Much Faster Pandas with cuDF GPU Processing - CPU vs GPU Speed Benchmarks
HTML-код
- Опубликовано: 12 июн 2024
- What if I told you that all this time we've been using Pandas wrong? 🐼 🐼 🐼
We keep running it on our CPU and wondering why it's slow - but what happens when we switch to GPU processing? 🤔
In this tutorial we will explore the brand new technology behind cuDF Pandas Accelerator Mode that allows us to use our graphic cards to make Pandas MUCH faster! We will:
1️⃣ 1. Install RAPIDS cuDF via Windows Subsystem for Linux @ 01:58 - 05:03
2️⃣ 2. Code a Data Science Workflow with cuDF Pandas & Sentiment140 @ 05:03 - 11:28
3️⃣ 3. Learn Basic Feature Engineering with Regex @ 11:28 - 15:12
4️⃣ 4. Conduct a CPU Pandas versus GPU cuDF Pandas Speed Test @ 15:12 - 19:08
Throughout the tutorial we will talk about the relationship between processors, memory and graphics cards with lots of colorful visualizations and examples. We will see how our software processes manifest on the hardware level and explore the effects of GPU parallel programming in the realm of sentiment analysis.
⭐ More cuDF Pandas Resources ⭐
---------------------------------------------------------------
📝 Official cuDF Pandas Colab Notebook Code Example (beginner friendly):
colab.research.google.com/git...
💻 cuDF Pandas Virtual Summit Sessions:
nvda.ws/3tkJLF0
🎥 Related Videos of Mine 🎥
---------------------------------------------------------------
⭐ CUDA Simply Explained - GPU vs CPU Parallel Computing:
• CUDA Simply Explained ...
⭐ Basic Guide to Pandas - Tricks, Shortcuts, Must Know Commands:
• Basic Guide to Pandas!...
⭐ FASTER Inference with Torch TensorRT - CPU vs CUDA:
• FASTER Inference with ...
⭐ Anaconda Beginners Guide for Linux and Windows:
• Anaconda Beginners Gui...
💻 Installation and Download Links 💻
---------------------------------------------------------------
⭐ cuDF Installation Guide @ 2:59:
docs.rapids.ai/install#conda
⭐ Sentiment140 Homepage @ 5:50:
help.sentiment140.com
🛑 PLEASE NOTE: Sentiment140 homepage is currently down (as to December 13th 2023). Therefore you won't be able to copy the URL to the zip file from there. I am not allowed to share it with you directly, so temporarily - please manually type the same URL you see in the video @ 5:55 (it is displayed in full! starts with cs.stanford.edu/)
Once the homepage is back - you'll be able to copy the link from their student section.
⏰ TIMESTAMPS ⏰
---------------------------------------------------------------
00:00 - 00:43 | intro
00:43 - 01:10 | what is CUDA?
01:10 - 02:25 | install WSL (Windows Subsystem for Linux)
02:25 - 03:42 | install Anaconda in WSL
03:42 - 05:03 | Install RAPIDS cuDF
05:03 - 05:30 | what is Sentiment Analysis?
05:30 - 07:12 | download and unzip Sentiment140 with code
07:12 - 07:51 | import cuDF Pandas Accelerator Mode
07:51 - 09:46 | loading and processing operations on GPU
09:46 - 10:24 | GPU and CPU commands profiling
10:24 - 10:40 | why do we need CPUs if GPUs runs faster?
10:40 - 11:28 | cuDF CPU fallback
11:28 - 13:27 | feature extraction with Regex patterns (Regular Expressions)
13:27 - 15:11 | feature reduction
15:11 - 19:09 | CPU Pandas versus GPU cuDF Pandas Speed Test
19:09 - 19:35 | Challenge
19:35 - 19:57 | Thanks for watching!
