CNN Tensor Shape Explained - Convolutional Neural Networks and Feature Maps

[@deeplizard]

Check out the corresponding blog and other resources for this video at: deeplizard.com/learn/video/k6ZF1TSniYk

[@cassiusnolan2851]

pro trick: you can watch movies on flixzone. Me and my gf have been using it for watching a lot of movies these days.

[@jonahfinn6852]

@Cassius Nolan Yea, I have been watching on Flixzone for months myself :D

[@ravihammond]

I'm loving your short clips at the end of each video. Great touch!

[@upiferico]

I rarely comment. But your content is so good I had to pause and make you know that!! thanks :)

[@lingshuaikong5644]

I just cannot be more grateful to you for providing such an clear and thorough and vivid illustration of the concepts in Deep Learning. Things you've down really make sense and help a lot of people who want to enter the Deep Learning field lacking of ways to kick in like me. Truely Thanks! May you good!

[@deeplizard]

Thank you for taking the time to let me know! 😊 Chris

[@reefcrazed2070]

I really like these snippets you put through your videos. They have a purpose and are well done.

[@deeplizard]

Hey Reefcrazed - Thank you for your comment! Appreciate your feedback. It is like fuel!

[@richarda1630]

you guys do such an incredible job of explaining, and the overall why videos at the end get the viewer all inspired and pumped up for more! :)

[@mohammedrakib3736]

Crystal 🔮 clear! Loved the animations! Finally understood how color channels get converted to feature maps using filters while convolving.

[@vaishnavideshpande2233]

I can watch these tutorials all day!

[@Vikram-wx4hg]

Lovely explanations!

[@rafaeloyamada2712]

Dude, such a great channel! I am even watching videos about topics that I already am familiar with, thanks for that

[@srijaljoshi3421]

Your channel, in addition to your website, is pure value! Just wanted to drop by and say that please don't stop making new content!

[@deeplizard]

Thank you, Srijal! Glad you're enjoying the content!

[@Brahma2012]

The insights to deep learning in this video is amazing. Explanation of CNN Tensor is lucid and clear!!

[@philipweslow565]

{ "question": "In a typical input tensor to a CNN, the first axis usually denotes what type of information?", "choices": [ "The feature map.", "The number of convolutional filters.", "The batch size.", "The color channel." ], "answer": "The batch size.", "creator": "PW", "creationDate": "2019-07-08T18:13:21.102Z" }

[@deeplizard]

Also, added "for pytorch" for this one too. I've just added them to the quiz. Let me know if you see them and if all is good. Thanks again! 🧠🚀

[@sqliu9489]

love this channel !

[@sahbatahsini4725]

Your channel is just AMAZING! full of comprehensive, updated and high-quality videos. Good job man!

[@deeplizard]

Thank you so much Sahba! Really appreciate that! 🧠

[@Sikuq]

That filter animation - wow.

[@Nikage23]

this channel is a treasure!

[@tianjoshua4079]

Explained masterly. Deeply appreciated!

[@yourxploit7858]

These videos are great

[@DanielWeikert]

Thanks so much. Very well explained and a perfect continuation from the last video. Again, you guys are simply awesome

[@tymo3391]

very well presented content!

[@philipweslow565]

{ "question": "What is the length of a typical input tensor for a CNN?", "choices": [ "100", "28", "4", "1" ], "answer": "1", "creator": "PW", "creationDate": "2019-07-08T18:01:41.732Z" }

[@deeplizard]

Hey Philip - Thanks for the question. For this one, did you mean the "shape" of an input tensor to a CNN has a length of "4"?

[@joebender9052]

I wish I'd had this explanation the first time I tried making a CNN in PyTorch.

[@Sikuq]

Great. Thank you.

