paste).shape)
    print(np.array(masks_forged).shape)
    print("-----------------")

    combined_mask = np.maximum(masks_copy, masks_paste)  # Take union of copy and paste masks
    combined_mask = np.maximum(combined_mask, masks_forged)  # Add forged regions
    return combined_mask

combined_masks = [combine_masks(copy, paste, forged) for copy, paste, forged in zip(masks_copy, masks_paste, masks_forged)]
print(np.array(combined_masks).shape)  # shape of combined masks
resized_images = [cv2.resize(image, (256, 256)) for image in original_images]

# Train the model
# model.fit(np.array(original_images), np.array(combined_masks), epochs=10, batch_size=1, validation_split=0.2)
model.fit(np.array(resized_images), np.array(combined_masks), epochs=10, batch_size=1, validation_split=0.2)

how should I resolve this error?

There is a mismatch between the shape of inputted mask (256x256) and the predicted mask from the model. The model outputs a mask of shape (1, 2048, 2048), while the target mask is of shape (1, 256, 256)?

Gaussian model code: https://paste.pythondiscord.com/N65Q

[mild dirge]

You should make it so the model outputs a tensor of shape (1, 256, 256)

Because that is what you are trying to predict

[fervent canopy]

BlazeChat - Fast CPU-Powered Chat & Image Generation: https://github.com/SanshruthR/CPU_BlazeChat

[ocean hinge]

Hello. I want to start course on data science. So I was thinking of either getting the DL or ML course by Andrew Ng on coursera. Can anyone tell me which course i am supposed to watch first? The DL one or ML

[plucky iron]

I made the Final Layer Adjustment in model -> Changed the final Conv2D layer to output a single channel with a sigmoid activation function. Ensures that the final layer of model outputs a tensor with the desired shape (1, 256, 256).
Still the same error I am getting.
Could you please look once into the updated model code?
https://paste.pythondiscord.com/5FJQ

I am working on a Copy-Move Image Forgery Detection DL model. I can give you full context of the code if u want for reference so u can understand the model little bit clearer.

I mean access to the code

[pulsar field]

@serene scaffold How can I start off with AI and machine learning from scratch? I have foundational python skills, and i kinda like maths and statistics, so what could be the roadmap??

[weary timber]

tell me if im wrong. i saw in 3b1b's video that you can calculate how the previous neurons activation should change in order to lower the cost at output. dont we use that value as desired and calculate the cost for that layer like we did at output? (current - desired)^2

if not please explain to me how exactly we do it im really stuck

i can calculate all but stuck at this

[serene scaffold]

Look at zestar's message in the pins

[weary timber]

@serene scaffold maybe you can help me with this. (sorry if im bothering too much tell me if i at any point)

[serene scaffold]

This is what we talked about yesterday. It seems that you understand it

What is your question?

[weary timber]

@serene scaffold im asking if everything i said here is correct

[serene scaffold]

There isn't a "cost for each layer"

[weary timber]

so i just use the cost at output everywhere?

[ocean hinge]

Anyone?

[weary timber]

can you (if you have time) explain to me how that actually works im so confused

in one video it says you propagate the error back

now you say this my mind is mixed

[wooden sail]

that's correct. there is one cost evaluated with the final output of the network. you propagate that cost back through the chain rule

[weary timber]

can you tell me or provide me some resources over how i do that?

cuz shit im stuck

[wooden sail]

the chain rule on wikipedia should be clear enough, i think

https://en.wikipedia.org/wiki/Chain_rule

[weary timber]

thanks i will check it out

[hidden fable]

Can someone make an basic script so I can see how I can start making AI

[wicked pine]

Can someone provide a link to all the available transfer learning models?

like from tensorflow hub

am struggling to find that particular link

[serene scaffold]

"AI" is a very general concept. What kind of AI do you want to make? (Don't say "Jarvis")

[hidden fable]

I just want to make an AI for my website as support

[serene scaffold]

What does "AI" mean to you?

are you talking about a chat bot? because that's a very small subset of what AI is.

[hidden fable]

Yea mb sorry

I am new into programming with Python

[serene scaffold]

That's okay. What kind of website is this?

[hidden fable]

Wdym what kind

Its not wordpress or Php

[serene scaffold]

What is the point of the website

[hidden fable]

Sorry if this is advertising, but you can check website here: https://noctyrium.online

[serene scaffold]

@hidden fable so it's a website for people to learn about your business and potentially establish a working relationship, and the bot should answer questions about what your business does?

[hidden fable]

Its like they can use our services to upgrade their company. But I want to make an AI which will make them bot so they don't need to wait

[serene scaffold]

They don't need to wait for what?

[hidden fable]

For me to give them bot and that

Like AI can do it with python (discord.py)

And I actually need chatbot too

[serene scaffold]

So your company is about making discord bots for clients?

