check_repo.py 31.8 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 collections import OrderedDict
22
from difflib import get_close_matches
23
from pathlib import Path
24

25
from transformers import is_flax_available, is_tf_available, is_torch_available
26
from transformers.models.auto import get_values
27
from transformers.utils import ENV_VARS_TRUE_VALUES
28

29
30
31
32
33

# 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"
34
PATH_TO_DOC = "docs/source/en"
35

36
37
38
# Update this list with models that are supposed to be private.
PRIVATE_MODELS = [
    "DPRSpanPredictor",
Daniel Stancl's avatar
Daniel Stancl committed
39
    "LongT5Stack",
Li-Huai (Allan) Lin's avatar
Li-Huai (Allan) Lin committed
40
    "RealmBertModel",
41
42
43
44
    "T5Stack",
    "TFDPRSpanPredictor",
]

45
46
# 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.
47
IGNORE_NON_TESTED = PRIVATE_MODELS.copy() + [
48
    # models to ignore for not tested
NielsRogge's avatar
NielsRogge committed
49
50
    "DeformableDetrEncoder",  # Building part of bigger (tested) model.
    "DeformableDetrDecoder",  # Building part of bigger (tested) model.
Younes Belkada's avatar
Younes Belkada committed
51
    "OPTDecoder",  # Building part of bigger (tested) model.
52
    "DecisionTransformerGPT2Model",  # Building part of bigger (tested) model.
NielsRogge's avatar
NielsRogge committed
53
    "SegformerDecodeHead",  # Building part of bigger (tested) model.
Gunjan Chhablani's avatar
Gunjan Chhablani committed
54
55
56
    "PLBartEncoder",  # Building part of bigger (tested) model.
    "PLBartDecoder",  # Building part of bigger (tested) model.
    "PLBartDecoderWrapper",  # Building part of bigger (tested) model.
Vasudev Gupta's avatar
Vasudev Gupta committed
57
58
59
    "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
60
61
62
    "DetrEncoder",  # Building part of bigger (tested) model.
    "DetrDecoder",  # Building part of bigger (tested) model.
    "DetrDecoderWrapper",  # Building part of bigger (tested) model.
63
64
    "ConditionalDetrEncoder",  # Building part of bigger (tested) model.
    "ConditionalDetrDecoder",  # Building part of bigger (tested) model.
Suraj Patil's avatar
Suraj Patil committed
65
66
    "M2M100Encoder",  # Building part of bigger (tested) model.
    "M2M100Decoder",  # Building part of bigger (tested) model.
Chan Woo Kim's avatar
Chan Woo Kim committed
67
    "MCTCTEncoder",  # Building part of bigger (tested) model.
Suraj Patil's avatar
Suraj Patil committed
68
69
    "Speech2TextEncoder",  # Building part of bigger (tested) model.
    "Speech2TextDecoder",  # Building part of bigger (tested) model.
Patrick von Platen's avatar
Patrick von Platen committed
70
71
    "LEDEncoder",  # Building part of bigger (tested) model.
    "LEDDecoder",  # Building part of bigger (tested) model.
72
    "BartDecoderWrapper",  # Building part of bigger (tested) model.
73
    "BartEncoder",  # Building part of bigger (tested) model.
74
    "BertLMHeadModel",  # Needs to be setup as decoder.
75
    "BlenderbotSmallEncoder",  # Building part of bigger (tested) model.
76
    "BlenderbotSmallDecoderWrapper",  # Building part of bigger (tested) model.
77
    "BlenderbotEncoder",  # Building part of bigger (tested) model.
78
    "BlenderbotDecoderWrapper",  # Building part of bigger (tested) model.
79
    "MBartEncoder",  # Building part of bigger (tested) model.
80
    "MBartDecoderWrapper",  # Building part of bigger (tested) model.
81
82
83
84
    "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.
StevenTang1998's avatar
StevenTang1998 committed
85
86
    "MvpDecoderWrapper",  # Building part of bigger (tested) model.
    "MvpEncoder",  # Building part of bigger (tested) model.
87
    "PegasusEncoder",  # Building part of bigger (tested) model.
88
    "PegasusDecoderWrapper",  # Building part of bigger (tested) model.
Jason Phang's avatar
Jason Phang committed
89
90
91
    "PegasusXEncoder",  # Building part of bigger (tested) model.
    "PegasusXDecoder",  # Building part of bigger (tested) model.
    "PegasusXDecoderWrapper",  # Building part of bigger (tested) model.
92
    "DPREncoder",  # Building part of bigger (tested) model.
93
    "ProphetNetDecoderWrapper",  # Building part of bigger (tested) model.
Li-Huai (Allan) Lin's avatar
Li-Huai (Allan) Lin committed
94
95
96
97
    "RealmBertModel",  # Building part of bigger (tested) model.
    "RealmReader",  # Not regular model.
    "RealmScorer",  # Not regular model.
    "RealmForOpenQA",  # Not regular model.
98
    "ReformerForMaskedLM",  # Needs to be setup as decoder.
99
    "Speech2Text2DecoderWrapper",  # Building part of bigger (tested) model.
Ratthachat (Jung)'s avatar
Ratthachat (Jung) committed
100
    "TFDPREncoder",  # Building part of bigger (tested) model.
101
102
    "TFElectraMainLayer",  # Building part of bigger (tested) model (should it be a TFPreTrainedModel ?)
    "TFRobertaForMultipleChoice",  # TODO: fix
103
    "TrOCRDecoderWrapper",  # Building part of bigger (tested) model.
abhishek thakur's avatar
abhishek thakur committed
104
    "SeparableConv1D",  # Building part of bigger (tested) model.
105
    "FlaxBartForCausalLM",  # Building part of bigger (tested) model.
106
    "FlaxBertForCausalLM",  # Building part of bigger (tested) model. Tested implicitly through FlaxRobertaForCausalLM.
Younes Belkada's avatar
Younes Belkada committed
107
    "OPTDecoderWrapper",
108
    "TFSegformerDecodeHead",  # Not a regular model.
109
110
111
112
113
]

