Code for Deep Learning - ArgMax and Reduction Tensor Ops

[@deeplizard]

Who's seen the movie IT? If you've seen it, tell us: 1) What country did you see it in? 2) On a scale from 0% - 100%, what's your rating of the movie? Check out the corresponding blog and other resources for this video at: deeplizard.com/learn/video/K3lX3Cltt4c

[@sourajitsaha4724]

Bangladesh, 95%

[@victortarnovskiy8407]

Russia, 90%. It is truly a masterpiece.

[@rajeshviky]

Finland 83%

[@林盟政]

Taiwan, IT version 85%, element-wise version 120%

[@min-youngchoi3833]

South Korea / 100% and might be overflow

[@thepresistence5935]

I literally wasted more than 6 hours watching this continuously and successfully completed 11 video. It's like watching a movie :) Thanks a lot, I am understanding everything.

[@ioannis.g.tzolas]

Congratulations on the terrific content and style of presentation. Your videos are like watching a TV series (with tensors as heroes). Can't stop watching !!! Congratulations !

[@deeplizard]

Wow! Tensors as super heros! Love it! Thank you! 📺

[@mohammadsadilkhan1875]

Correct me if I am wrong. This is the most interesting as well as a great resource for learning. Your idea of putting videos of pioneers in this field is truly marvellous.

[@stormzrift4575]

This has been fantastic so far. The memes make it so much more engaging

[@hzmuhabbet]

I am happy to find that channel, there's a great work here mate. Thanks for your efforts.

[@user-or7ji5hv8y]

Really finding this series rewarding. very much motivated to learn PyTorch.

[@deeplizard]

Thank you, James! I'm glad to hear that!

[@philtrem]

same !

[@_edwardedberg]

Dude.. your videos should be the world standard of neural network education.

[@deeplizard]

Thank you, Edward!

[@Brahma2012]

This is one of the best trainings on PyTorch. The coverage is highly relevant and exhaustive. Thank you

[@techfirsttamil]

No words to thank you guys. Always raising bar of quality content about deep learning. Heartly thanks from India

[@deeplizard]

Thank you, Venkatesh! Appreciate that!

[@bisnar1307]

Thank you for this series. It's extremely helpful. Please continue teaching people more!

[@sumanmondal2152]

You're way of teaching is amazing !!

[@harishjulapalli448]

Great Video. I got the concepts very clear now. Thanks a lot for your efforts.

[@kyle_bro]

This really is the highest level of content. 10/10

[@Fighter_Believer_Achiever]

Bro brilliant work. Legendary!!

[@DanielWeikert]

Love you guys. Great video again. Highly appreciated.

[@deeplizard]

Thanks Daniel! 🚀

[@Gee.U]

this channel is a blessing, thank you!

[@saptarshidattaaot]

Will never ever forget element-wise operation

[@ferielferiel5006]

OOOh my god !!! I was watching this video late at night !! and you hell scared me

[@deeplizard]

🤣🤣🤣🤣🤣

[@haroldsu1696]

really awesome lectures.

[@herrmann5384]

I dont get, why the order is changed, when you set a number for "dim": For example: t = torch.tensor([ [1,1,1,1], [2,2,2,2], [3,3,3,3] ]) print(t.shape) = [3,4] = 3 = first rows, 4 = then columns . But when you use the sum method: print(t.sum(dim = 0)) = tensor([6, 6, 6, 6]) dim = 0 : adds the columns and dim = 1 the rows.... why is it the other way around? Thanks so much!

[@deeplizard]

Thinking in terms of rows and columns can be problematic when dealing with tensors because they are more abstract data structures. Instead of saying we have 3 rows and 4 columns. Let's say we have 3 things and for each of these we have 4 things. Where are the 3 things? dim=0 What are the 3 things? rank-1 tensors (arrays) of length 4 How do we sum rank-1 tensors together? via element-wise addition The last part (element-wise) is the key. When we take the sum with respect to dim=0, we are summing the elements that run on that dimension. If these elements are not numbers, the sum is an element-wise sum. Since each array has a length of 4 (4 indicies), the result has a length of 4. The same logic applies to dim=1. What are the 4 things. numbers How many groups of 4 things do we have? 3 This means we get 3 sums of 4 things. This idea extends to higher rank tensors as well. Let me know if this helps.

[@hamedhojatian3098]

Perfect! Thank you!

[@deeplizard]

Hey Hamed - You are welcome!

[@elibelash8238]

Perfect, thank you.

[@deeplizard]

Hey Eli - You are welcome!

[@rubencg195]

Awesome!

