CSS-LM

1
# coding=utf-8
2
# Copyright 2020 The Facebook AI Research Team Authors and The HuggingFace Inc. team.
3
#
4
# Licensed under the Apache License, Version 2.0 (the "License");
5
# you may not use this file except in compliance with the License.
6
# You may obtain a copy of the License at
7
#
8
#     http://www.apache.org/licenses/LICENSE-2.0
9
#
10
# Unless required by applicable law or agreed to in writing, software
11
# distributed under the License is distributed on an "AS IS" BASIS,
12
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13
# See the License for the specific language governing permissions and
14
# limitations under the License.
15

16
import logging
17
from typing import List, Optional
18

19
from .tokenization_roberta import RobertaTokenizer, RobertaTokenizerFast
20
from .tokenization_utils import BatchEncoding
21
from .tokenization_xlm_roberta import XLMRobertaTokenizer
22

23

24
logger = logging.getLogger(__name__)
25

26

27
# vocab and merges same as roberta
28
vocab_url = "https://s3.amazonaws.com/models.huggingface.co/bert/roberta-large-vocab.json"
29
merges_url = "https://s3.amazonaws.com/models.huggingface.co/bert/roberta-large-merges.txt"
30
_all_bart_models = [
31
    "facebook/bart-base",
32
    "facebook/bart-large",
33
    "facebook/bart-large-mnli",
34
    "facebook/bart-large-cnn",
35
    "facebook/bart-large-xsum",
36
    "yjernite/bart_eli5",
37
]
38

39

40
class BartTokenizer(RobertaTokenizer):
41
    # merges and vocab same as Roberta
42
    max_model_input_sizes = {m: 1024 for m in _all_bart_models}
43
    pretrained_vocab_files_map = {
44
        "vocab_file": {m: vocab_url for m in _all_bart_models},
45
        "merges_file": {m: merges_url for m in _all_bart_models},
46
    }
47

48

49
class BartTokenizerFast(RobertaTokenizerFast):
50
    # merges and vocab same as Roberta
51
    max_model_input_sizes = {m: 1024 for m in _all_bart_models}
52
    pretrained_vocab_files_map = {
53
        "vocab_file": {m: vocab_url for m in _all_bart_models},
54
        "merges_file": {m: merges_url for m in _all_bart_models},
55
    }
56

57

58
_all_mbart_models = ["facebook/mbart-large-en-ro", "facebook/mbart-large-cc25"]
59
SPM_URL = "https://s3.amazonaws.com/models.huggingface.co/bert/facebook/mbart-large-en-ro/sentence.bpe.model"
60

61
FAIRSEQ_LANGUAGE_CODES = [
62
    "ar_AR",
63
    "cs_CZ",
64
    "de_DE",
65
    "en_XX",
66
    "es_XX",
67
    "et_EE",
68
    "fi_FI",
69
    "fr_XX",
70
    "gu_IN",
71
    "hi_IN",
72
    "it_IT",
73
    "ja_XX",
74
    "kk_KZ",
75
    "ko_KR",
76
    "lt_LT",
77
    "lv_LV",
78
    "my_MM",
79
    "ne_NP",
80
    "nl_XX",
81
    "ro_RO",
82
    "ru_RU",
83
    "si_LK",
84
    "tr_TR",
85
    "vi_VN",
86
    "zh_CN",
87
]
88

89

90
class MBartTokenizer(XLMRobertaTokenizer):
91
    """
92
    This inherits from XLMRobertaTokenizer. ``prepare_translation_batch`` should be used to encode inputs.
93
    Other tokenizer methods like ``encode`` do not work properly.
94
    The tokenization method is ``<tokens> <eos> <language code>`` for source language documents, and
95
    ``<language code> <tokens> <eos>``` for target language documents.
96

