I did my final project for my degree last year using TensorFlow. I'm tearing my hair out now learning about PyTorch, this looks like it would have saved me so much time and suffering 😭
There are actually *very* few differences between the 2. Their use cases may differ, but that really only because of what a company decided to use for their project. You can actually switch from one to the other with little effort.
I agree with the above commenters, but if you are familiar with Numpy, USE PYTORCH. Numpy is extremely similar to PyTorch, and they even share an astounding number of identical functions, where PyTorch’s version just operate on tensors instead of numpy arrays and can use your GPU. Creating an array in numpy can be done identically in PyTorch with the Zeros(), Ones(), and empty() functions. They also connect very well, there’s even a from_numpy() function to get convert between the two
IMO, the less code something has, the better. *Especially* when it's boilerplate code. Sometimes I believe people have a typing fetish... I actually have a coworker that likes typing and he writes a lot of repetitive code and does not recycle the code.
@@SI0AX You mean the less code _you_ have to type the better. We all want complexity abstracted away with nice easy to use APIs we can just chain to get things done, but there's still a bunch of code someone else had to write running in the background to get that thing done.
I did a project with Tensorflow before and was really struggling with Tensorflow and am looking into what to use for my thesis, so i think your comment should be reson enough to check it out 😂😅
Thinking of this, Pytorch is trending and why there is no 100 second video from Fireship. Just got into machine learning in pytorch. Absolutely 🔥🔥. Need a comparison with Tensorflow. 5days before the version 2 became stable.
It depends a bit on your use case, pytorch is better for prototyping/research/personal projects but tensorflow is more easily deployable in the real world As an ML researcher I love pytorch personally and can't recommend it enough but tensorflow has its upsides too
@@Imperial_Squid Never tried Tensorflow before. Can I ask you something?, Which is more compatible with mobile devices. Tried pytorch with coreml and onnx. As ML researcher, which one do you prefer in case of Mobile/Edge devices
Tensorflow has been around a lot longer. Its why its used in industry. It recently combined with the Keras library and got a lot of functionality from that. Pytorch itself is much newer and is used today more so as a way to build models. A lot of people, myself included prefer Pytorch for its usability. I personally think Pytorch is going to surpass Tensorflow because the magic of deploying a model has been demystified for awhile now. Its all about applied model accuracy; preventing overtraining etc etc. This is just flat out easier with Pytorch. Integration of a model with an API really is not an issue today.
For deploying I'd had to say Tensorflow... Buf. I'd rather go with Torch. Sometimes not even the toy examples from TF doc works. Too much change from version 1 to 2. I'd stick to Torch
MORE DEEP LEARNING STUFF IN 100 SECONDS! I remember going through courses that took literally freaking days... tons of crazy equations and weird explanations... this stuff is gold man.
The best video tutorial on PyTorch I have ever seen. All the courses on this topic are so complicated but your video is an amazing one. I learned more than I have in a year. Thank you so much Fireship!!!
But this does not teach any theory or reasoning as to why you would do certain things. And what the effects of these things are. But I guess if your school is trash, then you probably learn more here.
Alright man. Seriously? You have my computer bugged don't you? I swear, every god damn side project I start the next week we get a video with perfect relevance. I love you
Great little intro to CUDA in here! Would like to see a video giving an overview on GPU layers (AMD ROC and whatnot), and maybe the associated hardware situation right now. CUDA/NVIDIA gets way too much press, it seems.
“PyTorch is a world-class machine learning framework. It has been used to build image generation models and highly advanced voice recognition systems. To use it, just import torch, define some tensors et voilà … there‘s your self-driving car.”
I think you get way better results for an image classifier if you start with a convolutional layer instead of flattening the input instantly. I havn't tested it though.
2:29 Thanks for not including sample data for "some_data", code is dogshit without it. Tried providing my own tensor and it gives RuntimeError: mat1 and mat2 shapes cannot be multiplied
As a college student who recently got into deep learning with TensorFlow, I can understand and relate to PyTorch. Both are fundamentally pretty similar.
