Differences with Python version? about umap

[tkonopka]

Hi Maarten

That seems odd. To answer the question about synchronizing the parameters - the object 'umap.defaults' has all the arguments and values. The argument names match with those in umap-learn, so you can construct a python command setting the arguments one by one. But it sounds you have done that already.

The python group released a new version of umap-learn recently. I haven't checked yet how well the R package interfaces with that. It's a to-do item and I'll update if necessary. Curious, what version are you using?

(You closed this now - did you manage to resolve it?)

[slagtermaarten]

Thanks for your message! I made an embarrassing mistake in transferring the raw data over from R to python: the sample annotation got shuffled during a data merging step and as such samples were not correctly labeled in python's UMAP plot. I finally figured out that this was the problem after having painstakingly checked all other steps (was assuming my rusty python skills were the source of the problem). The UMAPs looks highly similar now, just rotated about 90 degrees from each other.

[slagtermaarten]

To quantitate the semblance between the two implementations on my data, I combined the 4 coordinate axes (2 from each UMAP) into one matrix and did a PCA on this matrix. I found the first two PCs to describe over 99% of all variance, consistent with rotational symmetry.

[tkonopka]

Glad it worked out. Good luck with your analysis