# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved. # 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 paddle from parameterized import parameterized_class from paddlenlp.transformers import ( UIE, ErnieConfig, ErnieForMaskedLM, ErnieForMultipleChoice, ErnieForPretraining, ErnieForQuestionAnswering, ErnieForSequenceClassification, ErnieForTokenClassification, ErnieModel, ) from ...testing_utils import slow from ..test_modeling_common import ( ModelTesterMixin, ModelTesterPretrainedMixin, ids_tensor, random_attention_mask, ) class ErnieModelTester: def __init__( self, parent, batch_size=13, seq_length=7, is_training=True, use_input_mask=True, use_token_type_ids=True, vocab_size=99, hidden_size=32, num_hidden_layers=5, 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=2, initializer_range=0.02, pad_token_id=0, type_sequence_label_size=2, num_labels=3, num_choices=4, num_classes=3, 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.vocab_size = vocab_size self.hidden_size = hidden_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.initializer_range = initializer_range self.pad_token_id = pad_token_id self.type_sequence_label_size = type_sequence_label_size self.num_classes = num_classes 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.parent.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 ErnieConfig( 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, 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, initializer_range=self.initializer_range, pad_token_id=self.pad_token_id, num_class=self.num_classes, num_labels=self.num_labels, num_choices=self.num_choices, ) def create_and_check_model( self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels, ): model = ErnieModel(config) model.eval() result = model( input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, return_dict=self.parent.return_dict ) result = model(input_ids, token_type_ids=token_type_ids, return_dict=self.parent.return_dict) result = model(input_ids, return_dict=self.parent.return_dict) self.parent.assertEqual(result[0].shape, [self.batch_size, self.seq_length, self.hidden_size]) self.parent.assertEqual(result[1].shape, [self.batch_size, self.hidden_size]) def create_and_check_for_masked_lm( self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels, ): model = ErnieForMaskedLM(config) model.eval() result = model( input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=token_labels, return_dict=self.parent.return_dict, ) if token_labels is not None: result = result[1:] elif paddle.is_tensor(result): result = [result] self.parent.assertEqual(result[0].shape, [self.batch_size, self.seq_length, self.vocab_size]) def create_and_check_for_multiple_choice( self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels, ): model = ErnieForMultipleChoice(config) model.eval() multiple_choice_inputs_ids = input_ids.unsqueeze(1).expand([-1, self.num_choices, -1]) multiple_choice_token_type_ids = token_type_ids.unsqueeze(1).expand([-1, self.num_choices, -1]) multiple_choice_input_mask = input_mask.unsqueeze(1).expand([-1, self.num_choices, -1]) result = model( multiple_choice_inputs_ids, attention_mask=multiple_choice_input_mask, token_type_ids=multiple_choice_token_type_ids, labels=choice_labels, return_dict=self.parent.return_dict, ) if choice_labels is not None: result = result[1:] elif paddle.is_tensor(result): result = [result] self.parent.assertEqual(result[0].shape, [self.batch_size, self.num_choices]) def create_and_check_for_question_answering( self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels, ): model = ErnieForQuestionAnswering(config) model.eval() result = model( input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, start_positions=sequence_labels, end_positions=sequence_labels, return_dict=self.parent.return_dict, ) if sequence_labels is not None: start_logits, end_logits = result[1], result[2] else: start_logits, end_logits = result[0], result[1] self.parent.assertEqual(start_logits.shape, [self.batch_size, self.seq_length]) self.parent.assertEqual(end_logits.shape, [self.batch_size, self.seq_length]) def create_and_check_for_uie( self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels, ): model = UIE(config) model.eval() start_prob, end_prob = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids) self.parent.assertEqual(start_prob.shape, [self.batch_size, self.seq_length]) self.parent.assertEqual(end_prob.shape, [self.batch_size, self.seq_length]) def create_and_check_for_sequence_classification( self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels, ): model = ErnieForSequenceClassification(config) model.eval() result = model( input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=sequence_labels, return_dict=self.parent.return_dict, ) if sequence_labels is not None: result = result[1:] elif paddle.is_tensor(result): result = [result] self.parent.assertEqual(result[0].shape, [self.batch_size, self.num_classes]) def create_and_check_for_token_classification( self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels, ): model = ErnieForTokenClassification(config) model.eval() result = model( input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=token_labels, return_dict=self.parent.return_dict, ) if token_labels is not None: result = result[1:] elif paddle.is_tensor(result): result = [result] self.parent.assertEqual(result[0].shape, [self.batch_size, self.seq_length, self.num_classes]) def create_and_check_model_cache( self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels ): model = ErnieModel(config) model.eval() # first