def _resample(self, swarm):
log_W = swarm.log_weights
log_W[np.isneginf(log_W)] = -1e10
log_l = scipy.optimize.bisect(_resample_opt_func,
log_W.min(),
1000,
args=(np.log(self.num_particles), log_W))
new_swarm = ParticleSwarm()
if log_W[0] >= log_l:
new_swarm.add_particle(log_W[0], swarm[0])
else:
new_swarm.add_particle(log_l, swarm[0])
for i in range(1, swarm.num_particles):
if log_W[i] >= log_l:
new_swarm.add_particle(log_W[i], swarm[i])
else:
if bernoulli_rvs_log(log_W[i] - log_l):
new_swarm.add_particle(log_l, swarm[i])
return new_swarm
def _update_row(self, cols, data, dist, feat_probs, params, row_idx, test_path):
T = len(cols)
conditional_path = params.Z[row_idx, cols].copy()
params.Z[row_idx, cols] = test_path
log_feat_probs = np.row_stack([np.log1p(-feat_probs), np.log(feat_probs)])
swarm = ParticleSwarm()
for _ in range(self.num_particles):
swarm.add_particle(0, None)
for t in range(T):
if t > 0:
try:
swarm = self._resample(swarm)
except ValueError:
params.Z[row_idx, cols] = conditional_path
return params
assert np.all(swarm.particles[0].path == conditional_path[:t])
new_swarm = ParticleSwarm()
annealing_factor = self._get_annealing_factor(t, T)
col = cols[t]
for i, (parent_particle, log_W) in enumerate(zip(swarm.particles, swarm.log_weights)):
if parent_particle is not None:
params.Z[row_idx, cols[:t]] = parent_particle.path
if i == 0:
value = conditional_path[t]
else:
value = None
particle = self._get_new_particle(
annealing_factor, col, data, dist, log_feat_probs, params, parent_particle, row_idx, value=value
)
new_swarm.add_particle(log_W + particle.log_w, particle)
swarm = new_swarm
params.Z[row_idx, cols] = swarm.sample().path
return params
def _update_row(self, cols, data, dist, feat_probs, params, row_idx,
test_path):
T = len(cols)
conditional_path = params.Z[row_idx, cols].copy()
params.Z[row_idx, cols] = test_path
log_feat_probs = np.row_stack(
[np.log1p(-feat_probs), np.log(feat_probs)])
swarm = ParticleSwarm()
swarm.add_particle(0, None)
for t in range(T):
if swarm.num_particles > self.num_particles:
swarm = self._resample(swarm)
new_swarm = ParticleSwarm()
annealing_factor = self._get_annealing_factor(t, T)
col = cols[t]
states = [conditional_path[t], 1 - conditional_path[t]]
for log_W, parent_particle in zip(swarm.log_weights,
swarm.particles):
if parent_particle is not None:
params.Z[row_idx, cols[:t]] = parent_particle.path
for s in states:
particle = self._get_new_particle(annealing_factor, col,
data, dist,
log_feat_probs, params,
parent_particle, row_idx,
s)
new_swarm.add_particle(log_W + particle.log_w, particle)
swarm = new_swarm
if swarm.num_particles > self.max_particles:
self.max_particles = swarm.num_particles
assert np.all(swarm[0].path == conditional_path)
params.Z[row_idx, cols] = swarm.sample().path
return params
def _resample(self, swarm):
if swarm.relative_ess <= self.resample_threshold:
new_swarm = ParticleSwarm()
if self.resample_scheme == 'multinomial':
idxs = multinomial_resampling(swarm.unnormalized_log_weights, self.num_particles)
elif self.resample_scheme == 'stratified':
idxs = stratified_resampling(swarm.unnormalized_log_weights, self.num_particles)
else:
raise Exception('Unknown resampling scheme: {}'.format(self.resample_scheme))
for idx in idxs:
new_swarm.add_particle(0, swarm.particles[idx])
else:
new_swarm = swarm
return new_swarm
def _update_row(self, cols, data, dist, feat_probs, params, row_idx, test_path):
T = len(cols)
old_path = params.Z[row_idx, cols].copy()
params.Z[row_idx, cols] = test_path
log_feat_probs = np.row_stack([np.log1p(-feat_probs), np.log(feat_probs)])
swarm = ParticleSwarm()
for _ in range(self.num_particles):
swarm.add_particle(0, None)
for t in range(T):
if t > 0:
try:
swarm = self._resample(swarm)
except ValueError:
params.Z[row_idx, cols] = old_path
return params
new_swarm = ParticleSwarm()
annealing_factor = self._get_annealing_factor(t, T)
col = cols[t]
for parent_particle, log_W in zip(swarm.particles, swarm.log_weights):
if parent_particle is not None:
params.Z[row_idx, cols[:t]] = parent_particle.path
particle = self._get_new_particle(
annealing_factor, col, data, dist, log_feat_probs, params, parent_particle, row_idx, value=None
)
new_swarm.add_particle(log_W + particle.log_w, particle)
swarm = new_swarm
params.Z[row_idx, cols] = swarm.sample().path
return params