🤝 Connect with me 🤝
----------------------------------------------------------------
🔗 Github:
github.com/mariyasha
🔗 Discord:
/ discord
🔗 LinkedIn:
/ mariyasha888
🔗 Twitter:
/ mariyasha888
🔗 Blog:
www.pythonsimplified.org
💳 Credits 💳
----------------------------------------------------------------
⭐ Beautiful titles, transitions, sound FX:
mixkit.co
⭐ Beautiful icons:
flaticon.com
⭐ Beautiful graphics:
freepik.com
#python #pythonprogramming #machinelearning #pandas #pythonpandas #pythonpattern #pattern #regex #regularexpression #gpu #cpu #processor #graphicscard #graphiccard #hardware #computerhardware #encoding #benchmark #benchmarks #cuda #rapids #nvidia #artificialintelligence #datascience #programming #coding #neuralnetworks #ml #ai #technology #computer #computerscience #data #dataanalytics #datastructures #gpucomputing #multiprocessing #rtx #rtx4080 #jupyterlab #sentimentanalysis #featureengineering #database #datasets #twitter 
Наука
My WSL is set up as Debian, and I have Python 3.11 installed on it, so I wasn't able to replicate this. Apparently 3.11 isn't supported yet 😢.
I will try some other time to set up a second WSL with Ubuntu and see what happens.
I have a Threadripper Pro 3945WX with 256GB RAM, and an RTX 3080Ti, and generally speaking, CUDA is about 4x faster than CPU. Only having 12GB of VRAM is a major problem though. Typically the datasets I work with are around 100m rows.
I reduced the exectution time of one Pandas project from 10 minutes to 0.1 seconds by re-coding it in Numpy, and Polars gives me not quite so good speed increases for a lot less effort.
You're 100% right, Katrina! cuDF doesn't support Python 3.11 just yet, so it might be the reason why the installation fails. I don't think there's a problem with Debian compatability as Miniconda should take care of installation command discrepancies 🐍🐍🐍
This cuDF Pandas accelerator mode was only released a month ago, so I'm sure Python 3.11 support is on the way... but I'll double-check it with Nvidia just in case! 😉
For 100m rows, I think it's worth jumping back to Python 3.10. Losing 10 seconds here and 10 seconds there doesn't sound too bad - but it compounds to minutes and hours very quickly. Plus, I didn't have time to cover it in the video, but cuDF has plenty of functionality beyond what Pandas can offer! I wanted to end the tutorial with stripping punctuation + stopwords and stemming the words, but it would have made the tutorial way too long... 🙃
Essentially, what took me 7 minutes to do on my recent midterm project - was done in seconds with cuDF!!! It's insane!! For preprocessing tasks - I don't think I'd be utilizing my CPU ever again.
But I still need to check out Polars and see how it compares with cuDF.
Sounds like a great idea for a new tutorial! 🤪
Hi Katrina, I have some great news! Just spoke with Nvidia, and they're currently working on Python 3.11 integration! 🥳🥳🥳
For now Python 3.9 and 3.10 are fully supported, and they will announce when they are done with 3.11 (I think because 3.11 is a relatively new version, and many of us haven't switched yet).
Also, quick 1 months of usage update: I'm not working with traditional Pandas anymore (or even Windows). Fully switched to WSL and cuDF! Will arrange a nice live stream soon t6o catch up 😉
It's amazing, I started to work on it more than a decade ago. But now, everyone can do it in seconds. Thank you.
Hi Maria ,Love the way you teach and I can understand everything you teach.
Love the idea of moving processing onto the GPU!
Great video! ❤
i love your way of teaching, this video came just in time
Yeyy!! Thank you so much!! Super happy to hear! 😀😀😀
@@PythonSimplified you are gorgeous !
CPU R9 5900x - 1min 3 sec,
GPU RTX 3090 - 1.6 sec.
Furthermore, Pandas is single-threaded by default, but there is a "modin" project/library that allows to scale pandas so that it utilizes all CPU cores and threads, which helps especially well on large data sets and when a CUDA device is not available. To accomplish this, simply replace the import statement with "import modin.pandas as pd" in the script (previously installed the library, of course). For the task in the video, this reduced the execution time by more than half comparing to ordinary pandas using.
Thank you for the video!
You´re always amazing M, great video!!! 🤘😸
Very nice! It helped me a lot. Thank you!❤❤❤
I love you explanation i love how you simplify things
Thank you. It's very helpful to me.:)
Wow those helped thank you man🌺
Excellent tutorial as always!
