test_tokenization_common.py 197 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
# coding=utf-8
# Copyright 2019 HuggingFace Inc.
#
# 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.
Aymeric Augustin's avatar
Aymeric Augustin committed
15

16

17
import inspect
18
import itertools
19
import json
thomwolf's avatar
thomwolf committed
20
import os
21
import pickle
22
import re
Aymeric Augustin's avatar
Aymeric Augustin committed
23
import shutil
24
import sys
25
import tempfile
Sylvain Gugger's avatar
Sylvain Gugger committed
26
import unittest
27
import unittest.mock as mock
28
from collections import OrderedDict
29
from itertools import takewhile
30
from pathlib import Path
31
from typing import TYPE_CHECKING, Any, Dict, List, Tuple, Union
Aymeric Augustin's avatar
Aymeric Augustin committed
32

33
from huggingface_hub import Repository, delete_repo, login
Sylvain Gugger's avatar
Sylvain Gugger committed
34
from requests.exceptions import HTTPError
35
from transformers import (
36
37
    AlbertTokenizer,
    AlbertTokenizerFast,
38
    AutoTokenizer,
Sylvain Gugger's avatar
Sylvain Gugger committed
39
    BertTokenizer,
40
    BertTokenizerFast,
41
42
43
    PreTrainedTokenizer,
    PreTrainedTokenizerBase,
    PreTrainedTokenizerFast,
44
    SpecialTokensMixin,
45
46
    Trainer,
    TrainingArguments,
47
    is_tf_available,
48
    is_tokenizers_available,
49
50
    is_torch_available,
)
51
from transformers.testing_utils import (
Sylvain Gugger's avatar
Sylvain Gugger committed
52
53
    PASS,
    USER,
54
55
    get_tests_dir,
    is_pt_tf_cross_test,
Sylvain Gugger's avatar
Sylvain Gugger committed
56
    is_staging_test,
57
58
59
60
61
    require_tf,
    require_tokenizers,
    require_torch,
    slow,
)
62
from transformers.tokenization_utils import AddedToken, Trie
63

64

65
66
67
68
if is_torch_available():
    import torch.nn as nn


69
if TYPE_CHECKING:
70
    from transformers import PretrainedConfig, PreTrainedModel, TFPreTrainedModel
71
72


73
74
75
76
77
78
79
80
81
sys.path.append(str(Path(__file__).parent.parent / "utils"))

from test_module.custom_tokenization import CustomTokenizer  # noqa E402


if is_tokenizers_available():
    from test_module.custom_tokenization_fast import CustomTokenizerFast


82
83
NON_ENGLISH_TAGS = ["chinese", "dutch", "french", "finnish", "german", "multilingual"]

84
85
86
87
88
SMALL_TRAINING_CORPUS = [
    ["This is the first sentence.", "This is the second one."],
    ["This sentence (contains #) over symbols and numbers 12 3.", "But not this one."],
]

89
90

def filter_non_english(_, pretrained_name: str):
Patrick von Platen's avatar
Patrick von Platen committed
91
    """Filter all the model for non-english language"""
92
93
94
95
96
97
98
    return not any([lang in pretrained_name for lang in NON_ENGLISH_TAGS])


def filter_roberta_detectors(_, pretrained_name: str):
    return "detector" not in pretrained_name


99
def merge_model_tokenizer_mappings(
LysandreJik's avatar
LysandreJik committed
100
101
102
103
104
105
    model_mapping: Dict["PretrainedConfig", Union["PreTrainedModel", "TFPreTrainedModel"]],
    tokenizer_mapping: Dict["PretrainedConfig", Tuple["PreTrainedTokenizer", "PreTrainedTokenizerFast"]],
) -> Dict[
    Union["PreTrainedTokenizer", "PreTrainedTokenizerFast"],
    Tuple["PretrainedConfig", Union["PreTrainedModel", "TFPreTrainedModel"]],
]:
106
107
108
109
    configurations = list(model_mapping.keys())
    model_tokenizer_mapping = OrderedDict([])

    for configuration in configurations:
110
111
112
113
114
115
116
117
        if configuration in model_mapping and configuration in tokenizer_mapping:
            model = model_mapping[configuration]
            tokenizer = tokenizer_mapping[configuration][0]
            tokenizer_fast = tokenizer_mapping[configuration][1]

            model_tokenizer_mapping.update({tokenizer: (configuration, model)})
            if tokenizer_fast is not None:
                model_tokenizer_mapping.update({tokenizer_fast: (configuration, model)})
118
119
120
121

    return model_tokenizer_mapping


122
class TokenizerTesterMixin:
123

124
    tokenizer_class = None
125
    rust_tokenizer_class = None
126
127
    test_slow_tokenizer = True
    test_rust_tokenizer = True
128
    space_between_special_tokens = False
129
130
131
    from_pretrained_kwargs = None
    from_pretrained_filter = None
    from_pretrained_vocab_key = "vocab_file"
132
    test_seq2seq = True
133

134
135
136
137
138
139
140
    # set to True to test a sentencepiece tokenizer
    test_sentencepiece = False

    # set to True to ignore casing when testing a sentencepiece tokenizer
    # test_sentencepiece must also be set to True
    test_sentencepiece_ignore_case = False

141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
    def setUp(self) -> None:
        # Tokenizer.filter makes it possible to filter which Tokenizer to case based on all the
        # information available in Tokenizer (name, rust class, python class, vocab key name)
        if self.test_rust_tokenizer:
            tokenizers_list = [
                (
                    self.rust_tokenizer_class,
                    pretrained_name,
                    self.from_pretrained_kwargs if self.from_pretrained_kwargs is not None else {},
                )
                for pretrained_name in self.rust_tokenizer_class.pretrained_vocab_files_map[
                    self.from_pretrained_vocab_key
                ].keys()
                if self.from_pretrained_filter is None
                or (self.from_pretrained_filter is not None and self.from_pretrained_filter(pretrained_name))
            ]
            self.tokenizers_list = tokenizers_list[:1]  # Let's just test the first pretrained vocab for speed
        else:
            self.tokenizers_list = []
        with open(f"{get_tests_dir()}/fixtures/sample_text.txt", encoding="utf-8") as f_data:
            self._data = f_data.read().replace("\n\n", "\n").strip()
162

163
        self.tmpdirname = tempfile.mkdtemp()
164

165
166
    def tearDown(self):
        shutil.rmtree(self.tmpdirname)
167

168
169
170
171
    def get_input_output_texts(self, tokenizer):
        input_txt = self.get_clean_sequence(tokenizer)[0]
        return input_txt, input_txt

172
    def get_clean_sequence(self, tokenizer, with_prefix_space=False, max_length=20, min_length=5) -> Tuple[str, list]:
173
174
175
176
177
        toks = [(i, tokenizer.decode([i], clean_up_tokenization_spaces=False)) for i in range(len(tokenizer))]
        toks = list(filter(lambda t: re.match(r"^[ a-zA-Z]+$", t[1]), toks))
        toks = list(filter(lambda t: [t[0]] == tokenizer.encode(t[1], add_special_tokens=False), toks))
        if max_length is not None and len(toks) > max_length:
            toks = toks[:max_length]
178
179
180
        if min_length is not None and len(toks) < min_length and len(toks) > 0:
            while len(toks) < min_length:
                toks = toks + toks
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
        # toks_str = [t[1] for t in toks]
        toks_ids = [t[0] for t in toks]

        # Ensure consistency
        output_txt = tokenizer.decode(toks_ids, clean_up_tokenization_spaces=False)
        if " " not in output_txt and len(toks_ids) > 1:
            output_txt = (
                tokenizer.decode([toks_ids[0]], clean_up_tokenization_spaces=False)
                + " "
                + tokenizer.decode(toks_ids[1:], clean_up_tokenization_spaces=False)
            )
        if with_prefix_space:
            output_txt = " " + output_txt
        output_ids = tokenizer.encode(output_txt, add_special_tokens=False)
        return output_txt, output_ids

197
    def get_tokenizers(self, fast=True, **kwargs) -> List[PreTrainedTokenizerBase]:
198
        if fast and self.test_rust_tokenizer and self.test_slow_tokenizer:
199
            return [self.get_tokenizer(**kwargs), self.get_rust_tokenizer(**kwargs)]
200
201
202
203
204
205
        elif fast and self.test_rust_tokenizer:
            return [self.get_rust_tokenizer(**kwargs)]
        elif self.test_slow_tokenizer:
            return [self.get_tokenizer(**kwargs)]
        else:
            raise ValueError("This tokenizer class has no tokenizer to be tested.")
206

207
208
    def get_tokenizer(self, **kwargs) -> PreTrainedTokenizer:
        return self.tokenizer_class.from_pretrained(self.tmpdirname, **kwargs)
209

210
    def get_rust_tokenizer(self, **kwargs) -> PreTrainedTokenizerFast:
211
        return self.rust_tokenizer_class.from_pretrained(self.tmpdirname, **kwargs)
212

213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
    def tokenizer_integration_test_util(
        self,
        expected_encoding: Dict,
        model_name: str,
        revision: str = None,
        sequences: List[str] = None,
        decode_kwargs: Dict[str, Any] = None,
        padding: bool = True,
    ):
        """
        Util for integration test.

        Text is tokenized and then reverted back to text. Both results are then checked.

        Args:
            expected_encoding:
                The expected result of the tokenizer output.
            model_name:
                The model name of the tokenizer to load and use.
            revision:
                The full git revision number of the model. This is to pin the
                tokenizer config and to avoid that tests start to fail if the
                config gets changed upstream.
            sequences:
                Can overwrite the texts that are used to check the tokenizer.
                This is useful if the tokenizer supports non english languages
                like france.
            decode_kwargs:
                Additional args for the ``decode`` function which reverts the
                tokenized text back to a string.
            padding:
                Activates and controls padding of the tokenizer.
        """
        decode_kwargs = {} if decode_kwargs is None else decode_kwargs

        if sequences is None:
            sequences = [
                "Transformers (formerly known as pytorch-transformers and pytorch-pretrained-bert) provides "
                "general-purpose architectures (BERT, GPT-2, RoBERTa, XLM, DistilBert, XLNet...) for Natural "
                "Language Understanding (NLU) and Natural Language Generation (NLG) with over 32+ pretrained "
                "models in 100+ languages and deep interoperability between Jax, PyTorch and TensorFlow.",
                "BERT is designed to pre-train deep bidirectional representations from unlabeled text by jointly "
                "conditioning on both left and right context in all layers.",
                "The quick brown fox jumps over the lazy dog.",
            ]

259
260
261
        if self.test_sentencepiece_ignore_case:
            sequences = [sequence.lower() for sequence in sequences]

262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
        tokenizer_classes = [self.tokenizer_class]
        if self.test_rust_tokenizer:
            tokenizer_classes.append(self.rust_tokenizer_class)

        for tokenizer_class in tokenizer_classes:
            tokenizer = tokenizer_class.from_pretrained(
                model_name,
                revision=revision,  # to pin the tokenizer version
            )

            encoding = tokenizer(sequences, padding=padding)
            decoded_sequences = [
                tokenizer.decode(seq, skip_special_tokens=True, **decode_kwargs) for seq in encoding["input_ids"]
            ]

            encoding_data = encoding.data
            self.assertDictEqual(encoding_data, expected_encoding)

            for expected, decoded in zip(sequences, decoded_sequences):
                if self.test_sentencepiece_ignore_case:
                    expected = expected.lower()
                self.assertEqual(expected, decoded)
thomwolf's avatar
thomwolf committed
284

285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
    def assert_padded_input_match(self, input_r: list, input_p: list, max_length: int, pad_token_id: int):
        # Ensure we match max_length
        self.assertEqual(len(input_r), max_length)
        self.assertEqual(len(input_p), max_length)

        # Ensure the number of padded tokens is the same
        padded_tokens_r = list(takewhile(lambda i: i == pad_token_id, reversed(input_r)))
        padded_tokens_p = list(takewhile(lambda i: i == pad_token_id, reversed(input_p)))
        self.assertSequenceEqual(padded_tokens_r, padded_tokens_p)

    def assert_batch_padded_input_match(
        self,
        input_r: dict,
        input_p: dict,
        max_length: int,
        pad_token_id: int,
        model_main_input_name: str = "input_ids",
    ):
        for i_r in input_r.values():
            self.assertEqual(len(i_r), 2), self.assertEqual(len(i_r[0]), max_length), self.assertEqual(
                len(i_r[1]), max_length
            )
            self.assertEqual(len(i_r), 2), self.assertEqual(len(i_r[0]), max_length), self.assertEqual(
                len(i_r[1]), max_length
            )

        for i_r, i_p in zip(input_r[model_main_input_name], input_p[model_main_input_name]):
            self.assert_padded_input_match(i_r, i_p, max_length, pad_token_id)

        for i_r, i_p in zip(input_r["attention_mask"], input_p["attention_mask"]):
            self.assertSequenceEqual(i_r, i_p)

317
318
319
    @staticmethod
    def convert_batch_encode_plus_format_to_encode_plus(batch_encode_plus_sequences):
        # Switch from batch_encode_plus format:   {'input_ids': [[...], [...]], ...}
320
        # to the list of examples/ encode_plus format: [{'input_ids': [...], ...}, {'input_ids': [...], ...}]
321
322
        return [
            {value: batch_encode_plus_sequences[value][i] for value in batch_encode_plus_sequences.keys()}
Lysandre Debut's avatar
Lysandre Debut committed
323
            for i in range(len(batch_encode_plus_sequences["input_ids"]))
324
325
        ]

326
327
328
    # TODO: this test can be combined with `test_sentencepiece_tokenize_and_convert_tokens_to_string` after the latter is extended to all tokenizers.
    def test_tokenize_special_tokens(self):
        """Test `tokenize` with special tokens."""
329
        tokenizers = self.get_tokenizers(fast=True, do_lower_case=True)
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
        for tokenizer in tokenizers:
            with self.subTest(f"{tokenizer.__class__.__name__}"):
                SPECIAL_TOKEN_1 = "[SPECIAL_TOKEN_1]"
                SPECIAL_TOKEN_2 = "[SPECIAL_TOKEN_2]"

                # TODO:
                # Can we combine `unique_no_split_tokens` and `all_special_tokens`(and properties related to it)
                # with one variable(property) for a better maintainability?

                # `add_tokens` method stores special tokens only in `tokenizer.unique_no_split_tokens`. (in tokenization_utils.py)
                tokenizer.add_tokens([SPECIAL_TOKEN_1], special_tokens=True)
                # `add_special_tokens` method stores special tokens in `tokenizer.additional_special_tokens`,
                # which also occur in `tokenizer.all_special_tokens`. (in tokenization_utils_base.py)
                tokenizer.add_special_tokens({"additional_special_tokens": [SPECIAL_TOKEN_2]})

                token_1 = tokenizer.tokenize(SPECIAL_TOKEN_1)
                token_2 = tokenizer.tokenize(SPECIAL_TOKEN_2)

                self.assertEqual(len(token_1), 1)
                self.assertEqual(len(token_2), 1)
                self.assertEqual(token_1[0], SPECIAL_TOKEN_1)
                self.assertEqual(token_2[0], SPECIAL_TOKEN_2)

353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
    # TODO: this test could be extended to all tokenizers - not just the sentencepiece
    def test_sentencepiece_tokenize_and_convert_tokens_to_string(self):
        """Test ``_tokenize`` and ``convert_tokens_to_string``."""
        if not self.test_sentencepiece:
            return

        tokenizer = self.get_tokenizer()
        text = "This is text to test the tokenizer."

        if self.test_sentencepiece_ignore_case:
            text = text.lower()

        tokens = tokenizer.tokenize(text)

        self.assertTrue(len(tokens) > 0)

        # check if converting back to original text works
        reverse_text = tokenizer.convert_tokens_to_string(tokens)

        if self.test_sentencepiece_ignore_case:
            reverse_text = reverse_text.lower()

        self.assertEqual(reverse_text, text)

377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
    def test_subword_regularization_tokenizer(self) -> None:
        if not self.test_sentencepiece:
            return

        # Subword regularization is only available for the slow tokenizer.
        sp_model_kwargs = {"enable_sampling": True, "alpha": 0.1, "nbest_size": -1}
        tokenizer = self.get_tokenizer(sp_model_kwargs=sp_model_kwargs)

        self.assertTrue(hasattr(tokenizer, "sp_model_kwargs"))
        self.assertIsNotNone(tokenizer.sp_model_kwargs)
        self.assertTrue(isinstance(tokenizer.sp_model_kwargs, dict))
        self.assertEqual(tokenizer.sp_model_kwargs, sp_model_kwargs)
        self.check_subword_sampling(tokenizer)

    def test_pickle_subword_regularization_tokenizer(self) -> None:
        if not self.test_sentencepiece:
            return

        """Google pickle __getstate__ __setstate__ if you are struggling with this."""
        # Subword regularization is only available for the slow tokenizer.
        sp_model_kwargs = {"enable_sampling": True, "alpha": 0.1, "nbest_size": -1}
        tokenizer = self.get_tokenizer(sp_model_kwargs=sp_model_kwargs)
        tokenizer_bin = pickle.dumps(tokenizer)
        del tokenizer
        tokenizer_new = pickle.loads(tokenizer_bin)

        self.assertTrue(hasattr(tokenizer_new, "sp_model_kwargs"))
        self.assertIsNotNone(tokenizer_new.sp_model_kwargs)
        self.assertTrue(isinstance(tokenizer_new.sp_model_kwargs, dict))
        self.assertEqual(tokenizer_new.sp_model_kwargs, sp_model_kwargs)
        self.check_subword_sampling(tokenizer_new)

409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
    def test_save_sentencepiece_tokenizer(self) -> None:
        if not self.test_sentencepiece or not self.test_slow_tokenizer:
            return
        # We want to verify that we will be able to save the tokenizer even if the original files that were used to
        # build the tokenizer have been deleted in the meantime.
        text = "This is text to test the tokenizer."

        tokenizer_slow_1 = self.get_tokenizer()
        encoding_tokenizer_slow_1 = tokenizer_slow_1(text)

        tmpdirname_1 = tempfile.mkdtemp()
        tmpdirname_2 = tempfile.mkdtemp()

        tokenizer_slow_1.save_pretrained(tmpdirname_1)
        tokenizer_slow_2 = self.tokenizer_class.from_pretrained(tmpdirname_1)
        encoding_tokenizer_slow_2 = tokenizer_slow_2(text)

        shutil.rmtree(tmpdirname_1)
        tokenizer_slow_2.save_pretrained(tmpdirname_2)

        tokenizer_slow_3 = self.tokenizer_class.from_pretrained(tmpdirname_2)
        encoding_tokenizer_slow_3 = tokenizer_slow_3(text)
        shutil.rmtree(tmpdirname_2)

        self.assertEqual(encoding_tokenizer_slow_1, encoding_tokenizer_slow_2)
        self.assertEqual(encoding_tokenizer_slow_1, encoding_tokenizer_slow_3)

436
437
438
439
440
441
442
443
444
445
446
447
    def test_model_input_names_signature(self):
        accepted_model_main_input_names = [
            "input_ids",  # nlp models
            "input_values",  # speech models
        ]

        tokenizers = self.get_tokenizers()
        for tokenizer in tokenizers:
            # first name of model_input_names has to correspond to main model input name
            # to make sure `tokenizer.pad(...)` works correctly
            self.assertTrue(tokenizer.model_input_names[0] in accepted_model_main_input_names)

448
449
450
451
452
453
454
455
456
    def test_rust_tokenizer_signature(self):
        if not self.test_rust_tokenizer:
            return

        signature = inspect.signature(self.rust_tokenizer_class.__init__)

        self.assertIn("tokenizer_file", signature.parameters)
        self.assertIsNone(signature.parameters["tokenizer_file"].default)

457
    def test_tokenizer_slow_store_full_signature(self):
458
459
460
        if not self.test_slow_tokenizer:
            return

461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
        signature = inspect.signature(self.tokenizer_class.__init__)
        tokenizer = self.get_tokenizer()

        for parameter_name, parameter in signature.parameters.items():
            if parameter.default != inspect.Parameter.empty:
                self.assertIn(parameter_name, tokenizer.init_kwargs)

    def test_tokenizer_fast_store_full_signature(self):
        if not self.test_rust_tokenizer:
            return

        signature = inspect.signature(self.rust_tokenizer_class.__init__)
        tokenizer = self.get_rust_tokenizer()

        for parameter_name, parameter in signature.parameters.items():
476
477
478
479
480
            if parameter.default != inspect.Parameter.empty and parameter_name not in [
                "vocab_file",
                "merges_file",
                "tokenizer_file",
            ]:
481
482
                self.assertIn(parameter_name, tokenizer.init_kwargs)

483
484
485
486
    def test_rust_and_python_full_tokenizers(self):
        if not self.test_rust_tokenizer:
            return

487
488
489
490
        if not self.test_slow_tokenizer:
            # as we don't have a slow version, we can't compare the outputs between slow and fast versions
            return

491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
        tokenizer = self.get_tokenizer()
        rust_tokenizer = self.get_rust_tokenizer()

        sequence, _ = self.get_input_output_texts(tokenizer)

        # We don't have an exact equivalence on `tokenize()` between Rust and Slow
        # Slow tokenizer only split tokens, Rust tokenizers will replace with <unk>
        # tokens = tokenizer.tokenize(sequence)
        # rust_tokens = rust_tokenizer.tokenize(sequence)
        # self.assertListEqual(tokens, rust_tokens)

        ids = tokenizer.encode(sequence, add_special_tokens=False)
        rust_ids = rust_tokenizer.encode(sequence, add_special_tokens=False)
        self.assertListEqual(ids, rust_ids)

        ids = tokenizer.encode(sequence, add_special_tokens=True)
        rust_ids = rust_tokenizer.encode(sequence, add_special_tokens=True)
        self.assertListEqual(ids, rust_ids)

510
    def test_tokenizers_common_properties(self):
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
        tokenizers = self.get_tokenizers()
        for tokenizer in tokenizers:
            with self.subTest(f"{tokenizer.__class__.__name__}"):
                attributes_list = [
                    "bos_token",
                    "eos_token",
                    "unk_token",
                    "sep_token",
                    "pad_token",
                    "cls_token",
                    "mask_token",
                ]
                for attr in attributes_list:
                    self.assertTrue(hasattr(tokenizer, attr))
                    self.assertTrue(hasattr(tokenizer, attr + "_id"))

                self.assertTrue(hasattr(tokenizer, "additional_special_tokens"))
                self.assertTrue(hasattr(tokenizer, "additional_special_tokens_ids"))

                attributes_list = [
                    "model_max_length",
                    "init_inputs",
                    "init_kwargs",
                ]
                if not isinstance(tokenizer, PreTrainedTokenizerFast):
                    attributes_list += [
                        "added_tokens_encoder",
                        "added_tokens_decoder",
                    ]
                for attr in attributes_list:
                    self.assertTrue(hasattr(tokenizer, attr))
542

543
544
    def test_save_and_load_tokenizer(self):
        # safety check on max_len default value so we are sure the test works
545
        tokenizers = self.get_tokenizers()
546
547
        for tokenizer in tokenizers:
            with self.subTest(f"{tokenizer.__class__.__name__}"):
Stas Bekman's avatar
Stas Bekman committed
548
                self.assertNotEqual(tokenizer.model_max_length, 42)
549

550
        # Now let's start the test
551
        tokenizers = self.get_tokenizers()
552
553
        for tokenizer in tokenizers:
            with self.subTest(f"{tokenizer.__class__.__name__}"):
554
555
556
557
                # Isolate this from the other tests because we save additional tokens/etc
                tmpdirname = tempfile.mkdtemp()

                sample_text = " He is very happy, UNwant\u00E9d,running"
558
                before_tokens = tokenizer.encode(sample_text, add_special_tokens=False)
559
560
561
562
563
564
565
566
567
568
                before_vocab = tokenizer.get_vocab()
                tokenizer.save_pretrained(tmpdirname)

                after_tokenizer = tokenizer.__class__.from_pretrained(tmpdirname)
                after_tokens = after_tokenizer.encode(sample_text, add_special_tokens=False)
                after_vocab = after_tokenizer.get_vocab()
                self.assertListEqual(before_tokens, after_tokens)
                self.assertDictEqual(before_vocab, after_vocab)

                shutil.rmtree(tmpdirname)
569

570
571
572
573
574
575
576
577
578
579
580
581
582
583
        tokenizers = self.get_tokenizers(model_max_length=42)
        for tokenizer in tokenizers:
            with self.subTest(f"{tokenizer.__class__.__name__}"):
                # Isolate this from the other tests because we save additional tokens/etc
                tmpdirname = tempfile.mkdtemp()

                sample_text = " He is very happy, UNwant\u00E9d,running"
                tokenizer.add_tokens(["bim", "bambam"])
                additional_special_tokens = tokenizer.additional_special_tokens
                additional_special_tokens.append("new_additional_special_token")
                tokenizer.add_special_tokens({"additional_special_tokens": additional_special_tokens})
                before_tokens = tokenizer.encode(sample_text, add_special_tokens=False)
                before_vocab = tokenizer.get_vocab()
                tokenizer.save_pretrained(tmpdirname)
584

