def test_sparse_solution():
Phi, y, w_gen = generate_fake_data(100, 20, .5)
model = LinARD(100)
model.train(Phi, y)
assert np.sum(abs(model.coef_) > 0) > .4
assert np.sum(abs(model.coef_) > 0) > .6
def test_recover_sparseness():
Phi, y, w_gen = generate_fake_data(10000, 20, .5)
model = LinARD(200)
model.train(Phi, y)
# allow for up to 3 random mistake
print "This test is based on randomness. It may occasionally fail," \
" but you should get suspicious if it fails twice in a row."
assert np.sum((abs(model.coef_) < 1e-7) == (abs(w_gen) > 1e-7)) <= 3
def test_more_data_better():
Phi, y, w_gen = generate_fake_data(20000, 20, .5)
model_small_data = LinARD(100)
model_big_data = LinARD(100)
model_small_data.train(Phi[:1000], y[:1000])
model_big_data.train(Phi[:15000], y[:15000])
y_small = model_small_data.predict(Phi[15000:])
y_big = model_big_data.predict(Phi[15000:])
assert np.sum((y[15000:]-y_small)**2) > np.sum((y[15000:]-y_big)**2)