Where convolutions have been the doing great at what it does, involution symmetrically inverts the inherent properties of convolutions. Where
convs are spatial-agnostic and channel-specific operations,
invs are spatial-specific and channel-agnostic operations.
My take on Involutions: GitHub - ariG23498/involution-tf: TensorFlow implementation of involution.
Here one can find the
Involution Layer which has all the necessary code to build the kernel dynamically and also apply it on the input feature space. One can also pick the code up and apply the layer to any
tf based architecture.