# 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 = [
Yih-Dar's avatar
Yih-Dar committed
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
    "models/decision_transformer/test_modeling_decision_transformer.py",
    "models/camembert/test_modeling_camembert.py",
    "models/mt5/test_modeling_flax_mt5.py",
    "models/mbart/test_modeling_mbart.py",
    "models/mt5/test_modeling_mt5.py",
    "models/pegasus/test_modeling_pegasus.py",
    "models/camembert/test_modeling_tf_camembert.py",
    "models/mt5/test_modeling_tf_mt5.py",
    "models/xlm_roberta/test_modeling_tf_xlm_roberta.py",
    "models/xlm_roberta/test_modeling_flax_xlm_roberta.py",
    "models/xlm_prophetnet/test_modeling_xlm_prophetnet.py",
    "models/xlm_roberta/test_modeling_xlm_roberta.py",
    "models/vision_text_dual_encoder/test_modeling_vision_text_dual_encoder.py",
    "models/vision_text_dual_encoder/test_modeling_flax_vision_text_dual_encoder.py",
    "models/decision_transformer/test_modeling_decision_transformer.py",
129
130
]

131
132
# 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.
133
IGNORE_NON_AUTO_CONFIGURED = PRIVATE_MODELS.copy() + [
134
    # models to ignore for model xxx mapping
Jason Phang's avatar
Jason Phang committed
135
136
137
138
139
140
    "PegasusXEncoder",
    "PegasusXDecoder",
    "PegasusXDecoderWrapper",
    "PegasusXEncoder",
    "PegasusXDecoder",
    "PegasusXDecoderWrapper",
NielsRogge's avatar
NielsRogge committed
141
    "DPTForDepthEstimation",
142
    "DecisionTransformerGPT2Model",
NielsRogge's avatar
NielsRogge committed
143
    "GLPNForDepthEstimation",
NielsRogge's avatar
NielsRogge committed
144
145
    "ViltForImagesAndTextClassification",
    "ViltForImageAndTextRetrieval",
146
    "ViltForTokenClassification",
NielsRogge's avatar
NielsRogge committed
147
    "ViltForMaskedLM",
Suraj Patil's avatar
Suraj Patil committed
148
149
150
    "XGLMEncoder",
    "XGLMDecoder",
    "XGLMDecoderWrapper",
NielsRogge's avatar
NielsRogge committed
151
152
    "PerceiverForMultimodalAutoencoding",
    "PerceiverForOpticalFlow",
NielsRogge's avatar
NielsRogge committed
153
    "SegformerDecodeHead",
154
    "TFSegformerDecodeHead",
Kamal Raj's avatar
Kamal Raj committed
155
    "FlaxBeitForMaskedImageModeling",
Gunjan Chhablani's avatar
Gunjan Chhablani committed
156
157
158
    "PLBartEncoder",
    "PLBartDecoder",
    "PLBartDecoderWrapper",
NielsRogge's avatar
NielsRogge committed
159
    "BeitForMaskedImageModeling",
Suraj Patil's avatar
Suraj Patil committed
160
161
    "CLIPTextModel",
    "CLIPVisionModel",
162
163
    "GroupViTTextModel",
    "GroupViTVisionModel",
Yih-Dar's avatar
Yih-Dar committed
164
165
    "TFCLIPTextModel",
    "TFCLIPVisionModel",
Suraj Patil's avatar
Suraj Patil committed
166
167
    "FlaxCLIPTextModel",