[hidden fable]

And websites too

And it is for hosting others websites but it is not still available

[serene scaffold]

And you want to use generative AI to produce the source code for the bot or website?

[hidden fable]

Yes! It can be only for bot for now

Is it even possible to make

[serene scaffold]

Yes, but it will be worse than ChatGPT, so you shouldn't bother.

[bleak glacier]

hey, what can i do if my image segmentation model training has become super slow?

it was initially fast, but as the model got better and better the training going slower

[small wedge]

what do you mean by training slower, like the actual epochs are taking longer or the model is getting diminishing returns on its accuracy/other metrics increases?

[bleak glacier]

no the epochs run pretty fast (i run it on x2 rtx4090 and the dataset isnt super large, 7.5 images) but improvements are super slow

and minor at that

[small wedge]

are you doing any kind of hyperparameter scheduling like learning rate decay?

[bleak glacier]

its my first time training a model, i've been writing in python for over a year but pytorch is nothing like anything ive done before so im not super fimilar with all the terms, used alot of googling, samples from github and (shamefully) GPT as well

thats my scheduler
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer, mode='max', patience=3, factor=0.5, verbose=True)

[small wedge]

your factor is pretty high

the default is .1

there could be a lot happening here

[bleak glacier]

i will reduce it and try to run again

[small wedge]

beyond the fact that models will tend to plateau in performance as they approach convergence just by the nature of needing to make smaller adjustments to fit into a local optimum, the lr being cut early might be causing the model's updates to be "too careful" slowing down the convergence, or forcing it into a suboptimal local minimum

[bleak glacier]

umm im not sure what that means to be completely honest

[small wedge]

do you have a geometric intuition of gradient descent?

like a ball rolling down a hill

[bleak glacier]

im making sure to not save the model unless theres an improvment over the baseline model from which point it has started

[bleak glacier]

no

can i share the entire code here and get some suggestions on points to improve?

[small wedge]

https://datamapu.com/images/20231102_ai_ml_dl/gradient_descent.gif

[small wedge]

sure

[bleak glacier]

!paste

[arctic wedge]

Pasting large amounts of code

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

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

[bleak glacier]

https://paste.pythondiscord.com/BGZQ

[small wedge]

you can imagine the little black balls as being the model, on the yellow landscape that represents the negative gradient

each update the model moves a step down the gradient in a direction

[bleak glacier]

please keep in mind that my first ever ML related code

[small wedge]

how far that ball moves is dependant on your learning rate

[bleak glacier]

oh okay i understand

so like layers?

[small wedge]

wdym by layers

[bleak glacier]

by moving a step down, do you mean removing layers in the image?

ive read online the the decoder does that (?)

[small wedge]

https://www.youtube.com/watch?v=IHZwWFHWa-w

[small wedge]

no, I mean adjusting the model to make its outputs have a lower cost

the landscape is representing the cost of the average guess a model would make

so if it's really high on the landscape any guesses it makes will be pretty bad

and vice versa

[bleak glacier]

oh i understand

optimizing the training basically

[small wedge]

yeah gradient descent is the basis of optimizing most neural nets

[bleak glacier]

my issue isnt with time is i rent cloud gpus and can basically let them run for days, but the issue is even with LOTS of time it still barely improve

[small wedge]

another thing I'm noticing that you could mess with is your batch size

you have it set to 8 which is very low for most problems

[bleak glacier]

for example for the last 2 hours it isnt improved at all

[small wedge]

this would suggest you're stuck in a local minimum

[bleak glacier]

thats the default argument but i pass to it 30

[small wedge]

https://www.makerluis.com/content/images/2023/12/Momentum_Optimizations_Vanishing_Gradient-1.gif

[small wedge]

I see, 30 should be fine

[bleak glacier]

i made it 8 by default as i ran tests on my laptop (rtx 3090, not exactly super strong)

[small wedge]

here you can see the balls on the left don't take very big steps and end up getting stuck in a little ridge

this can represent what a low learning rate could do

batch size on the other hand determines how well your model actually estimates what the landscape looks like

a true gradient would be calculated by passing every single sample in your dataset and finding the average cost

[bleak glacier]

so higher batch size = better predictions ?

