check_repo.py 26.3 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
# coding=utf-8
# Copyright 2020 The HuggingFace Inc. team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

16
17
18
19
import importlib
import inspect
import os
import re
20
import warnings
21
from pathlib import Path
22

23
24
from transformers import is_flax_available, is_tf_available, is_torch_available
from transformers.file_utils import ENV_VARS_TRUE_VALUES
25
26
from transformers.models.auto import get_values

27
28
29
30
31

# All paths are set with the intent you should run this script from the root of the repo with the command
# python utils/check_repo.py
PATH_TO_TRANSFORMERS = "src/transformers"
PATH_TO_TESTS = "tests"
32
PATH_TO_DOC = "docs/source"
33

34
35
36
37
38
39
40
# Update this list with models that are supposed to be private.
PRIVATE_MODELS = [
    "DPRSpanPredictor",
    "T5Stack",
    "TFDPRSpanPredictor",
]

41
42
# Update this list for models that are not tested with a comment explaining the reason it should not be.
# Being in this list is an exception and should **not** be the rule.
43
IGNORE_NON_TESTED = PRIVATE_MODELS.copy() + [
44
    # models to ignore for not tested
NielsRogge's avatar
NielsRogge committed
45
    "SegformerDecodeHead",  # Building part of bigger (tested) model.
Vasudev Gupta's avatar
Vasudev Gupta committed
46
47
48
    "BigBirdPegasusEncoder",  # Building part of bigger (tested) model.
    "BigBirdPegasusDecoder",  # Building part of bigger (tested) model.
    "BigBirdPegasusDecoderWrapper",  # Building part of bigger (tested) model.
NielsRogge's avatar
NielsRogge committed
49
50
51
    "DetrEncoder",  # Building part of bigger (tested) model.
    "DetrDecoder",  # Building part of bigger (tested) model.
    "DetrDecoderWrapper",  # Building part of bigger (tested) model.
Suraj Patil's avatar
Suraj Patil committed
52
53
    "M2M100Encoder",  # Building part of bigger (tested) model.
    "M2M100Decoder",  # Building part of bigger (tested) model.
Suraj Patil's avatar
Suraj Patil committed
54
55
    "Speech2TextEncoder",  # Building part of bigger (tested) model.
    "Speech2TextDecoder",  # Building part of bigger (tested) model.
Patrick von Platen's avatar
Patrick von Platen committed
56
57
    "LEDEncoder",  # Building part of bigger (tested) model.
    "LEDDecoder",  # Building part of bigger (tested) model.
58
    "BartDecoderWrapper",  # Building part of bigger (tested) model.
59
    "BartEncoder",  # Building part of bigger (tested) model.
60
    "BertLMHeadModel",  # Needs to be setup as decoder.
61
    "BlenderbotSmallEncoder",  # Building part of bigger (tested) model.
62
    "BlenderbotSmallDecoderWrapper",  # Building part of bigger (tested) model.
63
    "BlenderbotEncoder",  # Building part of bigger (tested) model.
64
    "BlenderbotDecoderWrapper",  # Building part of bigger (tested) model.
65
    "MBartEncoder",  # Building part of bigger (tested) model.
66
    "MBartDecoderWrapper",  # Building part of bigger (tested) model.
67
68
69
70
    "MegatronBertLMHeadModel",  # Building part of bigger (tested) model.
    "MegatronBertEncoder",  # Building part of bigger (tested) model.
    "MegatronBertDecoder",  # Building part of bigger (tested) model.
    "MegatronBertDecoderWrapper",  # Building part of bigger (tested) model.
71
    "PegasusEncoder",  # Building part of bigger (tested) model.
72
    "PegasusDecoderWrapper",  # Building part of bigger (tested) model.
73
    "DPREncoder",  # Building part of bigger (tested) model.
74
    "ProphetNetDecoderWrapper",  # Building part of bigger (tested) model.
75
    "ReformerForMaskedLM",  # Needs to be setup as decoder.
76
    "Speech2Text2DecoderWrapper",  # Building part of bigger (tested) model.
Ratthachat (Jung)'s avatar
Ratthachat (Jung) committed
77
    "TFDPREncoder",  # Building part of bigger (tested) model.
78
79
    "TFElectraMainLayer",  # Building part of bigger (tested) model (should it be a TFPreTrainedModel ?)
    "TFRobertaForMultipleChoice",  # TODO: fix
80
    "TrOCRDecoderWrapper",  # Building part of bigger (tested) model.
abhishek thakur's avatar
abhishek thakur committed
81
    "SeparableConv1D",  # Building part of bigger (tested) model.
82
83
84
85
86
87
]

