Linear models

Speed improvments in linear models.

Multiple targets

This makes linear models accept 2d arrays for y intelligently, by moving redundant computation out of the loop. This does NOT mean multitask learning, the correlation between the targets is not taken into account, it's the same as y.shape[1] different copies of the estimator.

Lars

In [8]: from sklearn.datasets import make_regression
In [9]: from sklearn.linear_model import LassoLars

In [10]: X, y = make_regression(1000, 1000, 10, n_targets=50)

In [11]: def fit_iter(X, Y):
    ....:     for y in Y.T:
    ....:         LassoLars(1.0, precompute=True).fit(X, y)
    ....:         
In [12]: def fit_once(X, y):
    ....:    LassoLars(1.0, precompute=True).fit(X, y)
    ....:     

In [13]: %timeit fit_once(X, y)
1 loops, best of 3: 824 ms per loop

In [14]: %timeit fit_iter(X, y)
1 loops, best of 3: 10.1 s per loop