def create_and_check_for_token_classification(
self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
):
config.num_labels = self.num_labels
model = AlbertForTokenClassification(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=token_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.num_labels))
def create_and_check_albert_for_token_classification(
self, config, input_ids, token_type_ids, input_mask,
sequence_labels, token_labels, choice_labels):
config.num_labels = self.num_labels
model = AlbertForTokenClassification(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids,
attention_mask=input_mask,
token_type_ids=token_type_ids,
labels=token_labels)
self.parent.assertListEqual(
list(result["logits"].size()),
[self.batch_size, self.seq_length, self.num_labels])
self.check_loss_output(result)
class AlbertNER:
""" Class to use Albert to extract named entities.
TODO: Update model and checkpoints to work with last versions of transformers """
def __init__(self, path: str ,device: str = 'cpu'):
""" Init the NER Albert """
if not os.path.exists(path):
raise NotADirectoryError(
f"{os.path.abspath(path)} must be a directory containing the model files: config, tokenizer, weights.")
files = os.listdir(path)
if CONFIG_JSON_FILE not in files:
raise FileNotFoundError(f"{CONFIG_JSON_FILE} must be in {path}.")
if WEIGHTS_FILE not in files:
raise FileNotFoundError(f"{WEIGHTS_FILE} must be in {path}.")
with open(os.path.join(path, CONFIG_JSON_FILE), "r") as f:
config = json.load(f)
self.tokenizer = AutoTokenizer.from_pretrained(path)
weights = torch.load(os.path.join(path, WEIGHTS_FILE),
map_location=lambda storage, loc: storage)
# Load pretrained model/tokenizer
config = AlbertConfig.from_dict(config)
self.model = AlbertForTokenClassification(config)
self.model.load_state_dict(weights)
self.model = self.model.eval()
self.args = albert_args_ner
if device == "cuda":
logger.debug("Setting model with CUDA")
self.args['device'] = 'cuda'
self.model.to('cuda')
def extract(self, text: str, **kwargs: dict) -> List[Tuple[str, str]]:
""" Extract named entities from text
Keyword Arguments:
:param text: Text to extract entities from
:return: List of named entities extacted
"""
for key in kwargs:
if key in self.args:
self.args[key] = kwargs[key]
tokens = self.tokenizer.tokenize(self.tokenizer.decode(self.tokenizer.encode(text)))
inputs = self.tokenizer.encode(text, return_tensors="pt")
outputs = self.model(inputs, **kwargs)[0]
predictions = torch.argmax(outputs, dim=2)
return [(token, label_list[prediction]) for token, prediction in zip(tokens, predictions[0].tolist())]