def sis_r(db, ob, particles):
num_particles = cmd_args.num_particles
weights = torch.ones([num_particles, 1],
dtype=torch.float).to(DEVICE) / num_particles
# re-weight
reweight = torch.zeros(weights.shape, dtype=weights.dtype).to(DEVICE)
for ob_i in ob:
reweight += db.log_likelihood(ob_i.reshape([1, -1]), particles)
reweight = torch.exp(reweight)
weights = weights * reweight
# re-sample
weights = weights / torch.sum(weights)
mnrnd = torch.distributions.multinomial.Multinomial(
num_particles, torch.reshape(torch.relu(weights), [num_particles]))
idx = mnrnd.sample().int()
nonzero_cluster = idx.nonzero()
indx = 0
new_xi = torch.zeros(particles.shape, dtype=particles.dtype)
for iter in range(nonzero_cluster.shape[0]):
nonzero_idx = nonzero_cluster[iter][0]
new_xi[indx:indx + idx[nonzero_idx], :] = particles[
nonzero_cluster[iter][0]].repeat(int(idx[nonzero_idx]), 1)
indx += idx[nonzero_cluster[iter][0]]
particles = new_xi
ess = torch.sum(weights)**2 / torch.sum(weights * weights)
if ess <= cmd_args.threshold * num_particles:
# generate new location
kde = KDE(particles)
particles = kde.get_samples(cmd_args.num_particles)
return particles
def onepass_smc(db, ob, particles, alpha):
num_particles = cmd_args.num_particles
# re-weight
reweight = torch.zeros(alpha.shape, dtype=alpha.dtype)
for ob_i in ob:
reweight += db.log_likelihood(ob_i.reshape([1, -1]),
particles).view(reweight.size())
reweight = torch.exp(reweight)
alpha = alpha * reweight
# re-sample
ess = torch.sum(alpha)**2 / torch.sum(alpha * alpha)
if ess <= cmd_args.threshold * num_particles:
alpha = alpha / torch.sum(alpha)
mnrnd = torch.distributions.multinomial.Multinomial(
num_particles, torch.reshape(torch.abs(alpha), [num_particles]))
idx = mnrnd.sample().int()
nonzero_cluster = idx.nonzero()
indx = 0
new_xi = torch.zeros(particles.shape, dtype=particles.dtype).to(DEVICE)
for iter in range(nonzero_cluster.shape[0]):
nonzero_idx = nonzero_cluster[iter][0]
new_xi[indx:indx + idx[nonzero_idx], :] = particles[
nonzero_cluster[iter][0]].repeat(int(idx[nonzero_idx]), 1)
indx += idx[nonzero_cluster[iter][0]]
# generate new location
kde = KDE(new_xi)
xi = kde.get_samples(cmd_args.num_particles)
alpha = torch.ones([num_particles, 1],
dtype=torch.float) / num_particles
else:
alpha = alpha / torch.sum(alpha)
xi = particles
return xi, alpha