transformers/tests/test_modeling_xlm_roberta_xl.py

# coding=utf-8
# Copyright 2022 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

from transformers import XLMRobertaXLConfig, is_torch_available
from transformers.testing_utils import require_torch, slow, torch_device

from .test_configuration_common import ConfigTester
from .test_generation_utils import GenerationTesterMixin
from .test_modeling_common import ModelTesterMixin, floats_tensor, ids_tensor, random_attention_mask


if is_torch_available():
    import torch

    from transformers import (
        XLMRobertaXLForCausalLM,
        XLMRobertaXLForMaskedLM,
        XLMRobertaXLForMultipleChoice,
        XLMRobertaXLForQuestionAnswering,
        XLMRobertaXLForSequenceClassification,
        XLMRobertaXLForTokenClassification,
        XLMRobertaXLModel,
    )
    from transformers.models.xlm_roberta_xl.modeling_xlm_roberta_xl import (
        XLMRobertaXLEmbeddings,
        create_position_ids_from_input_ids,
    )


class XLMRobertaXLModelTester:
    def __init__(
        self,
        parent,
    ):
        self.parent = parent
        self.batch_size = 13
        self.seq_length = 7
        self.is_training = True
        self.use_input_mask = True
        self.use_token_type_ids = True
        self.use_labels = True
        self.vocab_size = 99
        self.hidden_size = 32
        self.num_hidden_layers = 5
        self.num_attention_heads = 4
        self.intermediate_size = 37
        self.hidden_act = "gelu"
        self.hidden_dropout_prob = 0.1
        self.attention_probs_dropout_prob = 0.1
        self.max_position_embeddings = 512
        self.type_vocab_size = 16
        self.type_sequence_label_size = 2
        self.initializer_range = 0.02
        self.num_labels = 3
        self.num_choices = 4
        self.scope = None

    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 XLMRobertaXLConfig(
            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,
        )

    def prepare_config_and_inputs_for_decoder(self):
        (
            config,
            input_ids,
            token_type_ids,
            input_mask,
            sequence_labels,
            token_labels,
            choice_labels,
        ) = self.prepare_config_and_inputs()

        config.is_decoder = True
        encoder_hidden_states = floats_tensor([self.batch_size, self.seq_length, self.hidden_size])
        encoder_attention_mask = ids_tensor([self.batch_size, self.seq_length], vocab_size=2)

        return (
            config,
            input_ids,
            token_type_ids,
            input_mask,
            sequence_labels,
            token_labels,
            choice_labels,
            encoder_hidden_states,
            encoder_attention_mask,
        )

    def create_and_check_model(
        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
    ):
        model = XLMRobertaXLModel(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_model_as_decoder(
        self,
        config,
        input_ids,
        token_type_ids,
        input_mask,
        sequence_labels,
        token_labels,
        choice_labels,
        encoder_hidden_states,
        encoder_attention_mask,
    ):
        config.add_cross_attention = True
        model = XLMRobertaXLModel(config)
        model.to(torch_device)
        model.eval()
        result = model(
            input_ids,
            attention_mask=input_mask,
            token_type_ids=token_type_ids,
            encoder_hidden_states=encoder_hidden_states,
            encoder_attention_mask=encoder_attention_mask,
        )
        result = model(
            input_ids,
            attention_mask=input_mask,
            token_type_ids=token_type_ids,
            encoder_hidden_states=encoder_hidden_states,
        )
        result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_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_for_causal_lm(
        self,
        config,
        input_ids,
        token_type_ids,
        input_mask,
        sequence_labels,
        token_labels,
        choice_labels,
        encoder_hidden_states,
        encoder_attention_mask,
    ):
        model = XLMRobertaXLForCausalLM(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_decoder_model_past_large_inputs(
        self,
        config,
        input_ids,
        token_type_ids,
        input_mask,
        sequence_labels,
        token_labels,
        choice_labels,
        encoder_hidden_states,
        encoder_attention_mask,
    ):
        config.is_decoder = True
        config.add_cross_attention = True
        model = XLMRobertaXLForCausalLM(config=config).to(torch_device).eval()

        # make sure that ids don't start with pad token
        mask = input_ids.ne(config.pad_token_id).long()
        input_ids = input_ids * mask

