Is there any way I can get the gradient value of the alignment loss

[SSSRgo]

I am going to analyze which features (neurons) contribute mostly to the alignment of multi days data. I know the a significant difference between UMAP and AlignedUMAP is the added alignment loss term. I want to get the gradient of the alignment loss term and then compute the contribution of neurons to the alignment loss mostly.

[lmcinnes]

I think you'll find that one is buried in the code itself. If you look at the aligned umap layout functions in layouts.py you should ebe able to find the innards. I don't think it will be easy to just pull out a term.