# Copyright 2020 The HuggingFace Team. All rights reserved. # # 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. import unittest import pytest from packaging import version from transformers import AutoTokenizer, MobileBertConfig, MobileBertForMaskedLM, is_torch_available from transformers.models.auto import get_values from transformers.testing_utils import require_sentencepiece, require_tokenizers, require_torch, slow, torch_device from ...test_configuration_common import ConfigTester from ...test_modeling_common import ModelTesterMixin, ids_tensor, random_attention_mask from ...test_pipeline_mixin import PipelineTesterMixin if is_torch_available(): import torch from transformers import ( MODEL_FOR_PRETRAINING_MAPPING, MobileBertForMaskedLM, MobileBertForMultipleChoice, MobileBertForNextSentencePrediction, MobileBertForPreTraining, MobileBertForQuestionAnswering, MobileBertForSequenceClassification, MobileBertForTokenClassification, MobileBertModel, ) class MobileBertModelTester: def __init__( self, parent, batch_size=13, seq_length=7, is_training=True, use_input_mask=True, use_token_type_ids=True, use_labels=True, vocab_size=99, hidden_size=64, embedding_size=32, num_hidden_layers=2, num_attention_heads=4, intermediate_size=37, hidden_act="gelu", hidden_dropout_prob=0.1, attention_probs_dropout_prob=0.1, max_position_embeddings=512, type_vocab_size=16, type_sequence_label_size=2, initializer_range=0.02, num_labels=3, num_choices=4, scope=None, ): self.parent = parent self.batch_size = batch_size self.seq_length = seq_length self.is_training = is_training self.use_input_mask = use_input_mask self.use_token_type_ids = use_token_type_ids self.use_labels = use_labels self.vocab_size = vocab_size self.hidden_size = hidden_size self.embedding_size = embedding_size self.num_hidden_layers = num_hidden_layers self.num_attention_heads = num_attention_heads self.intermediate_size = intermediate_size self.hidden_act = hidden_act self.hidden_dropout_prob = hidden_dropout_prob self.attention_probs_dropout_prob = attention_probs_dropout_prob self.max_position_embeddings = max_position_embeddings self.type_vocab_size = type_vocab_size self.type_sequence_label_size = type_sequence_label_size self.initializer_range = initializer_range self.num_labels = num_labels self.num_choices = num_choices self.scope = scope def prepare_config_and_inputs(self): input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size) input_mask = None if self.use_input_mask: input_mask = random_attention_mask([self.batch_size, self.seq_length]) token_type_ids = None if self.use_token_type_ids: token_type_ids = ids_tensor([self.batch_size, self.seq_length], self.type_vocab_size) sequence_labels = None token_labels = None choice_labels = None if self.use_labels: sequence_labels = ids_tensor([self.batch_size], self.type_sequence_label_size) token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels) choice_labels = ids_tensor([self.batch_size], self.num_choices) config = self.get_config() return config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels def get_config(self): return MobileBertConfig( vocab_size=self.vocab_size, hidden_size=self.hidden_size, num_hidden_layers=self.num_hidden_layers, num_attention_heads=self.num_attention_heads, intermediate_size=self.intermediate_size, embedding_size=self.embedding_size, hidden_act=self.hidden_act, hidden_dropout_prob=self.hidden_dropout_prob, attention_probs_dropout_prob=self.attention_probs_dropout_prob, max_position_embeddings=self.max_position_embeddings, type_vocab_size=self.type_vocab_size, is_decoder=False, initializer_range=self.initializer_range, ) def create_and_check_mobilebert_model( self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels ): model = MobileBertModel(config=config) model.to(torch_device) model.eval() result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids) result = model(input_ids, token_type_ids=token_type_ids) result = model(input_ids) self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size)) self.parent.assertEqual(result.pooler_output.shape, (self.batch_size, self.hidden_size)) def create_and_check_mobilebert_for_masked_lm( self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels ): model = MobileBertForMaskedLM(config=config) model.to(torch_device) model.eval() result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=token_labels) self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size)) def