def propose_assignments2(model, data, state, update=False):
N, K = state.Z.shape
state = state.copy()
cache = ibp.IBPCache.from_state(model, data, state, np.zeros(N, dtype=bool))
proposal_prob = 0.
for i in range(N):
obs = data.mask[i, :]
x = state.X[i, :]
evidence = np.zeros(4)
prior_odds = np.zeros(4)
for c, (z1, z2) in enumerate(CHOICES):
z = np.array([z1, z2])
mu = cache.fpost.predictive_mu(z)
ssq = cache.fpost.predictive_ssq(z) + state.sigma_sq_n
evidence[c] = ibp.gauss_loglik_vec_C2(x[obs], mu[obs], ssq)
for k in [0, 1]:
if cache.counts[k] > 0:
prior_odds[c] += np.log(cache.counts[k]) - np.log(cache.num_included - cache.counts[k] + 1)
else:
prior_odds[c] += np.log(model.alpha) - np.log(cache.num_included + 1)
odds = evidence + prior_odds
dist = distributions.MultinomialDistribution.from_odds(odds)
if update:
state.Z[i, :] = CHOICES[dist.sample().argmax()]
proposal_prob += dist.loglik(CHOICES.index(tuple(state.Z[i, :])))
cache.add(i, state.Z[i, :], state.X[i, :])
assert np.isfinite(proposal_prob)
return state, proposal_prob
def next_assignment_proposal(model, data, state, cache, Sigma_info, i, k):
assert not cache.rows_included[i]
x = state.X[i, :]
evidence = np.zeros(2)
for assignment in [0, 1]:
mu = Sigma_info.mu_for(k, assignment)
ssq = Sigma_info.sigma_sq_for(k, assignment) + state.sigma_sq_n
evidence[assignment] = ibp.gauss_loglik_vec_C2(x, mu, ssq)
data_odds = evidence[1] - evidence[0]
if cache.counts[k] > 0:
prior_odds = np.log(cache.counts[k]) - np.log(cache.num_included - cache.counts[k] + 1)
else:
#prior_odds = poisson(1, 0.5 * model.alpha / (i+1)) - poisson(0, 0.5 * model.alpha / (i+1))
prior_odds = np.log(model.alpha) - np.log(cache.num_included + 1)
return distributions.BernoulliDistribution.from_odds(data_odds + prior_odds)
def next_assignment_proposal(model, data, state, cache, Sigma_info, i, k):
assert not cache.rows_included[i]
x = state.X[i, :]
evidence = np.zeros(2)
for assignment in [0, 1]:
mu = Sigma_info.mu_for(k, assignment)
ssq = Sigma_info.sigma_sq_for(k, assignment) + state.sigma_sq_n
evidence[assignment] = ibp.gauss_loglik_vec_C2(x, mu, ssq)
data_odds = evidence[1] - evidence[0]
if cache.counts[k] > 0:
prior_odds = np.log(
cache.counts[k]) - np.log(cache.num_included - cache.counts[k] + 1)
else:
#prior_odds = poisson(1, 0.5 * model.alpha / (i+1)) - poisson(0, 0.5 * model.alpha / (i+1))
prior_odds = np.log(model.alpha) - np.log(cache.num_included + 1)
return distributions.BernoulliDistribution.from_odds(data_odds +
prior_odds)
def propose_assignments2(model, data, state, update=False):
N, K = state.Z.shape
state = state.copy()
cache = ibp.IBPCache.from_state(model, data, state, np.zeros(N,
dtype=bool))
proposal_prob = 0.
for i in range(N):
obs = data.mask[i, :]
x = state.X[i, :]
evidence = np.zeros(4)
prior_odds = np.zeros(4)
for c, (z1, z2) in enumerate(CHOICES):
z = np.array([z1, z2])
mu = cache.fpost.predictive_mu(z)
ssq = cache.fpost.predictive_ssq(z) + state.sigma_sq_n
evidence[c] = ibp.gauss_loglik_vec_C2(x[obs], mu[obs], ssq)
for k in [0, 1]:
if cache.counts[k] > 0:
prior_odds[c] += np.log(
cache.counts[k]) - np.log(cache.num_included -
cache.counts[k] + 1)
else:
prior_odds[c] += np.log(
model.alpha) - np.log(cache.num_included + 1)
odds = evidence + prior_odds
dist = distributions.MultinomialDistribution.from_odds(odds)
if update:
state.Z[i, :] = CHOICES[dist.sample().argmax()]
proposal_prob += dist.loglik(CHOICES.index(tuple(state.Z[i, :])))
cache.add(i, state.Z[i, :], state.X[i, :])
assert np.isfinite(proposal_prob)
return state, proposal_prob
