Do the last FC layers get out of the Hyperbolic space? about hyperim HOT 3 CLOSED

The final MLP uses the similarity scores to do the prediction part so its parameters are in the Euclidean space and are updated via vanille optimization methods.

Only the word embeddings and label embeddings (and the bias in the hyperbolic RNN/GRU) are located in the hyperbolic space and are updated with Riemannian optimization methods.

from hyperim.

Does Riemannian Optimizer distinguish Euclidean parameters with Hyperbolic parameters?
Because only one optimizer is used as:

from hyperim.

For parameters in the hyperbolic manifold, RiemannianAdam is applied, while vanilla Adam is applied to other parameters in the Euclidean space.

from hyperim.