# Update this list with test files that don't have a tester with a `all_model_classes` variable and which don't
# trigger the common tests.
TEST_FILES_WITH_NO_COMMON_TESTS = [
    "test_modeling_camembert.py",
88
    "test_modeling_flax_mt5.py",
Stas Bekman's avatar
Stas Bekman committed
89
    "test_modeling_mbart.py",
Patrick von Platen's avatar
Patrick von Platen committed
90
    "test_modeling_mt5.py",
Stas Bekman's avatar
Stas Bekman committed
91
    "test_modeling_pegasus.py",
92
    "test_modeling_tf_camembert.py",
Sylvain Gugger's avatar
Sylvain Gugger committed
93
    "test_modeling_tf_mt5.py",
94
    "test_modeling_tf_xlm_roberta.py",
Weizhen's avatar
Weizhen committed
95
    "test_modeling_xlm_prophetnet.py",
96
    "test_modeling_xlm_roberta.py",
Suraj Patil's avatar
Suraj Patil committed
97
98
    "test_modeling_vision_text_dual_encoder.py",
    "test_modeling_flax_vision_text_dual_encoder.py",
99
100
]

101
102
# Update this list for models that are not in any of the auto MODEL_XXX_MAPPING. Being in this list is an exception and
# should **not** be the rule.
103
IGNORE_NON_AUTO_CONFIGURED = PRIVATE_MODELS.copy() + [
104
    # models to ignore for model xxx mapping
NielsRogge's avatar
NielsRogge committed
105
106
    "PerceiverForMultimodalAutoencoding",
    "PerceiverForOpticalFlow",
NielsRogge's avatar
NielsRogge committed
107
108
    "SegformerDecodeHead",
    "SegformerForSemanticSegmentation",
109
    "BeitForSemanticSegmentation",
Kamal Raj's avatar
Kamal Raj committed
110
    "FlaxBeitForMaskedImageModeling",
NielsRogge's avatar
NielsRogge committed
111
    "BeitForMaskedImageModeling",
Suraj Patil's avatar
Suraj Patil committed
112
113
    "CLIPTextModel",
    "CLIPVisionModel",
Yih-Dar's avatar
Yih-Dar committed
114
115
    "TFCLIPTextModel",
    "TFCLIPVisionModel",
Suraj Patil's avatar
Suraj Patil committed
116
117
    "FlaxCLIPTextModel",
    "FlaxCLIPVisionModel",
118
    "FlaxWav2Vec2ForCTC",
NielsRogge's avatar
NielsRogge committed
119
    "DetrForSegmentation",
120
121
122
    "DPRReader",
    "FlaubertForQuestionAnswering",
    "GPT2DoubleHeadsModel",
Ryokan RI's avatar
Ryokan RI committed
123
    "LukeForMaskedLM",
NielsRogge's avatar
NielsRogge committed
124
125
126
    "LukeForEntityClassification",
    "LukeForEntityPairClassification",
    "LukeForEntitySpanClassification",
127
128
129
130
    "OpenAIGPTDoubleHeadsModel",
    "RagModel",
    "RagSequenceForGeneration",
    "RagTokenForGeneration",
Ratthachat (Jung)'s avatar
Ratthachat (Jung) committed
131
    "TFDPRReader",
132
133
    "TFGPT2DoubleHeadsModel",
    "TFOpenAIGPTDoubleHeadsModel",
Ratthachat (Jung)'s avatar
Ratthachat (Jung) committed
134
135
136
    "TFRagModel",
    "TFRagSequenceForGeneration",
    "TFRagTokenForGeneration",
137
    "Wav2Vec2ForCTC",
Patrick von Platen's avatar
Patrick von Platen committed
138
    "HubertForCTC",
139
140
    "SEWForCTC",
    "SEWDForCTC",
141
142
    "XLMForQuestionAnswering",
    "XLNetForQuestionAnswering",
abhishek thakur's avatar
abhishek thakur committed
143
    "SeparableConv1D",
Gunjan Chhablani's avatar
Gunjan Chhablani committed
144
145
146
147
    "VisualBertForRegionToPhraseAlignment",
    "VisualBertForVisualReasoning",
    "VisualBertForQuestionAnswering",
    "VisualBertForMultipleChoice",
Will Rice's avatar
Will Rice committed
148
    "TFWav2Vec2ForCTC",
Will Rice's avatar
Will Rice committed
149
    "TFHubertForCTC",
150
151
]