    "FlaxCLIPVisionModel",
168
    "FlaxWav2Vec2ForCTC",
NielsRogge's avatar
NielsRogge committed
169
    "DetrForSegmentation",
170
    "ConditionalDetrForSegmentation",
171
172
    "DPRReader",
    "FlaubertForQuestionAnswering",
173
174
175
176
    "FlavaImageCodebook",
    "FlavaTextModel",
    "FlavaImageModel",
    "FlavaMultimodalModel",
177
    "GPT2DoubleHeadsModel",
178
    "LayoutLMForQuestionAnswering",
Ryokan RI's avatar
Ryokan RI committed
179
    "LukeForMaskedLM",
NielsRogge's avatar
NielsRogge committed
180
181
182
    "LukeForEntityClassification",
    "LukeForEntityPairClassification",
    "LukeForEntitySpanClassification",
183
    "OpenAIGPTDoubleHeadsModel",
184
185
186
    "OwlViTTextModel",
    "OwlViTVisionModel",
    "OwlViTForObjectDetection",
187
188
189
    "RagModel",
    "RagSequenceForGeneration",
    "RagTokenForGeneration",
Li-Huai (Allan) Lin's avatar
Li-Huai (Allan) Lin committed
190
191
192
193
    "RealmEmbedder",
    "RealmForOpenQA",
    "RealmScorer",
    "RealmReader",
Ratthachat (Jung)'s avatar
Ratthachat (Jung) committed
194
    "TFDPRReader",
195
    "TFGPT2DoubleHeadsModel",
196
    "TFLayoutLMForQuestionAnswering",
197
    "TFOpenAIGPTDoubleHeadsModel",
Ratthachat (Jung)'s avatar
Ratthachat (Jung) committed
198
199
200
    "TFRagModel",
    "TFRagSequenceForGeneration",
    "TFRagTokenForGeneration",
201
    "Wav2Vec2ForCTC",
Patrick von Platen's avatar
Patrick von Platen committed
202
    "HubertForCTC",
203
204
    "SEWForCTC",
    "SEWDForCTC",
205
206
    "XLMForQuestionAnswering",
    "XLNetForQuestionAnswering",
abhishek thakur's avatar
abhishek thakur committed
207
    "SeparableConv1D",
Gunjan Chhablani's avatar
Gunjan Chhablani committed
208
209
210
211
    "VisualBertForRegionToPhraseAlignment",
    "VisualBertForVisualReasoning",
    "VisualBertForQuestionAnswering",
    "VisualBertForMultipleChoice",
Will Rice's avatar
Will Rice committed
212
    "TFWav2Vec2ForCTC",
Will Rice's avatar
Will Rice committed
213
    "TFHubertForCTC",
214
    "MaskFormerForInstanceSegmentation",
NielsRogge's avatar
NielsRogge committed
215
216
    "XCLIPVisionModel",
    "XCLIPTextModel",
217
218
]

219
220
221
222
223
224
# Update this list for models that have multiple model types for the same
# model doc
MODEL_TYPE_TO_DOC_MAPPING = OrderedDict(
    [
        ("data2vec-text", "data2vec"),
        ("data2vec-audio", "data2vec"),
225
        ("data2vec-vision", "data2vec"),
NielsRogge's avatar
NielsRogge committed
226
        ("donut-swin", "donut"),
227
228
229
230
    ]
)


231
232
233
234
235
236
237
238
239
# 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()


240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
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)}."
        )