585
586
587
                after_tokenizer = tokenizer.__class__.from_pretrained(tmpdirname)
                after_tokens = after_tokenizer.encode(sample_text, add_special_tokens=False)
                after_vocab = after_tokenizer.get_vocab()
588
                self.assertListEqual(before_tokens, after_tokens)
589
590
591
592
593
                self.assertDictEqual(before_vocab, after_vocab)
                self.assertIn("bim", after_vocab)
                self.assertIn("bambam", after_vocab)
                self.assertIn("new_additional_special_token", after_tokenizer.additional_special_tokens)
                self.assertEqual(after_tokenizer.model_max_length, 42)
594

595
                tokenizer = tokenizer.__class__.from_pretrained(tmpdirname, model_max_length=43)
596
                self.assertEqual(tokenizer.model_max_length, 43)
597

598
599
                shutil.rmtree(tmpdirname)

600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
        # Test that we can also use the non-legacy saving format for fast tokenizers
        tokenizers = self.get_tokenizers(model_max_length=42)
        for tokenizer in tokenizers:
            if not tokenizer.is_fast:
                continue
            with self.subTest(f"{tokenizer.__class__.__name__}"):
                # Isolate this from the other tests because we save additional tokens/etc
                tmpdirname = tempfile.mkdtemp()

                sample_text = " He is very happy, UNwant\u00E9d,running"
                tokenizer.add_tokens(["bim", "bambam"])
                additional_special_tokens = tokenizer.additional_special_tokens
                additional_special_tokens.append("new_additional_special_token")
                tokenizer.add_special_tokens({"additional_special_tokens": additional_special_tokens})
                before_tokens = tokenizer.encode(sample_text, add_special_tokens=False)
                before_vocab = tokenizer.get_vocab()
                tokenizer.save_pretrained(tmpdirname)

                after_tokenizer = tokenizer.__class__.from_pretrained(tmpdirname)
                after_tokens = after_tokenizer.encode(sample_text, add_special_tokens=False)
                after_vocab = after_tokenizer.get_vocab()
                self.assertListEqual(before_tokens, after_tokens)
                self.assertDictEqual(before_vocab, after_vocab)
                self.assertIn("bim", after_vocab)
                self.assertIn("bambam", after_vocab)
                self.assertIn("new_additional_special_token", after_tokenizer.additional_special_tokens)
                self.assertEqual(after_tokenizer.model_max_length, 42)

                tokenizer = tokenizer.__class__.from_pretrained(tmpdirname, model_max_length=43)
                self.assertEqual(tokenizer.model_max_length, 43)

                shutil.rmtree(tmpdirname)

633
    def test_pickle_tokenizer(self):
634
        """Google pickle __getstate__ __setstate__ if you are struggling with this."""
635
636
637
638
        tokenizers = self.get_tokenizers()
        for tokenizer in tokenizers:
            with self.subTest(f"{tokenizer.__class__.__name__}"):
                self.assertIsNotNone(tokenizer)
639

640
641
                text = "Munich and Berlin are nice cities"
                subwords = tokenizer.tokenize(text)
642

643
644
645
                filename = os.path.join(self.tmpdirname, "tokenizer.bin")
                with open(filename, "wb") as handle:
                    pickle.dump(tokenizer, handle)
646

647
648
                with open(filename, "rb") as handle:
                    tokenizer_new = pickle.load(handle)
649

650
                subwords_loaded = tokenizer_new.tokenize(text)
651

652
                self.assertListEqual(subwords, subwords_loaded)
653

654
    @require_tokenizers
Anthony MOI's avatar
Anthony MOI committed
655
656
657
658
659
660
    def test_pickle_added_tokens(self):
        tok1 = AddedToken("<s>", rstrip=True, lstrip=True, normalized=False, single_word=True)
        tok2 = pickle.loads(pickle.dumps(tok1))

        self.assertEqual(tok1.__getstate__(), tok2.__getstate__())

661
    def test_added_tokens_do_lower_case(self):
662
        tokenizers = self.get_tokenizers(do_lower_case=True)
663
664
        for tokenizer in tokenizers:
            with self.subTest(f"{tokenizer.__class__.__name__}"):
665
666
667
                if not hasattr(tokenizer, "do_lower_case") or not tokenizer.do_lower_case:
                    continue

668
                special_token = tokenizer.all_special_tokens[0]
669

670
671
                text = special_token + " aaaaa bbbbbb low cccccccccdddddddd l " + special_token
                text2 = special_token + " AAAAA BBBBBB low CCCCCCCCCDDDDDDDD l " + special_token
672

673
                toks_before_adding = tokenizer.tokenize(text)  # toks before adding new_toks
674

675
                new_toks = ["aaaaa bbbbbb", "cccccccccdddddddd", "AAAAA BBBBBB", "CCCCCCCCCDDDDDDDD"]
676
                added = tokenizer.add_tokens([AddedToken(tok, lstrip=True, rstrip=True) for tok in new_toks])
677

678
679
                toks_after_adding = tokenizer.tokenize(text)
                toks_after_adding2 = tokenizer.tokenize(text2)
680

681
682
683
684
685
686
687
688
                # Rust tokenizers dont't lowercase added tokens at the time calling `tokenizer.add_tokens`,
                # while python tokenizers do, so new_toks 0 and 2 would be treated as the same, so do new_toks 1 and 3.
                self.assertIn(added, [2, 4])

                self.assertListEqual(toks_after_adding, toks_after_adding2)
                self.assertTrue(
                    len(toks_before_adding) > len(toks_after_adding),  # toks_before_adding should be longer
                )
689

690
691
                # Check that none of the special tokens are lowercased
                sequence_with_special_tokens = "A " + " yEs ".join(tokenizer.all_special_tokens) + " B"
692
693
694
695
                # Convert the tokenized list to str as some special tokens are tokenized like normal tokens
                # which have a prefix spacee e.g. the mask token of Albert, and cannot match the original
                # special tokens exactly.
                tokenized_sequence = "".join(tokenizer.tokenize(sequence_with_special_tokens))
696

697
698
                for special_token in tokenizer.all_special_tokens:
                    self.assertTrue(special_token in tokenized_sequence)
699

700
        tokenizers = self.get_tokenizers(do_lower_case=True)
701
702
        for tokenizer in tokenizers:
            with self.subTest(f"{tokenizer.__class__.__name__}"):
703
704
705
                if hasattr(tokenizer, "do_lower_case") and tokenizer.do_lower_case:
                    continue

706
                special_token = tokenizer.all_special_tokens[0]
707

708
709
                text = special_token + " aaaaa bbbbbb low cccccccccdddddddd l " + special_token
                text2 = special_token + " AAAAA BBBBBB low CCCCCCCCCDDDDDDDD l " + special_token
710

711
                toks_before_adding = tokenizer.tokenize(text)  # toks before adding new_toks
thomwolf's avatar
thomwolf committed
712

713
714
                new_toks = ["aaaaa bbbbbb", "cccccccccdddddddd", "AAAAA BBBBBB", "CCCCCCCCCDDDDDDDD"]
                added = tokenizer.add_tokens([AddedToken(tok, lstrip=True, rstrip=True) for tok in new_toks])
715
                self.assertIn(added, [2, 4])
716

717
718
                toks_after_adding = tokenizer.tokenize(text)
                toks_after_adding2 = tokenizer.tokenize(text2)
719

720
721
722
723
724
725
726
                self.assertEqual(len(toks_after_adding), len(toks_after_adding2))  # Length should still be the same
                self.assertNotEqual(
                    toks_after_adding[1], toks_after_adding2[1]
                )  # But at least the first non-special tokens should differ
                self.assertTrue(
                    len(toks_before_adding) > len(toks_after_adding),  # toks_before_adding should be longer
                )
727

728
729
730
731
732
733
734
735
    def test_add_tokens_tokenizer(self):
        tokenizers = self.get_tokenizers(do_lower_case=False)
        for tokenizer in tokenizers:
            with self.subTest(f"{tokenizer.__class__.__name__}"):
                vocab_size = tokenizer.vocab_size
                all_size = len(tokenizer)

                self.assertNotEqual(vocab_size, 0)
736
737
738
739

                # We usually have added tokens from the start in tests because our vocab fixtures are
                # smaller than the original vocabs - let's not assert this
                # self.assertEqual(vocab_size, all_size)
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777

                new_toks = ["aaaaa bbbbbb", "cccccccccdddddddd"]
                added_toks = tokenizer.add_tokens(new_toks)
                vocab_size_2 = tokenizer.vocab_size
                all_size_2 = len(tokenizer)

                self.assertNotEqual(vocab_size_2, 0)
                self.assertEqual(vocab_size, vocab_size_2)
                self.assertEqual(added_toks, len(new_toks))
                self.assertEqual(all_size_2, all_size + len(new_toks))

                tokens = tokenizer.encode("aaaaa bbbbbb low cccccccccdddddddd l", add_special_tokens=False)

                self.assertGreaterEqual(len(tokens), 4)
                self.assertGreater(tokens[0], tokenizer.vocab_size - 1)
                self.assertGreater(tokens[-2], tokenizer.vocab_size - 1)

                new_toks_2 = {"eos_token": ">>>>|||<||<<|<<", "pad_token": "<<<<<|||>|>>>>|>"}
                added_toks_2 = tokenizer.add_special_tokens(new_toks_2)
                vocab_size_3 = tokenizer.vocab_size
                all_size_3 = len(tokenizer)

                self.assertNotEqual(vocab_size_3, 0)
                self.assertEqual(vocab_size, vocab_size_3)
                self.assertEqual(added_toks_2, len(new_toks_2))
                self.assertEqual(all_size_3, all_size_2 + len(new_toks_2))

                tokens = tokenizer.encode(
                    ">>>>|||<||<<|<< aaaaabbbbbb low cccccccccdddddddd <<<<<|||>|>>>>|> l", add_special_tokens=False
                )

                self.assertGreaterEqual(len(tokens), 6)
                self.assertGreater(tokens[0], tokenizer.vocab_size - 1)
                self.assertGreater(tokens[0], tokens[1])
                self.assertGreater(tokens[-2], tokenizer.vocab_size - 1)
                self.assertGreater(tokens[-2], tokens[-3])
                self.assertEqual(tokens[0], tokenizer.eos_token_id)
                self.assertEqual(tokens[-2], tokenizer.pad_token_id)
778

779
    def test_add_special_tokens(self):
780
781
782
783
        tokenizers = self.get_tokenizers(do_lower_case=False)
        for tokenizer in tokenizers:
            with self.subTest(f"{tokenizer.__class__.__name__}"):
                input_text, ids = self.get_clean_sequence(tokenizer)
784

785
                special_token = "[SPECIAL_TOKEN]"
786

787
788
789
                tokenizer.add_special_tokens({"cls_token": special_token})
                encoded_special_token = tokenizer.encode(special_token, add_special_tokens=False)
                self.assertEqual(len(encoded_special_token), 1)
790

791
792
                text = tokenizer.decode(ids + encoded_special_token, clean_up_tokenization_spaces=False)
                encoded = tokenizer.encode(text, add_special_tokens=False)
793

794
795
796
                input_encoded = tokenizer.encode(input_text, add_special_tokens=False)
                special_token_id = tokenizer.encode(special_token, add_special_tokens=False)
                self.assertEqual(encoded, input_encoded + special_token_id)
797

798
799
                decoded = tokenizer.decode(encoded, skip_special_tokens=True)
                self.assertTrue(special_token not in decoded)
800

801
    def test_internal_consistency(self):
802
803
804
805
        tokenizers = self.get_tokenizers()
        for tokenizer in tokenizers:
            with self.subTest(f"{tokenizer.__class__.__name__}"):
                input_text, output_text = self.get_input_output_texts(tokenizer)
806

807
808
809
810
                tokens = tokenizer.tokenize(input_text)
                ids = tokenizer.convert_tokens_to_ids(tokens)
                ids_2 = tokenizer.encode(input_text, add_special_tokens=False)
                self.assertListEqual(ids, ids_2)
811

812
813
814
815
                tokens_2 = tokenizer.convert_ids_to_tokens(ids)
                self.assertNotEqual(len(tokens_2), 0)
                text_2 = tokenizer.decode(ids)
                self.assertIsInstance(text_2, str)
816

817
                self.assertEqual(text_2, output_text)
818

819
    @require_tokenizers
820
    def test_encode_decode_with_spaces(self):
821
822
823
        tokenizers = self.get_tokenizers(do_lower_case=False)
        for tokenizer in tokenizers:
            with self.subTest(f"{tokenizer.__class__.__name__}"):
LysandreJik's avatar
LysandreJik committed
824

825
826
827
828
829
                new_toks = [
                    AddedToken("[ABC]", normalized=False),
                    AddedToken("[DEF]", normalized=False),
                    AddedToken("GHI IHG", normalized=False),
                ]
830
                tokenizer.add_tokens(new_toks)
831
                input = "[ABC][DEF][ABC]GHI IHG[DEF]"
832
                if self.space_between_special_tokens:
833
                    output = "[ABC] [DEF] [ABC] GHI IHG [DEF]"
834
835
                else:
                    output = input
836
                encoded = tokenizer.encode(input, add_special_tokens=False)
837
838
                decoded = tokenizer.decode(encoded, spaces_between_special_tokens=self.space_between_special_tokens)
                self.assertIn(decoded, [output, output.lower()])
839

840
    def test_pretrained_model_lists(self):
841
842
843
844
845
846
847
848
849
        # We should have at least one default checkpoint for each tokenizer
        # We should specify the max input length as well (used in some part to list the pretrained checkpoints)
        self.assertGreaterEqual(len(self.tokenizer_class.pretrained_vocab_files_map), 1)
        self.assertGreaterEqual(len(list(self.tokenizer_class.pretrained_vocab_files_map.values())[0]), 1)
        self.assertEqual(
            len(list(self.tokenizer_class.pretrained_vocab_files_map.values())[0]),
            len(self.tokenizer_class.max_model_input_sizes),
        )

850
851
852
853
        weights_list = list(self.tokenizer_class.max_model_input_sizes.keys())
        weights_lists_2 = []
        for file_id, map_list in self.tokenizer_class.pretrained_vocab_files_map.items():
            weights_lists_2.append(list(map_list.keys()))
854

855
856
        for weights_list_2 in weights_lists_2:
            self.assertListEqual(weights_list, weights_list_2)
LysandreJik's avatar
LysandreJik committed
857

858
    def test_mask_output(self):
859
        tokenizers = self.get_tokenizers(do_lower_case=False)
860
861
862
863
864
865
866
867
868
869
870
871
        for tokenizer in tokenizers:
            with self.subTest(f"{tokenizer.__class__.__name__}"):

                if (
                    tokenizer.build_inputs_with_special_tokens.__qualname__.split(".")[0] != "PreTrainedTokenizer"
                    and "token_type_ids" in tokenizer.model_input_names
                ):
                    seq_0 = "Test this method."
                    seq_1 = "With these inputs."
                    information = tokenizer.encode_plus(seq_0, seq_1, add_special_tokens=True)
                    sequences, mask = information["input_ids"], information["token_type_ids"]
                    self.assertEqual(len(sequences), len(mask))
872

873
874
875
876
877
878
879
880
    def test_token_type_ids(self):
        tokenizers = self.get_tokenizers()
        for tokenizer in tokenizers:
            with self.subTest(f"{tokenizer.__class__.__name__}"):
                seq_0 = "Test this method."

                # We want to have sequence 0 and sequence 1 are tagged
                # respectively with 0 and 1 token_ids
NielsRogge's avatar
NielsRogge committed
881
                # (regardless of whether the model use token type ids)
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
                # We use this assumption in the QA pipeline among other place
                output = tokenizer(seq_0, return_token_type_ids=True)
                self.assertIn(0, output["token_type_ids"])

    def test_sequence_ids(self):
        tokenizers = self.get_tokenizers()
        for tokenizer in tokenizers:
            if not tokenizer.is_fast:
                continue
            with self.subTest(f"{tokenizer.__class__.__name__}"):
                seq_0 = "Test this method."
                seq_1 = "With these inputs."

                # We want to have sequence 0 and sequence 1 are tagged
                # respectively with 0 and 1 token_ids
NielsRogge's avatar
NielsRogge committed
897
                # (regardless of whether the model use token type ids)
898
899
900
901
902
903
904
905
906
907
908
                # We use this assumption in the QA pipeline among other place
                output = tokenizer(seq_0)
                self.assertIn(0, output.sequence_ids())

                output = tokenizer(seq_0, seq_1)
                self.assertIn(0, output.sequence_ids())
                self.assertIn(1, output.sequence_ids())

                if tokenizer.num_special_tokens_to_add(pair=True):
                    self.assertIn(None, output.sequence_ids())

909
    def test_number_of_added_tokens(self):
910
911
912
        tokenizers = self.get_tokenizers(do_lower_case=False)
        for tokenizer in tokenizers:
            with self.subTest(f"{tokenizer.__class__.__name__}"):
913

914
915
                seq_0 = "Test this method."
                seq_1 = "With these inputs."
916

917
                sequences = tokenizer.encode(seq_0, seq_1, add_special_tokens=False)
918
                attached_sequences = tokenizer.encode(seq_0, seq_1, add_special_tokens=True)
919

920
921
922
923
924
                # Method is implemented (e.g. not GPT-2)
                if len(attached_sequences) != 2:
                    self.assertEqual(
                        tokenizer.num_special_tokens_to_add(pair=True), len(attached_sequences) - len(sequences)
                    )
925
926

    def test_maximum_encoding_length_single_input(self):
927
        tokenizers = self.get_tokenizers(do_lower_case=False, model_max_length=100)
928
929
        for tokenizer in tokenizers:
            with self.subTest(f"{tokenizer.__class__.__name__}"):
930
                seq_0, ids = self.get_clean_sequence(tokenizer, max_length=20)
931
932
933

                sequence = tokenizer.encode(seq_0, add_special_tokens=False)
                total_length = len(sequence)
934

Nicolas Patry's avatar
Nicolas Patry committed
935
                self.assertGreater(total_length, 4, "Issue with the testing sequence, please update it it's too short")
936
937
938
939
940
941
942
943

                # Test with max model input length
                model_max_length = tokenizer.model_max_length
                self.assertEqual(model_max_length, 100)
                seq_1 = seq_0 * model_max_length

                sequence1 = tokenizer(seq_1, add_special_tokens=False)
                total_length1 = len(sequence1["input_ids"])
Nicolas Patry's avatar
Nicolas Patry committed
944
945
946
                self.assertGreater(
                    total_length1, model_max_length, "Issue with the testing sequence, please update it it's too short"
                )
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962

                # Simple
                padding_strategies = (
                    [False, True, "longest"] if tokenizer.pad_token and tokenizer.pad_token_id >= 0 else [False]
                )
                for padding_state in padding_strategies:
                    with self.subTest(f"Padding: {padding_state}"):
                        for truncation_state in [True, "longest_first", "only_first"]:
                            with self.subTest(f"Truncation: {truncation_state}"):
                                output = tokenizer(seq_1, padding=padding_state, truncation=truncation_state)
                                self.assertEqual(len(output["input_ids"]), model_max_length)

                                output = tokenizer([seq_1], padding=padding_state, truncation=truncation_state)
                                self.assertEqual(len(output["input_ids"][0]), model_max_length)

                        # Simple with no truncation
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
                        # Reset warnings
                        tokenizer.deprecation_warnings = {}
                        with self.assertLogs("transformers", level="WARNING") as cm:
                            output = tokenizer(seq_1, padding=padding_state, truncation=False)
                            self.assertNotEqual(len(output["input_ids"]), model_max_length)
                        self.assertEqual(len(cm.records), 1)
                        self.assertTrue(
                            cm.records[0].message.startswith(
                                "Token indices sequence length is longer than the specified maximum sequence length for this model"
                            )
                        )

                        tokenizer.deprecation_warnings = {}
                        with self.assertLogs("transformers", level="WARNING") as cm:
                            output = tokenizer([seq_1], padding=padding_state, truncation=False)
                            self.assertNotEqual(len(output["input_ids"][0]), model_max_length)
                        self.assertEqual(len(cm.records), 1)
                        self.assertTrue(
                            cm.records[0].message.startswith(
                                "Token indices sequence length is longer than the specified maximum sequence length for this model"
                            )
                        )
985
986
987
988

                # Overflowing tokens
                stride = 2
                information = tokenizer(
989
990
991
992
993
994
                    seq_0,
                    max_length=total_length - 2,
                    add_special_tokens=False,
                    stride=stride,
                    truncation="longest_first",
                    return_overflowing_tokens=True,
995
                    # add_prefix_space=False,
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
                )

                # Overflowing tokens are handled quite differently in slow and fast tokenizers
                if isinstance(tokenizer, PreTrainedTokenizerFast):
                    truncated_sequence = information["input_ids"][0]
                    overflowing_tokens = information["input_ids"][1]
                    self.assertEqual(len(information["input_ids"]), 2)

                    self.assertEqual(len(truncated_sequence), total_length - 2)
                    self.assertEqual(truncated_sequence, sequence[:-2])

                    self.assertEqual(len(overflowing_tokens), 2 + stride)
                    self.assertEqual(overflowing_tokens, sequence[-(2 + stride) :])
                else:
                    truncated_sequence = information["input_ids"]
                    overflowing_tokens = information["overflowing_tokens"]
1012

1013
1014
                    self.assertEqual(len(truncated_sequence), total_length - 2)
                    self.assertEqual(truncated_sequence, sequence[:-2])
1015

1016
                    self.assertEqual(len(overflowing_tokens), 2 + stride)
1017
                    self.assertEqual(overflowing_tokens, sequence[-(2 + stride) :])
1018

1019
    def test_maximum_encoding_length_pair_input(self):
1020
        tokenizers = self.get_tokenizers(do_lower_case=False, model_max_length=100)
1021
1022
1023
1024
        for tokenizer in tokenizers:
            with self.subTest(f"{tokenizer.__class__.__name__}"):
                # Build a sequence from our model's vocabulary
                stride = 2
1025
                seq_0, ids = self.get_clean_sequence(tokenizer, max_length=20)
1026
                if len(ids) <= 2 + stride:
1027
1028
                    seq_0 = (seq_0 + " ") * (2 + stride)
                    ids = None
1029
1030

                seq0_tokens = tokenizer.encode(seq_0, add_special_tokens=False)
Nicolas Patry's avatar
Nicolas Patry committed
1031
                self.assertGreater(len(seq0_tokens), 2 + stride)
1032
1033
1034

                seq_1 = "This is another sentence to be encoded."
                seq1_tokens = tokenizer.encode(seq_1, add_special_tokens=False)
1035
                if abs(len(seq0_tokens) - len(seq1_tokens)) <= 2:
1036
1037
1038
1039
                    seq1_tokens = seq1_tokens + seq1_tokens
                    seq_1 = tokenizer.decode(seq1_tokens, clean_up_tokenization_spaces=False)
                seq1_tokens = tokenizer.encode(seq_1, add_special_tokens=False)

Nicolas Patry's avatar
Nicolas Patry committed
1040
                self.assertGreater(len(seq1_tokens), 2 + stride)
1041
1042
1043
1044
1045

                smallest = seq1_tokens if len(seq0_tokens) > len(seq1_tokens) else seq0_tokens

                # We are not using the special tokens - a bit too hard to test all the tokenizers with this
                # TODO try this again later
1046
                sequence = tokenizer.encode(seq_0, seq_1, add_special_tokens=False)  # , add_prefix_space=False)
1047
1048
1049
1050
1051

                # Test with max model input length
                model_max_length = tokenizer.model_max_length
                self.assertEqual(model_max_length, 100)
                seq_2 = seq_0 * model_max_length
Nicolas Patry's avatar
Nicolas Patry committed
1052
                self.assertGreater(len(seq_2), model_max_length)
1053
1054
1055
1056
1057

                sequence1 = tokenizer(seq_1, add_special_tokens=False)
                total_length1 = len(sequence1["input_ids"])
                sequence2 = tokenizer(seq_2, seq_1, add_special_tokens=False)
                total_length2 = len(sequence2["input_ids"])
Nicolas Patry's avatar
Nicolas Patry committed
1058
1059
1060
1061
1062
1063
                self.assertLess(
                    total_length1, model_max_length - 10, "Issue with the testing sequence, please update it."
                )
                self.assertGreater(
                    total_length2, model_max_length, "Issue with the testing sequence, please update it."
                )
1064
1065
1066
1067
1068
1069

                # Simple
                padding_strategies = (
                    [False, True, "longest"] if tokenizer.pad_token and tokenizer.pad_token_id >= 0 else [False]
                )
                for padding_state in padding_strategies:
1070
                    with self.subTest(f"{tokenizer.__class__.__name__} Padding: {padding_state}"):
1071
                        for truncation_state in [True, "longest_first", "only_first"]:
1072
                            with self.subTest(f"{tokenizer.__class__.__name__} Truncation: {truncation_state}"):
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
                                output = tokenizer(seq_2, seq_1, padding=padding_state, truncation=truncation_state)
                                self.assertEqual(len(output["input_ids"]), model_max_length)