Tensorflow is more for scaling up with other systems, pytorch is most useful for stationary or researching different models. I find tensor flow much easier to learn, whereas pytorch requires more knowledge and memorization of the syntax. Both are useful for separate things, but you can use them in the place of another, except tensor flow since pytorch does not support a lot of the platforms still. It will support mobile in a couple years fully though. TensorFlow is my favourite one
Install pytorch and optionally CUDA. Getting the right CUDA version and pytorch version and the other things for your GPU to run properly. Was a pain in the ass for me.
>pip install torch...cu117 >pip install thing wtf is version "2.0.0+cu117"? No package management for you >pip uninstall torch >pip install thing The thing is CUDAless >pip uninstall torch >pip install torch...cu117 Please tell me this doesn't only affect me.
PyTorch is a more advanced version of "Machine Learning". Learning PyTorch without learning ML with scikitlearn is like learning to run without having learned to walk. AI is basically ML but scaled up and with extra algorithms on top, but it still applies ML logic.
I regret using tensorflow now, I had recently tried updating to latest tensorflow and found gpu support was dropped for windows. Now have to use wsdl and other workarounds just to speed up the training.
PyTorch can use CUDA, HIP or plain C++ for it's backends. Meaning it can run on NVidia GPUs, AMD GPUs and use their accelerators, or simply run on the CPU. So not just CUDA
I did my final project for my degree last year using TensorFlow. I'm tearing my hair out now learning about PyTorch, this looks like it would have saved me so much time and suffering 😭
Is TensorFlow not worth learning than PyTorch?
@@suyashshrestha1099 Just pick one and go with it. Once you understand one of them, you can pick up the other in 2 hours or something.
There are actually *very* few differences between the 2. Their use cases may differ, but that really only because of what a company decided to use for their project. You can actually switch from one to the other with little effort.
I agree with the above commenters, but if you are familiar with Numpy, USE PYTORCH. Numpy is extremely similar to PyTorch, and they even share an astounding number of identical functions, where PyTorch’s version just operate on tensors instead of numpy arrays and can use your GPU. Creating an array in numpy can be done identically in PyTorch with the Zeros(), Ones(), and empty() functions. They also connect very well, there’s even a from_numpy() function to get convert between the two
Pytorch is majorly used in academia and TF in the industry. Personally I found TF easier to learn than pytorch. Both are great in their own ways.
I absolutely have loved your videos for so long. Thank you for the incredible production quality and fast-paced information made simple!
Let's develop a deep neural network to find who made the chicken.
botted subs momenttt
Jaden moment
@@whit3rose A dinosaur made the egg that mutated and made a chicken
@@fahd2372 not really
The more I learn about Python, the more it seems like good python code should contain as little code in python as possible.
Try to build a table completely out of glue. Same thing, you want to use as little as possible
Good python code seems to follow the following pattern
import taskDoer
doer = taskDoer
doer.do_task()
IMO, the less code something has, the better. *Especially* when it's boilerplate code. Sometimes I believe people have a typing fetish... I actually have a coworker that likes typing and he writes a lot of repetitive code and does not recycle the code.
@@SI0AX now he can beautify his code using gpt4. Don't show it to him it'll break his heart
@@SI0AX You mean the less code _you_ have to type the better. We all want complexity abstracted away with nice easy to use APIs we can just chain to get things done, but there's still a bunch of code someone else had to write running in the background to get that thing done.
kinda wish this was PyTorch in 1000 seconds
the best kind of python code is written in c
@@benwright4096Eh?
