SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

Implement Fast Tokenization for Deberta (#11387)

Browse Source
This commit is contained in:
Shubham Sanghavi
2021-04-30 07:08:15 -05:00
committed by GitHub
parent db9dd09cf9
commit 30ede8994e
9 changed files with 271 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
docs/source/index.rst
View File

@@ -284,7 +284,7 @@ Flax), PyTorch, and/or TensorFlow.
+-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
| DPR | ✅ | ✅ | ✅ | ✅ | ❌ |
+-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
| DeBERTa | ✅ | ❌ | ✅ | ❌ | ❌ |
| DeBERTa | ✅ | ✅ | ✅ | ❌ | ❌ |
+-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
| DeBERTa-v2 | ✅ | ❌ | ✅ | ❌ | ❌ |
+-----------------------------+----------------+----------------+-----------------+--------------------+--------------+

6
docs/source/model_doc/deberta.rst
View File

@@ -56,6 +56,12 @@ DebertaTokenizer
:members: build_inputs_with_special_tokens, get_special_tokens_mask,
create_token_type_ids_from_sequences, save_vocabulary
DebertaTokenizerFast
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
.. autoclass:: transformers.DebertaTokenizerFast
:members: build_inputs_with_special_tokens, create_token_type_ids_from_sequences
DebertaModel
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

2
src/transformers/__init__.py
View File

@@ -315,6 +315,7 @@ if is_tokenizers_available():
_import_structure["models.barthez"].append("BarthezTokenizerFast")
_import_structure["models.bert"].append("BertTokenizerFast")
_import_structure["models.camembert"].append("CamembertTokenizerFast")
_import_structure["models.deberta"].append("DebertaTokenizerFast")
_import_structure["models.distilbert"].append("DistilBertTokenizerFast")
_import_structure["models.dpr"].extend(
["DPRContextEncoderTokenizerFast", "DPRQuestionEncoderTokenizerFast", "DPRReaderTokenizerFast"]
@@ -1661,6 +1662,7 @@ if TYPE_CHECKING:
from .models.bert import BertTokenizerFast
from .models.camembert import CamembertTokenizerFast
from .models.convbert import ConvBertTokenizerFast
from .models.deberta import DebertaTokenizerFast
from .models.distilbert import DistilBertTokenizerFast
from .models.dpr import DPRContextEncoderTokenizerFast, DPRQuestionEncoderTokenizerFast, DPRReaderTokenizerFast
from .models.electra import ElectraTokenizerFast

32
src/transformers/convert_slow_tokenizer.py
View File

@@ -296,6 +296,37 @@ class RobertaConverter(Converter):
return tokenizer
class DebertaConverter(Converter):
def converted(self) -> Tokenizer:
ot = self.original_tokenizer
vocab = ot.encoder
merges = list(ot.bpe_ranks.keys())
tokenizer = Tokenizer(
BPE(
vocab=vocab,
merges=merges,
dropout=None,
continuing_subword_prefix="",
end_of_word_suffix="",
fuse_unk=False,
)
)
tokenizer.pre_tokenizer = pre_tokenizers.ByteLevel(add_prefix_space=ot.add_prefix_space)
tokenizer.decoder = decoders.ByteLevel()
tokenizer.post_processor = processors.TemplateProcessing(
single="[CLS]:0 $A:0 [SEP]:0",
pair="[CLS]:0 $A:0 [SEP]:0 $B:0 [SEP]:0",
special_tokens=[
("[CLS]", self.original_tokenizer.convert_tokens_to_ids("[CLS]")),
("[SEP]", self.original_tokenizer.convert_tokens_to_ids("[SEP]")),
],
)
return tokenizer
class SpmConverter(Converter):
def __init__(self, *args):
requires_backends(self, "protobuf")
@@ -654,6 +685,7 @@ SLOW_TO_FAST_CONVERTERS = {
"BertTokenizer": BertConverter,
"CamembertTokenizer": CamembertConverter,
"ConvBertTokenizer": BertConverter,
"DebertaTokenizer": DebertaConverter,
"DistilBertTokenizer": BertConverter,
"DPRReaderTokenizer": BertConverter,
"DPRQuestionEncoderTokenizer": BertConverter,

