Consistent ordering of cached data files between deeplearning and Great Lakes

[timwhite0]

os.listdir() does not necessarily list files in the same order on Great Lakes as it does on the deeplearning server, even if the contents of the directory are the same in both places and the same random seed is set in both places.

This behavior is documented, e.g., here and here.

This issue is relevant for any BLISS case study that trains a network on Great Lakes and then evaluates the trained network on the deeplearning server. In CachedSimulatedDataModule, files are assigned to train/val/test based on the order in which they are read in by os.listdir(). So because of the behavior mentioned above, one could obtain different train/val/test splits on deeplearning and Great Lakes even if the directory contents and random seed are the same in both places.

The proposed fix just wraps sorted() around the list of file names.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 76.15%. Comparing base (1baf4d4) to head (6dd65d9).
Report is 1 commits behind head on master.

Additional details and impacted files

@@           Coverage Diff           @@
##           master    #1075   +/-   ##
=======================================
  Coverage   76.15%   76.15%           
=======================================
  Files          22       22           
  Lines        3091     3091           
=======================================
  Hits         2354     2354           
  Misses        737      737           

Flag	Coverage Δ
unittests	76.15% <100.00%> (ø)

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[jeff-regier]

Nice!