Closed odow closed 2 months ago
The other option would be to add a reduced_space::Bool
flag to each predictor.
I guess the question is: how fine-grained to people expect these to be? Do I want all layers in a NN to be reduced-space? Only some? If so, which ones? All Affine but not ReLU?
A third option would be to add separate predictors, so something like ReducedSpace{ReLU}
.
Yeah, I think I like https://github.com/lanl-ansi/MathOptAI.jl/pull/72 more.
Closing in favor of #72
One option for #70