[small wedge]

but we take a shortcut in minibatch gradient descent where we only feed the model a number of samples equal to the batch size before assuming the gradient is good enough and making an update

[small wedge]

yeah basically

[bleak glacier]

i noticed that too hence i upped it to 30

[small wedge]

but bigger batch size also means slower training (less actual updates per epoch)

[bleak glacier]

tried 50, it ran out of memory xD

[small wedge]

yikes lol

[bleak glacier]

maybe i should hire someone to do this for me, im too clueless at ML, i usually do web related stuff (django)

this is way complex and require actual scientific knowlage

[small wedge]

yeah training them can be tricky

plus you seem to have jumped in the deep end with your first model

[bleak glacier]

its no joke! harder than any python stuff ive done before, even C extentions are easier

[small wedge]

most people do a simple mnist classifier for their hello world model

[bleak glacier]

well im trying to train a model for background removal and publish it for free under MIT license, and keep it lightweight

most existing models are heavy, and are not free for commercial use

[small wedge]

quantization could be an avenue to look down in that case, although that will require diving even deeper into the science rabbit hole

basically you can take really big models trained with standard float precisions (ex. 32) and then once they are trained convert their weights to small precisions like 1, 2, 3, 4, 8 bit float/integers

and you only lose a few percent of preformance on benchmarks if done right, but reduce the size and compute required to use the model by many fold

[bleak glacier]

mine currently get around 0.9763% accuracy (where 1 being the highest) and weights only 50mb but its a super slow progress in the past 2 days

[bleak glacier]

thats seems even more complex than what im doing now xD

[small wedge]

yeahh

[bleak glacier]

maybe i should just increase the dataset from 7.5k images to 20k

although getting good datasets is a pain

[small wedge]

oh you only have 7.5k?

[bleak glacier]

yep

[small wedge]

the dataset is the foundation, no matter how good your model is without a good dataset you will get shit results

[bleak glacier]

well its not even that bad now

maybe thats the main issue then, the dataaset

[small wedge]

have you tried any augmentation?

oh nvm you do

a lot

[bleak glacier]

lol yeah some transform stuff

[small wedge]

actually almost an alarming amount, another thing is to be 100% sure that none of those features you're adding noise to could be useful to the model

[bleak glacier]

well i do need the end model to handle any quality, zoom, crop, color image as the end user will upload whatever

[small wedge]

say for example I'm classifying a picture as either an apple or a banana, it's okay for me to rotate those images because the model should be able to tell the difference no matter the orientation, but say I now randomize all the colors on screen; I'm taking away a large portion of what makes apples and bananas easily distiguishable by adding noise there

[small wedge]

yeah it might be unavoidable in your case, the only real solution to what I'm mentioning would probably be more real samples

[bleak glacier]

But the model doesn't classify an object type

It just classifies what's the main object and what's the background

[small wedge]

yeah I gotchu, I'm just giving an example

saying you have to be sure that all of ```
A.Resize(512, 512),
A.HorizontalFlip(p=0.5),
A.VerticalFlip(p=0.5),
A.ColorJitter(brightness=0.2, contrast=0.2, saturation=0.2, hue=0.1, p=0.5),
A.RandomRotate90(p=0.5),
A.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225)),

[bleak glacier]

It gives decent results, I'll show ya

That's not super bad is it? And keep in mind that's an image he was not trained on

[serene scaffold]

remember that if you post images, they should be closely cropped around the part of interest.

[small wedge]

looks pretty good to me

[bleak glacier]

And it only weights 50mb, comparable models weight north of 1gb and are not under MIT license

[bleak glacier]

I'll keep that in mind, cheers!

[bleak glacier]

Yeah so I think I got a good shot to actually do somthing good here and produce an actual free and open source background removal model / tool, just getting stuck on the last bit of model training which is frustrating

[spare badger]

Hi guys. I need some help. I've one-way multi-layer neural
where i need y_t as function y_t=sin(PI*x1 *x2)
for x1,x2 = <-1,1>. And i need these graphs drawed while my "AI" is learning, but don't know why it isn't working

import numpy as np
import nnet as net
import matplotlib.pyplot as plt


x = np.array([[-1, -0.9, -0.8, -0.7, -0.6, -0.5, -0.4, -0.3, -0.2, -0.1, -0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0]])
y_t = np.array([[1.2246467991473532e-16, 0.5620833778521305, 0.9048270524660195, 0.9995065603657316, 0.9048270524660196, 0.7071067811865476, 0.4817536741017153,
                 0.2789911060392293, 0.12533323356430426, 0.03141075907812829, 0.0, 0.03141075907812829, 0.12533323356430426, 0.2789911060392293, 0.4817536741017153,
                 0.7071067811865476, 0.9048270524660196, 0.9995065603657316, 0.9048270524660195, 0.5620833778521305, 1.2246467991473532e-16]])


max_epoch = 200000
err_goal = 1e-50
disp_freq = 1000
lr = 0.01
L = x.shape[0]
K1 = 13
K2 = 7
K3 = y_t.shape[0]
SSE_vec = []
w1, b1 = net.nwtan(K1, L)
w2, b2 = net.nwtan(K2, K1)
w3, b3 = net.rands(K3, K2)
# hkl.dump([w1,b1,w2,b2,w3,b3], 'wagi2w.hkl')
# w1, b1, w2, b2 ,w3, b3 = hkl.load('wagi2w.hkl')