        # first forward pass
        outputs = model(
            input_ids,
            attention_mask=input_mask,
            encoder_hidden_states=encoder_hidden_states,
            encoder_attention_mask=encoder_attention_mask,
            use_cache=True,
        )
        past_key_values = outputs.past_key_values

        # create hypothetical multiple next token and extent to next_input_ids
        next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)

        # make sure that ids don't start with pad token
        mask = next_tokens.ne(config.pad_token_id).long()
        next_tokens = next_tokens * mask
        next_mask = ids_tensor((self.batch_size, 3), vocab_size=2)

        # append to next input_ids and
        next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
        next_attention_mask = torch.cat([input_mask, next_mask], dim=-1)

        output_from_no_past = model(
            next_input_ids,
            attention_mask=next_attention_mask,
            encoder_hidden_states=encoder_hidden_states,
            encoder_attention_mask=encoder_attention_mask,
            output_hidden_states=True,
        )["hidden_states"][0]
        output_from_past = model(
            next_tokens,
            attention_mask=next_attention_mask,
            encoder_hidden_states=encoder_hidden_states,
            encoder_attention_mask=encoder_attention_mask,
            past_key_values=past_key_values,
            output_hidden_states=True,
        )["hidden_states"][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(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))

    def create_and_check_for_masked_lm(
        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
    ):
        model = XLMRobertaXLForMaskedLM(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_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 = XLMRobertaXLForTokenClassification(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_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 = XLMRobertaXLForMultipleChoice(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 create_and_check_for_question_answering(
        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
    ):
        model = XLMRobertaXLForQuestionAnswering(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 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 XLMRobertaXLModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase):

    all_model_classes = (
        (
            XLMRobertaXLForCausalLM,
            XLMRobertaXLForMaskedLM,
            XLMRobertaXLModel,
            XLMRobertaXLForSequenceClassification,
            XLMRobertaXLForTokenClassification,
            XLMRobertaXLForMultipleChoice,
            XLMRobertaXLForQuestionAnswering,
        )
        if is_torch_available()
        else ()
    )
    all_generative_model_classes = (XLMRobertaXLForCausalLM,) if is_torch_available() else ()

    def setUp(self):
        self.model_tester = XLMRobertaXLModelTester(self)
        self.config_tester = ConfigTester(self, config_class=XLMRobertaXLConfig, hidden_size=37)

    def test_config(self):
        self.config_tester.run_common_tests()

    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_model_various_embeddings(self):
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
        for type in ["absolute", "relative_key", "relative_key_query"]:
            config_and_inputs[0].position_embedding_type = type
            self.model_tester.create_and_check_model(*config_and_inputs)

    def test_model_as_decoder(self):
        config_and_inputs = self.model_tester.prepare_config_and_inputs_for_decoder()
        self.model_tester.create_and_check_model_as_decoder(*config_and_inputs)

    def test_model_as_decoder_with_default_input_mask(self):
        # This regression test was failing with PyTorch < 1.3
        (
            config,
            input_ids,
            token_type_ids,
            input_mask,
            sequence_labels,
            token_labels,
            choice_labels,
            encoder_hidden_states,
            encoder_attention_mask,
        ) = self.model_tester.prepare_config_and_inputs_for_decoder()

        input_mask = None

        self.model_tester.create_and_check_model_as_decoder(
            config,
            input_ids,
            token_type_ids,
            input_mask,
            sequence_labels,
            token_labels,
            choice_labels,
            encoder_hidden_states,
            encoder_attention_mask,
        )

    def test_for_causal_lm(self):
        config_and_inputs = self.model_tester.prepare_config_and_inputs_for_decoder()
        self.model_tester.create_and_check_for_causal_lm(*config_and_inputs)

    def test_decoder_model_past_with_large_inputs(self):
        config_and_inputs = self.model_tester.prepare_config_and_inputs_for_decoder()
        self.model_tester.create_and_check_decoder_model_past_large_inputs(*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_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_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_create_position_ids_respects_padding_index(self):
        """Ensure that the default position ids only assign a sequential . This is a regression
        test for https://github.com/huggingface/transformers/issues/1761

