test_modeling_megatron_bert.py

# Copyright 2021 The HuggingFace Inc. team. All rights reserved.
# Copyright 2021 NVIDIA Corporation. 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.
""" Testing suite for the PyTorch MegatronBERT model. """


import math
import os
import unittest

from transformers import 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


if is_torch_available():
    import torch

    from transformers import (
        MODEL_FOR_PRETRAINING_MAPPING,
        MegatronBertConfig,
        MegatronBertForCausalLM,
        MegatronBertForMaskedLM,
        MegatronBertForMultipleChoice,
        MegatronBertForNextSentencePrediction,
        MegatronBertForPreTraining,
        MegatronBertForQuestionAnswering,
        MegatronBertForSequenceClassification,
        MegatronBertForTokenClassification,
        MegatronBertModel,
    )


class MegatronBertModelTester:
    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=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=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 = MegatronBertConfig(
            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,
        )

        return config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels

    def create_and_check_megatron_bert_model(
        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
    ):
        model = MegatronBertModel(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_megatron_bert_for_masked_lm(
        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
    ):
        model = MegatronBertForMaskedLM(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_causal_lm(
        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
    ):
        model = MegatronBertForCausalLM(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_megatron_bert_for_next_sequence_prediction(
        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
    ):
        model = MegatronBertForNextSentencePrediction(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_megatron_bert_for_pretraining(
        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
    ):
        model = MegatronBertForPreTraining(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_megatron_bert_for_question_answering(
        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
    ):
        model = MegatronBertForQuestionAnswering(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_megatron_bert_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 = MegatronBertForSequenceClassification(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_megatron_bert_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 = MegatronBertForTokenClassification(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_megatron_bert_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 = MegatronBertForMultipleChoice(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 MegatronBertModelTest(ModelTesterMixin, unittest.TestCase):

    all_model_classes = (
        (
            MegatronBertModel,
            MegatronBertForMaskedLM,
            MegatronBertForCausalLM,
            MegatronBertForMultipleChoice,
            MegatronBertForNextSentencePrediction,
            MegatronBertForPreTraining,
            MegatronBertForQuestionAnswering,
            MegatronBertForSequenceClassification,
            MegatronBertForTokenClassification,
        )
        if is_torch_available()
        else ()
    )
    fx_ready_model_classes = all_model_classes

    # test_resize_embeddings = False
    test_head_masking = False

    # 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

    def setUp(self):
        self.model_tester = MegatronBertModelTester(self)
        self.config_tester = ConfigTester(self, config_class=MegatronBertConfig, hidden_size=37)

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

    def test_megatron_bert_model(self):
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
        self.model_tester.create_and_check_megatron_bert_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_megatron_bert_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_megatron_bert_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_megatron_bert_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_megatron_bert_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_megatron_bert_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_megatron_bert_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_megatron_bert_for_token_classification(*config_and_inputs)


def _long_tensor(tok_lst):
    return torch.tensor(
        tok_lst,
        dtype=torch.long,
        device=torch_device,
    )


TOLERANCE = 1e-4


@require_torch
@require_sentencepiece
@require_tokenizers
class MegatronBertModelIntegrationTests(unittest.TestCase):
    @slow
    def test_inference_no_head(self):
        directory = "nvidia/megatron-bert-uncased-345m"
        if "MYDIR" in os.environ:
            directory = os.path.join(os.environ["MYDIR"], directory)
        model = MegatronBertModel.from_pretrained(directory)
        model.to(torch_device)
        model.half()
        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, 1024))
        self.assertEqual(output.shape, expected_shape)

        expected = [-0.6040, -0.2517, -0.1025, 0.3420, -0.6758, -0.0017, -0.1089, -0.1990, 0.5728]
        for ii in range(3):
            for jj in range(3):
                a = output[0, ii, jj]
                b = expected[3 * ii + jj]
                msg = "ii={} jj={} a={} b={}".format(ii, jj, a, b)
                self.assertTrue(math.isclose(a, b, rel_tol=TOLERANCE, abs_tol=TOLERANCE), msg=msg)