for epoch in range(1, max_epoch + 1):
    y1 = net.tansig(np.dot(w1, x), b1)
    y2 = net.tansig(np.dot(w2, y1), b2)
    y3 = net.purelin(np.dot(w3, y2), b3)
    e = y_t - y3
    d3 = net.deltalin(y3, e)
    d2 = net.deltatan(y2, d3, w3)
    d1 = net.deltatan(y1, d2, w2)

    dw1, db1 = net.learnbp(x, d1, lr)
    dw2, db2 = net.learnbp(y1, d2, lr)
    dw3, db3 = net.learnbp(y2, d3, lr)
    w1 += dw1
    b1 += db1
    w2 += dw2
    b2 += db2
    w3 += dw3
    b3 += db3

    SSE = net.sumsqr(e)
    if np.isnan(SSE):
        break

    SSE_vec.append(SSE)

    if SSE < err_goal:
        break
    if (epoch % disp_freq) == 0:
        print("Epoch: %5d | SSE: %5.5e " % (epoch, SSE))
        plt.clf()
        plt.plot(x[0],y_t[0],'r',x[0],y3[0],'g')
        plt.grid()
        plt.draw()
        plt.pause(1e-2)

print("Epoch: %5d | SSE: %5.5e " % (epoch, SSE))

fig = plt.figure(figsize=(8,8))
ax = fig.add_subplot(111, projection='3d')
surf = ax.plot_surface(x, y_t, x, cmap='viridis')

plt.ioff()
plt.figure()
plt.plot(SSE_vec)
plt.ylabel('SSE')
plt.yscale('linear')
plt.title('epoch')
plt.grid(True)
plt.show()

[lapis sequoia]

everything going on here is cool

im a beginner

how can i learn

[faint quail]

trying to train a neural network is like trying to play golf but you only know the general direction of the hole, and you have to guess how much power to use, what type of club to use and the conditions of the field

[bleak glacier]

Well I'm increasing the dataset to 40k let's see if that helps 😂

From 7.5k to 40k, that hopefully will do the job

[faint quail]

Before trying to go on to a larger and probably more complex dataset, see if you can overfit a much smaller dataset of a few select examples so you can train much faster and see if its a dataset issue, this also makes it easier to find good hyper-parameters

there may be bad examples in your dataset especially if it's self labeled

[unkempt wigeon]

I am beginning to machine learning or coding in general? My apologies

[faint quail]

you should defintely learn the fundamental ideas of ml
here's some videos that helped me learn:

https://www.youtube.com/watch?v=aircAruvnKk&list=PLZHQObOWTQDNU6R1_67000Dx_ZCJB-3pi
https://www.youtube.com/watch?v=hfMk-kjRv4c
https://www.youtube.com/watch?v=Lakz2MoHy6o

then learn the frameworks like tensorflow or pytorch for actually programming these models, start with the simplest of models with an XoR, then i'd reccoment move on to a mnist solver and then from there wherever you want

[unkempt wigeon]

https://pytorch.org/docs/stable/index.html my apologies

How far should I read into pytorch?

[serene scaffold]

completely stop reading pytorch docs until you figure out what you're trying to do.

[unkempt wigeon]

MNIST

[wicked pine]

If my pre-trained model is trained on images with the size 224, 224 for example, will i be forced to stick with that size or i can adjust it to be for example 150,150?

[serene scaffold]

I don't specialize in image-related AI, but I imagine you'd need adapter layers, or something

[wicked pine]

okok

[unkempt apex]

For MNIST , there are lot of articles to start out

So read them

[cursive wing]

some help me to make this data beutiful like data would be aligned in sigle rows rather then a huge screen and rows should be in black and white background colour

sorted out alhamdulillah

[normal meadow]

Hey, I was wondering after learning pandas and numpy, how should I start building real world projects?

I am new to the field so I would like a some sort of guidance

[delicate edge]

Do you have something you are into that you would like to apply your newfound knowledge on? I would start with something tangential or maybe something that relates to the industry you are trying to get into

[normal meadow]

Well not really, but I’ve done a quick research and results were to practice and apply what I know with kaggle

But I don’t have something particular in mind I want to get into

[delicate edge]

Kaggles not a bad place to start either, it's got a pretty bright community and you'll learn lots of cool things reading other peoples code while competing.

Participating in any of those competitions is pretty cool on it's own. I've personally have been able to use past competitions in job interviews even if I didn't place super high on any of the leaderboards

[normal meadow]

Sounds great tbh

So I can use kaggle in my CV and job interviews?

