Conv2DTranspose

[keilalopezbl]

Hi Jocicmarko, I have this question...

What is exactly doing the Conv2DTranspose layer in your code? I assume that is increasing the dimension by a factor of 2 ( I mean, performing the opposite operation that Maxpooling layer does). Am I correct?

[mrkolarik]

Hi there,

you are partially correct - Conv2DTranspose is something like the oposite of convolution layer - you can call it deconvolution. Increase of the dimension by factor 2 is done by parameter strides=(2, 2) at each transpose layer, not by the mathematical operation transpose convolution itself. Nice visualization is here: https://datascience.stackexchange.com/questions/6107/what-are-deconvolutional-layers. If I'm wrong, somebody please correct me :)

[keilalopezbl]

@mrkolarik Thanks for sharing!

[nabsabraham]

Hi all,
I don't understand why after the Conv2DTranspose operation we are again using Conv2D - this will keep the feature map size but what is the point of doing it? Is it learning in the upsampling process too?

u6 = Conv2DTranspose(256, (2, 2), strides=(2, 2), padding='same') (c5)
u6 = concatenate([u6, c4])
c6 = Conv2D(256, (3, 3), activation='relu', padding='same') (u6)
c6 = Conv2D(256, (3, 3), activation='relu', padding='same') (c6)