def log_likelihood(self):
if self._normalizer is None:
self._normalizer, _, _ = kalman_filter(
self.mu_init, self.sigma_init,
self.A, self.sigma_states, self.C, self.sigma_obs,
self.data)
return self._normalizer
def sample_predictions(self, Tpred, states_noise, obs_noise):
_, filtered_mus, filtered_sigmas = kalman_filter(
self.mu_init, self.sigma_init, self.A, self.sigma_states, self.C,
self.sigma_obs, self.data)
init_mu = self.A.dot(filtered_mus[-1])
init_sigma = self.sigma_states + self.A.dot(
filtered_sigmas[-1]).dot(self.A.T)
randseq = np.zeros((Tpred-1, self.n))
if states_noise:
L = np.linalg.cholesky(self.sigma_states)
randseq += np.random.randn(Tpred-1, self.n).dot(L.T)
states = np.empty((Tpred, self.n))
states[0] = np.random.multivariate_normal(init_mu, init_sigma)
for t in xrange(1,Tpred):
states[t] = self.A.dot(states[t-1]) + randseq[t-1]
obs = states.dot(self.C.T)
if obs_noise:
L = np.linalg.cholesky(self.sigma_obs)
obs += np.random.randn(Tpred, self.p).dot(L.T)
return obs
def _filter(self):
self._normalizer, self.filtered_mus, self.filtered_sigmas = \
kalman_filter(
self.mu_init, self.sigma_init,
self.A, self.sigma_states,
self.C, self.d, self.sigma_obs,
self.data)
def sample_predictions(self, Tpred, states_noise, obs_noise):
_, filtered_mus, filtered_sigmas = kalman_filter(
self.mu_init, self.sigma_init, self.A, self.sigma_states, self.C,
self.sigma_obs, self.data)
init_mu = self.A.dot(filtered_mus[-1])
init_sigma = self.sigma_states + self.A.dot(filtered_sigmas[-1]).dot(
self.A.T)
randseq = np.zeros((Tpred - 1, self.n))
if states_noise:
L = np.linalg.cholesky(self.sigma_states)
randseq += np.random.randn(Tpred - 1, self.n).dot(L.T)
states = np.empty((Tpred, self.n))
states[0] = np.random.multivariate_normal(init_mu, init_sigma)
for t in xrange(1, Tpred):
states[t] = self.A.dot(states[t - 1]) + randseq[t - 1]
obs = states.dot(self.C.T)
if obs_noise:
L = np.linalg.cholesky(self.sigma_obs)
obs += np.random.randn(Tpred, self.p).dot(L.T)
return obs
def log_likelihood(self):
if self._normalizer is None:
self._normalizer, _, _ = kalman_filter(self.mu_init,
self.sigma_init, self.A,
self.sigma_states, self.C,
self.sigma_obs, self.data)
return self._normalizer
def _filter(self):
self._normalizer, self.filtered_mus, self.filtered_sigmas = \
kalman_filter(
self.mu_init, self.sigma_init,
self.A, self.sigma_states,
self.C, self.d, self.sigma_obs,
self.data)
def sample_predictions(self, Tpred, Npred, obs_noise=True):
_, filtered_mus, filtered_sigmas = kalman_filter(
self.mu_init, self.sigma_init, self.A, self.sigma_states, self.C,
self.sigma_obs, self.data)
init_mu = self.A.dot(filtered_mus[-1])
init_sigma = self.sigma_states + self.A.dot(filtered_sigmas[-1]).dot(
self.A.T)
randseq = np.einsum('tjn,ij->tin',
np.random.randn(Tpred - 1, self.n, Npred),
np.linalg.cholesky(self.sigma_states))
states = np.empty((Tpred, self.n, Npred))
states[0] = np.random.multivariate_normal(init_mu,
init_sigma,
size=Npred).T
for t in xrange(1, Tpred):
states[t] = self.A.dot(states[t - 1]) + randseq[t - 1]
obs = np.einsum('ij,tjn->tin', self.C, states)
if obs_noise:
obs += np.einsum('tjn,ij->tin',
np.random.randn(Tpred, self.p, Npred),
np.linalg.cholesky(self.sigma_obs))
return obs
def _ll_diag(self):
if self._normalizer is None:
self._normalizer, _, _ = kalman_filter(
self.mu_init, self.sigma_init,
self.A, self.sigma_states,
self.C, self.d, self.sigma_obs,
self.data)
return self._normalizer
def _ll_diag(self):
if self._normalizer is None:
self._normalizer, _, _ = kalman_filter(self.mu_init,
self.sigma_init, self.A,
self.sigma_states, self.C,
self.d, self.sigma_obs,
self.data)
return self._normalizer