[delicate edge]

I don't see why not

[normal meadow]

Kaggle projects done through competitions

[delicate edge]

Just be ready to speak to your work

[normal meadow]

I am still a high schooler 😂😭

[delicate edge]

Hey the earlier the better! Right on.

[normal meadow]

True

Thanks for the help mate ✌️❤️

[delicate edge]

Anyone here familiar with pyspark? I'm starting out a new project for work and having troubles finding resources on project structures and organizational patterns.

I've kinda defaulted to the typical project structure. Where my main.py runs code in each of my packages.

foo/
-src/
--package1/
--package2/
--main.py

I don't love having to use magic commands to run code as I'm finding it causes issues with imports and variables being changed inadvertedly.

Anyways, just looking for resources on how to structure projects, I would also eventually love to implement unit tests in my codebase. So will take any suggestions!

[last nest]

Hey, any recommendations for courses or resources to learn pandas and xarray, from basic to advanced ?

[trail otter]

Here is a youtube small course I watched to get rather good in pandas, covers a lot of things: https://www.youtube.com/watch?v=8xUgesdShE8

[valid void]

Does anyone know what is the size of the tensorflow gpu and pytorch gpu libraries?

[last nest]

oh thanks a lot ! I can't return the favor for the ai field sorry 😦

[trail otter]

Haha, it is fine, I ll hope someone else can cover that

[ocean hinge]

Hello can someone explain what cost function is? and how it is represented in 3d?

[ocean hinge]

I get that the formula is y-hat(prediction) minus target value(actual value). So is it how wrong or the error between the prediction and target value for the same input?

[agile cobalt]

most of the time when you see something represented in 2d / 3d, that's an oversimplification

the cost function is what tells the model how wrong it is, and that value is used to determine how much it needs to change its parameters (as well as which parameters to change in which direction)

quoting https://developers.google.com/machine-learning/glossary#loss-function
(cost function and loss function are synonyms)

During training or testing, a mathematical function that calculates the loss on a batch of examples. A loss function returns a lower loss for models that makes good predictions than for models that make bad predictions.

The goal of training is typically to minimize the loss that a loss function returns.

Many different kinds of loss functions exist. Pick the appropriate loss function for the kind of model you are building.

[trail otter]

Hello I want to get into the AI field, but dont know were to start. Does anyone have a good roadmap towards learning machine learning and AI? However, I have heard that Andrew NG has a good course on corsea, but should I take any courses before that, or anything after that course - and how long do you think it would take to complete that course? Btw, I know a lot in pandas and numpy already. Or do you guys have a completely other roadmap you can recommend for a noobie like myself?

[ocean hinge]

Had the same question. I just started AI. 😆
I would suggest look at pins in this thread, Look at where you stand. Try the free courses first and then get the andrew course on coursera. I did the mistake of getting DL course first, You are supposed to learn ML first and then DL.

[spring hamlet]

Hello, i am trying to scrape several data on the brand "Polene", i'm still a beginner in python. i tried with reddit and get the error : raise ResponseException(response)
prawcore.exceptions.ResponseException: received 401 HTTP response, you can see my code on the screen

[lapis sequoia]

!code

[arctic wedge]

Formatting code on Discord

Here's how to format Python code on Discord:

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

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

For long code samples, you can use our pastebin.

[lapis sequoia]

Please copy the code, don't post a screen shot...

[left tartan]

401 means authorization error/failure. Your tokens are probably wrong (I'm guessing, no experience with that lib). Check the examples for that library and check your keys/etc.

[left tartan]

Open a #❓｜how-to-get-help help thread if you need more help, this isn't really on topic for this channel.

[spring hamlet]

ok!

[weary timber]

should i give up on trying to create a neural network without looking at any code if my brain isnt functioning and i have been stuck for the past 1 week?

[elder anchor]

@weary timber When I started working with neural networks, at the start I had lots of trouble to understand what was going on. What worked to me what to take a step back, research how neural networks work and I created an entire Neural Network training algorithm from scratch on my own. That helped me a lot to understand how they worked. I think doing something similar can help you a lot.

[weary timber]

who tf reacted to this lol

[weary timber]

i cant seem to understand the part where we have to propagate the error backwards.

[elder anchor]

What I mean is that you might be stuck because you are trying to understand complex stuff before understanding properly the basics.

So understanding everything from the lowest level might help you

[weary timber]

can you provide me a resource that made you understand?

[elder anchor]

https://youtu.be/bfmFfD2RIcg?si=FHD5e6T5QrcSn28N and https://youtu.be/aircAruvnKk?si=--BkVmGDgjwQKucV

These two helped me a ton

[weary timber]

i have already watched the 3b1b so i will look at the first one

[elder anchor]

The first one is very simple and a bit childish, but it was my first step and it built a great base for me

[brazen badger]

docker or kubernotes for a beginner in ml ?

