test_modeling_common.py 54.8 KB
Newer Older
thomwolf's avatar
thomwolf committed
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
import copy
17
import gc
18
import inspect
19
import os.path
Aymeric Augustin's avatar
Aymeric Augustin committed
20
import random
21
import tempfile
thomwolf's avatar
thomwolf committed
22
import unittest
23
from typing import List, Tuple
thomwolf's avatar
thomwolf committed
24

25
from transformers import is_torch_available
26
from transformers.file_utils import WEIGHTS_NAME
27
from transformers.testing_utils import require_torch, require_torch_multi_gpu, slow, torch_device
28

Aymeric Augustin's avatar
Aymeric Augustin committed
29

30
if is_torch_available():
31
    import numpy as np
32
    import torch
thomwolf's avatar
thomwolf committed
33

34
    from transformers import (
35
        BERT_PRETRAINED_MODEL_ARCHIVE_LIST,
36
37
        MODEL_FOR_CAUSAL_LM_MAPPING,
        MODEL_FOR_MASKED_LM_MAPPING,
38
        MODEL_FOR_MULTIPLE_CHOICE_MAPPING,
39
        MODEL_FOR_NEXT_SENTENCE_PREDICTION_MAPPING,
40
        MODEL_FOR_QUESTION_ANSWERING_MAPPING,
41
42
43
        MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING,
        MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING,
        MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING,
44
        MODEL_MAPPING,
45
46
47
48
49
        AdaptiveEmbedding,
        BertConfig,
        BertModel,
        PretrainedConfig,
        PreTrainedModel,
50
    )
thomwolf's avatar
thomwolf committed
51

52

53
54
55
def _config_zero_init(config):
    configs_no_init = copy.deepcopy(config)
    for key in configs_no_init.__dict__.keys():
56
        if "_range" in key or "_std" in key or "initializer_factor" in key:
Lysandre Debut's avatar
Lysandre Debut committed
57
            setattr(configs_no_init, key, 1e-10)
58
59
    return configs_no_init

thomwolf's avatar
thomwolf committed
60

61
62
63
64
65
@require_torch
class ModelTesterMixin:

    model_tester = None
    all_model_classes = ()
66
    all_generative_model_classes = ()
Patrick von Platen's avatar
Patrick von Platen committed
67
68
69
70
    test_torchscript = True
    test_pruning = True
    test_resize_embeddings = True
    test_head_masking = True
71
    test_missing_keys = True
72
    test_model_parallel = False
73
74
    is_encoder_decoder = False

75
76
    def _prepare_for_class(self, inputs_dict, model_class, return_labels=False):
        inputs_dict = copy.deepcopy(inputs_dict)
77
        if model_class in MODEL_FOR_MULTIPLE_CHOICE_MAPPING.values():
78
            inputs_dict = {
79
                k: v.unsqueeze(1).expand(-1, self.model_tester.num_choices, -1).contiguous()
80
                if isinstance(v, torch.Tensor) and v.ndim > 1
Sylvain Gugger's avatar
Sylvain Gugger committed
81
                else v
82
83
                for k, v in inputs_dict.items()
            }
84
85
86
87
88
89
90
91
92
93
94

        if return_labels:
            if model_class in MODEL_FOR_MULTIPLE_CHOICE_MAPPING.values():
                inputs_dict["labels"] = torch.ones(self.model_tester.batch_size, dtype=torch.long, device=torch_device)
            elif model_class in MODEL_FOR_QUESTION_ANSWERING_MAPPING.values():
                inputs_dict["start_positions"] = torch.zeros(
                    self.model_tester.batch_size, dtype=torch.long, device=torch_device
                )
                inputs_dict["end_positions"] = torch.zeros(
                    self.model_tester.batch_size, dtype=torch.long, device=torch_device
                )
95
96
97
98
            elif model_class in [
                *MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING.values(),
                *MODEL_FOR_NEXT_SENTENCE_PREDICTION_MAPPING.values(),
            ]:
99
100
101
102
103
104
105
106
107
108
109
110
                inputs_dict["labels"] = torch.zeros(
                    self.model_tester.batch_size, dtype=torch.long, device=torch_device
                )
            elif model_class in [
                *MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING.values(),
                *MODEL_FOR_CAUSAL_LM_MAPPING.values(),
                *MODEL_FOR_MASKED_LM_MAPPING.values(),
                *MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING.values(),
            ]:
                inputs_dict["labels"] = torch.zeros(
                    (self.model_tester.batch_size, self.model_tester.seq_length), dtype=torch.long, device=torch_device
                )
111
112
        return inputs_dict

Patrick von Platen's avatar
Patrick von Platen committed
113
    def test_save_load(self):
114
115
116
117
118
119
120
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            model = model_class(config)
            model.to(torch_device)
            model.eval()
            with torch.no_grad():
121
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
Weizhen's avatar
Weizhen committed
122

123
            out_2 = outputs[0].cpu().numpy()
124
            out_2[np.isnan(out_2)] = 0
125

126
            with tempfile.TemporaryDirectory() as tmpdirname:
127
128
                model.save_pretrained(tmpdirname)
                model = model_class.from_pretrained(tmpdirname)
129
                model.to(torch_device)
130
                with torch.no_grad():
131
                    after_outputs = model(**self._prepare_for_class(inputs_dict, model_class))
thomwolf's avatar
thomwolf committed
132

133
134
135
                # Make sure we don't have nans
                out_1 = after_outputs[0].cpu().numpy()
                out_1[np.isnan(out_1)] = 0
thomwolf's avatar
thomwolf committed
136
137
                max_diff = np.amax(np.abs(out_1 - out_2))
                self.assertLessEqual(max_diff, 1e-5)
138

139
    def test_save_load__keys_to_ignore_on_save(self):
140
141
142
143
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            model = model_class(config)
144
145
            _keys_to_ignore_on_save = getattr(model, "_keys_to_ignore_on_save", None)
            if _keys_to_ignore_on_save is None:
146
147
148
                continue

            # check the keys are in the original state_dict
149
            for k in _keys_to_ignore_on_save:
150
151
152
153
154
155
156
                self.assertIn(k, model.state_dict())

            # check that certain keys didn't get saved with the model
            with tempfile.TemporaryDirectory() as tmpdirname:
                model.save_pretrained(tmpdirname)
                output_model_file = os.path.join(tmpdirname, WEIGHTS_NAME)
                state_dict_saved = torch.load(output_model_file)
157
                for k in _keys_to_ignore_on_save:
158
159
                    self.assertNotIn(k, state_dict_saved)

Patrick von Platen's avatar
Patrick von Platen committed
160
    def test_initialization(self):
161
162
163
164
165
166
167
168
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        configs_no_init = _config_zero_init(config)
        for model_class in self.all_model_classes:
            model = model_class(config=configs_no_init)
            for name, param in model.named_parameters():
                if param.requires_grad:
                    self.assertIn(
Lysandre Debut's avatar
Lysandre Debut committed
169
                        ((param.data.mean() * 1e9).round() / 1e9).item(),
170
171
172
                        [0.0, 1.0],
                        msg="Parameter {} of model {} seems not properly initialized".format(name, model_class),
                    )
thomwolf's avatar
thomwolf committed
173