                                output = tokenizer(
                                    [seq_2], [seq_1], padding=padding_state, truncation=truncation_state
                                )
                                self.assertEqual(len(output["input_ids"][0]), model_max_length)

                        # Simple
                        output = tokenizer(seq_1, seq_2, padding=padding_state, truncation="only_second")
                        self.assertEqual(len(output["input_ids"]), model_max_length)

                        output = tokenizer([seq_1], [seq_2], padding=padding_state, truncation="only_second")
                        self.assertEqual(len(output["input_ids"][0]), model_max_length)

                        # Simple with no truncation
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
                        # Reset warnings
                        tokenizer.deprecation_warnings = {}
                        with self.assertLogs("transformers", level="WARNING") as cm:
                            output = tokenizer(seq_1, seq_2, padding=padding_state, truncation=False)
                            self.assertNotEqual(len(output["input_ids"]), model_max_length)
                        self.assertEqual(len(cm.records), 1)
                        self.assertTrue(
                            cm.records[0].message.startswith(
                                "Token indices sequence length is longer than the specified maximum sequence length for this model"
                            )
                        )

                        tokenizer.deprecation_warnings = {}
                        with self.assertLogs("transformers", level="WARNING") as cm:
                            output = tokenizer([seq_1], [seq_2], padding=padding_state, truncation=False)
                            self.assertNotEqual(len(output["input_ids"][0]), model_max_length)
                        self.assertEqual(len(cm.records), 1)
                        self.assertTrue(
                            cm.records[0].message.startswith(
                                "Token indices sequence length is longer than the specified maximum sequence length for this model"
                            )
                        )
1111

1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
                truncated_first_sequence = tokenizer.encode(seq_0, add_special_tokens=False)[:-2] + tokenizer.encode(
                    seq_1, add_special_tokens=False
                )
                truncated_second_sequence = (
                    tokenizer.encode(seq_0, add_special_tokens=False)
                    + tokenizer.encode(seq_1, add_special_tokens=False)[:-2]
                )
                truncated_longest_sequence = (
                    truncated_first_sequence if len(seq0_tokens) > len(seq1_tokens) else truncated_second_sequence
                )

                overflow_first_sequence = tokenizer.encode(seq_0, add_special_tokens=False)[
                    -(2 + stride) :
                ] + tokenizer.encode(seq_1, add_special_tokens=False)
                overflow_second_sequence = (
                    tokenizer.encode(seq_0, add_special_tokens=False)
                    + tokenizer.encode(seq_1, add_special_tokens=False)[-(2 + stride) :]
                )
                overflow_longest_sequence = (
                    overflow_first_sequence if len(seq0_tokens) > len(seq1_tokens) else overflow_second_sequence
                )

                # Overflowing tokens are handled quite differently in slow and fast tokenizers
                if isinstance(tokenizer, PreTrainedTokenizerFast):
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
                    information = tokenizer(
                        seq_0,
                        seq_1,
                        max_length=len(sequence) - 2,
                        add_special_tokens=False,
                        stride=stride,
                        truncation="longest_first",
                        return_overflowing_tokens=True,
                        # add_prefix_space=False,
                    )
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
                    truncated_sequence = information["input_ids"][0]
                    overflowing_tokens = information["input_ids"][1]
                    self.assertEqual(len(information["input_ids"]), 2)

                    self.assertEqual(len(truncated_sequence), len(sequence) - 2)
                    self.assertEqual(truncated_sequence, truncated_longest_sequence)

                    self.assertEqual(len(overflowing_tokens), 2 + stride + len(smallest))
                    self.assertEqual(overflowing_tokens, overflow_longest_sequence)
                else:
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
                    # No overflowing tokens when using 'longest' in python tokenizers
                    with self.assertRaises(ValueError) as context:
                        information = tokenizer(
                            seq_0,
                            seq_1,
                            max_length=len(sequence) - 2,
                            add_special_tokens=False,
                            stride=stride,
                            truncation="longest_first",
                            return_overflowing_tokens=True,
                            # add_prefix_space=False,
                        )
1168

1169
1170
1171
1172
1173
1174
1175
                    self.assertTrue(
                        context.exception.args[0].startswith(
                            "Not possible to return overflowing tokens for pair of sequences with the "
                            "`longest_first`. Please select another truncation strategy than `longest_first`, "
                            "for instance `only_second` or `only_first`."
                        )
                    )
1176
1177

                # Overflowing tokens are handled quite differently in slow and fast tokenizers
1178
                if isinstance(tokenizer, PreTrainedTokenizerFast):
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
                    information = tokenizer(
                        seq_0,
                        seq_1,
                        max_length=len(sequence) - 2,
                        add_special_tokens=False,
                        stride=stride,
                        truncation=True,
                        return_overflowing_tokens=True,
                        # add_prefix_space=False,
                    )
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
                    truncated_sequence = information["input_ids"][0]
                    overflowing_tokens = information["input_ids"][1]
                    self.assertEqual(len(information["input_ids"]), 2)

                    self.assertEqual(len(truncated_sequence), len(sequence) - 2)
                    self.assertEqual(truncated_sequence, truncated_longest_sequence)

                    self.assertEqual(len(overflowing_tokens), 2 + stride + len(smallest))
                    self.assertEqual(overflowing_tokens, overflow_longest_sequence)
                else:
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
                    # No overflowing tokens when using 'longest' in python tokenizers
                    with self.assertRaises(ValueError) as context:
                        information = tokenizer(
                            seq_0,
                            seq_1,
                            max_length=len(sequence) - 2,
                            add_special_tokens=False,
                            stride=stride,
                            truncation=True,
                            return_overflowing_tokens=True,
                            # add_prefix_space=False,
                        )
1211

1212
1213
1214
1215
1216
1217
1218
                    self.assertTrue(
                        context.exception.args[0].startswith(
                            "Not possible to return overflowing tokens for pair of sequences with the "
                            "`longest_first`. Please select another truncation strategy than `longest_first`, "
                            "for instance `only_second` or `only_first`."
                        )
                    )
1219

1220
                information_first_truncated = tokenizer(
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
                    seq_0,
                    seq_1,
                    max_length=len(sequence) - 2,
                    add_special_tokens=False,
                    stride=stride,
                    truncation="only_first",
                    return_overflowing_tokens=True,
                    # add_prefix_space=False,
                )
                # Overflowing tokens are handled quite differently in slow and fast tokenizers
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
                if isinstance(tokenizer, PreTrainedTokenizerFast):
                    truncated_sequence = information_first_truncated["input_ids"][0]
                    overflowing_tokens = information_first_truncated["input_ids"][1]
                    self.assertEqual(len(information_first_truncated["input_ids"]), 2)

                    self.assertEqual(len(truncated_sequence), len(sequence) - 2)
                    self.assertEqual(truncated_sequence, truncated_first_sequence)

                    self.assertEqual(len(overflowing_tokens), 2 + stride + len(seq1_tokens))
                    self.assertEqual(overflowing_tokens, overflow_first_sequence)
                else:
                    truncated_sequence = information_first_truncated["input_ids"]
                    overflowing_tokens = information_first_truncated["overflowing_tokens"]

                    self.assertEqual(len(truncated_sequence), len(sequence) - 2)
                    self.assertEqual(truncated_sequence, truncated_first_sequence)

                    self.assertEqual(len(overflowing_tokens), 2 + stride)
                    self.assertEqual(overflowing_tokens, seq0_tokens[-(2 + stride) :])

1251
                information_second_truncated = tokenizer(
1252
1253
1254
1255
1256
1257
1258
                    seq_0,
                    seq_1,
                    max_length=len(sequence) - 2,
                    add_special_tokens=False,
                    stride=stride,
                    truncation="only_second",
                    return_overflowing_tokens=True,
1259
                    # add_prefix_space=False,
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
                )
                # Overflowing tokens are handled quite differently in slow and fast tokenizers
                if isinstance(tokenizer, PreTrainedTokenizerFast):
                    truncated_sequence = information_second_truncated["input_ids"][0]
                    overflowing_tokens = information_second_truncated["input_ids"][1]
                    self.assertEqual(len(information_second_truncated["input_ids"]), 2)

                    self.assertEqual(len(truncated_sequence), len(sequence) - 2)
                    self.assertEqual(truncated_sequence, truncated_second_sequence)

                    self.assertEqual(len(overflowing_tokens), 2 + stride + len(seq0_tokens))
                    self.assertEqual(overflowing_tokens, overflow_second_sequence)
                else:
                    truncated_sequence = information_second_truncated["input_ids"]
                    overflowing_tokens = information_second_truncated["overflowing_tokens"]

                    self.assertEqual(len(truncated_sequence), len(sequence) - 2)
                    self.assertEqual(truncated_sequence, truncated_second_sequence)
1278

1279
1280
                    self.assertEqual(len(overflowing_tokens), 2 + stride)
                    self.assertEqual(overflowing_tokens, seq1_tokens[-(2 + stride) :])
1281

1282
1283
1284
1285
1286
    # def test_encode_input_type(self):
    #     tokenizers = self.get_tokenizers(do_lower_case=False)
    #     for tokenizer in tokenizers:
    #         with self.subTest(f"{tokenizer.__class__.__name__}"):
    #             sequence = "Let's encode this sequence"
1287

1288
1289
1290
    #             tokens = sequence.split()  # tokenizer.tokenize(sequence)
    #             # input_ids = tokenizer.convert_tokens_to_ids(tokens)
    #             formatted_input = tokenizer.encode(sequence, add_special_tokens=True, add_prefix_space=False)
1291

1292
    #             self.assertEqual(
1293
    #                 tokenizer.encode(tokens, is_split_into_words=True, add_special_tokens=True), formatted_input
1294
1295
1296
    #             )
    #             # This is not supported with the Rust tokenizers
    #             # self.assertEqual(tokenizer.encode(input_ids, add_special_tokens=True), formatted_input)
1297

1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
    # def test_swap_special_token(self):
    #     tokenizers = self.get_tokenizers(do_lower_case=False)
    #     for tokenizer in tokenizers:
    #         with self.subTest(f"{tokenizer.__class__.__name__}"):
    #             # Our mask token
    #             mask = "<mask>"
    #             # We take a single word in the middle of the vocabulary
    #             all_tokens = sorted(tokenizer.get_vocab().keys())
    #             word = tokenizer.decode(tokenizer.encode(all_tokens[len(all_tokens)//2], add_special_tokens=False)[:1])

    #             sequence_0 = "Encode " + word + " sequence"
    #             sequence_masked_0 = "Encode " + mask + " sequence"

    #             sequence_1 = word + " this sequence"
    #             sequence_masked_1 = mask + " this sequence"

    #             # Add tokens so that masked token isn't split
    #             # tokens = [AddedToken(t, lstrip=True, normalized=False) for t in sequence.split()]
    #             # tokenizer.add_tokens(tokens)
    #             tokenizer.add_special_tokens(
    #                 {"mask_token": AddedToken(mask, normalized=False)}
    #             )  # Eat left space on Byte-level BPE tokenizers
    #             mask_ind = tokenizer.convert_tokens_to_ids(mask)

    #             # Test first masked sequence
    #             encoded_0 = tokenizer.encode(sequence_0, add_special_tokens=False)
    #             encoded_masked = tokenizer.encode(sequence_masked_0, add_special_tokens=False)
Nicolas Patry's avatar
Nicolas Patry committed
1325
    #             self.assertEqual(len(encoded_masked), len(encoded_0))
1326
1327
1328
1329
1330
1331
1332
1333
    #             mask_loc = encoded_masked.index(mask_ind)
    #             encoded_masked[mask_loc] = encoded_0[mask_loc]

    #             self.assertEqual(encoded_masked, encoded_0)

    #             # Test second masked sequence
    #             encoded_1 = tokenizer.encode(sequence_1, add_special_tokens=False)
    #             encoded_masked = tokenizer.encode(sequence_masked_1, add_special_tokens=False)
Nicolas Patry's avatar
Nicolas Patry committed
1334
    #             self.assertEqual(len(encoded_masked), len(encoded_1))
1335
1336
1337
1338
    #             mask_loc = encoded_masked.index(mask_ind)
    #             encoded_masked[mask_loc] = encoded_1[mask_loc]

    #             self.assertEqual(encoded_masked, encoded_1)
1339

1340
    def test_special_tokens_mask(self):
1341
1342
1343
1344
1345
1346
1347
        tokenizers = self.get_tokenizers(do_lower_case=False)
        for tokenizer in tokenizers:
            with self.subTest(f"{tokenizer.__class__.__name__}"):
                sequence_0 = "Encode this."
                # Testing single inputs
                encoded_sequence = tokenizer.encode(sequence_0, add_special_tokens=False)
                encoded_sequence_dict = tokenizer.encode_plus(
1348
                    sequence_0, add_special_tokens=True, return_special_tokens_mask=True  # , add_prefix_space=False
1349
1350
1351
1352
1353
1354
1355
                )
                encoded_sequence_w_special = encoded_sequence_dict["input_ids"]
                special_tokens_mask = encoded_sequence_dict["special_tokens_mask"]
                self.assertEqual(len(special_tokens_mask), len(encoded_sequence_w_special))

                filtered_sequence = [x for i, x in enumerate(encoded_sequence_w_special) if not special_tokens_mask[i]]
                self.assertEqual(encoded_sequence, filtered_sequence)
1356

1357
    def test_special_tokens_mask_input_pairs(self):
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
        tokenizers = self.get_tokenizers(do_lower_case=False)
        for tokenizer in tokenizers:
            with self.subTest(f"{tokenizer.__class__.__name__}"):
                sequence_0 = "Encode this."
                sequence_1 = "This one too please."
                encoded_sequence = tokenizer.encode(sequence_0, add_special_tokens=False)
                encoded_sequence += tokenizer.encode(sequence_1, add_special_tokens=False)
                encoded_sequence_dict = tokenizer.encode_plus(
                    sequence_0,
                    sequence_1,
                    add_special_tokens=True,
                    return_special_tokens_mask=True,
1370
                    # add_prefix_space=False,
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
                )
                encoded_sequence_w_special = encoded_sequence_dict["input_ids"]
                special_tokens_mask = encoded_sequence_dict["special_tokens_mask"]
                self.assertEqual(len(special_tokens_mask), len(encoded_sequence_w_special))

                filtered_sequence = [
                    (x if not special_tokens_mask[i] else None) for i, x in enumerate(encoded_sequence_w_special)
                ]
                filtered_sequence = [x for x in filtered_sequence if x is not None]
                self.assertEqual(encoded_sequence, filtered_sequence)
1381

1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
    def test_padding_side_in_kwargs(self):
        for tokenizer, pretrained_name, kwargs in self.tokenizers_list:
            with self.subTest(f"{tokenizer.__class__.__name__} ({pretrained_name})"):
                if self.test_rust_tokenizer:
                    tokenizer_r = self.rust_tokenizer_class.from_pretrained(
                        pretrained_name, padding_side="left", **kwargs
                    )
                    self.assertEqual(tokenizer_r.padding_side, "left")

                    tokenizer_r = self.rust_tokenizer_class.from_pretrained(
                        pretrained_name, padding_side="right", **kwargs
                    )
                    self.assertEqual(tokenizer_r.padding_side, "right")

                    self.assertRaises(
                        ValueError,
                        self.rust_tokenizer_class.from_pretrained,
                        pretrained_name,
                        padding_side="unauthorized",
                        **kwargs,
                    )

                if self.test_slow_tokenizer:
                    tokenizer_p = self.tokenizer_class.from_pretrained(pretrained_name, padding_side="left", **kwargs)
                    self.assertEqual(tokenizer_p.padding_side, "left")

                    tokenizer_p = self.tokenizer_class.from_pretrained(pretrained_name, padding_side="right", **kwargs)
                    self.assertEqual(tokenizer_p.padding_side, "right")

                    self.assertRaises(
                        ValueError,
                        self.tokenizer_class.from_pretrained,
                        pretrained_name,
                        padding_side="unauthorized",
                        **kwargs,
                    )

1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
    def test_truncation_side_in_kwargs(self):
        for tokenizer, pretrained_name, kwargs in self.tokenizers_list:
            with self.subTest(f"{tokenizer.__class__.__name__} ({pretrained_name})"):
                if self.test_rust_tokenizer:
                    tokenizer_r = self.rust_tokenizer_class.from_pretrained(
                        pretrained_name, truncation_side="left", **kwargs
                    )
                    self.assertEqual(tokenizer_r.truncation_side, "left")

                    tokenizer_r = self.rust_tokenizer_class.from_pretrained(
                        pretrained_name, truncation_side="right", **kwargs
                    )
                    self.assertEqual(tokenizer_r.truncation_side, "right")

                    self.assertRaises(
                        ValueError,
                        self.rust_tokenizer_class.from_pretrained,
                        pretrained_name,
                        truncation_side="unauthorized",
                        **kwargs,
                    )

                if self.test_slow_tokenizer:
                    tokenizer_p = self.tokenizer_class.from_pretrained(
                        pretrained_name, truncation_side="left", **kwargs
                    )
                    self.assertEqual(tokenizer_p.truncation_side, "left")

                    tokenizer_p = self.tokenizer_class.from_pretrained(
                        pretrained_name, truncation_side="right", **kwargs
                    )
                    self.assertEqual(tokenizer_p.truncation_side, "right")

                    self.assertRaises(
                        ValueError,
                        self.tokenizer_class.from_pretrained,
                        pretrained_name,
                        truncation_side="unauthorized",
                        **kwargs,
                    )

1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
    def test_right_and_left_padding(self):
        tokenizers = self.get_tokenizers(do_lower_case=False)
        for tokenizer in tokenizers:
            with self.subTest(f"{tokenizer.__class__.__name__}"):
                sequence = "Sequence"
                padding_size = 10

                # check correct behaviour if no pad_token_id exists and add it eventually
                self._check_no_pad_token_padding(tokenizer, sequence)

                padding_idx = tokenizer.pad_token_id

                # RIGHT PADDING - Check that it correctly pads when a maximum length is specified along with the padding flag set to True
                tokenizer.padding_side = "right"
                encoded_sequence = tokenizer.encode(sequence)
                sequence_length = len(encoded_sequence)
                padded_sequence = tokenizer.encode(
                    sequence, max_length=sequence_length + padding_size, padding="max_length"
                )
                padded_sequence_length = len(padded_sequence)
Nicolas Patry's avatar
Nicolas Patry committed
1480
1481
                self.assertEqual(sequence_length + padding_size, padded_sequence_length)
                self.assertEqual(encoded_sequence + [padding_idx] * padding_size, padded_sequence)
1482
1483
1484
1485
1486
1487
1488
1489
1490

                # LEFT PADDING - Check that it correctly pads when a maximum length is specified along with the padding flag set to True
                tokenizer.padding_side = "left"
                encoded_sequence = tokenizer.encode(sequence)
                sequence_length = len(encoded_sequence)
                padded_sequence = tokenizer.encode(
                    sequence, max_length=sequence_length + padding_size, padding="max_length"
                )
                padded_sequence_length = len(padded_sequence)
Nicolas Patry's avatar
Nicolas Patry committed
1491
1492
                self.assertEqual(sequence_length + padding_size, padded_sequence_length)
                self.assertEqual([padding_idx] * padding_size + encoded_sequence, padded_sequence)
1493
1494
1495
1496
1497
1498
1499
1500

                # RIGHT & LEFT PADDING - Check that nothing is done for 'longest' and 'no_padding'
                encoded_sequence = tokenizer.encode(sequence)
                sequence_length = len(encoded_sequence)

                tokenizer.padding_side = "right"
                padded_sequence_right = tokenizer.encode(sequence, padding=True)
                padded_sequence_right_length = len(padded_sequence_right)
Nicolas Patry's avatar
Nicolas Patry committed
1501
1502
                self.assertEqual(sequence_length, padded_sequence_right_length)
                self.assertEqual(encoded_sequence, padded_sequence_right)
1503
1504
1505
1506

                tokenizer.padding_side = "left"
                padded_sequence_left = tokenizer.encode(sequence, padding="longest")
                padded_sequence_left_length = len(padded_sequence_left)
Nicolas Patry's avatar
Nicolas Patry committed
1507
1508
                self.assertEqual(sequence_length, padded_sequence_left_length)
                self.assertEqual(encoded_sequence, padded_sequence_left)
1509
1510
1511
1512

                tokenizer.padding_side = "right"
                padded_sequence_right = tokenizer.encode(sequence)
                padded_sequence_right_length = len(padded_sequence_right)
Nicolas Patry's avatar
Nicolas Patry committed
1513
1514
                self.assertEqual(sequence_length, padded_sequence_right_length)
                self.assertEqual(encoded_sequence, padded_sequence_right)
1515
1516
1517
1518

                tokenizer.padding_side = "left"
                padded_sequence_left = tokenizer.encode(sequence, padding=False)
                padded_sequence_left_length = len(padded_sequence_left)
Nicolas Patry's avatar
Nicolas Patry committed
1519
1520
                self.assertEqual(sequence_length, padded_sequence_left_length)
                self.assertEqual(encoded_sequence, padded_sequence_left)
1521

1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
    def test_right_and_left_truncation(self):
        tokenizers = self.get_tokenizers(do_lower_case=False)
        for tokenizer in tokenizers:
            with self.subTest(f"{tokenizer.__class__.__name__}"):
                sequence = "This is a test sequence"

                # RIGHT PADDING - Check that it correctly pads when a maximum length is specified along with the padding flag set to True
                truncation_size = 3
                tokenizer.truncation_side = "right"
                encoded_sequence = tokenizer.encode(sequence, add_special_tokens=False)
                sequence_length = len(encoded_sequence)
                # Remove EOS/BOS tokens
                truncated_sequence = tokenizer.encode(
                    sequence, max_length=sequence_length - truncation_size, truncation=True, add_special_tokens=False
                )
                truncated_sequence_length = len(truncated_sequence)
                self.assertEqual(sequence_length, truncated_sequence_length + truncation_size)
                self.assertEqual(encoded_sequence[:-truncation_size], truncated_sequence)

                # LEFT PADDING - Check that it correctly pads when a maximum length is specified along with the truncation flag set to True
                tokenizer.truncation_side = "left"
                sequence_length = len(encoded_sequence)
                truncated_sequence = tokenizer.encode(
                    sequence, max_length=sequence_length - truncation_size, truncation=True, add_special_tokens=False
                )
                truncated_sequence_length = len(truncated_sequence)
                self.assertEqual(sequence_length, truncated_sequence_length + truncation_size)
                self.assertEqual(encoded_sequence[truncation_size:], truncated_sequence)

                # RIGHT & LEFT PADDING - Check that nothing is done for 'longest' and 'no_truncation'
                sequence_length = len(encoded_sequence)

                tokenizer.truncation_side = "right"
                truncated_sequence_right = tokenizer.encode(sequence, truncation=True, add_special_tokens=False)
                truncated_sequence_right_length = len(truncated_sequence_right)
                self.assertEqual(sequence_length, truncated_sequence_right_length)
                self.assertEqual(encoded_sequence, truncated_sequence_right)

                tokenizer.truncation_side = "left"
                truncated_sequence_left = tokenizer.encode(
                    sequence, truncation="longest_first", add_special_tokens=False
                )
                truncated_sequence_left_length = len(truncated_sequence_left)
                self.assertEqual(sequence_length, truncated_sequence_left_length)
                self.assertEqual(encoded_sequence, truncated_sequence_left)

                tokenizer.truncation_side = "right"
                truncated_sequence_right = tokenizer.encode(sequence, add_special_tokens=False)
                truncated_sequence_right_length = len(truncated_sequence_right)
                self.assertEqual(sequence_length, truncated_sequence_right_length)
                self.assertEqual(encoded_sequence, truncated_sequence_right)

                tokenizer.truncation_side = "left"
                truncated_sequence_left = tokenizer.encode(sequence, truncation=False, add_special_tokens=False)
                truncated_sequence_left_length = len(truncated_sequence_left)
                self.assertEqual(sequence_length, truncated_sequence_left_length)
                self.assertEqual(encoded_sequence, truncated_sequence_left)

1580
    def test_padding_to_max_length(self):
1581
        """We keep this test for backward compatibility but it should be remove when `pad_to_max_length` is deprecated."""
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
        tokenizers = self.get_tokenizers(do_lower_case=False)
        for tokenizer in tokenizers:
            with self.subTest(f"{tokenizer.__class__.__name__}"):
                sequence = "Sequence"
                padding_size = 10

                # check correct behaviour if no pad_token_id exists and add it eventually
                self._check_no_pad_token_padding(tokenizer, sequence)

                padding_idx = tokenizer.pad_token_id

                # Check that it correctly pads when a maximum length is specified along with the padding flag set to True
                tokenizer.padding_side = "right"
                encoded_sequence = tokenizer.encode(sequence)
                sequence_length = len(encoded_sequence)
1597
                # FIXME: the next line should be padding(max_length) to avoid warning
1598
1599
1600
1601
                padded_sequence = tokenizer.encode(
                    sequence, max_length=sequence_length + padding_size, pad_to_max_length=True
                )
                padded_sequence_length = len(padded_sequence)
Nicolas Patry's avatar
Nicolas Patry committed
1602
1603
                self.assertEqual(sequence_length + padding_size, padded_sequence_length)
                self.assertEqual(encoded_sequence + [padding_idx] * padding_size, padded_sequence)
1604
1605
1606
1607
1608
1609
1610
1611

