def test_graphnet_wts(X, Xlist, Y, l1=500.0, l2=2, l3=3.5, tol=1e-4):
print "graphnet_wts", l1, l2, l3
n = len(Xlist)
wts = np.random.normal(0, 1, n)
Xlist2 = [Xlist[i] * wts[i] for i in range(n)]
A, Afull = gen_adj(X.shape[1])
l = regreg.cwpath(regreg.graphnet_wts((Xlist, Y, A), rowweights=wts)) # initial_coefs = np.array([7.]*10))
l.problem.assign_penalty(l1=l1, l2=l2, l3=l3)
l.fit(tol=tol, max_its=400)
beta = l.problem.coefficients
def f(beta):
return (
np.linalg.norm(Y - np.dot(np.vstack(Xlist2), beta)) ** 2 / (2 * len(Y))
+ np.fabs(beta).sum() * l1
+ l2 * np.linalg.norm(beta) ** 2 / 2
+ l3 * np.dot(beta, np.dot(Afull, beta)) / 2
)
v = scipy.optimize.fmin_powell(f, np.zeros(X.shape[1]), ftol=1.0e-10, xtol=1.0e-10, maxfun=100000)
v = np.asarray(v)
print np.round(100 * v) / 100, "\n", np.round(100 * beta) / 100
print f(beta), f(v)
if f(v) < f(beta) + 1e-10:
assert np.fabs(f(v) - f(beta)) / np.fabs(f(v) + f(beta)) < 1.0e-04
if np.linalg.norm(v) > 1e-8:
assert np.linalg.norm(v - beta) / np.linalg.norm(v) < 1.0e-04
else:
assert np.linalg.norm(beta) < 1e-8
def test_graphnet_wts(X, Xlist, Y, l1=500., l2=2, l3=3.5, tol=1e-4):
print "graphnet_wts", l1, l2, l3
n = len(Xlist)
wts = np.random.normal(0, 1, n)
Xlist2 = [Xlist[i] * wts[i] for i in range(n)]
A, Afull = gen_adj(X.shape[1])
l = regreg.cwpath(regreg.graphnet_wts(
(Xlist, Y, A), rowweights=wts)) # initial_coefs = np.array([7.]*10))
l.problem.assign_penalty(l1=l1, l2=l2, l3=l3)
l.fit(tol=tol, max_its=400)
beta = l.problem.coefficients
def f(beta):
return np.linalg.norm(Y - np.dot(np.vstack(Xlist2), beta))**2 / (
2 * len(Y)) + np.fabs(beta).sum() * l1 + l2 * np.linalg.norm(
beta)**2 / 2 + l3 * np.dot(beta, np.dot(Afull, beta)) / 2
v = scipy.optimize.fmin_powell(f,
np.zeros(X.shape[1]),
ftol=1.0e-10,
xtol=1.0e-10,
maxfun=100000)
v = np.asarray(v)
print np.round(100 * v) / 100, '\n', np.round(100 * beta) / 100
print f(beta), f(v)
if f(v) < f(beta) + 1e-10:
assert (np.fabs(f(v) - f(beta)) / np.fabs(f(v) + f(beta)) < 1.0e-04)
if np.linalg.norm(v) > 1e-8:
assert (np.linalg.norm(v - beta) / np.linalg.norm(v) < 1.0e-04)
else:
assert (np.linalg.norm(beta) < 1e-8)
def test_graphnet(X, Xlist, Y, l1=500.0, l2=2, l3=3.5, tol=1e-4):
print "GraphNet", l1, l2, l3
A, Afull = gen_adj(X.shape[1])
# l = regreg.regreg((Xlist, Y, A),regreg.graphnet,regreg.cwpath)#, initial_coefs= np.array([7.]*10))
l = regreg.cwpath(regreg.graphnet((Xlist, Y, A)))
l.problem.assign_penalty(l1=l1, l2=l2, l3=l3)
l.fit(tol=tol, inner_its=50, max_its=5000)