Patrick von Platen's avatar
Patrick von Platen committed
174
    def test_determinism(self):
175
176
177
178
179
180
181
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            model = model_class(config)
            model.to(torch_device)
            model.eval()
            with torch.no_grad():
182
183
                first = model(**self._prepare_for_class(inputs_dict, model_class))[0]
                second = model(**self._prepare_for_class(inputs_dict, model_class))[0]
Weizhen's avatar
Weizhen committed
184

185
186
187
188
189
190
191
            out_1 = first.cpu().numpy()
            out_2 = second.cpu().numpy()
            out_1 = out_1[~np.isnan(out_1)]
            out_2 = out_2[~np.isnan(out_2)]
            max_diff = np.amax(np.abs(out_1 - out_2))
            self.assertLessEqual(max_diff, 1e-5)

192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
    def test_forward_signature(self):
        config, _ = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            model = model_class(config)
            signature = inspect.signature(model.forward)
            # signature.parameters is an OrderedDict => so arg_names order is deterministic
            arg_names = [*signature.parameters.keys()]

            if model.config.is_encoder_decoder:
                expected_arg_names = [
                    "input_ids",
                    "attention_mask",
                    "decoder_input_ids",
                    "decoder_attention_mask",
                ]
208
209
210
211
212
213
                expected_arg_names.extend(
                    ["head_mask", "decoder_head_mask", "encoder_outputs"]
                    if "head_mask" and "decoder_head_mask" in arg_names
                    else ["encoder_outputs"]
                )
                self.assertListEqual(arg_names[: len(expected_arg_names)], expected_arg_names)
214
215
216
217
            else:
                expected_arg_names = ["input_ids"]
                self.assertListEqual(arg_names[:1], expected_arg_names)

218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
    def test_training(self):
        if not self.model_tester.is_training:
            return

        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
        config.return_dict = True

        for model_class in self.all_model_classes:
            if model_class in MODEL_MAPPING.values():
                continue
            model = model_class(config)
            model.to(torch_device)
            model.train()
            inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
            loss = model(**inputs).loss
            loss.backward()

    def test_training_gradient_checkpointing(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
        if not self.model_tester.is_training or not hasattr(config, "gradient_checkpointing"):
            return

        config.gradient_checkpointing = True
241
        config.use_cache = False
242
243
244
245
246
247
248
249
250
251
252
253
        config.return_dict = True

        for model_class in self.all_model_classes:
            if model_class in MODEL_MAPPING.values():
                continue
            model = model_class(config)
            model.to(torch_device)
            model.train()
            inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
            loss = model(**inputs).loss
            loss.backward()

Patrick von Platen's avatar
Patrick von Platen committed
254
    def test_attention_outputs(self):
255
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
Weizhen's avatar
Weizhen committed
256
257
        config.return_dict = True

sshleifer's avatar
sshleifer committed
258
        seq_len = getattr(self.model_tester, "seq_length", None)
sshleifer's avatar
sshleifer committed
259
260
        decoder_seq_length = getattr(self.model_tester, "decoder_seq_length", seq_len)
        encoder_seq_length = getattr(self.model_tester, "encoder_seq_length", seq_len)
Weizhen's avatar
Weizhen committed
261
        decoder_key_length = getattr(self.model_tester, "decoder_key_length", decoder_seq_length)
262
        encoder_key_length = getattr(self.model_tester, "key_length", encoder_seq_length)
Patrick von Platen's avatar
Patrick von Platen committed
263
264
265
        chunk_length = getattr(self.model_tester, "chunk_length", None)
        if chunk_length is not None and hasattr(self.model_tester, "num_hashes"):
            encoder_seq_length = encoder_seq_length * self.model_tester.num_hashes
266
267

        for model_class in self.all_model_classes:
268
            inputs_dict["output_attentions"] = True
Joseph Liu's avatar
Joseph Liu committed
269
            inputs_dict["output_hidden_states"] = False
270
            config.return_dict = True
271
272
273
274
            model = model_class(config)
            model.to(torch_device)
            model.eval()
            with torch.no_grad():
275
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
276
            attentions = outputs.encoder_attentions if config.is_encoder_decoder else outputs.attentions
277
278
279
280
281
282
283
284
285
            self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)

            # check that output_attentions also work using config
            del inputs_dict["output_attentions"]
            config.output_attentions = True
            model = model_class(config)
            model.to(torch_device)
            model.eval()
            with torch.no_grad():
286
287
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
            attentions = outputs.encoder_attentions if config.is_encoder_decoder else outputs.attentions
288
            self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)
Patrick von Platen's avatar
Patrick von Platen committed
289
290
291
292
293
294
295
296
297
298
299

            if chunk_length is not None:
                self.assertListEqual(
                    list(attentions[0].shape[-4:]),
                    [self.model_tester.num_attention_heads, encoder_seq_length, chunk_length, encoder_key_length],
                )
            else:
                self.assertListEqual(
                    list(attentions[0].shape[-3:]),
                    [self.model_tester.num_attention_heads, encoder_seq_length, encoder_key_length],
                )
300
            out_len = len(outputs)
thomwolf's avatar
thomwolf committed
301

302
            if self.is_encoder_decoder:
303
                correct_outlen = 5
304

305
306
307
308
309
310
                # loss is at first position
                if "labels" in inputs_dict:
                    correct_outlen += 1  # loss is added to beginning
                # Question Answering model returns start_logits and end_logits
                if model_class in MODEL_FOR_QUESTION_ANSWERING_MAPPING.values():
                    correct_outlen += 1  # start_logits and end_logits instead of only 1 output
311
312
                if "past_key_values" in outputs:
                    correct_outlen += 1  # past_key_values have been returned
Weizhen's avatar
Weizhen committed
313

Sam Shleifer's avatar
Sam Shleifer committed
314
315
                self.assertEqual(out_len, correct_outlen)

316
                # decoder attentions
317
                decoder_attentions = outputs.decoder_attentions
Sam Shleifer's avatar
Sam Shleifer committed
318
                self.assertIsInstance(decoder_attentions, (list, tuple))
319
                self.assertEqual(len(decoder_attentions), self.model_tester.num_hidden_layers)
thomwolf's avatar
thomwolf committed
320
                self.assertListEqual(
321
322
                    list(decoder_attentions[0].shape[-3:]),
                    [self.model_tester.num_attention_heads, decoder_seq_length, decoder_key_length],
323
                )
thomwolf's avatar
thomwolf committed
324

325
326
327
328
329
330
331
332
333
334
335
336
337
                # cross attentions
                cross_attentions = outputs.cross_attentions
                self.assertIsInstance(cross_attentions, (list, tuple))
                self.assertEqual(len(cross_attentions), self.model_tester.num_hidden_layers)
                self.assertListEqual(
                    list(cross_attentions[0].shape[-3:]),
                    [
                        self.model_tester.num_attention_heads,
                        decoder_seq_length,
                        encoder_key_length,
                    ],
                )