                # Check that nothing is done when a maximum length is not specified
                encoded_sequence = tokenizer.encode(sequence)
                sequence_length = len(encoded_sequence)

                tokenizer.padding_side = "right"
                padded_sequence_right = tokenizer.encode(sequence, pad_to_max_length=True)
                padded_sequence_right_length = len(padded_sequence_right)
Nicolas Patry's avatar
Nicolas Patry committed
1612
1613
                self.assertEqual(sequence_length, padded_sequence_right_length)
                self.assertEqual(encoded_sequence, padded_sequence_right)
1614

1615
1616
1617
    def test_padding_to_multiple_of(self):
        tokenizers = self.get_tokenizers()
        for tokenizer in tokenizers:
1618
1619
1620
1621
            with self.subTest(f"{tokenizer.__class__.__name__}"):
                if tokenizer.pad_token is None:
                    self.skipTest("No padding token.")
                else:
1622
1623
1624
                    empty_tokens = tokenizer("", padding=True, pad_to_multiple_of=8)
                    normal_tokens = tokenizer("This is a sample input", padding=True, pad_to_multiple_of=8)
                    for key, value in empty_tokens.items():
1625
                        self.assertEqual(len(value) % 8, 0, f"BatchEncoding.{key} is not multiple of 8")
1626
                    for key, value in normal_tokens.items():
1627
                        self.assertEqual(len(value) % 8, 0, f"BatchEncoding.{key} is not multiple of 8")
1628
1629
1630

                    normal_tokens = tokenizer("This", pad_to_multiple_of=8)
                    for key, value in normal_tokens.items():
1631
                        self.assertNotEqual(len(value) % 8, 0, f"BatchEncoding.{key} is not multiple of 8")
1632
1633
1634
1635

                    # Should also work with truncation
                    normal_tokens = tokenizer("This", padding=True, truncation=True, pad_to_multiple_of=8)
                    for key, value in normal_tokens.items():
1636
                        self.assertEqual(len(value) % 8, 0, f"BatchEncoding.{key} is not multiple of 8")
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648

                    # truncation to something which is not a multiple of pad_to_multiple_of raises an error
                    self.assertRaises(
                        ValueError,
                        tokenizer.__call__,
                        "This",
                        padding=True,
                        truncation=True,
                        max_length=12,
                        pad_to_multiple_of=8,
                    )

1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
    def test_padding_with_attention_mask(self):
        tokenizers = self.get_tokenizers()
        for tokenizer in tokenizers:
            with self.subTest(f"{tokenizer.__class__.__name__}"):
                if tokenizer.pad_token is None:
                    self.skipTest("No padding token.")
                if "attention_mask" not in tokenizer.model_input_names:
                    self.skipTest("This model does not use attention mask.")

                features = [
                    {"input_ids": [1, 2, 3, 4, 5, 6], "attention_mask": [1, 1, 1, 1, 1, 0]},
                    {"input_ids": [1, 2, 3], "attention_mask": [1, 1, 0]},
                ]
                padded_features = tokenizer.pad(features)
                if tokenizer.padding_side == "right":
                    self.assertListEqual(padded_features["attention_mask"], [[1, 1, 1, 1, 1, 0], [1, 1, 0, 0, 0, 0]])
                else:
                    self.assertListEqual(padded_features["attention_mask"], [[1, 1, 1, 1, 1, 0], [0, 0, 0, 1, 1, 0]])

1668
    def test_encode_plus_with_padding(self):
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
        tokenizers = self.get_tokenizers(do_lower_case=False)
        for tokenizer in tokenizers:
            with self.subTest(f"{tokenizer.__class__.__name__}"):
                sequence = "Sequence"

                # check correct behaviour if no pad_token_id exists and add it eventually
                self._check_no_pad_token_padding(tokenizer, sequence)

                padding_size = 10
                padding_idx = tokenizer.pad_token_id
                token_type_padding_idx = tokenizer.pad_token_type_id

                encoded_sequence = tokenizer.encode_plus(sequence, return_special_tokens_mask=True)
                input_ids = encoded_sequence["input_ids"]
                special_tokens_mask = encoded_sequence["special_tokens_mask"]
                sequence_length = len(input_ids)

                # Test 'longest' and 'no_padding' don't do anything
                tokenizer.padding_side = "right"

Lysandre's avatar
Lysandre committed
1689
1690
1691
1692
1693
                not_padded_sequence = tokenizer.encode_plus(
                    sequence,
                    padding=True,
                    return_special_tokens_mask=True,
                )
1694
1695
1696
1697
1698
                not_padded_input_ids = not_padded_sequence["input_ids"]

                not_padded_special_tokens_mask = not_padded_sequence["special_tokens_mask"]
                not_padded_sequence_length = len(not_padded_input_ids)

Nicolas Patry's avatar
Nicolas Patry committed
1699
1700
1701
                self.assertEqual(sequence_length, not_padded_sequence_length)
                self.assertEqual(input_ids, not_padded_input_ids)
                self.assertEqual(special_tokens_mask, not_padded_special_tokens_mask)
1702

Lysandre's avatar
Lysandre committed
1703
1704
1705
1706
1707
                not_padded_sequence = tokenizer.encode_plus(
                    sequence,
                    padding=False,
                    return_special_tokens_mask=True,
                )
1708
1709
1710
1711
1712
                not_padded_input_ids = not_padded_sequence["input_ids"]

                not_padded_special_tokens_mask = not_padded_sequence["special_tokens_mask"]
                not_padded_sequence_length = len(not_padded_input_ids)

Nicolas Patry's avatar
Nicolas Patry committed
1713
1714
1715
                self.assertEqual(sequence_length, not_padded_sequence_length)
                self.assertEqual(input_ids, not_padded_input_ids)
                self.assertEqual(special_tokens_mask, not_padded_special_tokens_mask)
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730

                # Test right padding
                tokenizer.padding_side = "right"

                right_padded_sequence = tokenizer.encode_plus(
                    sequence,
                    max_length=sequence_length + padding_size,
                    padding="max_length",
                    return_special_tokens_mask=True,
                )
                right_padded_input_ids = right_padded_sequence["input_ids"]

                right_padded_special_tokens_mask = right_padded_sequence["special_tokens_mask"]
                right_padded_sequence_length = len(right_padded_input_ids)

Nicolas Patry's avatar
Nicolas Patry committed
1731
1732
1733
                self.assertEqual(sequence_length + padding_size, right_padded_sequence_length)
                self.assertEqual(input_ids + [padding_idx] * padding_size, right_padded_input_ids)
                self.assertEqual(special_tokens_mask + [1] * padding_size, right_padded_special_tokens_mask)
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746

                # Test left padding
                tokenizer.padding_side = "left"
                left_padded_sequence = tokenizer.encode_plus(
                    sequence,
                    max_length=sequence_length + padding_size,
                    padding="max_length",
                    return_special_tokens_mask=True,
                )
                left_padded_input_ids = left_padded_sequence["input_ids"]
                left_padded_special_tokens_mask = left_padded_sequence["special_tokens_mask"]
                left_padded_sequence_length = len(left_padded_input_ids)

Nicolas Patry's avatar
Nicolas Patry committed
1747
1748
1749
                self.assertEqual(sequence_length + padding_size, left_padded_sequence_length)
                self.assertEqual([padding_idx] * padding_size + input_ids, left_padded_input_ids)
                self.assertEqual([1] * padding_size + special_tokens_mask, left_padded_special_tokens_mask)
1750
1751
1752
1753
1754
1755

                if "token_type_ids" in tokenizer.model_input_names:
                    token_type_ids = encoded_sequence["token_type_ids"]
                    left_padded_token_type_ids = left_padded_sequence["token_type_ids"]
                    right_padded_token_type_ids = right_padded_sequence["token_type_ids"]

Nicolas Patry's avatar
Nicolas Patry committed
1756
1757
1758
1759
1760
1761
                    self.assertEqual(
                        token_type_ids + [token_type_padding_idx] * padding_size, right_padded_token_type_ids
                    )
                    self.assertEqual(
                        [token_type_padding_idx] * padding_size + token_type_ids, left_padded_token_type_ids
                    )
1762
1763
1764
1765
1766
1767

                if "attention_mask" in tokenizer.model_input_names:
                    attention_mask = encoded_sequence["attention_mask"]
                    right_padded_attention_mask = right_padded_sequence["attention_mask"]
                    left_padded_attention_mask = left_padded_sequence["attention_mask"]

Nicolas Patry's avatar
Nicolas Patry committed
1768
1769
                    self.assertEqual(attention_mask + [0] * padding_size, right_padded_attention_mask)
                    self.assertEqual([0] * padding_size + attention_mask, left_padded_attention_mask)
1770
1771
1772
1773
1774

    def test_separate_tokenizers(self):
        # This tests that tokenizers don't impact others. Unfortunately the case where it fails is when
        # we're loading an S3 configuration from a pre-trained identifier, and we have no way of testing those today.

1775
1776
1777
1778
1779
        tokenizers = self.get_tokenizers(random_argument=True)
        new_tokenizers = self.get_tokenizers(random_argument=False)

        for tokenizer, new_tokenizer in zip(tokenizers, new_tokenizers):
            with self.subTest(f"{tokenizer.__class__.__name__}"):
Nicolas Patry's avatar
Nicolas Patry committed
1780
1781
1782
                self.assertTrue(tokenizer.init_kwargs["random_argument"])
                self.assertTrue(tokenizer.init_kwargs["random_argument"])
                self.assertFalse(new_tokenizer.init_kwargs["random_argument"])
1783
1784

    def test_get_vocab(self):
1785
1786
1787
        tokenizers = self.get_tokenizers(do_lower_case=False)
        for tokenizer in tokenizers:
            with self.subTest(f"{tokenizer.__class__.__name__}"):
1788
1789
1790
                vocab_dict = tokenizer.get_vocab()
                self.assertIsInstance(vocab_dict, dict)
                self.assertGreaterEqual(len(tokenizer), len(vocab_dict))
1791

1792
                vocab = [tokenizer.convert_ids_to_tokens(i) for i in range(len(tokenizer))]
1793
                self.assertEqual(len(vocab), len(tokenizer))
1794

1795
                tokenizer.add_tokens(["asdfasdfasdfasdf"])
1796
                vocab = [tokenizer.convert_ids_to_tokens(i) for i in range(len(tokenizer))]
1797
                self.assertEqual(len(vocab), len(tokenizer))
1798

1799
    def test_conversion_reversible(self):
1800
1801
1802
1803
1804
        tokenizers = self.get_tokenizers(do_lower_case=False)
        for tokenizer in tokenizers:
            with self.subTest(f"{tokenizer.__class__.__name__}"):
                vocab = tokenizer.get_vocab()
                for word, ind in vocab.items():
1805
1806
                    if word == tokenizer.unk_token:
                        continue
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
                    self.assertEqual(tokenizer.convert_tokens_to_ids(word), ind)
                    self.assertEqual(tokenizer.convert_ids_to_tokens(ind), word)

    def test_call(self):
        # Tests that all call wrap to encode_plus and batch_encode_plus
        tokenizers = self.get_tokenizers(do_lower_case=False)
        for tokenizer in tokenizers:
            with self.subTest(f"{tokenizer.__class__.__name__}"):
                sequences = [
                    "Testing batch encode plus",
                    "Testing batch encode plus with different sequence lengths",
                    "Testing batch encode plus with different sequence lengths correctly pads",
                ]

                # Test not batched
                encoded_sequences_1 = tokenizer.encode_plus(sequences[0])
                encoded_sequences_2 = tokenizer(sequences[0])
                self.assertEqual(encoded_sequences_1, encoded_sequences_2)

                # Test not batched pairs
                encoded_sequences_1 = tokenizer.encode_plus(sequences[0], sequences[1])
                encoded_sequences_2 = tokenizer(sequences[0], sequences[1])
                self.assertEqual(encoded_sequences_1, encoded_sequences_2)

                # Test batched
                encoded_sequences_1 = tokenizer.batch_encode_plus(sequences)
                encoded_sequences_2 = tokenizer(sequences)
                self.assertEqual(encoded_sequences_1, encoded_sequences_2)

                # Test batched pairs
                encoded_sequences_1 = tokenizer.batch_encode_plus(list(zip(sequences, sequences)))
                encoded_sequences_2 = tokenizer(sequences, sequences)
                self.assertEqual(encoded_sequences_1, encoded_sequences_2)
1840
1841
1842

    def test_batch_encode_plus_batch_sequence_length(self):
        # Tests that all encoded values have the correct size
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
        tokenizers = self.get_tokenizers(do_lower_case=False)
        for tokenizer in tokenizers:
            with self.subTest(f"{tokenizer.__class__.__name__}"):
                sequences = [
                    "Testing batch encode plus",
                    "Testing batch encode plus with different sequence lengths",
                    "Testing batch encode plus with different sequence lengths correctly pads",
                ]

                encoded_sequences = [tokenizer.encode_plus(sequence) for sequence in sequences]
                encoded_sequences_batch = tokenizer.batch_encode_plus(sequences, padding=False)
                self.assertListEqual(
                    encoded_sequences, self.convert_batch_encode_plus_format_to_encode_plus(encoded_sequences_batch)
                )

                maximum_length = len(
                    max([encoded_sequence["input_ids"] for encoded_sequence in encoded_sequences], key=len)
                )

                # check correct behaviour if no pad_token_id exists and add it eventually
                self._check_no_pad_token_padding(tokenizer, sequences)

                encoded_sequences_padded = [
                    tokenizer.encode_plus(sequence, max_length=maximum_length, padding="max_length")
                    for sequence in sequences
                ]

                encoded_sequences_batch_padded = tokenizer.batch_encode_plus(sequences, padding=True)
                self.assertListEqual(
                    encoded_sequences_padded,
                    self.convert_batch_encode_plus_format_to_encode_plus(encoded_sequences_batch_padded),
                )

                # check 'longest' is unsensitive to a max length
                encoded_sequences_batch_padded_1 = tokenizer.batch_encode_plus(sequences, padding=True)
                encoded_sequences_batch_padded_2 = tokenizer.batch_encode_plus(
                    sequences, max_length=maximum_length + 10, padding="longest"
                )
                for key in encoded_sequences_batch_padded_1.keys():
                    self.assertListEqual(
Lysandre's avatar
Lysandre committed
1883
1884
                        encoded_sequences_batch_padded_1[key],
                        encoded_sequences_batch_padded_2[key],
1885
1886
1887
1888
1889
1890
1891
1892
1893
                    )

                # check 'no_padding' is unsensitive to a max length
                encoded_sequences_batch_padded_1 = tokenizer.batch_encode_plus(sequences, padding=False)
                encoded_sequences_batch_padded_2 = tokenizer.batch_encode_plus(
                    sequences, max_length=maximum_length + 10, padding=False
                )
                for key in encoded_sequences_batch_padded_1.keys():
                    self.assertListEqual(
Lysandre's avatar
Lysandre committed
1894
1895
                        encoded_sequences_batch_padded_1[key],
                        encoded_sequences_batch_padded_2[key],
1896
                    )
1897

1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
    @require_tokenizers
    def test_added_token_are_matched_longest_first(self):
        if not self.test_slow_tokenizer:
            self.skipTest("This test is only for slow tokenizers")
            return
        tokenizers = self.get_tokenizers(fast=False)
        for tokenizer in tokenizers:
            with self.subTest(f"{tokenizer.__class__.__name__}"):
                try:
                    tokenizer.add_tokens([AddedToken("extra_id_1")])
                    tokenizer.add_tokens([AddedToken("extra_id_100")])
                except Exception:
                    # Canine cannot add tokens which are not codepoints
                    self.skipTest("Cannot add those Added tokens")

                # XXX: This used to split on `extra_id_1` first we're matching
                # longest first now.
                tokens = tokenizer.tokenize("This is some extra_id_100")
                self.assertIn("extra_id_100", tokens)

        for tokenizer in tokenizers:
            with self.subTest(f"{tokenizer.__class__.__name__}"):
                tokenizer.add_tokens([AddedToken("extra_id_100")])
                tokenizer.add_tokens([AddedToken("extra_id_1")])

                tokens = tokenizer.tokenize("This is some extra_id_100")
                self.assertIn("extra_id_100", tokens)

1926
    @require_tokenizers
1927
1928
1929
    def test_added_token_serializable(self):
        tokenizers = self.get_tokenizers(do_lower_case=False)
        for tokenizer in tokenizers:
1930
1931
1932
            with self.subTest(f"{tokenizer.__class__.__name__}"):
                new_token = AddedToken("new_token", lstrip=True)
                tokenizer.add_special_tokens({"additional_special_tokens": [new_token]})
1933

1934
1935
1936
                with tempfile.TemporaryDirectory() as tmp_dir_name:
                    tokenizer.save_pretrained(tmp_dir_name)
                    tokenizer.from_pretrained(tmp_dir_name)
1937

1938
1939
1940
1941
    def test_batch_encode_plus_padding(self):
        # Test that padded sequences are equivalent between batch_encode_plus and encode_plus

        # Right padding tests
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
        tokenizers = self.get_tokenizers(do_lower_case=False)
        for tokenizer in tokenizers:
            with self.subTest(f"{tokenizer.__class__.__name__}"):
                sequences = [
                    "Testing batch encode plus",
                    "Testing batch encode plus with different sequence lengths",
                    "Testing batch encode plus with different sequence lengths correctly pads",
                ]

                max_length = 100

                # check correct behaviour if no pad_token_id exists and add it eventually
                self._check_no_pad_token_padding(tokenizer, sequences)

                encoded_sequences = [
                    tokenizer.encode_plus(sequence, max_length=max_length, padding="max_length")
                    for sequence in sequences
                ]
                encoded_sequences_batch = tokenizer.batch_encode_plus(
                    sequences, max_length=max_length, padding="max_length"
                )
                self.assertListEqual(
                    encoded_sequences, self.convert_batch_encode_plus_format_to_encode_plus(encoded_sequences_batch)
                )
1966
1967

        # Left padding tests
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
        tokenizers = self.get_tokenizers(do_lower_case=False)
        for tokenizer in tokenizers:
            with self.subTest(f"{tokenizer.__class__.__name__}"):
                tokenizer.padding_side = "left"
                sequences = [
                    "Testing batch encode plus",
                    "Testing batch encode plus with different sequence lengths",
                    "Testing batch encode plus with different sequence lengths correctly pads",
                ]

                max_length = 100

                # check correct behaviour if no pad_token_id exists and add it eventually
                self._check_no_pad_token_padding(tokenizer, sequences)

                encoded_sequences = [
                    tokenizer.encode_plus(sequence, max_length=max_length, padding="max_length")
                    for sequence in sequences
                ]
                encoded_sequences_batch = tokenizer.batch_encode_plus(
                    sequences, max_length=max_length, padding="max_length"
                )
                self.assertListEqual(
                    encoded_sequences, self.convert_batch_encode_plus_format_to_encode_plus(encoded_sequences_batch)
                )

    def test_pretokenized_inputs(self):
        # Test when inputs are pretokenized

1997
        tokenizers = self.get_tokenizers(do_lower_case=False)  # , add_prefix_space=True)
1998
1999
2000
        for tokenizer in tokenizers:
            with self.subTest(f"{tokenizer.__class__.__name__}"):

2001
2002
2003
                if hasattr(tokenizer, "add_prefix_space") and not tokenizer.add_prefix_space:
                    continue

2004
2005
2006
2007
2008
2009
2010
                # Prepare a sequence from our tokenizer vocabulary
                sequence, ids = self.get_clean_sequence(tokenizer, with_prefix_space=True, max_length=20)
                # sequence = " " + sequence  # To be sure the byte-level tokenizers are feeling good
                token_sequence = sequence.split()
                # sequence_no_prefix_space = sequence.strip()

                # Test encode for pretokenized inputs
2011
                output = tokenizer.encode(token_sequence, is_split_into_words=True, add_special_tokens=False)
2012
2013
2014
                output_sequence = tokenizer.encode(sequence, add_special_tokens=False)
                self.assertEqual(output, output_sequence)

2015
                output = tokenizer.encode(token_sequence, is_split_into_words=True, add_special_tokens=True)
2016
2017
2018
2019
                output_sequence = tokenizer.encode(sequence, add_special_tokens=True)
                self.assertEqual(output, output_sequence)

                # Test encode_plus for pretokenized inputs
2020
                output = tokenizer.encode_plus(token_sequence, is_split_into_words=True, add_special_tokens=False)
2021
2022
2023
                output_sequence = tokenizer.encode_plus(sequence, add_special_tokens=False)
                for key in output.keys():
                    self.assertEqual(output[key], output_sequence[key])
2024
                output = tokenizer.encode_plus(token_sequence, is_split_into_words=True, add_special_tokens=True)
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034
                output_sequence = tokenizer.encode_plus(sequence, add_special_tokens=True)
                for key in output.keys():
                    self.assertEqual(output[key], output_sequence[key])

                # Test batch_encode_plus for pretokenized inputs
                sequence_batch = [sequence.strip()] * 2 + [sequence.strip() + " " + sequence.strip()]
                token_sequence_batch = [s.split() for s in sequence_batch]
                sequence_batch_cleaned_up_spaces = [" " + " ".join(s) for s in token_sequence_batch]

                output = tokenizer.batch_encode_plus(
2035
                    token_sequence_batch, is_split_into_words=True, add_special_tokens=False
2036
2037
2038
2039
2040
2041
2042
                )
                output_sequence = tokenizer.batch_encode_plus(
                    sequence_batch_cleaned_up_spaces, add_special_tokens=False
                )
                for key in output.keys():
                    self.assertEqual(output[key], output_sequence[key])
                output = tokenizer.batch_encode_plus(
2043
                    token_sequence_batch, is_split_into_words=True, add_special_tokens=True
2044
2045
2046
2047
2048
2049
2050
2051
2052
                )
                output_sequence = tokenizer.batch_encode_plus(
                    sequence_batch_cleaned_up_spaces, add_special_tokens=True
                )
                for key in output.keys():
                    self.assertEqual(output[key], output_sequence[key])

                # Test encode for pretokenized inputs pairs
                output = tokenizer.encode(
2053
                    token_sequence, token_sequence, is_split_into_words=True, add_special_tokens=False
2054
2055
2056
2057
                )
                output_sequence = tokenizer.encode(sequence, sequence, add_special_tokens=False)
                self.assertEqual(output, output_sequence)
                output = tokenizer.encode(
2058
                    token_sequence, token_sequence, is_split_into_words=True, add_special_tokens=True
2059
2060
2061
2062
2063
2064
                )
                output_sequence = tokenizer.encode(sequence, sequence, add_special_tokens=True)
                self.assertEqual(output, output_sequence)

                # Test encode_plus for pretokenized inputs pairs
                output = tokenizer.encode_plus(
2065
                    token_sequence, token_sequence, is_split_into_words=True, add_special_tokens=False
2066
2067
2068
2069
2070
                )
                output_sequence = tokenizer.encode_plus(sequence, sequence, add_special_tokens=False)
                for key in output.keys():
                    self.assertEqual(output[key], output_sequence[key])
                output = tokenizer.encode_plus(
2071
                    token_sequence, token_sequence, is_split_into_words=True, add_special_tokens=True
2072
2073
2074
2075
2076
2077
2078
2079
2080
2081
2082
2083
2084
2085
2086
                )
                output_sequence = tokenizer.encode_plus(sequence, sequence, add_special_tokens=True)
                for key in output.keys():
                    self.assertEqual(output[key], output_sequence[key])

                # Test batch_encode_plus for pretokenized inputs pairs
                sequence_pair_batch = [(sequence.strip(), sequence.strip())] * 2 + [
                    (sequence.strip() + " " + sequence.strip(), sequence.strip())
                ]
                token_sequence_pair_batch = [tuple(s.split() for s in pair) for pair in sequence_pair_batch]
                sequence_pair_batch_cleaned_up_spaces = [
                    tuple(" " + " ".join(s) for s in pair) for pair in token_sequence_pair_batch
                ]

                output = tokenizer.batch_encode_plus(
2087
                    token_sequence_pair_batch, is_split_into_words=True, add_special_tokens=False
2088
2089
2090
2091
2092
2093
2094
                )
                output_sequence = tokenizer.batch_encode_plus(
                    sequence_pair_batch_cleaned_up_spaces, add_special_tokens=False
                )
                for key in output.keys():
                    self.assertEqual(output[key], output_sequence[key])
                output = tokenizer.batch_encode_plus(
2095
                    token_sequence_pair_batch, is_split_into_words=True, add_special_tokens=True
2096
2097
2098
2099
2100
2101
                )
                output_sequence = tokenizer.batch_encode_plus(
                    sequence_pair_batch_cleaned_up_spaces, add_special_tokens=True
                )
                for key in output.keys():
                    self.assertEqual(output[key], output_sequence[key])
2102