My Alienware R11 has an i9 10th gen and only 16Gb RAM and a 1660ti. With regular pandas the kernel crashed on a couple of tries because it couldn't handle that amount of data (I think) but with cudf it did the labelling task on the 25 mil rows in 1min 35s.
You also have Aurora R11???? Me too!!! 🤩🤩🤩
My theory is that regular Pandas crashed because of the amount of RAM (I have the same CPU and same kind of RAM, but 32GB). You can always utilize cool techniques inside read_csv to avoid crashing:
1. Add a columns property and only read the columns you need.
2. Add a chunksize property and set it to some integer. Read_csv will then break the data into read-only segments in the size you specified. You can iterate over them and turn them into standalone DFs.
3. Don't worry about it, and keep running everything on GPU 🤪 hahaha
Amazing video :)
Dark Wolf??? Is that you????? How's it going?? 😃😃😃
Thks from Brazil! 😃
I really love ur vids!!!!
3:18 "We'll of course carefully read the license agreement" 😆
When I skip reading those - I always think of a South Park episode where Kyle failed to read the Apple license agreement 🤣🤣🤣 (I don't want to spoil it for those who didn't watch, but checkout the Human centiPad episode if you're curious)
@@PythonSimplified I'll check it out!
Dude, you finally discovered cuDF
It's Amazing!!! Very good!
Love you maria my best python teacher,
For me CPU : 9.64 s and GPU: 798ms
Thank you so much. I love your videos.
Thanks
Hi Maria.
Have you tried the Polars library? It is inspired by Pandas and works similarly but much faster because it is made by using Rust programming language.
Greetings from Ukraine.
I'll cover Polars on the channel soon, would love to test it against Pandas, cuDF Pandas and DuckDB 😉
Once and for all, we'll see which is better for which tasks! It's definitely on my to do list 😃
Cheers from Vancouver, Vasyl!! 🍁🍁🍁
I love your videos :)
It would be great to have the install workflows with pip and Python venv. Mixing use of Conda and pip sometimes beaks things. I have far less environment issues since stopped using Conda!
Could you also indicate the Python version it was tested on?
Absolutley! Python 3.9 and 3.10 are fully supported, and Nvidia is currently working on integrading 3.11 as well (they'll announce when it's good to go 😉)
I rarely use conda install anymore, switched to pip in terms of package distributions (I find that some of their packages are outdated), but I really like the way Anaconda manages virtual environments. It's super convenient to me, while VENV feels a bit unintuitive... but as usual, it's all a matter of personal preference 😃
Also, super happy you liked the video!! thank you so much for the lovely comment! 😁😁😁
Ohw Beautiful system 🤖🤯
Thanks for the content. I try to follow along. I set up the faster_pandas environment in WSL ( or do i need to install it in powershell). When i run the %load_ext cudf.pandas, i get an error message saying No module named 'cudf'. I wonder what did i do wrong?
Can you please make a video on flask appBuilder framework which is made over flask framework ?
Like the way you speak ❤
pls can you tell behinds background device use
So, I was attempting to follow this and the cuDF install cannot be installed without anaconda apparently.
A geek beautiful woman talking about very interesting topics. Like! 😀
Puking in my mouth ❤️ 🤢
HI, I am trying to download test data zip file but it looks it doesn't exist. any alternative?
Hi, we need a video about the streamlit framework. Regards
One month is too much to wait for the next video =///////
are you planning a set up tutorial of cudf for linux? been looking all over for one using cuda 12.3…
This one should work for Linux too, as I'm using WSL (well, it works with Ubuntu for sure, not sure about other distros... my Kali linux machine is no longer operational so I can't really test it).
And for CUDA 12.3 just select CUDA 12 in the RAPIDS installation guide, I had the same situation, and everything worked 😉
On Ubuntu 20.04 I failed to install cuDF using pip in venv, but when tried using Conda (similar steps as in the video) then it launched successfully.
Can you help me i want to code inteface with kivy of substraction
Please make video on creating openCv python stanalone applications.
I saw ur prev videos but they didn't work.
I always get error to include a file though there is nothing wrong.
