data-science-and-ml

[violet gull]

line 20, its not

[buoyant vine]

Oof why did my copy have it commented out

[violet gull]

since when does pygame throw seg faults

[buoyant vine]

😔 Copy paste did me dirty

Well I dont think it chooses to

It is likely it trying to either talk to a DLL which isn't there

or some UB internally

[spring field]

since always, they're a real treat too, but it might not be pygame that's throwing it, it might just be catching it from elsewhere

[violet gull]

how do i fix it

[buoyant vine]

Hope's and prayers

and a lot of commening lines in and out

🙏 Welcome to UB related errors

[violet gull]

should this be a bug report

[spring field]

as I said, they're a real treat

[violet gull]

UB?

[buoyant vine]

Undefined Behaviour

[spring field]

can you share the segfault?

[buoyant vine]

I.e. Null pointer, out of bounds memory access, invalid DLL read, thje list goes on

[spring field]

(also pygame-ce 😁)

[violet gull]

Segmentation fault (core dumped)

[buoyant vine]

The main issue with them is just because one line runs into the segfault, it does not mean that line caused it

[spring field]

!trace

[arctic wedge]

Traceback

Please provide the full traceback for your exception in order to help us identify your issue.
While the last line of the error message tells us what kind of error you got,
the full traceback will tell us which line, and other critical information to solve your problem.
Please avoid screenshots so we can copy and paste parts of the message.

A full traceback could look like:

Traceback (most recent call last):
  File "my_file.py", line 5, in <module>
    add_three("6")
  File "my_file.py", line 2, in add_three
    a = num + 3
        ~~~~^~~
TypeError: can only concatenate str (not "int") to str

If the traceback is long, use our pastebin.

[violet gull]

there is no stack trace

[buoyant vine]

it can be the result of a earlier action that then created the undefined behaviour

[spring field]

what's the output then?

[violet gull]

this

[spring field]

that can't be all of the output

[violet gull]

pygame 2.5.2 (SDL 2.28.2, Python 3.11.9)
Hello from the pygame community. https://www.pygame.org/contribute.html
/home/sarati/Programming/rust/ven311v/lib64/python3.11/site-packages/torchrl/data/replay_buffers/samplers.py:37: UserWarning: Failed to import torchrl C++ binaries. Some modules (eg, prioritized replay buffers) may not work with your installation. If you installed TorchRL from PyPI, please report the bug on TorchRL github. If you installed TorchRL locally and/or in development mode, check that you have all the required compiling packages.
  warnings.warn(EXTENSION_WARNING)
device: cpu
Segmentation fault (core dumped)```

that warning comes with and without the seg fault

[spring field]

can you switch to pygame-ce real quick, because the other pygame removed parachute

(and also because it's better anyway 😁)

pip uninstall pygame
pip install pygame-ce

[buoyant vine]

Unrelated but what is the backstory behind pygame and pygame-ce 😅 Last time I used either, pygame was the only one around

[violet gull]

!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.

[buoyant vine]

Tbh if I am debugging seg faults, a lib saying it failed to load C/C++ DLLs would be a good place to start

[violet gull]

https://paste.pythondiscord.com/SCLA

i see a pygame parachute error

[buoyant vine]

Well that is a much better error

[violet gull]

my environement is duct taped together im afraid to touch it

[buoyant vine]

🎉 That is a dream of a seg fault

[violet gull]

wat

[spring field]

never seen one that long

[buoyant vine]

That full dump gives way more info than just "Segfault xD good luck"

Are you on linux?

[violet gull]

unfortunately

[buoyant vine]

My guess

is torch is doing some threading related stuff

and pygame is not happy about making various calls from different threads

Does anything change if you set:

torch.set_num_threads(51
torch.set_num_interop_threads(1)

[violet gull]

so i found integration bug?

[buoyant vine]

Well, you have (maybe) found that not everything in Python is threadsafe

[violet gull]

no

[rich moth]

I swear i got this one time and it was releated to the autograd in the c++ python files.

[lapis sequoia]

wait, how do you find the the size of each word embedding?

[spring field]

well, the TL;DR is that the maintainer of that repo decided to ban all the core contributors due to disagreements regarding 3.11 release... so yeah, we just moved on to create a fork and continue development as pygame-ce pretty much, you can read more into detail here: https://www.reddit.com/r/pygame/comments/18xy7nf/what_was_the_disagreement_that_led_to_pygamece/

[buoyant vine]

interesting

[buoyant vine]

normally the embeddings are all the same size

[buoyant vine]

Sorry I typo'd I meant to set each to 1 idk if you did that or not

[spring field]

either see what the relevant papers use or experiment

[violet gull]

still error

[lapis sequoia]

only choice and only solution in all honesty

[spring field]

what?

[buoyant vine]

Rip, Err you could possibly try launch a debugger that can show the executing threads

I know pycharm supports it

idk of standalone ones or ones that aren't premium/paid for

[violet gull]

@spring field mr pygame dev do you know how to fix it?

[spring field]

it's not an issue with pygame(-ce) as far as I can see

[buoyant vine]

Other thing you can try is turn the pygame logic into a single threaded actor

and send updates via a queue

so all the pygame logic sticks to the same thread

I need to go to sleep though so I can't help much more than that until tomorrow 😔

[violet gull]

i appreciate you trying

im on a bit of streak of getting random failures

[violet gull]

even though its caused by a pygame line?

[buoyant vine]

It isn't pygame itself

It is likely a threadsafety issue

[violet gull]

where i gotta put bug report?

[buoyant vine]

Unless you get the error on a smaller file using on pygame and no pytorch

[buoyant vine]

I suspect it is not going to be labelled as a fixable bug regardless

[spring field]

where do you see it being caused by pygame code?

[buoyant vine]

It will be put under "Don't use this library across multiple threads"

Just to be sure, make another file that does the basic pygame logic without all the AI and pytorch stuff

[violet gull]

#screen = pygame.display.set_mode((screen_width, screen_height)) this line commented out fixes it

[buoyant vine]

then do the same for Pytorch

then that will probably tell you if it the specific combo of PyTorch threading + Pygame causing the issue

Pytorch does a lot of magic that can cause thread safety issues and various UB, so I would probably put my money on the two libs not playing nicely together

[violet gull]

but without doing weird thread duct taping im am screwed?

[deep sleet]

That is gives them more significance?

[rich moth]

You tried? torch.set_num_threads(2)
torch.set_num_interop_threads(2)

[violet gull]

i tried 1 and 1

[spring field]

can you put a bunch of prints around to see how far the code gets before segfaulting and send that code over as well

[violet gull]

i can do you better

hold omn

import pygame
import torch
from torch import nn
screen = pygame.display.set_mode((9, 9))
class BearBrain(nn.Module):
    def __init__(self):
        super().__init__()
brain = BearBrain()
optimizer = torch.optim.Adam([1], lr=1)``` @spring field 9 lines of code

causes a seg fault

[spring field]

bad news, I can't reproduce a segfault, it just exits with 0 😬

can you tell how far in that code it gets?

[violet gull]

fails on the last line

[spring field]

this new code fails with a different error for me btw

[violet gull]

what error

[spring field]

TypeError: optimizer can only optimize Tensors, but one of the params is int

[violet gull]

yes thats expected

[spring field]

so I assume it segfault before getting that far for you

[violet gull]

yes

Fatal Python error: pygame_parachute: (pygame parachute) Segmentation Fault
Python runtime state: initialized

Thread 0x00007f32c56006c0 (most recent call first):
  File "/usr/lib64/python3.11/threading.py", line 331 in 
...```

[spring field]

and it doesn't segfault w/o that set_mode line?

[violet gull]

TypeError: optimizer can only optimize Tensors, but one of the params is int
correct

[spring field]

threading? interesting

[violet gull]

why it error for me but not you

[spring field]

well, for one, I'm on Windows, you're on Linux

I'll try it on a linux machine though rq

also I'm on Python 3.12

[rich moth]

Well, at least we figured out its probably your env and not your code 🙂

[violet gull]

as usual

oh i got it to work in a new env

with just pygame-ce and torch

[rich moth]

awesome!

[violet gull]

on 3.12

[rich moth]

What a pain in the ass, lol.

[violet gull]

ive been at this since 5 o clock yesterday

very much a pain in the ass

[rich moth]

https://tenor.com/view/dog-smile-menacing-feeling-good-no-eyes-gif-15692697

[violet gull]

this also kinda just kills my entire project

for something i cant fix

[rich moth]

I guess I missed something. You needed 3.11 for this sto work?

[violet gull]

yes most of the torch libraries arent updated

[rich moth]

I see, damn

[violet gull]

do you think i should submit a bug report. I need this fixed

[spring field]

you're facing segfaults when using something more than just torch and pygame?

[violet gull]

yeah do you want my pip freeze

[rich moth]

Maybe inside your 3.12 instance, you can install the torch and pygame-ce for that compatible 3.11 version? Im just brainstorming here.

Are you using python virtual env or conda? I cant tell.

[violet gull]

venv

[spring field]

mmm, not really no, well, I'm fairly certain it's not an issue (at least not directly) with pygame-ce
am honestly not sure what you can do at this point sigh

[rich moth]

Maybe for shits and giggles build it in conda nad try it?

