ENH: Lightning seems to be slow when loss=log

[MechCoder]

@mblondel I'm not sure if this is meant to be, but I ran a quick few benchmarks.

# Load News20 dataset from scikit-learn. 
bunch = fetch_20newsgroups_vectorized(subset="all")
X = bunch.data
y = bunch.target

# To remove the effect of parallelization
y[y != 1] = -1
time_logistic = []
time_lightning = []
Cs = np.logspace(-4, 4, 10)

for C in Cs:
    print C
    t = time()
    clf = LogisticRegression(penalty='l1', tol=0.0001, fit_intercept=False, C=C)
    t = time()
    clf.fit(X, y)
    time_logistic.append(time() - t)
    print time_logistic
    cl = CDClassifier(loss='log', tol=0.0001, max_iter=100, max_steps=0, C=C, penalty='l1')
    t = time()
    cl.fit(X, y)
    time_lightning.append(time() - t)
    print time_lightning

I get times like these for a grid of 10 Cs from np.logspace(-4, 4, 10)

time_lightning

[0.20100116729736328, 0.6052899360656738,  0.7211019992828369,
 2.470484972000122, 4.043258190155029, 7.791965007781982,
 10.92172908782959, 13.969007968902588, 12.534989833831787,
  5.275091886520386]

time_logistic

[0.08612680435180664, 0.22542500495910645, 0.5105628967285156,
 0.5970029830932617, 0.642221212387085, 0.8863811492919922,
 1.241279125213623, 1.1004469394683838, 0.9302711486816406,
 0.8940119743347168]

[mblondel]

What is your point?

liblinear doesn't implement the same algorithm as lightning...

[MechCoder]

Liblinear implements the CD + glmnet right in the same paper? I just wanted to clarify if the new GLMnet is inherently much faster than CDN, or if you think there are places in the lightning code where we could speed up.

[mblondel]

Indeed liblinear uses CD + glmnet. Don't they compare with CD in the paper?

One bottleneck is the computation of logs and exponentials. Using some kind of approximation could result in a big speed up.

lightning uses dataset and loss function abstractions. I am not sure what is the overhead of virtual method calls.

[mblondel]

Something else to be careful about is the stopping criterion. If liblinear doesn't use the same as lightning, the meaning of tol will differ. The only objective way to compare is to plot the objective value over time.