338
            # Check attention is always last and order is fine
339
            inputs_dict["output_attentions"] = True
Joseph Liu's avatar
Joseph Liu committed
340
            inputs_dict["output_hidden_states"] = True
341
342
343
344
            model = model_class(config)
            model.to(torch_device)
            model.eval()
            with torch.no_grad():
345
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
346

Weizhen's avatar
Weizhen committed
347
348
349
350
351
352
353
354
            if hasattr(self.model_tester, "num_hidden_states_types"):
                added_hidden_states = self.model_tester.num_hidden_states_types
            elif self.is_encoder_decoder:
                added_hidden_states = 2
            else:
                added_hidden_states = 1
            self.assertEqual(out_len + added_hidden_states, len(outputs))

355
356
            self_attentions = outputs.encoder_attentions if config.is_encoder_decoder else outputs.attentions

357
            self.assertEqual(len(self_attentions), self.model_tester.num_hidden_layers)
Patrick von Platen's avatar
Patrick von Platen committed
358
359
360
361
362
363
364
365
366
367
            if chunk_length is not None:
                self.assertListEqual(
                    list(self_attentions[0].shape[-4:]),
                    [self.model_tester.num_attention_heads, encoder_seq_length, chunk_length, encoder_key_length],
                )
            else:
                self.assertListEqual(
                    list(self_attentions[0].shape[-3:]),
                    [self.model_tester.num_attention_heads, encoder_seq_length, encoder_key_length],
                )
thomwolf's avatar
thomwolf committed
368

Patrick von Platen's avatar
Patrick von Platen committed
369
    def test_torchscript(self):
370
371
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
        self._create_and_check_torchscript(config, inputs_dict)
thomwolf's avatar
thomwolf committed
372

Patrick von Platen's avatar
Patrick von Platen committed
373
    def test_torchscript_output_attentions(self):
374
375
376
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
        config.output_attentions = True
        self._create_and_check_torchscript(config, inputs_dict)
thomwolf's avatar
thomwolf committed
377

Patrick von Platen's avatar
Patrick von Platen committed
378
    def test_torchscript_output_hidden_state(self):
379
380
381
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
        config.output_hidden_states = True
        self._create_and_check_torchscript(config, inputs_dict)
thomwolf's avatar
thomwolf committed
382

383
    def _create_and_check_torchscript(self, config, inputs_dict):
Patrick von Platen's avatar
Patrick von Platen committed
384
        if not self.test_torchscript:
385
            return
386

387
388
389
390
391
392
        configs_no_init = _config_zero_init(config)  # To be sure we have no Nan
        configs_no_init.torchscript = True
        for model_class in self.all_model_classes:
            model = model_class(config=configs_no_init)
            model.to(torch_device)
            model.eval()
393
            inputs = self._prepare_for_class(inputs_dict, model_class)
thomwolf's avatar
thomwolf committed
394

395
            try:
396
                if model.config.is_encoder_decoder:
397
                    model.config.use_cache = False  # FSTM still requires this hack -> FSTM should probably be refactored similar to BART afterward
398
399
400
401
402
403
404
405
406
407
                    input_ids = inputs["input_ids"]
                    attention_mask = inputs["attention_mask"]
                    decoder_input_ids = inputs["decoder_input_ids"]
                    decoder_attention_mask = inputs["decoder_attention_mask"]
                    traced_model = torch.jit.trace(
                        model, (input_ids, attention_mask, decoder_input_ids, decoder_attention_mask)
                    )
                else:
                    input_ids = inputs["input_ids"]
                    traced_model = torch.jit.trace(model, input_ids)
408
409
            except RuntimeError:
                self.fail("Couldn't trace module.")
thomwolf's avatar
thomwolf committed
410

411
            with tempfile.TemporaryDirectory() as tmp_dir_name:
412
                pt_file_name = os.path.join(tmp_dir_name, "traced_model.pt")
thomwolf's avatar
thomwolf committed
413

414
                try:
415
                    torch.jit.save(traced_model, pt_file_name)
416
417
                except Exception:
                    self.fail("Couldn't save module.")
thomwolf's avatar
thomwolf committed
418

419
420
421
422
                try:
                    loaded_model = torch.jit.load(pt_file_name)
                except Exception:
                    self.fail("Couldn't load module.")
LysandreJik's avatar
LysandreJik committed
423

424
425
            model.to(torch_device)
            model.eval()
thomwolf's avatar
thomwolf committed
426

427
428
            loaded_model.to(torch_device)
            loaded_model.eval()
thomwolf's avatar
thomwolf committed
429

430
431
432
433
            model_state_dict = model.state_dict()
            loaded_model_state_dict = loaded_model.state_dict()

            self.assertEqual(set(model_state_dict.keys()), set(loaded_model_state_dict.keys()))
thomwolf's avatar
thomwolf committed
434

435
            models_equal = True
436
437
            for layer_name, p1 in model_state_dict.items():
                p2 = loaded_model_state_dict[layer_name]
438
439
                if p1.data.ne(p2.data).sum() > 0:
                    models_equal = False
thomwolf's avatar
thomwolf committed
440

441
            self.assertTrue(models_equal)
thomwolf's avatar
thomwolf committed
442

Patrick von Platen's avatar
Patrick von Platen committed
443
444
    def test_headmasking(self):
        if not self.test_head_masking:
445
            return
446

447
448
449
        global_rng.seed(42)
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
        global_rng.seed()
LysandreJik's avatar
LysandreJik committed
450

451
        inputs_dict["output_attentions"] = True
452
453
454
455
456
457
        config.output_hidden_states = True
        configs_no_init = _config_zero_init(config)  # To be sure we have no Nan
        for model_class in self.all_model_classes:
            model = model_class(config=configs_no_init)
            model.to(torch_device)
            model.eval()
LysandreJik's avatar
LysandreJik committed
458

459
460
461
            # Prepare head_mask
            # Set require_grad after having prepared the tensor to avoid error (leaf variable has been moved into the graph interior)
            head_mask = torch.ones(
Lysandre's avatar
Lysandre committed
462
463
464
                self.model_tester.num_hidden_layers,
                self.model_tester.num_attention_heads,
                device=torch_device,
465
466
467
468
            )
            head_mask[0, 0] = 0
            head_mask[-1, :-1] = 0
            head_mask.requires_grad_(requires_grad=True)
469
            inputs = self._prepare_for_class(inputs_dict, model_class).copy()
470
            inputs["head_mask"] = head_mask
471
472
473
474
475
            if model.config.is_encoder_decoder:
                signature = inspect.signature(model.forward)
                arg_names = [*signature.parameters.keys()]
                if "decoder_head_mask" in arg_names:  # necessary diferentiation because of T5 model
                    inputs["decoder_head_mask"] = head_mask
476
            outputs = model(**inputs, return_dict=True)
477
478
479
480
481
482
483
484
485