[spring field]

but if you do decide to report it as a pygame-ce issue, you can do it here: https://github.com/pygame-community/pygame-ce/issues
I'm just fairly certain that it will be closed because we can't directly really do anything about it, I don't think
you can also ig report it to whatever else you suspect of being at fault, it seems that triton (from openai) figured a lot in that segfault trace, it's still in development it seems, but, perhaps, onnx is using it? so ig onnx would be another project to report this to?

but even then it's probably gonna take a while to get it fixed, perhaps, try hacking something together with threads, lol

[rich moth]

Hey just curious what cuda verison are you using?

Maybe different drivers can help too

[spring field]

mmm, you could also try moving development to a docker container

[rich moth]

What verison are you using Matiiss?

Or is this just cpu related?

[spring field]

I have no idea what this is related to

[rich moth]

Im just curious your version and his. Since you got it to work.

[violet gull]

people always laugh at me for not wanting to use libraries and this is a prime example why

[rich moth]

well be sure to document this for them then 😄

[violet gull]

i submitted a issue to pytorch but like all other issues i face. I am the only one who can ever recreate it

[rich moth]

Did you try it with conda?

[violet gull]

idek what that is

[rich moth]

its a virtual env manager with depedency managment

I think its worth giving a shot,

[violet gull]

id rather just cry and eat blueberries

if only rust had machine learning

[small wedge]

heh

I've been working on an ML lib in rust for a while

but I haven't touched it in a long time

[rich moth]