5
src/transformers/models/auto/tokenization_auto.py
View File

@@ -157,6 +157,7 @@ if is_tokenizers_available():
from ..bert.tokenization_bert_fast import BertTokenizerFast
from ..camembert.tokenization_camembert_fast import CamembertTokenizerFast
from ..convbert.tokenization_convbert_fast import ConvBertTokenizerFast
from ..deberta.tokenization_deberta_fast import DebertaTokenizerFast
from ..distilbert.tokenization_distilbert_fast import DistilBertTokenizerFast
from ..dpr.tokenization_dpr_fast import DPRQuestionEncoderTokenizerFast
from ..electra.tokenization_electra_fast import ElectraTokenizerFast
@@ -181,6 +182,7 @@ if is_tokenizers_available():
from ..t5.tokenization_t5_fast import T5TokenizerFast
from ..xlm_roberta.tokenization_xlm_roberta_fast import XLMRobertaTokenizerFast
from ..xlnet.tokenization_xlnet_fast import XLNetTokenizerFast
else:
AlbertTokenizerFast = None
BartTokenizerFast = None
@@ -188,6 +190,7 @@ else:
BertTokenizerFast = None
CamembertTokenizerFast = None
ConvBertTokenizerFast = None
DebertaTokenizerFast = None
DistilBertTokenizerFast = None
DPRQuestionEncoderTokenizerFast = None
ElectraTokenizerFast = None
@@ -253,7 +256,7 @@ TOKENIZER_MAPPING = OrderedDict(
(CTRLConfig, (CTRLTokenizer, None)),
(FSMTConfig, (FSMTTokenizer, None)),
(BertGenerationConfig, (BertGenerationTokenizer, None)),
(DebertaConfig, (DebertaTokenizer, None)),
(DebertaConfig, (DebertaTokenizer, DebertaTokenizerFast)),
(DebertaV2Config, (DebertaV2Tokenizer, None)),
(RagConfig, (RagTokenizer, None)),
(XLMProphetNetConfig, (XLMProphetNetTokenizer, None)),

8
src/transformers/models/deberta/__init__.py
View File

@@ -18,7 +18,7 @@
from typing import TYPE_CHECKING
from ...file_utils import _BaseLazyModule, is_torch_available
from ...file_utils import _BaseLazyModule, is_tokenizers_available, is_torch_available
_import_structure = {
@@ -26,6 +26,9 @@ _import_structure = {
"tokenization_deberta": ["DebertaTokenizer"],
}
if is_tokenizers_available():
_import_structure["tokenization_deberta_fast"] = ["DebertaTokenizerFast"]
if is_torch_available():
_import_structure["modeling_deberta"] = [
"DEBERTA_PRETRAINED_MODEL_ARCHIVE_LIST",
@@ -42,6 +45,9 @@ if TYPE_CHECKING:
from .configuration_deberta import DEBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP, DebertaConfig
from .tokenization_deberta import DebertaTokenizer
if is_tokenizers_available():
from .tokenization_deberta_fast import DebertaTokenizerFast
if is_torch_available():
from .modeling_deberta import (
DEBERTA_PRETRAINED_MODEL_ARCHIVE_LIST,

207
src/transformers/models/deberta/tokenization_deberta_fast.py Normal file
View File

