How to change features size without padding in CNN?

1

Hi,

I’m using a 2d CNN. My input size is a square image of 20x20 pixels and my desired output has size 400x3.

The most common thing I see is to introduce padding over the convolutional layers in order to reach the desired output size and let the nn train.

The output I get in this way is characterized by lots of zeros (sparse matrix) and obviously the performance is very low. So what else can I do? I tried to introduce more layers, each one characterized by smaller padding values but it doesn’t improve…

any hint will be really appreciated, thanks!

2

Is your input image actually 20x20x3? Why would you need an output to be the same size as your input?

3

My input image is just 20x20 and they have not the same size, since I my net input is 20x20 and the output is 400x3.

thank you