When using transposedConv2dLayer, how do you calculate the size of the output given an input?

Let's say you have an array of size [a b c] as input to a transposedConv2dLayer. What would be the output based on the stride, crop, and filter-size numbers?

Answer 1

[Matt J]

I believe it would be

newsize = [stride.*[a-1,b-1] + filtersize - 2*crop, numFilters]

[Onur Kilic]

Thank you for the answer. What is the reason for why the third element is just the number of filters?

The reason for my question is that I am trying to understand how the 5 transposed convolution layers take an input of size [4 1 1025] and put out an output of size [1201 1 1] in this example:

https://www.mathworks.com/help/signal/ug/generate-synthetic-pump-signals-using-conditional-generative-adversarial-network.html#GenerateSyntheticPumpSignalsUsingCGANExample-3

[Matt J]

What is the reason for why the third element is just the number of filters?

Each filter is used to generate one channel of the output array. In other words, the number of filters is always the number of output channels

[Onur Kilic]

Thanks. Also, it seems the following edited formula gives the correct sizes for the linked example:

newsize = [stride.*[a-1,b-1] + filtersize - 2*crop, numFilters]

If you agree with this formula, can you edit your first answer so that I can accept it? Could be useful for others.

[Matt J]

Yes, I agree.