beta = l.problem.coefficients
def f(beta):
return (
np.linalg.norm(Y - np.dot(X, beta)) ** 2 / (2 * len(Y))
+ np.fabs(beta).sum() * l1
+ l2 * np.linalg.norm(beta) ** 2 / 2
+ l3 * np.dot(beta, np.dot(Afull, beta)) / 2
)
v = scipy.optimize.fmin_powell(f, np.zeros(X.shape[1]), ftol=1.0e-10, xtol=1.0e-10, maxfun=100000)
v = np.asarray(v)
# print np.round(100*v)/100,'\n', np.round(100*beta)/100
# print f(v), f(beta)
if f(v) < f(beta):
assert np.fabs(f(v) - f(beta)) / np.fabs(f(v) + f(beta)) < 1.0e-04
if np.linalg.norm(v) > 1e-8:
assert np.linalg.norm(v - beta) / np.linalg.norm(v) < 1.0e-04
else:
assert np.linalg.norm(beta) < 1e-8
def test_graphnet(X, Xlist, Y, l1=500., l2=2, l3=3.5, tol=1e-4):
print "GraphNet", l1, l2, l3
A, Afull = gen_adj(X.shape[1])
#l = regreg.regreg((Xlist, Y, A),regreg.graphnet,regreg.cwpath)#, initial_coefs= np.array([7.]*10))
l = regreg.cwpath(regreg.graphnet((Xlist, Y, A)))
l.problem.assign_penalty(l1=l1, l2=l2, l3=l3)
l.fit(tol=tol, inner_its=50, max_its=5000)
beta = l.problem.coefficients
def f(beta):
return np.linalg.norm(Y - np.dot(X, beta))**2 / (2 * len(Y)) + np.fabs(
beta).sum() * l1 + l2 * np.linalg.norm(beta)**2 / 2 + l3 * np.dot(
beta, np.dot(Afull, beta)) / 2
v = scipy.optimize.fmin_powell(f,
np.zeros(X.shape[1]),
ftol=1.0e-10,
xtol=1.0e-10,
maxfun=100000)
v = np.asarray(v)
#print np.round(100*v)/100,'\n', np.round(100*beta)/100
#print f(v), f(beta)
if f(v) < f(beta):
assert (np.fabs(f(v) - f(beta)) / np.fabs(f(v) + f(beta)) < 1.0e-04)
if np.linalg.norm(v) > 1e-8:
assert (np.linalg.norm(v - beta) / np.linalg.norm(v) < 1.0e-04)
else:
assert (np.linalg.norm(beta) < 1e-8)
def test_v_graphnet(X, Xlist, Y, l1=500., l2=2, l3=3.5, tol=1e-4):
print "v_graphnet", l1, l2, l3
A, Afull = gen_adj(X.shape[1])
vec = np.dot(Y, X)
l = regreg.cwpath(regreg.v_graphnet((vec, A)))
l.problem.assign_penalty(l1=l1, l2=l2, l3=l3)
l.fit(tol=tol, max_its=5000)
beta = l.problem.coefficients
lv2 = regreg.cwpath(regreg.lin_graphnet((Xlist, Y, A)))
lv2.problem.assign_penalty(l1=l1, l2=l2, l3=l3)
lv2.fit(tol=tol, max_its=5000)
beta2 = lv2.problem.coefficients
if np.linalg.norm(beta) > 1e-8:
assert (np.linalg.norm(beta - beta2) / np.linalg.norm(beta) < 1.0e-04)
else:
assert (np.linalg.norm(beta2) < 1e-8)
def f(beta):
return -np.dot(vec, beta) + np.fabs(
beta).sum() * l1 + l2 * np.linalg.norm(beta)**2 / 2 + l3 * np.dot(
beta, np.dot(Afull, beta)) / 2
v = scipy.optimize.fmin_powell(f,
np.zeros(X.shape[1]),
ftol=1.0e-10,
xtol=1.0e-10,
maxfun=100000)
