data-science-and-ml

[proud wing]

i was just trying to come up with a general case but i figured it would be way slower.

well you mean splitting the array into sub arrays right?

[topaz turtle]

well yes

i'm calling them tensors but ye they're essentially just n-dimensional arrays

[proud wing]

thats just being efficient

i already played with that too 🙂

@serene scaffold Good day to you!

[topaz turtle]

the splitting into subtensors to fit into cache part?

[proud wing]

yea

i started with a max_dimension of 32

to see if I could contract

[serene scaffold]

Good day, fellow intelligentleman

[proud wing]

The maximum size of an array is limited by the sys.maxsize variable, which is typically around 2^31 bytes.

If you're passing that up rn... you'd need an einstein and a chalk board with extensive proofs experience to prove the case beyond that... 😄

I'm a computer science lover with a nerdery for math so if the computer cant do it, I move on haha

[topaz turtle]

i mean sure, it's just that in my mind if i can come up with a way to do that for like 3 or 4 dimensional tensors contractions, i can easily generalise that to any arbitrary numbe or dimensions that i'd want my library to support

currently attempting to fish out of numpy developers how they acc did it lol

bc their C code is incomprehensible mess bc it essentially only functions to support the python wrapper

so i can't just take a look at it myself lol

😭

[proud wing]

The buffer protocol, which is an abtraction basically

I was thinking if my method using TRIS works faster than numpy even for the case we're evaluating

that might be quite valuable

[topaz turtle]

ye def

[proud wing]

it would probably become adopted by a lot of people as it could speed up inference

[topaz turtle]

i noticed that an arbitrary numpy tensor contraction takes about the same as a matrix multiplication with the same number of elements, so their tensor contraction is at the limit of matmul speed

so it would be quite astonishing if you made it faster

[proud wing]

Here's a useful trick with numpy btw. Use x.array_interface to explore the array

We might need to use numpsys dispatch mechanism

numpys

To check if a Numpy function can be overridden via array_ufunc, you can use allows_array_ufunc_override.

Looks like if you want to push numpy to its breaking point, there's some workaround.

I haven't explored it much though

If you go really low level with numpy, you get into some fun words like jiffies

@serene scaffold How many jiffies does it take to turn on a light bulb::D

https://github.com/numpy/numpy/blob/v1.26.0/numpy/core/code_generators/generate_numpy_api.py shows examples how they wrote C to abstract the api

They seem to have designed things fairly well to handle cases up to 3 dimensions

https://github.com/numpy/numpy/blob/d35cd07ea997f033b2d89d349734c61f5de54b0d/numpy/core/shape_base.py#L855

[arctic wedge]

numpy/core/shape_base.py line 855

# It was found through benchmarking that making an array of final size```

[proud wing]

Also worth reading: https://github.com/numpy/numpy/blob/v1.26.0/numpy/core/einsumfunc.py

I didn't know this but apparently _can_dot Checks if we can use BLAS (np.tensordot) call and its beneficial to do so.

They wrote: " If the operations is BLAS level 1 or 2 and is not already aligned we default back to einsum as the memory movement to copy is more costly than the operation itself."

Also you can see their exploration of einsum and testing it with various algorithm pathings (eg. greedy) or pre-computing the optimal path and repeatedly applying it https://github.com/numpy/numpy/blob/d35cd07ea997f033b2d89d349734c61f5de54b0d/numpy/core/einsumfunc.py#L1334

[arctic wedge]

numpy/core/einsumfunc.py line 1334

Chained array operations. For more complicated contractions, speed ups```

[proud wing]

Einops is a really big advancement in easing tensor manipulation

[pearl barn]

guys how can i run this pages locally on my anaconda

https://jovian.com/learn/data-analysis-with-python-zero-to-pandas/lesson/lesson-1-introduction-to-programming-with-python

is this better or should i switch to alex the analyst videos because i spent along time to only learn the basics and i still didn't finish learning braching with if else elif

[proud wing]

@pearl barn

You mean like: conda -n create "your_environment" ?

Just do something like conda -n create testingenv python=3.10
conda activate testingenv

then just type python on the console or make a new python script

python on the local console will allow you to explore those functions

@pearl barn

[pearl barn]

Didn't understand anything you are so good in this

[exotic star]

how can u make a program in python that can scan 100 plants(a list of plants u made) and recognize them?
my big brother works on a project with a few other people they need to hire a programmer for this and asked me if i can all though i'm few months in and idk much, my big brother is in nanotech so he doesn't know anything about programming and i'm curious about this

[trim saddle]

1. Create a anaconda environment, see @proud wing video, or google for how to do that
2. install jupyterlab with conda install -c conda-forge jupyterlab
3. start jupyter lab and open the jupyter notebook from your tutorial (have to download the .ipynb file)

[proud wing]

@exotic star There's already apps that do this on the app store. You'd need to train a classifier model on plants that are labeled, with an extensive exsting database of plants. It's not easily done with limited experience from scratch.

[exotic star]

If you were to hire a programmer, how much would it cost

to make handle the programming only(not the design of the app)

[proud wing]

You could check out fiverr or toptal and ask around there, you might get a quote from someone capable.

[exotic star]

got it, ty for the help

[covert finch]

hello everyone! I have been a stats programmer working (almost) exclusively with pandas for the last 7 years. I recently began a new job working out of a databricks environment, in which 95% of the notebooks are writted in pyspark. I am looking to sharpen my knowledge of pyspark, but I am having trouble setting up an environment to practice on my local machine since. Does anyone have any questions on how I can do a little pyspark studying outside of my org?

[agile owl]

you have to install spark separately then install pyspark

[covert finch]

yeah, Java too right? I just cant seem to get my environment correct

[agile owl]

why dont you say what the actual issue is because there's a lot of guides out there on how to set it up including the one from the spark docs themselves

[covert finch]

sure just give me one second if you dont mind

import pyspark
import os
os.environ["JAVA_HOME"] = r"C:\Program Files (x86)\Java\jre-1.8"
os.environ["SPARK_HOME"] = r"C:\Users\Vince\Downloads\spark-3.5.0-bin-hadoop3.tgz"
import findspark
findspark.init()
from pyspark.sql import SparkSession
from pyspark import SQLContext

spark = SparkSession.builder.master("local[*]").getOrCreate()
df=spark.read.options(delimiter=",", header=True).csv("fake_ae1.csv")
df.show()

FileNotFoundError Traceback (most recent call last)
Cell In[3], line 10
7 from pyspark.sql import SparkSession
8 from pyspark import SQLContext
---> 10 spark = SparkSession.builder.master("local[*]").getOrCreate()

this was in jupyter notebook in a virtual env

[pearl barn]

How to download .ipynb it's already on jovian website but don't wanna to use it

[trim saddle]

Idk if theresan option on jovian, but worst case, you can just copy the cells from online into a new local notebook

[pearl barn]

It shows run code locally but nothing happens after that

[trim saddle]

What does it show in your jupyter notebook locally? Can you provide code/picture, what it Shows and ehat you expected

[potent pollen]

Guys I'm desperate

I've created a librairy to create and use neural networks, and one to make my IA play Tetirs. I've been careful to the details that could cause a disfunction on my IA. The librairy I created is for unsupervised learning, and the problem is that after 500 generations of 124 agents, they don't play any better. I search on internet, and some tell me that Tetris cannot be learned with an unsupervised algorithm. It really bother me, since I've been working on it for months : are you able to explain the fact that my IA don't learn, or is it just that she cannot learn by herself Tetris?

[serene scaffold]

by unsupervised, do you mean reinforcement?

[keen delta]

Hey guys, i am trying to do a machine learning project and I'm stuck in part and not able to figure out what to do. The project is about pdf question answering using llms. I need help

[potent pollen]

Yes

[serene scaffold]

you can train a tetris playing agent with reinforcement learning.
Remember that artificial intelligence is AI in English, not IA.

[serene scaffold]

how far did you get, and what's the current stumbling block?
Remember that using LLMs is quite challenging.

[potent pollen]

Ok so I'm just bad, I will work on it. Thank you !

[serene scaffold]

You are great

[potent pollen]

Thanks, you're right. I tried to do my AI with a RNN, is it compulsory to use a CNN in our opinion ?

[serene scaffold]

I've never done anything with reinforcement learning, so idk

[small wedge]

It usually takes way more time to use a CNN in my experience. Generally it's easier for a model to learn when the input space is smaller.

[past meteor]

If you're operating on raw pixels for Tetris then a CNN might make more sense than an RNN

[past meteor]

It seems like this isn't even reinforcement learning but it's actually a genetic algorithm, is this correct?

[potent pollen]

Well thats what I did, I entered 34 neurons with processed data, but no results. My algorithm is mabe just too simple, as I was saying I'm not using Tensor flow or anything

[potent pollen]

Yes, my bad

Well done

[past meteor]

Why are you using a genetic algorithm for this? Any particular reason?

[potent pollen]

It was just an algotihm that I knowed how to code, I thought it would have work

Is that the problem ?

[past meteor]

It can work but it's a bit "wasteful". It's a black box optimiser (look this up) and there're options that can actually use more information of the game and learn in a more "directed" way, for instance deep Q Net (DQN)

A genetic algo might work still, mostly because Tetris isn't a super hard problem I suppose

[potent pollen]

Ok then I will check on what you tell me, even though it would be a relief for me that my alg works

[keen delta]

so i am trying to make a pdf question answering chatbot using google gemini's llm. I have made one with openai and falcon-7b llm and they're working perfectly however when i'm trying to make one with google gemini, i'm getting a lot of errors. I am using streamlit and the issue is that when i'm trying to run the code, i am getting no output

[agile owl]

what can cause this type of results in reinforcement learning

I have a branch for the reward wonder if the branch and nonlinearity there is causing this

it's an accumulator and if it falls below 70% of its original balance I give it a heavy reward penalty

otherwise the reward is the net return divided by an estimate of the variance of the return

I'm using PPO with LSTM from sb3-contribs

the observations are the variance driver, a variance estimate, and a handful of exogeneous state variables

Not sure how to design the reward to get what I want if I need to take out the branch for it to work properly

I played with genetic algorithms for a bit and they showed some promise at first but I think they are just variance machines

any information I gleaned from them that stood up out-of-sample was marginal

I ended up using reinforcement learning for my problem instead even though I had some fun messing around with evolutionary algos

for genetic algos to work I would actually have to be able to parameterize the underlying process in a way that aligns with what is causing variance in reality which is a nearly impossible task

reinforcement learning basically does that part for you

I've come to the conclusion after a couple months that genetic algos are basically trash for most things like zestar said

my reinforcement learner is trying to put the appropriate coefficient on the variance driver over time to form a series of products that optimize mean over variance over time but I have some special rules I want to add on top of that

I'd rather it do nothing than explore into spaces where it can end up in a certain state, basically like a game over for a video game

but even penalizing ending up in that state doesn't seem to prevent it reliably and it also has some really weird results on the learning curve

is that just the nature of the beast with reinforcement learning and I need to tinker with the weights of the reward if I want to branch it?

[final kiln]

got my gpu quota request accepted, cloudwatch all setup, logs broadcast back to the master node which I just log.info so that it gets printed to the runner logs. two missing pieces are the runner workflow with secrets and finish implementing the fault tolerance in the training loop, it's actually already implemented and tested, I just need some way of having aws notify the program that the instance is gonna be terminated in 2 min, other than that signal its all done.

