Rank, Axes, and Shape Explained - Tensors for Deep Learning

[@deeplizard]

Check out the corresponding blog and other resources for this video at: deeplizard.com/learn/video/AiyK0idr4uM

[@reefcrazed2070]

Very well thought out and explained well. I like these 10 minute chunks too, just enough to not be overwhelmed with too much new information.

[@deeplizard]

Thank you Reefcrazed!

[@joebender9052]

I really like your teaching style. I feel like I'm getting a much better understanding of things I thought I already knew.

[@muayas9602]

Very nice explanations of tensors. Plus those clips from developers gives the final touch.

[@sebgrootus]

I've been studying AI for 2 years, but i gained a better understanding thanks to your teaching style. Very nice :)

[@akarshjain7141]

A new perspective to my existing knowledge of rank, axes and shape of tensors ! Thanks deep lizard !

[@Brahma2012]

Reshaping has been explained here concisely at its best. Thanks

[@yousufborno3875]

Wow !! you guys made it so simple. Great job ! Please keep making vedio. This channel deserves a lot of subscribers and should be a go to channel for ML/DL developers.

[@abyssus9934]

Very intuitive explanation! Thank you so much!

[@Sikuq]

Excellent foundation recap. Thanks.

[@ashutoshchauhan5586]

Very nice explanation

[@deeplizard]

Hey Ashutosh - Thank you!

[@iamparadox8885]

this tutorial is awesome

[@atmismahir]

at 3:16 the comment should be "The third index of axis 2" . Thanks for the tutorials :)

[@s25412]

It threw me off when watching the video, but it all makes sense after this comment

[@Spencer-to9gu]

seriously, that confused the shit out of me

[@qzorn4440]

So after PyTorch is setup what would be some simple test examples to make sure it is functioning correctly? 😎 Thanks.

[@ernieblessed9510]

so intelligent, I felt smart now

[@Nissearne12]

I like this serie of PyTorch tutorial much.

[@Fatima-kj9ws]

Awesome

[@mjackstewart]

Hmm … A mathematical tensor is an object that remains unchanged under a change of coordinates, and its components change in predictable ways under a change of coordinates. So when you say a neural network tensor retains the same information but can take on different shapes, is THIS the underlying parallel between math an NN tensors?

[@abhijeetsharma5533]

3:02 each element along the 2nd axis is ..? shouldn't it be an array as well(and not a number)? because only when we access a particular index of that [element along the 2nd axis], we'll get a number. Isn't it? Correct me if I am wrong.

[@forvm2051]

8:47 I really laughed...

[@tonihuhtiniemi1222]

I

[@deeplizard]

Hey Toni - Thank you for the ❤️ and joining the Hivemind on Patreon. Chamath is cool!

[@BenR-qj3fn]

Is that the music from Kerbal Space Program in the background?

[@SOMADLOVECLUB]

1:40 lmao KSP music escootit

[@iamparadox8885]

this tutorials are awesome

[@zes7215]

wrg, no such thing as growing intelligenx or not, nonex, doesn't matter, cpux, any infinitely intelligent nmw, learn or machine learning or not doesn't matter, think,do,canthink do any nmw and any can be perfx

[@aakashsingh8120]

What's adding along axis = 1 mean here?

[@deeplizard]

timestamp?

[@shanakaj007]

glad to see a Sri Lankan in the video...

[@deeplizard]

Thanks for mentioning this. I previously thought Chamath was from India. Are you also Sri Lankan?

[@shanakaj007]

@@deeplizard Yes I'm and that's a fair confusion. hahaaa i'm huge fan of your videos series because you are so talented in delivering advanced stuff in a simple manner with graphics..

[@Floodbait_117]

I hate math but want to learn about deep learning my curiosity is out weighting my hate of math

[@prateekyadav9811]

I am wondering if I specially dumb or something. I am yet to finish this video but since I am totally confused I felt compelled to make this comment. In previous video of this series, you said that the dimension of a tensor is the number of indexes needed to obtain an element from the tensor. From the slide displayed at 1:00, it seems that rank is the same as dimension because you are asserting that a 2d array (indexes = 2 and thus, dimension = 2) has rank 2. Then, from 1:47 to 1:56, you state that a tensor of rank 2 means that tensor has 2 dimensions or 2 axes. You didn't at all define what these terms mean but merely pointed to the notion of indexes. In all this, I only understood what an indexes are, that's all. I am sorry but, I, personally, found this to be a frustratingly poor explanation.

