I’m told to have one dimension less than expected since the 7th block. I always use the previous’ 0-th as the expensive part and the opposite’s 1-th as the contraction part, while the problem persists if I use the previous’ last layer as the expensive part. Could you help me? Thanks in advance.
The point is that the downsampling layers have two outputs, so you have to index the output to select which one you want. But notice that the upsampling layers are different: they only produce one output. If you index the output of one of the upsampling layers, that’s a mistake: it just means you strip off one dimension of the tensor with the unhappy result that you see.
Excellent solution! Can’t thank you enough.