Open sophiamaedler opened 5 months ago
This issue arises from different version of scikit-image being installed.
We've observed discrepancies between 0.19.3/0.20.0/0.21.0 and 0.22.0/0.24.0. 0.22.0 has given consistent results over different environments and machines. 0.19.3 has given two different set of results.
Upgrading to 0.22.0 results in problems with installing ashlar into the same environment as with the latest release it requires < 0.20.
As ashlar has been implemented to also use scikit-image versions > 0.20 I think we can close this and pin scikit-image to >= 0.22.0.
Long term it would be good to set up a test for this to ensure that any future changes in scikit-image do not break our reproducibility.
To do's before closing:
Running the same notebook on the same data results in slightly differing number of nuclei between different environments/machines.
These results should be fully reproducible. The source of the discrepancy needs to be identified.