[lapis sequoia]

What problem are you trying to solve? Kubernetes uses docker (most of the time). You probably don’t need either until you get into some serious ML

[brazen badger]

well, im trying to start a career in ml

sooooooooooooooooooooooooo

which one ?

[lapis sequoia]

Don’t worry about k8s or docker for a while. Most of your day to day will not be using either. The infrastructure team will handle that

[brazen badger]

it will help for the resume

[lapis sequoia]

But you should know docker before k8s or learning k8s will be way harder

[brazen badger]

after evalualition i need to use that, no ?

docker

ok great

thanks

[lapis sequoia]

Learn docker, then k8s. But there are a lot of other things that are more important in a DS/ML resume imo.

[brazen badger]

yeahhh

do i type the hyperlinks to my notebooks in github or kaggle ?

in my resume

or the project names and details or both

which area do u work in ?

[lapis sequoia]

Basic docker isn’t hard. You can get a decent grasp in a week or so. Advanced docker can be brutal. You get into some deep Linux networking and kernel level stuff.

[calm thicket]

I don't think there's any reason for a person to learn docker while getting into machine learning

[brazen badger]

hmmmmm

alright thanks

one thing off the list to learn lol

[lapis sequoia]

I definitely agree with this. You could learn docker and k8s if you have been applying for jobs for months with no responses and you want something extra to standout. But it is very unusual for a DS person to do any infrastructure like that. There are normally teams who handle it.

That being said… docker makes development and production deployment much easier even for example projects. I HATE when people try and give me Jupyter notebooks.

[calm thicket]

there are a lot more careers in ML than just DS to be fair. an mle for example might frequently use docker. but when just starting in ML, docker is not really relevant

[lapis sequoia]

Exactly

[limber spear]

is Cuda a programming language? How come GitHub lists it as a programming language

[serene scaffold]

if it is, it's not one that you need to know

[lapis sequoia]

Cuda is c

[serene scaffold]

CUDA is just the API for the GPU as hardware. you don't really have to think about it except for when you install the CUDA drivers or when you update pytorch, jax, etc.

[limber spear]

Well GitHub lists it as a language which misleads my belief in what Cuda actually is

😂

i said programming language idk im sure the programming community can chime in on that

[iron basalt]

It's the Nvidia Graphics API/SDK and also a language (both under the same name "CUDA"), which is very similar to C++.

The implementation of that language is really ugly, it kind of high-jacks the C++ compiler.

You don't need to know it unless you want to write compute shaders (programs that run on the GPU).

Usually there are a bunch of these programs (kernels) written for you and wrapped by a library, e.g. pytorch.

(Or generated as needed by some libraries for you)

(There are other options too, such as OpenCL, Vulkan (but on Nvidia GPUs, CUDA will always have priority on new features))

[faint quail]

#data-science-and-ml message

[stone coral]

When should I use DRL vs Deep Learning?

I've looked it up and genuinely there are mixed answers and I really could go either or with my project on sports prediction. The way im looking at this is I have 10 years of data. I implemented Deep Learning which is fairly ok. But is it worth pushing to do Deep Reinforcement Learning? I can just use all those years to try to see if it can learn on its own. But again Im not sure what to do

[rugged wraith]

Hey guys so I have a doubt with NLP, I am working on text classification (it's binary classification) and I'm using BioBert for the text embedding and it seems that LSTM has less to no learning rate, the predictions are all stagnant i.e there's no variance in prediction. However when I use the Bidirectional layer on LSTM the results are great with a 90%+ accuracy and a confidence level (0.9) of 90%. Is having a bidirectional layer a must for bert embeddings?

[unkempt apex]

I think it depends on your project
so tell more about your sports prediction project

[coarse nacelle]

How can I improve the images? So far I've done this -- Is it possible to refine the images any further than this?

Here's what my model does:

def build_text2image_model():
    encoder_inputs = Input(shape=(max_query_length, len(unique_characters)))

    encoder = LSTM(256, return_state=True, return_sequences=False)
    encoder_outputs, state_h, state_c = encoder(encoder_inputs)

    decoder_inputs = RepeatVector(max_answer_length)(encoder_outputs)
    decoder_lstm = LSTM(256, return_sequences=True)
    decoder_outputs = decoder_lstm(decoder_inputs)

    decoder_dense = TimeDistributed(Dense(7 * 7 * 64, activation='relu'))
    decoder_outputs = decoder_dense(decoder_outputs)

    decoder_reshape = TimeDistributed(Reshape((7, 7, 64)))
    decoder_outputs = decoder_reshape(decoder_outputs)