152
153
154
155
156
157
158
159
160
# This is to make sure the transformers module imported is the one in the repo.
spec = importlib.util.spec_from_file_location(
    "transformers",
    os.path.join(PATH_TO_TRANSFORMERS, "__init__.py"),
    submodule_search_locations=[PATH_TO_TRANSFORMERS],
)
transformers = spec.loader.load_module()


161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
def check_model_list():
    """Check the model list inside the transformers library."""
    # Get the models from the directory structure of `src/transformers/models/`
    models_dir = os.path.join(PATH_TO_TRANSFORMERS, "models")
    _models = []
    for model in os.listdir(models_dir):
        model_dir = os.path.join(models_dir, model)
        if os.path.isdir(model_dir) and "__init__.py" in os.listdir(model_dir):
            _models.append(model)

    # Get the models from the directory structure of `src/transformers/models/`
    models = [model for model in dir(transformers.models) if not model.startswith("__")]

    missing_models = sorted(list(set(_models).difference(models)))
    if missing_models:
        raise Exception(
            f"The following models should be included in {models_dir}/__init__.py: {','.join(missing_models)}."
        )


181
182
183
# If some modeling modules should be ignored for all checks, they should be added in the nested list
# _ignore_modules of this function.
def get_model_modules():
Patrick von Platen's avatar
Patrick von Platen committed
184
    """Get the model modules inside the transformers library."""
185
186
187
188
189
190
191
192
    _ignore_modules = [
        "modeling_auto",
        "modeling_encoder_decoder",
        "modeling_marian",
        "modeling_mmbt",
        "modeling_outputs",
        "modeling_retribert",
        "modeling_utils",
Sylvain Gugger's avatar
Sylvain Gugger committed
193
        "modeling_flax_auto",
194
        "modeling_flax_encoder_decoder",
Stas Bekman's avatar
Stas Bekman committed
195
        "modeling_flax_utils",
196
        "modeling_speech_encoder_decoder",
197
        "modeling_flax_vision_encoder_decoder",
198
199
        "modeling_transfo_xl_utilities",
        "modeling_tf_auto",
200
        "modeling_tf_encoder_decoder",
201
202
203
204
        "modeling_tf_outputs",
        "modeling_tf_pytorch_utils",
        "modeling_tf_utils",
        "modeling_tf_transfo_xl_utilities",
205
        "modeling_vision_encoder_decoder",
206
207
    ]
    modules = []
Sylvain Gugger's avatar
Sylvain Gugger committed
208
209
210
211
212
213
214
215
216
    for model in dir(transformers.models):
        # There are some magic dunder attributes in the dir, we ignore them
        if not model.startswith("__"):
            model_module = getattr(transformers.models, model)
            for submodule in dir(model_module):
                if submodule.startswith("modeling") and submodule not in _ignore_modules:
                    modeling_module = getattr(model_module, submodule)
                    if inspect.ismodule(modeling_module):
                        modules.append(modeling_module)
217
218
219
    return modules