@@youtubeacc666 leave him. He's just a normie who has just learnt about computer
@@architmishra015 they were prob thinking of CPython
@@architmishra015 Look at me! i know c is faster than python! i'm so clever
Me, a web dev: "Hmm, I know some of those words"
I used PyTorch for my thesis and it was really easy to use, I loved the modularity of everything
YoloV8 and Mindsdb is a great thing for us smooth brains. 😊
how did you got to learn pytorch can you tell me. I am interested in learning it but not able to find best source
what was the thesis on?
@@ramandeepkaurbanvat7583 If you finally got a good source, can you also share please
I did a project with Tensorflow before and was really struggling with Tensorflow and am looking into what to use for my thesis, so i think your comment should be reson enough to check it out 😂😅
your 100 second videos seem to come out exactly when I need them. I just started a AI, ML project this week.
And I am starting now
Will you please share details about your AI ML project
@@mujibshaikh7494SHUT up
0:18 it's "scalar" not "scaler"
Such large concept in 100 seconds? Just brilliant work. Thanks!
Meant for those who already understand the concept
Absolutely brilliant combination of conciseness and coverage.
Amazing explanation as always. Short yet concise. I hope you live 100+ more years.
Thinking of this, Pytorch is trending and why there is no 100 second video from Fireship. Just got into machine learning in pytorch. Absolutely 🔥🔥. Need a comparison with Tensorflow. 5days before the version 2 became stable.
It depends a bit on your use case, pytorch is better for prototyping/research/personal projects but tensorflow is more easily deployable in the real world
As an ML researcher I love pytorch personally and can't recommend it enough but tensorflow has its upsides too
@@Imperial_Squid Never tried Tensorflow before. Can I ask you something?, Which is more compatible with mobile devices. Tried pytorch with coreml and onnx. As ML researcher, which one do you prefer in case of Mobile/Edge devices
Tensorflow has been around a lot longer. Its why its used in industry. It recently combined with the Keras library and got a lot of functionality from that. Pytorch itself is much newer and is used today more so as a way to build models. A lot of people, myself included prefer Pytorch for its usability. I personally think Pytorch is going to surpass Tensorflow because the magic of deploying a model has been demystified for awhile now. Its all about applied model accuracy; preventing overtraining etc etc. This is just flat out easier with Pytorch. Integration of a model with an API really is not an issue today.
For deploying I'd had to say Tensorflow... Buf. I'd rather go with Torch. Sometimes not even the toy examples from TF doc works. Too much change from version 1 to 2. I'd stick to Torch
MORE DEEP LEARNING STUFF IN 100 SECONDS! I remember going through courses that took literally freaking days... tons of crazy equations and weird explanations... this stuff is gold man.
Learning PyTorch currently and it was nice seeing you make a video on this topic.
Book recommendation: "A Primer to the 42 Most commonly used Machine Learning Algorithms (With Code Samples)."
The timing of this video couldn't be any better because I'm working on some AI project with Pytorch right now :)
Haha me too. I am downloading the ai models for the first time at this time just before the video
Just curious, what are you working on?
@@valentinoedits1963 some analysis about anime music cd sales
now everything have that AI badge. you worked on web3 project before, right?
@@BlueEighthNote i want to get into AI where do I start, already learning from coursera.
Me clicking in 1 sec the notification drops
Parkour
Here for you🏆
@@dhirajnavale3861 😂😂😂
Script
Me here complaining how my notifications are always 1hr late. Does anyone know why?
0:18 "... tensors, which are basically just multidimensional arrays" cries in mathematics, hearing computer science
The code looks so much easier to understand then I expected. That's what makes me so excited and scared to even try this kind of thing.
@0:19 I think it is meant to say scalar rather than scaler
The best two minute and 43 second long video about PyTorch.
Now I'm ready to start my Data Science career.
Thank you Fireship!
For all both of us out there using Haskell, there is an implementation of Torch being developed for the language called HaskTorch.
Also exist other for Tensorflow.
The best video tutorial on PyTorch I have ever seen. All the courses on this topic are so complicated but your video is an amazing one. I learned more than I have in a year.
