def build_tokenizer(self, tokenizer: TransformerSequenceTokenizer):
"""Build a transformer tokenizer for this task.
Args:
tokenizer: A tokenizer which is shared but can be adjusted to provide per-task settings.
Returns:
A TransformerSequenceTokenizer.
"""
if tokenizer.cls_is_bos != self.cls_is_bos or tokenizer.sep_is_eos != self.sep_is_eos:
tokenizer = copy(tokenizer)
tokenizer.cls_is_bos = self.cls_is_bos
tokenizer.sep_is_eos = self.sep_is_eos
return tokenizer
def build_dataloader(self,
data,
transform: TransformList = None,
training=False,
device=None,
logger: logging.Logger = None,
tokenizer: PreTrainedTokenizer = None,
**kwargs) -> DataLoader:
assert tokenizer
dataset = TextTokenizingDataset(data, cache=isinstance(data, str), delimiter=self.config.sent_delimiter,
generate_idx=isinstance(data, list),
max_seq_len=self.config.max_seq_len,
sent_delimiter=self.config.sent_delimiter,
transform=[
TransformerSequenceTokenizer(tokenizer,
'text',
ret_prefix_mask=True,
ret_subtokens=True,
),
FieldLength('text_input_ids', 'text_input_ids_length', delta=-2),
generate_token_span_tuple])
return PadSequenceDataLoader(
batch_sampler=self.sampler_builder.build(self.compute_lens(data, dataset, 'text_input_ids', 'text'),
shuffle=training),
device=device,
dataset=dataset)
def make_batch_for_bart(augmented_concept,
ret,
tokenizer,
device,
training=True):
token_field = 'concept'
tokenizer = TransformerSequenceTokenizer(tokenizer.tokenizer,
token_field,
cls_is_bos=True,
sep_is_eos=None)
encodings = [
tokenizer({token_field: x[:-1] if training else x})
for x in augmented_concept
]
ret.update(merge_list_of_dict(encodings))
decoder_mask = []
max_seq_len = len(max(ret['concept_input_ids'], key=len))
last_concept_offset = []
for spans, concepts in zip(ret['concept_token_span'], augmented_concept):
mask = ~SelfAttentionMask.get_mask(
max_seq_len, device, ret_parameter=False)
for group in spans:
for i in range(len(group)):
for j in range(i + 1, len(group)):
mask[group[i], group[j]] = True
decoder_mask.append(mask)
last_concept_offset.append(len(concepts) - 1)
ret['decoder_mask'] = torch.stack(decoder_mask)
if not training:
ret['last_concept_offset'] = torch.tensor(last_concept_offset,
device=device,
dtype=torch.long)
subtoken_to_tensor(token_field, ret)
def tokenizer_transform(self) -> TransformerSequenceTokenizer:
if not self._tokenizer_transform:
self._tokenizer_transform = TransformerSequenceTokenizer(
self.transformer_tokenizer,
self.config.token_key,
ret_token_span=True)
return self._tokenizer_transform
def tokenizer_transform(self):
if not self._tokenizer_transform:
self._tokenizer_transform = TransformerSequenceTokenizer(self.transformer_tokenizer,
self.config.token_key,
ret_subtokens=True,
ret_subtokens_group=True,
ret_token_span=False)
return self._tokenizer_transform
def build_tokenizer_transform(self):
return TransformerSequenceTokenizer(self.transformer_tokenizer,
'token',
'',
ret_token_span=True,
cls_is_bos=True,
max_seq_length=self.config.get(
'max_sequence_length', 512),
truncate_long_sequences=False)
def build_tokenizer(self, tokenizer: TransformerSequenceTokenizer):