[@vipulmishra554]

At 7:43, can't we interpret it as reshaping to a 2d-array with 1 row and 9 columns? Because reshaping it to a tensor with length of first axis as 1 and second axis as 9 makes sense to me only if 1st axis refers to columns and 2nd one to rows. So, is that the case about axes in pytorch?

[@Sikuq]

Fantastic PyTorch introduction. I downloaded your code via your Patreon, run it step-by-step in VSCode, and I now realize just how far I still have to go. Thanks.

[@prachijadhav7578]

{ "question": "How many dimensions given tensor t has? t = [[[1,2], [3,4], [5,6]]]", "choices": [ "3", "4", "5", "2" ], "answer": "3", "creator": "PrachiUR", "creationDate": "2021-11-25T07:30:53.447Z" }

[@lingshuaikong5644]

Thanks for this great series, Sir! It really helps. Also, I thought maybe there's a little typo in the video when you say the third index of axes 3 in the time of 3:23 because axes 3 has only 1 index.

[@plebblesrollingston]

Minor Corrections: @03:16: "The third index of axis 2". Instead of 'thrid', it should be 'third' and instead of '3', it should be '2'.

[@DanielWeikert]

Thank you guys! I am really looking forward to future videos about reshaping. Honestly that is one of the most difficult things to understand for me. When do I need what shape of my data? Hope you will enlighten me ;) Great work again!

[@deeplizard]

Hey Daniel - You're welcome! We're with you on that one. Shaping is tricky. Lot's to look forward to on shaping coming up. Especially when we start building the CNN. We've got you! ;)

[@DanielWeikert]

Awesome

[@richarda1630]

The Chamath cameo caught me by surprise :) how profound to watch this now in 2021

[@griffingibson4389]

t.reshape{9,4} would be how you get the neural network shape at the end of the the video then? Never typed a line of code in my life have mercy.

[@WangLiaoYuan]

the example (3,3) is not good. If it could be (3,4), so the rank is 4, axes is 3. Am I right?

[@fahimtaibahmed1335]

Which video is the indices? is there a link thank you

[@deeplizard]

Hi Fahim - You can find an organized list of all the videos here: deeplizard.com/learn/video/v5cngxo4mIg Tensor episodes start on number 5. Good luck!

[@viictorsperanta5023]

You are a great explainer. Thank you.

[@het314]

A multidimensional Array?

[@Vikram-wx4hg]

Excellent video!

[@wilsonzhang5258]

At 4:43, shouldn't that be "we have multiple words for the same concepts" instead of "different concepts"?

[@deeplizard]

You are on to something there 😂 I think I was trying to articulate that the "multiple words for the same concept" phenomenon was occurring multiple times. 😄 Seems to have came out wrong.

[@louerleseigneur4532]

merci merci

[@leiferickson4841]

thrid index

[@amdenis]

Do you have more of these types of videos on TensorFlow? They are great videos, but the market has really shifted strongly to TF over the past couple years-- at least for our US teams. I run consulting firm HQ'd in Silicon Valley and Research Triangle. with 16 teams of people, who are hiring on new DS and AI people every month now. Although PyTorch still accounts for some of the projects, about 80-90% of all new projects we get across commercial, defense and government are all TensorFlow 2.0 now. As such, we like to provide support and learning materials for the new people coming on, such as books and videos. For better or worse, the market is growing faster than the average skill level of people, and many of the new people who we develop and groom over time have a year or less AI dev experience. So, if you do not already have all of these sorts of videos in Tensorflow rather than PyTorch you should, as your instructional style and knowledge level are perfect for many of our new hires coming out of school.

[@deeplizard]

Hey Andrew - Thanks for your comment and these details 😊 We have a set of courses that are detailed here: deeplizard.com/ My view is that AI devs should be encouraged to learn multiple libraries. The knowledge should hopefully generalize. Especially for the tensor videos in the PyTorch course, Part 1 in particular (all about tensors), the information is definitely general, the knowledge gained applies to TensorFlow as well, even though the demonstration is with PyTorch.