v = np.asarray(v)
#print np.round(100*v)/100,'\n', np.round(100*beta)/100
if f(v) < f(beta):
assert (np.fabs(f(v) - f(beta)) / np.fabs(f(v) + f(beta)) < 1.0e-04)
if np.linalg.norm(v) > 1e-8:
assert (np.linalg.norm(v - beta) / np.linalg.norm(v) < 1.0e-04)
else:
assert (np.linalg.norm(beta) < 1e-8)
def test_v_graphnet(X, Xlist, Y, l1=500.0, l2=2, l3=3.5, tol=1e-4):
print "v_graphnet", l1, l2, l3
A, Afull = gen_adj(X.shape[1])
vec = np.dot(Y, X)
l = regreg.cwpath(regreg.v_graphnet((vec, A)))
l.problem.assign_penalty(l1=l1, l2=l2, l3=l3)
l.fit(tol=tol, max_its=5000)
beta = l.problem.coefficients
lv2 = regreg.cwpath(regreg.lin_graphnet((Xlist, Y, A)))
lv2.problem.assign_penalty(l1=l1, l2=l2, l3=l3)
lv2.fit(tol=tol, max_its=5000)
beta2 = lv2.problem.coefficients
if np.linalg.norm(beta) > 1e-8:
assert np.linalg.norm(beta - beta2) / np.linalg.norm(beta) < 1.0e-04
else:
assert np.linalg.norm(beta2) < 1e-8
def f(beta):
return (
-np.dot(vec, beta)
+ np.fabs(beta).sum() * l1
+ l2 * np.linalg.norm(beta) ** 2 / 2
+ l3 * np.dot(beta, np.dot(Afull, beta)) / 2
)
v = scipy.optimize.fmin_powell(f, np.zeros(X.shape[1]), ftol=1.0e-10, xtol=1.0e-10, maxfun=100000)
v = np.asarray(v)
# print np.round(100*v)/100,'\n', np.round(100*beta)/100
if f(v) < f(beta):
assert np.fabs(f(v) - f(beta)) / np.fabs(f(v) + f(beta)) < 1.0e-04
if np.linalg.norm(v) > 1e-8:
assert np.linalg.norm(v - beta) / np.linalg.norm(v) < 1.0e-04
else:
assert np.linalg.norm(beta) < 1e-8
def test_lin_graphnet(X, Xlist, Y, l1=500.0, l2=2, l3=3.5, tol=1e-4):
print "lin_graphnet", l1, l2, l3
A, Afull = gen_adj(X.shape[1])
orth = np.random.normal(0, 1, X.shape[1])
eta = np.random.normal(0, 1, 1)[0]
# l = regreg.regreg((Xlist, Y, A),regreg.lin_graphnet,regreg.cwpath)#, initial_coefs= np.array([7.]*10))
l = regreg.cwpath(regreg.lin_graphnet((Xlist, Y, A)))
l.problem.assign_penalty(l1=l1, l2=l2, l3=l3, eta=eta)
l.problem.orth = orth
l.fit(tol=tol, max_its=5000)
beta = l.problem.coefficients
def f(beta):
q = np.dot(X, beta)
return (
-np.dot(Y, q)
+ np.fabs(beta).sum() * l1
+ l2 * np.linalg.norm(beta) ** 2 / 2
+ l3 * np.dot(beta, np.dot(Afull, beta)) / 2
+ eta * np.dot(beta, orth)
)
v = scipy.optimize.fmin_powell(f, np.zeros(X.shape[1]), ftol=1.0e-10, xtol=1.0e-10, maxfun=100000)
v = np.asarray(v)
print np.round(100 * v) / 100, "\n", np.round(100 * beta) / 100
print eta, np.sum(beta), np.sum(v)
if f(v) < f(beta):
assert np.fabs(f(v) - f(beta)) / np.fabs(f(v) + f(beta)) < 1.0e-04
if np.linalg.norm(v) > 1e-8:
