Reproducibility issue of MRVI

[Nusob888]

Dear Yosef lab members,

Thanks for MRVI, I have been trying it out finally.

Unfortunately I am experiencing some reproducibility issues. I first noticed this when I accidentally refreshed a Jupyter notebook and re-ran the code. The embeddings were very different.
Since then I have tried a numerous amount of methods to set seed and I cannot seem to produce anything consistent.

This would be a big issue for reproducibility if the u embeddings were ever to be used for clustering (which I had hoped to do) and I presume it would also effect how the abundance and differential expression analysis perform?

I have attached 4 htmls of my code run on the PBMC 3K dataset.

Any updates on this would be great to help me keep track of the tool, as I would very much like to use it, but have concerns around this issue.

SCVI_MRVI_reproducibility_scvi_1.pdf
SCVI_MRVI_reproducibility_mrvi_2.pdf
SCVI_MRVI_reproducibility_mrvi_1.pdf
SCVI_MRVI_reproducibility_scvi_2.pdf