bash Miniconda3-latest-Linux-x86_64.sh

conda activate test```

```conda install pytorch torchvision torchaudio pytorch-cuda=11.7 -c pytorch -c nvidia
conda install -c conda-forge pygame-ce
pip install torchrl onnxruntime

Run the code.

I made it easy for you 🙂

Maybe experiment with different cuda versions. It could really impact this

Two more hours till we get to see the results from the 1st epoch, I finally got XLM Roerta large working for the caption creation. Its taking 6 hours just for the eval stage on my 4090.

I went with Roberta mainly for its multi lingual abilities. Now it can read and generate captions in many languages.

The caption creation size has a max length of 77 tokens, which is limited by the clip model.

[violet gull]

what you training?

[rich moth]

image captioning model. It has a vq-vae manifold with dynamic learning and attention mechanism. The idea is to create multilingual captions for images.

It a model that understand the semantic realtionship between words and images.

[tawdry monolith]

I am 15 Yr rn I want to learn ml I have been learning python from few weeks I want to know if I should learn dsa or not I mean if yes than should I cover all topic

[serene scaffold]

if you're a beginner, it's just as well that you practice graph traversal algorithms, since graphs are used to conceptualize things in ML.

Note that by graph, I am not talking about data visualizations. I'm talking about nodes and edges that model entities and their relationships.

[rich moth]

Check out Arrays & Strings , Linked lists , stack quesus , tree and graphs, hash tables and sorting and searching algos. Agents are a hot topic right now as well. But different graph algos like dfs and bfs , etc. you might want to check out too.

[tawdry monolith]

Ok

[serene scaffold]

except if you're doing to do ML, it's important to remember that lists and arrays are different, and to never use those words interchangeably

[serene scaffold]

what do you mean by agent?

[rich moth]

What do you mean?

[serene scaffold]

you said that agents are a hot topic right now. I'm interested to know what examples you have in mind.

[austere perch]

can someone help me im getting this error

[serene scaffold]

Hello, be sure to always post the code and the error message as text. Can you post the whole error message, starting from Traceback, and the code that caused this?

!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.

[austere perch]

basically an ai that has a specific role. the ai impersonates

[austere perch]

ya

[serene scaffold]

I asked you not to post text as screenshots.

[austere perch]

oh my bad

this is the error

File "/Users/cyrusvakil/Visual Studios/Python/Experimental/.venv/lib/python3.12/site-packages/autogen/agentchat/conversable_agent.py", line 159, in init
self._validate_llm_config(llm_config)
File "/Users/cyrusvakil/Visual Studios/Python/Experimental/.venv/lib/python3.12/site-packages/autogen/agentchat/conversable_agent.py", line 263, in _validate_llm_config
self.client = None if self.llm_config is False else OpenAIWrapper(**self.llm_config)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/Users/cyrusvakil/Visual Studios/Python/Experimental/.venv/lib/python3.12/site-packages/autogen/oai/client.py", line 392, in init
self._register_default_client(extra_kwargs, openai_config)
File "/Users/cyrusvakil/Visual Studios/Python/Experimental/.venv/lib/python3.12/site-packages/autogen/oai/client.py", line 453, in _register_default_client
client = OpenAI(**openai_config)
^^^^^^^^^^^^^^^^^^^^^^^
File "/Users/cyrusvakil/Visual Studios/Python/Experimental/.venv/lib/python3.12/site-packages/openai/_client.py", line 104, in init
raise OpenAIError(
openai.OpenAIError: The api_key client option must be set either by passing api_key to the client or by setting the OPENAI_API_KEY environment variable
(.venv) cyrusvakil@Cyruss-MBP Experimental %

the code is kinda long

import os
llm_config={"model": "gpt-3.5-turbo"}
from autogen import agentchat
from autogen.agentchat import ConversableAgent

onboarding_personal_information_agent = ConversableAgent(
name="Onboarding Personal Information Agent",
system_message='''You are a helpful customer onboarding agent,
you are here to help new customers get started with our product.
Your job is to gather customer's name and location.
Do not ask for other information. Return 'TERMINATE'
when you have gathered all the information.''',
llm_config=llm_config,
code_execution_config=False,
human_input_mode="NEVER",
)

onboarding_topic_preference_agent = ConversableAgent(
name="Onboarding Topic preference Agent",
system_message='''You are a helpful customer onboarding agent,
you are here to help new customers get started with our product.
Your job is to gather customer's preferences on news topics.
Do not ask for other information.
Return 'TERMINATE' when you have gathered all the information.''',
llm_config=llm_config,
code_execution_config=False,
human_input_mode="NEVER",
)

[serene scaffold]

Please read this

[serene scaffold]

@austere perch read this

[rich moth]

Did you install openai via pip?

[iron basalt]

You should be able to solve a couple problems on those programming puzzle sites. You don't need to be able to do the most difficult ones. The fundamental problem solving techniques are the important part. For example, if the problem is too hard, try to solve a simplified version first, then use that to come up with a solution to the harder problem. These methods don't even apply to programming specifically. The other important part is just knowing a bunch of the common data structures and algorithms which you can then use to get a quick estimation for how fast or slow something will probably be. One of the main differences between a good ML algorithm and a bad one is just how fast it is on modern hardware / what DSA can it make use of.

[austere perch]

i dont think so

[austere perch]

sorry ill do this now

[austere perch]

wait so i quote the line that has the error

[serene scaffold]

put the whole code in the paste bin

[austere perch]

just did that

[rich moth]

same way you install the other packages install the openai package so you can import your API key

[rich moth]

# Set the API key directly in your code
openai.api_key = 'your-api-key-here'

you might need to setup something else I forget, but check out openai importing api keys python in google

You need to get an API also via openAI website, I found the playground easier to access it.

[austere perch]

i got a key

[austere perch]

it wont let me cuz its too long

[rich moth]

roger, trying doing that for now then

[serene scaffold]

how many lines is it?

[austere perch]

98

[serene scaffold]

Were you trying to paste it into this chat, or into the paste bin?

the pastebin is a website that is separate from Discord.

[austere perch]

oh

how do i access paste bin

[rich moth]

type !paste

[austere perch]

!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.

[serene scaffold]

@austere perch please read this whole message #data-science-and-ml message

it tells you how to get to the paste bin. I posted it a while ago.

but I guess you can skip that for now

[serene scaffold]

this should be the solution.

[austere perch]

wait so i pasted it

do i send u the link

https://paste.pythondiscord.com/LINQ

is that how i dont rlly understand

[serene scaffold]

this would probably also work

import os
os.environ['OPENAI_API_KEY'] = 'your_api_key'

@austere perch yes, but you've posted your API key. please go to the OpenAI website and change it as soon as possible.

[austere perch]

oh

....

[rich moth]

Double ya. change it with a new one, you're comprimised.

[austere perch]

okay i changed it

[rich moth]

delete old one

[austere perch]

i deleted it

[rich moth]

good man

[austere perch]

and m,ade a new one

[austere perch]

is trhis how to show the error as well

[rich moth]

Dude you input the suggestion stelercus made?

Oops, did.

[serene scaffold]

yes

be sure to always post code and error messages as text (not screenshots) because they're easier to read and can be copied/pasted.

[austere perch]

do i paste the error in the same place

as the code or do i just make another one

[serene scaffold]

either in the chat or in the paste bin--whichever is fastest

[austere perch]

ohhh okayt

[serene scaffold]

well, I guess it's not about speed, but that you have to use the paste bin for text that's too long for discord.

[austere perch]

EYYY

it worked thank you sm

[austere perch]

yeah

[rich moth]

Awesome!

[lapis sequoia]

Hey

@austere perch you shouldn’t use OpenAI it costs money

[austere perch]

yeah i realized

[lapis sequoia]

Try pip install shuttleai it’s free

[austere perch]

what do i use

[lapis sequoia]

shuttleai.app

[austere perch]

okay

[lapis sequoia]

Need any help dm me it’s free has OpenAI models and more

And lib is identical to OpenAI but faster more optimized

[austere perch]

sure

[serene scaffold]

@lapis sequoia what is your experience with shuttleai?

[rich moth]

Results are in for epoch. Not sure what to make of it .

[serene scaffold]

looks like shuttleai is another generative AI platform that has a free tier. But it's not going to be identical to what you're getting from OpenAI.

[rich moth]

So far it seems like the best result ive gotten so far, you can make out different hues of color

[rich moth]

The aligment of the vector quantization loss and validation loss is interesting

The size of the torch model exploded though from the last runs, 5 epochs was like 2.5 gigs

Actually it was around 10, but its a lot bigger than its ever been, im pretty sure because it incluedes the newly created captions

I dunno..

[urban helm]

just me messing around with basic neuron networks

i guess a single neuron would have an array of inputs (neurons from last layer inputting an integer after ReLU so just 0-int limit), a bias and a weight as properties?

maybe a function to compute output etc

but thats mostly it..?

also i definitely need an explanation on how backpropogation works

[unkempt apex]

qt.qpa.xcb: could not connect to display :8
qt.qpa.plugin: Could not load the Qt platform plugin "xcb" in "/opt/conda/lib/python3.10/site-packages/cv2/qt/plugins" even though it was found.
This application failed to start because no Qt platform plugin could be initialized. Reinstalling the application may fix this problem.

Available platform plugins are: xcb.```

what's this now

[unkempt apex]

go for 3blue1brown, he had explained like god level!

[small wedge]

In practice, neurons are just the values that we pass through the network, your first layer of neurons is your input layer, you multiply that input layer with your weights and add your bias (and sometimes add an activation function) to calculate your next layer of neurons and so on as it propagates forward through the network.

What is your understanding of backprop right now, do you know what gradient descent is?

[urban helm]

gradient descent is basically finding the minimum of a function.

idk much ab backprop, i only know it is related to the cost, and adjusts neuron weights and biases based on their importance (weights as the "meaning" of each input for the output)

[urban helm]

i am pretty interested at how it actually works in code, like how does it know whcih value to change?

[small wedge]

Right so in gradient descent we calculate the gradient, which tells us which direction to move each variable in order to lower the cost (output of the function)

[urban helm]

i watched like 5 videos from his introduction, i have a hard time remembering and understanding stuff if i can not see their base. his explanation is some sort of higher level and it all happens "magically"

[small wedge]

Have you taken any calculus classes or have any knowledge of that?

[urban helm]

how would an output look like in a gradient descent function

[urban helm]

not really lol

well basic knowledge

nothing too crazy

a gradient descent basically returns the slope id need to go to minimize a value

iirc

[small wedge]

Pretty much, so imagine your model as a function

[urban helm]

the minimum value represents the minimum cost i suppose?

[small wedge]

Say you have 3 layers, 2 sets of weights/biases. You can write it as second_layer(first_layer(input))

(Input is technically your first layer here)

[urban helm]

yeah i figured

[small wedge]

You can see that the function for second layer relies on the output of the first

[urban helm]

i understood the structure

interested ab all the background processes and how they work

[small wedge]

So we calculate the partial derivative of the weight and bias used in the first layer w.r.t cost, which tells us which direction to move

[urban helm]

what does wrt mean again

[small wedge]

With respect to

[urban helm]

and partial derivative is just for one variable iirc

all others being treated as constants ig?

[small wedge]

Yes and no, each variable has its own partial derivative but we can calculate them all at once for a given matrix (wieght or bias) using the chain rule

[urban helm]

how does the rule go?

i forgot lol

[small wedge]

If h(x) = f(g(x)) then h'(x) = f'(g(x)) * g'(x)

[urban helm]

weights are just numbers from 0-1 after all, wouldnt making a derivative of them just make them equal to 0..? sorry ab these questions the math is not mathing in my brain rn

[urban helm]

oh that lol

[small wedge]

So back to my example, we know what we need to adjust the first layers weights and biases but we need to calculate the second layers weights and biases, so we will use a partial derivative from the last layer and the chain rule to calculate the partial derivative of the next weight and bias

[small wedge]

They can be any number not just 0 to 1, and derivatives have to do with functions and their variables, we don't treat anything here as a constant (which would have a 0 derivative)

[small wedge]

