Assume that I have a CNNNode with the dimensions h1xw1xchannel1 (kernel sizes do not matter). This node is "converted" by a fully connected node to the dimensions 1x1xchannel2. The condition is that h1xw1xchannel1 >> channel2. Afterwards, there are some batch normalizations, ReLU and one or two more fully connected layers.
Is there any performance benefit to do the reshaping from 1x1xchannel2 to channel2x1x1 or doesn't it matter at all?
Is there a memory benefit?
Is there any performance benefit to do the reshaping from 1x1xchannel2 to channel2x1x1 or doesn't it matter at all?
Is there a memory benefit?