2103
2104
2105
    def test_prepare_for_model(self):
        tokenizers = self.get_tokenizers(do_lower_case=False)
        for tokenizer in tokenizers:
2106
2107
2108
2109
2110
2111
            with self.subTest(f"{tokenizer.__class__.__name__}"):
                string_sequence = "Testing the prepare_for_model method."
                ids = tokenizer.encode(string_sequence, add_special_tokens=False)
                prepared_input_dict = tokenizer.prepare_for_model(ids, add_special_tokens=True)

                input_dict = tokenizer.encode_plus(string_sequence, add_special_tokens=True)
2112

2113
                self.assertEqual(input_dict, prepared_input_dict)
2114

2115
2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
2126
    def test_batch_encode_plus_overflowing_tokens(self):
        tokenizers = self.get_tokenizers(do_lower_case=False)
        for tokenizer in tokenizers:
            string_sequences = ["Testing the prepare_for_model method.", "Test"]

            if tokenizer.pad_token is None:
                tokenizer.add_special_tokens({"pad_token": "[PAD]"})

            tokenizer.batch_encode_plus(
                string_sequences, return_overflowing_tokens=True, truncation=True, padding=True, max_length=3
            )

2127
    @is_pt_tf_cross_test
2128
    def test_batch_encode_plus_tensors(self):
2129
2130
2131
2132
2133
2134
2135
2136
2137
2138
2139
2140
2141
2142
2143
        tokenizers = self.get_tokenizers(do_lower_case=False)
        for tokenizer in tokenizers:
            with self.subTest(f"{tokenizer.__class__.__name__}"):
                sequences = [
                    "Testing batch encode plus",
                    "Testing batch encode plus with different sequence lengths",
                    "Testing batch encode plus with different sequence lengths correctly pads",
                ]

                # A Tensor cannot be build by sequences which are not the same size
                self.assertRaises(ValueError, tokenizer.batch_encode_plus, sequences, return_tensors="pt")
                self.assertRaises(ValueError, tokenizer.batch_encode_plus, sequences, return_tensors="tf")

                if tokenizer.pad_token_id is None:
                    self.assertRaises(
Lysandre's avatar
Lysandre committed
2144
2145
2146
2147
2148
                        ValueError,
                        tokenizer.batch_encode_plus,
                        sequences,
                        padding=True,
                        return_tensors="pt",
2149
2150
                    )
                    self.assertRaises(
Lysandre's avatar
Lysandre committed
2151
2152
2153
2154
2155
                        ValueError,
                        tokenizer.batch_encode_plus,
                        sequences,
                        padding="longest",
                        return_tensors="tf",
2156
2157
2158
2159
2160
                    )
                else:
                    pytorch_tensor = tokenizer.batch_encode_plus(sequences, padding=True, return_tensors="pt")
                    tensorflow_tensor = tokenizer.batch_encode_plus(sequences, padding="longest", return_tensors="tf")
                    encoded_sequences = tokenizer.batch_encode_plus(sequences, padding=True)
2161

2162
2163
2164
2165
                    for key in encoded_sequences.keys():
                        pytorch_value = pytorch_tensor[key].tolist()
                        tensorflow_value = tensorflow_tensor[key].numpy().tolist()
                        encoded_value = encoded_sequences[key]
2166

2167
                        self.assertEqual(pytorch_value, tensorflow_value, encoded_value)
2168
2169
2170
2171
2172
2173

    def _check_no_pad_token_padding(self, tokenizer, sequences):
        # if tokenizer does not have pad_token_id, an error should be thrown
        if tokenizer.pad_token_id is None:
            with self.assertRaises(ValueError):
                if isinstance(sequences, list):
2174
                    tokenizer.batch_encode_plus(sequences, padding="longest")
2175
                else:
2176
                    tokenizer.encode_plus(sequences, padding=True)
2177
2178
2179

            # add pad_token_id to pass subsequent tests
            tokenizer.add_special_tokens({"pad_token": "<PAD>"})
2180

2181
2182
2183
2184
2185
2186
2187
2188
2189
2190
2191
2192
2193
2194
2195
2196
2197
2198
2199
2200
2201
2202
2203
2204
2205
2206
2207
2208
2209
2210
2211
2212
2213
2214
2215
2216
2217
2218
2219
2220
    def check_subword_sampling(
        self,
        tokenizer: PreTrainedTokenizer,
        text: str = None,
    ) -> None:
        """
        Check if the tokenizer generates different results when subword regularization is enabled.

        Subword regularization augments training data with subword sampling.
        This has a random component.

        Args:
            tokenizer: The tokenizer to check.
            text: The text to use for the checks.
        """
        text = "This is a test for subword regularization." if text is None else text
        if self.test_sentencepiece_ignore_case:
            text = text.lower()

        tokens_list = []
        for _ in range(5):
            tokens_list.append(tokenizer.tokenize(text))

        # the list of different pairs of tokens_list
        combinations = itertools.combinations(tokens_list, 2)

        # check of sampling is done
        subword_sampling_found = False
        for combination in combinations:
            if combination[0] != combination[1]:
                subword_sampling_found = True
        self.assertTrue(subword_sampling_found)

        # check if converting back to original text works
        for tokens in tokens_list:
            if self.test_sentencepiece_ignore_case:
                self.assertEqual(text, tokenizer.convert_tokens_to_string(tokens).lower())
            else:
                self.assertEqual(text, tokenizer.convert_tokens_to_string(tokens))

2221
    @require_torch
Sylvain Gugger's avatar
Sylvain Gugger committed
2222
    @slow
2223
    def test_torch_encode_plus_sent_to_model(self):
2224
        import torch
2225

2226
2227
2228
2229
        from transformers import MODEL_MAPPING, TOKENIZER_MAPPING

        MODEL_TOKENIZER_MAPPING = merge_model_tokenizer_mappings(MODEL_MAPPING, TOKENIZER_MAPPING)

2230
2231
2232
        tokenizers = self.get_tokenizers(do_lower_case=False)
        for tokenizer in tokenizers:
            with self.subTest(f"{tokenizer.__class__.__name__}"):
2233

2234
2235
                if tokenizer.__class__ not in MODEL_TOKENIZER_MAPPING:
                    return
2236

2237
2238
                config_class, model_class = MODEL_TOKENIZER_MAPPING[tokenizer.__class__]
                config = config_class()
2239

2240
2241
                if config.is_encoder_decoder or config.pad_token_id is None:
                    return
2242

2243
                model = model_class(config)
2244

2245
2246
                # Make sure the model contains at least the full vocabulary size in its embedding matrix
                is_using_common_embeddings = hasattr(model.get_input_embeddings(), "weight")
Nicolas Patry's avatar
Nicolas Patry committed
2247
2248
                if is_using_common_embeddings:
                    self.assertGreaterEqual(model.get_input_embeddings().weight.shape[0], len(tokenizer))
2249

2250
2251
2252
2253
                # Build sequence
                first_ten_tokens = list(tokenizer.get_vocab().keys())[:10]
                sequence = " ".join(first_ten_tokens)
                encoded_sequence = tokenizer.encode_plus(sequence, return_tensors="pt")
2254
2255
2256
2257

                # Ensure that the BatchEncoding.to() method works.
                encoded_sequence.to(model.device)

2258
2259
                batch_encoded_sequence = tokenizer.batch_encode_plus([sequence, sequence], return_tensors="pt")
                # This should not fail
2260

2261
2262
2263
                with torch.no_grad():  # saves some time
                    model(**encoded_sequence)
                    model(**batch_encoded_sequence)
2264

2265
2266
2267
2268
2269
2270
2271
        # if self.test_rust_tokenizer:
        #     fast_tokenizer = self.get_rust_tokenizer()
        #     encoded_sequence_fast = fast_tokenizer.encode_plus(sequence, return_tensors="pt")
        #     batch_encoded_sequence_fast = fast_tokenizer.batch_encode_plus([sequence, sequence], return_tensors="pt")
        #     # This should not fail
        #     model(**encoded_sequence_fast)
        #     model(**batch_encoded_sequence_fast)
2272
2273

    @require_tf
Sylvain Gugger's avatar
Sylvain Gugger committed
2274
    @slow
2275
2276
2277
2278
2279
    def test_tf_encode_plus_sent_to_model(self):
        from transformers import TF_MODEL_MAPPING, TOKENIZER_MAPPING

        MODEL_TOKENIZER_MAPPING = merge_model_tokenizer_mappings(TF_MODEL_MAPPING, TOKENIZER_MAPPING)

2280
2281
2282
2283
2284
        tokenizers = self.get_tokenizers(do_lower_case=False)
        for tokenizer in tokenizers:
            with self.subTest(f"{tokenizer.__class__.__name__}"):
                if tokenizer.__class__ not in MODEL_TOKENIZER_MAPPING:
                    return
2285

2286
2287
                config_class, model_class = MODEL_TOKENIZER_MAPPING[tokenizer.__class__]
                config = config_class()
2288

2289
2290
                if config.is_encoder_decoder or config.pad_token_id is None:
                    return
2291

2292
                model = model_class(config)
2293

2294
                # Make sure the model contains at least the full vocabulary size in its embedding matrix
Nicolas Patry's avatar
Nicolas Patry committed
2295
                self.assertGreaterEqual(model.config.vocab_size, len(tokenizer))
2296

2297
2298
2299
2300
2301
                # Build sequence
                first_ten_tokens = list(tokenizer.get_vocab().keys())[:10]
                sequence = " ".join(first_ten_tokens)
                encoded_sequence = tokenizer.encode_plus(sequence, return_tensors="tf")
                batch_encoded_sequence = tokenizer.batch_encode_plus([sequence, sequence], return_tensors="tf")
2302

2303
2304
2305
                # This should not fail
                model(encoded_sequence)
                model(batch_encoded_sequence)
2306
2307
2308

    # TODO: Check if require_torch is the best to test for numpy here ... Maybe move to require_flax when available
    @require_torch
Sylvain Gugger's avatar
Sylvain Gugger committed
2309
    @slow
2310
2311
2312
2313
2314
    def test_np_encode_plus_sent_to_model(self):
        from transformers import MODEL_MAPPING, TOKENIZER_MAPPING

        MODEL_TOKENIZER_MAPPING = merge_model_tokenizer_mappings(MODEL_MAPPING, TOKENIZER_MAPPING)

2315
2316
2317
2318
2319
        tokenizers = self.get_tokenizers()
        for tokenizer in tokenizers:
            with self.subTest(f"{tokenizer.__class__.__name__}"):
                if tokenizer.__class__ not in MODEL_TOKENIZER_MAPPING:
                    return
2320

2321
2322
                config_class, model_class = MODEL_TOKENIZER_MAPPING[tokenizer.__class__]
                config = config_class()
2323

2324
2325
                if config.is_encoder_decoder or config.pad_token_id is None:
                    return
2326

2327
2328
2329
2330
2331
2332
2333
2334
2335
2336
2337
2338
2339
2340
2341
2342
2343
2344
2345
2346
                # Build sequence
                first_ten_tokens = list(tokenizer.get_vocab().keys())[:10]
                sequence = " ".join(first_ten_tokens)
                encoded_sequence = tokenizer.encode_plus(sequence, return_tensors="np")
                batch_encoded_sequence = tokenizer.batch_encode_plus([sequence, sequence], return_tensors="np")

                # TODO: add forward through JAX/Flax when PR is merged
                # This is currently here to make flake8 happy !
                if encoded_sequence is None:
                    raise ValueError("Cannot convert list to numpy tensor on  encode_plus()")

                if batch_encoded_sequence is None:
                    raise ValueError("Cannot convert list to numpy tensor on  batch_encode_plus()")

                if self.test_rust_tokenizer:
                    fast_tokenizer = self.get_rust_tokenizer()
                    encoded_sequence_fast = fast_tokenizer.encode_plus(sequence, return_tensors="np")
                    batch_encoded_sequence_fast = fast_tokenizer.batch_encode_plus(
                        [sequence, sequence], return_tensors="np"
                    )
2347

2348
2349
2350
2351
                    # TODO: add forward through JAX/Flax when PR is merged
                    # This is currently here to make flake8 happy !
                    if encoded_sequence_fast is None:
                        raise ValueError("Cannot convert list to numpy tensor on  encode_plus() (fast)")
2352

2353
2354
                    if batch_encoded_sequence_fast is None:
                        raise ValueError("Cannot convert list to numpy tensor on  batch_encode_plus() (fast)")
2355
2356
2357

    @require_torch
    def test_prepare_seq2seq_batch(self):
2358
2359
2360
        if not self.test_seq2seq:
            return

2361
2362
2363
2364
2365
2366
2367
2368
2369
2370
2371
2372
2373
2374
2375
2376
2377
2378
2379
2380
2381
2382
2383
2384
2385
2386
2387
2388
2389
2390
2391
2392
2393
        tokenizers = self.get_tokenizers()
        for tokenizer in tokenizers:
            with self.subTest(f"{tokenizer.__class__.__name__}"):
                # Longer text that will definitely require truncation.
                src_text = [
                    " UN Chief Says There Is No Military Solution in Syria",
                    " Secretary-General Ban Ki-moon says his response to Russia's stepped up military support for Syria is that 'there is no military solution' to the nearly five-year conflict and more weapons will only worsen the violence and misery for millions of people.",
                ]
                tgt_text = [
                    "Şeful ONU declară că nu există o soluţie militară în Siria",
                    "Secretarul General Ban Ki-moon declară că răspunsul său la intensificarea sprijinului militar al Rusiei "
                    'pentru Siria este că "nu există o soluţie militară" la conflictul de aproape cinci ani şi că noi arme nu '
                    "vor face decât să înrăutăţească violenţele şi mizeria pentru milioane de oameni.",
                ]
                try:
                    batch = tokenizer.prepare_seq2seq_batch(
                        src_texts=src_text,
                        tgt_texts=tgt_text,
                        max_length=3,
                        max_target_length=10,
                        return_tensors="pt",
                        src_lang="en_XX",  # this should be ignored (for all but mbart) but not cause an error
                    )
                except NotImplementedError:
                    return
                self.assertEqual(batch.input_ids.shape[1], 3)
                self.assertEqual(batch.labels.shape[1], 10)
                # max_target_length will default to max_length if not specified
                batch = tokenizer.prepare_seq2seq_batch(
                    src_text, tgt_texts=tgt_text, max_length=3, return_tensors="pt"
                )
                self.assertEqual(batch.input_ids.shape[1], 3)
                self.assertEqual(batch.labels.shape[1], 3)
2394

2395
2396
2397
2398
2399
2400
                batch_encoder_only = tokenizer.prepare_seq2seq_batch(
                    src_texts=src_text, max_length=3, max_target_length=10, return_tensors="pt"
                )
                self.assertEqual(batch_encoder_only.input_ids.shape[1], 3)
                self.assertEqual(batch_encoder_only.attention_mask.shape[1], 3)
                self.assertNotIn("decoder_input_ids", batch_encoder_only)
2401
2402
2403

    def test_is_fast(self):
        for tokenizer, pretrained_name, kwargs in self.tokenizers_list:
2404
            with self.subTest(f"{tokenizer.__class__.__name__} ({pretrained_name})"):
2405
2406
2407
2408
                tokenizer_r = self.rust_tokenizer_class.from_pretrained(pretrained_name, **kwargs)
                # Check is_fast is set correctly
                self.assertTrue(tokenizer_r.is_fast)

2409
2410
2411
2412
                if self.test_slow_tokenizer:
                    tokenizer_p = self.tokenizer_class.from_pretrained(pretrained_name, **kwargs)
                    self.assertFalse(tokenizer_p.is_fast)

2413
2414
    def test_fast_only_inputs(self):
        for tokenizer, pretrained_name, kwargs in self.tokenizers_list:
2415
            with self.subTest(f"{tokenizer.__class__.__name__} ({pretrained_name})"):
2416
2417
2418
2419
2420
2421
2422
2423
2424
2425
                tokenizer_r = self.rust_tokenizer_class.from_pretrained(pretrained_name, **kwargs)

                # Ensure None raise an error
                self.assertRaises(TypeError, tokenizer_r.tokenize, None)
                self.assertRaises(TypeError, tokenizer_r.encode, None)
                self.assertRaises(TypeError, tokenizer_r.encode_plus, None)
                self.assertRaises(TypeError, tokenizer_r.batch_encode_plus, None)

    def test_alignement_methods(self):
        for tokenizer, pretrained_name, kwargs in self.tokenizers_list:
2426
            with self.subTest(f"{tokenizer.__class__.__name__} ({pretrained_name})"):
2427
2428
2429
2430
2431
2432
2433
2434
2435
2436
2437
2438
2439
2440
2441
2442
2443
2444
2445
2446
2447
2448
2449
2450
2451
2452
2453
2454
2455
2456
2457
2458
2459
2460
2461
2462
2463
2464
2465
2466
2467
2468
2469
2470
2471
2472
2473
2474
2475
2476
2477
2478
2479
2480
2481
2482
2483
2484
2485
2486
2487
2488
2489
2490
2491
2492
2493
2494
2495
2496
2497
2498
2499
2500
2501
2502
2503
2504
2505
2506
2507
2508
2509
2510
2511
                tokenizer_r = self.rust_tokenizer_class.from_pretrained(pretrained_name, **kwargs)

                words = ["Wonderful", "no", "inspiration", "example", "with", "subtoken"]
                text = " ".join(words)
                batch_size = 3

                encoding = tokenizer_r.encode_plus(text, add_special_tokens=False)

                batch_encoding = tokenizer_r.batch_encode_plus([text] * batch_size, add_special_tokens=False)
                num_tokens = len(encoding["input_ids"])

                last_word_index = len(words) - 1
                last_token_index = num_tokens - 1
                last_batch_index = batch_size - 1
                last_char_index = len(text) - 1

                # words, tokens
                self.assertEqual(len(encoding.words(0)), num_tokens)
                self.assertEqual(max(encoding.words(0)), last_word_index)
                self.assertEqual(min(encoding.words(0)), 0)
                self.assertEqual(len(batch_encoding.words(last_batch_index)), num_tokens)
                self.assertEqual(max(batch_encoding.words(last_batch_index)), last_word_index)
                self.assertEqual(min(batch_encoding.words(last_batch_index)), 0)
                self.assertEqual(len(encoding.tokens(0)), num_tokens)

                # Assert token_to_word
                self.assertEqual(encoding.token_to_word(0), 0)
                self.assertEqual(encoding.token_to_word(0, 0), 0)
                self.assertEqual(encoding.token_to_word(last_token_index), last_word_index)
                self.assertEqual(encoding.token_to_word(0, last_token_index), last_word_index)
                self.assertEqual(batch_encoding.token_to_word(1, 0), 0)
                self.assertEqual(batch_encoding.token_to_word(0, last_token_index), last_word_index)
                self.assertEqual(batch_encoding.token_to_word(last_batch_index, last_token_index), last_word_index)

                # Assert word_to_tokens
                self.assertEqual(encoding.word_to_tokens(0).start, 0)
                self.assertEqual(encoding.word_to_tokens(0, 0).start, 0)
                self.assertEqual(encoding.word_to_tokens(last_word_index).end, last_token_index + 1)
                self.assertEqual(encoding.word_to_tokens(0, last_word_index).end, last_token_index + 1)
                self.assertEqual(batch_encoding.word_to_tokens(1, 0).start, 0)
                self.assertEqual(batch_encoding.word_to_tokens(0, last_word_index).end, last_token_index + 1)
                self.assertEqual(
                    batch_encoding.word_to_tokens(last_batch_index, last_word_index).end, last_token_index + 1
                )

                # Assert token_to_chars
                self.assertEqual(encoding.token_to_chars(0).start, 0)
                self.assertEqual(encoding.token_to_chars(0, 0).start, 0)
                self.assertEqual(encoding.token_to_chars(last_token_index).end, last_char_index + 1)
                self.assertEqual(encoding.token_to_chars(0, last_token_index).end, last_char_index + 1)
                self.assertEqual(batch_encoding.token_to_chars(1, 0).start, 0)
                self.assertEqual(batch_encoding.token_to_chars(0, last_token_index).end, last_char_index + 1)
                self.assertEqual(
                    batch_encoding.token_to_chars(last_batch_index, last_token_index).end, last_char_index + 1
                )

                # Assert char_to_token
                self.assertEqual(encoding.char_to_token(0), 0)
                self.assertEqual(encoding.char_to_token(0, 0), 0)
                self.assertEqual(encoding.char_to_token(last_char_index), last_token_index)
                self.assertEqual(encoding.char_to_token(0, last_char_index), last_token_index)
                self.assertEqual(batch_encoding.char_to_token(1, 0), 0)
                self.assertEqual(batch_encoding.char_to_token(0, last_char_index), last_token_index)
                self.assertEqual(batch_encoding.char_to_token(last_batch_index, last_char_index), last_token_index)

                # Assert char_to_word
                self.assertEqual(encoding.char_to_word(0), 0)
                self.assertEqual(encoding.char_to_word(0, 0), 0)
                self.assertEqual(encoding.char_to_word(last_char_index), last_word_index)
                self.assertEqual(encoding.char_to_word(0, last_char_index), last_word_index)
                self.assertEqual(batch_encoding.char_to_word(1, 0), 0)
                self.assertEqual(batch_encoding.char_to_word(0, last_char_index), last_word_index)
                self.assertEqual(batch_encoding.char_to_word(last_batch_index, last_char_index), last_word_index)

                # Assert word_to_chars
                self.assertEqual(encoding.word_to_chars(0).start, 0)
                self.assertEqual(encoding.word_to_chars(0, 0).start, 0)
                self.assertEqual(encoding.word_to_chars(last_word_index).end, last_char_index + 1)
                self.assertEqual(encoding.word_to_chars(0, last_word_index).end, last_char_index + 1)
                self.assertEqual(batch_encoding.word_to_chars(1, 0).start, 0)
                self.assertEqual(batch_encoding.word_to_chars(0, last_word_index).end, last_char_index + 1)
                self.assertEqual(
                    batch_encoding.word_to_chars(last_batch_index, last_word_index).end, last_char_index + 1
                )

2512
2513
2514
2515
2516
2517
2518
2519
2520
2521
2522
2523
2524
2525
2526
2527
2528
2529
2530
2531
2532
2533
2534
2535
2536
2537
2538
2539
2540
2541
2542
2543
2544
2545
2546
2547
2548
2549
2550
2551
2552
2553
2554
2555
2556
2557
2558
2559
2560
2561
2562
2563
2564
2565
2566
2567
2568
2569
2570
2571
2572
2573
2574
2575
2576
2577
2578
2579
2580
2581
2582
2583
2584
2585
2586
2587
2588
2589
2590
2591
2592
2593
2594
2595
2596
2597
2598
2599
2600
2601
2602
2603
2604
2605
2606
2607
2608
2609
2610
2611
2612
2613
2614
2615
2616
2617
2618
2619
2620
2621
2622
2623
2624
2625
2626
2627
2628
2629
2630
2631
2632
2633
2634
2635
2636
2637
2638
2639
2640
2641
2642
2643
2644
2645
2646
2647
2648
2649
                # Assert token_to_sequence
                self.assertEqual(encoding.token_to_sequence(num_tokens // 2), 0)
                self.assertEqual(encoding.token_to_sequence(0, num_tokens // 2), 0)
                self.assertEqual(batch_encoding.token_to_sequence(1, num_tokens // 2), 0)
                self.assertEqual(batch_encoding.token_to_sequence(0, num_tokens // 2), 0)
                self.assertEqual(batch_encoding.token_to_sequence(last_batch_index, num_tokens // 2), 0)

                # Pair of input sequences

                words = ["Wonderful", "no", "inspiration", "example", "with", "subtoken"]
                text = " ".join(words)
                pair_words = ["Amazing", "example", "full", "of", "inspiration"]
                pair_text = " ".join(pair_words)
                batch_size = 3
                index_word_in_first_seq = words.index("inspiration")
                index_word_in_pair_seq = pair_words.index("inspiration")
                index_char_in_first_seq = text.find("inspiration")
                index_char_in_pair_seq = pair_text.find("inspiration")

                pair_encoding = tokenizer_r.encode_plus(text, pair_text, add_special_tokens=False)

                pair_batch_encoding = tokenizer_r.batch_encode_plus(
                    [(text, pair_text)] * batch_size, add_special_tokens=False
                )
                num_tokens = len(encoding["input_ids"])

                last_word_index = len(words) - 1
                last_token_index = num_tokens - 1
                last_batch_index = batch_size - 1
                last_char_index = len(text) - 1

                # Assert word_to_tokens
                self.assertNotEqual(
                    pair_encoding.word_to_tokens(index_word_in_first_seq, sequence_index=0).start,
                    pair_encoding.word_to_tokens(index_word_in_pair_seq, sequence_index=1).start,
                )
                self.assertEqual(
                    pair_encoding["input_ids"][
                        pair_encoding.word_to_tokens(index_word_in_first_seq, sequence_index=0).start
                    ],
                    pair_encoding["input_ids"][
                        pair_encoding.word_to_tokens(index_word_in_pair_seq, sequence_index=1).start
                    ],
                )
                self.assertNotEqual(
                    pair_batch_encoding.word_to_tokens(1, index_word_in_first_seq, sequence_index=0).start,
                    pair_batch_encoding.word_to_tokens(1, index_word_in_pair_seq, sequence_index=1).start,
                )
                self.assertEqual(
                    pair_batch_encoding["input_ids"][1][
                        pair_batch_encoding.word_to_tokens(1, index_word_in_first_seq, sequence_index=0).start
                    ],
                    pair_batch_encoding["input_ids"][1][
                        pair_batch_encoding.word_to_tokens(1, index_word_in_pair_seq, sequence_index=1).start
                    ],
                )