[agile owl]

like what does this mean in reinforcement learning

when the rewards actually go down if you train it too long

if I keep letting it explore in theory should it eventually hit a new optimum

or is it "stuck"

[worldly dawn]

There are quite a few dependencies on how you model the agents, how you combine/mutate/evolve your population, on how you measure fitness, etc.

[potent pollen]

Yeah I know, but I didn't have a choice but to create my own, since I wanted to make it run on my calculator

[worldly dawn]

Sure, but there are entire books dedicated to genetic algorithms.
Without providing more info on your implementation, there won't be much that can be said

[potent pollen]

I am just modifing the weights and biases of the network using a uniform function, so it's basically random modifications ig

[worldly dawn]

oh so no selection or combination

[echo mesa]

Hello guys I'm facing with a problem. I'm currently in secondary school and at the end of school we are gonna learn calculus, currently I'm doing machine learning compilers and so on which would require the knowledge of calculus what I am wondering about is whether I should learn calculus on my own right now or I should wait till I learn it in school, would learning calculus be a waste of time on my own if I'm gonna learn it in school anyways or should I pursue it?

[proud wing]

@echo mesa why are you asking us?

@echo mesa You should learn what you are truly interested and excited to learn. Not what 'school' tells you to learn.

Some areas of math that will be useful for you if you want to do some really wild stuff with machine learning include linear algebra and tensor calculus.

[final kiln]

Yes

[proud wing]

Hey @final kiln where did you get your quota incrceased at?

[final kiln]

In the quota page. Just search quota on AWS console

[proud wing]

I want to test my general approach for calculating tensor contractions in C vs Numpys... if its faster I will probably need to publish a paper on it

@final kiln oh i just meant which provider are you using 🙂

@final kiln I use Google cloud for most things cloud, except ML

[final kiln]

I have funding for AWS. I co founded a 501c3 and we get tons of free stuff to do open source

[final kiln]

C is usually faster

What results did you get ?

[iron basalt]

If your C code is slower than Numpy then your C implementation is wrong.

[final kiln]

That is usually how it goes

[iron basalt]

(And/or compiler flags)

[iron basalt]

This goes for pretty much everything vs C, except maybe assembly if you try hard enough.

(Although you then usually are doing assembly inline in C via compiler extensions)

[final kiln]

My impression is that you need to be some sort of wizard to beat the compiler

[iron basalt]

Not really, compilers are good at repetitive optimizations that would take a human way too long to apply to the whole huge code base, but with time a human will always win.

It's like saying that ChatGPT writes better code (which is probably something people will start saying too).

It's about time spent doing it though.

But most of the time performance issues are from the 80% gains given by computational complexity (big O needs to be reasonable (but not best)) and not having the CPU do more work than needed in general (e.g. by choosing C over Python) for the same end result.

[worldly dawn]

it gets muddy very quickly. There are all sorts of optimizations, including those that benefit from observing the behavior of the program

[iron basalt]

(And knowing the general stuff like the CPU cache / are you compute bound or memory bound?)

[final kiln]

Uhm, all I know is avoid branches, keep data local and compact, inline stuff.

Has served me quite well

[iron basalt]

The gist is that your CPU can operate on data so fast that what often really matters is getting the data to the CPU in the first place. The CPU has its own local memory (the cache) to make this faster. To help the CPU you want to make memory access predictable and also it fetches memory in chunks, so contiguous memory.

(e.g. loading a single integer from RAM takes so long that in that time your CPU can do hundreds of additions / multiplications)

[final kiln]

I am aware. RAM fetches r slow

Like some hundred cycles or something like that

[worldly dawn]

also pipelines have an impact

[iron basalt]

Since your CPU has SIMD now and multiple threads and more, it can do a ridiculous amount of stuff in that same amount of time.

For that to become the bottleneck you need to do a ridiculous amount of stuff that touches the same memory (already in cache then in that case) over and over, e.g. matrix multiplication of large matrices.

[hollow sentinel]

so i got this data from this website: https://catalog.data.gov/dataset/national-student-loan-data-system-722b0

i'm kinda overwhelmed

i downloaded all their spreadsheets

i had a hypothesis that students from certain demographic groups (age, gender, ethnicity) might exhibit different loan default rates.

the problem is that the data that i'm looking at does not show any demographics

[hollow sentinel]

https://nces.ed.gov/programs/digest/d22/tables/dt22_331.95.asp?current=yes i switched to this dataset

trying to get clean data off this is not fun

[unique summit]

anyone want to spend the new year helping me find where the issue is with a gesture recognizer calculation:)

[agile owl]

It's very easy to pin a CPU with anything that's inherently iterative and if it's something that can be parallelized you can easily pin a server chip of arbitrary size with a large enough problem

i disagree with this idea that cpus are not a bottleneck you have to worry about, it depends on what you're trying to do

there's a lot of problem spaces that can't be vectorized to a GPU atleast not very easily

if you are doing evolutionary algos you can easily pin a cpu for hours and end up with garbage because it's a variance machine

reward = self.sharpe_ratio - prev_sharpe_ratio
if self.initial_margin > 0.2 * current_portfolio_value:
    reward = min(reward, 0)

so things like these in reinforcement learning, is it right to make them part of the reward return or should I make them deterministic constraints such that if the agent runs into them it will always do a corrective action

[worldly dawn]

I would rather describe them as meta heuristic.

Whether that runs on GPU or CPU would depend more on the problem to be solved

And whether or not you end up with garbage is more linked to how they are used, no different than any other method.

[agile owl]

I think EA are more prone to overfitting than anything else

unless you happen to know the problem has a certain structure and you're just solving for the parameters it's like shooting in the dark that the parameters you are solving for actually mean anything

[worldly dawn]

not really

That said, feature engineering is an important step for any ML/AI related problem

[agile owl]

engineering the features is different from engineering a parameterization of them

for RL you engineer features and the algorithm handles the structure

for EA you have to also provide a structure to the problem

and if it doesn't match reality, then you get garbage

which is easy to have happen I think from my experience with them but maybe other people have had better experiences

[worldly dawn]

In my experience, if you get garbage with EA, is that you are doing something really wrong

[agile owl]

let's say it's an agent problem like playing a game or optimizing some return over time

for RL you have to define the game and the features, sure, but you don't have to tell it what to do as a result of the features being some value or another etc.

if you want to use EA you have to have an idea of the causal relationship between the features and whatever the reward is

and learn whatever weights/parameters you apply to them using the EA

to me it seems easy to get that part wrong

[worldly dawn]

take NEAT for instance

you don't model the structure

you have your input and output

like any other ML model

but it could be anything else you think is appropriate

You don't need any causal relationship. That's not how EA work

[agile owl]

EA is just a way to solve for parameters right

the genes that solve a problem the best

[worldly dawn]

What matter is you need to make it friendly to your mutation/selection/combination operators

[agile owl]

what I got was something that fit in-sample time series data extremely well

and generalized very poorly out of sample

whereas with RL I get comparable performances

that's why I called it pejoratively a variance machine

I have no doubt that EA could reproduce what I'm doing with RL if I understood the right way to use the genes

but if I knew that why would I be using ML methods to begin with

[worldly dawn]

Right. And similarly, had you gotten great results with EA and terrible results with RL (even for a single typo), you might be stating the opposite.

What I am reacting to is not that EA should be the solution to all the problems. Far from that. I am reacting to calling it garbage. It's a great tool which works very well for what it's made

[agile owl]

I understood it to just be a way to fit parameters for experimental models

things where you aren't quite sure how to solve it using something like differentiation and gradients

[worldly dawn]

It's great when you are ok with approximate solutions, it's very expensive to derive a model or have to deal with non-linear stuff

And also it has a cool factor

I mean, it's a meta heuristic after all

not a random generator

[agile owl]

I think unless there's a clearer way to control overfitting then it's of limited use to complex problems like trading

and even if there was one, and I spent a lot of time thinking about it and talking with people about it, you still end up stuck if you find out you are overfitting and you don't know how to change the model to generalize better

it's just a lot more work

also even though you can parallelize it I don't think you can vectorize it in general

EA is only a partial solution I guess is the point and the other part of the solution can be arbitarily complex and something like RL can handle that better if you can't start to put into words how you should use the features you have available to solve the problem

you have to be searching in the right places to start with

I'd like something like NEAT for RL though

[worldly dawn]

overfitting is no different than other methods

[agile owl]

I think the whole point of EAs is it depends on what you are using the genes for

it seems like with NEAT the genes themselves are weights in a neural network that will handle the structure of the problem for you

but then it's really a neural network but you are using EA as the optimization methodology instead of gradient descent

so EA is just an optimization method

you still need something else to provide the structure to the problem

anyway feel like I'm just repeating myself at this point not sure what our difference in understanding is

like, how do you use EA?

[worldly dawn]

the contribution to NEAT is that you aren't just modifying weights, but also the whole structure

[worldly dawn]

I have used it in many contexts, from generative arts, agents in games, to generating code in b2b products

[agile owl]

right, where the structure is represented by the genes

you are providing the neural network and the ways its structure can be represented as a way to abstract away the structuring of the problem though

I was thinking of EA with freeform modeling by hand and solving for parameters

[worldly dawn]

I mean, only the game agent was using NEAT. The rest was using other structures

[agile owl]

I am curious how they are able to do both the structures and the weights with EA though

given that the number of weights youd have to solve for would be different depending on the structural parameters

so the individuals would be of different sizes wouldnt they

[worldly dawn]

I would recommend to read the paper about NEAT. It's a pretty cool paper and easy read too

And if you are in the mood, you could take it further with CPPNs

[agile owl]

I'm having some success with sb3 is there a library like that for NEAT

and is it as documented and easy to use

[worldly dawn]

there are some python neat libraries

I have been using java though for the NEAT stuff

[agile owl]

that's often scary because it means there's a diffusion and duplication of efforts

and none of them end up being obviously the right choice

[paper aurora]

Hi there, looking for someone as an intern who is good in image processing and know how to deal with PTZ/Network cameras.
Should be available for next 3 months as an intern. Hit me up!!

[final kiln]

donzies

gotta test it on a gpu machine tho

should work the same if the os is the same and the drivers are pre installed

its going for 100 epochs, so I got a lot of time for testing

im gonna see if I can get aws to send the spot notice of termination to verify that the fault tolerance isworking

ah no way to do it, I'd have to mock it

[final kiln]

mlflow is goated

it's backing up my models automatically to s3, providing fault tolerance and facilitating master-slave communication between my actions workflow and spot instance

goated

yep, like everything in the internet

works quite well tho, what happened

don't see how this can possibly substitute an ML Eng, it's more like a logbook or a database

im a bit confused tho

looking at their website they dont seem to claim that it is a substitute

it's possible they offer some sort of consulting service to maintain ml infra

well in any case, im super happy with my setup

I can run the experiments right from actions

then see the stats on mlflow, and if I dont like it I can redo it by setting the same params and some other adjustment

all with spot instances, which is super cost effective

and I can always change the model in code, test it locally with cpu, and then go on to gpu via the actions workflow

[final kiln]

I think this is a good point to stop and move on to the Shakespeare dataset, gpu is not fully figured out, but I'm sure it's some detail that I can handle once I need it, I also got quotas for the expensive gpus only so Im gonna have to wait a bit more on that

[past meteor]

I use MLFLOW a bit myself at work. I do it on-premise though, it's a fine tool.

