def score(self, x, y, x_prev, n):
"""
Score a particular proposal
"""
x_det = x_prev['x'] + x_prev['xdot'] * self.DELTA_T
y_det = x_prev['y'] + x_prev['ydot'] * self.DELTA_T
score = 0.0
nd = ssm.util.log_norm_dens
score += nd(x['x'], x_det, self.POS_NOISE_STD**2)
score += nd(x['y'], y_det, self.POS_NOISE_STD**2)
score += nd(x['xdot'], x_prev['xdot'], self.VELOCITY_NOISE_STD**2)
score += nd(x['ydot'], x_prev['ydot'], self.VELOCITY_NOISE_STD**2)
score += util.log_vonmises_dens(x['phi'], x_prev['phi'],
self.PHI_NOISE_STD**2)
# now the theta likelihood is fun because it's like, the product
# of two things
t_o = x['theta'] - self.THETA_OFFSET
tn_o = x_prev['theta'] - self.THETA_OFFSET
score += nd(tn_o, t_o, self.THETA_DRIFT_SIZE)
score += nd(tn_o, 0, self.THETA_ENVELOPE_SIZE)
return score
def score(self, x, y, x_prev, n):
"""
Score a particular proposal
"""
candidate_points = self.cached_candidate_points(y, x_prev, n)
MIX_COMP = len(candidate_points)
if MIX_COMP == 0 or n > self.DEBUG_STOP_PROPOSING:
return self.base_proposal.score(x, y, x_prev, n)
score = 0.0
nd = ssm.util.log_norm_dens
scores = np.zeros(MIX_COMP)
for mci, mc in enumerate(candidate_points):
score = 0.0
score += nd(x['x'], mc['x'], self.POS_STD**2)
score += nd(x['y'], mc['y'], self.POS_STD**2)
# FIXME shoudl be von mises
score += util.log_vonmises_dens(x['phi'], mc['phi'],
self.PHI_STD**2)
scores[mci] = score
const = np.ones(MIX_COMP) * np.log(1. / MIX_COMP)
scores += const
score_accum = scores[0]
for s in scores[1:]:
score_accum = np.logaddexp(score_accum, s)
return score_accum
def score_trans(self, x, xn, i):
x_det = x["x"] + x["xdot"] * self.DELTA_T
y_det = x["y"] + x["ydot"] * self.DELTA_T
score = 0.0
nd = ssm.util.log_norm_dens
score += nd(xn["x"], x_det, self.POS_NOISE_STD ** 2)
score += nd(xn["y"], y_det, self.POS_NOISE_STD ** 2)
score += nd(xn["xdot"], x["xdot"], self.VELOCITY_NOISE_STD ** 2)
score += nd(xn["ydot"], x["ydot"], self.VELOCITY_NOISE_STD ** 2)
score += util.log_vonmises_dens(xn["phi"], x["phi"], self.PHI_NOISE_STD ** 2)
# now the theta likelihood is fun because it's like, the product
# of two things
t_o = x["theta"] - self.THETA_OFFSET
tn_o = xn["theta"] - self.THETA_OFFSET
score += nd(tn_o, t_o, self.THETA_DRIFT_SIZE)
score += nd(tn_o, 0, self.THETA_ENVELOPE_SIZE)
# FIXME automatic validation of this somehow? generate a shit-ton
# of samples and then measure marginals?
return score
def score(self, x, y, x_prev, n):
"""
Score a particular proposal
"""
x_det = x_prev['x'] + x_prev['xdot'] * self.DELTA_T
y_det = x_prev['y'] + x_prev['ydot'] * self.DELTA_T
score = 0.0
nd = ssm.util.log_norm_dens
score += nd(x['x'], x_det, self.POS_NOISE_STD**2)
score += nd(x['y'], y_det, self.POS_NOISE_STD**2)
score += nd(x['xdot'], x_prev['xdot'],
self.VELOCITY_NOISE_STD**2)
score += nd(x['ydot'], x_prev['ydot'],
self.VELOCITY_NOISE_STD**2)
score += util.log_vonmises_dens(x['phi'], x_prev['phi'],
self.PHI_NOISE_STD**2)
# now the theta likelihood is fun because it's like, the product
# of two things
t_o = x['theta'] - self.THETA_OFFSET
tn_o = x_prev['theta'] - self.THETA_OFFSET
score += nd(tn_o, t_o, self.THETA_DRIFT_SIZE)
score += nd(tn_o, 0, self.THETA_ENVELOPE_SIZE)
return score
def score(self, x, y, x_prev, n):
"""
Score a particular proposal
"""
candidate_points = self.cached_candidate_points(y, x_prev, n)
MIX_COMP = len(candidate_points)
if MIX_COMP == 0 or n > self.DEBUG_STOP_PROPOSING:
return self.base_proposal.score(x, y, x_prev, n)
score = 0.0
nd = ssm.util.log_norm_dens
scores = np.zeros(MIX_COMP)
for mci, mc in enumerate(candidate_points):
score = 0.0
score += nd(x['x'], mc['x'], self.POS_STD**2)
score += nd(x['y'], mc['y'], self.POS_STD**2)
# FIXME shoudl be von mises
score += util.log_vonmises_dens(x['phi'],
mc['phi'], self.PHI_STD**2)
scores[mci] = score
const = np.ones(MIX_COMP) * np.log(1./MIX_COMP)
scores += const
score_accum = scores[0]
for s in scores[1:]:
score_accum = np.logaddexp(score_accum, s)
return score_accum
def score_trans(self, x, xn, i):
x_det = x['x'] + x['xdot'] * self.DELTA_T
y_det = x['y'] + x['ydot'] * self.DELTA_T
score = 0.0
nd = ssm.util.log_norm_dens
score += nd(xn['x'], x_det, self.POS_NOISE_STD**2)
score += nd(xn['y'], y_det, self.POS_NOISE_STD**2)
score += nd(xn['xdot'], x['xdot'], self.VELOCITY_NOISE_STD**2)
score += nd(xn['ydot'], x['ydot'], self.VELOCITY_NOISE_STD**2)
score += util.log_vonmises_dens(xn['phi'], x['phi'],
self.PHI_NOISE_STD**2)
# now the theta likelihood is fun because it's like, the product
# of two things
t_o = x['theta'] - self.THETA_OFFSET
tn_o = xn['theta'] - self.THETA_OFFSET
score += nd(tn_o, t_o, self.THETA_DRIFT_SIZE)
score += nd(tn_o, 0, self.THETA_ENVELOPE_SIZE)
# FIXME automatic validation of this somehow? generate a shit-ton
# of samples and then measure marginals?
return score