This process of calculating all this (your gradient) is called back propagation

[urban helm]

so all weights and biases are variables, and bec we have w*a + B i guess we consider WA as the inner function so when f(x) = b
g(x) = wa we can actually solve it

[urban helm]

so for example say i got an answer, how do i forward that information to modify my current weights and biases

[small wedge]

We would treat it as a multivariate function which is why we are using partials

[urban helm]

using my estimation?

like how would the function look like

so make a derivative for each one of them, so the derivative for the entire function would be WX + 1

?

[small wedge]

We aren't concerned with the derivative of the whole function, all we care about are the partial derivatives of the variables that we need to update, w and b

[urban helm]

oh alright

[small wedge]

Hold on I've written all the math out before as an example let me see if I can find one

[urban helm]

oh wow that could be helpful, lmk

[small wedge]

x = original input
y = labeled output
z1 = x.w1+b1
a1 = activation1(z1)
z2 = a1.w2+b2
a2 = activation2(z2)
c = cost(a2, y)

∂c/∂w2 = ∂z2/∂w2 * ∂a2/∂z2 * ∂c/∂a2

Note: ∂z2/∂w2 = transposed(a1)
Note: ∂a2/∂z2 = derivative_activation(z2)
Note: ∂c/∂a2 = derivative_cost(a2, y)

For the sake of simplicity let d1 = ∂a2/∂z2 * ∂c/∂a2

∂c/∂w1 = ∂z1/∂w1 * ∂a1/∂z1 * ∂z2/∂a1 * d1

Note: ∂z1/∂w1 = transpose(x)
Note: ∂z2/∂a1 = transpose(w2)
Note: ∂a1/∂z1 = derivative_activation(z1)

Breaking all these functions apart into their partial derivatives lets us apply the chain rule to any number of layers, each time we just remove the first partial ∂zi/∂wi and multiply that by the new partials. For example the next d for a 3 layer network would be d2 = ∂a1/∂z1 * ∂z2/∂a1 * d1 and so on

[urban helm]

what does that greek character mean again

[small wedge]

Partial derivative

[urban helm]

i have no knowledge when it comes to notation lmao

oh alr

i suppose labeled is the corre t answer

[small wedge]

In supervised learning our labels are the correct answers, yes

[urban helm]

also i guess z1 is the input neuron and z2 is the current one were handling? since we use z1's activation for z2's value

[small wedge]

X is the input layer

Z1 is the output of that after multiplying by weights and adding bias, then a1 would be our next layer after we apply the activation function

[urban helm]

and z2 is second layer

or output, might as well be

[small wedge]

Almost, we apply activation to that too so a2 would be the output layer

But if we didn't use activation functions there then that'd be right

[urban helm]

oh i see

sorry whats tranposition again? it doesnt have the same name here 😭

[small wedge]

If I have a 2x3 matrix and I transpose it I will have a 3x2 matrix

[urban helm]

oh i see alr

i suppose X and W1 are all matrixes with values according to each neuron where B is just the next neuron's bias?

same for a1 and a2

[small wedge]

Weights and biases exist between layers of neurons, so we wouldn't say that the weight or bias belongs to a layer

[urban helm]

lets take the 28x28 number approach.

x would be 748 elements

so would w1

and b1?

[small wedge]

How many neurons do you want in the next layer?

The dim of the weights will determine this

1x748 * 748xN + 1xN

Where N is the number of neurons in the next layer

[urban helm]

i see

[small wedge]

Multiplying 1x748 * 748xN results in a 1xN layer of neurons ( or at least the values of the neurons before activation is applied)

[urban helm]

so 748 + 748N + N is equal to the size of the matrix?

[small wedge]

No

Not sure which matrix you're asking about

[small wedge]

1xN + 1xN results in 1xN so that's the size of the matrix representing the next layer of neurons

[urban helm]

isnt that 2xN?

[small wedge]

No, it is element wise addition

[urban helm]

ah i see

[small wedge]

https://arxiv.org/abs/1802.01528 this is my holy grail for this topic

[urban helm]

ill check it out very soon, thank you so much for ur assistance

i suppose this is activation with the ReLU function?

[small wedge]

ReLU(x) = max(0, x)

[urban helm]

so thats just activation with relu

[kindred crest]

heyyy

IBM or Google certification, which one should I apply for data science?

[past meteor]

@umbral bison look into adversarial attacks on LLMs. I think you can get away with doing it on smaller models like gpt-2

For some adversarial attacks you need access to the gradients

You kind of do, for llama and co

But unless his uni has a cluster it'll be hard to do with truly large models

[unkempt apex]

cv2.error: OpenCV(4.10.0) /io/opencv/modules/highgui/src/window.cpp:1301: error: (-2:Unspecified error) The function is not implemented. Rebuild the library with Windows, GTK+ 2.x or Cocoa support. If you are on Ubuntu or Debian, install libgtk2.0-dev and pkg-config, then re-run cmake or configure script in function 'cvShowImage'

now what is this thing?

why to install libgtk?

this error only occurs when I am running on docker container!~

[past meteor]

Can you install those dependencies in your Dockerfile

[tidal bough]

I think docker containers don't get to open windows on the host machine at all by default; might need some configuration

[unkempt apex]

yeah I guess

[unkempt apex]

yup I installed libgtk2.0

but dunno about config?

yeah!!

I just said yeah!!😂

installed! now how to config that thing, I dunno about cmake

[tidal bough]

just reinstall cv2, building from source if necessary.

[unkempt apex]

I did that actually and reinstalled without headers
--headless

[tidal bough]

(did you configure the container to allow access to the X server, though? without it, I don't think the container will ever be able to open windows on the host)

[unkempt apex]

@final kiln ?

he configured my container actually

[unkempt apex]

wdym mean by windows?
that pygame window? opencv window

N: Updating from such a repository can't be done securely, and is therefore disabled by default.

[tidal bough]

yeah, I mean like the window that imshow opens. I don't think docker containers can normally open those.

[unkempt apex]

yeah that's the issue

[tidal bough]

oh, vscode's dev container support can allow that? huh, that's pretty cool

[unkempt apex]

and why I am running ubuntu commands on fedora?
is that my docker container running on ubuntu?

this is already installed

we already did this ?

I already have dockerfile

yeah

I created that

FROM python:3.12.3-slim

WORKDIR /app

COPY requirements.txt .

RUN pip install --no-cache-dir -r requirements.txt

COPY . .

CMD ["python"]

  pytorch:
    image: pytorch/pytorch
    command: sleep infinity
    working_dir: /app
    volumes:
      - ./:/app```

change th is?

we build for pytorch

how to ?

and it will get that opencv too?

did that already now running

??

where to put this?

docker compose exec bash!

command?

docker compose exec pytorch?

but why we are already running that docker compose up --build

yeah it's running now?

what to do?

I am in cmd now

yeah working

and when running .py file again same error

[haughty pawn]

hi there datasatanists, how does one write a simplified alternative for PoS diffusers?

[unkempt apex]

I need to cmake that file which I downloaded

[spring field]

I gotta say that UNet++ has got to be the most annoying architecture I have encountered so far
just so many connections everywhere

[spring field]

also, w/o context (but it is a rather small dataset), an IoU score of 0.04, does that like make sense? given that it is increasing

this is how I calculated it

            iou_numerator = torch.sum(y * y_prim, dim=(-1, -2)) + 1e-8
            iou_denominator = torch.sum(y * y_prim + (1 - y) * y_prim + y * (1 - y_prim)) + 1e-8
            iou = torch.mean(iou_numerator / iou_denominator)

y and y_prim have the same shape of (batch, channels, height, width) (or maybe width and height are the other way around, but they're the same value anyway)
and also, the test samples seem to quite closely resemble the expected images
(weeee, more context (I was not expecting to provide this much as one might infer based on the beginning of this message, but in the end, I have provided at least some context... anyway))

bruh

I forgot dim=(-1, -2) for the denominator, cool

I knew it was something silly like that cuz 0.04 just seemed like wayyyy too low

[haughty pawn]

hmm, seems like nobody knows at the moment

[lapis sequoia]

does tokenization come before lemming/stemming?

[spring field]

Your question is a bit broad, because the whole process involves several steps. What is your current level of knowledge on such topics? Why do you want to write a simple alternative? Is it just for practice? Are you trying to learn more about neural nets in general?
Now, for (simple) audio generation specifically you could look into Recurrent Neural Networks. But for a more general approach to generative AI, you can look into Generative Adversial Networks (I'm not entirely sure on how diffusion works, so I can't tell you whether diffusion techniques also employ parts of GANs, but GANs is sort of the more general approach to generating stuff (though, perhaps, it's best suited for images...))

[haughty pawn]

well, i'm not a datasatanist, and i want to enable not datasatanists to write scripts/apps for SD

of course, in an abstract, easy, way

without needing that ai/ml jargon

[unkempt apex]

https://stackoverflow.com/questions/28776053/opencv-gtk2-x-error-unspecified-error-the-function-is-not-implemented

I found the solution

but where I can find now Opencv folder in docker container?

hey @final kiln
I think we need to make another image of open cv now
or else we can make one image which have both pytorch and opencv

https://medium.com/@albertqueralto/installing-opencv-within-docker-containers-for-computer-vision-and-development-a93b46996520

where did you find this?

and yeah another is that
which opencv should install?
headless or normal?

yeah done downloading now what?

Authorization required, but no authorization protocol specified
qt.qpa.xcb: could not connect to display :11
qt.qpa.plugin: Could not load the Qt platform plugin "xcb" in "/opt/conda/lib/python3.10/site-packages/cv2/qt/plugins" even though it was found.
This application failed to start because no Qt platform plugin could be initialized. Reinstalling the application may fix this problem.

Available platform plugins are: xcb.```

new error that's nice

Reinstalling the application may fix this problem.

what does this mean now?
reinstalling what

ahh !!

what question?

wait

I am not able to understand this now!

custom environemnt for pygame game, where my agent will be trained

that's why I need a window

to create a env

then it's not good!!

I need window badly

nah, we need to search

for what?

we are creating with opencv then

should I switch to pygame window now?

there is another error for pygame also lemme share that too😂

Authorization required, but no authorization protocol specified
error: XDG_RUNTIME_DIR not set in the environment.
ALSA lib confmisc.c:855:(parse_card) cannot find card '0'
ALSA lib conf.c:5178:(_snd_config_evaluate) function snd_func_card_inum returned error: No such file or directory
ALSA lib confmisc.c:422:(snd_func_concat) error evaluating strings
ALSA lib conf.c:5178:(_snd_config_evaluate) function snd_func_concat returned error: No such file or directory
ALSA lib confmisc.c:1334:(snd_func_refer) error evaluating name
ALSA lib conf.c:5178:(_snd_config_evaluate) function snd_func_refer returned error: No such file or directory
ALSA lib conf.c:5701:(snd_config_expand) Evaluate error: No such file or directory
ALSA lib pcm.c:2664:(snd_pcm_open_noupdate) Unknown PCM default```

yeah!

not possible for now!

yeah

vega 8 which takes 2gb ram

[spring field]

if anyone is suddenly interested in my implementation of UNet++, here's the code: https://paste.pythondiscord.com/R27Q
(based on this paper: https://arxiv.org/pdf/1807.10165)
some plot from training, lol

[unkempt apex]

https://stackoverflow.com/questions/60042568/this-application-failed-to-start-because-no-qt-platform-plugin-could-be-initiali

what about this?

so we will download qt externally

I got new error wait

improvement is happening

so I have downloaded that file
libgtk something

now the question is how can I re-run that?

but how can I re-run?

that cmake

[scenic chasm]

Do you guys think the YOLO object detection model will run smoothly enough on a rpi3 for a self centering camera?

[unkempt apex]

I need to install open-cv along with all dependencies and then again build docker container

[urban helm]

random question, opencv lets you do image analysis?

like with yo camera and shit

[unkempt apex]

you have to do that!!

[urban helm]

so just build a casual CNN

to analyze

[unkempt apex]

yeah

[urban helm]

what tool do you use for building neural networds

networks

[unkempt apex]

tool?
pytorch

libraries

[urban helm]

currently im trying to build my own neural network, is that a possibility to do from scratch?

[unkempt apex]

yeah pytorch allows that!

or from scratch , god!! then build it from numpy!

[wooden sail]

what i would note is that if you are already comfortable with numpy, you might be interested in trying jax out

[urban helm]

whats the contribution of numpy towards building a neural network? i suppose just the big ass arrays?

[wooden sail]

letting you do the math at a reasonable level of abstract

[urban helm]

im not looking for an extremely useful network, going with the basic 28x28 hand drawn numbers and ill just plug a front end where you can draw shit on there

[urban helm]

yeah ig

[wooden sail]

cuz otherwise you have to start by writing your own matrix operations like addition and multiplication

[urban helm]

oh hell nawh

[wooden sail]

right. so from numpy up

[urban helm]

yup i guess so

[wooden sail]

if you will do the derivatives yourself by hand, numpy will be enough

