def log_prob(self, value):
if self._validate_args:
self._validate_sample(value)
normalize_term = self.total_count * logsumexp(self.logits, axis=-1) \
- gammaln(self.total_count + 1)
return np.sum(value * self.logits - gammaln(value + 1),
axis=-1) - normalize_term
def semi_supervised_hmm(transition_prior, emission_prior,
supervised_categories, supervised_words,
unsupervised_words):
num_categories, num_words = transition_prior.shape[
0], emission_prior.shape[0]
transition_prob = numpyro.sample(
'transition_prob',
dist.Dirichlet(
np.broadcast_to(transition_prior,
(num_categories, num_categories))))
emission_prob = numpyro.sample(
'emission_prob',
dist.Dirichlet(
np.broadcast_to(emission_prior, (num_categories, num_words))))
# models supervised data;
# here we don't make any assumption about the first supervised category, in other words,
# we place a flat/uniform prior on it.
numpyro.sample('supervised_categories',
dist.Categorical(
transition_prob[supervised_categories[:-1]]),
obs=supervised_categories[1:])
numpyro.sample('supervised_words',
dist.Categorical(emission_prob[supervised_categories]),
obs=supervised_words)
# computes log prob of unsupervised data
transition_log_prob = np.log(transition_prob)
emission_log_prob = np.log(emission_prob)
init_log_prob = emission_log_prob[:, unsupervised_words[0]]
log_prob = forward_log_prob(init_log_prob, unsupervised_words[1:],
transition_log_prob, emission_log_prob)
log_prob = logsumexp(log_prob, axis=0, keepdims=True)
# inject log_prob to potential function
numpyro.factor('forward_log_prob', log_prob)
def log_prob(self, value):
if self._validate_args:
self._validate_sample(value)
value = np.expand_dims(value, -1)
log_pmf = self.logits - logsumexp(self.logits, axis=-1, keepdims=True)
value, log_pmf = promote_shapes(value, log_pmf)
value = value[..., :1]
return np.take_along_axis(log_pmf, value, -1)[..., 0]
def predict(model, at_bats, hits, z, rng, player_names, train=True):
header = model.__name__ + (' - TRAIN' if train else ' - TEST')
model = substitute(seed(model, rng), z)
model_trace = trace(model).get_trace(at_bats)
predictions = model_trace['obs']['value']
print_results('=' * 30 + header + '=' * 30,
predictions,
player_names,
at_bats,
hits)
if not train:
model = substitute(model, z)
model_trace = trace(model).get_trace(at_bats, hits)
log_joint = 0.
for site in model_trace.values():
site_log_prob = site['fn'].log_prob(site['value'])
log_joint = log_joint + np.sum(site_log_prob.reshape(site_log_prob.shape[:1] + (-1,)),
-1)
log_post_density = logsumexp(log_joint) - np.log(np.shape(log_joint)[0])
print('\nPosterior log density: {:.2f}\n'.format(log_post_density))
def dual_moon_pe(x):
term1 = 0.5 * ((np.linalg.norm(x, axis=-1) - 2) / 0.4) ** 2
term2 = -0.5 * ((x[..., :1] + np.array([-2., 2.])) / 0.6) ** 2
return term1 - logsumexp(term2, axis=-1)
def forward_one_step(prev_log_prob, curr_word, transition_log_prob,
emission_log_prob):
log_prob_tmp = np.expand_dims(prev_log_prob, axis=1) + transition_log_prob
log_prob = log_prob_tmp + emission_log_prob[:, curr_word]
return logsumexp(log_prob, axis=0)
def log_prob(self, x):
term1 = 0.5 * ((np.linalg.norm(x, axis=-1) - 2) / 0.4) ** 2
term2 = -0.5 * ((x[..., :1] + np.array([-2., 2.])) / 0.6) ** 2
pe = term1 - logsumexp(term2, axis=-1)
return -pe