Dominik Klein

Hi Marco, Thanks for your response. I might be wrong but reading equation 24 in https://arxiv.org/pdf/1910.12958.pdf the term multiplied by epsilon in the entropic penalization should be non-positive because `f_i...

Hi Marco, Thanks, yeah, I spotted my mistake and your reasoning makes perfect sense!

Hi, Thank you very much for your response. I tried to replicate the instructions but still run into the error RuntimeError: (PreconditionNotMet) Cannot load cudnn shared library. Cannot invoke method...

I think you can establish norms here. I just take the combosciplex dataset of pertpy. Here, we have two obs-columns `Drug1` and `Drug2`

> https://github.com/theislab/moscot/actions/runs/9221997646/job/25372233453?pr=662 complains that it can't find `diffrax` with the minimum version from `ott-jax[neural]` on python 3.8, and I'm not sure what to make of that since `ott-jax` tests against...

should we make it more flexible then? Here, the control should always be the source.

Hi @AlinaKurjan , Have you tried sparsifying the matrix using https://github.com/theislab/moscot/blob/ba231ea4f7511caaa1915e2b7e28e88d709d45ad/src/moscot/base/output.py#L170 / https://moscot.readthedocs.io/en/latest/genapi/moscot.backends.ott.OTTOutput.sparsify.html#moscot.backends.ott.OTTOutput.sparsify ? As Marius mentioned, CellRank needs to materialize the transport matrix, hence it becomes quadratic in memory....

Hi @AlinaKurjan , Thanks for this, it's super helpful feedback. Would you mind copy-pasting the code which you tried? This would be easier for us to track down the error.

Hi @AlinaKurjan , It is literally the code you used? There might be an error as you use `sparcify_kwargs` instead of `sparsify_kwargs`. Hence, the `sparsify_kwargs` are not passed. I will...

> @AlinaKurjan there are 2 batch sizes, one in `sparsify` and the other one in `problem.solve`. > > The one in the `solve` method only materialized to point cloud of...