PyTorch in 5 Minutes

[@example.com.]

PyTorch is far more easier to use as a beginner. Writing code for most part looks like writing code, instead of fighting against the library.

[@Inflationforthemasses]

Wop, did not intend to offend anyone; just nicely interact: deleted my comment. Sorry man.

[@LeCoolCroco]

sklearn is even more straight-forward

[@SirajRaval]

whatever gets the job done

[@enntimessix]

Mochizuki-san, face it, beneath your clever Teichmüller camouflage, you're actually also a computer geek. When will you admit you're Satoshi Nakamoto ? ;-) kzbin.info/www/bejne/m56ne4d9idd4o5I

[@ANTIMONcom]

Estimator API and the Predictor class is a big fight. Not fun :(

[@CharltonLepkofker]

You know you’re entertaining when you get someone to watch every one of your videos and that person literally never understands what’s going on

[@hamzaleb9215]

hahahaha

[@_productivity__nill_1131]

I got the first part but the part when he was rambling about the code was foreign to me

[@DodaGarcia]

I mean he's also hot so that helps

[@jurriensi]

@@DodaGarcia :s

[@MrSupergingerman]

I spent several hours digging through documentation and issue logs to learn the same information that you present here in 5 minutes. Wish this was out a few weeks ago! Great content as always, Siraj!

[@quant-trader-010]

After watching this video, I am now leaning toward TensorFlow, as the impression I get is that TensorFlow is closer to Functional Programming Paradigm, while PyTorch is Imperative. I have a definite preference of FP code over OOP/Imperative code. I'd also prefer static graph over dynamic graph, as the former is easier to reason about than the latter. Thanks for making the video, Siraj!

[@ebeisaac7700]

Seriously, Siraj, I haven't seen any DL video quite as much entertaining as yours. What an inspiration!

[@ZedCraftHD]

Congrats for reaching 100k subscribers... You absolutely deserve them! Keep up the good content... and raps ;)

[@echoRohit]

I just learnt Python and I use to each your video just make myself feel as machine learning programmer 😀

[@sararojas7999]

love you and love your videos greetings from Colombia ❤️

[@SirajRaval]

Thanks! I also use instagram

[@edwardmitchell6581]

In under four minutes I learned the difference between imperative and symbolic models, the benefits of each, and sold on the Pytorch over Tensor Flow. I was really expecting click bait, but

[@feixyzliu5432]

Really cool! Looking forward to your releasing of more pytorch-related videos, since nearly all your videos are tensorflow based.

[@SirajRaval]

will do thx

[@akathevip]

Harold & Kumar is the name of a series of American stoner comedy films starring John Cho and Kal Penn. The first film, Harold & Kumar Go to White Castle, was released in 2004 by New Line Cinema and was followed by two sequels, Harold & Kumar Escape from Guantanamo Bay and A Very Harold & Kumar 3D Christmas.

[@conorgildea7441]

Discovered your channel, love your channel

[@rajathshetty325]

Checkout "Enter Siraj" :P

[@SirajRaval]

thanks Conor

[@andreialmasan7054]

I am a beginner programmer and I started watching your videos since the channel had just 2 or 3 videos and I must say that the quality of this videos keep getting better and better, keep it up

[@SirajRaval]

thanks Andrei!

[@michelaka6836]

Hello from Ghana! Siraj, we love your videos here too... Siraj.SetWorldOptimizerFunction(God_Mode=True, Simple_Explain="Always")

[@user-or7ji5hv8y]

awesome video! the way you structure your videos and get at the essence of the issue is really helpful!

[@janzaibmbaloch5484]

Has PyTorch changed or what? This code is not working anymore. AttributeError: 'NoneType' object has no attribute 'data'

[@bwallace722]

That was really clear and useful, thanks!

[@vishalmotwanitsec]

Thats correct. Pytorch is great for research. However, I wished there was a way to get over slow looping of python. Let me explain this... For example, If you need to use a standard cell like LSTM or GRU for RNN then Pytorch is great and super fast. But if you have to define a custom cell for your RNN, then Pytorch will slow down your code like crazy because now what is executed is the python loops that I have written(instead of underlying C or C++ loops) This is by far the only drawback I have come across in pytorch and I would guess this drawback would exist for other deep learning libraries too (or may be not)

[@OmarMiranda]

vishal motwani if it is open source, you can always check where and how the LSTM cell is, code yours following the same and recompile pytorch. Every research on arquitectural innovations must follow the same path, because hasn't been created yet :)

[@vishalmotwanitsec]

Thanks Omar, I may perhaps end up doing what you suggested for my task.

[@commentsanitizer7929]

I see how the tables are turned in these 3 years.