Please help
Which specific video and which specific error? I can help, but I need to know the details to do so.
Still good to run it during training model like pytorch or tensorflow?
Yup! When you install RAPIDS and select cuDF in the "additional packages" section - please make sure you select pytorch and tensorflow as well, and you're good to go! 😉
I am currently doing university final project on ward to sign language translator app, i need a little guidance from you.
Thanks
Can i use this for password cracking?
hey Ms-Py! 😉
Mariya your beauty is a distraction for learners😂
I'm not able to replicate your code. Can you specify the version of everything that you have used mainly wsl's version.
thank you so much for this video
I'm using WSL 2 and RAPIDS 23.10.
What errors are you getting and at which stage of the tutorial?
Will Python 3.11 be supported soon?
❤
Good onne 🎉
Thank you!
Hi.., i am big fan you.here i have a small doubt?, can you make video with brief explain of pop up frame😢 example user profile is there in top right corner. I will click this action literly pop up a new window with borderless and without window title buttons.,
If you can't make video plz rply this comment for what i do this 😢
please make video on how to .py file of computer vision to .exe file. The normal python file is converted to .exe file but the py file of computer vision doesn't do!!!!!!! Please help
I watched your previous videos but they too didn't work
I want to deploy my kivy program into app but I haven't Linux system as you used it I have already tried many other ways but they all are not working it get crashed when I used Collab please make a detailed and separate video on this topic
I actually have a really nice tutorial that shows you how to do it with Windows!
⭐ Convert Python to Android with WINDOWS & LINUX + Fix Common Bugs:
ruclips.net/video/VsTaM057rdc/видео.html
It also uses WSL, just like we do in this cuDF Pandas tutorial. The Colab conversion way doesn't work for the past few years, but you don't need a full blown Linux Machine either. Just install WSL from your command prompt with wsl --install and you have Linux installed on your Windows system 😉
Good luck and I hope it helps! 😀
@@PythonSimplified thank you so much for replying. Let me check it...
Do you have any videos that teach making android apps from python. I have to make an app for a contest where judges will score contestants in their respective android devices and in the end consolidated result will be given can you please help
Great thanks. Where are you???
Wrapping my last university exams, just moved to the Fraser Valley where I have a much bigger office that I'm still yet to properly set up, and getting ready to fly to San Jose for GTC-2024 next week 🤯🤯🤯
In other words - I'm running around like a chicken without a head hahahaha but the good news is -> I wrote an incredible cloud computing tutorial (on CPU and GPU!) and will film it as soon as I'm back to Canada 😉 Hopefully you'll get to see it this month! fingers crossed 🤞
Lets Roll 👉👉
This time I was about to end the intro with 30 **zoom to 105%**. times **zoom to 108%**. faster! **zoom to 111%**. Then the Intro music kicks in.
Changed my mind last minute and included the usual catchphrase instead (was worried that folks will be upset about the abrupt sentence ending) hahahaha 😅😅😅
It is Zaporozhets, but with sporty spoilers. Might go faster.
I love you ❤
CPU Time : (i7 9700K and 32GB memory ) 1 minute and 48 seconds using my GPU 3.45 seconds (2080 TI)
I suddenly want to start learning python again.
In the 2000s A Boy looking for old graphics cpu
speeeeeeeeeedddddddddddddddd
I heard pandas and came as soon as I could
I think I'm first here 😊
Almost! @tanishaness got here first 🤪
(but still very fast!!!! a perfect match to the spirit of this tutorial hahaha 🤣)
Now we need support for AMD ROCm and OneAPI for Intel
You'll need something like OpenCL for AMD based GPUs 😉
Anything CUDA related is proprietary to Nvidia on the hardware level, so it can't be replicated with equipment from other manufacturers 🙃
thank you@@PythonSimplified
I love you 😍
Me: "Looks at my Intel Irisx. Well, I can always bookmark the video I guess..."
😂😂 Seems we're in same predicament! I know better now to get a pc with Nvidia GPU when next I wanna buy a new laptop.