[cosmic plover]

hello, I want to change my field to bioinformatics. can anyone guide for the starters>

[left tartan]

!rule 9 6 No recruiting here, please.

[arctic wedge]

Rules

6. Do not post unapproved advertising.

9. Do not offer or ask for paid work of any kind.

[final kiln]

Even with this setup, Im looking at 200-300 dollars just to reproduce the smallest gpt 2, so I think I'm gonna stop at the Shakespeare dataset

[wide tendon]

What algorithm would i want to use for beating a platformer DQN DDPG or PPO

[potent pollen]

I have 20% of parents (so ai of past generation), 60% of mutations and then some random ones

[worldly dawn]

yeah, in this configuration it feels more like random exploration than an EA.
I would suggest to stick with more classical approaches and to use a library. See https://www.geeksforgeeks.org/genetic-algorithms/

For instance, the top performers from the previous generations are fine being only 2-5% of the new generations. and 60% mutations is pretty high.

Note also that the operators do matter. How do you select individuals, through tournaments, roulette wheel, other?

[worldly dawn]

Here is an example of parameters from a random paper on using EA to control cars:

[potent pollen]

You guys are so handsome TYSM
I also do think I need to use a library but it would be a shame I have created my own for nothing
For the selection of individuals, I let them play a party and rate their game based on how much line they made, how much block they have displayed, the bumpiness of their grid, the number of holes, the mean height of their columns...
I will try modifying the criteria of selection, you're right, but if that was the problem, I guess they would have evolved a bit, at least they would have been slightly better, which is not the case

[worldly dawn]

For the selection of individuals, I let them play a party and rate their game based on how much line they made, how much block they have displayed, the bumpiness of their grid, the number of holes, the mean height of their columns...

This is describing your fitness function, not the selection process.
Let's say you have 100 individuals, each with their fitness score as established by your fitness function.
Let's say you need to select 10 individuals out of this population, what is the process?

[potent pollen]

I take the 10 individuals with the best fitness

[worldly dawn]

I am asking this because it is important to apply pressure to the evolution.
For instance, if you select the top 10, it's a start, but it also means that they are considered the same regardless of their success.
In natural selection / evolution, the most successful entity is the one that tends to reproduce the most since it's the most fit for the environment. And as such, it would make sense to select multiple times the top individual if its fitness is multiple times better than the rest

That's why I would suggest to look into the "roulette wheel" selection operator or the "tournament" one

[potent pollen]

Yeah right, I will go check on that one thanks !

But I still think there is another problem to it

Maybe it's just me who messed up a thing on my code, though

[worldly dawn]

Tangential, it looks like your home made library still has some way to go before implementing to basics of GA. Coupled with a tough RL type problem you are trying to solve, that means mixing two different and very complex problems.

I would suggest to pause your problem for now and focus on something simpler until you get your GA library working with the basics of GA

Once you got your GA library solved, it will be easier to focus on your game ai problem

[potent pollen]

You are so impressive

[worldly dawn]

lol no

[potent pollen]

Thank you for your tips, I will modify this, can I call you afterwards if it still doesn't work (or if it does to thank you)?

[worldly dawn]

np, have fun!

[crisp salmon]

Ive uninstalled and reinstalled h5py multiple times now. Im trying to use tf 2.9 with nvidia cuda and cudnn so that I can utilize my gpu. Initially I did this project with my mac book and on its cpu. but after the first epoch I keep getting this error. I dont know what to do anymore and im beginning to get pretty frustrated

I think it has something to do with the version compatibilities but ive found nothing that says which version of h5py i need to use with tf 2.9 or anythintg of that matter

Im contemplating redoing this entire project in pytorch at this point

[paper aurora]

My bad!!

[agile owl]

    def reset(self, seed = 42):
        self.period = np.random.randint(0, len(self.periods)-1) if len(self.periods) > 1 else 0

    def reset(self, seed = 42):
        self.period = self.period + 1 if self.period < len(self.periods) - 1 else 0

What are the pros and cons of each of these approaches to iterating over instances of games/periods in RL training data (picking a random game/period each time you reset the environment vs iterating over them sequentially and looping when you hit the end)

[indigo moth]

hi guys, I came across this old exercice I did last year during my studies, but couldn't understand a part of the code:

a = np.random.randint(100, size=10)
print(a)
a = [*a]
a[a.index(max(a))] = None
print(a)

anyone can tell me what I meant with a = [*a]? it doesn't seem to do anything I can't remember what I did here

P.S: another example of why one should always comment code lol

[mild dirge]

It makes a list of the elements of a

* in that context means unpacking of an iterable (in this case a numpy array)

If the numpy array a has the elements [1 2 3 4] and you write [*a] that is the same as writing [1, 2, 3, 4]

Better to write list(a), a lot more understandable

@indigo moth

[final kiln]

today is the day I teach a machine to write poetry

[brittle storm]

how to make a command that would work like this:

if subject_code == "0417" and ranses == "w" and paper_number == "2" or "3" and ranyear == "2019":```


i want it to proceed only if all the conditions are accepted

elif subject_code == "0417" and ranses == "w" and paper_number == "2" or "3" and ranyear == "2019":
sesh = "November"
qpv = f"{paper_number}"
msv = f"{paper_number}"
qp = f"https://edupapers.store/wp-content/uploads/simple-file-list/CIE/{programme}/{subject_name}-{subject_code}/{ranyear}/{sesh}/{subject_code}_{ranses}{ranyear[2:5]}_qp_{qpv}.pdf"
ms = f"https://edupapers.store/wp-content/uploads/simple-file-list/CIE/{programme}/{subject_name}-{subject_code}/{ranyear}/{sesh}/{subject_code}_{ranses}{ranyear[2:5]}_ms_{msv}.pdf"
print(qp)
print(ms)```

[mild dirge]

The only thing you need to change is the or

!or-gotcha

[arctic wedge]

The or-gotcha

When checking if something is equal to one thing or another, you might think that this is possible:

# Incorrect...
if favorite_fruit == 'grapefruit' or 'lemon':
    print("That's a weird favorite fruit to have.")

While this makes sense in English, it may not behave the way you would expect. In Python, you should have complete instructions on both sides of the logical operator.

So, if you want to check if something is equal to one thing or another, there are two common ways:

# Like this...
if favorite_fruit == 'grapefruit' or favorite_fruit == 'lemon':
    print("That's a weird favorite fruit to have.")

# ...or like this.
if favorite_fruit in ('grapefruit', 'lemon'):
    print("That's a weird favorite fruit to have.")

[mild dirge]

@brittle storm

[brittle storm]

but now.. i asked to print the a random paper that has these:
subject_code == "0417" and ranses == "w" and paper_number == "2" or paper_number == "3" and ranyear == "2019":

it prints this now.. https://edupapers.store/wp-content/uploads/simple-file-list/CIE/IGCSE/Information-and-Communication-Technology-0417/2022/November/0417_**s**22_ms_2.pdf.

which is invalid cuz s is wrong

it is supposed to get on w

[mild dirge]

I'm not sure about the priority of the and and or

Try adding brackets

subject_code == "0417" and ranses == "w" and (paper_number == "2" or paper_number == "3") and ranyear == "2019":

[brittle storm]

i did that and now it doesn't do anything

it printed once just now

and i ran the command again

it didn't print

@mild dirge

[mild dirge]

I'm not sure, the syntax corresponds to the logic I think you want to implement. I don't know where the problem lies.

[neon field]

Anyone just give me a whole machine learning project for social good already please!!!
I don't wanna do this 😭

[odd meteor]

Have you checked online for possible suggestions? If yes, what did you find?

[agile owl]

how many more timesteps would you run this before you consider it converged

(this is PPO btw)

[final kiln]

that looks like a fractal time series

[agile owl]

well it's convolved with a window of 50 so it could look even more like a fractal if I didn't smooth it but that's how it looks when it converges slowly I think

[final kiln]

uhm, the overall trend seems to be a linear function, so no convergence to a given value

tho I dont really know the context here

[agile owl]

It looks like diminishing returns have set in to me

but hard to tell where that last oscillation will end up

I was trying to avoid running it for 2e6 because I will probably iterate on it again anyway

but I guess that's how I will tell for sure

[final kiln]

not 100% sure what you're doing, but I'd be tempted to run several experiments witth diff seeds and place the graphs on top of each other

[agile owl]

that's a good idea

this is what I'm doing, big business

[versed gulch]

I have a text file which looks like this:
number_of_nodes: 9
842 2578 0
842 2578 1
842 2578 2
842 2578 3
842 2578 4
842 2578 5
843 2579 6
843 2579 7
843 2579 8
number_of_nodes: 4
926 2206 0
927 2205 0
927 2204 0
927 2203 0

The lines number_of_nodes: 4 represent how many coordinates there are below, what I want to do is read this file in and have list of nested lists containing these coordinates, i.e the [[...], [(926, 2206, 0 ], ... (927, 2203, 0) ].
Any help will be much appreciated

[agile owl]

this is what it looked like with another seed, much more convincing diminishing returns

although the policies it came up with are quite different

[final kiln]

uhm, looks like you have a random walker

one way to look at this

look at the y axis, and see the difference in y from each step i to the next step i + 1

if you histogram it, you will likely see the gaussian distribution

[agile owl]

reinforcement learning is supposed to do that

to some extent

it's exploring

it doesn't monotonically increase

the trend is definitely higher up to a point though

importantly it gets into positive rewards territory

so you could say it "passes" the problem

it could be arbitrarily better though

I'm going to do some feature engineering next

especially since this is an on-policy learner

it only can learn from what it does

other methods try to figure out what the best action is instead of trying to make the one it's doing the best one possible if that makes any sense so they have nicer learning curves most of the time but they often fail to actually solve the problem

[final kiln]

again I don't really know the context, just gathering from the graph alone haha

shakespeare is coming along

[boreal gale]

i have not followed what you are doing so far. but

1. what is reward here?
2. isn't it bad that it's an on-policy learner? the reward still flip flops and some time goes into <0 reward region, isn't that bad for when you are actually trading?
3. will you be validating your agent against a dataset that it hasn't explicitly learnt from? i would be super anxious about overfitting here

[agile owl]

You only use the saved model from the best reward timestep

reward is change in portfolio value / value at risk

yes I am doing the out-of-sample evaluation but I am only looking at one year period out of sample whereas I'd like to have more to have a "distribution" of out-of-sample performance

importantly the on-policy learner is the only one that actually gets to positive rewards

the off-policy methods overleverage themselves

and blow up

namely PPO vs SAC

I'm looking at PPO and PPO with an LSTM architecture in the policy network vs SAC

and SAC gives a lot nicer training chart but i can't get it to stop overleveraging itself even if I try to penalize it in the reward

the on-policy results are much better out-of-sample too

[boreal gale]

that's cool 👍
reinforcement learning is not my strong suit.

what asset class are you trading if you don't mind me asking? hopefully you are getting some nice out of sample sharpe/calmar ratio already 😄

[agile owl]

this is actually doing the thing that everyone says is a fool's errand

I figured if I can trade 10y duration then I can find a model that can trade almost anything

I've got 1.5 sharpe ratio on a combination of 10y, 2s10s curve and usdjpy models

the correlations were low and the correlation between 10y and 2s10s was even negative

I figure if I do some feature engineering I can get the sharpes convincingly above 2

[boreal gale]

that would be sweet 😉 good luck!

