I need to unstride a tensor, i.e. input a tensor, insert zeros between all items and return a tensor twice as large (in the spatial dimensions). Is this possible to do?
Note I’ve found a hacky workaround by using max unpooling but it looks like a mess.
I’m working with the TensorComprehension integration and need such a function to backprop a strided convolution as this part can’t be done in TC.