220
def get_models(module, include_pretrained=False):
Patrick von Platen's avatar
Patrick von Platen committed
221
    """Get the objects in module that are models."""
222
    models = []
223
    model_classes = (transformers.PreTrainedModel, transformers.TFPreTrainedModel, transformers.FlaxPreTrainedModel)
224
    for attr_name in dir(module):
225
        if not include_pretrained and ("Pretrained" in attr_name or "PreTrained" in attr_name):
226
227
228
229
230
231
232
            continue
        attr = getattr(module, attr_name)
        if isinstance(attr, type) and issubclass(attr, model_classes) and attr.__module__ == module.__name__:
            models.append((attr_name, attr))
    return models


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
259
260
261
262
def is_a_private_model(model):
    """Returns True if the model should not be in the main init."""
    if model in PRIVATE_MODELS:
        return True

    # Wrapper, Encoder and Decoder are all privates
    if model.endswith("Wrapper"):
        return True
    if model.endswith("Encoder"):
        return True
    if model.endswith("Decoder"):
        return True
    return False


def check_models_are_in_init():
    """Checks all models defined in the library are in the main init."""
    models_not_in_init = []
    dir_transformers = dir(transformers)
    for module in get_model_modules():
        models_not_in_init += [
            model[0] for model in get_models(module, include_pretrained=True) if model[0] not in dir_transformers
        ]

    # Remove private models
    models_not_in_init = [model for model in models_not_in_init if not is_a_private_model(model)]
    if len(models_not_in_init) > 0:
        raise Exception(f"The following models should be in the main init: {','.join(models_not_in_init)}.")


263
264
265
# If some test_modeling files should be ignored when checking models are all tested, they should be added in the
# nested list _ignore_files of this function.
def get_model_test_files():
Patrick von Platen's avatar
Patrick von Platen committed
266
    """Get the model test files."""
267
268
269
    _ignore_files = [
        "test_modeling_common",
        "test_modeling_encoder_decoder",
270
        "test_modeling_flax_encoder_decoder",
271
272
        "test_modeling_marian",
        "test_modeling_tf_common",
273
        "test_modeling_tf_encoder_decoder",
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
    ]
    test_files = []
    for filename in os.listdir(PATH_TO_TESTS):
        if (
            os.path.isfile(f"{PATH_TO_TESTS}/{filename}")
            and filename.startswith("test_modeling")
            and not os.path.splitext(filename)[0] in _ignore_files
        ):
            test_files.append(filename)
    return test_files


# This is a bit hacky but I didn't find a way to import the test_file as a module and read inside the tester class
# for the all_model_classes variable.
def find_tested_models(test_file):
Patrick von Platen's avatar
Patrick von Platen committed
289
    """Parse the content of test_file to detect what's in all_model_classes"""
Sylvain Gugger's avatar
Sylvain Gugger committed
290
    # This is a bit hacky but I didn't find a way to import the test_file as a module and read inside the class
291
    with open(os.path.join(PATH_TO_TESTS, test_file), "r", encoding="utf-8", newline="\n") as f:
292
        content = f.read()
Sylvain Gugger's avatar
Sylvain Gugger committed
293
    all_models = re.findall(r"all_model_classes\s+=\s+\(\s*\(([^\)]*)\)", content)
294
295
    # Check with one less parenthesis as well
    all_models += re.findall(r"all_model_classes\s+=\s+\(([^\)]*)\)", content)
Sylvain Gugger's avatar
Sylvain Gugger committed
296
    if len(all_models) > 0:
297
        model_tested = []
Sylvain Gugger's avatar
Sylvain Gugger committed
298
299
300
301
302
        for entry in all_models:
            for line in entry.split(","):
                name = line.strip()
                if len(name) > 0:
                    model_tested.append(name)
303
304
305
306
        return model_tested


