def score(self,
hypothesis: List[str],
references: List[List[str]],
tags: Optional[List[List[str]]] = None) -> VizSeqScore:
self._update_n_workers(len(hypothesis))
sent_scores = _get_sent_cider(hypothesis,
references,
extra_args={
'n_workers': self.n_workers,
'verbose': self.verbose
})
corpus_score, group_scores = None, None
if self.corpus_level:
corpus_score = np.mean(sent_scores)
if not self.sent_level:
sent_scores = None
if tags is not None:
tag_set = self._unique(tags)
group_scores = {}
for t in tag_set:
indices = [i for i, cur in enumerate(tags) if t in cur]
group_scores[t] = np.mean([sent_scores[i] for i in indices])
return VizSeqScore.make(corpus_score=corpus_score,
sent_scores=sent_scores,
group_scores=group_scores)
def score(
self, hypothesis: List[str], references: List[List[str]],
tags: Optional[List[List[str]]] = None
) -> VizSeqScore:
self._update_n_workers(len(hypothesis))
corpus_score, group_scores, sent_scores = None, None, None
if self.sent_level:
sent_scores = self._score_sentences_multiprocess(
hypothesis, references, _get_sent_bleu
)
if self.corpus_level:
corpus_score = self.score_corpus_multiprocess(
hypothesis, references
)
if tags is not None:
tag_set = self._unique(tags)
group_scores = {}
for t in tag_set:
indices = [i for i, cur in enumerate(tags) if t in cur]
ref_slice = [[r[i] for i in indices] for r in references]
pred_slice = [hypothesis[i] for i in indices]
group_scores[t] = self.score_corpus_multiprocess(
pred_slice, ref_slice
)
return VizSeqScore.make(
corpus_score=corpus_score, sent_scores=sent_scores,
group_scores=group_scores
)
def score(
self, hypothesis: List[str], references: List[List[str]],
tags: Optional[List[List[str]]] = None
) -> VizSeqScore:
corpus_score, sent_scores, group_scores = None, None, None
import bert_score as bs
import langid
import logging
logging.getLogger('pytorch_pretrained_bert').setLevel(logging.WARNING)
logging.getLogger('langid').setLevel(logging.WARNING)
lang = langid.classify(references[0][0])[0]
sent_scores = bs.score(
hypothesis, references[0], nthreads=self.n_workers, lang=lang,
verbose=self.verbose
)[2].tolist()
if self.corpus_level:
corpus_score = np.mean(sent_scores)
if tags is not None:
tag_set = self._unique(tags)
group_scores = {}
for t in tag_set:
indices = [i for i, cur in enumerate(tags) if t in cur]
group_scores[t] = np.mean([sent_scores[i] for i in indices])
return VizSeqScore.make(
corpus_score=corpus_score, sent_scores=sent_scores,
group_scores=group_scores
)
def score(self,
hypothesis: List[str],
references: List[List[str]],
tags: Optional[List[List[str]]] = None) -> VizSeqScore:
self._update_n_workers(len(hypothesis))
corpus_score, group_scores, sent_scores = None, None, None
sent_scores = self._score_sentences_multiprocess(
hypothesis, references, _get_sent_wer)
sent_lens = None
if self.corpus_level:
sent_lens = self._score_sentences_multiprocess(
hypothesis, references, _get_sent_len_r)
n_incorrect = np.sum(
[s * l for s, l in zip(sent_scores, sent_lens)])
corpus_score = n_incorrect / np.sum(sent_lens)
if tags is not None:
tag_set = self._unique(tags)
group_scores = {}
if sent_lens is None:
sent_lens = self._score_sentences_multiprocess(
hypothesis, references, _get_sent_len_r)
for t in tag_set:
indices = [i for i, cur in enumerate(tags) if t in cur]
cur_sent_scores = [sent_scores[i] for i in indices]
cur_sent_lens = [sent_lens[i] for i in indices]
n_incorrect = np.sum(
[s * l for s, l in zip(cur_sent_scores, cur_sent_lens)])
group_scores[t] = n_incorrect / np.sum(sent_lens)
return VizSeqScore.make(corpus_score=corpus_score,
sent_scores=sent_scores,
group_scores=group_scores)
def score(
self, hypothesis: List[str], references: Optional[List[List[str]]] = None,
tags: Optional[List[List[str]]] = None
) -> VizSeqScore:
