WGR incorrect default alphas in RidgeRegression

[eric-czech]

In the second stage regression for WGR, the default alphas generated here based on the number of distinct headers are including the label values:

glow/python/glow/wgr/linear_model/ridge_model.py

Line 220 in f3edf5b

self.alphas = generate_alphas(blockdf)

This means that the alphas become a function of the number of outcomes (i.e. labels), which I assume isn't what was intended right? It seems odd to me that the results would change for one outcome if it was fit alone vs w/ others, so I think it should probably only be based on the number of features for the regression like stage 1 -- number of variant blocks * number of stage 1 alphas (not then also * number of outcomes).

[LelandBarnard]

Hi Eric, great catch, I only recently noticed this as well It is definitely unintended, we will fix it in an upcoming release.

[henrydavidge]

This is resolved.