260
261
262
# 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
263
    """Get the model modules inside the transformers library."""
264
265
266
267
268
269
270
271
    _ignore_modules = [
        "modeling_auto",
        "modeling_encoder_decoder",
        "modeling_marian",
        "modeling_mmbt",
        "modeling_outputs",
        "modeling_retribert",
        "modeling_utils",
Sylvain Gugger's avatar
Sylvain Gugger committed
272
        "modeling_flax_auto",
273
        "modeling_flax_encoder_decoder",
Stas Bekman's avatar
Stas Bekman committed
274
        "modeling_flax_utils",
275
        "modeling_speech_encoder_decoder",
276
        "modeling_flax_speech_encoder_decoder",
277
        "modeling_flax_vision_encoder_decoder",
278
279
        "modeling_transfo_xl_utilities",
        "modeling_tf_auto",
280
        "modeling_tf_encoder_decoder",
281
282
283
284
        "modeling_tf_outputs",
        "modeling_tf_pytorch_utils",
        "modeling_tf_utils",
        "modeling_tf_transfo_xl_utilities",
285
        "modeling_tf_vision_encoder_decoder",
286
        "modeling_vision_encoder_decoder",
287
288
    ]
    modules = []
Sylvain Gugger's avatar
Sylvain Gugger committed
289
290
291
292
293
294
295
296
297
    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)
298
299
300
    return modules


301
def get_models(module, include_pretrained=False):
Patrick von Platen's avatar
Patrick von Platen committed
302
    """Get the objects in module that are models."""
303
    models = []
304
    model_classes = (transformers.PreTrainedModel, transformers.TFPreTrainedModel, transformers.FlaxPreTrainedModel)
305
    for attr_name in dir(module):
306
        if not include_pretrained and ("Pretrained" in attr_name or "PreTrained" in attr_name):
307
308
309
310
311
312
313
            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


314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
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)}.")


344
345
346
# 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():
Yih-Dar's avatar
Yih-Dar committed
347
348
349
350
351
352
    """Get the model test files.

    The returned files should NOT contain the `tests` (i.e. `PATH_TO_TESTS` defined in this script). They will be
    considered as paths relative to `tests`. A caller has to use `os.path.join(PATH_TO_TESTS, ...)` to access the files.
    """

353
354
355
    _ignore_files = [
        "test_modeling_common",
        "test_modeling_encoder_decoder",
356
        "test_modeling_flax_encoder_decoder",
357
        "test_modeling_flax_speech_encoder_decoder",
358
359
        "test_modeling_marian",
        "test_modeling_tf_common",
360
        "test_modeling_tf_encoder_decoder",
361
362
    ]
    test_files = []
Yih-Dar's avatar
Yih-Dar committed
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
    # Check both `PATH_TO_TESTS` and `PATH_TO_TESTS/models`
    model_test_root = os.path.join(PATH_TO_TESTS, "models")
    model_test_dirs = []
    for x in os.listdir(model_test_root):
        x = os.path.join(model_test_root, x)
        if os.path.isdir(x):
            model_test_dirs.append(x)

    for target_dir in [PATH_TO_TESTS] + model_test_dirs:
        for file_or_dir in os.listdir(target_dir):
            path = os.path.join(target_dir, file_or_dir)
            if os.path.isfile(path):
                filename = os.path.split(path)[-1]
                if "test_modeling" in filename and not os.path.splitext(filename)[0] in _ignore_files:
                    file = os.path.join(*path.split(os.sep)[1:])
                    test_files.append(file)

380
381
382
383
384
385
    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
386
    """Parse the content of test_file to detect what's in all_model_classes"""
Sylvain Gugger's avatar
Sylvain Gugger committed
387
    # 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
388
    with open(os.path.join(PATH_TO_TESTS, test_file), "r", encoding="utf-8", newline="\n") as f:
389
        content = f.read()
Sylvain Gugger's avatar
Sylvain Gugger committed
390
    all_models = re.findall(r"all_model_classes\s+=\s+\(\s*\(([^\)]*)\)", content)
391
392
    # 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
393
    if len(all_models) > 0:
394
        model_tested = []
Sylvain Gugger's avatar
Sylvain Gugger committed
395
396
397
398
399
        for entry in all_models:
            for line in entry.split(","):
                name = line.strip()
                if len(name) > 0:
                    model_tested.append(name)
400
401
402
403
        return model_tested


