parallel sequentialfs, reproducibility, and substreams

hi again,

i have a question about the use of SubStreams for reproducibility.

theoretically speaking, sequentialfs should always output the same result, assuming the same folds/partitions and data are used on different runs.

i mean, the first feature that produces the highest accuracy for the given data/folds is not going to change because in order to make that determination, the remaining n-1 features must also be evaluated. i don't see how the random number seed would change the selection of these features (in a forward manner, at least).

so what does SubStreams do exactly? i'm corresponding with billy on figuring out why my macpro5,1 is using a crazy amount of ram for sequentialfs, and i'm wondering if this structure has something to do with it.

it's using a STUPID amount of memory. activity monitor in mac reports like ~50 gigs, but >100 gb is apparently being used and it's swapping out :/