[@amdenis]

​@@deeplizard - So true and thank you so much for the feedback. Please don't think I am criticizing your teaching. It is about the best I've come across in a long while. Also, we do use a bunch of different frameworks, languages and related tools, including some of the proprietary ones for certain .mil and closed architecture projects. It is just that we are finding an increasing number of the kids we are hiring have spent too much time on some tools and frameworks at the expense of what many of the major projects increasingly require. We don't mind starting good people who can learn efficiently at $100K with somewhat limited experience, but we are having to pay $150+ & equity and train them for 12 months because the universities and other institutions have exposed them mostly to less applicable tools. This is not a new issue; when I graduated some decades ago from MIT, and then even subsequently at U of I Champaign Urbana, they were teaching using languages like PL/I and Pascal, Scheme and Modula-2, despite the fact the business were clamoring for Fortran, COBOL and related alternatives. While on average I'm glad I learned everything from Ada to Zilog AL, I spent the next 3 years after graduating learning languages and frameworks that were most needed. Things moved much slower and budgets and expectations were lower then, so it worked out relatively okay. However, even then I wished I had learned more useful tools and languages. Since then, I have always tried to make sure that my background and the teams I hired are educated based on what's actually needed, as is reasonably possible; and all I can say is that technology is now moving too fast for even the "best" universities to stay current in what they teach-- let alone pro-active. It generally takes years for such institutions to propose new curriculum and associated tools and possibly get them approved. I know and agree with what you are saying, and we encourage breadth and depth, and still maintain dev suites running things like old versions of Matlab and Simulink on SGI Onyx's to about a dozen versions of virtualized COBOL and Fortran. However, and this is where groups like yours come in and really can shine, you CAN stay current and ahead of the game with tools like TF, Swift, etc, in order to make up for the limitations and deficiencies of our educational institutions. Of course we hire deep theory people with doctorates, and seasoned (read 'old like me') researchers who have deep specialties. However, like most others in the field (AI, Quantum-AI, crypto and related), we have little problem finding great available specialists with deep, somewhat less mainstream skills. What is difficult to find are people with what 80%+ are hiring specifically for: deep, efficient TensorFlow ML and DL hosted, platform and embedded skills. Maybe 1 out of 9 or 10 of our last hires were of this ilk and didn't require 1-2 years of concurrent OTJ training. We are always happy to pay more than double what the typically trained hires get plus incentives for such recent grads or self-trains. The issue is that over the past 5 years it has gone from finding them every week, to every month and now just two last year. This is why I am ecstatic to find resources like yours, where the material is taught better than virtually anything I have come across anywhere. Your blending of theory with practical, and the accuracy of what you put forth are better than almost anything we come across. Also, you do have a good mix, and you definitely teach some of the most useful tools. As I said, about 10% of our projects still involve some amount of PyTorch. I just mentioned the TF as a note to you to ensure you're aware the bigger role that you play given your amazing teaching abilities and knowledge set.

[@nehadyounis]

hey isn't it the music from kerbal space program?

[@deeplizard]

KSP creators used the same music library as we did :D

[@nehadyounis]

deeplizard that might explain why both of you are awesome

[@silversword7528]

Hi, very good video. Just to note that there is an error at mn 3:19. You have written axis 3 where is should be axis 2 I guess.

[@maheshvenk]

A big Thank You

[@deeplizard]

You are welcome!

[@iamparadox8885]

.

[@kyle_bro]

I had to check and see if I had ksp running in the background lol

[@deeplizard]

Lol

[@aipreacher9378]

Where are the notebooks available???

[@deeplizard]

Hey Sourangshu - Download access to code files and notebooks is available as a perk for the deeplizard hivemind. Check out the details regarding deeplizard perks and rewards at: deeplizard.com/hivemind If you choose to join, you will gain access to the code from this series at the link below: www.patreon.com/posts/code-for-pytorch-21607032

[@Небудьбараном-к1м]

Had a hard time figuring what leength is