def test_group_lasso_separable():
"""
This test verifies that the specification of a separable
penalty yields the same results as having two linear_atoms
with selector matrices. The penalty here is a group_lasso, i.e. l2
penalty.
"""
X = np.random.standard_normal((100,20))
Y = np.random.standard_normal((100,)) + np.dot(X, np.random.standard_normal(20))
penalty1 = rr.l2norm(10, lagrange=.2)
penalty2 = rr.l2norm(10, lagrange=.2)
penalty = rr.separable((20,), [penalty1, penalty2], [slice(0,10), slice(10,20)])
# solve using separable
loss = rr.quadratic.affine(X, -Y, coef=0.5)
problem = rr.separable_problem.fromatom(penalty, loss)
solver = rr.FISTA(problem)
solver.fit(min_its=200, tol=1.0e-12)
coefs = solver.composite.coefs
# solve using the selectors
penalty_s = [rr.linear_atom(p, rr.selector(g, (20,))) for p, g in
zip(penalty.atoms, penalty.groups)]
problem_s = rr.container(loss, *penalty_s)
solver_s = rr.FISTA(problem_s)
solver_s.fit(min_its=200, tol=1.0e-12)
coefs_s = solver_s.composite.coefs
np.testing.assert_almost_equal(coefs, coefs_s)
def test_nonnegative_positive_part(debug=False):
"""
This test verifies that using nonnegative constraint
with a linear term, with some unpenalized terms yields the same result
as using separable with constrained_positive_part and nonnegative
"""
import numpy as np
import regreg.api as rr
import regreg.atoms as rra
# N - number of data points
# P - number of columns in design == number of betas
N, P = 40, 30
# an arbitrary positive offset for data and design
offset = 2
# data
Y = np.random.normal(size=(N, )) + offset
# design - with ones as last column
X = np.ones((N, P))
X[:, :-1] = np.random.normal(size=(N, P - 1)) + offset
# coef for loss
coef = 0.5
# lagrange for penalty
lagrange = .1
# Loss function (squared difference between fitted and actual data)
loss = rr.quadratic_loss.affine(X, -Y, coef=coef)
# Penalty using nonnegative, leave the last 5 unpenalized but
# nonnegative
weights = np.ones(P) * lagrange
weights[-5:] = 0
linq = rr.identity_quadratic(0, 0, weights, 0)
penalty = rr.nonnegative(P, quadratic=linq)
# Solution
composite_form = rr.separable_problem.singleton(penalty, loss)
solver = rr.FISTA(composite_form)
solver.debug = debug
solver.fit(tol=1.0e-12, min_its=200)
coefs = solver.composite.coefs
# using the separable penalty, only penalize the first
# 25 coefficients with constrained_positive_part
penalties_s = [
rr.constrained_positive_part(25, lagrange=lagrange),
rr.nonnegative(5)
]
groups_s = [slice(0, 25), slice(25, 30)]
penalty_s = rr.separable((P, ), penalties_s, groups_s)
composite_form_s = rr.separable_problem.singleton(penalty_s, loss)
solver_s = rr.FISTA(composite_form_s)
solver_s.debug = debug
solver_s.fit(tol=1.0e-12, min_its=200)
coefs_s = solver_s.composite.coefs
nt.assert_true(
np.linalg.norm(coefs - coefs_s) / np.linalg.norm(coefs) < 1.0e-02)
def test_group_lasso_separable():
"""
This test verifies that the specification of a separable
penalty yields the same results as having two linear_atoms
with selector matrices. The penalty here is a group_lasso, i.e. l2
penalty.
