def eval_nllh(beta):
llh = l2_clogistic_llh(X=self.solver.grouping_designs["judge"],
Y=self.solver.response,
alpha=alpha,
beta=beta,
offset=None,
penalty_matrix=self.solver.sparse_inv_covs["judge"].sparse_matrix)
return -1 * llh
def eval_nllh(alpha_beta):
alpha = alpha_beta[:4]
beta = alpha_beta[4:]
llh = l2_clogistic_llh(X=self.solver.main_design,
Y=self.solver.response,
alpha=alpha,
beta=beta,
offset=None,
penalty_matrix=PI)
return -1 * llh
def test_gradient_ranef(self, h=1e-6):
" Use finite differences to test gradient of random effects "
grad = \
l2_clogistic_gradient(X=self.solver.grouping_designs["judge"],
Y=self.solver.response,
alpha=self.results["intercepts"],
beta=self.results["judge"]["intercept"].values,
penalty_matrix=self.solver.sparse_inv_covs["judge"].sparse_matrix,
# adjust for main effects
offset=self.solver.main_design.dot(self.results["main"]["main_value"].values))
# this function includes derivatives of the intercepts, which we checked above
# so restrict to just the random effects
grad = grad[4:]
for i in xrange(len(grad)):
delta = np.zeros(len(grad))
delta[i] = h
grad_fwd = \
l2_clogistic_llh(X=self.solver.grouping_designs["judge"],
Y=self.solver.response,
alpha=self.results["intercepts"],
beta=self.results["judge"]["intercept"].values + delta,
penalty_matrix=self.solver.sparse_inv_covs["judge"].sparse_matrix,
offset=self.solver.main_design.dot(self.results["main"]["main_value"].values)
)
grad_bwd = \
l2_clogistic_llh(X=self.solver.grouping_designs["judge"],
Y=self.solver.response,
alpha=self.results["intercepts"],
beta=self.results["judge"]["intercept"].values - delta,
penalty_matrix=self.solver.sparse_inv_covs["judge"].sparse_matrix,
offset=self.solver.main_design.dot(self.results["main"]["main_value"].values)
)
grad_fd = (grad_fwd - grad_bwd) / (2 * h)
self.assertAlmostEqual(grad[i], grad_fd, places=6)
def test_gradient_fixef(self, h=1e-6):
" Use finite differences to test gradient of main effects "
grad = l2_clogistic_gradient(
X=self.solver.main_design,
Y=self.solver.response,
intercept=True,
beta=self.results["main"]["main_value"].values,
alpha=self.results["intercepts"],
# need to adjust for random effects
offset=self.solver.grouping_designs["judge"].dot(self.results["judge"]["intercept"].values)
)
for i in xrange(len(grad)):
delta = np.zeros(len(grad))
delta[i] = h
grad_fwd = \
l2_clogistic_llh(X=self.solver.main_design,
Y=self.solver.response,
alpha=self.results["intercepts"] + delta[:4],
beta=self.results["main"]["main_value"].values + delta[4:],
penalty_matrix=np.zeros((2, 2)),
offset=self.solver.grouping_designs["judge"].dot(
self.results["judge"]["intercept"].values)
)
grad_bwd = \
l2_clogistic_llh(X=self.solver.main_design,
Y=self.solver.response,
alpha=self.results["intercepts"] - delta[:4],
beta=self.results["main"]["main_value"].values - delta[4:],
penalty_matrix=np.zeros((2, 2)),
offset=self.solver.grouping_designs["judge"].dot(
self.results["judge"]["intercept"].values)
)
grad_fd = (grad_fwd - grad_bwd) / (2 * h)
self.assertAlmostEqual(grad[i], grad_fd, places=6)