corpus_score, group_scores, sent_scores = None, None, None
selfbleu_scores = compute_self_bleu(hypothesis)
if self.corpus_level:
# implement corpus-level score
corpus_score = np.mean(selfbleu_scores)
if self.sent_level:
# implement sentence-level score
sent_scores = selfbleu_scores
if tags is not None:
raise NotImplementedError
# tag_set = self._unique(tags)
# implement group-level (by sentence tags) score
# group_scores={t: 99.9 for t in tag_set}
return VizSeqScore.make(
corpus_score=corpus_score, sent_scores=sent_scores,
group_scores=group_scores
)
def score(self,
hypothesis: List[str],
references: List[List[str]],
tags: Optional[List[List[str]]] = None) -> VizSeqScore:
corpus_score, group_scores, sent_scores = None, None, None
sent_scores = _get_sent_laser(hypothesis, references)
if self.corpus_level:
corpus_score = np.mean(sent_scores)
if tags is not None:
tag_set = self._unique(tags)
group_scores = {}
for t in tag_set:
indices = [i for i, cur in enumerate(tags) if t in cur]
group_scores[t] = np.mean([sent_scores[i] for i in indices])
return VizSeqScore.make(corpus_score=corpus_score,
sent_scores=sent_scores,
group_scores=group_scores)
def score(
self,
hypothesis: List[str],
references: List[List[str]],
tags: Optional[List[List[str]]] = None,
sources: Optional[List[List[str]]] = None,
) -> VizSeqScore:
problem = self.extra_args["problem"]
# Only relevant if predicting tasks
assert problem in (
"TargetProductAndRequirements_TO_Tasks",
"Requirements_TO_TargetProductAndTasks",
"TargetProductAndRequirementsAndTasks",
"RequirementsAndTargetProductAndTasks",
)
corpus_score, sent_scores, group_scores = None, None, None
requirements = sources[0]
hypotheses = hypothesis
sent_scores = []
for req_str, hypo in zip(requirements, hypotheses):
try:
score = compute_requirement_coverage(hypo,
req_str,
essential=True,
problem=problem)
except ValueError:
score = 0
sent_scores.append(score)
if self.corpus_level:
corpus_score = np.mean(sent_scores) * 100
return VizSeqScore.make(corpus_score=corpus_score,
sent_scores=sent_scores,
group_scores={})
def score(
self,
hypothesis: List[str],
references: List[List[str]],
tags: Optional[List[List[str]]] = None,
sources: Optional[List[List[str]]] = None,
) -> VizSeqScore:
# global bert_scorer
problem = self.extra_args["problem"]
# Only relevant if predicting tasks
assert problem in (
"TargetProductAndRequirements_TO_Tasks",
"Requirements_TO_TargetProductAndTasks",
"TargetProductAndRequirementsAndTasks",
"RequirementsAndTargetProductAndTasks",
)
corpus_score, sent_scores, group_scores = None, None, None
# requirements = sources[0]
references = references[0]
hypotheses = hypothesis
sent_scores = []
for ref, hypo in zip(references, hypotheses):
# if problem == 'TargetProductAndRequirementsAndTasks':
# # HACK: RecipeGPT grammar
# tasks_gt = re.split('\. |! ', ref.rstrip(' <end-directions>'))
# tasks_pred = re.split('\. |! ',
# hypo.rstrip(' <end-directions>'))
if problem in (
"Requirements_TO_TargetProductAndTasks",
"RequirementsAndTargetProductAndTasks",
):
# Requirements_TO_TargetProductAndTasks
tgt_prod_and_tasks_gt = string_to_tasks(ref, parse_tp=True)
tasks_gt = tgt_prod_and_tasks_gt[1:]
try:
tgt_prod_and_tasks_pred = string_to_tasks(hypo,
parse_tp=True)
tasks_pred = tgt_prod_and_tasks_pred[1:]
except ValueError:
continue
else:
# TargetProductAndRequirements_TO_Tasks or TargetProductAndRequirementsAndTasks
tasks_gt = string_to_tasks(ref)
tasks_pred = string_to_tasks(hypo)
try:
score = compute_task_order_score(tasks_gt, tasks_pred)
sent_scores.append(score)
except ScoreComputationError:
continue
if self.corpus_level:
print(len(sent_scores))
corpus_score = np.mean(sent_scores)
return VizSeqScore.make(corpus_score=corpus_score,
sent_scores=sent_scores,
group_scores={})