"""
X = np.random.standard_normal((100,20))
Y = np.random.standard_normal((100,)) + np.dot(X, np.random.standard_normal(20))
penalty1 = rr.l2norm(10, lagrange=.2)
penalty2 = rr.l2norm(10, lagrange=.2)
penalty = rr.separable((20,), [penalty1, penalty2], [slice(0,10), slice(10,20)])
# solve using separable
loss = rr.quadratic_loss.affine(X, -Y, coef=0.5)
problem = rr.separable_problem.fromatom(penalty, loss)
solver = rr.FISTA(problem)
solver.fit(min_its=200, tol=1.0e-12)
coefs = solver.composite.coefs
# solve using the selectors
penalty_s = [rr.linear_atom(p, rr.selector(g, (20,))) for p, g in
zip(penalty.atoms, penalty.groups)]
problem_s = rr.container(loss, *penalty_s)
solver_s = rr.FISTA(problem_s)
solver_s.fit(min_its=200, tol=1.0e-12)
coefs_s = solver_s.composite.coefs
np.testing.assert_almost_equal(coefs, coefs_s)
def test_nonnegative_positive_part(debug=False):
"""
This test verifies that using nonnegative constraint
with a linear term, with some unpenalized terms yields the same result
as using separable with constrained_positive_part and nonnegative
"""
import numpy as np
import regreg.api as rr
import regreg.atoms as rra
# N - number of data points
# P - number of columns in design == number of betas
N, P = 40, 30
# an arbitrary positive offset for data and design
offset = 2
# data
Y = np.random.normal(size=(N,)) + offset
# design - with ones as last column
X = np.ones((N,P))
X[:,:-1] = np.random.normal(size=(N,P-1)) + offset
# coef for loss
coef = 0.5
# lagrange for penalty
lagrange = .1
# Loss function (squared difference between fitted and actual data)
loss = rr.quadratic.affine(X, -Y, coef=coef)
# Penalty using nonnegative, leave the last 5 unpenalized but
# nonnegative
weights = np.ones(P) * lagrange
weights[-5:] = 0
linq = rr.identity_quadratic(0,0,weights,0)
penalty = rr.nonnegative(P, quadratic=linq)
# Solution
composite_form = rr.separable_problem.singleton(penalty, loss)
solver = rr.FISTA(composite_form)
solver.debug = debug
solver.fit(tol=1.0e-12, min_its=200)
coefs = solver.composite.coefs
# using the separable penalty, only penalize the first
# 25 coefficients with constrained_positive_part
penalties_s = [rr.constrained_positive_part(25, lagrange=lagrange),
rr.nonnegative(5)]
groups_s = [slice(0,25), slice(25,30)]
penalty_s = rr.separable((P,), penalties_s,
groups_s)
composite_form_s = rr.separable_problem.singleton(penalty_s, loss)
solver_s = rr.FISTA(composite_form_s)
solver_s.debug = debug
solver_s.fit(tol=1.0e-12, min_its=200)
coefs_s = solver_s.composite.coefs
nt.assert_true(np.linalg.norm(coefs - coefs_s) / np.linalg.norm(coefs) < 1.0e-02)
def test_lasso_separable():
"""
This test verifies that the specification of a separable
penalty yields the same results as having two linear_atoms
with selector matrices. The penalty here is a lasso, i.e. l1
penalty.
"""
X = np.random.standard_normal((100,20))
Y = np.random.standard_normal((100,)) + np.dot(X, np.random.standard_normal(20))
penalty1 = rr.l1norm(10, lagrange=1.2)
penalty2 = rr.l1norm(10, lagrange=1.2)
penalty = rr.separable((20,), [penalty1, penalty2], [slice(0,10), slice(10,20)], test_for_overlap=True)
# ensure code is tested
print(penalty1.latexify())
print(penalty.latexify())
print(penalty.conjugate)
print(penalty.dual)
print(penalty.seminorm(np.ones(penalty.shape)))
print(penalty.constraint(np.ones(penalty.shape), bound=2.))