# The transform for tokenizer needs very special settings, ensure these settings are set properly.
return TransformerSequenceTokenizer(
tokenizer.tokenizer,
tokenizer.input_key,
tokenizer.output_key,
tokenizer.max_seq_length,
tokenizer.truncate_long_sequences,
ret_subtokens=True,
ret_subtokens_group=True,
ret_token_span=True,
cls_is_bos=True,
sep_is_eos=True,
use_fast=tokenizer.tokenizer.is_fast,
dict_force=self.dict_force,
strip_cls_sep=False,
)
def __init__(self,
field: str,
transformer: str,
average_subwords=False,
scalar_mix: Union[ScalarMixWithDropoutBuilder, int] = None,
word_dropout: Optional[Union[float, Tuple[float,
str]]] = None,
max_sequence_length=None,
truncate_long_sequences=False,
cls_is_bos=False,
sep_is_eos=False,
ret_token_span=True,
ret_subtokens=False,
ret_subtokens_group=False,
ret_prefix_mask=False,
ret_raw_hidden_states=False,
transformer_args: Dict[str, Any] = None,
use_fast=True,
do_basic_tokenize=True,
trainable=True) -> None:
"""A contextual word embedding builder which builds a
:class:`~hanlp.layers.embeddings.contextual_word_embedding.ContextualWordEmbeddingModule` and a
:class:`~hanlp.transform.transformer_tokenizer.TransformerSequenceTokenizer`.
Args:
field: The field to work on. Usually some token fields.
transformer: An identifier of a ``PreTrainedModel``.
average_subwords: ``True`` to average subword representations.
scalar_mix: Layer attention.
word_dropout: Dropout rate of randomly replacing a subword with MASK.
max_sequence_length: The maximum sequence length. Sequence longer than this will be handled by sliding
window.
truncate_long_sequences: ``True`` to return hidden states of each layer.
cls_is_bos: ``True`` means the first token of input is treated as [CLS] no matter what its surface form is.
``False`` (default) means the first token is not [CLS], it will have its own embedding other than
the embedding of [CLS].
sep_is_eos: ``True`` means the last token of input is [SEP].
``False`` means it's not but [SEP] will be appended,
``None`` means it dependents on `input[-1] == [EOS]`.
ret_token_span: ``True`` to return span of each token measured by subtoken offsets.
ret_subtokens: ``True`` to return list of subtokens belonging to each token.
ret_subtokens_group: ``True`` to return list of offsets of subtokens belonging to each token.
ret_prefix_mask: ``True`` to generate a mask where each non-zero element corresponds to a prefix of a token.
ret_raw_hidden_states: ``True`` to return hidden states of each layer.
transformer_args: Extra arguments passed to the transformer.
use_fast: Whether or not to try to load the fast version of the tokenizer.
do_basic_tokenize: Whether to do basic tokenization before wordpiece.
trainable: ``False`` to use static embeddings.
"""
super().__init__()
self.truncate_long_sequences = truncate_long_sequences
self.transformer_args = transformer_args
self.trainable = trainable
self.ret_subtokens_group = ret_subtokens_group
self.ret_subtokens = ret_subtokens
self.ret_raw_hidden_states = ret_raw_hidden_states
self.sep_is_eos = sep_is_eos
self.cls_is_bos = cls_is_bos
self.max_sequence_length = max_sequence_length
self.word_dropout = word_dropout
self.scalar_mix = scalar_mix
self.average_subwords = average_subwords
self.transformer = transformer
self.field = field
self._transformer_tokenizer = AutoTokenizer_.from_pretrained(
self.transformer,
use_fast=use_fast,
do_basic_tokenize=do_basic_tokenize)
self._tokenizer_transform = TransformerSequenceTokenizer(
self._transformer_tokenizer,
field,
truncate_long_sequences=truncate_long_sequences,
ret_prefix_mask=ret_prefix_mask,
ret_token_span=ret_token_span,
cls_is_bos=cls_is_bos,
sep_is_eos=sep_is_eos,
ret_subtokens=ret_subtokens,
ret_subtokens_group=ret_subtokens_group,
max_seq_length=self.max_sequence_length)
