I think the FLOPS for a CNN filter would be (H' x W' x D) x ( 2 x K x K x C). For each pixel in the output feature map of size (H', W', D), we compute the dot product of a (K x K) kernel across C channels.

I think the FLOPS for a CNN filter would be (H' x W' x D) x ( 2 x K x K x C). For each pixel in the output feature map of size (H', W', D), we compute the dot product of a (K x K) kernel across C channels.