def check_models_are_tested(module, test_file):
Patrick von Platen's avatar
Patrick von Platen committed
307
    """Check models defined in module are tested in test_file."""
308
    # XxxPreTrainedModel are not tested
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
    defined_models = get_models(module)
    tested_models = find_tested_models(test_file)
    if tested_models is None:
        if test_file in TEST_FILES_WITH_NO_COMMON_TESTS:
            return
        return [
            f"{test_file} should define `all_model_classes` to apply common tests to the models it tests. "
            + "If this intentional, add the test filename to `TEST_FILES_WITH_NO_COMMON_TESTS` in the file "
            + "`utils/check_repo.py`."
        ]
    failures = []
    for model_name, _ in defined_models:
        if model_name not in tested_models and model_name not in IGNORE_NON_TESTED:
            failures.append(
                f"{model_name} is defined in {module.__name__} but is not tested in "
                + f"{os.path.join(PATH_TO_TESTS, test_file)}. Add it to the all_model_classes in that file."
                + "If common tests should not applied to that model, add its name to `IGNORE_NON_TESTED`"
                + "in the file `utils/check_repo.py`."
            )
    return failures


def check_all_models_are_tested():
Patrick von Platen's avatar
Patrick von Platen committed
332
    """Check all models are properly tested."""
333
334
335
336
    modules = get_model_modules()
    test_files = get_model_test_files()
    failures = []
    for module in modules:
Sylvain Gugger's avatar
Sylvain Gugger committed
337
        test_file = f"test_{module.__name__.split('.')[-1]}.py"
338
339
340
341
342
343
344
345
346
        if test_file not in test_files:
            failures.append(f"{module.__name__} does not have its corresponding test file {test_file}.")
        new_failures = check_models_are_tested(module, test_file)
        if new_failures is not None:
            failures += new_failures
    if len(failures) > 0:
        raise Exception(f"There were {len(failures)} failures:\n" + "\n".join(failures))


347
def get_all_auto_configured_models():
Patrick von Platen's avatar
Patrick von Platen committed
348
    """Return the list of all models in at least one auto class."""
349
    result = set()  # To avoid duplicates we concatenate all model classes in a set.
350
351
    if is_torch_available():
        for attr_name in dir(transformers.models.auto.modeling_auto):
352
            if attr_name.startswith("MODEL_") and attr_name.endswith("MAPPING_NAMES"):
353
354
355
                result = result | set(get_values(getattr(transformers.models.auto.modeling_auto, attr_name)))
    if is_tf_available():
        for attr_name in dir(transformers.models.auto.modeling_tf_auto):
356
            if attr_name.startswith("TF_MODEL_") and attr_name.endswith("MAPPING_NAMES"):
357
358
359
                result = result | set(get_values(getattr(transformers.models.auto.modeling_tf_auto, attr_name)))
    if is_flax_available():
        for attr_name in dir(transformers.models.auto.modeling_flax_auto):
360
            if attr_name.startswith("FLAX_MODEL_") and attr_name.endswith("MAPPING_NAMES"):
361
                result = result | set(get_values(getattr(transformers.models.auto.modeling_flax_auto, attr_name)))
362
    return [cls for cls in result]
363
364


365
366
367
368
369
370
371
372
373
374
375
def ignore_unautoclassed(model_name):
    """Rules to determine if `name` should be in an auto class."""
    # Special white list
    if model_name in IGNORE_NON_AUTO_CONFIGURED:
        return True
    # Encoder and Decoder should be ignored
    if "Encoder" in model_name or "Decoder" in model_name:
        return True
    return False