def check_models_are_tested(module, test_file):
Patrick von Platen's avatar
Patrick von Platen committed
404
    """Check models defined in module are tested in test_file."""
405
    # XxxPreTrainedModel are not tested
406
407
408
    defined_models = get_models(module)
    tested_models = find_tested_models(test_file)
    if tested_models is None:
409
        if test_file.replace(os.path.sep, "/") in TEST_FILES_WITH_NO_COMMON_TESTS:
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
            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
429
    """Check all models are properly tested."""
430
431
432
433
    modules = get_model_modules()
    test_files = get_model_test_files()
    failures = []
    for module in modules:
434
435
        test_file = [file for file in test_files if f"test_{module.__name__.split('.')[-1]}.py" in file]
        if len(test_file) == 0:
436
            failures.append(f"{module.__name__} does not have its corresponding test file {test_file}.")
437
438
439
440
        elif len(test_file) > 1:
            failures.append(f"{module.__name__} has several test files: {test_file}.")
        else:
            test_file = test_file[0]
441
442
443
            new_failures = check_models_are_tested(module, test_file)
            if new_failures is not None:
                failures += new_failures
444
445
446
447
    if len(failures) > 0:
        raise Exception(f"There were {len(failures)} failures:\n" + "\n".join(failures))


448
def get_all_auto_configured_models():
Patrick von Platen's avatar
Patrick von Platen committed
449
    """Return the list of all models in at least one auto class."""
450
    result = set()  # To avoid duplicates we concatenate all model classes in a set.
451
452
    if is_torch_available():
        for attr_name in dir(transformers.models.auto.modeling_auto):
453
            if attr_name.startswith("MODEL_") and attr_name.endswith("MAPPING_NAMES"):
454
455
456
                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):
457
            if attr_name.startswith("TF_MODEL_") and attr_name.endswith("MAPPING_NAMES"):
458
459
460
                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):
461
            if attr_name.startswith("FLAX_MODEL_") and attr_name.endswith("MAPPING_NAMES"):
462
                result = result | set(get_values(getattr(transformers.models.auto.modeling_flax_auto, attr_name)))
463
    return [cls for cls in result]
464
465


466
467
468
469
470
471
472
473
474
475
476
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


477
def check_models_are_auto_configured(module, all_auto_models):
Patrick von Platen's avatar
Patrick von Platen committed
478
    """Check models defined in module are each in an auto class."""
479
480
481
    defined_models = get_models(module)
    failures = []
    for model_name, _ in defined_models:
482
        if model_name not in all_auto_models and not ignore_unautoclassed(model_name):
483
484
485
486
487
488
489
490
491
            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
492
    """Check all models are each in an auto class."""
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
    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."
            )
514
515
516
517
518
519
520
521
522
523
524
    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
525
526
527
528
_re_decorator = re.compile(r"^\s*@(\S+)\s+$")


def check_decorator_order(filename):
Patrick von Platen's avatar
Patrick von Platen committed
529
    """Check that in the test file `filename` the slow decorator is always last."""
530
    with open(filename, "r", encoding="utf-8", newline="\n") as f:
Sylvain Gugger's avatar
Sylvain Gugger committed
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
        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
547
    """Check that in all test files, the slow decorator is always last."""
Sylvain Gugger's avatar
Sylvain Gugger committed
548
549
550
551
552
553
554
555
556
    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(
Sylvain Gugger's avatar
Sylvain Gugger committed
557
558
            "The parameterized decorator (and its variants) should always be first, but this is not the case in the"
            f" following files:\n{msg}"
Sylvain Gugger's avatar
Sylvain Gugger committed
559
560
561
        )


562
def find_all_documented_objects():
Patrick von Platen's avatar
Patrick von Platen committed
563
    """Parse the content of all doc files to detect which classes and functions it documents"""
564
565
    documented_obj = []
    for doc_file in Path(PATH_TO_DOC).glob("**/*.rst"):
Julien Plu's avatar
Julien Plu committed
566
        with open(doc_file, "r", encoding="utf-8", newline="\n") as f:
567
568
569
            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
570
571
572
573
574
    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]
575
576
577
578
579
580
    return documented_obj