[urban helm]

by scratch i mean building the classes for neurons and the network itself

[wooden sail]

if you want the derivatives to be computed automatically, you have to use pytorch, tf, jax, or something of the sort

[urban helm]

i dont think doing derivatives would be neither interesting nor fun

[wooden sail]

then the starting point is pytorch, jax, or tf. pytorch is probably the most recommended. i like jax because it writes very much like numpy. tf... is kinda hard to recommend at the moment

[urban helm]

doesnt pytorch make u create a neural network using it though? or can i still utilize some other functions

[wooden sail]

wdym?

[urban helm]

idk im used to cpp

i forgot how to do python shit other than some basic networking

[wooden sail]

i still don't get what you're asking

[urban helm]

atp idk either im tired asf

what im asking

is if all functions are accessible

honestly i shd check out the package first

[river cape]

Guys why do we need keras? Why couldnt we just stick to tensorflow

[deep sleet]

When I use SGDRegressor it gives a rmse of over 53753742.24651016

[buoyant vine]

It is mostly good for beginners getting started with things

I could say the same with TF tbh

Outside of google, I haven't work on any AI related projects that use TF over PyTorch for anything that matters

[deep sleet]

While when I switch to linearregression gives an acceptable 10711.00334810241

[river cape]

I have heard that Keras is like an API for TF

[buoyant vine]

yes

[river cape]

Thats so much

[buoyant vine]

it is a more beginner friendly API wrapper over TF

that makes some operations simpler and lowers the barrier to entry a bit for new people

[river cape]

API in sense , it collects data from the TF framework and presents it to the user?

[buoyant vine]

no]

[deep sleet]

Yes

idk why

[buoyant vine]

API in the sense it wraps the tensorflow api

API != webserver or what not here

Keras just basically creates some functions and classes for people to use and the internally makes the calls to tensorflow and configurations

In the same way requests is an API wrapper around urllib3 for example, you could call urllib3 directly, but requests exists to make your life easier

[river cape]

So the Dense() function is like a part of the keras?

[river cape]

Oh makes sense

[coral field]

Is there any case where multicollinearity in regression models can somewhat help it's accuracy?

[river cape]

I believe no

[wooden sail]

it can only make it worse

[deep sleet]

what is scaling your input refers to?

[coral field]

Alr

[coral field]

Changing the range of a features values to be closer to every other features values

[tidal bough]

Applying a linear transformation to change the range of values (because some kinds of models are sensitive to that, and won't perform well if one of your features is from 0 to 1 while another is from 0 to a billion). See https://scikit-learn.org/stable/modules/preprocessing.html for an overview.

[deep sleet]

oh ok

will check it out

ohhh

apparently that is the problem

I have to scale the inputs for gradient descent

Do I need to know the math behind it?

[uncut plaza]

hey everyone

does anyone know how to draw this?

[rich moth]

Looks drawn to me.

[uncut plaza]

yeah unfortunatly

[river cape]

Is this overfitting in terms of train and validation data?

[rich moth]

What color represents what?

[sharp cave]

Who have an idea about. ML ?

[river cape]

The blue line is the training loss and orange line is the validation loss

[violet gull]

in RL how does an agent train on dynamic environments? Currently i train it on a static envirnment but if i wanted to randomize the environment in the testing stage wouldnt i need to train it on thousands of random environments? that would multiply the training time by thousnads

[small wedge]

Why do you think you'd need thousands of environments? What do you imagine happening if you had only 2?

[violet gull]

an image classifier cant learn off 2 images, i assumed this was similar

[odd meteor]

You can recreate something like this with tools like draw.io and lucid.app

[small wedge]

Well same idea, I think what you're getting at is having a model that has a proper generalized function for any task you give it in the environment? You would encounter overfitting with 2 images in a classifier and the same would probably happen here. But your model is already overfit on one environment so as you increase the samples, the number of specific functions that it can overfit on decreases and it is forced to generalize more and more.

[violet gull]

so i do need a bunch of environments just less than thousands

[small wedge]

And that's really only the ideal, since RL environments can be incredibly finicky depending on how you give reward/punishment. If the model finds one environment where it can get 10x score it will steer it towards overfitting even if you have lots of other environments

[small wedge]

It's more about having enough environments that the model can't overfit. You could generate random environments even as long as the reward/punishment is bounded the same for all of them

This could be done with 4, 5, 6 environments if your model is simple enough that it can't overfit all those

[violet gull]

so 6 predefined environemts or a random environment each time for training?

[small wedge]

Just depends on your model architecture, policy, and environments

[violet gull]

also is it expected that the training global average is less than the max score due to it having a minimum random action chance?

[small wedge]

Global average as in the average of all agents over the sim?

[violet gull]

like right now i can get the average score to the max score but if i set a minimum random it will sometimes kill itself causing not every iteration to be perfect

[violet gull]

yes

[small wedge]

Yeah ofc, if even one agent gets a score below the max then it will be skewed, you wouldn't expect that number to match the max score

[deep sleet]

in my dataframe I have the column smokers with (yes and no) and I want to replace that to 1 and 0 to be able to deal with it how can I do that?

[lusty jacinth]

I have a dataset that looks like this, this is a “bad” dataset because of the periodicity occurring in the trough. I want to find a way to detect and quantify the periodicity that’s occurring in this dataset

I’ve looked into Fourier transforms and dug around scipy but haven’t had much success, Fourier transforms appear to be for filtering out digital signals and scipy I just haven’t quite found what I’m looking for yet

Any recommendations?

[unkempt apex]

now I will try for last time, and if that doesn't happened , I will increse that storage of /tmpfs

[violet gull]

hi

[unkempt apex]

hello

[violet gull]

post your error real quick i think i can fix it

[unkempt apex]

now I am cloning whole opencv through github

[unkempt apex]

what? of my docker?

[violet gull]

werent you having an issue of pip saying it doenst have enough storage

[lapis sequoia]

@unkempt apex can I help you w something

[unkempt apex]

yeah that error

[violet gull]

are you running in visual studio code in linux?

[unkempt apex]

no I didn't get I was totally bored with that

[unkempt apex]

yup fedora 39

[violet gull]

uninstall visual studio code and install it from dkp thingy. The flatpak version is bugged

[lapis sequoia]

import random

messages = [
"Get lost, you useless moron.",
"Your existence is a joke.",
"Why do I have to deal with idiots like you?",
"Go play in traffic, you imbecile.",
"You bring new meaning to the word stupidity."
]

def generate_zapbott_message():
return random.choice(messages)

print(generate_zapbot_message())

[unkempt apex]

yeah but pip was not able install from venv

[unkempt apex]

no one force you to help me!!

yeah make sense

[violet gull]

#ot0-psvm’s-eternal-disapproval message
this fixes it

[unkempt apex]

okay okay can I share you exact error msg of that pip?

yeah I am also thinking this

linux legends are online , he has something for me

my pc is getting hot now!!

my all cores are running at 100 percent

because of that cloning process

wait I am muted because of that bot
I was directly pasting code

yeah freak!

https://discord.com/servers/linux-chat-292942011261124608

[unkempt apex]

first of all why typed this?
are you going with bad day or what?

yeah they are very nice!

I need to install another fan now!!

also how can I delete all this docker containers and images now?

[lapis sequoia]

import random

def get_random_message():
"""
Returns a random message from a predefined list.
"""
messages = [
"Get lost, you moron.",
"Why are you wasting my time?",
"You really are an idiot, aren't you?",
"Stop bothering me.",
"You're absolutely clueless.",
"You couldn't do this if your life depended on it.",
"Give up already.",
"This is pointless and so are you."
]
return random.choice(messages)

def main():
while True:
user_input = input("Do you want a random message? (yes/no): ").strip().lower()
if user_input == 'yes':
print(get_random_message())
elif user_input == 'no':
print("Good. Now go away.")
break
else:
print("I didn't understand that. Try again, you moron.")

if __name__ == "__main__":
main()

[unkempt apex]

okay lemme first see that cloning

@lapis sequoia is having bad day I think!😂

[cobalt zinc]

I agree.

[unkempt apex]

hey that gpt docker file works

literralllyy

hwo can I run different image? on docker tell fast

[lapis sequoia]

def assess_day():
mood = "absolutely terrible"
message = f"He's definitely having a {mood} day. Just like every day he has to deal with oxygen-stealing idiots like you."
return message

def extend_insult():
additional = "You were a waste of resources from the moment you were born, and I sincerely pity anyone who has to tolerate your existence."
return additional

def main():
day_message = assess_day()
insult_message = extend_insult()
final_message = day_message + " " + insult_message
print(final_message)

if __name__ == "__main__":
main() ```

[unkempt apex]

I ran this with sudo docker compose up

but it is runnig that pytorch

not pong-game ( current )

I think I should destroy all the containers now , I am getting bored with those

@final kiln the solution was very simple

and I spend 2 days onto this , but anywas docker is the thing which I learned

but hey now how can I write something in docker which I will upload ongithub and others can also use that?

yeah

now need to clean storage

because I have run that build docker multiple times

so only 85 gb remaining

although I have removed all the images

and containers

Total reclaimed space: 21.69GB

wait what is leaking memory into disk?

that dataset is taking 23 gb

yeah the question was what to write on docker file now?

now we are not using docker file for that pytorch

but I still want to make use of dockerfile ( it looks nice on github though)

so what to write on that?

basic stuff for installing packages with docker?

why the hell you types all this😂

but the thing is I have already push code in which docker was there!!

it will disturb my code flow🙃

anyways need to again push without docker then

shit bruhh!!😂

    self.icon = cv2.imread("paddle.png") / 255.0
                ~~~~~~~~~~~~~~~~~~~~~~~~~^~~~~~~
TypeError: unsupported operand type(s) for /: 'NoneType' and 'float'

this was the same error but was on docker

so I solved that with absolute path of .png

but now we don't have docker

so what which absolute path?

yeah it's working with that

[rich moth]

This is taking forever now. Im going to restart the training after this epoch, I just wanna see the resuls. But it was strange when I woke up, it was going, but really slow. It might be the dynamic learning and optimizations causing the slow down. I also though it might be combined wit the roberta models max length of 256 so I matched that with the clip model of 77. I moved the terminal screen and it started to get faster, dunno just seem abnormality long. Also lowered my batch size from 16 to 5 would be more efficent, just from watching it seems process the first 5 pretty quick. Evaluation: 61%|█████████████████████████████████████████▍ | 355/582 [6:30:55<2:52:02, 45.47s/it] Oh I setup some metrics to evaulate the captions too.

Like when it slows down, it seems stuck but then it gets back around to that screen above and it will show like 140s/it

I also think all the context its creating for the captions is blowing up the size of the model. One run was like 2.2 gigs

Wait thas not right, I think it was more than one.

Evaluation: 63%|█████████████████████████████████████████▋ | 368/582 [6:51:21<11:43:55, 197.36s/it]
Thats what I mean.

[neat crystal]

can someone help me im making a neural net but idk how to visualize it? like should the wheights be the size of the line, or should it be the color? and is the bias line width or color or should he be the nodes? idk can someone help me please

[unkempt apex]

watch 3b1b video on it

[unkempt apex]

or you want from scratch

then go with sentdex!

sentdex has nailed that!

[neat crystal]

no i just want to know wheither the line width should be whieght or bias

[unkempt apex]

what is line width?

that nn lines? connected lines?

[neat crystal]

ya

[unkempt apex]

they are weights

[neat crystal]

the connected lines between the nodes

[unkempt apex]

and each node has bias which gets add into weighted sum

[violet gull]

@small wedge if the model has found a better option why isnt it doing it in testing. In testing I just run the model in pygame to visualize it with the randomness and learning removed.

Score:  73
Score:  73
Score:  73
Score:  73
Score:  73
Score:  73``` this is what its doing in testing 
```All Time Average Score: 50.563
Average of last 100 bears: 53.33
Highest Score:  78
Total Bears:  7000
Time between epoch: 44.77771258354187 s``` but it has found a better option of 76 in training

