BirthDeathProblem.solve Error: tp1[8, 9].set_graph_xy(dfs[8, 9].astype("float"), t=100.0)

[Yulab2022]

Dear contributors,

Thank you for providing such an excellent tool!
However, I’ve recently encountered some issues while using it. When following the code from the section “Disentangling lineage relationships of delta and epsilon cells in pancreatic development with the TemporalProblem,” I was able to successfully run the previous steps. However, after running .set_graph_xy(), an error occurs when I run .solve(). Interestingly, if I skip set_graph_xy, .solve() runs without any problem. Could you please help me understand what might be causing this issue?

tp1[8,9].xy is the result after running .set_graph_xy(),while tp0[8,9].xy is not.

I’m looking forward to your response. Below is my Python ipykernel version, which should not be the issue.

[MUCDK]

Hi,

Which GPU are you running the code on? I assume it’s an Out of memory error , as the geodesic cost required squared memory , whereas the default . The squared Euclidean cost, can run in linear memory complexity.

Can you try with larger memory?

[Yulab2022]

Dear contributors,
Thank you for your previous response.
However, I'm still encountering an issue. Our server is currently using driver version 525.147.05, which only supports up to CUDA 12.0.

However, moscot by default installs jax==0.6.2, and during execution, I get errors indicating that JAX cannot detect the GPU.

I suspect the issue might be due to the JAX version being too high and incompatible with the CUDA version on our system. However, I’ve tried several different versions of JAX and encountered the same problem each time.

I will continue working on resolving this issue.But what I want to know now is whether the .set_graph_xy() step is essential. Can we still get results from the .solve() step even if we don’t run .set_graph_xy()?

Or could you share the versions of CUDA, cuDNN, and JAX that you use when running on GPU?

[MUCDK]

Did you install jax with gpu support, see here: https://docs.jax.dev/en/latest/installation.html

As we write in the tutorials, we recommend not using the geodesic cost (i.e. set_graph_xy() ), but the Squared Eucl. cost as a first try. Hence, you should run the code without set_graph_xy as a first try.

[batalha23]

Following up on this issue, I'd just like to point out that it is not clear from the tutorial what are the steps required when one intends to use the square Euclidean cost instead of geodesic, neither the use cases where it would be more appropriate to use one of the other.
Could you please clarify this? Thank you very much!