[@deeplizard]

Thanks Ruben! Really appreciate your comments!

[@AnmolGargmed]

at 3:40, isn`t that rank 1 matrix ?

[@deeplizard]

Hey Anmol - The tensor t at 3:40 is a rank 2 tensor. There are two indices required to access an element inside t. Refer here: deeplizard.com/learn/video/AiyK0idr4uM If it still isn't clear, let me know more about how you are thinking about it. Hope this helps.

[@reb5642]

Thank you so much sir. Your video is so great

[@deeplizard]

Hey REB - Thank you!

[@ilfat_khairullin]

Thank you!

[@shanthalvasanth7782]

It makes sense that with t being a 3x4 tensor, the output of t.sum(dim=0) has the shape (1x4) but I don't get why the output tensor resulting from t.sum(dim=1) does not have the shape (3x1) ? It is unintuitive that sum over dim=1 yields a 1x3 tensor. EDIT : Just looked up Pytorch documentation - apparently you need to pass 'keepdim=True' to retain the dimensions and get a (3x1) tensor when summing over dim=1.

[@tingnews7273]

I thought I have figure out the dim torch.sum(dim=0)。But when I see the torch.argmax(dim=0)。I confused again。 It appears to me the logic is oppersite about this two. When go to the nd-array. Things seem to go more complex。Is there any way I can get it crystal

[@deeplizard]

你好刘新新 - The key is to remember that the reduction is element-wise unless we are working on the last dim. The last dim is always just an array of numbers (the scalar components). Suppose we have this tensor: t = [ [1,2,3,4], [0,0,0,0] ] How do we calculate t.sum(dim=0)? We do it by summing corresponding elements across dim-0. There exist two elements running along dim-0. This one: [1,2,3,4] And this one: [0,0,0,0] Since these are arrays, the sum must be element-wise. This gives us the following: [1+0,2+0,3+0,4+0] = [1,2,3,4] Now, with the argmax(dim=0) operation, it's the same only the operation is different. This time the operation is not sum, but argmax. Argmax gives us the index location of the max value. We have two possible index values, 0 or 1. This is because we are working with two elements. t[0] = [1,2,3,4] t[1] = [0,0,0,0] When we do the argmax, we argmax corresponding elements (element-wise) across dim-0. [argmax(1,0), argmax(2,0), argmax(3,0), argmax(4,0)] = [0,0,0,0] Let me know if this helps and if you still have questions. I think the best way to make this crystal is to create some random examples and practice working them out. It is tricky.

[@tingnews7273]

@@deeplizard Thank you for your reply. I spend some time to watch the video again and again ,read your reply and try it in code and calculate by pen. I think I get some intuition . But not that clear. I take it as a little progress . I think the most important thing to understand it is element-wise. base different dim . different element. I can figure out the argmax(dims=0) , use this perspective is element is a tensor ,so when we go through every scalar in the tensor , we ask the index of the tensor element. I found in the deep learning , if you want to understand more . you must get the basic ideas clear. I have doing the job for about 1 year more. Still so many basic concepts not that clear. Thank you very much.

[@thepresistence5935]

4:53 best moment :)

[@donfeto7636]

love this

[@Nikage23]

I wish I'd found this explanations earlier

[@marwaneboudraa9917]

best explanations ever

[@onePunch95]

tensor([1, 2, 3, 4]) >>> t.sum(dim=1) tensor([10, 0]) In this case, why don't we get [10]?

[@deeplizard]

Hi Satwik - I can't be sure because I cannot see the value of the tensor t. Note that the tensor shown tensor([1,2,3,4]) is a single dimensional tensor.

[@gactve2110]

I would drop the bassy sound effect when you type, it's anoyying when usinf earphones. Except that, Great content! Keep it up!

[@deeplizard]

Hey gactve - I appreciate the straight forward way you described the issue with the sound effect. Thanks! That sound has been removed in newer releases. 😊

[@takeyoshix]

"You gotta step up your game, mate!"... care to share the link to that hilarious woman?

[@deeplizard]

Link to her video in the description :D (Name: Sorelle)

[@heller4196]

4:30 LMAO

[@deeplizard]

Hey Omkar - 🤣 Thanks for your comments!

[@ennardzhao1871]

stop putting scary clips into your tutorial videos!!!!

[@jacoblawrence1138]

the sound you have added when typing the code i mean like a bhooo...... kind of sound is really annoying and it gets irritated when watching the videos for continuously. just take this as a suggestion from a viewer.

[@deeplizard]

Thanks Jacob. Are you using headphones? Which ones?