def _combine_posteriors_simple(self,
non_zero_words,
posterior,
unk_prob,
top_n=0,
original_posterior=None):
"""
Args:
non_zero_words (set): All words with positive probability
posteriors: Predictor posterior distributions calculated
with ``predict_next()``
unk_probs: UNK probabilities of the predictors, calculated
with ``get_unk_probability``
Returns:
combined,score_breakdown: like in ``apply_predictors()``
"""
if top_n > 0:
non_zero_words = utils.argmax_n(posterior, top_n)
scores_func = np.vectorize(
lambda x: utils.common_get(posterior, x, unk_prob))
scores = scores_func(non_zero_words)
orig_scores = None
if original_posterior is not None:
scores_func = np.vectorize(
lambda x: utils.common_get(original_posterior, x, unk_prob))
orig_scores = scores_func(non_zero_words)
return non_zero_words, scores, orig_scores
def _get_next_hypos(self, all_hypos, size, other_groups=None):
"""Get hypos for the next iteration. """
all_scores = np.array([self.get_adjusted_score(hypo) for hypo in all_hypos])
if other_groups:
all_scores = all_scores + self.lmbda*self.hamming_distance_penalty(all_hypos,
utils.flattened(other_groups))
inds = utils.argmax_n(all_scores, size)
return [all_hypos[ind] for ind in inds]
def _scale_combine_non_zero_scores(non_zero_word_count,
posteriors,
unk_probs,
pred_weights,
top_n=0):
scaled_posteriors = []
for posterior, unk_prob, weight in zip(
posteriors, unk_probs, pred_weights):
if isinstance(posterior, dict):
arr = np.full(non_zero_word_count, unk_prob)
for word, score in posterior.items():
if word < non_zero_word_count:
arr[word] = score
scaled_posteriors.append(arr * weight)
else:
n_unks = non_zero_word_count - len(posterior)
if n_unks > 0:
posterior = np.concatenate((
posterior, np.full(n_unks, unk_prob)))
elif n_unks < 0:
posterior = posterior[:n_unks]
scaled_posteriors.append(posterior * weight)
combined_scores = np.sum(scaled_posteriors, axis=0)
return utils.argmax_n(combined_scores, top_n)
def _get_next_hypos(self, all_hypos, all_scores):
"""Get hypos for the next iteration. """
inds = utils.argmax_n(all_scores, self.beam_size)
return [all_hypos[ind] for ind in inds]
def _best_eos(self, hypos):
"""Returns true if the best hypothesis ends with </S>"""
ln_scores = [self.get_adjusted_score(hypo) for hypo in hypos]
best_inds = utils.argmax_n(ln_scores, self.nbest)
return all(
[hypos[ind].get_last_word() == utils.EOS_ID for ind in best_inds])