[agile owl]

right now I have a pretty simple model where it's a natural gradient boosting fair value model based on macro variables, a garch-like variance estimate using Light GBM, and some indicators for economic events like NFP and CPI days

I figure if I add some more of the inputs I have on a daily frequency directly to the agent's observations that alone might improve performance

right now I have a lot of inputs that just go into the fair value model without being exposed directly to the RL agent

what I'd like and don't have is historical consensus estimates so it could try to learn a response to actual economic releases in a meaningful way

there's like one company that sells it and it's a ridiculous amount of money

[boreal gale]

what kind of estimates?

like CPI? estimated by a panel of analysts?

[agile owl]

yeah

banks put out estimates for those things like earnings

and they get compiled into a consensus number

so for each release of significance there is a number to compare it to

and say "is this higher or lower than expected"

that delta is what drives the market reaction not the value itself

so if CPI goes down and everyone expected it to go down your model shouldn't react to that like CPI is low

[boreal gale]

yeah, the expected value is all baked into the price already

[agile owl]

it needs to know what the expectation was before it makes a judgment

[final kiln]

couldn't you train a giant language model to do this stuff, feed it a bunch of documents and google searches and have it come up with predictions

have it decide what is important

[agile owl]

I've thought about that but the market isn't democratic

[final kiln]

well

[agile owl]

you could maybe get economic sentiment in general from that

but I'm sure it's noisy as hell

[final kiln]

from what I know it does require domain knowledge to be predicted

like, people who dont know about it are discouraged to participate in the first place

thus those who stay know their stuff

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

[agile owl]

You could maybe in theory extract historically significant economic relases from news stories

from places like Bloomberg and WSJ

[final kiln]

I say feed it the entire internet and have it converge to a decision, or at least, a slice of the internet, and during training you give it the internet

[agile owl]

I would actually like to set up a sentiment model for different assets using bloomberg news stories

the reaction of the internet and the reaction of the market are not similar enough to just use the entire internet

financial media is the way to go

[final kiln]

right, the model would filter out what's not relevant

like, training a language model is a compression procedure

[agile owl]

when people talk about economic data in general it's much more likely to be about politics than actual economics for instance

since economic data gets politicized

[final kiln]

yeah, I think economics as a field seems to interplay a lot with politics, but idk much about either of those topics

still goin

[agile owl]

nice

so what does it do again?

[final kiln]

well

at the moment it spits out non-sense

To be or not to be, that is'###--/00.033300--///>3<:!3030..<=/::.++0O3..3.:L0

the part that is not nonsense is the prompt

im gonna let it run for longer ig

[agile owl]

what is it meant to do, generate from shakespeare?

[final kiln]

yes

should write poetry

YES

I was just sampling it incorrectly

To be or not to be chopped,
He cannot tell thee to come to have, and we will play him?
If everlastingly king to deny straight: therefore, methinks you,
Must myself against you hear his presence.

LADY CAPULE

it's beautiful 🥲

im gonna have endless fun with this thing

[hollow sentinel]

i'm a bit confused

type                                    object
unit                                    object
creationdate                            object
startdate                               object
enddate                                 object
value                                  float64
HKMetadataKeyHeartRateMotionContext    float64
HKMetadataKeySyncVersion               float64
HKMetadataKeySyncIdentifier             object

most of these columns are objects

i'd like to convert creationdate, startdate, enddate to datetime

[agile owl]

pd.to_datetime

[hollow sentinel]

heart_df["creationdate"] = pd.to_datetime(heart_df["creationdate"])

[agile owl]

is it telling you it's an invalid format or something?

[hollow sentinel]

i tried this and tried testing again to see if it would change to datetime

UserWarning: Could not infer format, so each element will be parsed individually, falling back to dateutil. To ensure parsing is consistent and as-expected, please specify a format.

[agile owl]

well, is there a consistent format?

[hollow sentinel]

5/27/2022 8:02:00 AM

[agile owl]

did you check and try specifying it?

is it consistent though?

sounds like it might not be from that warning

[hollow sentinel]

is there any way to check if it's consistent besides scrolling through the entire file?

[agile owl]

you can do some string processing

how big is the file

[hollow sentinel]

31.4 mb

[agile owl]

sounds like you would probably need to come up with a way to figure out using python string processing methods

maybe regex

or

you can just try to specify that format

and see if it tells you there's an invalid value given that format

that might give you some hints

[hollow sentinel]

any way i can call datetime.strptime on a pandas column?

no, incompatible types.

i can try using a .apply function

[agile owl]

there might be a way with the dt accessor

I mean that's what to_datetime is supposed to be doing

[boreal gale]

!d pandas.to_datetime

[arctic wedge]

pandas.to\_datetime

pandas.to_datetime(arg, errors='raise', dayfirst=False, yearfirst=False, utc=False, format=None, exact=_NoDefault.no_default, unit=None, ...)```
Convert argument to datetime.

This function converts a scalar, array-like, [`Series`](https://pandas.pydata.org/pandas-docs/stable/reference/api/pandas.Series.html#pandas.Series) or [`DataFrame`](https://pandas.pydata.org/pandas-docs/stable/reference/api/pandas.DataFrame.html#pandas.DataFrame)/dict-like to a pandas datetime object.

[boreal gale]

use the format argument here just like how you would in datetime.strptime

[final kiln]

letting it go for 500 epochs

[agile owl]

I figured out why the off-policy learners were sucking

[final kiln]

there's a lot of things im not doing there that could improve this

[agile owl]

I forgot to tell the agent what its var was so even though it was using it in the reward function it had no way of directly observing it or using it to constrain its behavior

that's why they kept overleveraging

I'm amazed PPO worked anyway

engineered some new features and our rewards are higher than ever

[final kiln]

now that I have a first model that writes poems, I'm gonna setup a more serious infra to experiment with this, I'm mostly gonna copy what I did until now for the gpt array sorter, might need to adapt it a bit for gpu. After that it's time to explore my idea of using metric tensors for performing self attention and compare the results. I mean, tbh I kinda wanna do that now, I'm gonna run a quick loop to see what happen

it was following it a little bit too closely, so I restarted the run, got a print in there just to be certain the mod is there

but it seems that it was right

kinda weird that there's no difference, I expect that at least it takes less time training

for my last unsuccessful prediction, the newer version will start plateauing at a higher value

nvm I was right, mine is gonna end up being a bit faster

if this stays like this I'm gonna start getting excited. But in all likelihood it's gonna converge to a higher loss value because there's less parameters around and the values they can take are constrained

[agile owl]

much better

I like reinforcement learning because it's optimistic

it wants rewards

[final kiln]

That does look better, but you should probably use a fixed y-axis size

[final kiln]

This one doesn't reach 10

So it is better

I'm assuming higher = better

[agile owl]

I am only interested in the shape really

of course the value to

but this chart is for the shape

if I were to plot th em all to compare then it would be normalized obviously

[final kiln]

Yeah the other one looked like a random walker on the y axis

[agile owl]

it is supposed to random walk a bit

but I wasn't giving it enough information to reliably solve the problem that was my bad

I also switched to an off-policy method

[final kiln]

Right, but I'm guessing that a pure random walker would mean something is wrong

I remember studying this class of graphs in college. Even made a bunch of sims that generate them

[agile owl]

it wasn't a pure random walker

it had a definite up trend in the beginning

then it random walked around like 1

but it became definitely positive from negative

so there was something, and it was only with ppo because ppo is risk averse

compared to the off-policy learners

[final kiln]

The only way to confirm would've been to run several experiments. A random walker can produce any of the observed patterns via pure chance.

[final kiln]

This one is much more clear

[agile owl]

I ran it again with a differesnt seed and got the same pattern

[final kiln]

The other ones looked random

[agile owl]

nah that was just exploration

I'm sure if you looked at the stats that the moving average change was significant

difference is due to a couple factors including changing the model

[final kiln]

This is super random walker behaviour imo.

[agile owl]

the moving average definitely changes from negative to positive

and then it flatlines

[agile owl]

I ran it again with another seed and found a similar shape

it was two things: this one is actually truncated in the x-axis

and the model is different

and I didn't give it as much information

I mean the y -axis not x-axis sorry

it didn't start out as negative

that is just chance

[final kiln]

I'd fit it to gaussian + constant value

No

[agile owl]

the optimal value won't change

but the starting point did

[final kiln]

Uhm, + x

• c

Giving it a bias to go upwards

[agile owl]

well, it performed out-of-sample

just not as well

as the new one

reinforcement learning is different from supervised learning

there's a bit of "random walking" by design

it's changing the strategy to try to find new things to do to ultimately get to a higher end point

the on-policy ones tend to do that more aggressively

because they have limited information compared to off-policy

so they need to move around more to find a better policy

the other chart I showed more recently where it looked more like an inverse of your loss chart was from an off-learning policy

off-policy learner*

[final kiln]

That sounds ineficient tho

Trying out random stuff til it gets it right

[agile owl]

in a lot of cases they give more robust solutions

they also have better convergence criteria

even though you said it looks like a random walk

the other ones just became more negative

[final kiln]

It is one, you can even check that it's statistical properties are invariant under scale changes

Making it a fractal

[agile owl]

it goes through cycles of exploration and exploitation

but if you insist sure it was a pure random walk

I'm just glad I got the off-policy learner working

[pearl barn]

guys is maven analytics course for excel is good ??

[odd meteor]

I thought you were using Jovian to learn Python or so... You ditched it for excel now? How was your experience with Jovian?

[agile owl]

so here's another example

this is SAC (off-policy)

This is PPO (on-policy)

you might look at these and say wow PPO sucks in comparison but in reality it's not that clear on out-of-sample data

this is what they did respectively out of sample

the one on the right is a lot more stable and predictable in behavior

might want to run both and put a bigger weight on the less risky one

[agile owl]

I'm going to try to tune it to take less risk but the on-policy one takes less risks by default

[agile owl]

we've broken 2 sharpe it's time to hook it up to paper trading and demo it

[keen delta]

hey guys has anyone worked with llms here? i kinda need a small help

[lapis sequoia]

Hi, I'm excited to join here.
I'm a professional machine learning developer and have 8+ years of experience of developing data science, image processing, optimization projects.
Image preprocessing, deep learning, time series processing, dimension reduction and optimization are my major.
Nowadays I'm looking for employment opportunities and willing to do full-time/part-time.
Please DM me if any Employer is interested about me. thanks...

[final kiln]

read the rules

[whole egret]

what are you trying to do?

[lapis sequoia]

i wanna solve image processing or time series tasks...

[final kiln]

toda

my keyboard is\ going cra rn, the caps\ l.ock is\ del.eteing chars\

It's restarting, I think I'm gonna have to check for malware.

Anyway, today I'm gonna setup the rest of the infra for training GPT on a GPU machine without supervision and with fault tolerance.

It's mostly copy paste from the previous workflow + some corrections on a couple things I missed in the model code

[potent pollen]

I've implemented a GA, with elitism, crossover and mutation of parents, and some random agents. I've made a generation of 400 agents, with 8 elitists, 146 parents, 146 children and 100 random. Mutation probability is at 12.5% because I have my data on 8 bits. After ~60 generations, I could still not witness any upgrades.
Though, learning about GA and programming it was a fun task so thank you! But I still think the problem is something more simple...