376
def check_models_are_auto_configured(module, all_auto_models):
Patrick von Platen's avatar
Patrick von Platen committed
377
    """Check models defined in module are each in an auto class."""
378
379
380
    defined_models = get_models(module)
    failures = []
    for model_name, _ in defined_models:
381
        if model_name not in all_auto_models and not ignore_unautoclassed(model_name):
382
383
384
385
386
387
388
389
390
            failures.append(
                f"{model_name} is defined in {module.__name__} but is not present in any of the auto mapping. "
                "If that is intended behavior, add its name to `IGNORE_NON_AUTO_CONFIGURED` in the file "
                "`utils/check_repo.py`."
            )
    return failures


def check_all_models_are_auto_configured():
Patrick von Platen's avatar
Patrick von Platen committed
391
    """Check all models are each in an auto class."""
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
    missing_backends = []
    if not is_torch_available():
        missing_backends.append("PyTorch")
    if not is_tf_available():
        missing_backends.append("TensorFlow")
    if not is_flax_available():
        missing_backends.append("Flax")
    if len(missing_backends) > 0:
        missing = ", ".join(missing_backends)
        if os.getenv("TRANSFORMERS_IS_CI", "").upper() in ENV_VARS_TRUE_VALUES:
            raise Exception(
                "Full quality checks require all backends to be installed (with `pip install -e .[dev]` in the "
                f"Transformers repo, the following are missing: {missing}."
            )
        else:
            warnings.warn(
                "Full quality checks require all backends to be installed (with `pip install -e .[dev]` in the "
                f"Transformers repo, the following are missing: {missing}. While it's probably fine as long as you "
                "didn't make any change in one of those backends modeling files, you should probably execute the "
                "command above to be on the safe side."
            )
413
414
415
416
417
418
419
420
421
422
423
    modules = get_model_modules()
    all_auto_models = get_all_auto_configured_models()
    failures = []
    for module in modules:
        new_failures = check_models_are_auto_configured(module, all_auto_models)
        if new_failures is not None:
            failures += new_failures
    if len(failures) > 0:
        raise Exception(f"There were {len(failures)} failures:\n" + "\n".join(failures))


Sylvain Gugger's avatar
Sylvain Gugger committed
424
425
426
427
_re_decorator = re.compile(r"^\s*@(\S+)\s+$")


def check_decorator_order(filename):
Patrick von Platen's avatar
Patrick von Platen committed
428
    """Check that in the test file `filename` the slow decorator is always last."""
429
    with open(filename, "r", encoding="utf-8", newline="\n") as f:
Sylvain Gugger's avatar
Sylvain Gugger committed
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
        lines = f.readlines()
    decorator_before = None
    errors = []
    for i, line in enumerate(lines):
        search = _re_decorator.search(line)
        if search is not None:
            decorator_name = search.groups()[0]
            if decorator_before is not None and decorator_name.startswith("parameterized"):
                errors.append(i)
            decorator_before = decorator_name
        elif decorator_before is not None:
            decorator_before = None
    return errors


def check_all_decorator_order():
Patrick von Platen's avatar
Patrick von Platen committed
446
    """Check that in all test files, the slow decorator is always last."""
Sylvain Gugger's avatar
Sylvain Gugger committed
447
448
449
450
451
452
453
454
455
456
457
458
459
    errors = []
    for fname in os.listdir(PATH_TO_TESTS):
        if fname.endswith(".py"):
            filename = os.path.join(PATH_TO_TESTS, fname)
            new_errors = check_decorator_order(filename)
            errors += [f"- {filename}, line {i}" for i in new_errors]
    if len(errors) > 0:
        msg = "\n".join(errors)
        raise ValueError(
            f"The parameterized decorator (and its variants) should always be first, but this is not the case in the following files:\n{msg}"
        )


460
def find_all_documented_objects():
Patrick von Platen's avatar
Patrick von Platen committed
461
    """Parse the content of all doc files to detect which classes and functions it documents"""
462
463
    documented_obj = []
    for doc_file in Path(PATH_TO_DOC).glob("**/*.rst"):
Julien Plu's avatar
Julien Plu committed
464
        with open(doc_file, "r", encoding="utf-8", newline="\n") as f:
465
466
467
            content = f.read()
        raw_doc_objs = re.findall(r"(?:autoclass|autofunction):: transformers.(\S+)\s+", content)
        documented_obj += [obj.split(".")[-1] for obj in raw_doc_objs]
Sylvain Gugger's avatar
Sylvain Gugger committed
468
469
470
471
472
    for doc_file in Path(PATH_TO_DOC).glob("**/*.mdx"):
        with open(doc_file, "r", encoding="utf-8", newline="\n") as f:
            content = f.read()
        raw_doc_objs = re.findall("\[\[autodoc\]\]\s+(\S+)\s+", content)
        documented_obj += [obj.split(".")[-1] for obj in raw_doc_objs]
473
474
475
476
477
478
    return documented_obj


