blindml vs dod integration notes

[benclifford]

I've got dod and blindml talking to each other. There was a lot of fiddling. This issue documents potential action items on the blindml side. There is a corresponding dod issue, TheDataStation/aurum-dod-staging#3

Installability:

For blindml to be installed alongside other components then it should be packaged as a library. Mostly, this lets pip pick versions of dependencies that work for all of the components.

• add a setup.py to blindml so that pip can install it

Integration with DOD

Some parts of blindml assume numerical data. The test dataset I was using with dod was entirely categorical, which autosklearn supports, but I hacked some assumptions of numerical data out of blindml.

• add a test case for this style of data to the existing blindml tests, that demonstrates the problems i had
• fix blindml to support this

dod produces csv files with no extension. blindml then doesn't detect this as a csv. I hacked blindml to assume everything is a CSV. The right thing to do, I think, is that dod should produce outputs with a CSV extension: the corresponding DOD issue will have this as an action item.

Continuous Integration

• CI - blindml isn't particularly complicated installation/testing-wise, but the DOD issue will recommend that has CI, and it would make sense if doing that to do the same for blindml.