            # Test that we can get a gradient back for importance score computation
            output = sum(t.sum() for t in outputs[0])
            output = output.sum()
            output.backward()
            multihead_outputs = head_mask.grad

            self.assertIsNotNone(multihead_outputs)
            self.assertEqual(len(multihead_outputs), self.model_tester.num_hidden_layers)
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508

            def check_attentions_validity(attentions):
                # Remove Nan
                for t in attentions:
                    self.assertLess(
                        torch.sum(torch.isnan(t)), t.numel() / 4
                    )  # Check we don't have more than 25% nans (arbitrary)
                attentions = [
                    t.masked_fill(torch.isnan(t), 0.0) for t in attentions
                ]  # remove them (the test is less complete)

                self.assertAlmostEqual(attentions[0][..., 0, :, :].flatten().sum().item(), 0.0)
                self.assertNotEqual(attentions[0][..., -1, :, :].flatten().sum().item(), 0.0)
                if len(attentions) > 2:  # encoder-decoder models have only 2 layers in each module
                    self.assertNotEqual(attentions[1][..., 0, :, :].flatten().sum().item(), 0.0)
                self.assertAlmostEqual(attentions[-1][..., -2, :, :].flatten().sum().item(), 0.0)
                self.assertNotEqual(attentions[-1][..., -1, :, :].flatten().sum().item(), 0.0)

            if model.config.is_encoder_decoder:
                check_attentions_validity(outputs.encoder_attentions)
                check_attentions_validity(outputs.decoder_attentions)
            else:
                check_attentions_validity(outputs.attentions)
509

Patrick von Platen's avatar
Patrick von Platen committed
510
511
    def test_head_pruning(self):
        if not self.test_pruning:
512
513
514
            return

        for model_class in self.all_model_classes:
Lysandre's avatar
Lysandre committed
515
516
517
518
            (
                config,
                inputs_dict,
            ) = self.model_tester.prepare_config_and_inputs_for_common()
519

520
521
            if "head_mask" in inputs_dict:
                del inputs_dict["head_mask"]
522

523
            inputs_dict["output_attentions"] = True
524
525
526
527
            config.output_hidden_states = False
            model = model_class(config=config)
            model.to(torch_device)
            model.eval()
528
529
530
531
            heads_to_prune = {
                0: list(range(1, self.model_tester.num_attention_heads)),
                -1: [0],
            }
532
533
            model.prune_heads(heads_to_prune)
            with torch.no_grad():
534
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
535

536
            attentions = outputs[-1]
537

538
539
540
            self.assertEqual(attentions[0].shape[-3], 1)
            self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads)
            self.assertEqual(attentions[-1].shape[-3], self.model_tester.num_attention_heads - 1)
LysandreJik's avatar
LysandreJik committed
541

Patrick von Platen's avatar
Patrick von Platen committed
542
543
    def test_head_pruning_save_load_from_pretrained(self):
        if not self.test_pruning:
544
            return
LysandreJik's avatar
LysandreJik committed
545

546
        for model_class in self.all_model_classes:
Lysandre's avatar
Lysandre committed
547
548
549
550
            (
                config,
                inputs_dict,
            ) = self.model_tester.prepare_config_and_inputs_for_common()
551
552
553

            if "head_mask" in inputs_dict:
                del inputs_dict["head_mask"]
554

555
            inputs_dict["output_attentions"] = True
556
557
558
559
            config.output_hidden_states = False
            model = model_class(config=config)
            model.to(torch_device)
            model.eval()
560
561
562
563
            heads_to_prune = {
                0: list(range(1, self.model_tester.num_attention_heads)),
                -1: [0],
            }
564
            model.prune_heads(heads_to_prune)
565

566
            with tempfile.TemporaryDirectory() as temp_dir_name:
567
568
                model.save_pretrained(temp_dir_name)
                model = model_class.from_pretrained(temp_dir_name)
569
                model.to(torch_device)
570

571
            with torch.no_grad():
572
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
573
574
575
576
            attentions = outputs[-1]
            self.assertEqual(attentions[0].shape[-3], 1)
            self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads)
            self.assertEqual(attentions[-1].shape[-3], self.model_tester.num_attention_heads - 1)
577

Patrick von Platen's avatar
Patrick von Platen committed
578
579
    def test_head_pruning_save_load_from_config_init(self):
        if not self.test_pruning:
580
            return
581

582
        for model_class in self.all_model_classes:
Lysandre's avatar
Lysandre committed
583
584
585
586
            (
                config,
                inputs_dict,
            ) = self.model_tester.prepare_config_and_inputs_for_common()
587

588
589
            if "head_mask" in inputs_dict:
                del inputs_dict["head_mask"]
590

591
            inputs_dict["output_attentions"] = True
592
            config.output_hidden_states = False
593

594
595
596
597
            heads_to_prune = {
                0: list(range(1, self.model_tester.num_attention_heads)),
                -1: [0],
            }
598
            config.pruned_heads = heads_to_prune
599

600
601
602
            model = model_class(config=config)
            model.to(torch_device)
            model.eval()
603

604
            with torch.no_grad():
605
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
606
            attentions = outputs[-1]
607

608
609
610
            self.assertEqual(attentions[0].shape[-3], 1)
            self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads)
            self.assertEqual(attentions[-1].shape[-3], self.model_tester.num_attention_heads - 1)
611

Patrick von Platen's avatar
Patrick von Platen committed
612
613
    def test_head_pruning_integration(self):
        if not self.test_pruning:
614
            return
615

616
        for model_class in self.all_model_classes:
Lysandre's avatar
Lysandre committed
617
618
619
620
            (
                config,
                inputs_dict,
            ) = self.model_tester.prepare_config_and_inputs_for_common()
621

622
623
            if "head_mask" in inputs_dict:
                del inputs_dict["head_mask"]
624

625
            inputs_dict["output_attentions"] = True
626
            config.output_hidden_states = False
627

628
629
            heads_to_prune = {0: [0], 1: [1, 2]}
            config.pruned_heads = heads_to_prune
630

631
632
633
            model = model_class(config=config)
            model.to(torch_device)
            model.eval()
634

635
            with torch.no_grad():
636
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
637
            attentions = outputs[-1]
638

639
640
641
642
            self.assertEqual(attentions[0].shape[-3], self.model_tester.num_attention_heads - 1)
            self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads - 2)
            self.assertEqual(attentions[2].shape[-3], self.model_tester.num_attention_heads)
            self.assertEqual(attentions[3].shape[-3], self.model_tester.num_attention_heads)
thomwolf's avatar
thomwolf committed
643

644
            with tempfile.TemporaryDirectory() as temp_dir_name:
645
646
                model.save_pretrained(temp_dir_name)
                model = model_class.from_pretrained(temp_dir_name)
647
                model.to(torch_device)
thomwolf's avatar
thomwolf committed
648