                # Assert char_to_token
                self.assertNotEqual(
                    pair_encoding.char_to_token(index_char_in_first_seq, sequence_index=0),
                    pair_encoding.char_to_token(index_char_in_pair_seq, sequence_index=1),
                )
                self.assertEqual(
                    pair_encoding["input_ids"][pair_encoding.char_to_token(index_char_in_first_seq, sequence_index=0)],
                    pair_encoding["input_ids"][pair_encoding.char_to_token(index_char_in_pair_seq, sequence_index=1)],
                )
                self.assertNotEqual(
                    pair_batch_encoding.char_to_token(1, index_char_in_first_seq, sequence_index=0),
                    pair_batch_encoding.char_to_token(1, index_char_in_pair_seq, sequence_index=1),
                )
                self.assertEqual(
                    pair_batch_encoding["input_ids"][1][
                        pair_batch_encoding.char_to_token(1, index_char_in_first_seq, sequence_index=0)
                    ],
                    pair_batch_encoding["input_ids"][1][
                        pair_batch_encoding.char_to_token(1, index_char_in_pair_seq, sequence_index=1)
                    ],
                )

                # Assert char_to_word
                self.assertNotEqual(
                    pair_encoding.char_to_word(index_char_in_first_seq, sequence_index=0),
                    pair_encoding.char_to_word(index_char_in_pair_seq, sequence_index=1),
                )
                self.assertEqual(
                    words[pair_encoding.char_to_word(index_char_in_first_seq, sequence_index=0)],
                    pair_words[pair_encoding.char_to_word(index_char_in_pair_seq, sequence_index=1)],
                )
                self.assertNotEqual(
                    pair_batch_encoding.char_to_word(1, index_char_in_first_seq, sequence_index=0),
                    pair_batch_encoding.char_to_word(1, index_char_in_pair_seq, sequence_index=1),
                )
                self.assertEqual(
                    words[pair_batch_encoding.char_to_word(1, index_char_in_first_seq, sequence_index=0)],
                    pair_words[pair_batch_encoding.char_to_word(1, index_char_in_pair_seq, sequence_index=1)],
                )

                # Assert word_to_chars
                self.assertNotEqual(
                    pair_encoding.word_to_chars(index_word_in_first_seq, sequence_index=0).start,
                    pair_encoding.word_to_chars(index_word_in_pair_seq, sequence_index=1).start,
                )
                self.assertEqual(
                    text[pair_encoding.word_to_chars(index_word_in_first_seq, sequence_index=0).start],
                    pair_text[pair_encoding.word_to_chars(index_word_in_pair_seq, sequence_index=1).start],
                )
                self.assertNotEqual(
                    pair_batch_encoding.word_to_chars(1, index_word_in_first_seq, sequence_index=0).start,
                    pair_batch_encoding.word_to_chars(1, index_word_in_pair_seq, sequence_index=1).start,
                )
                self.assertEqual(
                    text[pair_batch_encoding.word_to_chars(1, index_word_in_first_seq, sequence_index=0).start],
                    pair_text[pair_batch_encoding.word_to_chars(1, index_word_in_pair_seq, sequence_index=1).start],
                )

                # Assert token_to_sequence
                pair_encoding = tokenizer_r.encode_plus(text, pair_text, add_special_tokens=True)

                pair_sequence_ids = [
                    pair_encoding.token_to_sequence(i) for i in range(len(pair_encoding["input_ids"]))
                ]
                self.assertIn(0, pair_sequence_ids)
                self.assertIn(1, pair_sequence_ids)
                if tokenizer_r.num_special_tokens_to_add(pair=True):
                    self.assertIn(None, pair_sequence_ids)

                pair_batch_encoding = tokenizer_r.batch_encode_plus(
                    [(text, pair_text)] * batch_size, add_special_tokens=True
                )
                pair_batch_sequence_ids = [
                    pair_batch_encoding.token_to_sequence(1, i)
                    for i in range(len(pair_batch_encoding["input_ids"][0]))
                ]
                self.assertIn(0, pair_batch_sequence_ids)
                self.assertIn(1, pair_batch_sequence_ids)
                if tokenizer_r.num_special_tokens_to_add(pair=True):
                    self.assertIn(None, pair_batch_sequence_ids)

2650
    def test_tokenization_python_rust_equals(self):
2651
2652
2653
2654
        if not self.test_slow_tokenizer:
            # as we don't have a slow version, we can't compare the outputs between slow and fast versions
            return

2655
        for tokenizer, pretrained_name, kwargs in self.tokenizers_list:
2656
            with self.subTest(f"{tokenizer.__class__.__name__} ({pretrained_name})"):
2657
2658
2659
2660
2661
2662
2663
2664
2665
2666
2667
2668
2669
2670
2671
2672
2673
2674
2675
2676
2677
2678
2679
2680
2681
2682
2683
2684
2685
2686
2687
2688
2689
2690
2691
                tokenizer_r = self.rust_tokenizer_class.from_pretrained(pretrained_name, **kwargs)
                tokenizer_p = self.tokenizer_class.from_pretrained(pretrained_name, **kwargs)

                # Ensure basic input match
                input_p = tokenizer_p.encode_plus(self._data)
                input_r = tokenizer_r.encode_plus(self._data)

                for key in filter(lambda x: x in ["input_ids", "token_type_ids", "attention_mask"], input_p.keys()):
                    self.assertSequenceEqual(input_p[key], input_r[key])

                input_pairs_p = tokenizer_p.encode_plus(self._data, self._data)
                input_pairs_r = tokenizer_r.encode_plus(self._data, self._data)

                for key in filter(lambda x: x in ["input_ids", "token_type_ids", "attention_mask"], input_p.keys()):
                    self.assertSequenceEqual(input_pairs_p[key], input_pairs_r[key])

                # Ensure truncation match
                input_p = tokenizer_p.encode_plus(self._data, max_length=512, truncation=True)
                input_r = tokenizer_r.encode_plus(self._data, max_length=512, truncation=True)

                for key in filter(lambda x: x in ["input_ids", "token_type_ids", "attention_mask"], input_p.keys()):
                    self.assertSequenceEqual(input_p[key], input_r[key])

                # Ensure truncation with stride match
                input_p = tokenizer_p.encode_plus(
                    self._data, max_length=512, truncation=True, stride=3, return_overflowing_tokens=True
                )
                input_r = tokenizer_r.encode_plus(
                    self._data, max_length=512, truncation=True, stride=3, return_overflowing_tokens=True
                )

                for key in filter(lambda x: x in ["input_ids", "token_type_ids", "attention_mask"], input_p.keys()):
                    self.assertSequenceEqual(input_p[key], input_r[key][0])

    def test_num_special_tokens_to_add_equal(self):
2692
2693
2694
2695
        if not self.test_slow_tokenizer:
            # as we don't have a slow version, we can't compare the outputs between slow and fast versions
            return

2696
        for tokenizer, pretrained_name, kwargs in self.tokenizers_list:
2697
            with self.subTest(f"{tokenizer.__class__.__name__} ({pretrained_name})"):
2698
2699
2700
2701
2702
2703
2704
2705
2706
2707
2708
2709
                tokenizer_r = self.rust_tokenizer_class.from_pretrained(pretrained_name, **kwargs)
                tokenizer_p = self.tokenizer_class.from_pretrained(pretrained_name, **kwargs)

                # Check we have the same number of added_tokens for both pair and non-pair inputs.
                self.assertEqual(
                    tokenizer_r.num_special_tokens_to_add(False), tokenizer_p.num_special_tokens_to_add(False)
                )
                self.assertEqual(
                    tokenizer_r.num_special_tokens_to_add(True), tokenizer_p.num_special_tokens_to_add(True)
                )

    def test_max_length_equal(self):
2710
2711
2712
2713
        if not self.test_slow_tokenizer:
            # as we don't have a slow version, we can't compare the outputs between slow and fast versions
            return

2714
        for tokenizer, pretrained_name, kwargs in self.tokenizers_list:
2715
            with self.subTest(f"{tokenizer.__class__.__name__} ({pretrained_name})"):
2716
2717
2718
2719
2720
2721
2722
2723
                tokenizer_r = self.rust_tokenizer_class.from_pretrained(pretrained_name, **kwargs)
                tokenizer_p = self.tokenizer_class.from_pretrained(pretrained_name, **kwargs)

                # Check we have the correct max_length for both pair and non-pair inputs.
                self.assertEqual(tokenizer_r.max_len_single_sentence, tokenizer_p.max_len_single_sentence)
                self.assertEqual(tokenizer_r.max_len_sentences_pair, tokenizer_p.max_len_sentences_pair)

    def test_special_tokens_map_equal(self):
2724
2725
2726
2727
        if not self.test_slow_tokenizer:
            # as we don't have a slow version, we can't compare the outputs between slow and fast versions
            return

2728
        for tokenizer, pretrained_name, kwargs in self.tokenizers_list:
2729
            with self.subTest(f"{tokenizer.__class__.__name__} ({pretrained_name})"):
2730
2731
2732
2733
2734
2735
2736
2737
2738
2739
2740
                tokenizer_r = self.rust_tokenizer_class.from_pretrained(pretrained_name, **kwargs)
                tokenizer_p = self.tokenizer_class.from_pretrained(pretrained_name, **kwargs)

                # Assert the set of special tokens match.
                self.assertSequenceEqual(
                    tokenizer_p.special_tokens_map.items(),
                    tokenizer_r.special_tokens_map.items(),
                )

    def test_add_tokens(self):
        for tokenizer, pretrained_name, kwargs in self.tokenizers_list:
2741
            with self.subTest(f"{tokenizer.__class__.__name__} ({pretrained_name})"):
2742
2743
2744
2745
2746
2747
2748
2749
2750
2751
2752
2753
2754
2755
2756
2757
2758
                tokenizer_r = self.rust_tokenizer_class.from_pretrained(pretrained_name, **kwargs)

                vocab_size = len(tokenizer_r)
                self.assertEqual(tokenizer_r.add_tokens(""), 0)
                self.assertEqual(tokenizer_r.add_tokens("testoken"), 1)
                self.assertEqual(tokenizer_r.add_tokens(["testoken1", "testtoken2"]), 2)
                self.assertEqual(len(tokenizer_r), vocab_size + 3)

                self.assertEqual(tokenizer_r.add_special_tokens({}), 0)
                self.assertEqual(tokenizer_r.add_special_tokens({"bos_token": "[BOS]", "eos_token": "[EOS]"}), 2)
                self.assertRaises(
                    AssertionError, tokenizer_r.add_special_tokens, {"additional_special_tokens": "<testtoken1>"}
                )
                self.assertEqual(tokenizer_r.add_special_tokens({"additional_special_tokens": ["<testtoken2>"]}), 1)
                self.assertEqual(
                    tokenizer_r.add_special_tokens({"additional_special_tokens": ["<testtoken3>", "<testtoken4>"]}), 2
                )
2759
2760
2761
2762
                self.assertIn("<testtoken3>", tokenizer_r.special_tokens_map["additional_special_tokens"])
                self.assertIsInstance(tokenizer_r.special_tokens_map["additional_special_tokens"], list)
                self.assertGreaterEqual(len(tokenizer_r.special_tokens_map["additional_special_tokens"]), 2)

2763
2764
2765
2766
                self.assertEqual(len(tokenizer_r), vocab_size + 8)

    def test_offsets_mapping(self):
        for tokenizer, pretrained_name, kwargs in self.tokenizers_list:
2767
            with self.subTest(f"{tokenizer.__class__.__name__} ({pretrained_name})"):
2768
2769
2770
2771
2772
2773
2774
2775
2776
2777
2778
2779
2780
2781
2782
2783
2784
2785
2786
2787
2788
2789
2790
2791
2792
2793
2794
2795
2796
2797
2798
2799
2800
2801
2802
2803
2804
2805
2806
2807
2808
2809
2810
2811
2812
                tokenizer_r = self.rust_tokenizer_class.from_pretrained(pretrained_name, **kwargs)

                text = "Wonderful no inspiration example with subtoken"
                pair = "Along with an awesome pair"

                # No pair
                tokens_with_offsets = tokenizer_r.encode_plus(
                    text, return_special_tokens_mask=True, return_offsets_mapping=True, add_special_tokens=True
                )
                added_tokens = tokenizer_r.num_special_tokens_to_add(False)
                offsets = tokens_with_offsets["offset_mapping"]

                # Assert there is the same number of tokens and offsets
                self.assertEqual(len(offsets), len(tokens_with_offsets["input_ids"]))

                # Assert there is online added_tokens special_tokens
                self.assertEqual(sum(tokens_with_offsets["special_tokens_mask"]), added_tokens)

                # Pairs
                tokens_with_offsets = tokenizer_r.encode_plus(
                    text, pair, return_special_tokens_mask=True, return_offsets_mapping=True, add_special_tokens=True
                )
                added_tokens = tokenizer_r.num_special_tokens_to_add(True)
                offsets = tokens_with_offsets["offset_mapping"]

                # Assert there is the same number of tokens and offsets
                self.assertEqual(len(offsets), len(tokens_with_offsets["input_ids"]))

                # Assert there is online added_tokens special_tokens
                self.assertEqual(sum(tokens_with_offsets["special_tokens_mask"]), added_tokens)

    def test_batch_encode_dynamic_overflowing(self):
        """
        When calling batch_encode with multiple sequence it can returns different number of
        overflowing encoding for each sequence:
        [
          Sequence 1: [Encoding 1, Encoding 2],
          Sequence 2: [Encoding 1],
          Sequence 3: [Encoding 1, Encoding 2, ... Encoding N]
        ]
        This needs to be padded so that it can represented as a tensor
        """
        for tokenizer, pretrained_name, kwargs in self.tokenizers_list:
            tokenizer = self.rust_tokenizer_class.from_pretrained(pretrained_name, **kwargs)

2813
            with self.subTest(f"{tokenizer.__class__.__name__} ({pretrained_name}, {tokenizer.__class__.__name__})"):
2814

2815
2816
2817
2818
2819
2820
                if is_torch_available():
                    returned_tensor = "pt"
                elif is_tf_available():
                    returned_tensor = "tf"
                else:
                    returned_tensor = "jax"
2821
2822
2823
2824
2825
2826
2827
2828
2829
2830
2831
2832
2833
2834
2835
2836
2837
2838
2839
2840
2841
2842
2843
2844
2845
2846
2847
2848
2849
2850
2851
2852
2853
2854
2855
2856
2857
2858
2859
2860
2861
2862
2863
2864
2865

                if not tokenizer.pad_token or tokenizer.pad_token_id < 0:
                    return

                tokens = tokenizer.encode_plus(
                    "HuggingFace is solving NLP one commit at a time",
                    max_length=6,
                    padding=True,
                    truncation=True,
                    return_tensors=returned_tensor,
                    return_overflowing_tokens=True,
                )

                for key in filter(lambda x: "overflow_to_sample_mapping" not in x, tokens.keys()):
                    self.assertEqual(len(tokens[key].shape), 2)

                # Mono sample
                tokens = tokenizer.batch_encode_plus(
                    ["HuggingFace is solving NLP one commit at a time"],
                    max_length=6,
                    padding=True,
                    truncation="only_first",
                    return_tensors=returned_tensor,
                    return_overflowing_tokens=True,
                )

                for key in filter(lambda x: "overflow_to_sample_mapping" not in x, tokens.keys()):
                    self.assertEqual(len(tokens[key].shape), 2)
                    self.assertEqual(tokens[key].shape[-1], 6)

                # Multi sample
                tokens = tokenizer.batch_encode_plus(
                    ["HuggingFace is solving NLP one commit at a time", "Very tiny input"],
                    max_length=6,
                    padding=True,
                    truncation="only_first",
                    return_tensors=returned_tensor,
                    return_overflowing_tokens=True,
                )

                for key in filter(lambda x: "overflow_to_sample_mapping" not in x, tokens.keys()):
                    self.assertEqual(len(tokens[key].shape), 2)
                    self.assertEqual(tokens[key].shape[-1], 6)

    def test_compare_pretokenized_inputs(self):
2866
2867
2868
2869
        if not self.test_slow_tokenizer:
            # as we don't have a slow version, we can't compare the outputs between slow and fast versions
            return

2870
        for tokenizer, pretrained_name, kwargs in self.tokenizers_list:
2871
            with self.subTest(f"{tokenizer.__class__.__name__} ({pretrained_name})"):
2872
2873
2874
2875
2876
2877
2878
2879
2880
2881
2882
2883
2884
2885
2886
2887
2888
2889
2890
2891
2892
2893
2894
2895
2896
2897
2898
2899
2900
2901
2902
2903
2904
2905
2906
2907
2908
2909
2910
2911
2912
2913
2914
2915
2916
2917
2918
2919
2920
2921
2922
2923
2924
2925
2926
2927
2928
2929
2930
2931
2932
2933
2934
2935
2936
2937
2938
2939
2940
2941
2942
2943
2944
2945
2946
2947
                tokenizer_r = self.rust_tokenizer_class.from_pretrained(pretrained_name, **kwargs)
                tokenizer_p = self.tokenizer_class.from_pretrained(pretrained_name, **kwargs)

                if hasattr(tokenizer_p, "add_prefix_space") and not tokenizer_p.add_prefix_space:
                    continue  # Too hard to test for now

                # Input string
                pretokenized_input_simple = "This is a sample input".split()
                pretokenized_input_pair = "This is a sample pair".split()

                # Test encode for pretokenized inputs
                output_r = tokenizer_r.encode(
                    pretokenized_input_simple, is_split_into_words=True, add_special_tokens=False
                )
                output_p = tokenizer_p.encode(
                    pretokenized_input_simple, is_split_into_words=True, add_special_tokens=False
                )
                self.assertEqual(output_p, output_r)

                kwargs = {
                    "is_split_into_words": True,
                    # "return_token_type_ids": True,  # Use the defaults for each tokenizers
                    # "return_attention_mask": True,  # Use the defaults for each tokenizers
                    "return_overflowing_tokens": False,
                    "return_special_tokens_mask": True,
                    "return_offsets_mapping": False,  # Not implemented in python tokenizers
                    # "add_special_tokens": False,
                }
                batch_kwargs = {
                    "is_split_into_words": True,
                    # "return_token_type_ids": True,  # Use the defaults for each tokenizers
                    # "return_attention_mask": True,  # Use the defaults for each tokenizers
                    "return_overflowing_tokens": False,
                    "return_special_tokens_mask": True,
                    "return_offsets_mapping": False,  # Not implemented in python tokenizers
                    # "add_special_tokens": False,
                }
                # Test encode_plus for pretokenized inputs
                output_r = tokenizer_r.encode_plus(pretokenized_input_simple, **kwargs)
                output_p = tokenizer_p.encode_plus(pretokenized_input_simple, **kwargs)
                for key in output_p.keys():
                    self.assertEqual(output_p[key], output_r[key])

                # Test batch_encode_plus for pretokenized inputs
                input_batch = ([pretokenized_input_simple] * 2) + [pretokenized_input_simple + pretokenized_input_pair]
                output_r = tokenizer_r.batch_encode_plus(input_batch, **batch_kwargs)
                output_p = tokenizer_p.batch_encode_plus(input_batch, **batch_kwargs)
                for key in output_p.keys():
                    self.assertEqual(output_p[key], output_r[key])

                # Test encode for pretokenized inputs pairs
                output_r = tokenizer_r.encode(
                    pretokenized_input_simple, pretokenized_input_pair, is_split_into_words=True
                )
                output_p = tokenizer_p.encode(
                    pretokenized_input_simple, pretokenized_input_pair, is_split_into_words=True
                )
                self.assertEqual(output_p, output_r)

                # Test encode_plus for pretokenized inputs
                output_r = tokenizer_r.encode_plus(pretokenized_input_simple, pretokenized_input_pair, **kwargs)
                output_p = tokenizer_p.encode_plus(pretokenized_input_simple, pretokenized_input_pair, **kwargs)
                for key in output_p.keys():
                    self.assertEqual(output_p[key], output_r[key])

                # Test batch_encode_plus for pretokenized inputs
                input_batch_pair = ([pretokenized_input_simple, pretokenized_input_pair] * 2) + [
                    pretokenized_input_simple + pretokenized_input_pair,
                    pretokenized_input_pair,
                ]
                output_r = tokenizer_r.batch_encode_plus(input_batch_pair, **batch_kwargs)
                output_p = tokenizer_p.batch_encode_plus(input_batch_pair, **batch_kwargs)
                for key in output_p.keys():
                    self.assertEqual(output_p[key], output_r[key])

    def test_create_token_type_ids(self):
2948
2949
2950
2951
        if not self.test_slow_tokenizer:
            # as we don't have a slow version, we can't compare the outputs between slow and fast versions
            return

2952
        for tokenizer, pretrained_name, kwargs in self.tokenizers_list:
2953
            with self.subTest(f"{tokenizer.__class__.__name__} ({pretrained_name})"):
2954
2955
2956
2957
2958
2959
2960
2961
2962
2963
2964
2965
2966
2967
2968
2969
                tokenizer_r = self.rust_tokenizer_class.from_pretrained(pretrained_name, **kwargs)
                tokenizer_p = self.tokenizer_class.from_pretrained(pretrained_name, **kwargs)
                input_simple = [1, 2, 3]
                input_pair = [1, 2, 3]

                # Generate output
                output_r = tokenizer_r.create_token_type_ids_from_sequences(input_simple)
                output_p = tokenizer_p.create_token_type_ids_from_sequences(input_simple)
                self.assertEqual(output_p, output_r)

                # Generate pair output
                output_r = tokenizer_r.create_token_type_ids_from_sequences(input_simple, input_pair)
                output_p = tokenizer_p.create_token_type_ids_from_sequences(input_simple, input_pair)
                self.assertEqual(output_p, output_r)

    def test_build_inputs_with_special_tokens(self):
2970
2971
2972
2973
        if not self.test_slow_tokenizer:
            # as we don't have a slow version, we can't compare the outputs between slow and fast versions
            return

2974
        for tokenizer, pretrained_name, kwargs in self.tokenizers_list:
2975
            with self.subTest(f"{tokenizer.__class__.__name__} ({pretrained_name})"):
2976
2977
2978
2979
2980
2981
2982
2983
2984
2985
2986
2987
2988
2989
2990
2991
2992
2993
2994
2995
2996
2997
2998
2999
3000
3001
3002
3003
3004
3005
3006
                tokenizer_r = self.rust_tokenizer_class.from_pretrained(pretrained_name, **kwargs)
                tokenizer_p = self.tokenizer_class.from_pretrained(pretrained_name, **kwargs)
                # # Input string
                # input_simple = tokenizer_p.tokenize("This is a sample input", add_special_tokens=False)
                # input_pair = tokenizer_p.tokenize("This is a sample pair", add_special_tokens=False)

                # # Generate output
                # output_r = tokenizer_r.build_inputs_with_special_tokens(input_simple)
                # output_p = tokenizer_p.build_inputs_with_special_tokens(input_simple)
                # self.assertEqual(output_p, output_r)

                # # Generate pair output
                # output_r = tokenizer_r.build_inputs_with_special_tokens(input_simple, input_pair)
                # output_p = tokenizer_p.build_inputs_with_special_tokens(input_simple, input_pair)
                # self.assertEqual(output_p, output_r)

                # Input tokens id
                input_simple = tokenizer_p.encode("This is a sample input", add_special_tokens=False)
                input_pair = tokenizer_p.encode("This is a sample pair", add_special_tokens=False)

                # Generate output
                output_r = tokenizer_r.build_inputs_with_special_tokens(input_simple)
                output_p = tokenizer_p.build_inputs_with_special_tokens(input_simple)
                self.assertEqual(output_p, output_r)

                # Generate pair output
                output_r = tokenizer_r.build_inputs_with_special_tokens(input_simple, input_pair)
                output_p = tokenizer_p.build_inputs_with_special_tokens(input_simple, input_pair)
                self.assertEqual(output_p, output_r)

    def test_padding(self, max_length=50):
3007
3008
3009
3010
        if not self.test_slow_tokenizer:
            # as we don't have a slow version, we can't compare the outputs between slow and fast versions
            return

3011
        for tokenizer, pretrained_name, kwargs in self.tokenizers_list:
3012
            with self.subTest(f"{tokenizer.__class__.__name__} ({pretrained_name})"):
3013
3014
3015
                tokenizer_r = self.rust_tokenizer_class.from_pretrained(pretrained_name, **kwargs)
                tokenizer_p = self.tokenizer_class.from_pretrained(pretrained_name, **kwargs)

3016
3017
                self.assertEqual(tokenizer_p.pad_token_id, tokenizer_r.pad_token_id)
                pad_token_id = tokenizer_p.pad_token_id
3018
3019
3020
3021