@@ -0,0 +1,207 @@
# coding=utf-8
# Copyright 2020 Microsoft and the HuggingFace Inc. team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
""" Fast Tokenization class for model DeBERTa."""
from typing import List, Optional
from ...tokenization_utils_base import AddedToken
from ...utils import logging
from ..gpt2.tokenization_gpt2_fast import GPT2TokenizerFast
from .tokenization_deberta import DebertaTokenizer
logger = logging.get_logger(__name__)
VOCAB_FILES_NAMES = {"vocab_file": "vocab.json", "merges_file": "merges.txt", "tokenizer_file": "tokenizer.json"}
PRETRAINED_VOCAB_FILES_MAP = {
"vocab_file": {
"microsoft/deberta-base": "https://huggingface.co/microsoft/deberta-base/resolve/main/vocab.json",
"microsoft/deberta-large": "https://huggingface.co/microsoft/deberta-large/resolve/main/vocab.json",
"microsoft/deberta-xlarge": "https://huggingface.co/microsoft/deberta-xlarge/resolve/main/vocab.json",
"microsoft/deberta-base-mnli": "https://huggingface.co/microsoft/deberta-base-mnli/resolve/main/vocab.json",
"microsoft/deberta-large-mnli": "https://huggingface.co/microsoft/deberta-large-mnli/resolve/main/vocab.json",
"microsoft/deberta-xlarge-mnli": "https://huggingface.co/microsoft/deberta-xlarge-mnli/resolve/main/vocab.json",
},
"merges_file": {
"microsoft/deberta-base": "https://huggingface.co/microsoft/deberta-base/resolve/main/merges.txt",
"microsoft/deberta-large": "https://huggingface.co/microsoft/deberta-large/resolve/main/merges.txt",
"microsoft/deberta-xlarge": "https://huggingface.co/microsoft/deberta-xlarge/resolve/main/merges.txt",
"microsoft/deberta-base-mnli": "https://huggingface.co/microsoft/deberta-base-mnli/resolve/main/merges.txt",
"microsoft/deberta-large-mnli": "https://huggingface.co/microsoft/deberta-large-mnli/resolve/main/merges.txt",
"microsoft/deberta-xlarge-mnli": "https://huggingface.co/microsoft/deberta-xlarge-mnli/resolve/main/merges.txt",
},
}
PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {
"microsoft/deberta-base": 512,
"microsoft/deberta-large": 512,
"microsoft/deberta-xlarge": 512,
"microsoft/deberta-base-mnli": 512,
"microsoft/deberta-large-mnli": 512,
"microsoft/deberta-xlarge-mnli": 512,
}
PRETRAINED_INIT_CONFIGURATION = {
"microsoft/deberta-base": {"do_lower_case": False},
"microsoft/deberta-large": {"do_lower_case": False},
}
class DebertaTokenizerFast(GPT2TokenizerFast):
"""
Constructs a "fast" DeBERTa tokenizer, which runs end-to-end tokenization: punctuation splitting + wordpiece. It is
backed by HuggingFace's `tokenizers` library.
Args:
vocab_file (:obj:`str`):
File containing the vocabulary.
do_lower_case (:obj:`bool`, `optional`, defaults to :obj:`True`):
Whether or not to lowercase the input when tokenizing.
unk_token (:obj:`str`, `optional`, defaults to :obj:`"[UNK]"`):
The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this
token instead.
sep_token (:obj:`str`, `optional`, defaults to :obj:`"[SEP]"`):
The separator token, which is used when building a sequence from multiple sequences, e.g. two sequences for
sequence classification or for a text and a question for question answering. It is also used as the last
token of a sequence built with special tokens.
pad_token (:obj:`str`, `optional`, defaults to :obj:`"[PAD]"`):
The token used for padding, for example when batching sequences of different lengths.
cls_token (:obj:`str`, `optional`, defaults to :obj:`"[CLS]"`):
The classifier token which is used when doing sequence classification (classification of the whole sequence
instead of per-token classification). It is the first token of the sequence when built with special tokens.
mask_token (:obj:`str`, `optional`, defaults to :obj:`"[MASK]"`):
The token used for masking values. This is the token used when training this model with masked language
modeling. This is the token which the model will try to predict.
"""
vocab_files_names = VOCAB_FILES_NAMES
pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
model_input_names = ["input_ids", "attention_mask", "token_type_ids"]
slow_tokenizer_class = DebertaTokenizer
def __init__(
self,
vocab_file,
merges_file,
tokenizer_file=None,
errors="replace",
bos_token="[CLS]",
eos_token="[SEP]",
sep_token="[SEP]",
cls_token="[CLS]",
unk_token="[UNK]",
pad_token="[PAD]",
mask_token="[MASK]",
add_prefix_space=False,
**kwargs
):
super().__init__(
vocab_file,
merges_file,
tokenizer_file=tokenizer_file,
errors=errors,
bos_token=bos_token,
eos_token=eos_token,
unk_token=unk_token,
sep_token=sep_token,
cls_token=cls_token,
pad_token=pad_token,
mask_token=mask_token,
add_prefix_space=add_prefix_space,
**kwargs,
)
@property
def mask_token(self) -> str:
"""
:obj:`str`: Mask token, to use when training a model with masked-language modeling. Log an error if used while
not having been set.
Deberta tokenizer has a special mask token to be used in the fill-mask pipeline. The mask token will greedily
comprise the space before the `[MASK]`.
"""
if self._mask_token is None and self.verbose:
logger.error("Using mask_token, but it is not set yet.")
return None
return str(self._mask_token)
@mask_token.setter
def mask_token(self, value):
"""
Overriding the default behavior of the mask token to have it eat the space before it.
"""
# Mask token behave like a normal word, i.e. include the space before it
# So we set lstrip to True
value = AddedToken(value, lstrip=True, rstrip=False) if isinstance(value, str) else value
self._mask_token = value
def build_inputs_with_special_tokens(
self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
) -> List[int]:
"""
Build model inputs from a sequence or a pair of sequence for sequence classification tasks by concatenating and
adding special tokens. A DeBERTa sequence has the following format:
- single sequence: [CLS] X [SEP]
- pair of sequences: [CLS] A [SEP] B [SEP]
Args:
token_ids_0 (:obj:`List[int]`):
List of IDs to which the special tokens will be added.
token_ids_1 (:obj:`List[int]`, `optional`):
Optional second list of IDs for sequence pairs.
Returns:
:obj:`List[int]`: List of `input IDs <../glossary.html#input-ids>`__ with the appropriate special tokens.
"""
if token_ids_1 is None:
return [self.cls_token_id] + token_ids_0 + [self.sep_token_id]
cls = [self.cls_token_id]
sep = [self.sep_token_id]
return cls + token_ids_0 + sep + token_ids_1 + sep
def create_token_type_ids_from_sequences(
self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
) -> List[int]:
"""
Create a mask from the two sequences passed to be used in a sequence-pair classification task. A DeBERTa
sequence pair mask has the following format:
::
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
| first sequence | second sequence |
If :obj:`token_ids_1` is :obj:`None`, this method only returns the first portion of the mask (0s).
Args:
token_ids_0 (:obj:`List[int]`):
List of IDs.
token_ids_1 (:obj:`List[int]`, `optional`):
Optional second list of IDs for sequence pairs.
Returns:
:obj:`List[int]`: List of `token type IDs <../glossary.html#token-type-ids>`_ according to the given
sequence(s).
"""
sep = [self.sep_token_id]
cls = [self.cls_token_id]
if token_ids_1 is None:
return len(cls + token_ids_0 + sep) * [0]
return len(cls + token_ids_0 + sep + token_ids_1 + sep) * [0]

