Lecture 11: Fractionally strided layer a.k.a Deconvolutional layer

[@saibharadwajasb]

Awesome video, you really broken down the mess into simple steps..... Thank you

[@thatipelli1]

Fantastic effort in breaking the topic into simple concepts. Thanks for all the efforts and keep this up.

[@learndlcodetf5331]

Thank you very much!

[@write2bishwarup]

3:56 - (3x3) filter applied on (4x4) with stride 2 is asymmetrical. Are you applying a ZeroPadding(1)? for (2x2) feature map with no padding the stride has to be 1.

[@learndlcodetf5331]

Hi, sorry for the confusion. Yes, I am applying zero padding else a stride of two won't be possible

[@sharansrivatsa210]

You need more subs! This is amazing.

[@saurabhsaxena2657]

Awesome video Vikramank ! Keep up the good work

[@learndlcodetf5331]

Thanks Saurabh :)

[@parikshitagarwal3901]

great efforts lead to awesome video, thanks.

[@anandsaha01]

At 4:05, the math doesn't add up when you calculate the size of feature map with a stride of 2. I think it should be stride of 1. i.e. 2 = (4 - 3) / 1 + 1

[@learndlcodetf5331]

Hi Anand, My apologies. I didn't mention the padding layer here. There is no way I can have a stride of 2 for a kernel of size 3 convolving an input of 4x4 without padding layer. After adding padding layers, due to the stride of 2, the size of feature map will naturally half the size of the input. Check out this awesome paper: A guide to convolution arithmetic for deep learning, Vincent Dumoulin and Francesco Visin : arxiv.org/abs/1603.07285

[@bijonguha2299]

5.01 explanation is needed on how you converting 3*3 to 4*16 .

[@amyhu7547]

7:33 the kernel size is 5 not 4, right?

[@learndlcodetf5331]

Hi, it is 5. My bad! Don't know how I drew 5 cells and wrote as 4. Sorry for the inconvenience

[@miftahbedru543]

Thank you for the post ..The part where the 3x3 matrix is converted to 4x16 by zero padding is a bit unclear, especially the part where raws of A are arranged with zeros inserted after each raws ... The zero padding is not the usual padding we do in CNN ? how do we get to 4x16 by zero-padding 3x3 weights? the first raw followed by one zero , 2nd raw followed by another zero , 3rd followed by another zero .. isn't this producing a raw of length 12?

[@trendzingo3164]

watching it exactly after an year!

[@learndlcodetf5331]

cool!

[@Suresh-sj9nk]

Hi thank you very much. how can i upsample an image of 64x64x3 to 128x128x3. please tell me code. What is Deconv2D in keras? How to use it?

[@learndlcodetf5331]

hi, you can use the tf.layers.Conv2DTranspose() function to upsample the image. for the image specification , the following should be your code output=tf.layers.Conv2DTranspose(3, kernel_size=(2, 2), strides=(2, 2), padding='same')(input) Deconv2D is mistakenly used in some cases. I think keras.layers.UpSampling2D(size=(2, 2), data_format=None) is the correct function for upsampling

[@Suresh-sj9nk]

for example I want to double the size of image or triple the size of an image. How to change the code accordingly

[@learndlcodetf5331]

play around with the strides. strides=(2,2) -> double (128x128x3) strides=(4,4) double of previous one and so on. It's basically upsampling by a factor of 2!. Remember this, while performing downsampling, the strides play a crucial role to determine the amount of downsampling. A stride of two will downsample the input into the half. Similarly, for upsampling, a stride of two will upsample by a factor of two. Hope this is clear

[@Suresh-sj9nk]

Thank you very much!

[@siddharthsvnit]

3:30 Inverses only make sense for square matrices so this interpretation here is factually wrong

[@learndlcodetf5331]

Hi Siddharth, yes are you right. That's why I assumed it is an orthogonal matrix. The idea behind is to give an intuition about the Fractionally strided convolution and since it is non-trivial I took this case because it is easier to explain.

[@kislaykunal8921]

deconvolution demystified.

[@learndlcodetf5331]

Hi Kunal, thanks for your feedback. With your permission, can I change the title of this video based on your comment?

[@kislaykunal8921]

yeah go ahead.

[@yanshen9318]

Good lecture, but with strong accent

[@learndlcodetf5331]

thank you !

[@S14S88]

wtf