649
            with torch.no_grad():
650
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
651
            attentions = outputs[-1]
LysandreJik's avatar
LysandreJik committed
652

653
654
655
656
            self.assertEqual(attentions[0].shape[-3], self.model_tester.num_attention_heads - 1)
            self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads - 2)
            self.assertEqual(attentions[2].shape[-3], self.model_tester.num_attention_heads)
            self.assertEqual(attentions[3].shape[-3], self.model_tester.num_attention_heads)
thomwolf's avatar
thomwolf committed
657

658
659
            heads_to_prune = {0: [0], 2: [1, 2]}
            model.prune_heads(heads_to_prune)
660

661
            with torch.no_grad():
662
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
663
            attentions = outputs[-1]
664

665
666
667
668
            self.assertEqual(attentions[0].shape[-3], self.model_tester.num_attention_heads - 1)
            self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads - 2)
            self.assertEqual(attentions[2].shape[-3], self.model_tester.num_attention_heads - 2)
            self.assertEqual(attentions[3].shape[-3], self.model_tester.num_attention_heads)
669

670
            self.assertDictEqual(model.config.pruned_heads, {0: [0], 1: [1, 2], 2: [1, 2]})
thomwolf's avatar
thomwolf committed
671

Patrick von Platen's avatar
Patrick von Platen committed
672
    def test_hidden_states_output(self):
Joseph Liu's avatar
Joseph Liu committed
673
        def check_hidden_states_output(inputs_dict, config, model_class):
674
            model = model_class(config)
675
            model.to(torch_device)
thomwolf's avatar
thomwolf committed
676
            model.eval()
Joseph Liu's avatar
Joseph Liu committed
677

thomwolf's avatar
thomwolf committed
678
            with torch.no_grad():
679
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
680
681

            hidden_states = outputs.encoder_hidden_states if config.is_encoder_decoder else outputs.hidden_states
Patrick von Platen's avatar
Patrick von Platen committed
682

Sylvain Gugger's avatar
Sylvain Gugger committed
683
684
685
686
            expected_num_layers = getattr(
                self.model_tester, "expected_num_hidden_layers", self.model_tester.num_hidden_layers + 1
            )
            self.assertEqual(len(hidden_states), expected_num_layers)
687

Patrick von Platen's avatar
Patrick von Platen committed
688
689
690
691
692
693
694
            if hasattr(self.model_tester, "encoder_seq_length"):
                seq_length = self.model_tester.encoder_seq_length
                if hasattr(self.model_tester, "chunk_length") and self.model_tester.chunk_length > 1:
                    seq_length = seq_length * self.model_tester.chunk_length
            else:
                seq_length = self.model_tester.seq_length

695
            self.assertListEqual(
Lysandre's avatar
Lysandre committed
696
697
                list(hidden_states[0].shape[-2:]),
                [seq_length, self.model_tester.hidden_size],
698
            )
thomwolf's avatar
thomwolf committed
699

700
701
702
703
704
705
706
707
708
709
710
711
712
            if config.is_encoder_decoder:
                hidden_states = outputs.decoder_hidden_states

                self.assertIsInstance(hidden_states, (list, tuple))
                self.assertEqual(len(hidden_states), expected_num_layers)
                seq_len = getattr(self.model_tester, "seq_length", None)
                decoder_seq_length = getattr(self.model_tester, "decoder_seq_length", seq_len)

                self.assertListEqual(
                    list(hidden_states[0].shape[-2:]),
                    [decoder_seq_length, self.model_tester.hidden_size],
                )

Joseph Liu's avatar
Joseph Liu committed
713
714
715
716
717
718
719
720
721
722
723
724
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            inputs_dict["output_hidden_states"] = True
            check_hidden_states_output(inputs_dict, config, model_class)

            # check that output_hidden_states also work using config
            del inputs_dict["output_hidden_states"]
            config.output_hidden_states = True

            check_hidden_states_output(inputs_dict, config, model_class)

725
726
727
728
729
730
731
732
733
734
735
736
737
    def test_retain_grad_hidden_states_attentions(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
        config.output_hidden_states = True
        config.output_attentions = True

        # no need to test all models as different heads yield the same functionality
        model_class = self.all_model_classes[0]
        model = model_class(config)
        model.to(torch_device)

        inputs = self._prepare_for_class(inputs_dict, model_class)

        outputs = model(**inputs)
738
739

        print(outputs)
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
        output = outputs[0]

        if config.is_encoder_decoder:
            # Seq2Seq models
            encoder_hidden_states = outputs.encoder_hidden_states[0]
            encoder_attentions = outputs.encoder_attentions[0]
            encoder_hidden_states.retain_grad()
            encoder_attentions.retain_grad()

            decoder_hidden_states = outputs.decoder_hidden_states[0]
            decoder_attentions = outputs.decoder_attentions[0]
            decoder_hidden_states.retain_grad()
            decoder_attentions.retain_grad()

            cross_attentions = outputs.cross_attentions[0]
            cross_attentions.retain_grad()

            output.flatten()[0].backward(retain_graph=True)

            self.assertIsNotNone(encoder_hidden_states.grad)
            self.assertIsNotNone(encoder_attentions.grad)
            self.assertIsNotNone(decoder_hidden_states.grad)
            self.assertIsNotNone(decoder_attentions.grad)
            self.assertIsNotNone(cross_attentions.grad)
        else:
            # Encoder-/Decoder-only models
            hidden_states = outputs.hidden_states[0]
            attentions = outputs.attentions[0]

            hidden_states.retain_grad()
            attentions.retain_grad()

            output.flatten()[0].backward(retain_graph=True)

            self.assertIsNotNone(hidden_states.grad)
            self.assertIsNotNone(attentions.grad)

Pradhy729's avatar
Pradhy729 committed
777
    def test_feed_forward_chunking(self):
Lysandre's avatar
Lysandre committed
778
779
780
781
        (
            original_config,
            inputs_dict,
        ) = self.model_tester.prepare_config_and_inputs_for_common()
Pradhy729's avatar
Pradhy729 committed
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
        for model_class in self.all_model_classes:
            torch.manual_seed(0)
            config = copy.deepcopy(original_config)
            model = model_class(config)
            model.to(torch_device)
            model.eval()

            hidden_states_no_chunk = model(**self._prepare_for_class(inputs_dict, model_class))[0]

            torch.manual_seed(0)
            config.chunk_size_feed_forward = 1
            model = model_class(config)
            model.to(torch_device)
            model.eval()

            hidden_states_with_chunk = model(**self._prepare_for_class(inputs_dict, model_class))[0]
            self.assertTrue(torch.allclose(hidden_states_no_chunk, hidden_states_with_chunk, atol=1e-3))

Patrick von Platen's avatar
Patrick von Platen committed
800
    def test_resize_tokens_embeddings(self):
Lysandre's avatar
Lysandre committed
801
802
803
804
        (
            original_config,
            inputs_dict,
        ) = self.model_tester.prepare_config_and_inputs_for_common()
Patrick von Platen's avatar
Patrick von Platen committed
805
        if not self.test_resize_embeddings:
806
807
808
809
810
            return

        for model_class in self.all_model_classes:
            config = copy.deepcopy(original_config)
            model = model_class(config)
811
            model.to(torch_device)
812

