def test_update_convergence(nlp, labels, docs_golds):
nlp.pipeline[-1][1]
optim = nlp.resume_training()
texts, labels = zip(*docs_golds)
prev = 1e9
for i in range(10):
docs = [nlp.make_doc(text) for text in texts]
nlp.update(docs, labels, optim)
assert prev > get_loss_from_docs(docs).item()
def _update_params(self,
docs: Sequence[Doc],
optimizer: Optimizer,
verbose: bool = False):
loss = get_loss_from_docs(docs)
optimizer.zero_grad()
loss.backward()
optimizer.step()
if verbose:
logger.info(f"Loss: {loss.detach().item()}")
def evaluate( # type: ignore
self,
docs_golds: Sequence[Tuple[Union[str, Doc], Union[Dict[str, Any],
GoldParse]]],
batch_size: int = 16,
scorer: Optional[Scorer] = None,
) -> Dict[str, Any]:
"""Overrided to compute loss"""
if scorer is None:
scorer = Scorer(pipeline=self.pipeline)
loss = 0.0
for batch in minibatch(docs_golds, size=batch_size):
docs, golds = zip(*batch)
docs, golds = self._format_docs_and_golds(docs,
golds) # type: ignore
for _, pipe in self.pipeline:
self._eval_pipe(pipe, docs, golds, batch_size=batch_size)
loss += cast(float, get_loss_from_docs(docs).cpu().float().item())
for doc, gold in zip(docs, golds):
scorer.score(doc, gold)
scores = scorer.scores
scores["loss"] = loss
return scores
def test_update(nlp, labels, docs_golds):
optim = nlp.resume_training()
texts, labels = zip(*docs_golds)
docs = [nlp.make_doc(text) for text in texts]
nlp.update(docs, labels, optim)
assert get_loss_from_docs(docs)
def test_update(nlp):
docs = [nlp.make_doc(text) for text in TEXTS]
optim = nlp.resume_training()
nlp.update(docs, [{} for _ in range(len(docs))], optim)
loss = get_loss_from_docs(docs)
assert loss.item() > 0