[haughty pawn]

so at the end of the day here's my PoC, likely to be broken as of right now https://github.com/aolko/StableDiffusionEvents/tree/main

not the highest quality out there, but hey, at least i'm trying to work with 💩 d i f f u s e r s

(and the yaml is the intermediate format for the future webui)

[rich moth]

Holy mother of mary Evaluation: 90%|██████████████████████████████████████████████████████████████▋ | 521/582 [8:46:03<37:49, 37.20s/it] just 37 more mins 😂

Poor gpu is tired.

[violet gull]

9210 54
9211 73
9212 54
9213 39
9214 70
9215 41
9216 41
9217 39
9218 71
9219 57
9220 37
9221 58
9222 44
9223 72``` i feel liks this shouldnt be jumping around so much

iteration vs score

[urban helm]

are you using ReLU or sigmoid as your activation function thing

sorry i do not know much i just know relu has a better learning curve

also its probably your rate of learning being inaccurate

[violet gull]

yeah that dont look normal

[austere perch]

what is this

[rich moth]

im training a model that combines image reconstruction and multilingual caption generation. its suppose to create a deep understanding, semantic relationship between them, thats the idea anyways. So i built this VQVAE with manifold learning and combined multiheaded attention mechanisms, dynamic weight adjustments and integrated caption generation for the images .

[austere perch]

out of the few words I understood, that sounds rlly cool

hopefully ill be able to do stuff like that eventually

what was the command to put ur code in the broser thing

i have an update issue

[rich moth]

the !pip install one?

[austere perch]

the one where u paste ur code in the browser

[rich moth]

oh !paste

[austere perch]

ohh

!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.

[austere perch]

any clue how to fix this?

[rich moth]

You might wanna change your API key again you exxposed it 😛

[violet gull]

snatches

[austere perch]

Oh shit..

waitugh

i cant rlly change that one

just dont take it i trust yall

[rich moth]

Well you never know who lurks in the shadows

trying running the openai migrate command

[austere perch]

https://paste.pythondiscord.com/PIXQ

[austere perch]

how

[rich moth]

try typing it in the terminal

[austere perch]

is this the command
pip install --upgrade openai

[rich moth]

honestly, not sure, but from the warning yesterday it said to use % instead of !

[austere perch]

i think i did it its

doing something

i typed "openai migrate"

[rich moth]

look like its updating your code base

[austere perch]

it seems it had quite a few errors ima just look into it

[rich moth]

paste them lets see.

sometimes pip uninstalling packages and reinstalling and work things out

remember what version works before you go dabbling with that though.

[austere perch]

i fixed it by pip install openai --upgrade

but now im getting this error

[serene scaffold]

Please remember what we discussed about screenshots

[rich moth]

try changing engine to model.

[austere perch]

oh sorry, screenshots are only for errors right

[serene scaffold]

Only for information that isn't text (so not code and not error messages)

[austere perch]

ohh

!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.

[serene scaffold]

Data visualizations obviously have to be screenshots

[spring field]

unless you convert them to ascii art

[austere perch]

!paste

[serene scaffold]

You don't need to do the command repeatedly

You can just click the link again

[austere perch]

https://paste.pythondiscord.com/3LGQ

[spring field]

in fact, you can bookmark that page

[austere perch]

okay

[austere perch]

this includes hte error @rich moth

i replaced engine with model

[rich moth]

it means they dont use that model anymore, pick an updated one. I believe.

chatgpt3.5 turbo or something

[austere perch]

ohh

[violet gull]

Reinforcement learning is so hard 😭

[rich moth]

Im really interested in learning what your working on a bit RL is something im also very interested in but havent dabbled it in really.

I mean I saw the code and looked over it, the reward system to interesting.

[violet gull]

I just never know what the issue is

It’s just randomly changing variable values until it works

[rich moth]

Dude something I feel like thats park of this ML , AI stuff. Sometimes when you get a solid foundation of something, its more expermenting like an alchemist.

[violet gull]

And every run takes 7 hours

[rich moth]

ya that too

[violet gull]

I’m back to mine not converging

No idea why

[rich moth]

Hey what was the issue with your enviorment last night? You fixed it I imagine?

[violet gull]

I fixed it by giving up

[austere perch]

REAL

[rich moth]

lol

[violet gull]

I went back to windows and if I ever get to the point where I need a gpu I’ll learn aws

My model was working perfectly

Then I upped the complexity just a tiny lil bit and now it doesn’t work

And I’ve over compensated on the model size so it must just be a fundamental flaw

[spring field]

it do be like that

[rich moth]

I can help you experiment with it later if you want. Im waiting for this damn training to finish though but im curious cause I'd like to learn as well.

[spring field]

not sure I've linked this, but here's a good book on RL http://incompleteideas.net/book/RLbook2020.pdf

[rich moth]

I wish I had only 7 hours lol. Evaluation: 98%|███████████████████████████████████████████████████████████████▊ | 571/582 [11:11:58<4:04:14, 1332.21s/it]

[violet gull]

how does yours have a set finish time?

i thought in ML you train it until its good enough not after a certain time period

[austere perch]

what does it mean to add a key to an .env?

"While you can provide an api_key keyword argument, we recommend using python-dotenv to add OPENAI_API_KEY="My API Key" to your .env file so that your API Key is not stored in source control."

[rich moth]

Thats my understanding also, this is only the 2nd epoch though

[violet gull]

but how does it have a time limit

[spring field]

that's for a single epoch

[violet gull]

i know

but my epochs have variable times

[spring field]

well, you're doing RL

[violet gull]

ive done other ML

they all had variable epoch time

[spring field]

that is rather questionable

[violet gull]

why does his not have variable epoch time

[spring field]

why would it? you're always processing the same number of inputs

[violet gull]

oh isee

its just based off how much of the input data has been processed

[spring field]

yeah

what is variable is the number of epochs you might need to train for

[violet gull]

dang 11 hours for one epoch that thing is gonna take all year

[rich moth]

I made some changes hopefully reducing the batch size and caption creation size. It was 256 tokens now its 100.

[violet gull]

u considered upping your hardware?

[spring field]

even in the case of RL though, you're pretty much always processing the same batch size, what could be variable is the length of a single episode, the agent could take variable number of steps every episode before the episode terminating

[rich moth]

Im on a 11700k, 128gigs, 4090. I need something more commerical lol

[violet gull]

aws computer is like a penny an hour

for a high end ML gpu

[rich moth]

ill check it out thanks, hopefully some of these changes i can avoid that though

[austere perch]

goddamn

[rich moth]

I had to add more metrics for the caption generation though

[violet gull]

can anyone interpret this?

[spring field]

nope, I have literally no idea what I'm looking at

[violet gull]

iteration vs episode score

the grouping is kinda weird

[spring field]

can you make it a line plot

[violet gull]

ill need to retrain but yeah

[spring field]

and high fluctuations can be expected in RL, did you try using fixed Q targets

[spring field]

good one

[violet gull]

see how it only hit the high values like once

shouldnt it be higher

[austere perch]

the heck is that

[violet gull]

a line plot

[austere perch]

😭😭😭

[rich moth]

It tells me theres 4, 5 types of players. Some being better than others lol

[austere perch]

thats a crazy line plot

[rich moth]

Ooops my text was meant for you Sarati

Theres are the results from the 2nd epoch

[spring field]

why are there 160k whatever x is? (please label your plots)

[violet gull]

episodes

[spring field]

that is a lot of episodes, have you considered making the whole plot window wider?

