mmilanifard
commented
3 years ago We don't support non-uniform lattices. The closest potential solution is to use a calibrated lattice structure, were you have piece-wise linear functions feeding into a lattice. The granularity of the overall composition changes over the range of inputs, but the non-uniformity is coordinate-wise independent across inputs. In 2d it looks something like this:
Is it possible to use nonuniform lattice cells with tensorflow lattice, i.e., to set lattice node unevenly distributed along each feature like the idea in the paper Nonuniform Lattice Regression for Modeling the Camera Imaging Pipeline.