I was looking to create a multi layer perceptron using tf lattice constraints of monotonicity on inputs. Based on the paper Monotonic calibrated interpolated look-up tables, it seems that lattice layers should only work with single layer in order to apply constraints. However, I was also going through the Deep Lattice Networks where ensemble lattices have been used. Is there a publicly available implementation of this in tf? All of the tutorials in tf lattice github seem to be a single layer constraint optimisation on features.
2 Answers
It is implemented as the random tiny lattices layers (rtl_layer). You may stack them on top of each other as to create more layers. Notice that you need to mark every layer except the bottom one monotonic with respect to all their inputs in order to make sure the monotonicity gets passed to your final prediction.
Related