I have been a Java developer, and did some C++ programming in my career, it is like you have person insight into everything I have done in my career up till now, I appreciate the work you have done here, though I did try and contribute an alternate method to your docker tutorial and you deleted it, I guess I don't understand why as there was nothing negative in it, a joke about not getting paid. Anyway, thanks.
Hi Louis! :) I never delete comments (unless they are phishing attacks), but RUclips does. If you include certain characters that very often appear in code (like HTML tags) - they may be flagged by RUclips. Your comments will then either be stored in a "held for review" section or they'd be removed automatically. I'll check this section on the Docker video shortly and see if I can find your comment there. It's just RUclips's way of handling bots and scrapers.
Thanks so much for the lovely feedback and I hope your next suggestions will pass through RUclips's filters with no issues! 😀😀😀
@@PythonSimplified I guess you learn something new every day, you would think they would have been able to master the use of tags so that they don't have to do something like that, maybe it was for email? Though I would think they could strip that out with something like a Baysian filter and simple AI, Or say removing tags from HTML from one of your screen scraping tutorials. If you ask Gemini what I typed in it would be able to you that "louis# podman " is a command line command, and that it is not likely a hacking attempt. It certainly would be easier to strip the characters like if you were building a prototype but I guess I viewed RUclips as a little more professional. I have seen you employ all the skills needed to solve all of these problems in code, and Google says they hire python developers, which begs the question if you make more on RUclips then working for Google directly? I had some ideas/features I thought were interesting or intriguing to me on this topic. Anyway, if I were in Google's HR, I would seriously consider hiring you, that was an excellent place to start looking for the cause. I didn't think there was anything offensive about it as it was clearly informative satire. Is there a link you could share that has this documented? Google has so much documentation and you are correct to assume I haven;t nor would I be able to read it all. I did have a joke in there that Gemini didn't provide a yes or no on offensiveness it just said it was very likely a joke than something offensive. The line was "I would have done more but I don't get paid for this sh1t" which it repeated back to me. Well since I checked to see if it would likely be protected speech in the US and Canada which it was it was less certain about if it violated Google's policy, though that would mean that a lot of comedy would be in violation. Maybe AI still needs a little help on protected speech. 8)
@@PythonSimplified Well I took a screen shot, the comments are still in my history even though they aren't listed on your site, I can post them on my linked in account if you want to see them there. If not no big deal I guess, I guess the truth just isn't that important to me.
when did u upgrade from RTX 3090 to RTX 4080 speeeeedd 👻👻
Where is she from?
Sadly, it's not just pandas-we seem to be mishandling the whole animal kingdom. It's a bit off-topic, but it's been on my mind. Great video anyway. You are a cute simplified snake girl... 😕😞🌎🌍
You are so funny your voice is like AI creature
or beautiful voice
salam dokhtari?
I must calm down and concentrate on studying,... studyyyyying.... Not the beautiful and dreamy blue eyed teacher girl, with hypnotic voice, alluring smile, wet gorgeous hair ... ehhm, ok where was I ?....
0 calls is quite impressive.
💯 Agreed!! At first - I was setting the encoding to "latin-1" (I always do it for this specific dataset).
And I couldn't understand why read_csv was falling back to CPU!!! It was driving me crazy!! But when I finally removed the encoding argument - the GPU finally kicked in and saved the day!!
Great Success!!! 😀😀😀
just use polars
Polars is definitely faster than Pandas... but is it 30 times faster like my cuDF example? 🤔
Checkout the challenge at minute 19:09 if you think that Polars beats cuDF, I'd love to see your solution! 😀
CPU -- 45s
GPU - 1.02s
reported to youtube, clickbait video, no actual pandas 🐼 in this video.
That's exactly why I hope RUclips doesn't recommend this video to bear lovers!!! 🤣🤣🤣🤣🤣
🤣@@PythonSimplified
How do I report this comment?
just ring your local police station.
@@munawarabbasi9683 Tobs was joking 😉
Python has serious security problems.
Generational humans, we love girls
Programmers : code , code, code................................and python chick
Share your LinkedIn profile please. I need to follow you and I have a bunch of questions plus I know you definitely have some resources I can use. Love your work
Could you make a collaboration with @NicholasRenotte. All the best insh'Allah