Thank you so much Fireship!!!
how the fuck lol
It amazes me to how much I can grasp in 100 seconds compared to what I might learn in a semester.
Nah you must be joking lol
@@tropicaldog430 it can be real, education in school don't teach very well especially on relatively new tech like this.
you're not wrong tbh, if you go to a garbage school then yeah, any youtube video is better than your semester at your school.
@@lottexy I go to the toughest school to get into. I go to IIT Kanpur which is like Indian MIT. Where I study statistics and data science
But this does not teach any theory or reasoning as to why you would do certain things. And what the effects of these things are. But I guess if your school is trash, then you probably learn more here.
Alright man. Seriously? You have my computer bugged don't you? I swear, every god damn side project I start the next week we get a video with perfect relevance. I love you
This is the only channel I have the bell 🔔 for notifications because of its pure quality of content 🔥
Great Video. Amazing demonstration of deep learning in Tesla Autopilot!
Great little intro to CUDA in here! Would like to see a video giving an overview on GPU layers (AMD ROC and whatnot), and maybe the associated hardware situation right now. CUDA/NVIDIA gets way too much press, it seems.
You summarized about 2 hours of Google searching in about 100 seconds. Nice job!
i really love this feeling of knowledge flowing through my brain, and none leaving after 2 seconds
let me cry a bit
Keep them coming! I love these 100 second videos!
i have no idea what i'm hearing but i'm definitely excited about it
The data you provide is invaluable when it comes to keeping me up to date with the tech around me. As a developer, I can't thank you enough.
It's Impressive that You Have Knowledge/Intuition in Many Field.
Keep On!
You are an amazing teacher♥️
Please don't stop making these kinda videos 🥹♥️♥️
"what you came here to do though..." is exactly what i came here to do
Great video! :) I.think it would be important to add the backforward step in pytorch so people could see the complete cycle of training a NN.
0:21 isn’t it “scalar” instead of “scaler” ?
“PyTorch is a world-class machine learning framework. It has been used to build image generation models and highly advanced voice recognition systems. To use it, just import torch, define some tensors et voilà … there‘s your self-driving car.”
Wonderful video!!! I would love to see Scala in 100 seconds next :)
1:26 miss-counted :-)
I was just about to request this one! Thanks a lot!
I think you get way better results for an image classifier if you start with a convolutional layer instead of flattening the input instantly. I havn't tested it though.
Agree, but this video is just a basic introduction to PyTorch in 100 seconds. Introducing the convolutional layer makes the video even longer.
Thank you for a brief overview, It was so helpful.
I just started learning pytorch and here is Fireship!
We need a full length tutorial on PyTorch!
Very good and concise explained, as always ^^
I’m going to pretend like I understood approximately 1 second in this video
Wow i thought i will learn pytorch but now i have learned neural networks as well as pytorch. Thanks!
Interesting video. Keep up the good work 👍
@0:20 -- *scalar
When Fireship is less confident about some cs topic he makes the video about it more serious and with less jokes
Excellent brief explanation at around 150 seconds
This video is a most excellent example of how to summarize!!!
Always great info on this channel. Thanks!
Pytorch Lightning is what you should use these days. It cleans up the code quite a lot.
How?
idk why I was expecting him to write a model that learns to say hello world
The fact that the tesla autopilot example had the car crashing and failing to do it's job is just chefs kiss 😘. (00:53)
Damn... started learning PyTorch for an interesting faculty, and you are here from the blue. THX for the vid, Torch really deserves more audience.
Hello @fireship,
Can you please make a video about the python framework "flask" ? I am still confused about the functionalities and benefits of flask.
Particularly the tensor library is a must have for ML/ Data engineers - totally agree
This video basically summarized 75% of my 100hour long applicated-development class into 100 seconds...
Always look forward to these videos
well that was more then 100 seconds more information then any videos!
My resume can basically say - I am subscribed to Fireship.