[@adamsnaider5535]

Hey Siraj, great video. It's been some time since you posted this, and since then TensorFlow added eager execution which defines-by-run like pytorch. Could you do a video on how the new interface works in TensorFlow?

[@snoopdoggfanclub]

great comparison of torch and tensorflow!... very talented presenter

[@cupajoesir]

This just blew my mind. I've heard about symbolic expressions vs computed/evaluated expressions but the idea of a dynamic computational graph is awesome. What other resources would you recommend to learn more about this topic?

[@javisartdesign]

Yeah, very clear explanation! thanks for that. Eager to see a becnhmark comparing speed and accuracy for current frameworks such as tensorflow, theano and pytorch.

[@artyte_b]

Tensorflow documentation is the best way for beginners to start learning? I had tons of problems trying to understand it and it was the first neural network api I worked with.

[@rajathshetty325]

Agree it isn;t actually.

[@arunghontale3189]

You should check out the pytorch documentations. They have really good tutorials as well.

[@EarlWallaceNYC]

As always, a great introduction

[@gabrielcostanzo9121]

Congrats on 100k subscribers!

[@SirajRaval]

thanks!

[@deraktdar]

This is actually a pretty neat example. Thanks! In recent PyTorch you have to move the zero to the bottom though, because grad tensors aren't initialized now until after the fist backprop. These fucking Python packages, and the never-ending bullshit with shuffling and deprecating stuff around for no good reason. It really blocks people from getting into it. Also fuck the haters - I don't care if you lifted some stuff from github. At least you made it understandable and in a lot of cases dealt with this deprecation crap. Continue doing what you do. Just because some douchbag dumped code on github without bothering to do the hard work of explaining what its actually doing or or how to use it, doesn't mean everyone who uses it as a basis for explaining to other has to reference their sorry asses.

[@Tomahawk1999]

Siraj Raval is like pointers in C - all pomp but no stuff. You just use them to know what and then go someplace else to actually learn anything.

[@vaarrunreddy5636]

thank you brother for ur videos...may God bless u

[@SirajRaval]

thank u

[@bryanpedrosa8061]

Pytorch in almost 6 minutes.

[@commentsanitizer7929]

The title is a clickbait

[@Inflationforthemasses]

PyTorch surely sounds like a thing I am interested on learning more about! Please use it more often in the near future! Siraj, does pytorch compare to tensorflow in training times?

[@kushalchawla4395]

Extremely helpful! Thanks.

[@moonreaderx]

Hey dude love the videos! Can you do a series for less advanced users of python? I feel that your current videos are more tailored to experienced users of programmining, this probably highly limits your viewer base as there are a lot more programmer newbies than veterans =p.

[@manust396]

I know nothing about neural networks nor machine learning but you really got me till the end

[@sithoidinh3891]

nice introduction to pytorch

[@core36]

So torch, or pytorch, is a framework that provides functions to create neural networks, without the need of defining what a node or a weight is yourself. Did I get this right?

[@tnuts92]

I preferred that video to some other you posted, since you didn't seem to be in a rush for explaining and hence were clearer

[@ericsaund6858]

Love how you are the model for how that guy on American Pickers gestures to the camera.

[@vulnerablegrowth3774]

I could definitely see myself using PyTorch for my research in physics, cool!

[@SirajRaval]

dope post github link to ur work in our slack

[@sabyasachimukhopadhyay1256]

Very useful!

[@chamisoufiane1054]

I love your channel !

[@shikharsharma02]

Alas!! Tried very hard but couldn't get what your videos are meant for.

[@MolecularArts]

Let us take a moment to appreciate the contributions of Einstein Pepe (at 0:28).

[@ricebastard]

Hey, thanks! And great job.

[@prismasasepaloalto1917]

Damn addicted for ur video s bro and I am getting more stuff.Thank you

[@Privacy-LOST]

Very to the point. Still Holds 2+ years later.

[@hanyuliangchina]

good tools for reinforcement learning， thanks siraj，

[@ajones1274]

Complete machine learning / AI newbie here. I wish to get into learning machine learning, probably through Tensor Flow from all that I have read. Going through your videos, and the content looks great, but a lot of it goes over my head. Where do you recommend starting for a beginner ? I started the Deep learning course Google has put together, but it seemed like it required some previous knowledge of machine learning and some Python libraries I have no idea about. My background is Javascript / C# and I have a very basic knowledge of Python. Any advice from anyone would be appreciated :) Thank you

[@catafest]

The simplest explanation of the effect of artificial intelligence on a bubble. I've worked a bit with the Torch and LUA and there are significant syntax differences at pytorch. I have no clues so I would have to work with the cpu.

