class knockoff_logit(knockoff_net):
""" Preforms the knockoff technique with logistic regression """
def fit(self,X_lrg=None):
""" Generates the knockoffs, fits the regression, and performs the FDR calculations """
# Generate knockoff as inherited from knockoff_net
if X_lrg is None:
if self.knockoff_type == 'original': self._original_knockoff()
elif self.knockoff_type == 'binary': self._binary_knockoff()
else:
self.X_lrg = X_lrg
# initialize and fit the glmnet object
self.lognet = LogisticNet(alpha=1,n_lambdas=self.p*20,frac_lg_lambda=min(.000001,.01/(self.p**2)))
self.lognet.fit(self.X_lrg,self.y,normalize=False,include_intercept=self.intercept)
# pull out some values from the glmnet object and clean
self.lambdas = self.lognet.out_lambdas
self.var_index_ent = np.sort(self.lognet._indices)
self.coef_matrix = np.zeros((2*self.p,self.lognet.n_lambdas))
self.coef_matrix[self.var_index_ent] = self.lognet._comp_coef.squeeze()[self.lognet._indices]
# figure out when different variables entered the model
self.var_entered = np.zeros(2*self.p).astype(bool)
self.var_entered[self.var_index_ent] = True
# Preform all the FDR calculations as inherited from knockoff_net
self._get_z()
self._get_w()
self._get_T()
self._get_S()
def fit(self, X_lrg=None):
""" Generates the knockoffs, fits the regression, and performs the FDR calculations """
# Generate knockoff as inherited from knockoff_net
if X_lrg is None:
if self.knockoff_type == 'original': self._original_knockoff()
elif self.knockoff_type == 'binary': self._binary_knockoff()
else:
self.X_lrg = X_lrg
# initialize and fit the glmnet object
self.lognet = LogisticNet(alpha=1,
n_lambdas=self.p * 20,
frac_lg_lambda=min(.000001,
.01 / (self.p**2)))
self.lognet.fit(self.X_lrg,
self.y,
normalize=False,
include_intercept=self.intercept)
# pull out some values from the glmnet object and clean
self.lambdas = self.lognet.out_lambdas
self.var_index_ent = np.sort(self.lognet._indices)
self.coef_matrix = np.zeros((2 * self.p, self.lognet.n_lambdas))
self.coef_matrix[self.var_index_ent] = self.lognet._comp_coef.squeeze(
)[self.lognet._indices]
# figure out when different variables entered the model
self.var_entered = np.zeros(2 * self.p).astype(bool)
self.var_entered[self.var_index_ent] = True
# Preform all the FDR calculations as inherited from knockoff_net
self._get_z()
self._get_w()
self._get_T()
self._get_S()
def test_unregularized_models(self):
Xdn = np.random.uniform(-1, 1, size=(50,10))
Xsp = csc_matrix(Xdn)
w = np.random.uniform(-1, 1, size=(10,))
y = (np.dot(Xdn, w) >= 0).astype(int)
for alpha in [0, .5, 1]:
for X in (Xdn, Xsp):
lnet = LogisticNet(alpha=alpha)
lnet.fit(X, y, lambdas=[0])
preds = (lnet.predict(X) >= .5).astype(int)
self.assertTrue(np.all(y == preds))
def test_unregularized_models(self):
Xdn = np.random.uniform(-1, 1, size=(50, 10))
Xsp = csc_matrix(Xdn)
w = np.random.uniform(-1, 1, size=(10, ))
y = (np.dot(Xdn, w) >= 0).astype(int)
for alpha in [0, .5, 1]:
for X in (Xdn, Xsp):
lnet = LogisticNet(alpha=alpha)
lnet.fit(X, y, lambdas=[0])
preds = (lnet.predict(X) >= .5).astype(int)
self.assertTrue(np.all(y == preds))
def test_ridge_models(self):
Xdn = np.random.uniform(-1, 1, size=(50000, 3))
Xsp = csc_matrix(Xdn)
w = (np.random.uniform(-1, 1, size=(3, )) >= 0).astype(int) - .5
for X in (Xdn, Xsp):
for lam in np.linspace(.1, 1, 10):
y = (np.dot(Xdn, w) >= 0).astype(int)
lnet = LogisticNet(alpha=0)
lnet.fit(X, y, lambdas=[lam])
ratios = lnet._coefficients.ravel() / w
norm_ratios = ratios / np.max(ratios)
test = np.allclose(norm_ratios, 1, atol=.05)
self.assertTrue(test)
def test_max_lambda(self):
Xdn = np.random.uniform(-1, 1, size=(50, 10))
Xsp = csc_matrix(Xdn)
w = np.random.uniform(-1, 1, size=(10, ))
y = (np.dot(Xdn, w) >= 0).astype(int)
for X in (Xdn, Xsp):
for alpha in [.01, .5, 1]:
lnet = LogisticNet(alpha=alpha)
lnet.fit(X, y)
ol = lnet.out_lambdas
max_lambda_from_fortran = ol[1] * (ol[1] / ol[2])
max_lambda_from_python = lnet._max_lambda(X, y)
self.assertAlmostEqual(max_lambda_from_fortran,
max_lambda_from_python, 4)
def test_lasso_models(self):
Xdn = np.random.uniform(-1, 1, size=(15000, 10))
Xsp = csc_matrix(Xdn)
w = (np.random.uniform(-1, 1, size=(10, )) >= 0).astype(int) - .5