assert np.linalg.norm(v - beta) / np.linalg.norm(v) < 1.0e-04
else:
assert np.linalg.norm(beta) < 1e-8
def test_lin_graphnet(X, Xlist, Y, l1=500., l2=2, l3=3.5, tol=1e-4):
print "lin_graphnet", l1, l2, l3
A, Afull = gen_adj(X.shape[1])
orth = np.random.normal(0, 1, X.shape[1])
eta = np.random.normal(0, 1, 1)[0]
#l = regreg.regreg((Xlist, Y, A),regreg.lin_graphnet,regreg.cwpath)#, initial_coefs= np.array([7.]*10))
l = regreg.cwpath(regreg.lin_graphnet((Xlist, Y, A)))
l.problem.assign_penalty(l1=l1, l2=l2, l3=l3, eta=eta)
l.problem.orth = orth
l.fit(tol=tol, max_its=5000)
beta = l.problem.coefficients
def f(beta):
q = np.dot(X, beta)
return -np.dot(Y, q) + np.fabs(
beta).sum() * l1 + l2 * np.linalg.norm(beta)**2 / 2 + l3 * np.dot(
beta, np.dot(Afull, beta)) / 2 + eta * np.dot(beta, orth)
v = scipy.optimize.fmin_powell(f,
np.zeros(X.shape[1]),
ftol=1.0e-10,
xtol=1.0e-10,
maxfun=100000)
v = np.asarray(v)
print np.round(100 * v) / 100, '\n', np.round(100 * beta) / 100
print eta, np.sum(beta), np.sum(v)
if f(v) < f(beta):
assert (np.fabs(f(v) - f(beta)) / np.fabs(f(v) + f(beta)) < 1.0e-04)
if np.linalg.norm(v) > 1e-8:
assert (np.linalg.norm(v - beta) / np.linalg.norm(v) < 1.0e-04)
else:
assert (np.linalg.norm(beta) < 1e-8)
def test_lasso(X, Xlist, Y, l1=500.0, tol=1e-4, M=0.0):
print "LASSO", l1
# l = regreg.regreg((Xlist, Y),regreg.lasso,regreg.cwpath)#, initial_coefs= np.array([7.]*10))
p1 = regreg.lasso((Xlist, Y))
p1.assign_penalty(l1=l1)
p2 = regreg.lasso((Xlist, Y))
p2.assign_penalty(l1=l1)
p3 = regreg.lasso((Xlist, Y))
p3.assign_penalty(l1=l1)
t1 = time.time()
o1 = regreg.cwpath(p1)
o1.fit(tol=tol, max_its=50)
beta1 = o1.problem.coefficients
t2 = time.time()
print "CWPATH", t2 - t1
t1 = time.time()
o2 = regreg.gengrad(p2)
o2.fit(M, tol=1e-10, max_its=1500)
beta2 = o2.problem.coefficients
t2 = time.time()
print "GENGRAD", t2 - t1
epsvec = [1e-0, 1e-3, 1e-6]
t1 = time.time()
o3 = regreg.nesterov(p3)
for eps in epsvec:
f_s = o3.fit(M, tol=1e-10, max_its=150, epsilon=eps)
f_s = o3.fit(M, tol=1e-10, max_its=5000, epsilon=eps)
beta3 = o3.problem.coefficients
t2 = time.time()
print "NEST", t2 - t1
def f(beta):
return np.linalg.norm(Y - np.dot(X, beta)) ** 2 / (2 * len(Y)) + np.fabs(beta).sum() * l1
v = scipy.optimize.fmin_powell(f, np.zeros(X.shape[1]), ftol=1.0e-10, xtol=1.0e-10, maxfun=100000)
v = np.asarray(v)
vs = scipy.optimize.fmin_powell(f_s, np.zeros(X.shape[1]), ftol=1.0e-10, xtol=1.0e-10, maxfun=100000)
vs = np.asarray(vs)
print np.round(10000 * beta1) / 10000
print np.round(10000 * beta2) / 10000