Patrick von Platen's avatar
Patrick von Platen committed
813
814
815
            if self.model_tester.is_training is False:
                model.eval()

816
817
818
819
820
821
822
823
824
825
            model_vocab_size = config.vocab_size
            # Retrieve the embeddings and clone theme
            model_embed = model.resize_token_embeddings(model_vocab_size)
            cloned_embeddings = model_embed.weight.clone()

            # Check that resizing the token embeddings with a larger vocab size increases the model's vocab size
            model_embed = model.resize_token_embeddings(model_vocab_size + 10)
            self.assertEqual(model.config.vocab_size, model_vocab_size + 10)
            # Check that it actually resizes the embeddings matrix
            self.assertEqual(model_embed.weight.shape[0], cloned_embeddings.shape[0] + 10)
826
            # Check that the model can still do a forward pass successfully (every parameter should be resized)
827
            model(**self._prepare_for_class(inputs_dict, model_class))
828
829
830
831
832
833
834

            # Check that resizing the token embeddings with a smaller vocab size decreases the model's vocab size
            model_embed = model.resize_token_embeddings(model_vocab_size - 15)
            self.assertEqual(model.config.vocab_size, model_vocab_size - 15)
            # Check that it actually resizes the embeddings matrix
            self.assertEqual(model_embed.weight.shape[0], cloned_embeddings.shape[0] - 15)

835
836
837
            # Check that the model can still do a forward pass successfully (every parameter should be resized)
            # Input ids should be clamped to the maximum size of the vocabulary
            inputs_dict["input_ids"].clamp_(max=model_vocab_size - 15 - 1)
838
839
840
841

            # make sure that decoder_input_ids are resized as well
            if "decoder_input_ids" in inputs_dict:
                inputs_dict["decoder_input_ids"].clamp_(max=model_vocab_size - 15 - 1)
842
            model(**self._prepare_for_class(inputs_dict, model_class))
843

844
845
846
847
848
849
850
851
            # Check that adding and removing tokens has not modified the first part of the embedding matrix.
            models_equal = True
            for p1, p2 in zip(cloned_embeddings, model_embed.weight):
                if p1.data.ne(p2.data).sum() > 0:
                    models_equal = False

            self.assertTrue(models_equal)

852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
    def test_resize_embeddings_untied(self):
        (
            original_config,
            inputs_dict,
        ) = self.model_tester.prepare_config_and_inputs_for_common()
        if not self.test_resize_embeddings:
            return

        original_config.tie_word_embeddings = False

        # if model cannot untied embeddings -> leave test
        if original_config.tie_word_embeddings:
            return

        for model_class in self.all_model_classes:
            config = copy.deepcopy(original_config)
            model = model_class(config).to(torch_device)

            # if no output embeddings -> leave test
            if model.get_output_embeddings() is None:
                continue

            # Check that resizing the token embeddings with a larger vocab size increases the model's vocab size
            model_vocab_size = config.vocab_size
            model.resize_token_embeddings(model_vocab_size + 10)
            self.assertEqual(model.config.vocab_size, model_vocab_size + 10)
            output_embeds = model.get_output_embeddings()
            self.assertEqual(output_embeds.weight.shape[0], model_vocab_size + 10)
            # Check bias if present
            if output_embeds.bias is not None:
                self.assertEqual(output_embeds.bias.shape[0], model_vocab_size + 10)
            # Check that the model can still do a forward pass successfully (every parameter should be resized)
            model(**self._prepare_for_class(inputs_dict, model_class))

            # Check that resizing the token embeddings with a smaller vocab size decreases the model's vocab size
            model.resize_token_embeddings(model_vocab_size - 15)
            self.assertEqual(model.config.vocab_size, model_vocab_size - 15)
            # Check that it actually resizes the embeddings matrix
            output_embeds = model.get_output_embeddings()
            self.assertEqual(output_embeds.weight.shape[0], model_vocab_size - 15)
            # Check bias if present
            if output_embeds.bias is not None:
                self.assertEqual(output_embeds.bias.shape[0], model_vocab_size - 15)
            # Check that the model can still do a forward pass successfully (every parameter should be resized)
            # Input ids should be clamped to the maximum size of the vocabulary
            inputs_dict["input_ids"].clamp_(max=model_vocab_size - 15 - 1)
            if "decoder_input_ids" in inputs_dict:
                inputs_dict["decoder_input_ids"].clamp_(max=model_vocab_size - 15 - 1)
            # Check that the model can still do a forward pass successfully (every parameter should be resized)
            model(**self._prepare_for_class(inputs_dict, model_class))

Patrick von Platen's avatar
Patrick von Platen committed
903
    def test_model_common_attributes(self):
904
905
906
907
908
909
910
911
912
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            model = model_class(config)
            self.assertIsInstance(model.get_input_embeddings(), (torch.nn.Embedding, AdaptiveEmbedding))
            model.set_input_embeddings(torch.nn.Embedding(10, 10))
            x = model.get_output_embeddings()
            self.assertTrue(x is None or isinstance(x, torch.nn.Linear))

913
    def test_correct_missing_keys(self):
914
915
        if not self.test_missing_keys:
            return
916
917
918
919
920
921
922
923
924
925
926
927
928
929
        config, _ = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            model = model_class(config)
            base_model_prefix = model.base_model_prefix

            if hasattr(model, base_model_prefix):
                with tempfile.TemporaryDirectory() as temp_dir_name:
                    model.base_model.save_pretrained(temp_dir_name)
                    model, loading_info = model_class.from_pretrained(temp_dir_name, output_loading_info=True)

                    with self.subTest(msg="Missing keys for {}".format(model.__class__.__name__)):
                        self.assertGreater(len(loading_info["missing_keys"]), 0)

930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
    def test_tie_model_weights(self):
        if not self.test_torchscript:
            return

        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        def check_same_values(layer_1, layer_2):
            equal = True
            for p1, p2 in zip(layer_1.weight, layer_2.weight):
                if p1.data.ne(p2.data).sum() > 0:
                    equal = False
            return equal

        for model_class in self.all_model_classes:
            config.torchscript = True
            model_not_tied = model_class(config)
            if model_not_tied.get_output_embeddings() is None:
                continue

            config_tied = copy.deepcopy(config)
            config_tied.torchscript = False
            model_tied = model_class(config_tied)
            params_tied = list(model_tied.parameters())
            # Check that the embedding layer and decoding layer are the same in size and in value
            # self.assertTrue(check_same_values(embeddings, decoding))

            # # Check that after modification, they remain the same.
            # embeddings.weight.data.div_(2)
            # # Check that the embedding layer and decoding layer are the same in size and in value
            # self.assertTrue(embeddings.weight.shape, decoding.weight.shape)
            # self.assertTrue(check_same_values(embeddings, decoding))

