Christopher Tessum

Yes, I mean the equivalent of `S2Polygon::InitToIntersection`. Thanks!

There is a port of (one of) the Martinez-Rueda algorithms [here](https://github.com/ctessum/polyclip-go). However, that one is for planar geometry, not sure that it would work here?

In case it's any help, I get the same error when I remove the final `S` layer from the neural network: ``` julia dudt = Chain( Dense(2,30,tanh), Dense(30,10), #S(10,2) )...

Thanks for your help with this! Here is another try at a reproducer: ```julia import Flux, DiffEqFlux, OrdinaryDiffEq using Test # Create a layer that is fully trainable struct TrainAll...

For what it's worth, a workaround is to make the non-trainable parameter `b` a global variable, so this: ``` julia b = ones(2, 2) struct TrainSome a end (s::TrainSome)(x) =...

Thanks for the responses! I think we can implement our system as a `NonlinearSystem` for now, but we would be a user of this feature if it ever gets implemented...

Sure, we can share the system once we have something put together.

That makes sense. I think, though, there still might be a rationale for separating the color map from the heat map, as having a standalone color map could enable functionality...

The functionality in question here is assigning colors to data points. Currently, HeatMap is the only part of the package that does this, but there are a bunch of potential...

Having Legend return a plot.Plot allows a lot of customizability with tick marks, titles, etc that would otherwise have to be duplicated, but it might be possible with some refactoring...