pencopy = copy(penalty)
pencopy.set_quadratic(rr.identity_quadratic(1,0,0,0))
pencopy.conjugate
# solve using separable
loss = rr.quadratic.affine(X, -Y, coef=0.5)
problem = rr.separable_problem.fromatom(penalty, loss)
solver = rr.FISTA(problem)
solver.fit(min_its=200, tol=1.0e-12)
coefs = solver.composite.coefs
# solve using the usual composite
penalty_all = rr.l1norm(20, lagrange=1.2)
problem_all = rr.container(loss, penalty_all)
solver_all = rr.FISTA(problem_all)
solver_all.fit(min_its=100, tol=1.0e-12)
coefs_all = solver_all.composite.coefs
# solve using the selectors
penalty_s = [rr.linear_atom(p, rr.selector(g, (20,))) for p, g in
zip(penalty.atoms, penalty.groups)]
problem_s = rr.container(loss, *penalty_s)
solver_s = rr.FISTA(problem_s)
solver_s.fit(min_its=500, tol=1.0e-12)
coefs_s = solver_s.composite.coefs
np.testing.assert_almost_equal(coefs, coefs_all)
np.testing.assert_almost_equal(coefs, coefs_s)
def test_lasso_separable():
"""
This test verifies that the specification of a separable
penalty yields the same results as having two linear_atoms
with selector matrices. The penalty here is a lasso, i.e. l1
penalty.
"""
X = np.random.standard_normal((100,20))
Y = np.random.standard_normal((100,)) + np.dot(X, np.random.standard_normal(20))
penalty1 = rr.l1norm(10, lagrange=1.2)
penalty2 = rr.l1norm(10, lagrange=1.2)
penalty = rr.separable((20,), [penalty1, penalty2], [slice(0,10), slice(10,20)], test_for_overlap=True)
# ensure code is tested
print(penalty1.latexify())
print(penalty.latexify())
print(penalty.conjugate)
print(penalty.dual)
print(penalty.seminorm(np.ones(penalty.shape)))
print(penalty.constraint(np.ones(penalty.shape), bound=2.))
pencopy = copy(penalty)
pencopy.set_quadratic(rr.identity_quadratic(1,0,0,0))
pencopy.conjugate
# solve using separable
loss = rr.quadratic_loss.affine(X, -Y, coef=0.5)
problem = rr.separable_problem.fromatom(penalty, loss)
solver = rr.FISTA(problem)
solver.fit(min_its=200, tol=1.0e-12)
coefs = solver.composite.coefs
# solve using the usual composite
penalty_all = rr.l1norm(20, lagrange=1.2)
problem_all = rr.container(loss, penalty_all)
solver_all = rr.FISTA(problem_all)
solver_all.fit(min_its=100, tol=1.0e-12)
coefs_all = solver_all.composite.coefs
# solve using the selectors
penalty_s = [rr.linear_atom(p, rr.selector(g, (20,))) for p, g in
zip(penalty.atoms, penalty.groups)]
problem_s = rr.container(loss, *penalty_s)
solver_s = rr.FISTA(problem_s)
solver_s.fit(min_its=500, tol=1.0e-12)
coefs_s = solver_s.composite.coefs
np.testing.assert_almost_equal(coefs, coefs_all)
np.testing.assert_almost_equal(coefs, coefs_s)
def test_path_group_lasso():
'''
this test looks at the paths of three different parameterizations
of the same problem
'''
n = 100
X = np.random.standard_normal((n, 10))
U = np.random.standard_normal((n, 2))
Y = np.random.standard_normal(100)
betaX = np.array([3, 4, 5, 0, 0] + [0] * 5)
betaU = np.array([10, -5])
Y += (np.dot(X, betaX) + np.dot(U, betaU)) * 5
Xn = rr.normalize(np.hstack([np.ones((100, 1)), X]),
inplace=True,
center=True,
scale=True,
intercept_column=0).normalized_array()
lasso = mixed_lasso.mixed_lasso_path.gaussian(Xn[:, 1:],
Y,
penalty_structure=[0] * 7 +
[1] * 3,
nstep=10)
sol = lasso.main(inner_tol=1.e-12, verbose=True)
beta = np.array(sol['beta'].todense())
sols = []
sols_sep = []
for l in sol['lagrange']:
loss = rr.glm.gaussian(Xn, Y)
penalty = rr.mixed_lasso([mixed_lasso.UNPENALIZED] + [0] * 7 + [1] * 3,
lagrange=l) # matrix contains an intercept...
