def _impulse(self, dt, mu, tau):
from pyhawkes.utils.utils import logit
# map dt onto support
x = self.translate_dt(dt)
# get prob at scaled x
Z = x * (1. - x) * np.sqrt(2 * np.pi / tau)
return (1. / Z) * np.exp(-tau / 2. * (logit(x) - mu)**2)
def impulse(self, dt, k1, k2):
"""
Impulse response induced by an event on process k1 on
the rate of process k2 at lag dt
"""
from pyhawkes.utils.utils import logit
mu, tau, dt_max = self.mu[k1,k2], self.tau[k1,k2], self.dt_max
Z = dt * (dt_max - dt)/dt_max * np.sqrt(2*np.pi/tau)
return 1./Z * np.exp(-tau/2. * (logit(dt/dt_max) - mu)**2)
def impulse(self, dt, k1, k2):
"""
Impulse response induced by an event on process k1 on
the rate of process k2 at lag dt
"""
from pyhawkes.utils.utils import logit
mu, tau, dt_max = self.mu[k1, k2], self.tau[k1, k2], self.dt_max
Z = dt * (dt_max - dt) / dt_max * np.sqrt(2 * np.pi / tau)
return 1. / Z * np.exp(-tau / 2. * (logit(dt / dt_max) - mu)**2)
def impulse(self, dt, k1, k2, delay=0):
"""
Impulse response induced by an event on process k1 on
the rate of process k2 at lag dt
"""
# delay added
dt = dt - self.delay[k1][k2]
from pyhawkes.utils.utils import logit
mu, tau, dt_max = self.mu[k1, k2], self.tau[k1, k2], self.dt_max
Z = dt * (dt_max - dt) / dt_max * np.sqrt(2 * np.pi / tau)
temp = 1. / Z * np.exp(-tau / 2. * (logit(dt / dt_max) - mu)**2)
# should be -inf
temp[dt < 0] = 0
return temp
def meanfieldupdate_p(self, stepsize=1.0):
"""
Update p given the network parameters and the current variational
parameters of the weight distributions.
:return:
"""
logit_p = self.network.expected_log_p() - self.network.expected_log_notp()
logit_p += self.network.kappa * self.network.expected_log_v() - gammaln(self.network.kappa)
logit_p += gammaln(self.mf_kappa_1) - self.mf_kappa_1 * np.log(self.mf_v_1)
logit_p += gammaln(self.kappa_0) - self.kappa_0 * np.log(self.nu_0)
logit_p += self.mf_kappa_0 * np.log(self.mf_v_0) - gammaln(self.mf_kappa_0)
# p_hat = logistic(logit_p)
# self.mf_p = (1.0 - stepsize) * self.mf_p + stepsize * p_hat
logit_p_hat = (1-stepsize) * logit(self.mf_p) + \
stepsize * logit_p
# self.mf_p = logistic(logit_p_hat)
self.mf_p = np.clip(logistic(logit_p_hat), 1e-8, 1-1e-8)
def meanfieldupdate_p(self, stepsize=1.0):
"""
Update p given the network parameters and the current variational
parameters of the weight distributions.
:return:
"""
logit_p = self.network.expected_log_p(
) - self.network.expected_log_notp()
logit_p += self.network.kappa * self.network.expected_log_v(
) - gammaln(self.network.kappa)
logit_p += gammaln(
self.mf_kappa_1) - self.mf_kappa_1 * np.log(self.mf_v_1)
logit_p += gammaln(self.kappa_0) - self.kappa_0 * np.log(self.nu_0)
logit_p += self.mf_kappa_0 * np.log(self.mf_v_0) - gammaln(
self.mf_kappa_0)
# p_hat = logistic(logit_p)
# self.mf_p = (1.0 - stepsize) * self.mf_p + stepsize * p_hat
logit_p_hat = (1-stepsize) * logit(self.mf_p) + \
stepsize * logit_p
# self.mf_p = logistic(logit_p_hat)
self.mf_p = np.clip(logistic(logit_p_hat), 1e-8, 1 - 1e-8)
def resample(self, data=[]):
"""
Resample the
:param data: a TxKxKxB array of parents. T time bins, K processes,
K parent processes, and B bases for each parent process.
"""
mu_0, lmbda_0, alpha_0, beta_0 = self.mu_0, self.lmbda_0, self.alpha_0, self.beta_0
assert data is None or isinstance(data, list)
# 0: count, # 1: Sum of scaled dt, #2: Sum of sq scaled dt
ss = np.zeros((3, self.K, self.K))
for d in data:
ss += d.compute_imp_suff_stats()
n = ss[0]
xbar = np.nan_to_num(ss[1] / n)
xvar = ss[2]
alpha_post = alpha_0 + n / 2.
# beta_post = beta_0 + 0.5 * xvar
# beta_post += 0.5 * lmbda_0 * n / (lmbda_0 + n) * (xbar-mu_0)**2
beta_post = beta_0 + 0.5 * xvar + 0.5 * lmbda_0 * n * (
xbar - mu_0)**2 / (lmbda_0 + n)
lmbda_post = lmbda_0 + n
mu_post = (lmbda_0 * mu_0 + n * xbar) / (lmbda_0 + n)
from pyhawkes.utils.utils import sample_nig
self.mu, self.tau = \
sample_nig(mu_post, lmbda_post, alpha_post, beta_post)
# Z is parent
# C is process
# S is event time
S = [x.S for x in self.model.data_list]
C = [x.C for x in self.model.data_list]
Z = [x.Z for x in self.model.data_list]
N_pts = 50
t = np.linspace(0, self.dt_max, N_pts)
pt = np.zeros(N_pts)
normal_sig = self.normal_sig
normal_mu = self.normal_mu
pt_prior = np.exp(-(t - normal_mu)**2 / (2 * normal_sig**2)) / np.sqrt(
2 * np.pi * normal_sig**2)
dt_max = self.dt_max
for k1 in range(self.K):
for k2 in range(self.K):
tt = np.zeros(N_pts)
gt = np.ones(N_pts)
for i in range(len(S)):
s, c, z = S[i], C[i], Z[i]
cind, pind = (c == k2), (c[z] == k1)
igmask = (z != -1) # no parent
inds = cind & pind & igmask
if ~np.all(~inds):
ds = s[inds] - s[z[inds]]
'''
if min(ds) < min(t):
print k1,k2
exit(0)
'''
ll = -utils.logit(
np.absolute(ds[None, :] - t[:, None]), dt_max)
# logit grows too fast
ll = np.minimum(ll, 100)
ll = np.maximum(ll, -100)
ll = (ll * self.tau[k1][k2] / 2 - self.mu[k1][k2])**2
tt = tt + np.sum(ll, 1)
gt[t > min(ds)] = 0.0
tt = np.cumsum(pt_prior * np.exp(tt - np.max(tt)) * gt)
if (tt == 0).all():
self.delay[k1][k2] = 0
#tt = np.cumsum(pt_prior)
#delay = t[np.flatnonzero(tt > np.random.uniform(0,tt[-1]))[0]]
#self.delay[k1][k2] = delay
else:
# exp grows too fast, normalize by e^-max(tt)
delay = -t[np.flatnonzero(
tt > np.random.uniform(0, tt[-1]))[0]]
self.delay[k1][k2] = delay
assert np.isfinite(self.mu).all()
assert np.isfinite(self.tau).all()