[spark nimbus]

I'm noticing pycharm has intelligent schema autocomplete with Spark with read statements, is there a way to inform it about schemas of dataframes passed to functions?

[final kiln]

the first self attention module of the metric tensor network

[final kiln]

[pearl barn]

How can get Maven Analytics courses for free or by someone who can share thier Udemy username and password

[long locust]

Do not request other people's credentials, that is not appropriate

[pearl barn]

Learning with high quality should be for everyone

[spark nimbus]

and teachers deserve to get paid

[long locust]

Asking for someone's login credentials is not appropriate, not sure how that is related

[tidal bough]

"maven" sure is a confusing name for an analytics platform, I was like "wtf, an entire course on, what, doing analytics on Maven downloads?"

[stark bay]

What is the best way to smoothen out or make good predictive accuracy score of linear regression model.... how to eliminate fluctuations/noise in a data that has a lot of hips and hops due to being in micro scale

[simple plinth]

from where should i start machine learning? i know python basics, im c++ programmer and im good at maths and logic building

[stark bay]

Datacamp and kaggle have a lot of courses for such... thats a good point

[simple plinth]

can you please share the links?

[elfin sinew]

Hello

Can anyone suggest me any YouTube channel for python ?

[stark bay]

Link

[elfin sinew]

@stark bay ?

[stark bay]

Yes?

[elfin sinew]

Plzz can you suggest me any YouTube channel for python ?

[stark bay]

I dont use yt

[elfin sinew]

Did you learn py or advanced py ?

I really need some suggestions

[stark bay]

I think u will get more relavant info here since i am not the best person to ask for that

What is the best way to smoothen out or make good predictive accuracy score of linear regression model.... how to eliminate fluctuations/noise in a data that has a lot of hips and hops due to being in micro scale

[simple plinth]

it is redirecting me to google.com

[elfin sinew]

Do you work something ?

I can help you if you want

BeCause i want to learn something

I just want any experience how to work or what to work

Sorry if i am disturbing you

[junior spruce]

Hello, I am a high school student interested in artificial intelligence. I know some Python. What can I do at this stage without entering the field? I mean I need a strong foundation to enter it (skills). What are these skills and what is the foundation that I need?

[simple plinth]

wanna start together?

[junior spruce]

Yes

[serene scaffold]

Take the most advanced math courses that you can and get into a CS university program with an AI concentration

And do well in school in general. But in STEM especially

[junior spruce]

Well, sir, I will definitely take the mathematics programs for those in my country who never accept high school students in universities

[simple plinth]

are u indian?

[junior spruce]

No iam algerian maybe you heard about it or maybe you didn't

[simple plinth]

IK algeria very well, its a north african country more closely related with arabs

[junior spruce]

Haha, you seem to know it, but there are many of its residents who are not Arabs, including me

[simple plinth]

cooool so you are african?

[mossy sable]

anyone know the best way to locally host an llm cuz i dont have a openai llama thingy

[odd meteor]

Get a major in CS or Statistics while you build on your python skill gradually.

[junior spruce]

You can say it

[agile owl]

trading USDJPY

[mossy sable]

live trading bot?

[junior spruce]

Ok sir thank u so much for this

[agile owl]

it's in development

I'm hooking it up to start paper trading now

[mossy sable]

what training api u using for trading playform or r u

im assuming ur not running it on a live account

[agile owl]

what do you mean training api

it's a stochastic actor-critic model from sb3

[mossy sable]

to actually make the trades are u using a sim api

[agile owl]

oh ibkr

[mossy sable]

that has actual prices but not actual moneyt

[agile owl]

they have paper trading

[mossy sable]

k

[odd meteor]

Is local your last resolve? Why not try some cloud options? Some of them even have free tier plan.

• Heroku, Streamlit Cloud, Cerebruim, etc.

Check out https://www.cerebrium.ai/

[odd meteor]

You're welcome sir ✌️. If you'd have gotten to >= 18 years by 3rd quarter of 2024, I'll recommend you consider applying to attend Deep Learning Indaba 2024.

All the best in your endeavour

[stark bay]

What is the best way to smoothen out or make good predictive accuracy score of linear regression model.... how to eliminate fluctuations/noise in a data that has a lot of hips and hops due to being in micro scale

[agile owl]

you can try rescaling

using log scale or something like that

note that you will have to exponentiate the prediction from the log-scale model to get the prediction in normal terms

[hollow magnet]

Hi ! Does anyone has any tips about good books to learn and deepen the knowledge in data science

[long locust]

there should be a few in the pinned messages in this channel

[hollow magnet]

didn't know it was a thing, ty

[real goblet]

Hey buds. So I have excel spreadsheet named data.xlsx

df = pd.read_excel("data.xlsx")

df.drop(df.columns[df.columns.str.contains('unnamed', case=False)], axis=1, inplace=True)

I try to do this but there is still freaking indexes

could someone fix it please

[hollow magnet]

@long locust Have you read the books that are recommended in the pin message ?

For a litlle question

Or did someone ?

[long locust]

I have not gone through them, what question do you have?

[hollow magnet]

Well if I read 1 or 2 of them in their entirety, I'm good to go right ?

Because I read that the best way to learn is to apply the knowledge, but in datascience, I literraly have 0 idea of what exercize to do to apply and learn

[odd meteor]

Sure. Just focus on one book at a time, If you can successfully finish https://mml-book.github.io/ and https://www.statlearning.com/ you're good to go.

[hollow magnet]

Thanks !

[odd meteor]

The D2L book from Microsoft is also a piece of art. https://d2l.ai/chapter_introduction/index.html

[serene scaffold]

df = pd.read_excel("data.xlsx")
df = df.drop(df.columns[df.columns.str.contains('unnamed', case=False)].tolist(), axis=1)

Try this.
Also, just never use in-place

[real goblet]

There is still freaking index

df = pd.read_excel("data.xlsx")
df = df.drop(df.columns[df.columns.str.contains('unnamed', case=False)].tolist(), axis=1)

df = df[['Game Name', 'Region', 'Group Override']]

        Game Name   Region Group Override

0 Dead Space Ukraine No
1 Rust Russia No
2 Lethal Company Russia Yes
3 Grand Theft Auto V Ukraine No

[serene scaffold]

by "freaking index", do you mean the 0, 1, 2, 3 on the left? because there must be an index. you can never stop having it no matter what.

[real goblet]

Yes

why cant I have it

[serene scaffold]

every row always has an index no matter what. you can print it without showing the index, but it's still there.

do print(df[['Game Name', 'Region', 'Group Override']].head().to_string(index=False))

[real goblet]

So there is no sense in droping index?

[serene scaffold]

"dropping the index" is impossible.

that's like saying "I don't want the columns to be labeled or numbered"

how would you get the first column of the dataframe if the column doesn't have a label, or a number?

[tidal bough]

perhaps you want to do .set_index("Game Name", drop=True)? having that column be the index would be a reasonable choice.

[pearl barn]

What are the best courses on Udemy to learn Python data analysis the basics and fundamentals to NumPy and pandas?

[agile owl]

I think what he wants is to set the index to game name

[final kiln]

by tomorrow I'll have a full variation on the transformer architecture, how do I go about making a fair comparison between them ?

[agile owl]

so it displays without the series index

[whole lichen]

who wants to practice matplotlib with me?

[worthy jewel]

Hello everyone,
Tyler here I’m a Comp Sci student currently on a data science internship. I also have my own startup McCarthy & Brogan Solutions.. we’ve been setup for around a year now and starting to get invited into various factories around the UK looking at how our services (primarily focused on maintenance & repair in this case) can increase efficiencies using AI amongst other things. We also have a subsidiary SmartFormAI with which we have just built a document automation application utilising LLM, OCR and GAR. I’d love to get to know some of you, please drop me a DM!

[serene scaffold]

This server is not for advertising or self-promotion

[worthy jewel]

It’s an introduction mate chill out

Others have done the same?

[final kiln]

I believe you are breaking the rules

[agile owl]

@serene scaffold can I promote you?

[serene scaffold]

No

[agile owl]

dang

[worthy jewel]

Apologies! Must have missed that

[final kiln]

    def forward(self, in_sequence_bwc: Tensor) -> Tensor:
        batch, words, coordinates = in_sequence_bwc.size()
        k_dimension = coordinates // self.NUMBER_OF_HEADS
        pre_metric_tensors_nww = self.pre_metric_tensors_nww.masked_fill(self.MASK_ww[:,:,:words,:words] == 0, 0)
        metric_tensors_nww = pre_metric_tensors_nww @ pre_metric_tensors_nww.transpose(-1, -2)  # ensures symmetry and positive definiteness


        all_projections_bwc = self.projections_cc(in_sequence_bwc)
        all_projections_bnwk = all_projections_bwc.view(batch, words, self.NUMBER_OF_HEADS, k_dimension).transpose(1, 2)

        all_dot_products_bnww = all_projections_bnwk.transpose(-1, -2) @ metric_tensors_nww @ all_projections_bnwk
        all_dot_products_bnww = all_dot_products_bnww / math.sqrt(k_dimension)
        all_dot_products_bnww = all_dot_products_bnww.masked_fill(self.MASK_ww[:,:,:words,:words] == 0, 0)

        nudged_vectors_bnwk = all_dot_products_bnww @ all_projections_bnwk
        nudged_vectors_bwnk = nudged_vectors_bnwk.transpose(1, 2).contiguous()
        nudged_vectors_bwc = nudged_vectors_bwnk.view(batch, words, coordinates)

        out_sequence_bwc = self.projection_cc(nudged_vectors_bwc)

        return out_sequence_bwc

this is my proposed self attention mechanism

every head has a projection matrix that compresses the embeddings, and a metric tensor that is used to calculate the dot product between all the elements in the sequence

and that's pretty much it, each projection gets scaled according to the dot products and then it's concatenated and mixed

[agile owl]

hey everyone I'd like to promote the Greenest Admin

he's very Green and very much an admin

[serene scaffold]

@agile owl don't shitpost in our wonderful data science chat

[ruby grail]

Hello. I have a question, I don't exactly understand what this X_embedded mean or how to use it. Can anyone please give a hint?

[worldly dawn]

Mixing elitists and parents does not look normal.
There are also too many pieces. I would still suggest to either use an off the shelf library or to validate your library on something simpler.

[shut slate]

Can someone please explain vectorization in Pandas and why do we not have to do a for loop? And how would I know which methods can be used for vectorization?

[agile owl]

it's setting up calculations to be applied across an axis in parallel instead of sequentially that's why you don't need a loop

in general if something is autoregressive or recursive it can't be vectorized

anything inherently sequential can't be vectorized

or at least not very easily afaik

[heady sierra]

Hello. I am working on a reinforcement learning model using a gym-anytrading environment. I am having this bug but I couldn't find a solution anywhere. Can someone help?

[agile owl]

the error seems pretty straightforward to me

you need to have a certain observation shape

[pastel lake]

hello guyzz, i learned python and wants to learn machine learning, can anyone share some advice or roadmap to give a great start at my machine leaning journey

[heady sierra]

Apparently while debugging the code, I forgot to rerun the cell above so that fixed the issue. But now I am trying to add some indicators to a custom stock env. It takes gym-anytrading's StockEnv. I am getting this error:

[agile owl]

sounds like it needs a data attribute. did you read the environment implementation or is there a specification for how to subclass it?