[@holly_hacker]

Damn, your subscriber count has been exploding lately. You're going to hit 100k in only a few hours :D

[@JimmyHendrixJR]

Why am I watching this I've never wrote a single line of code in my life but I'm sure this is a great video 🤣🤣.

[@7688abinash]

was thinking from a long time if you will ever shift from tensorflow... but cheers now you have :P

[@Eu_Sunt_Dracul]

one day I will learn everything I need to understand everything that was said in the video................one day. Keep learning.

[@momonga.]

You should look up Stanford's deep learning lectures they're all here on KZbin and really great

[@maximtheconqueror]

this video is so short. but, it feels so long

[@diwashshrestha4996]

Carlos Garcia because he is speaking slowly in this video.In other video he use to speak in lighting speed.

[@maxshibanov818]

that's what she said

[@sandeepkumarladi9127]

Hello Raj.Could you please upload a video tutorial on hyperspectral image classification using Deep learning.

[@tomyoung2674]

Yann Lecun loved this one on Facebook, bro. So do I.

[@shairuno]

I gave it a try and implemented a few models, and what frustrates me the most is its lack support of broadcasting.

[@BryceChudomelka]

There is an error in your print(t, loss.data[0]) line

[@humdinger8998]

Due to torch version. Just remove [0] and it works fine

[@BryceChudomelka]

@@humdinger8998 thank you

[@PascalD87]

@@BryceChudomelka For somebody stumbeling upon this in the future: better would be loss.data.item()

[@motivationmusic4821]

thank you for your video, I'm going to try PyTorch :)

[@SirajRaval]

dope thx

[@androidpc9656]

As an intermediate in python where to start learning machine learning ? *confused*

[@Yashpandey467]

I have a question, so what you mean by dynamic graphs is that we dont need multiple folds to compute our weights we can just train our model in 1 go or in just 1 fold?

[@MattiaPiola]

I keep watching your videos, and I never understand anything. Pffff I'm stupid.

[@mateos844]

Mattia Piola same Hahaha soon or later I will start to learn machine learning

[@mathephilia]

I suggest you take a look at neuralnetworksanddeeplearning.com/ it's what I started with. You don't need to understand every technical detail yet to understand most ideas behind the various models. You can easily build up from there (every time there's a new concept like RNNs, DCGANs, backpropagation, simulated annealing, look for blog posts, articles (wiki or scholar) and videos on the subject and take notes to explain what you are reading to yourself). Siraj's videos are great when you already have most of the vocabulary mastered (because they are fast, synthetic and present things in an often original and rich way), but before then it's a bit hard to follow. Or worse, you'll feel like "you're supposed to understand" but actually never will because there's a few words he'll use to which you don't REALLY know the definition (both formal and intuitive) or have never heard of before (offhanded references to some aspects of TensorFlow code, etc). Think of theses videos as an "intermediate" step to artificial neural networks.

[@MattiaPiola]

Tristan Duquesne thank you, I'll check it out.

[@mathephilia]

Glad I can help :)

[@SirajRaval]

nah this is advanced keep watching more beginner friendly stuff coming

[@pshubhaprasad]

Tenserflow or pytorch ? Which one is better ?

[@impaler42]

Hi Siraj, fantastic video. I'm getting the following error running the script on Jupyter Notebooks:

[@impaler42]

Index Error: invalid index of 0 dim tensor. Any idea what's going on? Just grabbed the code from your GitHub and tried running it without modification.

[@mostafamohsen250]

very informative, thank you

[@nikunjrastogi1806]

Hi Siraj, I am a data analyst and currently using windows 10. Please suggest me which is better for machine learning or artificial Intelligence in between windows and linux.

[@harshagarwal3855]

Linux or Mac OS would be best for Data Analysis or Data Science

[@kukuster]

Lua is hard to learn?????? Ok👌 Besides that, a very good and succinct video. Thanks!!

[@waynefilkins8394]

"Kids these days are so stupid and can't pay attention to anything for more than 6 seconds, so i'm going to put a shitload of memes in my videos to keep these idiots watching". I know you said this to yourself or to someone else at some point. Smart man because it's 100% true.

[@yogeshkulkarni]

Siraj, Good intro... one query: in the video (at 5.01) you ZERO the weights before loss.backward(), but that is giving me 'None' type error, whereas the github code you mentioned, works fine where weights are ZERO-ed at the end, inside 'for'. Please check.

[@withoutmalicexo]

I think I'll sign up for grad school now

[@mathephilia]

No need: just check resources online. You can start here if you feel overwhelmed by Siraj's videos, which are more for intermediate level in deep learning : neuralnetworksanddeeplearning.com/ :)

[@withoutmalicexo]

Thanks for the link/suggestion :) I appreciate challenges, and his videos challenge me.

