From 0bf053fbeb8cb9d498027c9a2151e8c812fd7a77 Mon Sep 17 00:00:00 2001
From: Alexandre Gramfort <alexandre.gramfort@inria.fr>
Date: Thu, 24 Feb 2011 17:44:31 -0500
Subject: [PATCH] FIX : fixing Lars lasso with early stopping using alph_min +
adding test for it
---
scikits/learn/linear_model/least_angle.py | 19 +++++++++++--------
.../linear_model/tests/test_least_angle.py | 16 ++++++++++++++++
2 files changed, 27 insertions(+), 8 deletions(-)
diff --git a/scikits/learn/linear_model/least_angle.py b/scikits/learn/linear_model/least_angle.py
index 6739451a22..affb85fa67 100644
--- a/scikits/learn/linear_model/least_angle.py
+++ b/scikits/learn/linear_model/least_angle.py
@@ -120,7 +120,17 @@ def lars_path(X, y, Xy=None, Gram=None, max_features=None,
alphas[n_iter] = C / n_samples
- if (C < alpha_min) or (n_active == max_features):
+ # Check for early stopping
+ if alphas[n_iter] < alpha_min: # interpolate
+ # interpolation factor 0 <= ss < 1
+ ss = (alphas[n_iter-1] - alpha_min) / (alphas[n_iter-1] -
+ alphas[n_iter])
+ coefs[n_iter] = coefs[n_iter-1] + ss*(coefs[n_iter] -
+ coefs[n_iter-1])
+ alphas[n_iter] = alpha_min
+ break
+
+ if n_active == max_features:
break
if not drop:
@@ -270,13 +280,6 @@ def lars_path(X, y, Xy=None, Gram=None, max_features=None,
if verbose:
print "%s\t\t%s\t\t%s\t\t%s\t\t%s" % (n_iter, '', drop_idx,
n_active, abs(temp))
- if alphas[n_iter] < alpha_min: # interpolate
- # interpolation factor 0 <= ss < 1
- ss = (alphas[n_iter-1] - alpha_min) / (alphas[n_iter-1] -
- alphas[n_iter])
- coefs[n_iter] = coefs[n_iter-1] + ss*(coefs[n_iter] - coefs[n_iter-1])
- alphas[n_iter] = alpha_min
-
# resize coefs in case of early stop
alphas = alphas[:n_iter+1]
diff --git a/scikits/learn/linear_model/tests/test_least_angle.py b/scikits/learn/linear_model/tests/test_least_angle.py
index c9bf228ca4..4521dbbab5 100644
--- a/scikits/learn/linear_model/tests/test_least_angle.py
+++ b/scikits/learn/linear_model/tests/test_least_angle.py
@@ -111,6 +111,22 @@ def test_lasso_lars_vs_lasso_cd(verbose=False):
error = np.linalg.norm(c - lasso_cd.coef_)
assert error < 0.01
+def test_lasso_lars_vs_lasso_cd_early_stopping(verbose=False):
+ """
+ Test that LassoLars and Lasso using coordinate descent give the
+ same results when early stopping is used.
+ (test : before, in the middle, and in the last part of the path)
+ """
+ alphas_min = [10, 0.9, 1e-4]
+ for alphas_min in alphas_min:
+ alphas, _, lasso_path = linear_model.lars_path(X, y, method='lasso',
+ alpha_min=0.9)
+ lasso_cd = linear_model.Lasso(fit_intercept=False)
+ lasso_cd.alpha = alphas[-1]
+ lasso_cd.fit(X, y, tol=1e-8)
+ error = np.linalg.norm(lasso_path[:,-1] - lasso_cd.coef_)
+ assert error < 0.01
+
if __name__ == '__main__':
import nose
--
GitLab