# One good reason for not being documented is to be deprecated. Put in this list deprecated objects.
DEPRECATED_OBJECTS = [
    "AutoModelWithLMHead",
479
    "BartPretrainedModel",
480
481
    "DataCollator",
    "DataCollatorForSOP",
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
    "GlueDataset",
    "GlueDataTrainingArguments",
    "LineByLineTextDataset",
    "LineByLineWithRefDataset",
    "LineByLineWithSOPTextDataset",
    "PretrainedBartModel",
    "PretrainedFSMTModel",
    "SingleSentenceClassificationProcessor",
    "SquadDataTrainingArguments",
    "SquadDataset",
    "SquadExample",
    "SquadFeatures",
    "SquadV1Processor",
    "SquadV2Processor",
    "TFAutoModelWithLMHead",
497
    "TFBartPretrainedModel",
498
499
    "TextDataset",
    "TextDatasetForNextSentencePrediction",
500
    "Wav2Vec2ForMaskedLM",
501
    "Wav2Vec2Tokenizer",
502
503
504
505
506
507
508
509
510
511
    "glue_compute_metrics",
    "glue_convert_examples_to_features",
    "glue_output_modes",
    "glue_processors",
    "glue_tasks_num_labels",
    "squad_convert_examples_to_features",
    "xnli_compute_metrics",
    "xnli_output_modes",
    "xnli_processors",
    "xnli_tasks_num_labels",
512
513
    "TFTrainer",
    "TFTrainingArguments",
514
515
516
517
518
519
520
]

# Exceptionally, some objects should not be documented after all rules passed.
# ONLY PUT SOMETHING IN THIS LIST AS A LAST RESORT!
UNDOCUMENTED_OBJECTS = [
    "AddedToken",  # This is a tokenizers class.
    "BasicTokenizer",  # Internal, should never have been in the main init.
521
    "CharacterTokenizer",  # Internal, should never have been in the main init.
522
    "DPRPretrainedReader",  # Like an Encoder.
523
    "MecabTokenizer",  # Internal, should never have been in the main init.
524
525
526
527
528
529
530
531
532
533
534
535
    "ModelCard",  # Internal type.
    "SqueezeBertModule",  # Internal building block (should have been called SqueezeBertLayer)
    "TFDPRPretrainedReader",  # Like an Encoder.
    "TransfoXLCorpus",  # Internal type.
    "WordpieceTokenizer",  # Internal, should never have been in the main init.
    "absl",  # External module
    "add_end_docstrings",  # Internal, should never have been in the main init.
    "add_start_docstrings",  # Internal, should never have been in the main init.
    "cached_path",  # Internal used for downloading models.
    "convert_tf_weight_name_to_pt_weight_name",  # Internal used to convert model weights
    "logger",  # Internal logger
    "logging",  # External module
536
    "requires_backends",  # Internal function
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
]

# This list should be empty. Objects in it should get their own doc page.
SHOULD_HAVE_THEIR_OWN_PAGE = [
    # Benchmarks
    "PyTorchBenchmark",
    "PyTorchBenchmarkArguments",
    "TensorFlowBenchmark",
    "TensorFlowBenchmarkArguments",
]