problem = rr.simple_problem(loss, penalty)
sols.append(problem.solve(tol=1.e-12).copy())
sep = rr.separable((11, ), [
rr.l2norm((7, ),
np.sqrt(7) * l),
rr.l2norm((3, ),
np.sqrt(3) * l)
], [np.arange(1, 8), np.arange(8, 11)])
sep_problem = rr.simple_problem(loss, sep)
sols_sep.append(sep_problem.solve(tol=1.e-12).copy())
sols = np.array(sols).T
sols_sep = np.array(sols_sep).T
nt.assert_true(
np.linalg.norm(beta - sols) / (1 + np.linalg.norm(beta)) <= 1.e-4)
nt.assert_true(
np.linalg.norm(beta - sols_sep) / (1 + np.linalg.norm(beta)) <= 1.e-4)
def test_path_group_lasso():
"""
this test looks at the paths of three different parameterizations
of the same problem
"""
n = 100
X = np.random.standard_normal((n, 10))
U = np.random.standard_normal((n, 2))
Y = np.random.standard_normal(100)
betaX = np.array([3, 4, 5, 0, 0] + [0] * 5)
betaU = np.array([10, -5])
Y += (np.dot(X, betaX) + np.dot(U, betaU)) * 5
Xn = rr.normalize(
np.hstack([np.ones((100, 1)), X]), inplace=True, center=True, scale=True, intercept_column=0
).normalized_array()
lasso = rr.lasso.squared_error(Xn[:, 1:], Y, penalty_structure=[0] * 7 + [1] * 3, nstep=10)
sol = lasso.main(inner_tol=1.0e-12, verbose=True)
beta = np.array(sol["beta"].todense())
sols = []
sols_sep = []
for l in sol["lagrange"]:
loss = rr.squared_error(Xn, Y, coef=1.0 / n)
penalty = rr.mixed_lasso([rr.UNPENALIZED] + [0] * 7 + [1] * 3, lagrange=l) # matrix contains an intercept...
problem = rr.simple_problem(loss, penalty)
sols.append(problem.solve(tol=1.0e-12).copy())
sep = rr.separable(
(11,),
[rr.l2norm((7,), np.sqrt(7) * l), rr.l2norm((3,), np.sqrt(3) * l)],
[np.arange(1, 8), np.arange(8, 11)],
)
sep_problem = rr.simple_problem(loss, sep)
sols_sep.append(sep_problem.solve(tol=1.0e-12).copy())
sols = np.array(sols).T
sols_sep = np.array(sols_sep).T
nt.assert_true(np.linalg.norm(beta - sols) / (1 + np.linalg.norm(beta)) <= 1.0e-4)
nt.assert_true(np.linalg.norm(beta - sols_sep) / (1 + np.linalg.norm(beta)) <= 1.0e-4)
def test_different_dim():
"""
This test checks that the reshape argument of separable
works properly.
"""
X = np.random.standard_normal((100, 20))
Y = (np.random.standard_normal(
(100, )) + np.dot(X, np.random.standard_normal(20)))
penalty1 = rr.nuclear_norm((5, 2), lagrange=1.2)
penalty2 = rr.l1norm(10, lagrange=1.2)
penalty = rr.separable((20, ), [penalty1, penalty2],
[slice(0, 10), slice(10, 20)],
test_for_overlap=True,
shapes=[(5, 2), None])
# ensure code is tested
print(penalty1.latexify())
print(penalty.latexify())
print(penalty.conjugate)
print(penalty.dual)
print(penalty.seminorm(np.ones(penalty.shape)))
print(penalty.constraint(np.ones(penalty.shape), bound=2.))