        The position ids should be masked with the embedding object's padding index. Therefore, the
        first available non-padding position index is XLMRobertaXLEmbeddings.padding_idx + 1
        """
        config = self.model_tester.prepare_config_and_inputs()[0]
        model = XLMRobertaXLEmbeddings(config=config)

        input_ids = torch.as_tensor([[12, 31, 13, model.padding_idx]])
        expected_positions = torch.as_tensor(
            [[0 + model.padding_idx + 1, 1 + model.padding_idx + 1, 2 + model.padding_idx + 1, model.padding_idx]]
        )

        position_ids = create_position_ids_from_input_ids(input_ids, model.padding_idx)
        self.assertEqual(position_ids.shape, expected_positions.shape)
        self.assertTrue(torch.all(torch.eq(position_ids, expected_positions)))

    def test_create_position_ids_from_inputs_embeds(self):
        """Ensure that the default position ids only assign a sequential . This is a regression
        test for https://github.com/huggingface/transformers/issues/1761

        The position ids should be masked with the embedding object's padding index. Therefore, the
        first available non-padding position index is XLMRobertaXLEmbeddings.padding_idx + 1
        """
        config = self.model_tester.prepare_config_and_inputs()[0]
        embeddings = XLMRobertaXLEmbeddings(config=config)

        inputs_embeds = torch.empty(2, 4, 30)
        expected_single_positions = [
            0 + embeddings.padding_idx + 1,
            1 + embeddings.padding_idx + 1,
            2 + embeddings.padding_idx + 1,
            3 + embeddings.padding_idx + 1,
        ]
        expected_positions = torch.as_tensor([expected_single_positions, expected_single_positions])
        position_ids = embeddings.create_position_ids_from_inputs_embeds(inputs_embeds)
        self.assertEqual(position_ids.shape, expected_positions.shape)
        self.assertTrue(torch.all(torch.eq(position_ids, expected_positions)))


@require_torch
class XLMRobertaModelXLIntegrationTest(unittest.TestCase):
    @slow
    def test_xlm_roberta_xl(self):
        model = XLMRobertaXLModel.from_pretrained("facebook/xlm-roberta-xl").to(torch_device)
        input_ids = torch.tensor(
            [[0, 581, 10269, 83, 99942, 136, 60742, 23, 70, 80583, 18276, 2]], device=torch_device
        )
        # The dog is cute and lives in the garden house

        expected_output_shape = torch.Size((1, 12, 2560))  # batch_size, sequence_length, embedding_vector_dim
        expected_output_values_last_dim = torch.tensor(
            [[0.0110, 0.0605, 0.0354, 0.0689, 0.0066, 0.0691, 0.0302, 0.0412, 0.0860, 0.0036, 0.0405, 0.0170]],
            device=torch_device,
        )

        output = model(input_ids)["last_hidden_state"].detach()
        self.assertEqual(output.shape, expected_output_shape)
        # compare the actual values for a slice of last dim
        self.assertTrue(torch.allclose(output[:, :, -1], expected_output_values_last_dim, atol=1e-3))

    @unittest.skip(reason="Model is too large to be tested on the CI")
    def test_xlm_roberta_xxl(self):
        model = XLMRobertaXLModel.from_pretrained("facebook/xlm-roberta-xxl").to(torch_device)
        input_ids = torch.tensor(
            [[0, 581, 10269, 83, 99942, 136, 60742, 23, 70, 80583, 18276, 2]], device=torch_device
        )
        # The dog is cute and lives in the garden house

        expected_output_shape = torch.Size((1, 12, 4096))  # batch_size, sequence_length, embedding_vector_dim
        expected_output_values_last_dim = torch.tensor(
            [[0.0046, 0.0146, 0.0227, 0.0126, 0.0219, 0.0175, -0.0101, 0.0006, 0.0124, 0.0209, -0.0063, 0.0096]],
            device=torch_device,
        )

        output = model(input_ids)["last_hidden_state"].detach()
        self.assertEqual(output.shape, expected_output_shape)
        # compare the actual values for a slice of last dim
        self.assertTrue(torch.allclose(output[:, :, -1], expected_output_values_last_dim, atol=1e-3))