royerlab/ultrack

Reproducing 2d stardist example conda env

Closed this issue · 4 comments

Trying to create the GPU conda env from the stardist_2d example provided I get:
(OS: Windows 10 64-bit, CUDA 12.3)

ResolvePackageNotFound:
  - rapidsai::cucim
  - coin-or-cbc
  - conda-forge::tensorflow-gpu=2.11.0

Hi @folterj, thanks for raising this issue.

I don't have a Windows machine to test this, so I can't provide a direct solution.
However, we can tweak the dependencies.

I recommend:

  • removing cucim and cupy from the environment file (they are not a hard requirement).
  • removing coin-or-cbc and making sure you have gurobi license, you can get a free academic license.
  • loosening the tensorflow dependency.

Note that some operational systems have issues installing TensorFlow and torch (track dependency), so a different example might be easier to run.

Hi @JoOkuma, thank you for this detail.
Reproducing this on a Windows machine appears complicated indeed. With your suggestions I was able to eventually create a conda env with both tensorflow (non-cuda) and pytorch (an old version for some reason), and added ultrack. It then it fails at the tracking. Perhaps this is related to the Gurobi license as it fails very close to where the example notebook shows output regarding the license. Obtaining & configuring the license is not trivial, and unclear how this is subsequently made available runtime.

Hi @folterj,

Gurobi has really helpful resources, this one might help you out.

If you're in a non-academic institution, you could try installing the other solver without conda; their instructions is here.

Note that CBC is much slower than gurobi.

Cheers,

Thank you, I was able to configure this now and set up the license etc.
Perhaps some more detail on this could be added to the example/readme?