pencopy = copy(penalty)
pencopy.set_quadratic(rr.identity_quadratic(1, 0, 0, 0))
pencopy.conjugate
# solve using separable
loss = rr.quadratic_loss.affine(X, -Y, coef=0.5)
problem = rr.separable_problem.fromatom(penalty, loss)
solver = rr.FISTA(problem)
solver.fit(min_its=200, tol=1.0e-12)
coefs = solver.composite.coefs
def test_centering_fit_inplace(debug=False):
# N - number of data points
# P - number of columns in design == number of betas
N, P = 40, 30
# an arbitrary positive offset for data and design
offset = 2
# design - with ones as last column
X = np.random.normal(size=(N, P)) + offset
L = rr.normalize(X, center=True, scale=False, inplace=True)
# X should have been normalized in place
np.testing.assert_almost_equal(np.sum(X, 0), 0)
# data
Y = np.random.normal(size=(N, )) + offset
# coef for loss
coef = 0.5
# lagrange for penalty
lagrange = .1
# Loss function (squared difference between fitted and actual data)
loss = rr.quadratic_loss.affine(L, -Y, coef=coef)
penalties = [
rr.constrained_positive_part(25, lagrange=lagrange),
rr.nonnegative(5)
]
groups = [slice(0, 25), slice(25, 30)]
penalty = rr.separable((P, ), penalties, groups)
initial = np.random.standard_normal(P)
composite_form = rr.separable_problem.fromatom(penalty, loss)
solver = rr.FISTA(composite_form)
solver.debug = debug
solver.fit(tol=1.0e-12, min_its=200)
coefs = solver.composite.coefs
# Solve the problem with X, which has been normalized in place
loss2 = rr.quadratic_loss.affine(X, -Y, coef=coef)
initial2 = np.random.standard_normal(P)
composite_form2 = rr.separable_problem.fromatom(penalty, loss2)
solver2 = rr.FISTA(composite_form2)
solver2.debug = debug
solver2.fit(tol=1.0e-12, min_its=200)
coefs2 = solver2.composite.coefs
for _ in range(10):
beta = np.random.standard_normal(P)
g1 = loss.smooth_objective(beta, mode='grad')
g2 = loss2.smooth_objective(beta, mode='grad')
np.testing.assert_almost_equal(g1, g2)
b1 = penalty.proximal(sq(1, beta, g1, 0))
b2 = penalty.proximal(sq(1, beta, g2, 0))
np.testing.assert_almost_equal(b1, b2)
f1 = composite_form.objective(beta)
f2 = composite_form2.objective(beta)
np.testing.assert_almost_equal(f1, f2)
np.testing.assert_almost_equal(composite_form.objective(coefs),
composite_form.objective(coefs2))
np.testing.assert_almost_equal(composite_form2.objective(coefs),
composite_form2.objective(coefs2))
nt.assert_true(
np.linalg.norm(coefs - coefs2) /
max(np.linalg.norm(coefs), 1) < 1.0e-04)
def test_centering_fit_inplace(debug=False):
# N - number of data points
# P - number of columns in design == number of betas
N, P = 40, 30
# an arbitrary positive offset for data and design
offset = 2
# design - with ones as last column
X = np.random.normal(size=(N,P)) + offset
L = rr.normalize(X, center=True, scale=False, inplace=True)
# X should have been normalized in place
np.testing.assert_almost_equal(np.sum(X, 0), 0)
# data
Y = np.random.normal(size=(N,)) + offset
# coef for loss
coef = 0.5
# lagrange for penalty
lagrange = .1
# Loss function (squared difference between fitted and actual data)
loss = rr.quadratic.affine(L, -Y, coef=coef)
penalties = [rr.constrained_positive_part(25, lagrange=lagrange),
rr.nonnegative(5)]
groups = [slice(0,25), slice(25,30)]
penalty = rr.separable((P,), penalties,
groups)
initial = np.random.standard_normal(P)