9
src/transformers/utils/dummy_tokenizers_objects.py
View File

@@ -56,6 +56,15 @@ class ConvBertTokenizerFast:
requires_backends(self, ["tokenizers"])
class DebertaTokenizerFast:
def __init__(self, *args, **kwargs):
requires_backends(self, ["tokenizers"])
@classmethod
def from_pretrained(self, *args, **kwargs):
requires_backends(self, ["tokenizers"])
class DistilBertTokenizerFast:
def __init__(self, *args, **kwargs):
requires_backends(self, ["tokenizers"])

5
tests/test_tokenization_deberta.py
View File

@@ -18,7 +18,7 @@ import json
import os
import unittest
from transformers import DebertaTokenizer
from transformers import DebertaTokenizer, DebertaTokenizerFast
from transformers.models.deberta.tokenization_deberta import VOCAB_FILES_NAMES
from transformers.testing_utils import slow
@@ -28,7 +28,8 @@ from .test_tokenization_common import TokenizerTesterMixin
class DebertaTokenizationTest(TokenizerTesterMixin, unittest.TestCase):
tokenizer_class = DebertaTokenizer
test_rust_tokenizer = False
test_rust_tokenizer = True
rust_tokenizer_class = DebertaTokenizerFast
def setUp(self):
super().setUp()