for w_mask in range(1, 10):
for X in (Xdn, Xsp):
w_masked = w.copy()
w_masked[w_mask:] = 0
y = (np.dot(Xdn, w_masked) >= 0).astype(int)
lnet = LogisticNet(alpha=1)
lnet.fit(X, y, lambdas=[.01])
lc = lnet._coefficients
self.assertTrue(
np.sum(np.abs(lc / np.max(np.abs(lc))) > .05) == w_mask)
def test_max_lambda(self):
Xdn = np.random.uniform(-1, 1, size=(50,10))
Xsp = csc_matrix(Xdn)
w = np.random.uniform(-1, 1, size=(10,))
y = (np.dot(Xdn, w) >= 0).astype(int)
for X in (Xdn, Xsp):
for alpha in [.01, .5, 1]:
lnet = LogisticNet(alpha=alpha)
lnet.fit(X, y)
ol = lnet.out_lambdas
max_lambda_from_fortran = ol[1] * (ol[1]/ol[2])
max_lambda_from_python = lnet._max_lambda(X, y)
self.assertAlmostEqual(
max_lambda_from_fortran, max_lambda_from_python, 4
)
def test_ridge_models(self):
Xdn = np.random.uniform(-1, 1, size=(50000,3))
Xsp = csc_matrix(Xdn)
w = (np.random.uniform(-1, 1, size=(3,)) >= 0).astype(int) - .5
for X in (Xdn, Xsp):
for lam in np.linspace(.1, 1, 10):
y = (np.dot(Xdn, w) >= 0).astype(int)
lnet = LogisticNet(alpha=0)
lnet.fit(X, y, lambdas=[lam])
ratios = lnet._coefficients.ravel() / w
norm_ratios = ratios / np.max(ratios)
test = np.allclose(
norm_ratios, 1, atol=.05
)
self.assertTrue(test)
def test_lasso_models(self):
Xdn = np.random.uniform(-1, 1, size=(15000,10))
Xsp = csc_matrix(Xdn)
w = (np.random.uniform(-1, 1, size=(10,)) >= 0).astype(int) - .5
for w_mask in range(1, 10):
for X in (Xdn, Xsp):
w_masked = w.copy()
w_masked[w_mask:] = 0
y = (np.dot(Xdn, w_masked) >= 0).astype(int)
lnet = LogisticNet(alpha=1)
lnet.fit(X, y, lambdas=[.01])
lc = lnet._coefficients
self.assertTrue(
np.sum(np.abs(lc / np.max(np.abs(lc))) > .05) == w_mask
)
def test_edge_cases(self):
'''Edge cases in model specification.'''
X = np.random.uniform(-1, 1, size=(50,10))
w = np.random.uniform(-1, 1, size=(10,))
y = (np.dot(X, w) >= 0).astype(int)
# Edge case
# A single lambda is so big that it sets all estimated coefficients
# to zero. This used to break the predict method.
lnet = LogisticNet(alpha=1)
lnet.fit(X, y, lambdas=[10**5])
_ = lnet.predict(X)
# Edge case
# Multiple lambdas are so big as to set all estiamted coefficients
# to zero. This used to break the predict method.
lnet = LogisticNet(alpha=1)
lnet.fit(X, y, lambdas=[10**5, 2*10**5])
_ = lnet.predict(X)
# Edge case:
# Some predictors have zero varaince. This used to break lambda
# max.
X = np.random.uniform(-1, 1, size=(50,10))
X[2,:] = 0; X[8,:] = 0
y = (np.dot(X, w) >= 0).astype(int)
lnet = LogisticNet(alpha=.1)
lnet.fit(X, y)
ol = lnet.out_lambdas
max_lambda_from_fortran = ol[1] * (ol[1]/ol[2])
max_lambda_from_python = lnet._max_lambda(X, y)
self.assertAlmostEqual(
max_lambda_from_fortran, max_lambda_from_python, 4
)
# Edge case.
# All predictors have zero variance. This is an error in
# sepcification.
with self.assertRaises(ValueError):
X = np.ones(shape=(50,10))
lnet = LogisticNet(alpha=.1)
lnet.fit(X, y)
def test_edge_cases(self):
'''Edge cases in model specification.'''
X = np.random.uniform(-1, 1, size=(50, 10))
w = np.random.uniform(-1, 1, size=(10, ))
y = (np.dot(X, w) >= 0).astype(int)
# Edge case
# A single lambda is so big that it sets all estimated coefficients
# to zero. This used to break the predict method.
lnet = LogisticNet(alpha=1)
lnet.fit(X, y, lambdas=[10**5])
_ = lnet.predict(X)
# Edge case
# Multiple lambdas are so big as to set all estiamted coefficients
# to zero. This used to break the predict method.
lnet = LogisticNet(alpha=1)
lnet.fit(X, y, lambdas=[10**5, 2 * 10**5])
_ = lnet.predict(X)
# Edge case:
# Some predictors have zero varaince. This used to break lambda
# max.
X = np.random.uniform(-1, 1, size=(50, 10))
X[2, :] = 0
X[8, :] = 0
y = (np.dot(X, w) >= 0).astype(int)
lnet = LogisticNet(alpha=.1)
lnet.fit(X, y)
ol = lnet.out_lambdas
max_lambda_from_fortran = ol[1] * (ol[1] / ol[2])
max_lambda_from_python = lnet._max_lambda(X, y)
self.assertAlmostEqual(max_lambda_from_fortran, max_lambda_from_python,
4)
# Edge case.
# All predictors have zero variance. This is an error in
# sepcification.
with self.assertRaises(ValueError):
X = np.ones(shape=(50, 10))
lnet = LogisticNet(alpha=.1)
lnet.fit(X, y)