97
    Examples::
98

99
        >>> from transformers import MBartTokenizer
100
        >>> tokenizer = MBartTokenizer.from_pretrained('facebook/mbart-large-en-ro')
101
        >>> example_english_phrase = " UN Chief Says There Is No Military Solution in Syria"
102
        >>> expected_translation_romanian = "Şeful ONU declară că nu există o soluţie militară în Siria"
103
        >>> batch: dict = tokenizer.prepare_translation_batch(
104
        ...     example_english_phrase, src_lang="en_XX", tgt_lang="ro_RO", tgt_texts=expected_translation_romanian
105
        ... )
106

107
    """
108

109
    vocab_files_names = {"vocab_file": "sentencepiece.bpe.model"}
110
    max_model_input_sizes = {m: 1024 for m in _all_mbart_models}
111
    pretrained_vocab_files_map = {"vocab_file": {m: SPM_URL for m in _all_mbart_models}}
112

113
    prefix_tokens: List[int] = []
114
    suffix_tokens: List[int] = []
115

116
    def __init__(self, *args, **kwargs):
117
        super().__init__(*args, **kwargs)
118

119
        self.sp_model_size = len(self.sp_model)
120
        self.lang_code_to_id = {
121
            code: self.sp_model_size + i + self.fairseq_offset for i, code in enumerate(FAIRSEQ_LANGUAGE_CODES)
122
        }
123
        self.id_to_lang_code = {v: k for k, v in self.lang_code_to_id.items()}
124
        self.cur_lang_code = self.lang_code_to_id["en_XX"]
125
        self.fairseq_tokens_to_ids["<mask>"] = len(self.sp_model) + len(self.lang_code_to_id) + self.fairseq_offset
126

127
        self.fairseq_tokens_to_ids.update(self.lang_code_to_id)
128
        self.fairseq_ids_to_tokens = {v: k for k, v in self.fairseq_tokens_to_ids.items()}
129
        self._additional_special_tokens = list(self.lang_code_to_id.keys())
130
        self.set_src_lang_special_tokens(kwargs.get("src_lang", "en_XX"))
131

132
    def build_inputs_with_special_tokens(
133
        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
134
    ) -> List[int]:
135
        """
136
        Build model inputs from a sequence or a pair of sequence for sequence classification tasks
137
        by concatenating and adding special tokens. The special tokens depend on calling set_lang.
138
        An MBART sequence has the following format, where ``X`` represents the sequence:
139
        - ``input_ids`` (for encoder) ``X [eos, src_lang_code]``
140
        - ``decoder_input_ids``: (for decoder) ``[tgt_lang_code] X [eos]``
141
        BOS is never used.
142
        Pairs of sequences are not the expected use case, but they will be handled without a separator.
143

144
        Args:
145
            token_ids_0 (:obj:`List[int]`):
146
                List of IDs to which the special tokens will be added
147
            token_ids_1 (:obj:`List[int]`, `optional`, defaults to :obj:`None`):
148
                Optional second list of IDs for sequence pairs.
149

150
        Returns:
151
            :obj:`List[int]`: list of `input IDs <../glossary.html#input-ids>`__ with the appropriate special tokens.
152
        """
153
        if token_ids_1 is None:
154
            return self.prefix_tokens + token_ids_0 + self.suffix_tokens
155
        # We don't expect to process pairs, but leave the pair logic for API consistency
156
        return self.prefix_tokens + token_ids_0 + token_ids_1 + self.suffix_tokens
157

158
    def get_special_tokens_mask(
159
        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None, already_has_special_tokens: bool = False
160
    ) -> List[int]:
161
        """
162
        Retrieves sequence ids from a token list that has no special tokens added. This method is called when adding
163
        special tokens using the tokenizer ``prepare_for_model`` methods.
164

165
        Args:
166
            token_ids_0 (:obj:`List[int]`):
167
                List of ids.
168
            token_ids_1 (:obj:`List[int]`, `optional`, defaults to :obj:`None`):
169
                Optional second list of IDs for sequence pairs.
170
            already_has_special_tokens (:obj:`bool`, `optional`, defaults to :obj:`False`):
171
                Set to True if the token list is already formatted with special tokens for the model
172