def ignore_undocumented(name):
    """Rules to determine if `name` should be undocumented."""
    # NOT DOCUMENTED ON PURPOSE.
    # Constants uppercase are not documented.
    if name.isupper():
        return True
    # PreTrainedModels / Encoders / Decoders / Layers / Embeddings / Attention are not documented.
    if (
        name.endswith("PreTrainedModel")
        or name.endswith("Decoder")
        or name.endswith("Encoder")
        or name.endswith("Layer")
        or name.endswith("Embeddings")
        or name.endswith("Attention")
    ):
        return True
    # Submodules are not documented.
    if os.path.isdir(os.path.join(PATH_TO_TRANSFORMERS, name)) or os.path.isfile(
        os.path.join(PATH_TO_TRANSFORMERS, f"{name}.py")
    ):
        return True
    # All load functions are not documented.
    if name.startswith("load_tf") or name.startswith("load_pytorch"):
        return True
    # is_xxx_available functions are not documented.
    if name.startswith("is_") and name.endswith("_available"):
        return True
    # Deprecated objects are not documented.
    if name in DEPRECATED_OBJECTS or name in UNDOCUMENTED_OBJECTS:
        return True
    # MMBT model does not really work.
    if name.startswith("MMBT"):
        return True
    if name in SHOULD_HAVE_THEIR_OWN_PAGE:
        return True
    return False


def check_all_objects_are_documented():
Patrick von Platen's avatar
Patrick von Platen committed
588
    """Check all models are properly documented."""
589
    documented_objs = find_all_documented_objects()
590
591
592
    modules = transformers._modules
    objects = [c for c in dir(transformers) if c not in modules and not c.startswith("_")]
    undocumented_objs = [c for c in objects if c not in documented_objs and not ignore_undocumented(c)]
593
594
595
596
597
    if len(undocumented_objs) > 0:
        raise Exception(
            "The following objects are in the public init so should be documented:\n - "
            + "\n - ".join(undocumented_objs)
        )
598
599
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
633
634
635
636
637
638
639
    check_docstrings_are_in_md()


# Re pattern to catch :obj:`xx`, :class:`xx`, :func:`xx` or :meth:`xx`.
_re_rst_special_words = re.compile(r":(?:obj|func|class|meth):`([^`]+)`")
# Re pattern to catch things between double backquotes.
_re_double_backquotes = re.compile(r"(^|[^`])``([^`]+)``([^`]|$)")
# Re pattern to catch example introduction.
_re_rst_example = re.compile(r"^\s*Example.*::\s*$", flags=re.MULTILINE)


def is_rst_docstring(docstring):
    """
    Returns `True` if `docstring` is written in rst.
    """
    if _re_rst_special_words.search(docstring) is not None:
        return True
    if _re_double_backquotes.search(docstring) is not None:
        return True
    if _re_rst_example.search(docstring) is not None:
        return True
    return False


def check_docstrings_are_in_md():
    """Check all docstrings are in md"""
    files_with_rst = []
    for file in Path(PATH_TO_TRANSFORMERS).glob("**/*.py"):
        with open(file, "r") as f:
            code = f.read()
        docstrings = code.split('"""')

        for idx, docstring in enumerate(docstrings):
            if idx % 2 == 0 or not is_rst_docstring(docstring):
                continue
            files_with_rst.append(file)
            break

    if len(files_with_rst) > 0:
        raise ValueError(
            "The following files have docstrings written in rst:\n"
            + "\n".join([f"- {f}" for f in files_with_rst])
Sylvain Gugger's avatar
Sylvain Gugger committed
640
            + "To fix this run `doc-builder convert path_to_py_file` after installing `doc-builder`\n"
641
642
            "(`pip install git+https://github.com/huggingface/doc-builder`)"
        )
643
644


645
def check_repo_quality():
Patrick von Platen's avatar
Patrick von Platen committed
646
    """Check all models are properly tested and documented."""
647
648
    print("Checking all models are included.")
    check_model_list()
649
650
    print("Checking all models are public.")
    check_models_are_in_init()
651
    print("Checking all models are properly tested.")
Sylvain Gugger's avatar
Sylvain Gugger committed
652
    check_all_decorator_order()
653
    check_all_models_are_tested()
654
    print("Checking all objects are properly documented.")
655
    check_all_objects_are_documented()
656
657
    print("Checking all models are in at least one auto class.")
    check_all_models_are_auto_configured()
658
659
660
661


if __name__ == "__main__":
    check_repo_quality()