def set_annotations(self, docs: Iterable[Doc], outputs: Dict):
"""Set udify's output, which is calculated in self.predict, to docs"""
try:
from udify.models.udify_model import OUTPUTS as UdifyOUTPUTS # type: ignore
except ImportError:
Errors.E0(package="unofficial-udify")
for sent, output in zip(flatten_docs_to_sents(docs), outputs):
words = output[UdifyOUTPUTS.words]
_doc_tokens = [token.text for token in sent]
if words != _doc_tokens:
raise ValueError("Internal error has occured."
f"Input text: {sent.text}\n"
f"Input tokens: {_doc_tokens}\n"
f"Model words: {words}")
for token, dep, upos, lemma in zip(
sent,
output[UdifyOUTPUTS.predicted_dependencies],
output[UdifyOUTPUTS.upos],
output[UdifyOUTPUTS.lemmas],
):
token.dep_ = dep
token.lemma_ = lemma
token.pos_ = upos
sent = set_heads(sent, output[UdifyOUTPUTS.predicted_heads])
sent.doc.is_parsed = True
def test_flatten_docs_to_sens(vocab):
sentencizer = Sentencizer(".")
nlp = spacy.blank("en")
nlp.add_pipe(sentencizer)
texts = ["Foo is bar. Bar is baz.", "It is a sentence."]
docs = nlp.pipe(texts)
all_sents = flatten_docs_to_sents(docs)
assert len(all_sents) == 3
def predict(self, docs: Iterable[Doc]) -> Dict:
self.model.eval()
all_sents = flatten_docs_to_sents(docs)
with torch.no_grad():
tokens_list = [[t.text for t in sent] for sent in all_sents]
instances = [
self.dataset_reader.text_to_instance(tokens)
for tokens in tokens_list
]
outputs = self.model.forward_on_instances(instances)
return outputs