I understand the concept of convolution.
Let's say that my input dimension is 3 x 10 x 10
And if I say that I will have 20 activation maps and a filter size of 5, I will end up with 20 different filters for my layer, each with the dimension of (3 x 5 x 5)
My output will therefor be (20 x ? x ?)
. I Put a "?" there, because it obviously depends on the filter stride etc.
Now I wanted to implement deconvolution but I am stuck at the following point:
For the following questions, let's assume that the input size for the deconvolution is (5 x 8 x 8)
,
- If we think about a filter in 3 dimensions. Can I choose any depth for the filter?
- How would the effect of the amount of filters (amount of activation maps) work with deconvolution? Do I only have one filter?
- How does the input depth (5) come into play. Would the output depth be equal to
(filter depth) * (input depth) ?
I am trying to find the symmetry to forward convolution but I do not understand how to use the amount of activation maps in deconvolution.
I am very thankful for any help.