Alternative network architecture for Parametric UMAP

[16thomja]

I am projecting 10s of millions of 200-dimensional text embeddings into 3 dimensions with the intent of clustering in the reduced space. Under my memory constraints, training PUMAP on a subset of these embeddings appears to be the best option. Can anyone suggest alternative NN architectures for this purpose? I am unsure whether the default 3-layer 100-neuron is sufficient.

[Mirage-deux]

maybe an auto-encoder of sorts comes to mind but not rlly sure