print np.round(10000 * beta3) / 10000
print "\n", np.round(10000 * v) / 10000
print np.round(10000 * vs) / 10000
print f_s(beta3), f_s(vs)
assert np.fabs(f(beta1) - f(beta2)) / np.fabs(f(beta1) + f(beta1)) < 1.0e-04
if np.linalg.norm(beta1) > 1e-8:
assert np.linalg.norm(beta2 - beta1) / np.linalg.norm(beta1) < 1.0e-04
else:
assert np.linalg.norm(beta2) < 1e-8
if f(v) < f(beta1):
assert np.fabs(f(v) - f(beta1)) / np.fabs(f(v) + f(beta1)) < 1.0e-04
if np.linalg.norm(v) > 1e-8:
assert np.linalg.norm(v - beta1) / np.linalg.norm(v) < 1.0e-04
else:
assert np.linalg.norm(beta1) < 1e-8
def test_lasso(X, Xlist, Y, l1=500., tol=1e-4, M=0.):
print "LASSO", l1
#l = regreg.regreg((Xlist, Y),regreg.lasso,regreg.cwpath)#, initial_coefs= np.array([7.]*10))
p1 = regreg.lasso((Xlist, Y))
p1.assign_penalty(l1=l1)
p2 = regreg.lasso((Xlist, Y))
p2.assign_penalty(l1=l1)
p3 = regreg.lasso((Xlist, Y))
p3.assign_penalty(l1=l1)
t1 = time.time()
o1 = regreg.cwpath(p1)
o1.fit(tol=tol, max_its=50)
beta1 = o1.problem.coefficients
t2 = time.time()
print "CWPATH", t2 - t1
t1 = time.time()
o2 = regreg.gengrad(p2)
o2.fit(M, tol=1e-10, max_its=1500)
beta2 = o2.problem.coefficients
t2 = time.time()
print "GENGRAD", t2 - t1
epsvec = [1e-0, 1e-3, 1e-6]
t1 = time.time()
o3 = regreg.nesterov(p3)
for eps in epsvec:
f_s = o3.fit(M, tol=1e-10, max_its=150, epsilon=eps)
f_s = o3.fit(M, tol=1e-10, max_its=5000, epsilon=eps)
beta3 = o3.problem.coefficients
t2 = time.time()
print "NEST", t2 - t1
def f(beta):
return np.linalg.norm(Y - np.dot(X, beta))**2 / (
2 * len(Y)) + np.fabs(beta).sum() * l1
v = scipy.optimize.fmin_powell(f,
np.zeros(X.shape[1]),
ftol=1.0e-10,
xtol=1.0e-10,
maxfun=100000)
v = np.asarray(v)
vs = scipy.optimize.fmin_powell(f_s,
np.zeros(X.shape[1]),
ftol=1.0e-10,
xtol=1.0e-10,
maxfun=100000)
vs = np.asarray(vs)
print np.round(10000 * beta1) / 10000
print np.round(10000 * beta2) / 10000
print np.round(10000 * beta3) / 10000
print "\n", np.round(10000 * v) / 10000
print np.round(10000 * vs) / 10000
print f_s(beta3), f_s(vs)
assert (np.fabs(f(beta1) - f(beta2)) / np.fabs(f(beta1) + f(beta1)) <
1.0e-04)
if np.linalg.norm(beta1) > 1e-8:
assert (np.linalg.norm(beta2 - beta1) / np.linalg.norm(beta1) <
1.0e-04)
else:
assert (np.linalg.norm(beta2) < 1e-8)
if f(v) < f(beta1):
assert (np.fabs(f(v) - f(beta1)) / np.fabs(f(v) + f(beta1)) < 1.0e-04)
if np.linalg.norm(v) > 1e-8:
assert (np.linalg.norm(v - beta1) / np.linalg.norm(v) < 1.0e-04)
else:
assert (np.linalg.norm(beta1) < 1e-8)