search

SpatialHackathon / SpaceHack2023

MIT No Attribution
16 stars 3 forks source link

Meringue config syntax fix #317

Closed shdam closed 2 months ago

shdam commented 4 months ago

Not sure if this will solve the other issue. But Meringue seems to run on the test set.

peicai commented 4 months ago

Not sure if this will solve the other issue. But Meringue seems to run on the test set.

Yep, I fixed this on the cluster, but thanks! I guess you used the calculated dimensionality reduction matrix instead of Meringue's preprocessing step if you used the current version of the meringue_optargs.json file.

shdam commented 4 months ago

I ran using a test script, which ignores the optargs.

Reading the error message again, it seems the answer is in this line: QH6239 Qhull precision error: Initial simplex is cocircular or cospherical. Use option 'Qz' for the Delaunay triangulation or Voronoi diagram of cocircular/cospherical points. Option 'Qz' adds a point "at infinity". Use option 'Qs' to search all points for the initial simplex.

Which dataset causes the error? - Perhaps the spots have a non-default geometry

peicai commented 4 months ago

The LIBD data... One option is that I could modify the getSpatialNeighbors function by testing different options, like 'Qz' or 'Qs'. Alternatively, I could apply Meringue to another dataset to see whether the issue is due to the spot resolution. However, in the SpaceFlow paper, they successfully applied Meringue to the LIBD data.

shdam commented 4 months ago

Hmm, the test script also uses LIBD. Are you running the script as it is on the native branch?

peicai commented 4 months ago

Yes, I ran the script as it is on the native branch, only fixing the syntax in the config and updating the optargs. It might be related to the SLURM cluster setup, as I tested it on JupyterHub and it works.

Jieran-S commented 2 months ago

Hi, forgot to mention here but I fixed the meringue bug by removing the identical z-axis of the dataset (which costs the breaking down of the Delaunay function) in #319. I will close this PR along with issue #316