def __call__(self, engine: Engine) -> None:
output = self._output_transform(engine.state.output)
def raise_error(x: Union[float, torch.Tensor]) -> None:
if isinstance(x, numbers.Number):
x = torch.tensor(x)
if isinstance(x, torch.Tensor) and not bool(torch.isfinite(x).all()):
raise RuntimeError("Infinite or NaN tensor found.")
try:
apply_to_type(output, (numbers.Number, torch.Tensor), raise_error)
except RuntimeError:
self.logger.warning(f"{self.__class__.__name__}: Output '{output}' contains NaN or Inf. Stop training")
engine.terminate()
def _update(engine, batch):
model.train()
optimizer.zero_grad()
batch = apply_to_type(batch, torch.Tensor, lambda x: x.to(device))
scores = model(batch)
loss = -scores.mean()
if max_grad_norm is not None:
torch.nn.utils.clip_grad_norm_(model.parameters(), max_grad_norm)
loss.backward()
optimizer.step()
return loss.item()
def _inference(engine, batch):
model.eval()
with torch.no_grad():
batch = apply_to_type(batch, torch.Tensor, lambda x: x.to(device))
scores = model(batch, reduce=False)
return scores
model.load_state_dict(torch.load("my_model/nmt_mymodel_28.pth"))
model.to(model.device)
test_iterator = BasicIterator(batch_size=1)
test_iterator.index_with(vocab)
all_hyps = []
all_refs = []
for idx, test_data_sample in tqdm(
enumerate(test_iterator(instances, shuffle=False, num_epochs=1)),
total=len(instances),
):
# noinspection PyTypeChecker
with torch.no_grad():
hyps = model.beam_search(
apply_to_type(test_data_sample, torch.Tensor,
lambda t: t.to(model.device)),
TokenCharactersIndexer(
"char_trg",
character_tokenizer=MyCharacterTokenizer(max_length=21),
))
best_hyp = hyps[0].value
# try:
ref = [str(t) for t in instances[idx]["target_sentence"].tokens[1:-1]]
all_hyps.append(best_hyp)
all_refs.append([ref])
if idx % 100 == 0:
print(" ".join(best_hyp), "->", " ".join(ref))
# except:
# pass
# if idx == 100:
# break
