robustify against 0 norm features

[mathurinm]

Something weird is happening: without this, the solver fails when X has a 0 column (which is normal, there is a division by 0)

But such a column should not be selected in the WS, right ?

Reproduce with

import libsvmdata
import numpy as np
from numpy.linalg import norm
from andersoncd import Lasso
X, y = libsvmdata.fetch_libsvm("rcv1.binary")
alpha_max = norm(X.T @ y, ord=np.inf) / len(y)
clf = Lasso(fit_intercept=False, alpha=alpha_max/10, verbose=1).fit(X, y)

[codecov-commenter]

Codecov Report

Merging #63 (ea11649) into master (b8ce7d3) will increase coverage by 8.90%. The diff coverage is 67.84%.

@@            Coverage Diff             @@
##           master      #63      +/-   ##
==========================================
+ Coverage   56.19%   65.09%   +8.90%     
==========================================
  Files          12       11       -1     
  Lines        1098      742     -356     
  Branches      242      117     -125     
==========================================
- Hits          617      483     -134     
+ Misses        406      228     -178     
+ Partials       75       31      -44     

Impacted Files	Coverage Δ
andersoncd/tests/test_docstring_parameters.py	73.91% <ø> (-0.73%)	:arrow_down:
andersoncd/penalties.py	44.76% <44.76%> (ø)
andersoncd/datafits.py	52.23% <52.23%> (ø)
andersoncd/solver.py	66.44% <66.44%> (ø)
andersoncd/data/synthetic.py	72.41% <69.23%> (-18.50%)	:arrow_down:
andersoncd/estimators.py	92.39% <92.39%> (ø)
andersoncd/__init__.py	100.00% <100.00%> (ø)
andersoncd/data/__init__.py	100.00% <100.00%> (ø)
andersoncd/tests/test_estimators.py	100.00% <100.00%> (ø)
andersoncd/utils.py	28.76% <0.00%> (-4.11%)	:arrow_down:
... and 1 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more Δ = absolute <relative> (impact), ø = not affected, ? = missing data Powered by Codecov. Last update 1d83596...ea11649. Read the comment docs.

[mathurinm]

Investigated a bit further : at the iteration where the solver fails, kkt has only 82 non zero values (I'm surprised)

by growth policy it turns out that we select 114 feats in subpb.. We are unlucky and amongst the 114 - 86 features with 0 kkt violation that we pick, there is one which is a 0 column, hence the failure.

It's wild that so many features have 0 kkt violation, no ? I see the easy fix of selecting at most (kkt != 0).sum() features in the ws, but I'm surprised it's 0 for such a large number of features

Your take on this @QB3 ?

[QB3]

I'm surprised it's 0 for such a large number of features

I also already observed that uncleaned rcv1 has a large number of zero columns. +1 for (kkt != 0).sum()