ENH use weights in penalties

[mathurinm]

paves the way for non convex solvers solved with MM approachs (MCP, AdaptiveLasso)
also allows finer policies in GroupLasso case

[codecov-commenter]

Codecov Report

Merging #131 into master will increase coverage by 0.53%.
The diff coverage is 96.96%.

@@            Coverage Diff             @@
##           master     #131      +/-   ##
==========================================
+ Coverage   79.26%   79.80%   +0.53%     
==========================================
  Files          12       12              
  Lines        1008     1030      +22     
  Branches      147      148       +1     
==========================================
+ Hits          799      822      +23     
+ Misses        168      167       -1     
  Partials       41       41              

Impacted Files	Coverage Δ
celer/homotopy.py	87.30% <94.73%> (+1.06%)	⬆️
celer/dropin_sklearn.py	89.20% <100.00%> (+0.07%)	⬆️
celer/tests/test_lasso.py	100.00% <100.00%> (ø)

---

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 20a06a4...93e402b. Read the comment docs.

[mathurinm]

Remains todo: implement weights usage for groups.
Prox Newton solver support should come after #109 has been fixed.

[mathurinm]

@QB3 this needs a benchmark to check that it does not slow down classic Lasso, but it should be useful for you.

[codecov-io]

Codecov Report

Merging #131 into master will increase coverage by 0.45%.
The diff coverage is 96.96%.

@@            Coverage Diff             @@
##           master     #131      +/-   ##
==========================================
+ Coverage   84.35%   84.81%   +0.45%     
==========================================
  Files          12       12              
  Lines         959      981      +22     
  Branches      136      137       +1     
==========================================
+ Hits          809      832      +23     
+ Misses        109      108       -1     
  Partials       41       41              

Impacted Files	Coverage Δ
celer/homotopy.py	87.30% <94.73%> (+1.06%)	⬆️
celer/dropin_sklearn.py	96.42% <100.00%> (+0.02%)	⬆️
celer/tests/test_lasso.py	100.00% <100.00%> (ø)

---

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 1783b7c...f168e29. Read the comment docs.

[mathurinm]

This PR:

In [1]: import numpy as np

In [2]: from libsvmdata import fetch_libsvm

In [3]: from celer import Lasso

In [4]: X, y = fetch_libsvm("news20")
Dataset: news20

In [5]: alpha = np.linalg.norm(X.T @ y, ord=np.inf) / len(y) / 100

In [6]: %timeit -n 1 -r 15 clf = Lasso(alpha=alpha).fit(X, y)
6.38 s ± 240 ms per loop (mean ± std. dev. of 15 runs, 1 loop each)

Master:

In [1]: import numpy as np

In [2]: from libsvmdata import fetch_libsvm

In [3]: from celer import Lasso

In [4]: X, y = fetch_libsvm("news20")
Dataset: news20

In [5]: alpha = np.linalg.norm(X.T @ y, ord=np.inf) / len(y) / 100

In [6]: %timeit -n 1 -r 15 clf = Lasso(alpha=alpha).fit(X, y)
6.38 s ± 68.9 ms per loop (mean ± std. dev. of 15 runs, 1 loop each)

There should be a way to automate this on every PR