Beching to measure the effect of random descent

bench_demo.py

import numpy as np
from scipy import signal

rng = np.random.RandomState(42)
n_samples, n_features, k = 500, 1000, 10

h = signal.gaussian(50, 15)
X = signal.convolve2d(np.eye(n_features), h[:, np.newaxis], 'same')  # convolutional design
X = X[::n_features // n_samples]

coef0 = rng.randn(n_features)
coef0[np.abs(coef0) < np.sort(np.abs(coef0))[-k]] = 0
y = np.dot(X, coef0)

alpha = 0.001

from sklearn.linear_model import Lasso
clf = Lasso(alpha=alpha, random_state=0)
clf.fit(X, y)