class PretrainedTransformerMismatchedIndexer(TokenIndexer):
"""
Use this indexer when (for whatever reason) you are not using a corresponding
`PretrainedTransformerTokenizer` on your input. We assume that you used a tokenizer that splits
strings into words, while the transformer expects wordpieces as input. This indexer splits the
words into wordpieces and flattens them out. You should use the corresponding
`PretrainedTransformerMismatchedEmbedder` to embed these wordpieces and then pull out a single
vector for each original word.
# Parameters
model_name : `str`
The name of the `transformers` model to use.
namespace : `str`, optional (default=`tags`)
We will add the tokens in the pytorch_transformer vocabulary to this vocabulary namespace.
We use a somewhat confusing default value of `tags` so that we do not add padding or UNK
tokens to this namespace, which would break on loading because we wouldn't find our default
OOV token.
"""
def __init__(self,
model_name: str,
namespace: str = "tags",
**kwargs) -> None:
super().__init__(**kwargs)
# The matched version v.s. mismatched
self._matched_indexer = PretrainedTransformerIndexer(
model_name, namespace, **kwargs)
# add_special_tokens=False since we don't want wordpieces to be surrounded by special tokens
self._allennlp_tokenizer = PretrainedTransformerTokenizer(
model_name, add_special_tokens=False)
self._tokenizer = self._allennlp_tokenizer.tokenizer
(
self._num_added_start_tokens,
self._num_added_end_tokens,
) = self._determine_num_special_tokens_added()
@overrides
def count_vocab_items(self, token: Token, counter: Dict[str, Dict[str,
int]]):
return self._matched_indexer.count_vocab_items(token, counter)
@overrides
def tokens_to_indices(self, tokens: List[Token],
vocabulary: Vocabulary) -> IndexedTokenList:
orig_token_mask = [1] * len(tokens)
tokens, offsets = self._intra_word_tokenize(tokens)
# {"token_ids": ..., "mask": ...}
output = self._matched_indexer.tokens_to_indices(tokens, vocabulary)
# Insert type ids for the special tokens.
output[
"type_ids"] = self._tokenizer.create_token_type_ids_from_sequences(
output["token_ids"])
# Insert the special tokens themselves.
output["token_ids"] = self._tokenizer.build_inputs_with_special_tokens(
output["token_ids"])
output["mask"] = orig_token_mask
output["offsets"] = [(start + self._num_added_start_tokens,
end + self._num_added_start_tokens)
for start, end in offsets]
output["wordpiece_mask"] = [1] * len(output["token_ids"])
return output
@overrides
def get_empty_token_list(self) -> IndexedTokenList:
output = self._matched_indexer.get_empty_token_list()
output["offsets"] = []
output["wordpiece_mask"] = []
return output
@overrides
def as_padded_tensor_dict(
self, tokens: IndexedTokenList,
padding_lengths: Dict[str, int]) -> Dict[str, torch.Tensor]:
tokens = tokens.copy()
padding_lengths = padding_lengths.copy()
offsets_tokens = tokens.pop("offsets")
offsets_padding_lengths = padding_lengths.pop("offsets")
tensor_dict = self._matched_indexer.as_padded_tensor_dict(
tokens, padding_lengths)
tensor_dict["offsets"] = torch.LongTensor(
pad_sequence_to_length(offsets_tokens,
offsets_padding_lengths,
default_value=lambda: (0, 0)))
return tensor_dict
def __eq__(self, other):
if isinstance(other, PretrainedTransformerMismatchedIndexer):
for key in self.__dict__:
if key == "tokenizer":
# This is a reference to a function in the huggingface code, which we can't
# really modify to make this clean. So we special-case it.
continue
if self.__dict__[key] != other.__dict__[key]:
return False
return True
return NotImplemented
def _intra_word_tokenize(
self,
tokens: List[Token]) -> Tuple[List[Token], List[Tuple[int, int]]]:
"""
Tokenizes each word into wordpieces separately. Also calculates offsets such that
wordpices[offsets[i][0]:offsets[i][1] + 1] corresponds to the original i-th token.
Does not insert special tokens.
"""
wordpieces: List[Token] = []
offsets = []
cumulative = 0
for token in tokens:
subword_wordpieces = self._allennlp_tokenizer.tokenize(token.text)
wordpieces.extend(subword_wordpieces)
start_offset = cumulative
cumulative += len(subword_wordpieces)
end_offset = cumulative - 1 # inclusive
offsets.append((start_offset, end_offset))
return wordpieces, offsets
def _determine_num_special_tokens_added(self) -> Tuple[int, int]:
"""
Determines the number of tokens self._tokenizer adds to a sequence (currently doesn't
consider sequence pairs) in the start & end.
# Returns
The number of tokens (`int`) that are inserted in the start & end of a sequence.
