test_modeling_deit.py

# coding=utf-8
# Copyright 2021 The HuggingFace Inc. 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.
""" Testing suite for the PyTorch DeiT model. """


import inspect
import unittest
import warnings

from transformers import DeiTConfig
from transformers.file_utils import cached_property, is_torch_available, is_vision_available
from transformers.models.auto import get_values
from transformers.testing_utils import require_torch, require_vision, slow, torch_device

from ..test_configuration_common import ConfigTester
from ..test_modeling_common import ModelTesterMixin, floats_tensor, ids_tensor


if is_torch_available():
    import torch
    from torch import nn

    from transformers import (
        MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING,
        MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING,
        MODEL_MAPPING,
        DeiTForImageClassification,
        DeiTForImageClassificationWithTeacher,
        DeiTForMaskedImageModeling,
        DeiTModel,
    )
    from transformers.models.deit.modeling_deit import DEIT_PRETRAINED_MODEL_ARCHIVE_LIST, to_2tuple


if is_vision_available():
    from PIL import Image

    from transformers import DeiTFeatureExtractor


class DeiTModelTester:
    def __init__(
        self,
        parent,
        batch_size=13,
        image_size=30,
        patch_size=2,
        num_channels=3,
        is_training=True,
        use_labels=True,
        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,
        type_sequence_label_size=10,
        initializer_range=0.02,
        num_labels=3,
        scope=None,
        encoder_stride=2,
    ):
        self.parent = parent
        self.batch_size = batch_size
        self.image_size = image_size
        self.patch_size = patch_size
        self.num_channels = num_channels
        self.is_training = is_training
        self.use_labels = use_labels
        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.type_sequence_label_size = type_sequence_label_size
        self.initializer_range = initializer_range
        self.scope = scope
        self.encoder_stride = encoder_stride

    def prepare_config_and_inputs(self):
        pixel_values = floats_tensor([self.batch_size, self.num_channels, self.image_size, self.image_size])

        labels = None
        if self.use_labels:
            labels = ids_tensor([self.batch_size], self.type_sequence_label_size)

        config = self.get_config()

        return config, pixel_values, labels

    def get_config(self):
        return DeiTConfig(
            image_size=self.image_size,
            patch_size=self.patch_size,
            num_channels=self.num_channels,
            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,
            is_decoder=False,
            initializer_range=self.initializer_range,
            encoder_stride=self.encoder_stride,
        )

    def create_and_check_model(self, config, pixel_values, labels):
        model = DeiTModel(config=config)
        model.to(torch_device)
        model.eval()
        result = model(pixel_values)
        # expected sequence length = num_patches + 2 (we add 2 for the [CLS] and distillation tokens)
        image_size = to_2tuple(self.image_size)
        patch_size = to_2tuple(self.patch_size)
        num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
        self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, num_patches + 2, self.hidden_size))

    def create_and_check_for_image_classification(self, config, pixel_values, labels):
        config.num_labels = self.type_sequence_label_size
        model = DeiTForImageClassification(config)
        model.to(torch_device)
        model.eval()
        result = model(pixel_values, labels=labels)
        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.type_sequence_label_size))

    def prepare_config_and_inputs_for_common(self):
        config_and_inputs = self.prepare_config_and_inputs()
        (
            config,
            pixel_values,
            labels,
        ) = config_and_inputs
        inputs_dict = {"pixel_values": pixel_values}
        return config, inputs_dict


@require_torch
class DeiTModelTest(ModelTesterMixin, unittest.TestCase):
    """
    Here we also overwrite some of the tests of test_modeling_common.py, as DeiT does not use input_ids, inputs_embeds,
    attention_mask and seq_length.
    """

    all_model_classes = (
        (
            DeiTModel,
            DeiTForImageClassification,
            DeiTForImageClassificationWithTeacher,
            DeiTForMaskedImageModeling,
        )
        if is_torch_available()
        else ()
    )

    test_pruning = False
    test_torchscript = False
    test_resize_embeddings = False
    test_head_masking = False

    def setUp(self):
        self.model_tester = DeiTModelTester(self)
        self.config_tester = ConfigTester(self, config_class=DeiTConfig, has_text_modality=False, hidden_size=37)

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

    def test_inputs_embeds(self):
        # DeiT does not use inputs_embeds
        pass

    def test_model_common_attributes(self):
        config, _ = 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(), (nn.Module))
            x = model.get_output_embeddings()
            self.assertTrue(x is None or isinstance(x, nn.Linear))