# One good reason for not being documented is to be deprecated. Put in this list deprecated objects.
DEPRECATED_OBJECTS = [
    "AutoModelWithLMHead",
581
    "BartPretrainedModel",
582
583
    "DataCollator",
    "DataCollatorForSOP",
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
    "GlueDataset",
    "GlueDataTrainingArguments",
    "LineByLineTextDataset",
    "LineByLineWithRefDataset",
    "LineByLineWithSOPTextDataset",
    "PretrainedBartModel",
    "PretrainedFSMTModel",
    "SingleSentenceClassificationProcessor",
    "SquadDataTrainingArguments",
    "SquadDataset",
    "SquadExample",
    "SquadFeatures",
    "SquadV1Processor",
    "SquadV2Processor",
    "TFAutoModelWithLMHead",
599
    "TFBartPretrainedModel",
600
601
    "TextDataset",
    "TextDatasetForNextSentencePrediction",
602
    "Wav2Vec2ForMaskedLM",
603
    "Wav2Vec2Tokenizer",
604
605
606
607
608
609
610
611
612
613
    "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",
614
615
    "TFTrainer",
    "TFTrainingArguments",
616
617
618
619
620
621
622
]

# 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.
623
    "CharacterTokenizer",  # Internal, should never have been in the main init.
624
    "DPRPretrainedReader",  # Like an Encoder.
Sylvain Gugger's avatar
Sylvain Gugger committed
625
    "DummyObject",  # Just picked by mistake sometimes.
626
    "MecabTokenizer",  # Internal, should never have been in the main init.
627
628
629
630
631
632
633
634
635
636
637
    "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.
    "convert_tf_weight_name_to_pt_weight_name",  # Internal used to convert model weights
    "logger",  # Internal logger
    "logging",  # External module
638
    "requires_backends",  # Internal function
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
]

# 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
690
    """Check all models are properly documented."""
691
    documented_objs = find_all_documented_objects()
692
693
694
    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)]
695
696
697
698
699
    if len(undocumented_objs) > 0:
        raise Exception(
            "The following objects are in the public init so should be documented:\n - "
            + "\n - ".join(undocumented_objs)
        )
700
    check_docstrings_are_in_md()
701
702
703
704
705
706
707
708
709
    check_model_type_doc_match()


def check_model_type_doc_match():
    """Check all doc pages have a corresponding model type."""
    model_doc_folder = Path(PATH_TO_DOC) / "model_doc"
    model_docs = [m.stem for m in model_doc_folder.glob("*.mdx")]

    model_types = list(transformers.models.auto.configuration_auto.MODEL_NAMES_MAPPING.keys())
710
    model_types = [MODEL_TYPE_TO_DOC_MAPPING[m] if m in MODEL_TYPE_TO_DOC_MAPPING else m for m in model_types]
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728

    errors = []
    for m in model_docs:
        if m not in model_types and m != "auto":
            close_matches = get_close_matches(m, model_types)
            error_message = f"{m} is not a proper model identifier."
            if len(close_matches) > 0:
                close_matches = "/".join(close_matches)
                error_message += f" Did you mean {close_matches}?"
            errors.append(error_message)

    if len(errors) > 0:
        raise ValueError(
            "Some model doc pages do not match any existing model type:\n"
            + "\n".join(errors)
            + "\nYou can add any missing model type to the `MODEL_NAMES_MAPPING` constant in "
            "models/auto/configuration_auto.py."
        )
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755


# 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"):
756
        with open(file, encoding="utf-8") as f:
757
758
759
760
761
762
763
764
765
766
767
768
769
            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])
Kamal Raj's avatar
Kamal Raj committed
770
            + "\nTo fix this run `doc-builder convert path_to_py_file` after installing `doc-builder`\n"
771
772
            "(`pip install git+https://github.com/huggingface/doc-builder`)"
        )
773
774


775
def check_repo_quality():
Patrick von Platen's avatar
Patrick von Platen committed
776
    """Check all models are properly tested and documented."""
777
778
    print("Checking all models are included.")
    check_model_list()
779
780
    print("Checking all models are public.")
    check_models_are_in_init()
781
    print("Checking all models are properly tested.")
Sylvain Gugger's avatar
Sylvain Gugger committed
782
    check_all_decorator_order()
783
    check_all_models_are_tested()
784
    print("Checking all objects are properly documented.")
785
    check_all_objects_are_documented()
786
787
    print("Checking all models are in at least one auto class.")
    check_all_models_are_auto_configured()
788
789
790
791


if __name__ == "__main__":
    check_repo_quality()