            # # Check that after modification, they remain the same.
            # decoding.weight.data.div_(4)
            # # Check that the embedding layer and decoding layer are the same in size and in value
            # self.assertTrue(embeddings.weight.shape, decoding.weight.shape)
            # self.assertTrue(check_same_values(embeddings, decoding))

            # Check that after resize they remain tied.
            model_tied.resize_token_embeddings(config.vocab_size + 10)
            params_tied_2 = list(model_tied.parameters())
            self.assertEqual(len(params_tied_2), len(params_tied))

            # decoding.weight.data.mul_(20)
            # # Check that the embedding layer and decoding layer are the same in size and in value
            # self.assertTrue(model.transformer.wte.weight.shape, model.lm_head.weight.shape)
            # self.assertTrue(check_same_values(model.transformer.wte, model.lm_head))

978
979
980
981
    def test_model_outputs_equivalence(self):

        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

Sam Shleifer's avatar
Sam Shleifer committed
982
983
984
985
        def set_nan_tensor_to_zero(t):
            t[t != t] = 0
            return t

986
987
988
989
990
991
992
993
994
995
996
997
998
        def check_equivalence(model, tuple_inputs, dict_inputs, additional_kwargs={}):
            with torch.no_grad():
                tuple_output = model(**tuple_inputs, return_dict=False, **additional_kwargs)
                dict_output = model(**dict_inputs, return_dict=True, **additional_kwargs).to_tuple()

                def recursive_check(tuple_object, dict_object):
                    if isinstance(tuple_object, (List, Tuple)):
                        for tuple_iterable_value, dict_iterable_value in zip(tuple_object, dict_object):
                            recursive_check(tuple_iterable_value, dict_iterable_value)
                    elif tuple_object is None:
                        return
                    else:
                        self.assertTrue(
Sam Shleifer's avatar
Sam Shleifer committed
999
1000
1001
                            torch.allclose(
                                set_nan_tensor_to_zero(tuple_object), set_nan_tensor_to_zero(dict_object), atol=1e-5
                            ),
1002
                            msg=f"Tuple and dict output are not equal. Difference: {torch.max(torch.abs(tuple_object - dict_object))}. Tuple has `nan`: {torch.isnan(tuple_object).any()} and `inf`: {torch.isinf(tuple_object)}. Dict has `nan`: {torch.isnan(dict_object).any()} and `inf`: {torch.isinf(dict_object)}.",
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
                        )

                recursive_check(tuple_output, dict_output)

        for model_class in self.all_model_classes:
            model = model_class(config)
            model.to(torch_device)
            model.eval()

            tuple_inputs = self._prepare_for_class(inputs_dict, model_class)
            dict_inputs = self._prepare_for_class(inputs_dict, model_class)
            check_equivalence(model, tuple_inputs, dict_inputs)

            tuple_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
            dict_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
            check_equivalence(model, tuple_inputs, dict_inputs)

            tuple_inputs = self._prepare_for_class(inputs_dict, model_class)
            dict_inputs = self._prepare_for_class(inputs_dict, model_class)
            check_equivalence(model, tuple_inputs, dict_inputs, {"output_hidden_states": True})

            tuple_inputs = self._prepare_for_class(inputs_dict, model_class)
            dict_inputs = self._prepare_for_class(inputs_dict, model_class)
            check_equivalence(model, tuple_inputs, dict_inputs, {"output_attentions": True})

            tuple_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
            dict_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
            check_equivalence(model, tuple_inputs, dict_inputs, {"output_hidden_states": True})

            tuple_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
            dict_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
            check_equivalence(model, tuple_inputs, dict_inputs, {"output_attentions": True})

            tuple_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
            dict_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
            check_equivalence(
                model, tuple_inputs, dict_inputs, {"output_hidden_states": True, "output_attentions": True}
            )

Patrick von Platen's avatar
Patrick von Platen committed
1042
    def test_inputs_embeds(self):
1043
1044
1045
1046
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            model = model_class(config)
1047
            model.to(torch_device)
thomwolf's avatar
thomwolf committed
1048
            model.eval()
1049

1050
            inputs = copy.deepcopy(self._prepare_for_class(inputs_dict, model_class))
Weizhen's avatar
Weizhen committed
1051

1052
1053
1054
1055
1056
1057
1058
1059
1060
            if not self.is_encoder_decoder:
                input_ids = inputs["input_ids"]
                del inputs["input_ids"]
            else:
                encoder_input_ids = inputs["input_ids"]
                decoder_input_ids = inputs.get("decoder_input_ids", encoder_input_ids)
                del inputs["input_ids"]
                inputs.pop("decoder_input_ids", None)

1061
1062
            wte = model.get_input_embeddings()
            if not self.is_encoder_decoder:
1063
                inputs["inputs_embeds"] = wte(input_ids)
1064
            else:
1065
1066
                inputs["inputs_embeds"] = wte(encoder_input_ids)
                inputs["decoder_inputs_embeds"] = wte(decoder_input_ids)
1067

thomwolf's avatar
thomwolf committed
1068
            with torch.no_grad():
Weizhen's avatar
Weizhen committed
1069
                model(**inputs)[0]
1070

1071
1072
    @require_torch_multi_gpu
    def test_multi_gpu_data_parallel_forward(self):
1073
1074
1075
1076
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        # some params shouldn't be scattered by nn.DataParallel
        # so just remove them if they are present.
Patrick von Platen's avatar
Patrick von Platen committed
1077
        blacklist_non_batched_params = ["head_mask", "decoder_head_mask"]
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
        for k in blacklist_non_batched_params:
            inputs_dict.pop(k, None)

        # move input tensors to cuda:O
        for k, v in inputs_dict.items():
            if torch.is_tensor(v):
                inputs_dict[k] = v.to(0)

        for model_class in self.all_model_classes:
            model = model_class(config=config)
            model.to(0)
            model.eval()

            # Wrap model in nn.DataParallel
            model = torch.nn.DataParallel(model)
            with torch.no_grad():
1094
                _ = model(**self._prepare_for_class(inputs_dict, model_class))
1095

1096
1097
1098
    @require_torch_multi_gpu
    def test_model_parallelization(self):
        if not self.test_model_parallel:
1099
            return
1100

1101
        # a candidate for testing_utils
1102
        def get_current_gpu_memory_use():
1103
1104
1105
1106
1107
1108
            """ returns a list of cuda memory allocations per GPU in MBs"""

            per_device_memory = []
            for id in range(torch.cuda.device_count()):
                with torch.cuda.device(id):
                    per_device_memory.append(torch.cuda.memory_allocated() >> 20)
1109
1110
1111
1112
1113
1114
1115
1116
1117

            return per_device_memory

        # Needs a large model to see the difference.
        config = self.model_tester.get_large_model_config()

        for model_class in self.all_parallelizable_model_classes:
            torch.cuda.empty_cache()