[@philipweslow565]

{ "question": "In a typical input tensor to a CNN, which two axes denote the height and width coordinates of a particular pixel?", "choices": [ "The first two.", "The last two.", "The middle two.", "There is no set convention." ], "answer": "The last two.", "creator": "PW", "creationDate": "2019-07-08T18:08:58.979Z" }

[@deeplizard]

The order of the axes is arbitrary, so it might be different depending on the framework. For example, Keras defaults to having the channels last. The question is still good. I just added "for PyTorch". 🧠🚀

[@anastassioskaragiannis8554]

{ "question": "Considering a tensor with the following shape [3500, 3, 1920, 1200]. How many samples do we have ?", "choices": [ "3500", "3", "1920", "1200" ], "answer": "3500", "creator": "AKTasos", "creationDate": "2020-02-10T09:05:52.275Z" }

[@deeplizard]

Thanks, Anastassios! Just added your question to deeplizard.com/learn/video/k6ZF1TSniYk :)

[@ferielferiel5006]

Finally found what I was looking for

[@mck1ing]

A question about 5:22, if your CNN kernel is taking in nxn blocks of the image does that mean it can mix and match between different "A3"s of your input tensor? because if you A2xA3 is 28x28 I would imagine each A3 represents a horizontal or vertical strip of the image.

[@bedeamadi9317]

{ "question": "Which of these represents a PyTorch CNN tensor input for a grayscale image", "choices": [ "[3200,1,32,32]", "[3200,3,32,32]", "[3200,32,32,3]", "[3200,32,32,1]" ], "answer": "[3200,1,32,32]", "creator": "B.A", "creationDate": "2022-01-17T09:07:56.413Z" }

[@sandeepladi5929]

Can you please create a video on create custom keras layer where we can replace the convolution operation with the other operations like Euclidean, sigmoid, Gaussian etc.

[@ליהיימליך]

wow, that's a great video!

[@GauravSharma-ui4yd]

hey can you send the link to nvidia talk presented at the end

[@deeplizard]

Hey Gaurav - I added it to the description: kzbin.info/www/bejne/fIGwZKGjocaBZpI You can also find other newer talks by going to Nvidia's youbute channel. Search for nvidia key note.

[@kritikaprasai1121]

So if I am using a 700*460 pixel image then input shape of my layer will be (3,700,460) and for batch size I can just take some random value based on what kind of batch I want(stochastic,mini batch)??

[@deeplizard]

Hey Anukritika - That's correct. 🚀

[@LakshmiSivaKarthikPadala-y9h]

meanwhile I was waiting for 2012 year as "end of the world", Alex did big bang in AI!

[@hrehann]

Hi, the convolution of a RGB image (32 x 32 x 3) with filter (3 x 3 x 3) results in a feature map (28 x 28 x 1). I see that the feature map is 2dimensional...Having seen a lot of lectures i still have a confusion whether the feature maps are grayscale or RGB. Since during the convolution each channel of image and filter are convolved separately and then added that results in a 2d feature map...but still addition of R+G+B channels results in a colour image. any help will be appreciated

[@deeplizard]

Hey hrehann - The feature maps (aka output channels) are arbitrary and depend on the number of filters inside the layer. In your example, the number of output channels is assumed to be 1. However, the number of output channels could be any number. The number of output channels depends on the network’s architecture. The way to think about this is to realize that the feature maps (output channels) are more abstract data structures (opposed to RGB and Grayscale). This is why we stop referring to images after the input layer and start using a different word (feature maps). We touch on this a bit in the series here: deeplizard.com/learn/video/IKOHHItzukk If you are interested in visualizing feature maps, this may be helpful: distill.pub/2017/feature-visualization/

[@hrehann]

@@deeplizard Hi, thanks for your response. Finally things are clear now!

[@louerleseigneur4532]

merci merci

[@magelauditore333]

Please Please start tutorial on RNNs too atleast shapes part

[@akashthoriya]

can you please elaborate, what is shape?

[@deeplizard]

Hey Aakash - The shape defines the structure of the tensor. It specifies the number of axes and their lengths. Have you seen these two yet? deeplizard.com/learn/video/AiyK0idr4uM deeplizard.com/learn/video/fCVuiW9AFzY They come before this video in the series and have many details on shape. Let me know if these help.

[@nsu3274]

can i say something *'gigantic love from india'*

[@yashas9974]

The person who downvoted was so excited that they had their head turned upside down when they hit the dislike button.

[@deeplizard]

Haha. Love it. 🙃 🤣

[@felixt2920]

Good explanation, but the problem is im really bad in english

[@deeplizard]

You can use the text version on the blog which might be easier 😃