Looking forward to more AI stuffs! Keep up the good work!
I have no idea what you just said, but I loved it.
Video Suggestion: Tensorflow vs Pytorch😊
2:29 Thanks for not including sample data for "some_data", code is dogshit without it. Tried providing my own tensor and it gives RuntimeError: mat1 and mat2 shapes cannot be multiplied
0:53 Damn, that Tesla Autopilot needs some more training…
I like how the Autopilot clip is a Tesla rear ending another car 😂
0:53 Whoa, what's the original context of that video?
I'm dying at the crash at 0:53
100s of pytorch, years of statistics and math to understand what's really happening.
Finally I can say you read minds and released a video about something I just started learning
As a college student who recently got into deep learning with TensorFlow, I can understand and relate to PyTorch. Both are fundamentally pretty similar.
The best way to put is is that TensorFlow is a library, whereas PyTorch is a framework. Also PyTorch is just way more pythonic than TensorFlow.
Tensorflow is more for scaling up with other systems, pytorch is most useful for stationary or researching different models. I find tensor flow much easier to learn, whereas pytorch requires more knowledge and memorization of the syntax. Both are useful for separate things, but you can use them in the place of another, except tensor flow since pytorch does not support a lot of the platforms still. It will support mobile in a couple years fully though. TensorFlow is my favourite one
Install pytorch and optionally CUDA. Getting the right CUDA version and pytorch version and the other things for your GPU to run properly. Was a pain in the ass for me.
0:52 I love these satirical inserts :D
I made a chess playing CNN with PyTorch! It played great opening, and atrocious endgames
watched at x2 speed, learned pytorch in 50 seconds
Can you make a tutorial on how to pick the right cuda version for use with Docker?
After learning Pytorch and seeing this vid, it gives every basic info
Instructions unclear, my computer is flying and shooting lasers
I laughed so hard when that car just crashed using some AI vision, as you were talking about a framework used to build famous AI products.
0:54 went from braking to breaking real quick
FINALLY! FINALLY! FINALLY! FINALLY! FINALLY!
Thank you Jeff!
>pip install torch...cu117
>pip install thing
wtf is version "2.0.0+cu117"? No package management for you
>pip uninstall torch
>pip install thing
The thing is CUDAless
>pip uninstall torch
>pip install torch...cu117
Please tell me this doesn't only affect me.
the quality of the analogy and abstraction to explain concepts is outstanding, kudos 👏👏👏
especially 0:41 - 0:42
woah, got an ad with your voice before this video, really trippy
Thanks for this. Been trying to do more than django/flask with python and this is my next learning step
I always consider that ML/AI/Data is the main usage of Python and web framework like Django/Flask is just "hey we can do web dev using Python too!"
PyTorch is a more advanced version of "Machine Learning". Learning PyTorch without learning ML with scikitlearn is like learning to run without having learned to walk. AI is basically ML but scaled up and with extra algorithms on top, but it still applies ML logic.
@@SI0AX gotcha so go ML first before pytorch. Thanks for the advise
understood everything perfectly, gonna code skyned
what is the name of the effect with the appearance of the logo at the beginning of the video?
I regret using tensorflow now, I had recently tried updating to latest tensorflow and found gpu support was dropped for windows. Now have to use wsdl and other workarounds just to speed up the training.
does it use CTypes for controlling it all so it is not as slow as python? it does use tensor cores so it should be ok?
But just wondering.
It uses numpy I think, and that uses C. Which makes it fast
So it’s the same as tensorflow? Including tensors, cuda an a computation graph
I have no idea what Fireship talked about in this video but he made it look awesome. 😂
Looking at a fireship video with a fireship ad lmao
video is wonderfully useful.❤
PyTorch can use CUDA, HIP or plain C++ for it's backends. Meaning it can run on NVidia GPUs, AMD GPUs and use their accelerators, or simply run on the CPU. So not just CUDA
and also MPS, Apples thing