"""
# Uses a slightly higher index to avoid tokenizer doing special things to lower-indexed
# tokens which might be special.
dummy = [1000]
inserted = self._tokenizer.build_inputs_with_special_tokens(dummy)
num_start = num_end = 0
seen_dummy = False
for idx in inserted:
if idx == dummy[0]:
if seen_dummy: # seeing it twice
raise ValueError(
"Cannot auto-determine the number of special tokens added."
)
seen_dummy = True
continue
if not seen_dummy:
num_start += 1
else:
num_end += 1
assert num_start + num_end == self._tokenizer.num_added_tokens()
return num_start, num_end
class PretrainedTransformerMismatchedIndexer(TokenIndexer):
"""
Use this indexer when (for whatever reason) you are not using a corresponding
`PretrainedTransformerTokenizer` on your input. We assume that you used a tokenizer that splits
strings into words, while the transformer expects wordpieces as input. This indexer splits the
words into wordpieces and flattens them out. You should use the corresponding
`PretrainedTransformerMismatchedEmbedder` to embed these wordpieces and then pull out a single
vector for each original word.
Registered as a `TokenIndexer` with name "pretrained_transformer_mismatched".
# Parameters
model_name : `str`
The name of the `transformers` model to use.
namespace : `str`, optional (default=`tags`)
We will add the tokens in the pytorch_transformer vocabulary to this vocabulary namespace.
We use a somewhat confusing default value of `tags` so that we do not add padding or UNK
tokens to this namespace, which would break on loading because we wouldn't find our default
OOV token.
max_length : `int`, optional (default = `None`)
If positive, split the document into segments of this many tokens (including special tokens)
before feeding into the embedder. The embedder embeds these segments independently and
concatenate the results to get the original document representation. Should be set to
the same value as the `max_length` option on the `PretrainedTransformerMismatchedEmbedder`.
tokenizer_kwargs : `Dict[str, Any]`, optional (default = `None`)
Dictionary with
[additional arguments](https://github.com/huggingface/transformers/blob/155c782a2ccd103cf63ad48a2becd7c76a7d2115/transformers/tokenization_utils.py#L691)
for `AutoTokenizer.from_pretrained`.
""" # noqa: E501
def __init__(
self,
model_name: str,
namespace: str = "tags",
max_length: int = None,
tokenizer_kwargs: Optional[Dict[str, Any]] = None,
**kwargs,
) -> None:
super().__init__(**kwargs)
# The matched version v.s. mismatched
self._matched_indexer = PretrainedTransformerIndexer(
model_name,
namespace=namespace,
max_length=max_length,
tokenizer_kwargs=tokenizer_kwargs,
**kwargs,
)
self._allennlp_tokenizer = self._matched_indexer._allennlp_tokenizer
self._tokenizer = self._matched_indexer._tokenizer
self._num_added_start_tokens = self._matched_indexer._num_added_start_tokens
self._num_added_end_tokens = self._matched_indexer._num_added_end_tokens
@overrides
def count_vocab_items(self, token: Token, counter: Dict[str, Dict[str,
int]]):
return self._matched_indexer.count_vocab_items(token, counter)
@overrides
def tokens_to_indices(self, tokens: List[Token],
vocabulary: Vocabulary) -> IndexedTokenList:
self._matched_indexer._add_encoding_to_vocabulary_if_needed(vocabulary)
wordpieces, offsets = self._allennlp_tokenizer.intra_word_tokenize(
[t.text for t in tokens])
# For tokens that don't correspond to any word pieces, we put (-1, -1) into the offsets.
# That results in the embedding for the token to be all zeros.
offsets = [x if x is not None else (-1, -1) for x in offsets]
output: IndexedTokenList = {
"token_ids": [t.text_id for t in wordpieces],
"mask":
[True] * len(tokens), # for original tokens (i.e. word-level)
"type_ids": [t.type_id for t in wordpieces],
"offsets": offsets,
"wordpiece_mask":
[True] * len(wordpieces), # for wordpieces (i.e. subword-level)
}
return self._matched_indexer._postprocess_output(output)
@overrides
def get_empty_token_list(self) -> IndexedTokenList:
output = self._matched_indexer.get_empty_token_list()
output["offsets"] = []
output["wordpiece_mask"] = []
return output
@overrides
def as_padded_tensor_dict(
self, tokens: IndexedTokenList,
padding_lengths: Dict[str, int]) -> Dict[str, torch.Tensor]:
tokens = tokens.copy()
padding_lengths = padding_lengths.copy()
offsets_tokens = tokens.pop("offsets")
offsets_padding_lengths = padding_lengths.pop("offsets")
tensor_dict = self._matched_indexer.as_padded_tensor_dict(
tokens, padding_lengths)
tensor_dict["offsets"] = torch.LongTensor(
pad_sequence_to_length(offsets_tokens,
offsets_padding_lengths,
default_value=lambda: (0, 0)))
return tensor_dict
def __eq__(self, other):
if isinstance(other, PretrainedTransformerMismatchedIndexer):
for key in self.__dict__:
if key == "_tokenizer":
# This is a reference to a function in the huggingface code, which we can't
# really modify to make this clean. So we special-case it.
continue
if self.__dict__[key] != other.__dict__[key]:
return False
return True
return NotImplemented
class PretrainedTransformerMismatchedIndexer(TokenIndexer):
"""
Use this indexer when (for whatever reason) you are not using a corresponding
`PretrainedTransformerTokenizer` on your input. We assume that you used a tokenizer that splits
strings into words, while the transformer expects wordpieces as input. This indexer splits the
words into wordpieces and flattens them out. You should use the corresponding
`PretrainedTransformerMismatchedEmbedder` to embed these wordpieces and then pull out a single
vector for each original word.
