===============================================================
Run GroupLasso and GroupLasso CV for structured sparse recovery
===============================================================
The example runs the GroupLasso scikit-learn like estimators.
"""
import numpy as np
import matplotlib.pyplot as plt
from celer import GroupLassoCV, LassoCV
from celer.datasets import make_correlated_data
from celer.plot_utils import configure_plt
print(__doc__)
configure_plt(fontsize=16)
# Generating X, y, and true regression coefs with 4 groups of 5 non-zero values
n_samples, n_features = 100, 50
w_true = np.zeros(n_features)
w_true[:5] = 1
w_true[10:15] = 1
w_true[30:35] = -1
w_true[45:] = 1
X, y, w_true = make_correlated_data(
n_samples, n_features, w_true=w_true, snr=5, random_state=0)
###############################################################################
# Get group Lasso's optimal alpha for prediction by cross validation
=====================================================
From the data, a good dual stepsize can be estimated in the case of sparse
recovery.
"""
import numpy as np
from numpy.linalg import norm
import matplotlib.pyplot as plt
from sklearn.metrics import f1_score
from celer.datasets import make_correlated_data
from celer.plot_utils import configure_plt
from iterreg.sparse import dual_primal
configure_plt()
# data for the experiment:
n_samples = 200
n_features = 500
X, y, w_true = make_correlated_data(n_samples=n_samples,
n_features=n_features,
corr=0.2,
density=0.1,
snr=10,
random_state=0)
###############################################################################
# In the L1 case, the Chambolle-Pock algorithm converges to the noisy Basis
# Pursuit solution, which has ``min(n_samples, n_features)`` non zero entries.