create_and_check_mobilebert_for_next_sequence_prediction( self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels ): model = MobileBertForNextSentencePrediction(config=config) model.to(torch_device) model.eval() result = model( input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=sequence_labels, ) self.parent.assertEqual(result.logits.shape, (self.batch_size, 2)) def create_and_check_mobilebert_for_pretraining( self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels ): model = MobileBertForPreTraining(config=config) model.to(torch_device) model.eval() result = model( input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=token_labels, next_sentence_label=sequence_labels, ) self.parent.assertEqual(result.prediction_logits.shape, (self.batch_size, self.seq_length, self.vocab_size)) self.parent.assertEqual(result.seq_relationship_logits.shape, (self.batch_size, 2)) def create_and_check_mobilebert_for_question_answering( self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels ): model = MobileBertForQuestionAnswering(config=config) model.to(torch_device) model.eval() result = model( input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, start_positions=sequence_labels, end_positions=sequence_labels, ) self.parent.assertEqual(result.start_logits.shape, (self.batch_size, self.seq_length)) self.parent.assertEqual(result.end_logits.shape, (self.batch_size, self.seq_length)) def create_and_check_mobilebert_for_sequence_classification( self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels ): config.num_labels = self.num_labels model = MobileBertForSequenceClassification(config) model.to(torch_device) model.eval() result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=sequence_labels) self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_labels)) def create_and_check_mobilebert_for_token_classification( self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels ): config.num_labels = self.num_labels model = MobileBertForTokenClassification(config=config) model.to(torch_device) model.eval() result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=token_labels) self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.num_labels)) def create_and_check_mobilebert_for_multiple_choice( self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels ): config.num_choices = self.num_choices model = MobileBertForMultipleChoice(config=config) model.to(torch_device) model.eval() multiple_choice_inputs_ids = input_ids.unsqueeze(1).expand(-1, self.num_choices, -1).contiguous() multiple_choice_token_type_ids = token_type_ids.unsqueeze(1).expand(-1, self.num_choices, -1).contiguous() multiple_choice_input_mask = input_mask.unsqueeze(1).expand(-1, self.num_choices, -1).contiguous() result = model( multiple_choice_inputs_ids, attention_mask=multiple_choice_input_mask, token_type_ids=multiple_choice_token_type_ids, labels=choice_labels, ) self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_choices)) def prepare_config_and_inputs_for_common(self): config_and_inputs = self.prepare_config_and_inputs() ( config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels, ) = config_and_inputs inputs_dict = {"input_ids": input_ids, "token_type_ids": token_type_ids, "attention_mask": input_mask} return config, inputs_dict @require_torch class MobileBertModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase): all_model_classes = ( ( MobileBertModel, MobileBertForMaskedLM, MobileBertForMultipleChoice, MobileBertForNextSentencePrediction, MobileBertForPreTraining, MobileBertForQuestionAnswering, MobileBertForSequenceClassification, MobileBertForTokenClassification, ) if is_torch_available() else () ) pipeline_model_mapping = ( { "feature-extraction": MobileBertModel, "fill-mask": MobileBertForMaskedLM, "question-answering": MobileBertForQuestionAnswering, "text-classification": MobileBertForSequenceClassification, "token-classification": MobileBertForTokenClassification, "zero-shot": MobileBertForSequenceClassification, } if is_torch_available() else {} ) # special case for ForPreTraining model def _prepare_for_class(self, inputs_dict, model_class, return_labels=False): inputs_dict = super()._prepare_for_class(inputs_dict, model_class, return_labels=return_labels) if return_labels: if model_class in get_values(MODEL_FOR_PRETRAINING_MAPPING): inputs_dict["labels"] = torch.zeros( (self.model_tester.batch_size, self.model_tester.seq_length), dtype=torch.long, device=torch_device ) inputs_dict["next_sentence_label"] = torch.zeros( self.model_tester.batch_size, dtype=torch.long, device=torch_device ) return inputs_dict # TODO (@SunMarc): Fix me @unittest.skip(reason="It's broken.") def test_resize_tokens_embeddings(self): super().test_resize_tokens_embeddings() def setUp(self): self.model_tester = MobileBertModelTester(self) self.config_tester = ConfigTester(self, config_class=MobileBertConfig, hidden_size=37) def test_config(self): self.config_tester.run_common_tests() def test_mobilebert_model(self): config_and_inputs = self.model_tester.prepare_config_and_inputs() self.model_tester.create_and_check_mobilebert_model(*config_and_inputs) def test_for_masked_lm(self): config_and_inputs = self.model_tester.prepare_config_and_inputs() self.model_tester.create_and_check_mobilebert_for_masked_lm(*config_and_inputs) def test_for_multiple_choice(self): config_and_inputs = self.model_tester.prepare_config_and_inputs() self.model_tester.create_and_check_mobilebert_for_multiple_choice(*config_and_inputs) def test_for_next_sequence_prediction(self): config_and_inputs = self.model_tester.prepare_config_and_inputs() self.model_tester.create_and_check_mobilebert_for_next_sequence_prediction(*config_and_inputs) def test_for_pretraining(self): config_and_inputs = self.model_tester.prepare_config_and_inputs() self.model_tester.create_and_check_mobilebert_for_pretraining(*config_and_inputs) def test_for_question_answering(self): config_and_inputs = self.model_tester.prepare_config_and_inputs() self.model_tester.create_and_check_mobilebert_for_question_answering(*config_and_inputs) def test_for_sequence_classification(self): config_and_inputs = self.model_tester.prepare_config_and_inputs() self.model_tester.create_and_check_mobilebert_for_sequence_classification(*config_and_inputs) def test_for_token_classification(self): config_and_inputs = self.model_tester.prepare_config_and_inputs() self.model_tester.create_and_check_mobilebert_for_token_classification(*config_and_inputs) def _long_tensor(tok_lst): return torch.tensor( tok_lst, dtype=torch.long, device=torch_device, ) TOLERANCE = 1e-3 @require_torch @require_sentencepiece @require_tokenizers class MobileBertModelIntegrationTests(unittest.TestCase): @slow def test_inference_no_head(self): model = MobileBertModel.from_pretrained("google/mobilebert-uncased", attn_implementation="eager").to( torch_device ) input_ids = _long_tensor([[101, 7110, 1005, 1056, 2023, 11333, 17413, 1029, 102]]) with torch.no_grad(): output = model(input_ids)[0] expected_shape = torch.Size((1, 9, 512)) self.assertEqual(output.shape, expected_shape) expected_slice = torch.tensor( [ [ [-2.4736526e07, 8.2691656e04, 1.6521838e05], [-5.7541704e-01, 3.9056022e00, 4.4011507e00], [2.6047359e00, 1.5677652e00, -1.7324188e-01], ] ], device=torch_device, ) # MobileBERT results range from 10e0 to 10e8. Even a 0.0000001% difference with a value of 10e8 results in a # ~1 difference, it's therefore not a good idea to measure using addition. # Here, we instead divide the expected result with the result in order to obtain ~1. We then check that the # result is held between bounds: 1 - TOLERANCE < expected_result / result < 1 + TOLERANCE lower_bound = torch.all((expected_slice / output[..., :3, :3]) >= 1 - TOLERANCE) upper_bound = torch.all((expected_slice / output[..., :3, :3]) <= 1 + TOLERANCE) self.assertTrue(lower_bound and upper_bound) @pytest.mark.torch_export_test @slow def test_export(self): if version.parse(torch.__version__) < version.parse("2.4.0"): self.skipTest(reason="This test requires torch >= 2.4 to run.") mobilebert_model = "google/mobilebert-uncased" device = "cpu" attn_implementation = "eager" max_length = 512 tokenizer = AutoTokenizer.from_pretrained(mobilebert_model) inputs = tokenizer( f"the man worked as a {tokenizer.mask_token}.", return_tensors="pt", padding="max_length", max_length=max_length, ) model = MobileBertForMaskedLM.from_pretrained( mobilebert_model, device_map=device, attn_implementation=attn_implementation, ) logits = model(**inputs).logits eg_predicted_mask = tokenizer.decode(logits[0, 6].topk(5).indices) self.assertEqual(eg_predicted_mask.split(), ["carpenter", "waiter", "mechanic", "teacher", "clerk"]) exported_program = torch.export.export( model, args=(inputs["input_ids"],), kwargs={"attention_mask": inputs["attention_mask"]}, strict=True, ) result = exported_program.module().forward(inputs["input_ids"], inputs["attention_mask"]) ep_predicted_mask = tokenizer.decode(result.logits[0, 6].topk(5).indices) self.assertEqual(eg_predicted_mask, ep_predicted_mask)