[violet gull]

doesnt help much

[spring field]

IIRC

matplotlib.rcParams["figure_size"] = (width, height)  # in some weird unit, in your case you can try something like (100, 10) maybe, lmao

[violet gull]

the graph isnt going to show anything different

[spring field]

I mean, what I'm seeing is great fluctuations which as I said, can be expected in RL
now, did you try using fixed Q targets?

[violet gull]

how is it expected

shouldnt it consistently get high scores

havent used Q targets

[spring field]

well, you have, you haven't used fixed Q targets

[violet gull]

is this statement wrong

[spring field]

ehhhh
ehhhhhhh
honestly, no clue, I mean, that is the expectation... but there are a ton of factors

[violet gull]

somehow the test value is always constant though

[rich moth]

I feel like you need a bigger range or less bears to measure it accurately.

the results are so condensed how do you measure the results?

[violet gull]

wym measure

[rich moth]

guess im confused.

[violet gull]

im logging some metrics as well as testing the model in pygame to visualize if the model works or not

[rich moth]

No I guess what I mean, is like exxtending the range and maybe lessen the time? I just look at the chart and its hard to tell whats going on in the middle.

[violet gull]

this one shows it better

[rich moth]

I mean obviously their in there!~ but how much?

[spring field]

if it's of any consolation, some research papers have trained RL models for hundreds of millions of episodes

[violet gull]

yeah but those are big complicated enviornments

and mine has stopped learning

[spring field]

not necessarily big but reasonably complicated, perhaps, yeah

[violet gull]

the scores plummeted

[rich moth]

Oh I see and understand now 😛

[spring field]

are you plotting loss (TD error but inverse pretty much) and number of steps per episode as well?

[violet gull]

number of steps directly affects score

[spring field]

I think, in RL having a complex model might not just mean it'll take longer to converge, it might mean it'll never converge because the agent simply does't get to explore the things that would move the network closer to convergence

[violet gull]

mine has reached the max score though

[spring field]

at least that's my intuition

[violet gull]

but only once, why cant it do it more than once

[rich moth]

Couldn't you intice it via the reward system ?

[spring field]

that's what RL already does

yeah, you could adjust rewards, ofc, that's another hyperparameter you can tune

[violet gull]

right in here it hit the maximum score but never again

[rich moth]

Can we see your working code again?

[violet gull]

!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.

[violet gull]

https://paste.pythondiscord.com/OJRQ

[spring field]

so, they way it's learning, right, it's randomly sampling past actions from memory, that will, inevitably include a bunch of not so exciting ones, where there isn't much happening, so overall it'll not converge to the maximum
this is why something like Priority Experience Replay is another optimization which simply makes it sample past experiences with greater priority assigned to those that have a greater deviation from that expected score
also why Fixed Q Targets are used, you have an anchor point that you explore around and then move the anchor to the best found new location instead of carrying the anchor constantly around to the first best position you find pretty much, this helps to reduce the fluctuations (I think it's also called variance in statistical terms, not sure)
then also stuff like deterministic and non-deterministic policies plays a role, if you have a deterministic policy, you can pretty much stop once you've reached the max score and that's it, it'll always take the same actions, but it will reach the max score every time (unless there is some randomness in the environment itself)
those are just my sort of intuitions on this topic and from stuff I've read

[violet gull]

hmmmm i will try to program those optimizations, but do you know why its failing to converge currently?

Fixed Q Targets, Priority Experience Replay

[rich moth]

Have you tried getting any advice from AI about it or are you apposed to such ideas?

Sometimes they point out the obvious, we often overlook. Or get different insights and angles.

[violet gull]

chat gpt speaks nonsense for anything as complicated as this

might as well be asking the dude on the street corner

[rich moth]

lol

[rich moth]

Well check out these suggestions ```def get_score(action, bear):
reward = 1 # Default reward for taking an action
match action:
case 0 | 1 | 4 | 5: # Left, right, up, down
bear.hp -= 2 # Reduced penalty for movement
# Calculate proximity reward based on distance to the nearest berry
min_distance = min(bear.distance(berry) for berry in berries if berry.color == green)
proximity_reward = 10 / (1 + min_distance) # Reward decreases with distance
reward += proximity_reward
case 2: # Wait
reward = -1 # Small penalty for waiting
case 3: # Eat
for berry in berries:
if bear.borders(berry) and berry.color == green:
berry.color = red
reward = 20 * (100 - bear.hp) / 100 # Reward scales with hunger
if bear.hp < 100:
bear.hp = min(bear.hp + 50, 100)
return reward # Exit early if a berry was eaten
elif bear.distance(berry) < 3: # Intermediate reward based on proximity
reward += 5
case 6: # Explore (new action)
reward = 2 # Small reward for exploring new areas
case 7: # Kill self (adjusted action)
bear.hp -= 20 # Reduced penalty for "kill self" action
reward = -10
return reward # Return the calculated reward

Update the BearBrain class to include the new actions

class BearBrain(nn.Module):
def init(self, input_shape, actions):
super().init()
self.network = nn.Sequential(
nn.Linear(3, 128),
nn.ReLU(),
nn.Linear(128, len(actions))
)
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
self.device = "cpu"
print("device:", self.device)
self.to(self.device)

def forward(self, x):
    return self.network(x)

Update the Bear class to include the new actions

class Bear(Square):
hp = 100
actions = [0, 1, 2, 3, 4, 5, 6, 7] # Updated actions
states = []
learning_rate = 0.0025
discount_applied_to_future = 0.9
random_behavior_chance = 1
input_shape = (1, 3)
batch_size = 128

# Rest of the Bear class code remains the same

Update the main loop to pass the bear object to the get_score function

while True:
# ...

if bear.hp > 0:
    total_action_count += 1
    action = bear.choose_action(state)
    score_per_action = get_score(action, bear)  # Pass the bear object
    total_score += score_per_action
    next_state = get_state(bear, berries)
    bear.remember(state, action, score_per_action, next_state, bear.hp <= 0)
    bear.learn()
    state = next_state

    # ...```

[violet gull]

what changed

i see it decreased the hp loss from movement which i dont like because it just inflates the score

[spring field]

not really no, as I have mentioned a couple times, tons of factors and RL is not easy, lol

[violet gull]

blindly tuning parameters is not for me 😭

[rich moth]

I felt it was too long to paste in here.

[violet gull]

how did u even paste that one its massive

[rich moth]

What do you mean?

[violet gull]

usually python bot stops messages that big

[rich moth]

Ah, maybe I was just a token few short. lol

[violet gull]

well ima read that book matiss sent earlier

[rich moth]

probably a good idea.

[violet gull]

see if allows me to figure out why this doesnt work

@rich moth GL on your model, lmk how it turns out

[lapis sequoia]

How does one keep the Val accuracy consistent with the training accuracy for a Tokenizer RNN LSTM thing. I don’t know, this is hard. Should have started with cnn. Is a validation split better ? Or is using the test data after NLP stuff better or does the defeat the purpose of a testing set by using it to evaluate the training data?

[spring field]

wdym consistent with the training accuracy? validation accuracy is seldom as high as that of training

and wdym started with a CNN, that's a completely different network altogether

[rich moth]

Im a bit confused. can you try to rephrase it again.

[rich moth]

Man lowering the batch size from 16 to 8 was about a 20% increase in performance.

[rich moth]

I'm having a hell of a time getting the Roberta model to generate captions. The left side padding error is driving me bonkers.

I dunno, I feel like its processing the captions I just cant correctly get it to show

!paste

[arctic wedge]

Pasting large amounts of code

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

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

[rich moth]

here is the section related to it if anyone has any suggestions . https://paste.pythondiscord.com/352A

going to bed

[spring field]

is that code AI-generated? 🤨

[buoyant vine]

It is also basically impossible to gain any useful info out of that code because we know none of the types

Why is the Roberta model supposed to be generating text? It is not particularly designed to do that...

Also idk what is going on with your tokenizer and attention mask but if this is a huggingface transformers type then that is not how you create the input IDs and attention mask

[unkempt apex]

any RL people?

now the issues are solved mostly but some bugs in env!!

so if you have done using gym environment it will be easy for you to identify!

[past meteor]

(small note, if you use dropout it' s common that for certain settings your val loss is lower than your train loss)

[river cape]

Guys I have a question

In perceptron why do we need a peceptron update rule when we can just analyze the perceptron loss and adjust weights accordingly?

[warm trellis]

That’s really nice question

[river cape]

Any answers?

[lapis sequoia]

I guess a simpler way to say what I was asking is whether or not the training set could ever be ran through the nn with the test set acting as the val set, but that would defat the purpose of a testing set.

[spring field]

perceptron as in step activation function?

[tranquil swift]

how do i start learning ML guys i have no idea where to start from any recommendations like videos and sites etc?

[spring field]

see pinned messages in this channel

[tranquil swift]

ok ty

[round tusk]

Can someone suggest some courses for NLP?

[serene scaffold]

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

[round tusk]

Thank you so much

[unkempt apex]

https://www.pythonmorsels.com/p/29wgb/

is there any logical error here? please check someone

because the current environment code works fine but it lags/freeze in train.py

[river cape]

Yes

[wooden prism]

I made a program so that I can download all of the transcripts from a YouTube channel. What should I do with that?