[@mathephilia]

Challenges are always good, and it's quite nice to keep yourself interested with the "good stuff", which is generally more complex. However, it's still extremely useful to really master the basics so you can get the most out of more advanced material !

[@ozzycampos8008]

Thanks for that link, Tristan

[@Dlaba1973]

Thanks. But was the video made before the advent of Keras?

[@MIraCFA]

Next one to do: Sonnet. Is it on the map for you to talk about?

[@skrmnghrd4520]

I was drunk when i stumbled upon this vid. soooo i played it on 0.5 speeeed sooo wee aaaaaare ooooon theeee saaaaaame waaaaveleeeength.

[@MarkJay]

Man I just started using tensorflow and there's already a new library! we're moving too fast!!

[@commentsanitizer7929]

Now this new library is the new norm

[@zegov94]

Using dynamic graphs, can I change the neural network during training as the epochs are being processed? Should it be useful for some application?

[@eglintonflats]

I was very surprised that all this arm flapping did not result in you flying away

[@trueperson22]

Please update your github code: print(t, loss.data) To avoid: invalid index of a 0-dim tensor. Use tensor.item() to convert a 0-dim tensor to a Python Thanks a lot

[@kumarsen88]

So, do you suggest Keras+Tensorflow or Pytorch???

[@StereogasmSounds]

You showed "the hard way" of implementing a NN in Pytorch. You didn't use the torch.nn.Module class nor the advantages it gives. With modules, code in Pytorch is substantially cleaner than TF, thus making it more suitable to beginners as well!

[@SirajRaval]

indeed good catch. thought about including that. hard to fit in time limit. perhaps future vid

[@ЭльмирЭльмир-д5с]

So, pytorch will be useful for reinforcement learning,amiright?

[@lasanihussain8810]

Sir, which book should I study for learning ML and deep learning ?

[@onehumanwasted4228]

How can Tesla use PyTorch as it has to run on distributed hardware?

[@maanvis81]

Thanks for pointing pytorch out. I already worked with the Lua implementation and saw the python variant as just a port but it appears that it has more going for it than I thought :).

[@jackbeefree2753]

How I unsubschribe for dont receive e-mails by pytorch github ?

[@madmotorcyclist]

So funny, that after 40 years what was done in LISP is catching on nowadays.

[@Skandawin78]

Is there a 'R' Torch 🙁 as I'm quite comfortable with R

[@smcalpha]

Tensorflow supports R

[@NoC4k3]

Cool! I learned almost nothing about PyTorch in 5 minutes, thank you

[@avareallymeow]

ok this hair is kinda fire lowkey

[@yarcowang]

(I'm the guy from after two years when you left your messages here. So, are you guys now experts in machine learning now?) Aha~

[@commentsanitizer7929]

No, not quite.

[@ARJUN-op2dh]

Next video, how to make a baby in 5 minutes...

[@Otnielush]

Thanks so much

I m willing to develop a chatbot able to benchmark from various sources and inputs (text, voice, image, graphics) that allow someone to ask information on various (directly and indirectly related) and the bot producing a summarized output. What would be the building blocks to consider to develop this project ?

[@ycyan3473]

nice video!. can you please make a video on caffe2 . from their website I came to know that caffe2 could be better in distributed system too which was not so great for old caffe

[@TheWild69Thing]

Seems like pytorch is very pythonic compared with tf

[@kudoamv]

So can AMD card users use Pytorch with gpu????????????

[@johnmark-ps8jy]

@@kudoamv Yes, very well...

[@larryteslaspacexboringlawr739]

thank you for making video, question about possible research topic, could you use generative adversarial network (with genetic programming (prof. john koza stanford) to help detect malicious binaries (below a certain size on a specific operating system) where the you could use the database of public known malicious binaries as a training data and one network tries to detect the malicious binary and the GAN actually compiles or writes the source code a that ends up being very similar to a malicious binary?

[@Mizar88]

is it just me or there is an extremely annoying low-volume background noise (music?) in this video?

[@chrisanderson1513]

Are you using clamp as a relu?

[@exploretheworld17]

tensorflow is hard to beginner, however I do think it is more powerful

[@isweartofuckinggod]

Is learning Lua that big of a barrier?

[@xNiickT]

What do you mean by "PyTorch is great for research" ?

[@smallbluemachine]

I want that answer too, and in 2 to 3 minutes. I thought 5 minutes would be ok, but there's too much flexing here and not enough plain-talking.

[@michaelphelps2350]

2:02 Aren't i2h and h2h labeled opposite from the code?