def continue_constrained(self, hyp):
assert hyp.unfinished_constraint is True, 'hyp must be part of an unfinished constraint'
# this should come from the model
score = np.random.random()
constraint_row_index = hyp.constraint_index[0]
# the index of the next token in the constraint
constraint_tok_index = hyp.constraint_index[1] + 1
constraint_index = (constraint_row_index, constraint_tok_index)
continued_constraint_token = hyp.constraints[constraint_index[0]][
constraint_index[1]]
coverage = copy.deepcopy(hyp.coverage)
coverage[constraint_row_index][constraint_tok_index] = 1
if len(hyp.constraints[constraint_row_index]
) > constraint_tok_index + 1:
unfinished_constraint = True
else:
unfinished_constraint = False
new_hyp = ConstraintHypothesis(
token=continued_constraint_token,
score=score,
coverage=coverage,
constraints=hyp.constraints,
payload=None,
backpointer=hyp,
constraint_index=constraint_index,
unfinished_constraint=unfinished_constraint)
return new_hyp
def generate_constrained(self, hyp):
"""Look at the coverage of the hyp to get constraint candidates"""
assert hyp.unfinished_constraint is not True, 'hyp must not be part of an unfinished constraint'
new_constraint_hyps = []
available_constraints = hyp.constraint_candidates()
for idx in available_constraints:
# starting a new constraint
constraint_token = hyp.constraints[idx][0]
# this should come from the model
score = np.random.random()
coverage = copy.deepcopy(hyp.coverage)
coverage[idx][0] = 1
if len(coverage[idx]) > 1:
unfinished_constraint = True
else:
unfinished_constraint = False
new_hyp = ConstraintHypothesis(
token=constraint_token,
score=score,
coverage=coverage,
constraints=hyp.constraints,
payload=None,
backpointer=hyp,
constraint_index=(idx, 0),
unfinished_constraint=unfinished_constraint)
new_constraint_hyps.append(new_hyp)
return new_constraint_hyps
def test_chaining_hypotheses(self):
next_symbol = 1
next_hyp = ConstraintHypothesis(token=next_symbol,
score=1.0,
coverage=self.start_hyp.coverage,
constraints=self.start_hyp.constraints,
constraint_index=None,
payload=None,
backpointer=self.start_hyp,
unfinished_constraint=False)
# test that hypotheses correctly output their sequences
token_seq = next_hyp.sequence
true_seq = [0, 1]
self.assertTrue(len(token_seq) == 2)
self.assertTrue(
all(true_sym == seq_sym
for true_sym, seq_sym in zip(true_seq, token_seq)))
constraint_idx_seq = next_hyp.constraint_index_sequence
true_seq = [None, None]
self.assertTrue(len(constraint_idx_seq) == 2)
self.assertTrue(
all(true_sym == seq_sym
for true_sym, seq_sym in zip(true_seq, constraint_idx_seq)))
def setUp(self):
self.vocabulary = [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]
sample_constraints = [[1, 2], [5, 6, 7]]
# create a start hypothesis
start_token = u'<S>'
p_start = 1.0
# dummy start hypothesis
self.start_hyp = ConstraintHypothesis(
token=start_token,
score=p_start,
coverage=init_coverage(sample_constraints),
constraints=sample_constraints,
payload=None,
backpointer=None,
unfinished_constraint=False)
def setUp(self):
start_symbol = 0
sample_constraints = [[1, 2], [5, 6, 7]]
coverage = [[0, 0], [0, 0, 0]]
p_start = 1.0
# dummy start hypothesis
self.start_hyp = ConstraintHypothesis(token=start_symbol,
score=p_start,
coverage=coverage,
constraints=sample_constraints,
constraint_index=None,
payload=None,
backpointer=None,
unfinished_constraint=False)
def generate(self, hyp, n_best=1):
# make k_best random hyp objects
next_tokens = np.random.choice(self.vocabulary, size=n_best)
next_scores = np.random.random(size=n_best)
new_hyps = []
for i in range(n_best):
new_hyp = ConstraintHypothesis(token=next_tokens[i],
score=next_scores[i],
coverage=copy.deepcopy(
hyp.coverage),
constraints=hyp.constraints,
payload=None,
backpointer=hyp,
constraint_index=None,
unfinished_constraint=False)
new_hyps.append(new_hyp)
return new_hyps