I'm currently studying about Resnet and I have question in downsampling.
In the paper, it is written that ( When the dimensions increase, the shortcut will perform identity mapping with extra zero entries padded in increasing dimension, or, use 1x1 convolution to match the dimensions.)
so, let us say the 'shortcut' has size (channel=50, height=56,width=56) and would like to add on to following feature map which is sized ( 100, 28, 28 ). Then, since they have different channel size and height, how can we add them using down sampling?
For example, if we use (1x1 with channel 100) Convolution, we can make the short cut to (100,56,56),but still since the height and the width is different we cannot add together.