I just make my own envs so I can't help you there

I only need to implement step and reset

i'm guessing the data attribute just needs to be some dataframe with certain column headers

[heady sierra]

I think I understand the error better. I need to check the documentation again.

Thanks for the help

[gloomy parrot]

hey everyone, i just want to ask on how can i stitch multiple images of receipt?

Does anyone tried it before?

[woeful fossil]

sup everyone, im new to this chat and Ai/data science. i know some basic stuff about Ai and data science but im wondering if someone would be willing to help me learn more on this topic.

[lapis sequoia]

Does anyone actually use feature mapping as oppose to kernels/gram matrix in SVM?

Cornell says that it's more efficient in lower dimensional feature maps but wouldn't you have to calculate the inner product anyway? So how is it more efficient when the kernel can do this without the explicit mapping

[stark ermine]

Good morning, gentlemen,

I'm writing to you because I can't find the solution to my problem after much research.

I'm running a log-linear regression on the adjusted price of a stock (dependent/target variable = Y).
in order to make the relationship between my variables (price/date) more linear (especially if the distribution doesn't really follow a normal distribution, but I'm not telling you anything...).

On the other hand, I'd like to display the "true price" and the true standard deviations (68/95/99.7%).

But I have no idea...

I use python with yfinance, plotly and streamlit.

Thanks for your help in advance : ) !

[zenith wing]

Any one has pc recommendation for data analytics??
I'm thinking ryzen 5 7600 and 6750xt with 32gb ram
Please advice my current amd laptop is quiet old and has 4gb ram which is hardly available half the time so I saved some money to buy a good pc to last few years

[stark ermine]

Your planned setup with a Ryzen 5 7600 CPU, Radeon RX 6750XT GPU, and 32GB of RAM seems quite decent for handling data analytics tasks.

[zenith wing]

Alright just one question should I go with 13500 in intel but i also prioritize power consumption because cant pay too high electricity bills right

[buoyant vine]

If power is something you care about

I would not use Intel, AMD in general has much better power efficiency per compute value in the newer chips in general

Also, in general, unless you're using 100% of the CPU all the time, your power bill is probably not going to be noticably different regardless of what CPU you go with

other than maybe if you put a Threadripper or server CPU in it 😅

If you really care about the power efficiency per compute and can afford it, the Ryzen x3D chips are incredible CPUs and have by far some of the best efficiency of the modern CPUs on the market rn.

But again, I don't think you will notice much difference in the power bill department

[wooden sail]

do note that you'll have to fiddle with your BLAS backend if you use AMD and want to do computations on cpu

and that AMD gpu's are not super well supported for any gpu computations

sadly history has favored intel and nvidia greatly in the area of scientific computing

check whether your target modules support ROCm instead of cuda, and read around about MKL vs openBLAS vs BLIS

[lapis sequoia]

Hello, anyone have experience with Python and AI/Machine Learning etc? or good and woorking tutorials? for prediction stock price. I have data in CSV

[hollow sentinel]

import pandas as pd 
import numpy as np 
import warnings
from datetime import datetime
import matplotlib.pyplot as plt 
import matplotlib.dates as dates

heart_df = pd.read_csv("/Users//Desktop/Apple Watch Data/HKQuantityTypeIdentifierHeartRate.csv")
pd.set_option('display.max_rows', None)
pd.set_option('display.max_columns', None)
print(heart_df.head(5))


heart_df["creationdate"] = pd.to_datetime(heart_df["creationdate"], format = "%m/%d/%y %H:%M")

heart_df["startdate"] = pd.to_datetime(heart_df["startdate"], format = "%m/%d/%y %H:%M")

heart_df["enddate"] = pd.to_datetime(heart_df["enddate"], format= "%m/%d/%y %H:%M")
print(heart_df.dtypes)


plt.plot(heart_df["startdate"], heart_df["value"], linestyle = "dotted")
# Add title and axis labels
plt.title('Time Series Plot')
plt.xlabel('Time')
plt.ylabel('Heart Rate')
plt.xticks(rotation=45)
plt.show()

is there anything else i can do with this visualization?

bc i can't really make sense of this, there's no underlying trend here

maybe combine it with something else?

[final kiln]

    def forward(self, in_sequence_bwc: Tensor) -> Tensor:

        batch, words, coordinates = in_sequence_bwc.size()
        pre_metric_tensors_nkk = self.pre_metric_tensors_nkk * self.MASK_11ww[0, :, self.K_DIMENSION, self.K_DIMENSION]
        metric_tensors_nkk = pre_metric_tensors_nkk @ pre_metric_tensors_nkk.transpose(-1, -2)  # ensures symmetry and positive definiteness

        all_projections_bwc = self.projections_cc(in_sequence_bwc)

        all_projections_bnwk = all_projections_bwc.view(batch, words, self.NUMBER_OF_HEADS, self.K_DIMENSION).transpose(1, 2)
        # all_projections_bnwk = F.normalize(all_projections_bnwk, p=2, dim=-1)

        all_dot_products_bnww = all_projections_bnwk @ metric_tensors_nkk @ all_projections_bnwk.transpose(-1, -2)
        all_dot_products_bnww = all_dot_products_bnww / math.sqrt(self.K_DIMENSION)
        all_dot_products_bnww = all_dot_products_bnww.masked_fill(self.MASK_11ww[:,:,:words,:words] == 0, float('-inf'))
        all_dot_products_bnww = F.softmax(all_dot_products_bnww, dim=-1)

        # all_dot_products_bnww = all_dot_products_bnww * self.MASK_11ww[:,:,:words,:words]

        nudged_vectors_bnwk = all_dot_products_bnww @ all_projections_bnwk
        nudged_vectors_bwnk = nudged_vec
tors_bnwk.transpose(1, 2).contiguous()
        nudged_vectors_bwc = nudged_vectors_bwnk.view(batch, words, coordinates)

        out_sequence_bwc = self.mixer_cc(nudged_vectors_bwc)

        return out_sequence_bwc

don't know what's up with that softmax stuff, but unlike Q, K, V, it seems to be essential. I've trimmed down Q, K, V to a lower triangular matrix (pre_metric_tensors_nkk, of which only half get updated during training ) and a projection matrix (projections_cc). results so far are very similar to Q, K, V which result from Wq, Qk and Wv, which total a larger number of params

[potent pollen]

Is there a problem to give some negative values in the input layer of a RNN ? I give a distance to the ai, and i tell me maybe the fact that it's negative will slow down the learning process, since I'm using ReLU for the hidden layers. Should I modify my way of counting the distance ?

[agile cobalt]

strictly speaking "distance" is always a positive metric, you probably should be abs()'ing it anyway

[agile cobalt]

but there is no problem in giving negative inputs, and in some cases you are even recommended to transform positive inputs into negative inputs as part of normalization - see https://datascience.stackexchange.com/questions/54296/should-input-images-be-normalized-to-1-to-1-or-0-to-1 for example

[potent pollen]

Ok thanks !

[final kiln]

this is the loss graph from the new transformer variant, which I'm calling metric tensor network (mtn), maybe I'm biased but it looks exactly like the transformers loss graph that I've been sharing here so far

the real advantage will come from the fact that I can double the amount of attention heads (since the metric tensor is symmetrical, half the space is being wasted rn) and still have less parameters than the transformer that produces this

[frigid owl]

Im looking for a tool to visuilaze training results. For now im just using matplotlib but is there any better libraries to get the job done?

[final kiln]

I've been using MLFlow, and it has been a blessing

[frigid owl]

Thanks a lot

[devout creek]

someone can help me with this ? : https://colab.research.google.com/drive/10g0pY3vv-mBu5t9yVJ2dQADoI-luOPa9

[frigid owl]

Cant get in, No access

[devout creek]

i want to output in srt format, someone know how to do this ?

[devout creek]

what ?

[frigid owl]

your collab notebook is private

make it public

[devout creek]

okay

https://colab.research.google.com/drive/10g0pY3vv-mBu5t9yVJ2dQADoI-luOPa9?usp=sharing

now it's public

@frigid owl

[frigid owl]

👍

[devout creek]

so can you help me or no ?

[frigid owl]

i dont really understand the problem

can you explain again please

[devout creek]

do you know what is insanely fast whisper ?

[frigid owl]

nah

[devout creek]

do you know what is whisper ?

[frigid owl]

yes

speech recognition model

[devout creek]

so insanely fast whisper is way way faster

but i don't know how to make srt file output instead of json file

[frigid owl]

have no idea how to do it

[devout creek]

okay

[frigid owl]

maybe go ask in #1035199133436354600

[devout creek]

no one knows here

[past meteor]

New(er) versions of matplotlib have a "plot training curve thing"

[frigid owl]

huh

can please send a link to documentation

i looked it up but didnt find anything

ik that sklearn has this but never seen something like that in matplotlib

[past meteor]

Oh sorry, I meant newer versions of sklearn

[faint cape]

Hi

I'm a junior python developer that do web scraping and data analysis

I'm new

[dusty valve]

hello, i wrote a mandelbrot zoom code and it pretty much zooms in anywhere.
The issue is, as it zooms it requires more and more iterations to get a clear image (as is intended). but i dont know exactly how many are required. so i just take perform iterations=frame_n*2.
This works fairly well, but is extremely slow as the video progresses.
Is there any formula or way to approximate the iterations required as i know the zoom in speed and point?

when it turns all white, the iterations are not high enough to get a clear frame. and later, it fuzzes out

[grave ledge]

I have a task at hand. I have to write a python script to build a log parser into JSON Format. The log files are taken from the MACBook. I have to ultimately feed it to an LLM model so that it can detect the issues from the log files and summarise it.
Can anyone point me to a good resource or help me understand how can I build a good log parser?

[final phoenix]

Hi
Please help me to extract 30 secs time interval from a column called created_time (ex 08:00:00 format) in python.

[indigo moth]

Ohhhhh I see ! Thanks a lot sir.

[woeful wren]

how do i plot the graph of a function that i calculated

[serene scaffold]

Hello, please show your code as text

!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.

[woeful wren]

i have these functions

fig    = plt.figure(figsize=(6, 6))
ax     = fig.add_subplot()

xx     = np.linspace(-5., 10., 1000)
# ------- Vul verder aan ------- 
vgl = opl.subs(a, 1/2)
vgl1 = vgl.subs(b, 2)
vgl2 = vgl.subs(b,4)
display(vgl1)
display(vgl2)

plt.show()

oh thats the wrong thing elt me fix

ye so this is my code so far

[serene scaffold]

and this is with sympy?