    # add Conv2DTranspose layers with batch normalization and activations
    decoder_deconv1 = TimeDistributed(Conv2DTranspose(64, (3, 3), strides=(2, 2), padding='same'))
    decoder_outputs = decoder_deconv1(decoder_outputs)
    decoder_outputs = TimeDistributed(BatchNormalization())(decoder_outputs)
    decoder_outputs = TimeDistributed(Activation('relu'))(decoder_outputs)

    # upsample to 28x28
    decoder_deconv2 = TimeDistributed(Conv2DTranspose(32, (3, 3), strides=(2, 2), padding='same'))
    decoder_outputs = decoder_deconv2(decoder_outputs)
    decoder_outputs = TimeDistributed(BatchNormalization())(decoder_outputs)
    decoder_outputs = TimeDistributed(Activation('relu'))(decoder_outputs)

    # final output layer
    decoder_deconv3 = TimeDistributed(Conv2DTranspose(1, (3, 3), activation='sigmoid', padding='same'))
    decoder_outputs = decoder_deconv3(decoder_outputs)

    text2image_model = Model(encoder_inputs, decoder_outputs)

    text2image_model.compile(optimizer='adam', loss='binary_crossentropy')

    text2image_model.summary()
    return text2image_model

[unkempt apex]

ahh interesting

so what are your inputs?
numbers or images

like, to get that output image as "113" are you using images or just numbers for addition operations?

[latent dirge]

is this a good place to ask about numpy?

[serene scaffold]

it's the best place

[latent dirge]

right, so I wrote this little piece

def part2(self):
  floor = 0
  position = 0
  # self.instruction is a ndarray
  for step in self.instructions:
      floor += step
      position += 1
      if floor == -1:
          break
  return position

I feel like there should be a better way to utilize the numpy array to get the position, but I'm not terribly familiar with numpy's idiosyncrasies to know if that's the case

[unkempt apex]

heh what is instruction array?

just give a small example

[latent dirge]

it's an array of ints with 3 different values: 1, -1 and 0, like [1, 1, -1, 0]

[calm thicket]

is this aoc? don't you just want to sum the entire array?

ah sorry. you want to do a cumulative sum and find where it gets to -1

[latent dirge]

ah, you know it, then

yes, exactly

[unkempt apex]

there is cumsum right

for numpy?

[calm thicket]

yes

[latent dirge]

well, how exactly do I use it? don't see a conditional for when the sum reaches -1

[calm thicket]

why are you using numpy for this?

[latent dirge]

used it for the first part, because it was just that, a sum of the whole thing

[coarse nacelle]

yes, its like query in integers and output is images. I made some improvement, does it look promising than the last?

[latent dirge]

what's the operation called for numpy, when you have 2d array, and each subarray consists of n elements and each element needs to be multiplied with another?
example:
[2, 5, 6] -> [2*5, 2*6, 5*6] -> [10, 12, 30]

[serene scaffold]

You just do normal * and you get that behavior automatically

Wait, you need to multiply every combination of two elements?

[latent dirge]

yes, so you get from [2, 5, 6] to [10, 12, 30]

[unkempt apex]

yeah, what improvements do you make?

[coarse nacelle]

I added Conv2DTranspose layers with upsampling and residual connections instead of batch normalization and activations, and converted grayscale to RGB for VGG16 perceptual loss then finally I reducded nooise

Also, I think this is good because last time when I put this in the supervised training model it yield only 47% accuracy now I got 85%

[dire horizon]

hello python community

how can i learn python for data science?

[serene scaffold]

Look in the pins

Keep in mind that data science is an academic subject unto itself. You have to think of yourself as learning data science, and using python to implement what you're doing. The code is not the objective.

[limber spear]

Hmm good question. I’m actually stumped. Besides here, anyone know? Coursera has really good certificate programs. Tho they are a couple months long. Tho I believe there are shorter term certificate courses

But it is not free to be transparent. Maybe some are?

[stone coral]

Should I use multiple targets in training to get all predictions at once or seperate them by each target? Im not sure if this affect accuracy in any way or not

I have 10 inputs and 10 targets.

[unkempt apex]

U wanna classify the 10 targets?

[limber spear]

That’s pretty much a multiclass classification task no?

mb it’s multioutput classification

[coarse nacelle]

I am trying to improve the validation accuracy of this model but I am not sure where I am wrong.
loss: 0.0119 - accuracy: 0.9976 - val_loss: 4.7028 - val_accuracy: 0.3508

This is what I did so far.. is it even possible to improve validation accuracy?

def build_image2text_model():
    input_shape = (X_img.shape[1], X_img.shape[2], X_img.shape[3], 1)

    image2text = tf.keras.Sequential()
    
    image2text.add(
        ConvLSTM2D(
            filters=64,
            kernel_size=(3, 3),
            activation='relu',
            input_shape=input_shape
        )
    )
    image2text.add(Dropout(0.2))
    image2text.add(Flatten())

    image2text.add(RepeatVector(max_answer_length))
    image2text.add(LSTM(256, return_sequences=True))
    image2text.add(Dropout(0.2))
    image2text.add(TimeDistributed(Dense(len(unique_characters), activation='softmax')))

    image2text.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
    image2text.summary()

    return image2text

image2text_model = build_image2text_model()
image2text_model.fit(
    X_img, y_text_onehot,
    epochs=40, batch_size=128,
    validation_split=0.1
)

[unkempt apex]

anyone have tried tensorflow object detection API?