composite_form = rr.separable_problem.fromatom(penalty, loss)
solver = rr.FISTA(composite_form)
solver.debug = debug
solver.fit(tol=1.0e-12, min_its=200)
coefs = solver.composite.coefs
# Solve the problem with X, which has been normalized in place
loss2 = rr.quadratic.affine(X, -Y, coef=coef)
initial2 = np.random.standard_normal(P)
composite_form2 = rr.separable_problem.fromatom(penalty, loss2)
solver2 = rr.FISTA(composite_form2)
solver2.debug = debug
solver2.fit(tol=1.0e-12, min_its=200)
coefs2 = solver2.composite.coefs
for _ in range(10):
beta = np.random.standard_normal(P)
g1 = loss.smooth_objective(beta, mode='grad')
g2 = loss2.smooth_objective(beta, mode='grad')
np.testing.assert_almost_equal(g1, g2)
b1 = penalty.proximal(sq(1, beta, g1,0))
b2 = penalty.proximal(sq(1, beta, g2,0))
np.testing.assert_almost_equal(b1, b2)
f1 = composite_form.objective(beta)
f2 = composite_form2.objective(beta)
np.testing.assert_almost_equal(f1, f2)
np.testing.assert_almost_equal(composite_form.objective(coefs), composite_form.objective(coefs2))
np.testing.assert_almost_equal(composite_form2.objective(coefs), composite_form2.objective(coefs2))
nt.assert_true(np.linalg.norm(coefs - coefs2) / max(np.linalg.norm(coefs),1) < 1.0e-04)
def test_scaling_and_centering_intercept_fit(debug=False):
# N - number of data points
# P - number of columns in design == number of betas
N, P = 40, 30
# an arbitrary positive offset for data and design
offset = 2
# design - with ones as last column
X = np.random.normal(size=(N, P)) + 0 * offset
X2 = X - X.mean(0)[None, :]
X2 = X2 / np.std(X2, 0, ddof=1)[None, :]
X2 = np.hstack([np.ones((X2.shape[0], 1)), X2])
L = rr.normalize(X, center=True, scale=True, intercept=True)
# data
Y = np.random.normal(size=(N, )) + offset
# lagrange for penalty
lagrange = .1
# Loss function (squared difference between fitted and actual data)
loss = rr.squared_error(L, Y)
penalties = [
rr.constrained_positive_part(25, lagrange=lagrange),
rr.nonnegative(5)
]
groups = [slice(0, 25), slice(25, 30)]
penalty = rr.separable((P + 1, ), penalties, groups)
initial = np.random.standard_normal(P + 1)
composite_form = rr.separable_problem.fromatom(penalty, loss)
solver = rr.FISTA(composite_form)
solver.debug = debug
solver.fit(tol=1.0e-12, min_its=200)
coefs = solver.composite.coefs
# Solve the problem with X2
loss2 = rr.squared_error(X2, Y)
initial2 = np.random.standard_normal(P + 1)
composite_form2 = rr.separable_problem.fromatom(penalty, loss2)
solver2 = rr.FISTA(composite_form2)
solver2.debug = debug
solver2.fit(tol=1.0e-12, min_its=200)
coefs2 = solver2.composite.coefs
for _ in range(10):
beta = np.random.standard_normal(P + 1)
g1 = loss.smooth_objective(beta, mode='grad')
g2 = loss2.smooth_objective(beta, mode='grad')
np.testing.assert_almost_equal(g1, g2)
b1 = penalty.proximal(sq(1, beta, g1, 0))
b2 = penalty.proximal(sq(1, beta, g2, 0))
np.testing.assert_almost_equal(b1, b2)
f1 = composite_form.objective(beta)
f2 = composite_form2.objective(beta)
np.testing.assert_almost_equal(f1, f2)
np.testing.assert_almost_equal(composite_form.objective(coefs),
composite_form.objective(coefs2))
np.testing.assert_almost_equal(composite_form2.objective(coefs),
composite_form2.objective(coefs2))
nt.assert_true(
np.linalg.norm(coefs - coefs2) /
max(np.linalg.norm(coefs), 1) < 1.0e-04)