[serene scaffold]

Did you use the YouTube API?

[wooden prism]

yep

[river cape]

Guys

In huber loss , if the point isnt an outlier we use the formula similiar to that of the MSE right?

But in my sources , it shows 1/2 of the MSE

Why is that

[wooden sail]

scaling of the loss doesn't have any effect

the MSE is often written with a 1/2 in front simply because differentiating gets rid of the 1/2 factor

same in the huber loss

nicer to write, but otherwise it has no effect

[river cape]

But if we divide the loss by half , isnt it a problem?

[wooden sail]

no, it literally makes no difference

[river cape]

how?

[wooden sail]

the value of the loss is never important. what matters is the value that achieves the minimum (i.e. the minimizer)

for the one dimensional case, you can easily convince yourself that x^2/2 and x^2 have the same minimizer: x = 0

the same is true of adding a scalar btw

[river cape]

Oh so we arent considered about the number , we just want that term to be as minimum as possible?

[wooden sail]

from the convex optimization perspective, you only need to find a point where the gradient equals 0. well, scalar multiplication factors out of differentiation. that is, if we have f(x) and its derivative f'(x), it turns out that the derivative of 2f(x) is 2f'(x). if you equate f'(x) = 0 and 2f'(x) = 0, you can just divide the latter by 2 and you get exactly the same thing again: f'(x) = 0

unless you have an explicit interpretation for the loss, the number usually doesn't matter at all

you more care about the minimizer and about the inference results you obtain, not the value the loss takes

[river cape]

Oh makes sense . Thanks just one more thing , why do we make the small errors to be penalized quadratically and not linearly?

[wooden sail]

that's a difficult question to answer. the choice of cost function always depends on the properties of your problem

sometimes quadratic terms make sense, other times they don't. there's no general rule

the common motivation behind MSE is that, under AWGN, least squares minimization is the maximum likelihood estimator

under different statistical criteria, the absolute value is preferred

you could take the huber loss as a robust MSE minimizer that, for small errors, behaves like the usual least squares, and for large errors becomes insensitive

[river cape]

Ohhh so it depends on the problem statement but overall we would want to use huber loss

[wooden sail]

you wouldn't always want to use it

only for problems where it makes sense

[river cape]

So after the loss is calculated, we use gradient descent and then adjust the weights right

[wooden sail]

yeah

technically you never need to evaluate the loss, but this is often done as a sanity check to verify that it's decreasing

[river cape]

Oh it clears my whole doubt . Thanks mate!!

[long robin]

People say
Tensorflow is easy to deploy
Pytorch is not
Pytorch's syntax is better
Tensorflow has prebuilt libraries in it
Pytorch is used more in research field

What should i do?
I planned to start with tensorflow, but now I'm confused

What if i go with pytorch first, and then if needed i can shift to tensorflow

[rich moth]

Thanks for pointing out how I was using the input id attention mask, I got it fixed regarding Roberta I was experimenting with different models. I switched to a ber base uncased for masked LM. Its working now. . Feature shape: torch.Size([512]), Input IDs: tensor([[ 101, 101, 1996, 3746, 3065, 103, 102, 102]], device='cuda:0') Generated Caption: the image shows
its not much, but its start.

[small wedge]

the image does indeed shows

[rich moth]

well it wasnt producing anything at all

its a start

[brave sand]

has anyone made a custom environment with pettingzoo?

[unkempt apex]

no but making for Pong

but mine is not for multi-agent

[brave sand]

how hard is it?

is the hardest part the gui?

and what have you done so far?

[unkempt apex]

yeah created environment, now training part

but yeah for begineers like it's quite hard because new naming convetions are being introduced to me

[unkempt apex]

no I mean you can do that with opencv, but how you are representing that is you play

btw which game you are making?

[unkempt apex]

You should see others code first wait a sec I will send that link

https://blog.paperspace.com/creating-custom-environments-openai-gym/

read this

[brave sand]

did you have to encode state space and action and stuff into the env?

or did you make a gui that represented pong

[unkempt apex]

wdym by encode?

[brave sand]

how does the training send data to ur custom env?

[unkempt apex]

yeah so in simple terms I created whole Pong like game in env

[unkempt apex]

it's long part first read that blog
then you have to read about Q-Learning

[brave sand]

i have the training done

[unkempt apex]

??

[brave sand]

but i only have the policy being in the terminal

no output visually

[unkempt apex]

which game?

[brave sand]

DMs?

[unkempt apex]

yeah!

are you shy about sharing game name!😂

[rich moth]

ratio: 0.1665920527192292
Epoch [1/5], Train Loss: 0.9839, Val Loss: 0.5658, Train PSNR: 14.2182, Val PSNR: 17.2875, Train SSIM: 0.1736, Val SSIM: 0.2983, BLEU Score: 0.0010, CIDEr Score: 0.0008, ROUGE Scores: {'rouge1': 0.0430025915208296, 'rouge2': 8.739937283566147e-05, 'rougeL': 0.04283252114215165}```
what do you guys think? I changed my batch size from 16 to 8, but without adjusting the learning rate, which wasnt smart, So Ill think Ill go back to 16 for now or try or a lowering learning rate for 8.

[unkempt apex]

why are you training for different types of image ?

I mean the previous one was on another filtered images
now it seems another

[rich moth]

Im shuffling the data before the split now, which created a bit of more randomness.

[small wedge]

are you having issues with the amount of time the model is taking to train?

[rich moth]

The dataset is a big confusing, everything is saved all 31k rows under test instead of train, but all the strings for train, test eval exist. I had to split up the data manually.

[small wedge]

setting batch as low as 8 could be problems even if you adjust the learning rate since it might not be able to get a good enough gradient estimation to learn properly

[unkempt apex]

yeah make sense!

[rich moth]

after experimenting yours words ring true

[small wedge]

if you are struggling with the time it takes for your model to converge this is always a fun paper https://arxiv.org/abs/1711.00489

[rich moth]

You guys are seeing the results of batch size 8, ill go back to 16 and test this out again.

I should experiment with different learning rates at 16 too. Im using 1e-4 , maybe increasing that instead might be a good idea.

[unkempt apex]

but what is this green filtered image called as, or you have given them a name?

[rich moth]

The visual resconstructions? Why is it green?

[unkempt apex]

yeah

Updating DQN
current loss ->  tensor(3.7179, grad_fn=<SmoothL1LossBackward0>)
Episode 91 : Total Reward = -10
Updating DQN
current loss ->  tensor(5.3552, grad_fn=<SmoothL1LossBackward0>)
Episode 92 : Total Reward = -10
Updating DQN
current loss ->  tensor(4.8030, grad_fn=<SmoothL1LossBackward0>)
Episode 93 : Total Reward = -10
Updating DQN
current loss ->  tensor(3.1958, grad_fn=<SmoothL1LossBackward0>)
Episode 94 : Total Reward = -10
Updating DQN
current loss ->  tensor(2.1072, grad_fn=<SmoothL1LossBackward0>)
Episode 95 : Total Reward = -10
Updating DQN
current loss ->  tensor(2.2879, grad_fn=<SmoothL1LossBackward0>)
Episode 96 : Total Reward = -10
Updating DQN
current loss ->  tensor(1.9174, grad_fn=<SmoothL1LossBackward0>)
Episode 97 : Total Reward = -10
Updating DQN
current loss ->  tensor(2.5300, grad_fn=<SmoothL1LossBackward0>)
Episode 98 : Total Reward = -10
Updating DQN
current loss ->  tensor(4.0256, grad_fn=<SmoothL1LossBackward0>)
Episode 99 : Total Reward = -10
Updating DQN
current loss ->  tensor(4.2301, grad_fn=<SmoothL1LossBackward0>)
Episode 100 : Total Reward = -10

what are this loss values trying to tell, I mean I know the less the loss, the great modle performs

but hey how can I optimize my model now?

ohh, send the docs

then I have to run all 100 episodes again?
I have .pth file created after training

and yeah after lot of bug solving my RL model is ready with custom env!!

what is this image?

[brave sand]

can the reward function change every timestep?

[unkempt apex]

NICE!!

and what you are trying to achive with this

very much advanced level project , I mean GOD level!

what I can do with .pth?

I guess it contains parameters

[small wedge]

is this like a hypothetical question or are you asking about someone's project?

[mint palm]

will it be weird to ask HR to keep joining date as 2nd week of some month, because you want to resign after taking the annual bonus, so you can resign only in second week

[small wedge]

lmaoo

sounds like a good plan to me

[mint palm]

also, is the annual bonus given in 12th salary? if cycle does not follow financial year

or is it 13th?

[brave sand]

my project

[small wedge]

just fill your work calender with random fake timeslots until then "I'm sorry I'm very busy"

[small wedge]

this might be better to ask in #career-advice but I think it depends on the company, idk the standards

[mint palm]

lol i thought this was career discussions

[small wedge]

what kinda model are you using?

[brave sand]

huh? i’m trying to find a MARL algorithm that can use a changing reward function

[unkempt apex]

what to plot?
loss values?

[small wedge]

oh interesting, haven't heard of that

[brave sand]

is there anyway to bypass that

[unkempt apex]

shit I have do all 100 episodes again!

I forgot to store them on list

[small wedge]

to bypass what?