I am trying to to fine-tune the model but getting version issues, although tried every possible way I guess

[small wedge]

Seems like you might be overfitting, how big is your dataset?

[weary timber]

!code

[arctic wedge]

Formatting code on Discord

Here's how to format Python code on Discord:

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

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

For long code samples, you can use our pastebin.

[weary timber]

def backpropagation(output,desired,learning_rate):
    dZ2 = cache["A2"] - desired.T
    dW2 = np.matmul(dZ2,cache["A1"].T)
    dB2 = np.sum(dZ2,keepdims=True,axis=1)
    dA1 = np.matmul(params["W2"].T,dZ2)
    dZ1 = dA1 * sigmoid(cache["Z1"],True)
    dW1 = np.matmul(dZ1,cache["A0"])
    dB1 = np.sum(dZ1,keepdims=True,axis=1)
    params["W1"] -= learning_rate*dW1
    params["W2"] -= learning_rate*dW2
    params["B1"] -= learning_rate*dB1
    params["B2"] -= learning_rate*dB2

this uses the backprop equations everyone uses but i somehow get an error, can this be fixed?
the error says that dz2 and a1 cant be multiplied since they are different sizes

@serene scaffold sorry for ping

@wooden sail

[serene scaffold]

thank you for showing the code. be sure to always show the full error message as well.

[weary timber]

ok wait

[weary timber]

sorry for the wait i deleted the training data so im loading again to get error

[weary timber]

Traceback (most recent call last):
File "c:\Users\mehme\OneDrive\Desktop\neural network\nlfromscratch.py", line 40, in <module>
feedforward(x)
File "c:\Users\mehme\OneDrive\Desktop\neural network\nlfromscratch.py", line 21, in feedforward
cache["Z1"] = np.matmul(params["W1"],cache["A0"].T) + params["B1"]
~~~~~~~~~~~~~~~~~~~~~~~~~~^~
File "C:\Users\mehme\AppData\Local\Programs\Python\Python312\Lib\site-packages\pandas\core\ops\common.py", line 76, in new_method
return method(self, other)
^^^^^^^^^^^^^^^^^^^
File "C:\Users\mehme\AppData\Local\Programs\Python\Python312\Lib\site-packages\pandas\core\arraylike.py", line 186, in add
return self._arith_method(other, operator.add)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\mehme\AppData\Local\Programs\Python\Python312\Lib\site-packages\pandas\core\frame.py", line 7910, in _arith_method
self, other = self._align_for_op(other, axis, flex=True, level=None)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\mehme\AppData\Local\Programs\Python\Python312\Lib\site-packages\pandas\core\frame.py", line 8141, in _align_for_op
right = to_series(right)
^^^^^^^^^^^^^^^^
File "C:\Users\mehme\AppData\Local\Programs\Python\Python312\Lib\site-packages\pandas\core\frame.py", line 8133, in to_series
raise ValueError(
ValueError: Unable to coerce to Series, length must be 70000: given 64

im using mnist from sklearn

[serene scaffold]

@weary timber look at the shapes of params["W1"] and cache["A0"].T

[weary timber]

they are not matching

but the thing i dont understand is i checked 2 codes and everyone uses the same as i do without an error

[serene scaffold]

do you know what the rule is for what matrix shapes can be multiplied

[weary timber]

idrk

[serene scaffold]

you should look into that.

[weary timber]

i understand the rules but how do i come up to a fix with them

[serene scaffold]

what is the rule?

[weary timber]

the number of columns on first matrix's should be equal to the second matrix's row number

a(m,n) b(n,p)

can be multiplied

[serene scaffold]

okay, and what are the snapes of your two matrices?

[weary timber]

W1 is 64x784 and A0 784

[serene scaffold]

784 by what?

[weary timber]

1

[serene scaffold]

and is that before or after you transpose it?

[weary timber]

what is transpose?

[serene scaffold]

the .T. if you have a matrix of shape (m, n), then it's transpose is where it's (n, m). the values are the same, just rearranged.

rows become columns, columns become rows.

[weary timber]

ohhh no its before

[serene scaffold]

if the shape of A0 is (784, 1) then what is the shape of its transpose?