# Parameters
model_name : `str`
The name of the `transformers` model to use.
namespace : `str`, optional (default=`tags`)
We will add the tokens in the pytorch_transformer vocabulary to this vocabulary namespace.
We use a somewhat confusing default value of `tags` so that we do not add padding or UNK
tokens to this namespace, which would break on loading because we wouldn't find our default
OOV token.
max_length : `int`, optional (default = None)
If positive, split the document into segments of this many tokens (including special tokens)
before feeding into the embedder. The embedder embeds these segments independently and
concatenate the results to get the original document representation. Should be set to
the same value as the `max_length` option on the `PretrainedTransformerMismatchedEmbedder`.
"""
def __init__(self,
model_name: str,
namespace: str = "tags",
max_length: int = None,
**kwargs) -> None:
super().__init__(**kwargs)
# The matched version v.s. mismatched
self._matched_indexer = PretrainedTransformerIndexer(
model_name, namespace, max_length, **kwargs)
self._tokenizer = self._matched_indexer._tokenizer
self._num_added_start_tokens = self._matched_indexer._num_added_start_tokens
self._num_added_end_tokens = self._matched_indexer._num_added_end_tokens
@overrides
def count_vocab_items(self, token: Token, counter: Dict[str, Dict[str,
int]]):
return self._matched_indexer.count_vocab_items(token, counter)
@overrides
def tokens_to_indices(self, tokens: List[Token],
vocabulary: Vocabulary) -> IndexedTokenList:
self._matched_indexer._add_encoding_to_vocabulary_if_needed(vocabulary)
indices, offsets = self._intra_word_tokenize(tokens)
# `create_token_type_ids_from_sequences()` inserts special tokens
type_ids = self._tokenizer.create_token_type_ids_from_sequences(
indices[self._num_added_start_tokens:-self._num_added_end_tokens])
output: IndexedTokenList = {
"token_ids": indices,
"mask": [1] * len(tokens), # for original tokens (i.e. word-level)
"type_ids": type_ids,
"offsets": offsets,
"wordpiece_mask":
[1] * len(indices), # for wordpieces (i.e. subword-level)
}
return self._matched_indexer._postprocess_output(output)
@overrides
def get_empty_token_list(self) -> IndexedTokenList:
output = self._matched_indexer.get_empty_token_list()
output["offsets"] = []
output["wordpiece_mask"] = []
return output
@overrides
def as_padded_tensor_dict(
self, tokens: IndexedTokenList,
padding_lengths: Dict[str, int]) -> Dict[str, torch.Tensor]:
tokens = tokens.copy()
padding_lengths = padding_lengths.copy()
offsets_tokens = tokens.pop("offsets")
offsets_padding_lengths = padding_lengths.pop("offsets")
tensor_dict = self._matched_indexer.as_padded_tensor_dict(
tokens, padding_lengths)
tensor_dict["offsets"] = torch.LongTensor(
pad_sequence_to_length(offsets_tokens,
offsets_padding_lengths,
default_value=lambda: (0, 0)))
return tensor_dict
def __eq__(self, other):
if isinstance(other, PretrainedTransformerMismatchedIndexer):
for key in self.__dict__:
if key == "tokenizer":
# This is a reference to a function in the huggingface code, which we can't
# really modify to make this clean. So we special-case it.
continue
if self.__dict__[key] != other.__dict__[key]:
return False
return True
return NotImplemented
def _intra_word_tokenize(
self,
tokens: List[Token]) -> Tuple[List[int], List[Tuple[int, int]]]:
"""
Tokenizes each word into wordpieces separately and returns the wordpiece IDs.
Also calculates offsets such that wordpices[offsets[i][0]:offsets[i][1] + 1]
corresponds to the original i-th token.
This function inserts special tokens.
"""
wordpieces: List[int] = []
offsets = []
cumulative = self._num_added_start_tokens
for token in tokens:
subword_wordpieces = self._tokenizer.encode(
token.text, add_special_tokens=False)
wordpieces.extend(subword_wordpieces)
start_offset = cumulative
cumulative += len(subword_wordpieces)
end_offset = cumulative - 1 # inclusive
offsets.append((start_offset, end_offset))
wordpieces = self._tokenizer.build_inputs_with_special_tokens(
wordpieces)
assert cumulative + self._num_added_end_tokens == len(wordpieces)
return wordpieces, offsets
