prediction on out of sample

[AleBitetto]

Hi,

I'm trying to fit the model on a dataset with densmap.fit() and then to evaluate the embedding on a new dataset with different number of rows with densmap.transform() but I get an error of "index out of bound" for the row index. Is it a bug or the algorithm, by definition, is not allowed to do out-of-sample prediction?
Thanks

[anashwin]

Hi,

Great question. Unfortunately in its current iteration, UMAP isn't parametric and thus cannot do out of sample transformation. You currently have to run fit_transform on the entire dataset rather than fit.

The current version of UMAP implements ParametricUMAP which can do out of sample prediction, and also allows you to use densMAP parameters, so you can do density preserving parametric UMAP.