forward pass outputs = model(input_ids, attention_mask=input_mask, use_cache=True, return_dict=self.parent.return_dict) past_key_values = outputs.past_key_values if self.parent.return_dict else outputs[2] # create hypothetical multiple next token and extent to next_input_ids next_tokens = ids_tensor((self.batch_size, 3), self.vocab_size) next_mask = ids_tensor((self.batch_size, 3), vocab_size=2) # append to next input_ids and next_input_ids = paddle.concat([input_ids, next_tokens], axis=-1) next_attention_mask = paddle.concat([input_mask, next_mask], axis=-1) outputs = model( next_input_ids, attention_mask=next_attention_mask, output_hidden_states=True, return_dict=self.parent.return_dict, ) output_from_no_past = outputs[2][0] outputs = model( next_tokens, attention_mask=next_attention_mask, past_key_values=past_key_values, output_hidden_states=True, return_dict=self.parent.return_dict, ) output_from_past = outputs[2][0] # select random slice random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item() output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach() output_from_past_slice = output_from_past[:, :, random_slice_idx].detach() self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1]) # test that outputs are equal for slice self.parent.assertTrue(paddle.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3)) 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 @parameterized_class( ("return_dict", "use_labels"), [ [False, False], [False, True], [True, False], [True, True], ], ) class ErnieModelTest(ModelTesterMixin, unittest.TestCase): base_model_class = ErnieModel return_dict: bool = False use_labels: bool = False use_test_inputs_embeds: bool = True test_tie_weights = True all_model_classes = ( ErnieModel, ErnieForMaskedLM, ErnieForMultipleChoice, ErnieForPretraining, ErnieForQuestionAnswering, ErnieForSequenceClassification, ErnieForTokenClassification, UIE, ) def setUp(self): self.model_tester = ErnieModelTester(self) def test_model(self): config_and_inputs = self.model_tester.prepare_config_and_inputs() self.model_tester.create_and_check_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_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_for_multiple_choice(*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_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_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_for_token_classification(*config_and_inputs) def test_for_uie(self): config_and_inputs = self.model_tester.prepare_config_and_inputs() self.model_tester.create_and_check_for_uie(*config_and_inputs) def test_for_model_cache(self): config_and_inputs = self.model_tester.prepare_config_and_inputs() self.model_tester.create_and_check_model_cache(*config_and_inputs) @slow def test_model_from_pretrained(self): for model_name in list(ErnieModel.pretrained_init_configuration)[:1]: model = ErnieModel.from_pretrained(model_name) self.assertIsNotNone(model) class ErnieModelIntegrationTest(unittest.TestCase, ModelTesterPretrainedMixin): base_model_class = ErnieModel hf_remote_test_model_path = "PaddleCI/tiny-random-ernie" paddlehub_remote_test_model_path = "__internal_testing__/tiny-random-ernie" @slow def test_inference_no_attention(self): model = ErnieModel.from_pretrained("ernie-1.0") model.eval() input_ids = paddle.to_tensor([[0, 345, 232, 328, 740, 140, 1695, 69, 6078, 1588, 2]]) with paddle.no_grad(): output = model(input_ids)[0] expected_shape = [1, 11, 768] self.assertEqual(output.shape, expected_shape) expected_slice = paddle.to_tensor( [[[-0.0664, -1.3266, -0.3922], [-0.2440, -1.3631, -1.0745], [-0.0986, -0.7947, -0.1932]]] ) self.assertTrue(paddle.allclose(output[:, 1:4, 1:4], expected_slice, atol=1e-4)) @slow def test_inference_with_attention(self): model = ErnieModel.from_pretrained("ernie-1.0") model.eval() input_ids = paddle.to_tensor([[0, 345, 232, 328, 740, 140, 1695, 69, 6078, 1588, 2]]) attention_mask = paddle.to_tensor([[0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]]) with paddle.no_grad(): output = model(input_ids, attention_mask=attention_mask)[0] expected_shape = [1, 11, 768] self.assertEqual(output.shape, expected_shape) expected_slice = paddle.to_tensor( [[[-0.0664, -1.3266, -0.3922], [-0.2440, -1.3631, -1.0745], [-0.0986, -0.7947, -0.1932]]] ) self.assertTrue(paddle.allclose(output[:, 1:4, 1:4], expected_slice, atol=1e-4)) @slow def test_inference_with_past_key_value(self): model = ErnieModel.from_pretrained("ernie-1.0") model.eval() input_ids = paddle.to_tensor([[0, 345, 232, 328, 740, 140, 1695, 69, 6078, 1588, 2]]) attention_mask = paddle.to_tensor([[0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]]) with paddle.no_grad(): output = model(input_ids, attention_mask=attention_mask, use_cache=True, return_dict=True) expected_shape = [1, 11, 768] self.assertEqual(output[0].shape, expected_shape) expected_slice = paddle.to_tensor( [ [ [-0.06635337, -1.32662833, -0.39223742], [-0.24396378, -1.36314595, -1.07446611], [-0.09860237, -0.79468340, -0.19317953], ] ] ) self.assertTrue(paddle.allclose(output[0][:, 1:4, 1:4], expected_slice, atol=1e-4)) # insert the past key value into model with paddle.no_grad(): output = model(input_ids, use_cache=True, past_key_values=output.past_key_values, return_dict=True) expected_slice = paddle.to_tensor( [ [ [0.07865857, -1.07012558, -1.02596498], [0.12576045, -1.76696026, -1.18682015], [-0.08606864, -1.37880838, -0.12364164], ] ] ) self.assertTrue(paddle.allclose(output[0][:, 1:4, 1:4], expected_slice, atol=1e-4)) if __name__ == "__main__": unittest.main()