1118
1119
1120
            # 1. single gpu memory load + unload + memory measurements
            # Retrieve initial memory usage (can easily be ~0.6-1.5GB if cuda-kernels have been preloaded by previous tests)
            memory_at_start = get_current_gpu_memory_use()
1121

1122
1123
            # Put model on device 0 and take a memory snapshot
            model = model_class(config)
1124
1125
1126
            model.to("cuda:0")
            memory_after_model_load = get_current_gpu_memory_use()

1127
1128
1129
            # The memory use on device 0 should be higher than it was initially.
            self.assertGreater(memory_after_model_load[0], memory_at_start[0])

1130
            del model
1131
            gc.collect()
1132
1133
            torch.cuda.empty_cache()

1134
1135
1136
            # 2. MP test
            # it's essential to re-calibrate the usage before the next stage
            memory_at_start = get_current_gpu_memory_use()
1137
1138

            # Spread model layers over multiple devices
1139
            model = model_class(config)
1140
1141
1142
1143
1144
            model.parallelize()
            memory_after_parallelization = get_current_gpu_memory_use()

            # Assert that the memory use on all devices is higher than it was when loaded only on CPU
            for n in range(torch.cuda.device_count()):
1145
                self.assertGreater(memory_after_parallelization[n], memory_at_start[n])
1146

1147
            # Assert that the memory use of device 0 is lower than it was when the entire model was loaded on it
1148
1149
            self.assertLess(memory_after_parallelization[0], memory_after_model_load[0])

1150
1151
            # Assert that the memory use of device 1 is higher than it was when the entire model was loaded
            # on device 0 and device 1 wasn't used at all
1152
1153
1154
            self.assertGreater(memory_after_parallelization[1], memory_after_model_load[1])

            del model
1155
            gc.collect()
1156
1157
1158
1159
1160
            torch.cuda.empty_cache()

    @require_torch_multi_gpu
    def test_model_parallel_equal_results(self):
        if not self.test_model_parallel:
1161
            return
1162
1163
1164
1165
1166
1167

        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_parallelizable_model_classes:
            inputs_dict = self._prepare_for_class(inputs_dict, model_class)

1168
            def cast_to_device(dictionary, device):
1169
1170
1171
                output = {}
                for k, v in dictionary.items():
                    if isinstance(v, torch.Tensor):
1172
                        output[k] = v.to(device)
1173
1174
1175
1176
1177
                    else:
                        output[k] = v

                return output

1178
1179
1180
1181
1182
1183
            model = model_class(config)
            output = model(**cast_to_device(inputs_dict, "cpu"))

            model.parallelize()

            parallel_output = model(**cast_to_device(inputs_dict, "cuda:0"))
1184
1185
1186
1187
1188
1189
1190
1191

            for value, parallel_value in zip(output, parallel_output):
                if isinstance(value, torch.Tensor):
                    self.assertTrue(torch.allclose(value, parallel_value.to("cpu"), atol=1e-7))
                elif isinstance(value, (Tuple, List)):
                    for value_, parallel_value_ in zip(value, parallel_value):
                        self.assertTrue(torch.allclose(value_, parallel_value_.to("cpu"), atol=1e-7))

1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
    @require_torch_multi_gpu
    def test_model_parallel_beam_search(self):
        if not self.test_model_parallel:
            return

        all_generative_and_parallelizable_model_classes = tuple(
            set(self.all_generative_model_classes).intersection(self.all_parallelizable_model_classes)
        )

        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in all_generative_and_parallelizable_model_classes:
            inputs_dict = self._prepare_for_class(inputs_dict, model_class)
            model = model_class(config)

            def cast_to_device(dictionary, device):
                output = {}
                for k, v in dictionary.items():
                    if isinstance(v, torch.Tensor):
                        output[k] = v.to(device)
                    else:
                        output[k] = v

                return output

            model.parallelize()
            model.generate(**cast_to_device(inputs_dict, "cuda:0"), num_beams=2)

1220

1221
global_rng = random.Random()
thomwolf's avatar
thomwolf committed
1222
1223


thomwolf's avatar
thomwolf committed
1224
def ids_tensor(shape, vocab_size, rng=None, name=None):
1225
    #  Creates a random int32 tensor of the shape within the vocab size
thomwolf's avatar
thomwolf committed
1226
    if rng is None:
1227
        rng = global_rng
thomwolf's avatar
thomwolf committed
1228

thomwolf's avatar
thomwolf committed
1229
1230
1231
    total_dims = 1
    for dim in shape:
        total_dims *= dim
thomwolf's avatar
thomwolf committed
1232

thomwolf's avatar
thomwolf committed
1233
1234
1235
    values = []
    for _ in range(total_dims):
        values.append(rng.randint(0, vocab_size - 1))
thomwolf's avatar
thomwolf committed
1236

1237
    return torch.tensor(data=values, dtype=torch.long, device=torch_device).view(shape).contiguous()
thomwolf's avatar
thomwolf committed
1238
1239


1240
1241
1242
1243
1244
1245
1246
def random_attention_mask(shape, rng=None, name=None):
    attn_mask = ids_tensor(shape, vocab_size=2, rng=None, name=None)
    # make sure that at least one token is attended to for each batch
    attn_mask[:, -1] = 1
    return attn_mask


1247
def floats_tensor(shape, scale=1.0, rng=None, name=None):
Patrick von Platen's avatar
Patrick von Platen committed
1248
    """Creates a random float32 tensor"""
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
    if rng is None:
        rng = global_rng

    total_dims = 1
    for dim in shape:
        total_dims *= dim

    values = []
    for _ in range(total_dims):
        values.append(rng.random() * scale)

1260
    return torch.tensor(data=values, dtype=torch.float, device=torch_device).view(shape).contiguous()
1261
1262


1263
@require_torch
thomwolf's avatar
thomwolf committed
1264
class ModelUtilsTest(unittest.TestCase):
1265
    @slow
Patrick von Platen's avatar
Patrick von Platen committed
1266
    def test_model_from_pretrained(self):
1267
        for model_name in BERT_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
thomwolf's avatar
thomwolf committed
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
            config = BertConfig.from_pretrained(model_name)
            self.assertIsNotNone(config)
            self.assertIsInstance(config, PretrainedConfig)

            model = BertModel.from_pretrained(model_name)
            model, loading_info = BertModel.from_pretrained(model_name, output_loading_info=True)
            self.assertIsNotNone(model)
            self.assertIsInstance(model, PreTrainedModel)
            for value in loading_info.values():
                self.assertEqual(len(value), 0)

            config = BertConfig.from_pretrained(model_name, output_attentions=True, output_hidden_states=True)
Lysandre Debut's avatar
Lysandre Debut committed
1280
1281
1282
1283

            # Not sure this is the intended behavior. TODO fix Lysandre & Thom
            config.name_or_path = model_name

thomwolf's avatar
thomwolf committed
1284
1285
1286
            model = BertModel.from_pretrained(model_name, output_attentions=True, output_hidden_states=True)
            self.assertEqual(model.config.output_hidden_states, True)
            self.assertEqual(model.config, config)