def fit_quasi_newton(self,
lamb_l2,
lamb_smth,
lamb_log,
w_init,
factr=1e7,
maxls=50,
m=10,
lb=-np.Inf,
ub=np.Inf,
maxiter=15000,
verbose=True):
"""
"""
obj = Objective(self.sumstats, self.graph)
obj.lamb_l2 = lamb_l2
obj.lamb_smth = lamb_smth
obj.lamb_log = lamb_log
res = fmin_l_bfgs_b(func=loss_wrapper,
x0=np.log(w_init),
args=[obj],
factr=factr,
m=m,
maxls=maxls,
maxiter=maxiter,
approx_grad=False,
bounds=[(lb, ub)
for _ in range(obj.graph.Delta.shape[1])])
if maxiter >= 100:
assert res[2]["warnflag"] == 0, "did not converge"
self.graph.w = np.exp(res[0])
# print update
self.train_loss, _ = loss_wrapper(res[0], obj)
self.train_nll = neg_log_lik_wrapper(res[0], obj)
self.pen = self.train_loss - self.train_nll
if verbose:
sys.stdout.write(
("lambda_l2={:.7f}, "
"lambda={:.7f}, "
"alpha={:.7f}, "
"converged in {} iterations, "
"train_loss={:.7f}\n").format(lamb_l2, lamb_smth, lamb_log,
res[2]["nit"], self.train_loss))
def comp_fit_cov(feems,
lamb_l2,
lamb_smth,
lamb_log,
projection=True,
include_var=False,
ind_level=False):
# create obj
obj = Objective(feems.sumstats, feems.graph)
obj.lamb_l2 = lamb_l2
obj.lamb_smth = lamb_smth
obj.lamb_log = lamb_log
# update laplacian
obj.graph.comp_lap(obj.graph.w)
# update nll and grad
obj.inv()
obj.grad()
if ind_level is True:
# number of individuals
n, p = obj.sumstats.data.shape
# row index of assn matrix
row = np.arange(n)
# col index of assn matrix
col = obj.graph.obs_ids
# fill up J
J = sp.csc_matrix(
(np.ones(n), (row, col)),
shape=(n, obj.graph.d))[:, obj.graph.perm_ids][:, :obj.graph.o]
# diagonal component
q_full_samples = 1.0 / obj.sumstats.s2 * np.ones(obj.graph.o)
# fitted covariance
fit_cov = J @ (obj.Linv_block['oo'] - 1 / obj.graph.d) @ J.T + np.diag(
J @ (1. / q_full_samples))
# empirical covariance
emp_cov = obj.sumstats.data @ obj.sumstats.data.T / p
else:
# fitted covariance
fit_cov = obj.Linv_block[
'oo'] - 1 / obj.graph.d + obj.sumstats.q_inv_diag.toarray()
# empirical covariance
emp_cov = obj.sumstats.S
# project to the space of contrast
if projection is True:
C = comp_contrasts(n) if ind_level is True else comp_contrasts(
obj.graph.o)
fit_cov = C @ fit_cov @ C.T
emp_cov = C @ emp_cov @ C.T
if include_var is True:
return (np.triu(fit_cov, k=0), np.triu(emp_cov, k=0))
else:
return (np.triu(fit_cov, k=1), np.triu(emp_cov, k=1))
