ModulatedDeformConv2dPack: DCNv2 breakdown

[janthmueller]

Can anybody explain the necessary matrix multiplications for the DCN layer used in this repo? Maybe using pytorchs fold and unfold.

[janthmueller]

To be more precise, I need the new input $X$ after performing the bilinear interpolation where i only need to perform the convolution and "masking". Can I utilize some function within a package to get $X$ with respect to the computed offsets?