def __init__(self, niter=10, verbose=10, theta0=None,
ssm_cls=None, prior=None, data=None, store_x=False):
for k in ['ssm_cls', 'prior', 'data', 'theta0', 'niter', 'store_x',
'verbose']:
setattr(self, k, locals()[k])
theta = np.empty(shape=niter, dtype=self.prior.dtype)
if store_x:
x = ssm_cls.state_container(niter, len(self.data))
self.chain = ssp.ThetaParticles(theta=theta, x=x)
else:
self.chain = ssp.ThetaParticles(theta=theta)
def __init__(self,
niter=10,
seed=None,
verbose=0,
theta0=None,
adaptive=True,
scale=1.,
rw_cov=None):
"""
Parameters
----------
niter: int
number of MCMC iterations
seed: int (default=None)
pseudo-random seed (if None, generator is not seeded)
verbose: int (default=0)
progress report printed every (niter/verbose) iterations (never if 0)
theta0: a structured array of size=1
starting point, simulated from the prior if set to None
adaptive: True/False
whether to use the adaptive version or not
scale: positive scalar (default = 1.)
in the adaptive case, covariance of the proposal is scale^2 times
(2.38 / d) times the current estimate of the target covariance
rw_cov: (d, d) array
covariance matrix of the random walk proposal (set to I_d if None)
"""
for k in ['niter', 'seed', 'verbose', 'theta0', 'adaptive']:
setattr(self, k, locals()[k])
self.chain = ssp.ThetaParticles(theta=np.empty(shape=niter,
dtype=self.prior.dtype),
lpost=np.empty(shape=niter))
self.nacc = 0
d = self.chain.dim
if self.adaptive:
optim_scale = 2.38 / np.sqrt(d)
self.scale = scale * optim_scale
self.cov_tracker = VanishCovTracker(dim=d, Sigma0=rw_cov)
self.L = self.scale * self.cov_tracker.L
else:
if rw_cov is None:
self.L = np.eye(d)
else:
self.L = cholesky(rw_cov, lower=True)
def step0(self):
th0 = self.prior.rvs(size=1) if self.theta0 is None else self.theta0
self.prop = ssp.ThetaParticles(theta=th0, lpost=np.zeros(1))
self.compute_post()
self.chain.copyto_at(0, self.prop, 0)