[woeful wren]

uh mathplotlib and sympy

the exam on monday is a sympy exam and this is a thing they gave us to prep

ax.plot(x, vgl1)

this is what i gotta use i think but it doesnt work

this is the error i get if i use thatt code

[serene scaffold]

import sympy as sp
x = symbols('x')
sp.plotting.plot(sp.cos(x), (x, -10, 10))

This worked for me

from your example, I don't know what opl is, so I can't make sense of what happens after that

[woeful wren]

ye alright but there is given code

sorry i forgot to mention the first part of the code is given

fig    = plt.figure(figsize=(6, 6))
ax     = fig.add_subplot()

xx     = np.linspace(-5., 10., 1000)
# ------- Vul verder aan ------- 

this is givenand the plt.show as well

[serene scaffold]

that's fine, but you don't define opl here, and then you use it in the next line. good examples have every variable defined, or follow common conventions for variable naming

(and maybe "opl" is a common name for something, but idk what)

[tidal bough]

either that or evaluate the function on some points and then plot them normally:

import sympy as sp, numpy as np
x = sp.symbols('x')
f = sp.cos(x)

X = np.linspace(-10,10,1000)
y = sp.lambdify(x,f)(X)

# plt.plot(X,y) or whatever

[woeful wren]

vgl = opl.subs(a, 1/2)
vgl1 = vgl.subs(b, 2)
vgl2 = vgl.subs(b,4)
display(vgl1)
display(vgl2)

all this pat of the code deos is calculate the fucntions

[woeful wren]

these functions

opl stands for oplossing which is dutch for solution and opl was the solution of my differential equation and then i sub a,b in the equation with the values that they tell us to sub

[woeful wren]

yo this works thankss

[soft lantern]

python for data analysis, a good book to start with?

[serene scaffold]

"data science from scratch" second edition

[soft lantern]

im not really a beginer, ive learned numpy and pandas on free code camp, but thats hasnt equiped me with problem solving

[serene scaffold]

then you can skip a few chapters

[soft lantern]

you still recomment that book?

[serene scaffold]

ya

[soft lantern]

um what makes you suggest that very book

[serene scaffold]

I've read it and it's a good overview of the space and doesn't have shitty code examples

[soft lantern]

ty admin

[dusty valve]

Bruh spy cam plot too?

sympy*

mpl still top tjo

[final kiln]

started drafting an explanation thing, if I get good results I'll expand it into a paper

[valid crow]

I'm relatively new to this field, but I'm eager to learn about data science and AI. If any of you have recommendations for the best resources to study data science and AI, I would greatly appreciate it if you could kindly share those details with me.

[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.

[dense yarrow]

how can i interpret this graph? the total is a column that combines the original "drug_recode" column with "drugs_imputated" column

the original drug recode column has 1 for often/sometimes use of drugs. 0 for never use of drugs

[tidal bough]

imputation increases drug use? 🥴

[dense yarrow]

does this mean imputating is not necessary for this dataset?

[tidal bough]

i feel like it's bad (suggests problems with imputation) if imputed points are so different from the real ones, but unsure.

[dense yarrow]

it seems that it increases the value in the drug_recode column

[dense yarrow]

i was thinking so too, but i wanted to get opinions from those more familiar with the topic

this is my first time working with imputation

[dense yarrow]

this is what i ended up writing:This graph shows the effect of imputation on the drug_recode value. As we see, the imputated values are very different than the reported or no imputatation values. The combined values better fit the reported values. As next steps, we would revisit the imputation method for significant error or bias as well as investigate how the structure of the data contributed to the accuracy of the imputation.

[tidal bough]

The combined values better fit the reported values
i wouldn't write this because this is literally always true

an average of datasets X and Y will always be more similar to X than Y is

[dense yarrow]

okay!

thanks!

[half mountain]

Hi, I want to use a machine learning model to predict future performance of players. I am using Quarterback stats from the NFL. The stats are week to week and player by player. I wanted to use past games to determine future games of a player. I decided to use a Random Forest Model. I will try to predict touchdowns using Features like Pass_YDs, Interceptions, and Pass Attempts.

What I am doing does not seem right. I will never have my Features (Pass_YDs, Interceptions, and Pass Attempts) before a game is played so I cannot predict Pass Touchdowns with those. Features seem to me like ideas I know before a game is played like opponent, Home or Away game, etc. What I am trying to do is predict a players future performance in games based on past performance. Can you help with ideas on how I would do this and if I am on the right track?

Thanks! 🙂

[agile owl]

i would probably regress each statistic using a model that ensures positivity on a vector of the past statistics of the player, their team, and the opposing team's defense

like the past 5 games or something

then i would fit that across quarterbacks in general so you have enough data

there's a lot of options for how to do the regression

I think with things like sports though the variance is huge

hard to model all the factors

[half mountain]

Okay, that should get me in the right direction. It is just an early project for a portfolio I am putting together. I am a data analyst so just practicing.

Thank you for the input!

[agile owl]

np

where are you getting the data?

trading usdjpy out of sample

hows shakespeare coming along

[half mountain]

ESPN

[agile owl]

do they have a way to download it easily or do you have to pull it from their summaries

[half mountain]

@agile owl I pull it from their summaries

[pale hemlock]

anyone interested in discussing the topic of conversation?

[agile owl]

trading the yield curve

trading the 10y treasury

add these all toegether and the sharpe ratio is well in excess of 2

I need to write the variance weighting code I've just been doing it by hand on the renders in excel

[barren fable]

I learned lots of different models in regression and classification in machine learning, like linear, polynomial, and svr for regression and logistic, svm, and kernel svm for classification. But all of these were intuitive explanations without, for instance, explaining the Gradient Descent and Convergence Algorithm, Ridge, Laso, or the math behind every model. So I don't want to dive into deep learning with our understandable machine. So first, what's your opinion? Should I dive more into machine learning or am I able to go to deep learning? Second, some people recommended to me some playlists for machine learning.

Krish Naik: https://www.youtube.com/watch?v=kEmnkUw0NTs&list=PLZoTAELRMXVPMbdMTjwolBI0cJcvASePD&index=3

Andrew NG: https://www.youtube.com/watch?v=vStJoetOxJg&list=PLkDaE6sCZn6FNC6YRfRQc_FbeQrF8BwGI

and some more, so is there a good resource explaining the models in depth in a simple way, like code implementation and projects? Thanks!

[wispy thorn]

Hi everyone, I'm trying to install an older version of matplotlib, i.e < 3.3.0 but everytime I try to install it, it's not able to gather the requirements to build the wheel

Anyone got any heads up as how this problem will be resolved?

[final kiln]

You probably need to update python to either an earlier or later version.

[vapid garden]

How can I generate classification dataset using either gaussian distribution or Poisson distribution.

...?

[past meteor]

You can do this using Numpy and Pandas

I've done this a couple of times in the past. Whenever I volunteer to give a data science training I make my own datasets and I do it with Numpy. I pick a case that seems interesting to the audience, define some variables and think about how they're generated (e.g., maybe age is gamma distributed, maybe the relationship between age and cost is a multimodal gaussian etc)

Maybe there's tools that do this in a more automated fashion but I do it manually 🙂

I don't respond to DMs sorry

[vapid garden]

Like a dataset if 10k rows each row is only Poisson distribution, can be any number of features

Can u give me code on how to do it

[past meteor]

I'm not a fan of giving you the full solution because then you learn the least

https://numpy.org/doc/stable/reference/random/generated/numpy.random.Generator.poisson.html <---- with this you variables following a poisson distribution

[desert oar]

think about how classification models work and do that in reverse: construct a continuous response/output/label, and convert that to discrete classes

[final kiln]

the larger scale test is taking a while, I'm almost done refactoring the repo for it. I'm going to search through these guys

https://paperswithcode.com/method/strided-attention
https://paperswithcode.com/method/fixed-factorized-attention
https://paperswithcode.com/method/dot-product-attention
https://paperswithcode.com/method/scaled

to try to see how I can evaluate the performance of my architecture, at least two of them seem to benchmark on language translation tasks, so I might have to adapt for that, the dot product attention seems to be where the 2017 paper ended coming from, I only read it for a bit, they don't seem to be using a lot of gpu which is a relief

ig my thing fits in as a generalization of the scaled dot product attention, or a simplification ig

[elfin swan]

bgf = BetaGeoFitter(penalizer_coef=0.001)
​
bgf.fit(rfm['frequency'],
rfm['recency_weekly_p'],
rfm['T_weekly'])

ConvergenceError:
The model did not converge. Try adding a larger penalizer to see if that helps convergence.

Please help me with this, I am getting this error

[serene scaffold]

can you show the import statement for BetaGeoFitter?

[heady sierra]

Hello. I am training a reinforcement learning model using stablebaselines3 but every time I train in vscode, I am using my cpu even though I have a rtx 4060. How can I make it use my gpu instead?

[twilit dove]

Hi guys! I've recently started my journey into data science and machine learning in general and one tip for exploring different applications I keep hearing is to read research papers and attempt to replicate the models created in the papers. Are there any websites/journals that ML researchers generally publish their papers on, or are ML research papers on more generic science journals. Thanks

[serene scaffold]

Can you show some code?

[serene scaffold]

I actually don't recommend that. Papers are about very specific developments, and are challenging to read, even for experienced practitioners. You won't build foundational knowledge in DS/ML by reading papers.

[heady sierra]

Its in vscode notebook so which part do you want to see?

[serene scaffold]

the part where you create the model. By the way, I won't look at screenshots.

[heady sierra]

class MyCustomEnv(StocksEnv):
    _process_data = signals

env2 = MyCustomEnv(df= data, window_size= 15, frame_bound=(15, 90))

env_maker = lambda: env2

env = DummyVecEnv([env_maker])

model = DQN("MlpPolicy", env, verbose=1)
model.learn(total_timesteps=500000)

while True:
    obs = obs[np.newaxis, ...]
    action, _states = model.predict(obs)
    obs, rewards, terminated, truncated, info = env.step(int(action))

    done = terminated or truncated
    if done:
        print("info", info)
        break

This is the output:

Using cpu device

| rollout/ | |
| exploration_rate | 0.993 |
| time/ | |
| episodes | 4 |
| fps | 16755 |
| time_elapsed | 0 |
| total_timesteps | 348 |

---

| rollout/ | |
| exploration_rate | 0.987 |
| time/ | |
| episodes | 8 |
| fps | 16891 |
| time_elapsed | 0 |
| total_timesteps | 696 |

---

| rollout/ | |
| exploration_rate | 0.98 |
| time/ | |
| episodes | 12 |
| fps | 16301 |
...
| learning_rate | 0.0001 |
| loss | 1.3e+05 |
| n_updates | 112431 |

[serene scaffold]