173
        Returns:
174
            :obj:`List[int]`: A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.
175
        """
176

177
        if already_has_special_tokens:
178
            if token_ids_1 is not None:
179
                raise ValueError(
180
                    "You should not supply a second sequence if the provided sequence of "
181
                    "ids is already formated with special tokens for the model."
182
                )
183
            return list(map(lambda x: 1 if x in [self.sep_token_id, self.cls_token_id] else 0, token_ids_0))
184
        prefix_ones = [1] * len(self.prefix_tokens)
185
        suffix_ones = [1] * len(self.suffix_tokens)
186
        if token_ids_1 is None:
187
            return prefix_ones + ([0] * len(token_ids_0)) + suffix_ones
188
        return prefix_ones + ([0] * len(token_ids_0)) + ([0] * len(token_ids_1)) + suffix_ones
189

190
    def prepare_translation_batch(
191
        self,
192
        src_texts: List[str],
193
        src_lang: str = "en_XX",
194
        tgt_texts: Optional[List[str]] = None,
195
        tgt_lang: str = "ro_RO",
196
        max_length: Optional[int] = None,
197
        max_target_length: Optional[int] = None,
198
        padding: str = "longest",
199
        return_tensors: str = "pt",
200
        **kwargs,
201
    ) -> BatchEncoding:
202
        """Prepare a batch that can be passed directly to an instance of MBartModel.
203
        Arguments:
204
            src_texts: list of src language texts
205
            src_lang: default en_XX (english), the language we are translating from
206
            tgt_texts: list of tgt language texts
207
            tgt_lang: default ro_RO (romanian), the language we are translating to
208
            max_length: (default=None, which defers to the config value of 1024 for facebook/mbart-large*
209
            padding: strategy for padding input_ids and decoder_input_ids. Should be max_length or longest.
210
            **kwargs: passed to self.__call__
211

212
        Returns:
213
            :obj:`BatchEncoding`: with keys input_ids, attention_mask, decoder_input_ids, decoder_attention_mask.
214
        """
215
        if max_length is None:
216
            max_length = self.max_len
217
        self.set_src_lang_special_tokens(src_lang)
218
        model_inputs: BatchEncoding = self(
219
            src_texts,
220
            add_special_tokens=True,
221
            return_tensors=return_tensors,
222
            max_length=max_length,
223
            padding=padding,
224
            truncation=True,
225
            **kwargs,
226
        )
227
        if tgt_texts is None:
228
            return model_inputs
229
        # Process tgt_texts
230
        if max_target_length is None:
231
            max_target_length = max_length
232
        self.set_tgt_lang_special_tokens(tgt_lang)
233
        decoder_inputs: BatchEncoding = self(
234
            tgt_texts,
235
            add_special_tokens=True,
236
            return_tensors=return_tensors,
237
            padding=padding,
238
            max_length=max_target_length,
239
            truncation=True,
240
            **kwargs,
241
        )
242
        for k, v in decoder_inputs.items():
243
            model_inputs[f"decoder_{k}"] = v
244

245
        self.set_src_lang_special_tokens(src_lang)  # sets to src_lang
246
        return model_inputs
247

248
    def set_src_lang_special_tokens(self, src_lang) -> None:
249
        """Reset the special tokens to the source lang setting. No prefix and suffix=[eos, cur_lang_code]."""
250
        self.cur_lang_code = self.lang_code_to_id[src_lang]
251
        self.prefix_tokens = []
252
        self.suffix_tokens = [self.eos_token_id, self.cur_lang_code]
253

254
    def set_tgt_lang_special_tokens(self, lang: str) -> None:
255
        """Reset the special tokens to the target language setting. Prefix [tgt_lang_code], suffix =[eos]."""
256
        self.cur_lang_code = self.lang_code_to_id[lang]
257
        self.prefix_tokens = [self.cur_lang_code]
258
        self.suffix_tokens = [self.eos_token_id]
259