                # Encode - Simple input
                input_r = tokenizer_r.encode("This is a simple input", max_length=max_length, pad_to_max_length=True)
                input_p = tokenizer_p.encode("This is a simple input", max_length=max_length, pad_to_max_length=True)
3022
                self.assert_padded_input_match(input_r, input_p, max_length, pad_token_id)
3023
3024
                input_r = tokenizer_r.encode("This is a simple input", max_length=max_length, padding="max_length")
                input_p = tokenizer_p.encode("This is a simple input", max_length=max_length, padding="max_length")
3025
                self.assert_padded_input_match(input_r, input_p, max_length, pad_token_id)
3026
3027
3028

                input_r = tokenizer_r.encode("This is a simple input", padding="longest")
                input_p = tokenizer_p.encode("This is a simple input", padding=True)
3029
                self.assert_padded_input_match(input_r, input_p, len(input_r), pad_token_id)
3030
3031
3032
3033
3034
3035
3036
3037

                # Encode - Pair input
                input_r = tokenizer_r.encode(
                    "This is a simple input", "This is a pair", max_length=max_length, pad_to_max_length=True
                )
                input_p = tokenizer_p.encode(
                    "This is a simple input", "This is a pair", max_length=max_length, pad_to_max_length=True
                )
3038
                self.assert_padded_input_match(input_r, input_p, max_length, pad_token_id)
3039
3040
3041
3042
3043
3044
                input_r = tokenizer_r.encode(
                    "This is a simple input", "This is a pair", max_length=max_length, padding="max_length"
                )
                input_p = tokenizer_p.encode(
                    "This is a simple input", "This is a pair", max_length=max_length, padding="max_length"
                )
3045
                self.assert_padded_input_match(input_r, input_p, max_length, pad_token_id)
3046
3047
                input_r = tokenizer_r.encode("This is a simple input", "This is a pair", padding=True)
                input_p = tokenizer_p.encode("This is a simple input", "This is a pair", padding="longest")
3048
                self.assert_padded_input_match(input_r, input_p, len(input_r), pad_token_id)
3049
3050
3051
3052
3053
3054
3055
3056

                # Encode_plus - Simple input
                input_r = tokenizer_r.encode_plus(
                    "This is a simple input", max_length=max_length, pad_to_max_length=True
                )
                input_p = tokenizer_p.encode_plus(
                    "This is a simple input", max_length=max_length, pad_to_max_length=True
                )
3057
                self.assert_padded_input_match(input_r["input_ids"], input_p["input_ids"], max_length, pad_token_id)
3058
3059
3060
3061
3062
3063
3064
                self.assertSequenceEqual(input_r["attention_mask"], input_p["attention_mask"])
                input_r = tokenizer_r.encode_plus(
                    "This is a simple input", max_length=max_length, padding="max_length"
                )
                input_p = tokenizer_p.encode_plus(
                    "This is a simple input", max_length=max_length, padding="max_length"
                )
3065
                self.assert_padded_input_match(input_r["input_ids"], input_p["input_ids"], max_length, pad_token_id)
3066
3067
3068
3069
                self.assertSequenceEqual(input_r["attention_mask"], input_p["attention_mask"])

                input_r = tokenizer_r.encode_plus("This is a simple input", padding="longest")
                input_p = tokenizer_p.encode_plus("This is a simple input", padding=True)
3070
3071
3072
                self.assert_padded_input_match(
                    input_r["input_ids"], input_p["input_ids"], len(input_r["input_ids"]), pad_token_id
                )
3073
3074
3075
3076
3077
3078
3079
3080
3081
3082

                self.assertSequenceEqual(input_r["attention_mask"], input_p["attention_mask"])

                # Encode_plus - Pair input
                input_r = tokenizer_r.encode_plus(
                    "This is a simple input", "This is a pair", max_length=max_length, pad_to_max_length=True
                )
                input_p = tokenizer_p.encode_plus(
                    "This is a simple input", "This is a pair", max_length=max_length, pad_to_max_length=True
                )
3083
                self.assert_padded_input_match(input_r["input_ids"], input_p["input_ids"], max_length, pad_token_id)
3084
3085
3086
3087
3088
3089
3090
                self.assertSequenceEqual(input_r["attention_mask"], input_p["attention_mask"])
                input_r = tokenizer_r.encode_plus(
                    "This is a simple input", "This is a pair", max_length=max_length, padding="max_length"
                )
                input_p = tokenizer_p.encode_plus(
                    "This is a simple input", "This is a pair", max_length=max_length, padding="max_length"
                )
3091
                self.assert_padded_input_match(input_r["input_ids"], input_p["input_ids"], max_length, pad_token_id)
3092
3093
3094
                self.assertSequenceEqual(input_r["attention_mask"], input_p["attention_mask"])
                input_r = tokenizer_r.encode_plus("This is a simple input", "This is a pair", padding="longest")
                input_p = tokenizer_p.encode_plus("This is a simple input", "This is a pair", padding=True)
3095
3096
3097
                self.assert_padded_input_match(
                    input_r["input_ids"], input_p["input_ids"], len(input_r["input_ids"]), pad_token_id
                )
3098
3099
3100
3101
3102
3103
3104
3105
3106
3107
3108
3109
3110
                self.assertSequenceEqual(input_r["attention_mask"], input_p["attention_mask"])

                # Batch_encode_plus - Simple input
                input_r = tokenizer_r.batch_encode_plus(
                    ["This is a simple input 1", "This is a simple input 2"],
                    max_length=max_length,
                    pad_to_max_length=True,
                )
                input_p = tokenizer_p.batch_encode_plus(
                    ["This is a simple input 1", "This is a simple input 2"],
                    max_length=max_length,
                    pad_to_max_length=True,
                )
3111
                self.assert_batch_padded_input_match(input_r, input_p, max_length, pad_token_id)
3112
3113
3114
3115
3116
3117
3118
3119
3120
3121
3122

                input_r = tokenizer_r.batch_encode_plus(
                    ["This is a simple input 1", "This is a simple input 2"],
                    max_length=max_length,
                    padding="max_length",
                )
                input_p = tokenizer_p.batch_encode_plus(
                    ["This is a simple input 1", "This is a simple input 2"],
                    max_length=max_length,
                    padding="max_length",
                )
3123
                self.assert_batch_padded_input_match(input_r, input_p, max_length, pad_token_id)
3124
3125
3126
3127
3128
3129
3130
3131
3132
3133
3134

                input_r = tokenizer_r.batch_encode_plus(
                    ["This is a simple input 1", "This is a simple input 2"],
                    max_length=max_length,
                    padding="longest",
                )
                input_p = tokenizer_p.batch_encode_plus(
                    ["This is a simple input 1", "This is a simple input 2"],
                    max_length=max_length,
                    padding=True,
                )
3135
                self.assert_batch_padded_input_match(input_r, input_p, len(input_r["input_ids"][0]), pad_token_id)
3136
3137
3138
3139
3140
3141
3142

                input_r = tokenizer_r.batch_encode_plus(
                    ["This is a simple input 1", "This is a simple input 2"], padding="longest"
                )
                input_p = tokenizer_p.batch_encode_plus(
                    ["This is a simple input 1", "This is a simple input 2"], padding=True
                )
3143
                self.assert_batch_padded_input_match(input_r, input_p, len(input_r["input_ids"][0]), pad_token_id)
3144
3145
3146
3147
3148
3149
3150
3151
3152
3153
3154
3155
3156
3157
3158
3159
3160
3161
3162
3163

                # Batch_encode_plus - Pair input
                input_r = tokenizer_r.batch_encode_plus(
                    [
                        ("This is a simple input 1", "This is a simple input 2"),
                        ("This is a simple pair 1", "This is a simple pair 2"),
                    ],
                    max_length=max_length,
                    truncation=True,
                    padding="max_length",
                )
                input_p = tokenizer_p.batch_encode_plus(
                    [
                        ("This is a simple input 1", "This is a simple input 2"),
                        ("This is a simple pair 1", "This is a simple pair 2"),
                    ],
                    max_length=max_length,
                    truncation=True,
                    padding="max_length",
                )
3164
                self.assert_batch_padded_input_match(input_r, input_p, max_length, pad_token_id)
3165
3166
3167
3168
3169
3170
3171
3172
3173
3174
3175
3176
3177
3178
3179

                input_r = tokenizer_r.batch_encode_plus(
                    [
                        ("This is a simple input 1", "This is a simple input 2"),
                        ("This is a simple pair 1", "This is a simple pair 2"),
                    ],
                    padding=True,
                )
                input_p = tokenizer_p.batch_encode_plus(
                    [
                        ("This is a simple input 1", "This is a simple input 2"),
                        ("This is a simple pair 1", "This is a simple pair 2"),
                    ],
                    padding="longest",
                )
3180
                self.assert_batch_padded_input_match(input_r, input_p, len(input_r["input_ids"][0]), pad_token_id)
3181
3182
3183
3184
3185

                # Using pad on single examples after tokenization
                input_r = tokenizer_r.encode_plus("This is a input 1")
                input_r = tokenizer_r.pad(input_r)

3186
3187
                input_p = tokenizer_p.encode_plus("This is a input 1")
                input_p = tokenizer_p.pad(input_p)
3188

3189
3190
3191
                self.assert_padded_input_match(
                    input_r["input_ids"], input_p["input_ids"], len(input_r["input_ids"]), pad_token_id
                )
3192
3193
3194
3195
3196

                # Using pad on single examples after tokenization
                input_r = tokenizer_r.encode_plus("This is a input 1")
                input_r = tokenizer_r.pad(input_r, max_length=max_length, padding="max_length")

3197
3198
                input_p = tokenizer_p.encode_plus("This is a input 1")
                input_p = tokenizer_p.pad(input_p, max_length=max_length, padding="max_length")
3199

3200
                self.assert_padded_input_match(input_r["input_ids"], input_p["input_ids"], max_length, pad_token_id)
3201
3202
3203
3204
3205
3206
3207

                # Using pad after tokenization
                input_r = tokenizer_r.batch_encode_plus(
                    ["This is a input 1", "This is a much longer input whilch should be padded"]
                )
                input_r = tokenizer_r.pad(input_r)

3208
                input_p = tokenizer_p.batch_encode_plus(
3209
3210
                    ["This is a input 1", "This is a much longer input whilch should be padded"]
                )
3211
                input_p = tokenizer_p.pad(input_p)
3212

3213
                self.assert_batch_padded_input_match(input_r, input_p, len(input_r["input_ids"][0]), pad_token_id)
3214
3215
3216
3217
3218
3219
3220

                # Using pad after tokenization
                input_r = tokenizer_r.batch_encode_plus(
                    ["This is a input 1", "This is a much longer input whilch should be padded"]
                )
                input_r = tokenizer_r.pad(input_r, max_length=max_length, padding="max_length")

3221
                input_p = tokenizer_p.batch_encode_plus(
3222
3223
                    ["This is a input 1", "This is a much longer input whilch should be padded"]
                )
3224
3225
                input_p = tokenizer_p.pad(input_p, max_length=max_length, padding="max_length")
                self.assert_batch_padded_input_match(input_r, input_p, max_length, pad_token_id)
3226

3227
3228
3229
                # Test padding nested empty lists (in some use-cases, there is no any token id in the `input_ids` list).
                input_r = tokenizer_r.pad({"input_ids": [[], []]}, max_length=max_length, padding="max_length")
                input_p = tokenizer_p.pad({"input_ids": [[], []]}, max_length=max_length, padding="max_length")
3230
3231
3232
                self.assert_batch_padded_input_match(input_r, input_p, max_length, pad_token_id)

    def test_padding_different_model_input_name(self):
3233
3234
3235
3236
        if not self.test_slow_tokenizer:
            # as we don't have a slow version, we can't compare the outputs between slow and fast versions
            return

3237
        for tokenizer, pretrained_name, kwargs in self.tokenizers_list:
3238
            with self.subTest(f"{tokenizer.__class__.__name__} ({pretrained_name})"):
3239
3240
3241
3242
3243
3244
3245
3246
3247
3248
3249
3250
3251
3252
3253
3254
3255
3256
3257
3258
3259
3260
3261
3262
3263
3264
3265
3266
3267
3268
                tokenizer_r = self.rust_tokenizer_class.from_pretrained(pretrained_name, **kwargs)
                tokenizer_p = self.tokenizer_class.from_pretrained(pretrained_name, **kwargs)
                self.assertEqual(tokenizer_p.pad_token_id, tokenizer_r.pad_token_id)
                pad_token_id = tokenizer_p.pad_token_id

                input_r = tokenizer_r.batch_encode_plus(
                    ["This is a input 1", "This is a much longer input whilch should be padded"]
                )
                input_p = tokenizer_r.batch_encode_plus(
                    ["This is a input 1", "This is a much longer input whilch should be padded"]
                )

                # rename encoded batch to "inputs"
                input_r["inputs"] = input_r[tokenizer_r.model_input_names[0]]
                del input_r[tokenizer_r.model_input_names[0]]

                input_p["inputs"] = input_p[tokenizer_p.model_input_names[0]]
                del input_p[tokenizer_p.model_input_names[0]]

                # Renaming `input_ids` to `inputs`
                tokenizer_r.model_input_names = ["inputs"] + tokenizer_r.model_input_names[1:]
                tokenizer_p.model_input_names = ["inputs"] + tokenizer_p.model_input_names[1:]

                input_r = tokenizer_r.pad(input_r, padding="longest")
                input_p = tokenizer_r.pad(input_p, padding="longest")

                max_length = len(input_p["inputs"][0])
                self.assert_batch_padded_input_match(
                    input_r, input_p, max_length, pad_token_id, model_main_input_name="inputs"
                )
3269
3270

    def test_save_pretrained(self):
3271
3272
3273
3274
        if not self.test_slow_tokenizer:
            # as we don't have a slow version, we can't compare the outputs between slow and fast versions
            return

3275
        for tokenizer, pretrained_name, kwargs in self.tokenizers_list:
3276
            with self.subTest(f"{tokenizer.__class__.__name__} ({pretrained_name})"):
3277
3278
3279
3280
3281
3282
3283
                tokenizer_r = self.rust_tokenizer_class.from_pretrained(pretrained_name, **kwargs)
                tokenizer_p = self.tokenizer_class.from_pretrained(pretrained_name, **kwargs)

                tmpdirname2 = tempfile.mkdtemp()

                tokenizer_r_files = tokenizer_r.save_pretrained(tmpdirname2)
                tokenizer_p_files = tokenizer_p.save_pretrained(tmpdirname2)
Sylvain Gugger's avatar
Sylvain Gugger committed
3284
3285
3286
3287

                # Checks it save with the same files + the tokenizer.json file for the fast one
                self.assertTrue(any("tokenizer.json" in f for f in tokenizer_r_files))
                tokenizer_r_files = tuple(f for f in tokenizer_r_files if "tokenizer.json" not in f)
3288
3289
3290
3291
3292
3293
3294
3295
3296
3297
3298
3299
3300
3301
                self.assertSequenceEqual(tokenizer_r_files, tokenizer_p_files)

                # Checks everything loads correctly in the same way
                tokenizer_rp = tokenizer_r.from_pretrained(tmpdirname2)
                tokenizer_pp = tokenizer_p.from_pretrained(tmpdirname2)

                # Check special tokens are set accordingly on Rust and Python
                for key in tokenizer_pp.special_tokens_map:
                    self.assertTrue(hasattr(tokenizer_rp, key))
                    # self.assertEqual(getattr(tokenizer_rp, key), getattr(tokenizer_pp, key))
                    # self.assertEqual(getattr(tokenizer_rp, key + "_id"), getattr(tokenizer_pp, key + "_id"))

                shutil.rmtree(tmpdirname2)

Sylvain Gugger's avatar
Sylvain Gugger committed
3302
3303
3304
3305
3306
3307
3308
3309
3310
3311
3312
3313
3314
3315
3316
3317
3318
3319
3320
3321
3322
3323
3324
3325
3326
3327
3328
3329
3330
3331
3332
3333
3334
3335
3336
3337
3338
3339
                # Save tokenizer rust, legacy_format=True
                tmpdirname2 = tempfile.mkdtemp()

                tokenizer_r_files = tokenizer_r.save_pretrained(tmpdirname2, legacy_format=True)
                tokenizer_p_files = tokenizer_p.save_pretrained(tmpdirname2)

                # Checks it save with the same files
                self.assertSequenceEqual(tokenizer_r_files, tokenizer_p_files)

                # Checks everything loads correctly in the same way
                tokenizer_rp = tokenizer_r.from_pretrained(tmpdirname2)
                tokenizer_pp = tokenizer_p.from_pretrained(tmpdirname2)

                # Check special tokens are set accordingly on Rust and Python
                for key in tokenizer_pp.special_tokens_map:
                    self.assertTrue(hasattr(tokenizer_rp, key))

                shutil.rmtree(tmpdirname2)

                # Save tokenizer rust, legacy_format=False
                tmpdirname2 = tempfile.mkdtemp()

                tokenizer_r_files = tokenizer_r.save_pretrained(tmpdirname2, legacy_format=False)
                tokenizer_p_files = tokenizer_p.save_pretrained(tmpdirname2)

                # Checks it saved the tokenizer.json file
                self.assertTrue(any("tokenizer.json" in f for f in tokenizer_r_files))

                # Checks everything loads correctly in the same way
                tokenizer_rp = tokenizer_r.from_pretrained(tmpdirname2)
                tokenizer_pp = tokenizer_p.from_pretrained(tmpdirname2)

                # Check special tokens are set accordingly on Rust and Python
                for key in tokenizer_pp.special_tokens_map:
                    self.assertTrue(hasattr(tokenizer_rp, key))

                shutil.rmtree(tmpdirname2)

3340
    def test_embeded_special_tokens(self):
3341
3342
3343
3344
        if not self.test_slow_tokenizer:
            # as we don't have a slow version, we can't compare the outputs between slow and fast versions
            return

3345
        for tokenizer, pretrained_name, kwargs in self.tokenizers_list:
3346
            with self.subTest(f"{tokenizer.__class__.__name__} ({pretrained_name})"):
3347
3348
3349
3350
3351
3352
3353
3354
3355
3356
3357
3358
3359
3360
3361
3362
3363
3364
3365
3366
3367
3368
3369
3370
                tokenizer_r = self.rust_tokenizer_class.from_pretrained(pretrained_name, **kwargs)
                tokenizer_p = self.tokenizer_class.from_pretrained(pretrained_name, **kwargs)
                sentence = "A, <mask> AllenNLP sentence."
                tokens_r = tokenizer_r.encode_plus(
                    sentence,
                    add_special_tokens=True,
                )
                tokens_p = tokenizer_p.encode_plus(
                    sentence,
                    add_special_tokens=True,
                )

                for key in tokens_p.keys():
                    self.assertEqual(tokens_r[key], tokens_p[key])

                if "token_type_ids" in tokens_r:
                    self.assertEqual(sum(tokens_r["token_type_ids"]), sum(tokens_p["token_type_ids"]))

                tokens_r = tokenizer_r.convert_ids_to_tokens(tokens_r["input_ids"])
                tokens_p = tokenizer_p.convert_ids_to_tokens(tokens_p["input_ids"])
                self.assertSequenceEqual(tokens_r, tokens_p)

    def test_compare_add_special_tokens(self):
        for tokenizer, pretrained_name, kwargs in self.tokenizers_list:
3371
            with self.subTest(f"{tokenizer.__class__.__name__} ({pretrained_name})"):
3372
3373
3374
3375
3376
3377
3378
3379
3380
3381
3382
3383
3384
3385
3386
3387
3388
3389
3390
3391
3392
3393
3394
3395
3396
3397
3398
3399
3400
3401
3402
3403
3404
3405
3406
3407
3408
                tokenizer_r = self.rust_tokenizer_class.from_pretrained(pretrained_name, **kwargs)

                simple_num_special_tokens_to_add = tokenizer_r.num_special_tokens_to_add(pair=False)
                # pair_num_special_tokens_to_add = tokenizer_r.num_special_tokens_to_add(pair=True)

                for text in ["", " "]:
                    # tokenize()
                    no_special_tokens = tokenizer_r.tokenize(text, add_special_tokens=False)
                    with_special_tokens = tokenizer_r.tokenize(text, add_special_tokens=True)
                    self.assertEqual(
                        len(no_special_tokens), len(with_special_tokens) - simple_num_special_tokens_to_add
                    )

                    # encode()
                    no_special_tokens = tokenizer_r.encode(text, add_special_tokens=False)
                    with_special_tokens = tokenizer_r.encode(text, add_special_tokens=True)
                    self.assertEqual(
                        len(no_special_tokens), len(with_special_tokens) - simple_num_special_tokens_to_add
                    )

                    # encode_plus()
                    no_special_tokens = tokenizer_r.encode_plus(text, add_special_tokens=False)
                    with_special_tokens = tokenizer_r.encode_plus(text, add_special_tokens=True)
                    for key in no_special_tokens.keys():
                        self.assertEqual(
                            len(no_special_tokens[key]),
                            len(with_special_tokens[key]) - simple_num_special_tokens_to_add,
                        )

                    # # batch_encode_plus
                    no_special_tokens = tokenizer_r.batch_encode_plus([text, text], add_special_tokens=False)
                    with_special_tokens = tokenizer_r.batch_encode_plus([text, text], add_special_tokens=True)
                    for key in no_special_tokens.keys():
                        for i_no, i_with in zip(no_special_tokens[key], with_special_tokens[key]):
                            self.assertEqual(len(i_no), len(i_with) - simple_num_special_tokens_to_add)

    def test_compare_prepare_for_model(self):
3409
3410
3411
3412
        if not self.test_slow_tokenizer:
            # as we don't have a slow version, we can't compare the outputs between slow and fast versions
            return

3413
        for tokenizer, pretrained_name, kwargs in self.tokenizers_list:
3414
            with self.subTest(f"{tokenizer.__class__.__name__} ({pretrained_name})"):
3415
3416
3417
3418
3419
3420
3421
3422
3423
3424
3425
                tokenizer_r = self.rust_tokenizer_class.from_pretrained(pretrained_name, **kwargs)
                tokenizer_p = self.tokenizer_class.from_pretrained(pretrained_name, **kwargs)
                string_sequence = "Asserting that both tokenizers are equal"
                python_output = tokenizer_p.prepare_for_model(
                    tokenizer_p.encode(string_sequence, add_special_tokens=False)
                )
                rust_output = tokenizer_r.prepare_for_model(
                    tokenizer_r.encode(string_sequence, add_special_tokens=False)
                )
                for key in python_output:
                    self.assertEqual(python_output[key], rust_output[key])
Sylvain Gugger's avatar
Sylvain Gugger committed
3426

Lysandre Debut's avatar
Lysandre Debut committed
3427
3428
3429
3430
3431
3432
3433
3434
3435
3436
3437
3438
3439
3440
    def test_special_tokens_initialization(self):
        for tokenizer, pretrained_name, kwargs in self.tokenizers_list:
            with self.subTest(f"{tokenizer.__class__.__name__} ({pretrained_name})"):

                added_tokens = [AddedToken("<special>", lstrip=True)]

                tokenizer_r = self.rust_tokenizer_class.from_pretrained(
                    pretrained_name, additional_special_tokens=added_tokens, **kwargs
                )
                r_output = tokenizer_r.encode("Hey this is a <special> token")

                special_token_id = tokenizer_r.encode("<special>", add_special_tokens=False)[0]

                self.assertTrue(special_token_id in r_output)
3441
3442
3443
3444
3445
3446
3447
3448
3449
3450
3451
3452
3453
3454
3455
3456
3457

                if self.test_slow_tokenizer:
                    tokenizer_cr = self.rust_tokenizer_class.from_pretrained(
                        pretrained_name, additional_special_tokens=added_tokens, **kwargs, from_slow=True
                    )
                    tokenizer_p = self.tokenizer_class.from_pretrained(
                        pretrained_name, additional_special_tokens=added_tokens, **kwargs
                    )

                    p_output = tokenizer_p.encode("Hey this is a <special> token")

                    cr_output = tokenizer_cr.encode("Hey this is a <special> token")

                    self.assertEqual(p_output, r_output)
                    self.assertEqual(cr_output, r_output)
                    self.assertTrue(special_token_id in p_output)
                    self.assertTrue(special_token_id in cr_output)
Lysandre Debut's avatar
Lysandre Debut committed
3458

3459
3460
3461
3462
3463
3464
3465
3466
3467
3468
3469
3470
3471
3472
3473
3474
3475
3476
3477
3478
3479
3480
3481
3482
3483
3484
3485
3486
3487
3488
3489
3490
3491
3492
3493
3494
3495
3496
3497
3498
3499
3500
3501
3502
3503
3504
3505
3506
3507
3508
3509
3510
3511
3512
3513
3514
3515
3516
    def test_special_tokens_initialization_with_non_empty_additional_special_tokens(self):
        tokenizer_list = []
        if self.test_slow_tokenizer:
            tokenizer_list.append((self.tokenizer_class, self.get_tokenizer()))

        if self.test_rust_tokenizer:
            tokenizer_list.append((self.rust_tokenizer_class, self.get_rust_tokenizer()))

        for tokenizer_class, tokenizer_utils in tokenizer_list:
            with tempfile.TemporaryDirectory() as tmp_dir:
                tokenizer_utils.save_pretrained(tmp_dir)

                with open(os.path.join(tmp_dir, "special_tokens_map.json"), encoding="utf-8") as json_file:
                    special_tokens_map = json.load(json_file)

                with open(os.path.join(tmp_dir, "tokenizer_config.json"), encoding="utf-8") as json_file:
                    tokenizer_config = json.load(json_file)

                special_tokens_map["additional_special_tokens"] = ["an_additional_special_token"]
                tokenizer_config["additional_special_tokens"] = ["an_additional_special_token"]

                with open(os.path.join(tmp_dir, "special_tokens_map.json"), "w", encoding="utf-8") as outfile:
                    json.dump(special_tokens_map, outfile)
                with open(os.path.join(tmp_dir, "tokenizer_config.json"), "w", encoding="utf-8") as outfile:
                    json.dump(tokenizer_config, outfile)