Can you add the import statements for everything used here?
Also, please use python highlighting.
```py

[twilit dove]

hmm ok then. Thanks!

[heady sierra]

import gymnasium as gym
import gym_anytrading

from stable_baselines3.common.vec_env import DummyVecEnv
from stable_baselines3 import DQN

import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import tensorflow as tf
import torch

class MyCustomEnv(StocksEnv):
    _process_data = signals

env2 = MyCustomEnv(df= data, window_size= 15, frame_bound=(15, 90))

env_maker = lambda: env2

env = DummyVecEnv([env_maker])

model = DQN("MlpPolicy", env, verbose=1)
model.learn(total_timesteps=500000)

while True:
    obs = obs[np.newaxis, ...]
    action, _states = model.predict(obs)
    obs, rewards, terminated, truncated, info = env.step(int(action))

    done = terminated or truncated
    if done:
        print("info", info)
        break

[serene scaffold]

modify these two lines accordingly

# before
model = DQN("MlpPolicy", env, verbose=1)
model.learn(total_timesteps=500000)

# after
cuda = torch.device('cuda')
model = DQN("MlpPolicy", env, verbose=1)
model.to(cuda)
model.learn(total_timesteps=500000)

then see if that moves it to the GPU

[elfin swan]

import datetime as dt
import pandas as pd
import matplotlib.pyplot as plt
from lifetimes import BetaGeoFitter
from lifetimes import GammaGammaFitter
from lifetimes.plotting import plot_period_transactions

[serene scaffold]

The short answer is that that model can't learn your rfm data with the penalizer_coef hyperparameter that you set.

btw @heady sierra, you might need to redefine obs as a tensor (it appears to be an array currently) and move that to the GPU as well.

[heady sierra]

ok got it. Thanks

[desert oar]

yeah if anything your technique is a specialization of it

great that you're following up on it. i'm looking forward to the benchmark results

[final kiln]

Yes indeed, it's a special case of it

[final kiln]

As long as I keep getting good results I'll keep digging

I've experimented with several designs now, and it seems that the network doesn't really care what you do as long as you calculate scores.

The exciting thing is the parameter reduction

And the design philosophy of forcing the network to do geometry to make decisions

I checked the hardware used ok the 2017 paper and I'll be able to reproduce a good chunk of the their table

These scores are already from the new thing

And its output is similar to the transformer

The meaning of life is to shake others.
With old only make an oath shallker thoughts; and says so defend:
but the fault out of thick-morrow, to ruin;
And holy clergymen must be needful
The younger slanders to be more than the hollow service
Known and defars; crave heaven,
Even as it would fear'd keep the time
Of love can admitges change;
So much lenity of soldiers,
Then thieves conn'd poison'd blanks,
Cry

So at least on a very small dataset it works fine and on par w/ the transformer

[agile owl]

Not sure if that means anything or not

lol

[final kiln]

Me either, I just ask gpt if it looks coherent

[agile owl]

I don't think shallker is a word?

then again I think shakespeare invented words so hey

[final kiln]

Wait

Yes it exists, it's an archaic word

[agile owl]

what about thick-morrow

[final kiln]

Ah, no it doesn't

Gpt hallucinated

[desert oar]

how do those compare to the attention matrices from the original?

not that i'd expect all 8 heads to find the same thing

but maybe we'd expect their sum/average to evaluate to something similar

or i suppose just comparing the final output distribution over tokens (rather than individual tokens that were chosen)

[final kiln]

I didn't graph the other scores, but I can do it tomorrow, same code practically

[desert oar]

that is, does your model generate quantitatively similar distributions over tokens (e.g. sum of squared differences akin to brier score)?

your projects have been very inspiring!

[final kiln]

But the point is that the output is the same non sense as the transformer

[final kiln]

That's a good idea actually

[desert oar]

i'm also really curious if this works just as well on bert-like models as it does on gpt-like models

[final kiln]

So like, feeding a token and directly comparing the output ?

[desert oar]

yeah, but instead of reducing the output to a token or sequence of tokens, leave it as a (sequence of) probability distribution(s) over tokens

[final kiln]

Right, ig it would make sense for them to be similar since they're both approximating the same probability law

[desert oar]

yeah hopefully

although you need to collapse it down to a single token to generate more than one token... so maybe just start with a single token prediction like you said

maybe walk through an existing valid english-language document and compare the output at each token?

that way you know the inputs make sense

averaging a bunch of one-step-ahead forecasts rather than averaging an entire forecasting procedure, if that makes any sense

[final kiln]

Yes it's a good idea. I'm gonna take note. Right now I'm preparing the repo for a series of larger scale and more systematic experiments. After that I'll go through a data analysis phase where I do these comparisons

[final kiln]

I didn't know bert was different

[desert oar]

i'm sure there are other details involved, but my understanding of the main difference is that bert doesn't do masking, so it's considered an "encoder-only" model compared to the gpt "decoder-only" design

that is, in bert every token can attend to every other token in the sequence without restriction, whereas in gpt tokens can only attend to tokens earlier in the sequence (which is why the upper triangle of the attention matrix in your output is all 0)

https://datascience.stackexchange.com/a/65242/1156

[agile owl]

what are the best models for sentiment extraction these days

I'm thinking about adding a new input stream to my reinforcement learner based on sentiment analysis

[desert oar]

interestingly it looks like microsoft was attempting to use a bert-like architecture for language generation, which i guess turned out to be a dead end since i've never heard anything about it (and gpt of course turned out to be a tremendous success) https://arxiv.org/abs/1905.02450

[final kiln]

Wait so you literally just don't mask it and have a special char marking a word as empty ?

The output is generated non-autoregressively (every token at the output is computed at the same time, without any self-attention mask), conditioning on the non-masked tokens, which are present in the same input sequence as the masked tokens.

So this is why it's so much faster

[shadow viper]

Good day everyone, is there a quicker way to train a model?
it takes too much time and the power outage affects it a lot

[proud wing]

@final kiln Thats a method... masking versus padding

It's just not supported by all the different tuning approaches out there.

Just finished one of the stages for my model ocr code dataset training-generator.

Model image detection:
[1] CPLUZPLUZ.png
[2] pyth0n.png
[3] ruzt.png
[4] c55.png
Auto-Selecting *.png
Auto-preprocess: (yes/no): yes
File Save Strategy: (txt/pdf/auto): auto
CPLUZPLUZ.png saved as CPLUZPLUZ-OCR.cpp (Cpp detected)
pyth0n.png saved as pyth0n-OCR.py (Py detected)
ruzt.png saved as ruzt-OCR.rs (Rs detected)
c55.png saved as c55-OCR.css (Css detected)

[slow vigil]

Anyone here use Polars? I'm having a weird issue. I have a program that is using multi-processing and threading in each process, and I'm trying to intermittently write a dataframe to a csv for each process once the dataframe reaches a certain height. The height check gets triggered and by printing the dataframe just before the call to df.write_csv(filename) I can see that the dataframe has data in it, but when I look in the file that gets written out it is only writing out the headers of the dataframe and doesn't contain any data

[past meteor]

Can I ask why you've structured your program like that?

[slow vigil]

There's a lot of data being gathered and I wanted to safeguard it in case the program fails at any point

I was being CPU bottlenecked so I set it up to run in multiple processes

Also, by flushing the dataframes periodically it keeps the in-mem size down

[past meteor]

Personally seems like a strange way to structure it, I think if you did it in a more orthodox way it wouldn't be as big of an issue

[slow vigil]

I would if I could

[past meteor]

Especially since Polars lets you stream data from 1 source to another

[slow vigil]

I had to do it this way

[past meteor]

So you can effectively work with larger-than-memory datasets

[slow vigil]

It was going to take 30 hours to run without multiprocessing

now it takes 4

[past meteor]

Polars uses multiple cores by default

[slow vigil]

Please just trust me lol

[proud wing]

Have you tried inspecting the dataframe prior to it reaching the height?

[past meteor]

I don't know what your specific issue is, you could have a race condition somewhere

[slow vigil]

I print the dataframe right before trying to write and it is normal. Full of data

[proud wing]

It sounds to me like the way you are writing the data is the issue

[past meteor]

Seems likely since you have multiple processes trying to write to the same file

[proud wing]

not the dataframe itself, since its printing without any issue

You might have a race condition if youre trying to write to the same file

You can use a writequeue

to ensure they dont compete for writing

[slow vigil]

if main_df.height >= 10:
  if os.path.exists(csv_name):
      existing = pl.read_csv(csv_name)
      main_df = pl.concat([existing, main_df])
  print(main_df)
  main_df.write_csv(csv_name)
  main_df = main_df.clear()

I used a height of 10 here just for testing purposes

But yeah it might be because of the threading

Maybe I'll just have to make it with bulletproof error handling and then return the df from each process

and do the write after it returns

pretty inconvenient though tbh

oh

[proud wing]

import os
import pandas as pd
import threading
import random
import time

def generate_test_data(thread_num):
    return pd.DataFrame({'Thread': [thread_num], 'Height': [random.randint(1, 15)]})

def process_and_write(main_df, csv_name, thread_num):
    with write_lock:
        if os.path.exists(csv_name):
            existing = pd.read_csv(csv_name)
            main_df = pd.concat([existing, main_df])
        print(f"Thread {thread_num} - Data put in write queue: {main_df}")
        main_df.to_csv(csv_name, mode='a', index=False, header=not os.path.exists(csv_name))
        print(f"Thread {thread_num} - Data written to the file: {main_df}")
        main_df = main_df.iloc[0:0]

def data_generation_thread(csv_name, thread_num):
    while True:
        main_df = generate_test_data(thread_num)
        if main_df.iloc[0]['Height'] >= 10:
            process_and_write(main_df, csv_name, thread_num)
        time.sleep(1)  


write_lock = threading.Lock()
num_threads = 3
threads = []
csv_name = 'output.csv'

for i in range(num_threads):
    thread = threading.Thread(target=data_generation_thread, args=(csv_name, i+1))
    threads.append(thread)
    thread.start()

for thread in threads:
    thread.join()

hope that helps.

[slow vigil]

hmmmm write lock

[past meteor]

Can you try writing to different files and just concatenating at the end?

[proud wing]

try that script ^

it will do what you basically described.

[slow vigil]

I'm writing to a separate csv for each process. I did plan to combine them at the end

[proud wing]

example output from it:

2,15
3,11
1,11
1,11
2,15
1,11
1,11
2,15
3,11
1,14
2,12
3,11
2,15
1,14
3,14

I have it writing the threadnumber to each line

[slow vigil]

I'm not using the threading library though

If I import it just for write lock will it still work?

[proud wing]

Its your script, test it:)

[slow vigil]

lol true

Another thing I'm thinking is I could use a multiprocessing Value object

idk if I can pass a dataframe into that lol

[past meteor]

Is this generated by AI?

I don't immediately see where/how this is failing but all I can say is that you're fighting against the API of polars

I would really really consider using the library as intended with, pl.source_x with x being your source and pl.sink_x to lazily read and write

[slow vigil]

Each process has 10 threads running at a time, so it's probably the case that somehow that is causing an issue

[past meteor]

Unless there's a very very specific reason why that is not possible

[slow vigil]

I will look into those. I'm new to polars so I'm not familiar with the workflow yet. I'm basically using it philosophically the same as Pandas, which I'm sure is wrong

[past meteor]

Yeah they're very different libraries. They only look similar on the surface

In the past I've reduced a workflow that took 1+ hour to run under <1 min using polars 🤷

[slow vigil]

Sheeesh

[past meteor]

This is due to various reasons, not just the libraries being different. (Pandas consumes a lot of memory so I had to batch my results, there was a lot of overhead of doing DB calls). All I'm saying is: read the documentation first and it'll pay off.

[slow vigil]

I caught a co-worker using nested calls to iterrows in Pandas the other day lol

I was like, "Nuh uh"

[frail arch]

so, I am trying to use standard baselines for some RL. Just learning.

env_name = "CartPole-v0"
env = gym.make(env_name)
env = DummyVecEnv([lambda: env])
model = PPO("MlpPolicy", env, verbose=1, tensorboard_log=log_path)
model.learn(total_timesteps=20000)

the last line is giving the following exception:

75 for env_idx in range(self.num_envs):
76 maybe_options = {"options": self._options[env_idx]} if self._options[env_idx] else {}
---> 77 obs, self.reset_infos[env_idx] = self.envs[env_idx].reset(seed=self._seeds[env_idx], **maybe_options)
78 self._save_obs(env_idx, obs)
79 # Seeds and options are only used once
ValueError: too many values to unpack (expected 2)
Any idea what I am doing wrong?

[desert oar]

i actually had a legitimate use case for nested itertuples in a project last year

never iterrows though. just use range and iloc for that

[slow vigil]

It appears that trying to write out from the process pool is causing each process to hang