                # the following checks allow us to verify that our test works as expected, i.e. that the tokenizer takes
                # into account the new value of additional_special_tokens given in the "tokenizer_config.json" and
                # "special_tokens_map.json" files
                tokenizer_without_change_in_init = tokenizer_class.from_pretrained(
                    tmp_dir,
                )
                self.assertIn(
                    "an_additional_special_token", tokenizer_without_change_in_init.additional_special_tokens
                )
                self.assertIn("an_additional_special_token", tokenizer_without_change_in_init.get_vocab())
                self.assertEqual(
                    ["an_additional_special_token"],
                    tokenizer_without_change_in_init.convert_ids_to_tokens(
                        tokenizer_without_change_in_init.convert_tokens_to_ids(["an_additional_special_token"])
                    ),
                )

                # Now we test that we can change the value of additional_special_tokens in the from_pretrained
                new_added_tokens = [AddedToken("a_new_additional_special_token", lstrip=True)]
                tokenizer = tokenizer_class.from_pretrained(
                    tmp_dir,
                    additional_special_tokens=new_added_tokens,
                )

                self.assertIn("a_new_additional_special_token", tokenizer.additional_special_tokens)
                self.assertEqual(
                    ["a_new_additional_special_token"],
                    tokenizer.convert_ids_to_tokens(
                        tokenizer.convert_tokens_to_ids(["a_new_additional_special_token"])
                    ),
                )

3517
3518
3519
3520
3521
3522
3523
3524
3525
3526
3527
3528
3529
3530
    def test_training_new_tokenizer(self):
        # This feature only exists for fast tokenizers
        if not self.test_rust_tokenizer:
            return

        tokenizer = self.get_rust_tokenizer()
        new_tokenizer = tokenizer.train_new_from_iterator(SMALL_TRAINING_CORPUS, 100)

        # Test we can use the new tokenizer with something not seen during training
        inputs = new_tokenizer(["This is the first sentence", "This sentence is different 🤗."])
        self.assertEqual(len(inputs["input_ids"]), 2)
        decoded_input = new_tokenizer.decode(inputs["input_ids"][0], skip_special_tokens=True)
        expected_result = "This is the first sentence"

3531
3532
        if tokenizer.backend_tokenizer.normalizer is not None:
            expected_result = tokenizer.backend_tokenizer.normalizer.normalize_str(expected_result)
3533
3534
3535
3536
3537
3538
3539
3540
3541
3542
3543
3544
3545
3546
3547
3548
3549
3550
3551
3552
3553
3554
3555
3556
3557
3558
3559
3560
3561
3562
3563
3564
3565
3566
3567
3568
3569
3570
3571
3572
3573
3574
3575
3576
3577
3578
3579
3580
3581
3582
3583
3584
3585
3586
3587
3588
3589
3590
3591
3592
3593
3594
3595
3596
3597
3598
3599
3600
3601
3602
3603
3604
3605
3606
3607
3608
3609
3610
3611
3612
3613
3614
3615
3616
3617
3618
3619
3620
3621
3622
3623
3624
3625
3626
3627
3628
3629
3630
3631
3632
3633
3634
3635
3636
3637
3638
3639
3640
3641
3642
3643
3644
3645
3646
        self.assertEqual(expected_result, decoded_input)

        # We check that the parameters of the tokenizer remained the same
        # Check we have the same number of added_tokens for both pair and non-pair inputs.
        self.assertEqual(tokenizer.num_special_tokens_to_add(False), new_tokenizer.num_special_tokens_to_add(False))
        self.assertEqual(tokenizer.num_special_tokens_to_add(True), new_tokenizer.num_special_tokens_to_add(True))

        # Check we have the correct max_length for both pair and non-pair inputs.
        self.assertEqual(tokenizer.max_len_single_sentence, new_tokenizer.max_len_single_sentence)
        self.assertEqual(tokenizer.max_len_sentences_pair, new_tokenizer.max_len_sentences_pair)

        # Assert the set of special tokens match as we didn't ask to change them
        self.assertSequenceEqual(
            tokenizer.all_special_tokens_extended,
            new_tokenizer.all_special_tokens_extended,
        )

        self.assertDictEqual(tokenizer.special_tokens_map, new_tokenizer.special_tokens_map)

    def test_training_new_tokenizer_with_special_tokens_change(self):
        # This feature only exists for fast tokenizers
        if not self.test_rust_tokenizer:
            return

        tokenizer = self.get_rust_tokenizer()
        # Test with a special tokens map
        class_signature = inspect.signature(tokenizer.__class__)
        if "cls_token" in class_signature.parameters:
            new_tokenizer = tokenizer.train_new_from_iterator(
                SMALL_TRAINING_CORPUS, 100, special_tokens_map={tokenizer.cls_token: "<cls>"}
            )
            cls_id = new_tokenizer.get_vocab()["<cls>"]
            self.assertEqual(new_tokenizer.cls_token, "<cls>")
            self.assertEqual(new_tokenizer.cls_token_id, cls_id)

        # Create a new mapping from the special tokens defined in the original tokenizer
        special_tokens_list = SpecialTokensMixin.SPECIAL_TOKENS_ATTRIBUTES.copy()
        special_tokens_list.remove("additional_special_tokens")
        special_tokens_map = {}
        for token in special_tokens_list:
            # Get the private one to avoid unnecessary warnings.
            if getattr(tokenizer, f"_{token}") is not None:
                special_token = getattr(tokenizer, token)
                special_tokens_map[special_token] = f"{special_token}a"

        # Train new tokenizer
        new_tokenizer = tokenizer.train_new_from_iterator(
            SMALL_TRAINING_CORPUS, 100, special_tokens_map=special_tokens_map
        )

        # Check the changes
        for token in special_tokens_list:
            # Get the private one to avoid unnecessary warnings.
            if getattr(tokenizer, f"_{token}") is None:
                continue
            special_token = getattr(tokenizer, token)
            if special_token in special_tokens_map:
                new_special_token = getattr(new_tokenizer, token)
                self.assertEqual(special_tokens_map[special_token], new_special_token)

                new_id = new_tokenizer.get_vocab()[new_special_token]
                self.assertEqual(getattr(new_tokenizer, f"{token}_id"), new_id)

        # Check if the AddedToken / string format has been kept
        for special_token in tokenizer.all_special_tokens_extended:
            if isinstance(special_token, AddedToken) and special_token.content not in special_tokens_map:
                # The special token must appear identically in the list of the new tokenizer.
                self.assertTrue(
                    special_token in new_tokenizer.all_special_tokens_extended,
                    f"'{special_token}' should be in {new_tokenizer.all_special_tokens_extended}",
                )
            elif isinstance(special_token, AddedToken):
                # The special token must appear in the list of the new tokenizer as an object of type AddedToken with
                # the same parameters as the old AddedToken except the content that the user has requested to change.
                special_token_str = special_token.content
                new_special_token_str = special_tokens_map[special_token_str]

                find = False
                for candidate in new_tokenizer.all_special_tokens_extended:
                    if (
                        isinstance(candidate, AddedToken)
                        and candidate.content == new_special_token_str
                        and candidate.lstrip == special_token.lstrip
                        and candidate.rstrip == special_token.rstrip
                        and candidate.normalized == special_token.normalized
                        and candidate.single_word == special_token.single_word
                    ):
                        find = True
                        break
                self.assertTrue(
                    find,
                    (
                        f"'{new_special_token_str}' doesn't appear in the list "
                        f"'{new_tokenizer.all_special_tokens_extended}' as an AddedToken with the same parameters as "
                        f"'{special_token}' in the list {tokenizer.all_special_tokens_extended}"
                    ),
                )
            elif special_token not in special_tokens_map:
                # The special token must appear identically in the list of the new tokenizer.
                self.assertTrue(
                    special_token in new_tokenizer.all_special_tokens_extended,
                    f"'{special_token}' should be in {new_tokenizer.all_special_tokens_extended}",
                )

            else:
                # The special token must appear in the list of the new tokenizer as an object of type string.
                self.assertTrue(special_tokens_map[special_token] in new_tokenizer.all_special_tokens_extended)

        # Test we can use the new tokenizer with something not seen during training
        inputs = new_tokenizer(["This is the first sentence", "This sentence is different 🤗."])
        self.assertEqual(len(inputs["input_ids"]), 2)
        decoded_input = new_tokenizer.decode(inputs["input_ids"][0], skip_special_tokens=True)
        expected_result = "This is the first sentence"

3647
3648
        if tokenizer.backend_tokenizer.normalizer is not None:
            expected_result = tokenizer.backend_tokenizer.normalizer.normalize_str(expected_result)
3649
3650
        self.assertEqual(expected_result, decoded_input)

3651
3652
3653
3654
3655
3656
3657
3658
3659
    def test_tokenizer_mismatch_warning(self):
        for tokenizer, pretrained_name, kwargs in self.tokenizers_list:
            with self.subTest(f"{tokenizer.__class__.__name__} ({pretrained_name})"):
                with self.assertLogs("transformers", level="WARNING") as cm:
                    try:
                        if self.tokenizer_class == BertTokenizer:
                            AlbertTokenizer.from_pretrained(pretrained_name)
                        else:
                            BertTokenizer.from_pretrained(pretrained_name)
3660
3661
3662
3663
                    except EnvironmentError as e:
                        # Some tokenizer will raised an error before reaching the logged warning because there are no
                        # corresponding files to load
                        error_message = str(e)
3664
3665
3666
3667
3668
                    except (TypeError, AttributeError):
                        # Some tokenizers cannot be loaded into the target tokenizer at all and errors are returned,
                        # here we just check that the warning has been logged before the error is raised
                        pass
                    finally:
3669
3670
3671
3672
3673
                        logged_msg_target = (
                            "The tokenizer class you load from this checkpoint is not the same type as the class "
                            "this function is called from."
                        )
                        raised_error_msg_target = "Can't load tokenizer for"
3674
                        self.assertTrue(
3675
3676
3677
                            cm.records[0].message.startswith(logged_msg_target)
                            if len(cm.records) > 0
                            else False or raised_error_msg_target in error_message
3678
3679
3680
3681
3682
3683
3684
3685
3686
3687
3688
3689
3690
3691
3692
3693
3694
                        )
                    try:
                        if self.rust_tokenizer_class == BertTokenizerFast:
                            AlbertTokenizerFast.from_pretrained(pretrained_name)
                        else:
                            BertTokenizerFast.from_pretrained(pretrained_name)
                    except (TypeError, AttributeError):
                        # Some tokenizers cannot be loaded into the target tokenizer at all and errors are returned,
                        # here we just check that the warning has been logged before the error is raised
                        pass
                    finally:
                        self.assertTrue(
                            cm.records[0].message.startswith(
                                "The tokenizer class you load from this checkpoint is not the same type as the class this function is called from."
                            )
                        )

3695
3696
3697
3698
3699
3700
3701
3702
3703
3704
3705
3706
3707
3708
3709
3710
3711
3712
3713
3714
3715
    @require_torch
    def test_saving_tokenizer_trainer(self):
        for tokenizer, pretrained_name, kwargs in self.tokenizers_list:
            with self.subTest(f"{tokenizer.__class__.__name__} ({pretrained_name})"):
                with tempfile.TemporaryDirectory() as tmp_dir:
                    # Save the fast tokenizer files in a temporary directory
                    tokenizer_old = self.rust_tokenizer_class.from_pretrained(pretrained_name, **kwargs, use_fast=True)
                    tokenizer_old.save_pretrained(tmp_dir, legacy_format=False)  # save only fast version

                    # Initialize toy model for the trainer
                    model = nn.Module()

                    # Load tokenizer from a folder without legacy files
                    tokenizer = self.rust_tokenizer_class.from_pretrained(tmp_dir)
                    training_args = TrainingArguments(output_dir=tmp_dir, do_train=True, no_cuda=True)
                    trainer = Trainer(model=model, args=training_args, tokenizer=tokenizer)

                    # Should not raise an error
                    trainer.save_model(os.path.join(tmp_dir, "checkpoint"))
                    self.assertIn("tokenizer.json", os.listdir(os.path.join(tmp_dir, "checkpoint")))

3716
3717
3718
3719
3720
3721
3722
3723
3724
3725
3726
3727
3728
3729
3730
3731
3732
3733
3734
3735
3736
3737
3738
3739
3740
3741
3742
3743
3744
    def test_save_slow_from_fast_and_reload_fast(self):
        if not self.test_slow_tokenizer or not self.test_rust_tokenizer:
            # we need both slow and fast versions
            return

        for tokenizer, pretrained_name, kwargs in self.tokenizers_list:
            with self.subTest(f"{tokenizer.__class__.__name__} ({pretrained_name})"):
                with tempfile.TemporaryDirectory() as tmp_dir_1:
                    # Here we check that even if we have initialized a fast tokenizer with a tokenizer_file we can
                    # still save only the slow version and use these saved files to rebuild a tokenizer
                    tokenizer_fast_old_1 = self.rust_tokenizer_class.from_pretrained(
                        pretrained_name, **kwargs, use_fast=True
                    )
                    tokenizer_file = os.path.join(tmp_dir_1, "tokenizer.json")
                    tokenizer_fast_old_1.backend_tokenizer.save(tokenizer_file)

                    tokenizer_fast_old_2 = self.rust_tokenizer_class.from_pretrained(
                        pretrained_name, **kwargs, use_fast=True, tokenizer_file=tokenizer_file
                    )

                    tokenizer_fast_old_2.save_pretrained(tmp_dir_1, legacy_format=True)  # save only slow version

                    tokenizer_slow = self.tokenizer_class.from_pretrained(tmp_dir_1)
                with tempfile.TemporaryDirectory() as tmp_dir_2:
                    tokenizer_slow.save_pretrained(tmp_dir_2)

                    # Should not raise an error
                    self.rust_tokenizer_class.from_pretrained(tmp_dir_2)

Sylvain Gugger's avatar
Sylvain Gugger committed
3745

3746
3747
3748
3749
3750
3751
3752
3753
3754
3755
3756
3757
3758
3759
3760
3761
3762
3763
class TokenizerUtilTester(unittest.TestCase):
    def test_cached_files_are_used_when_internet_is_down(self):
        # A mock response for an HTTP head request to emulate server down
        response_mock = mock.Mock()
        response_mock.status_code = 500
        response_mock.headers = []
        response_mock.raise_for_status.side_effect = HTTPError

        # Download this model to make sure it's in the cache.
        _ = BertTokenizer.from_pretrained("hf-internal-testing/tiny-random-bert")

        # Under the mock environment we get a 500 error when trying to reach the model.
        with mock.patch("transformers.utils.hub.requests.head", return_value=response_mock) as mock_head:
            _ = BertTokenizer.from_pretrained("hf-internal-testing/tiny-random-bert")
            # This check we did call the fake head request
            mock_head.assert_called()


Sylvain Gugger's avatar
Sylvain Gugger committed
3764
@is_staging_test
3765
class TokenizerPushToHubTester(unittest.TestCase):
Sylvain Gugger's avatar
Sylvain Gugger committed
3766
3767
3768
3769
    vocab_tokens = ["[UNK]", "[CLS]", "[SEP]", "[PAD]", "[MASK]", "bla", "blou"]

    @classmethod
    def setUpClass(cls):
3770
        cls._token = login(username=USER, password=PASS)
Sylvain Gugger's avatar
Sylvain Gugger committed
3771
3772
3773
3774

    @classmethod
    def tearDownClass(cls):
        try:
3775
            delete_repo(token=cls._token, name="test-tokenizer")
Sylvain Gugger's avatar
Sylvain Gugger committed
3776
3777
3778
3779
        except HTTPError:
            pass

        try:
3780
            delete_repo(token=cls._token, name="test-tokenizer-org", organization="valid_org")
Sylvain Gugger's avatar
Sylvain Gugger committed
3781
3782
3783
        except HTTPError:
            pass

3784
3785
3786
3787
3788
        try:
            delete_repo(token=cls._token, name="test-dynamic-tokenizer")
        except HTTPError:
            pass

Sylvain Gugger's avatar
Sylvain Gugger committed
3789
3790
3791
3792
3793
3794
    def test_push_to_hub(self):
        with tempfile.TemporaryDirectory() as tmp_dir:
            vocab_file = os.path.join(tmp_dir, "vocab.txt")
            with open(vocab_file, "w", encoding="utf-8") as vocab_writer:
                vocab_writer.write("".join([x + "\n" for x in self.vocab_tokens]))
            tokenizer = BertTokenizer(vocab_file)
3795
            tokenizer.save_pretrained(
3796
                os.path.join(tmp_dir, "test-tokenizer"), push_to_hub=True, use_auth_token=self._token
3797
            )
Sylvain Gugger's avatar
Sylvain Gugger committed
3798

3799
            new_tokenizer = BertTokenizer.from_pretrained(f"{USER}/test-tokenizer")
Sylvain Gugger's avatar
Sylvain Gugger committed
3800
3801
3802
3803
3804
3805
3806
3807
3808
            self.assertDictEqual(new_tokenizer.vocab, tokenizer.vocab)

    def test_push_to_hub_in_organization(self):
        with tempfile.TemporaryDirectory() as tmp_dir:
            vocab_file = os.path.join(tmp_dir, "vocab.txt")
            with open(vocab_file, "w", encoding="utf-8") as vocab_writer:
                vocab_writer.write("".join([x + "\n" for x in self.vocab_tokens]))
            tokenizer = BertTokenizer(vocab_file)
            tokenizer.save_pretrained(
3809
                os.path.join(tmp_dir, "test-tokenizer-org"),
Sylvain Gugger's avatar
Sylvain Gugger committed
3810
3811
3812
3813
3814
                push_to_hub=True,
                use_auth_token=self._token,
                organization="valid_org",
            )

3815
            new_tokenizer = BertTokenizer.from_pretrained("valid_org/test-tokenizer-org")
Sylvain Gugger's avatar
Sylvain Gugger committed
3816
            self.assertDictEqual(new_tokenizer.vocab, tokenizer.vocab)
3817

3818
    @require_tokenizers
3819
    def test_push_to_hub_dynamic_tokenizer(self):
3820
        CustomTokenizer.register_for_auto_class()
3821
3822
3823
3824
        with tempfile.TemporaryDirectory() as tmp_dir:
            vocab_file = os.path.join(tmp_dir, "vocab.txt")
            with open(vocab_file, "w", encoding="utf-8") as vocab_writer:
                vocab_writer.write("".join([x + "\n" for x in self.vocab_tokens]))
3825
            tokenizer = CustomTokenizer(vocab_file)
3826
3827
3828
3829
3830

        # No fast custom tokenizer
        with tempfile.TemporaryDirectory() as tmp_dir:
            repo = Repository(tmp_dir, clone_from=f"{USER}/test-dynamic-tokenizer", use_auth_token=self._token)
            tokenizer.save_pretrained(tmp_dir)
3831
3832
3833

            with open(os.path.join(tmp_dir, "tokenizer_config.json")) as f:
                tokenizer_config = json.load(f)
3834
3835
3836
            self.assertDictEqual(
                tokenizer_config["auto_map"], {"AutoTokenizer": ["custom_tokenization.CustomTokenizer", None]}
            )
3837
3838
3839
3840

            repo.push_to_hub()

        tokenizer = AutoTokenizer.from_pretrained(f"{USER}/test-dynamic-tokenizer", trust_remote_code=True)
3841
3842
        # Can't make an isinstance check because the new_model.config is from the CustomTokenizer class of a dynamic module
        self.assertEqual(tokenizer.__class__.__name__, "CustomTokenizer")
3843
3844

        # Fast and slow custom tokenizer
3845
3846
3847
3848
3849
3850
3851
3852
3853
3854
        CustomTokenizerFast.register_for_auto_class()
        with tempfile.TemporaryDirectory() as tmp_dir:
            vocab_file = os.path.join(tmp_dir, "vocab.txt")
            with open(vocab_file, "w", encoding="utf-8") as vocab_writer:
                vocab_writer.write("".join([x + "\n" for x in self.vocab_tokens]))

            bert_tokenizer = BertTokenizerFast.from_pretrained(tmp_dir)
            bert_tokenizer.save_pretrained(tmp_dir)
            tokenizer = CustomTokenizerFast.from_pretrained(tmp_dir)

3855
3856
3857
        with tempfile.TemporaryDirectory() as tmp_dir:
            repo = Repository(tmp_dir, clone_from=f"{USER}/test-dynamic-tokenizer", use_auth_token=self._token)
            tokenizer.save_pretrained(tmp_dir)
3858
3859
3860

            with open(os.path.join(tmp_dir, "tokenizer_config.json")) as f:
                tokenizer_config = json.load(f)
3861
            self.assertDictEqual(
3862
                tokenizer_config["auto_map"],
3863
3864
3865
3866
3867
3868
                {
                    "AutoTokenizer": [
                        "custom_tokenization.CustomTokenizer",
                        "custom_tokenization_fast.CustomTokenizerFast",
                    ]
                },
3869
            )
3870
3871
3872
3873
3874

            repo.push_to_hub()

        tokenizer = AutoTokenizer.from_pretrained(f"{USER}/test-dynamic-tokenizer", trust_remote_code=True)
        # Can't make an isinstance check because the new_model.config is from the FakeConfig class of a dynamic module
3875
        self.assertEqual(tokenizer.__class__.__name__, "CustomTokenizerFast")
3876
3877
3878
3879
        tokenizer = AutoTokenizer.from_pretrained(
            f"{USER}/test-dynamic-tokenizer", use_fast=False, trust_remote_code=True
        )
        # Can't make an isinstance check because the new_model.config is from the FakeConfig class of a dynamic module
3880
        self.assertEqual(tokenizer.__class__.__name__, "CustomTokenizer")
3881

3882
3883
3884
3885
3886
3887
3888
3889
3890
3891
3892
3893
3894
3895
3896
3897
3898

class TrieTest(unittest.TestCase):
    def test_trie(self):
        trie = Trie()
        trie.add("Hello 友達")
        self.assertEqual(trie.data, {"H": {"e": {"l": {"l": {"o": {" ": {"友": {"達": {"": 1}}}}}}}}})
        trie.add("Hello")
        trie.data
        self.assertEqual(trie.data, {"H": {"e": {"l": {"l": {"o": {"": 1, " ": {"友": {"達": {"": 1}}}}}}}}})

    def test_trie_split(self):
        trie = Trie()
        self.assertEqual(trie.split("[CLS] This is a extra_id_100"), ["[CLS] This is a extra_id_100"])
        trie.add("[CLS]")
        trie.add("extra_id_1")
        trie.add("extra_id_100")
        self.assertEqual(trie.split("[CLS] This is a extra_id_100"), ["[CLS]", " This is a ", "extra_id_100"])
3899
3900
3901
3902
3903
3904
3905
3906
3907
3908
3909
3910

    def test_trie_single(self):
        trie = Trie()
        trie.add("A")
        self.assertEqual(trie.split("ABC"), ["A", "BC"])
        self.assertEqual(trie.split("BCA"), ["BC", "A"])

    def test_trie_final(self):
        trie = Trie()
        trie.add("TOKEN]")
        trie.add("[SPECIAL_TOKEN]")
        self.assertEqual(trie.split("This is something [SPECIAL_TOKEN]"), ["This is something ", "[SPECIAL_TOKEN]"])
3911
3912
3913
3914
3915
3916
3917
3918
3919
3920
3921
3922
3923
3924
3925

    def test_trie_subtokens(self):
        trie = Trie()
        trie.add("A")
        trie.add("P")
        trie.add("[SPECIAL_TOKEN]")
        self.assertEqual(trie.split("This is something [SPECIAL_TOKEN]"), ["This is something ", "[SPECIAL_TOKEN]"])

    def test_trie_suffix_tokens(self):
        trie = Trie()
        trie.add("AB")
        trie.add("B")
        trie.add("C")
        self.assertEqual(trie.split("ABC"), ["AB", "C"])

3926
3927
3928
3929
3930
3931
3932
    def test_trie_skip(self):
        trie = Trie()
        trie.add("ABC")
        trie.add("B")
        trie.add("CD")
        self.assertEqual(trie.split("ABCD"), ["ABC", "D"])

3933
3934
3935
3936
3937
3938
    def test_cut_text_hardening(self):
        # Even if the offsets are wrong, we necessarily output correct string
        # parts.
        trie = Trie()
        parts = trie.cut_text("ABC", [0, 0, 2, 1, 2, 3])
        self.assertEqual(parts, ["AB", "C"])