    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()]

            expected_arg_names = ["pixel_values"]
            self.assertListEqual(arg_names[:1], expected_arg_names)

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

        # in DeiT, the seq_len equals the number of patches + 2 (we add 2 for the [CLS] and distillation tokens)
        image_size = to_2tuple(self.model_tester.image_size)
        patch_size = to_2tuple(self.model_tester.patch_size)
        num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
        seq_len = num_patches + 2
        encoder_seq_length = getattr(self.model_tester, "encoder_seq_length", seq_len)
        encoder_key_length = getattr(self.model_tester, "key_length", encoder_seq_length)
        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

        for model_class in self.all_model_classes:
            inputs_dict["output_attentions"] = True
            inputs_dict["output_hidden_states"] = False
            config.return_dict = True
            model = model_class(config)
            model.to(torch_device)
            model.eval()
            with torch.no_grad():
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
            attentions = outputs.encoder_attentions if config.is_encoder_decoder else outputs.attentions
            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():
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
            attentions = outputs.encoder_attentions if config.is_encoder_decoder else outputs.attentions
            self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)

            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],
                )
            out_len = len(outputs)

            # Check attention is always last and order is fine
            inputs_dict["output_attentions"] = True
            inputs_dict["output_hidden_states"] = True
            model = model_class(config)
            model.to(torch_device)
            model.eval()
            with torch.no_grad():
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))

            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))

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

            self.assertEqual(len(self_attentions), self.model_tester.num_hidden_layers)
            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],
                )

    def test_hidden_states_output(self):
        def check_hidden_states_output(inputs_dict, config, model_class):
            model = model_class(config)
            model.to(torch_device)
            model.eval()

            with torch.no_grad():
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))

            hidden_states = outputs.encoder_hidden_states if config.is_encoder_decoder else outputs.hidden_states

            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)

            # DeiT has a different seq_length
            image_size = to_2tuple(self.model_tester.image_size)
            patch_size = to_2tuple(self.model_tester.patch_size)
            num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
            seq_length = num_patches + 2

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

        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)

    # special case for DeiTForImageClassificationWithTeacher 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.__name__ == "DeiTForImageClassificationWithTeacher":
                del inputs_dict["labels"]

        return inputs_dict

    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:
            # DeiTForImageClassificationWithTeacher supports inference-only
            if (
                model_class in get_values(MODEL_MAPPING)
                or model_class.__name__ == "DeiTForImageClassificationWithTeacher"
            ):
                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:
            return

        config.use_cache = False
        config.return_dict = True

        for model_class in self.all_model_classes:
            if model_class in get_values(MODEL_MAPPING) or not model_class.supports_gradient_checkpointing:
                continue
            # DeiTForImageClassificationWithTeacher supports inference-only
            if model_class.__name__ == "DeiTForImageClassificationWithTeacher":
                continue
            model = model_class(config)
            model.gradient_checkpointing_enable()
            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_problem_types(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        problem_types = [
            {"title": "multi_label_classification", "num_labels": 2, "dtype": torch.float},
            {"title": "single_label_classification", "num_labels": 1, "dtype": torch.long},
            {"title": "regression", "num_labels": 1, "dtype": torch.float},
        ]

        for model_class in self.all_model_classes:
            if (
                model_class
                not in [
                    *get_values(MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING),
                    *get_values(MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING),
                ]
                or model_class.__name__ == "DeiTForImageClassificationWithTeacher"
            ):
                continue

            for problem_type in problem_types:
                with self.subTest(msg=f"Testing {model_class} with {problem_type['title']}"):

                    config.problem_type = problem_type["title"]
                    config.num_labels = problem_type["num_labels"]

                    model = model_class(config)
                    model.to(torch_device)
                    model.train()

                    inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)

                    if problem_type["num_labels"] > 1:
                        inputs["labels"] = inputs["labels"].unsqueeze(1).repeat(1, problem_type["num_labels"])

                    inputs["labels"] = inputs["labels"].to(problem_type["dtype"])

                    # This tests that we do not trigger the warning form PyTorch "Using a target size that is different
                    # to the input size. This will likely lead to incorrect results due to broadcasting. Please ensure
                    # they have the same size." which is a symptom something in wrong for the regression problem.
                    # See https://github.com/huggingface/transformers/issues/11780
                    with warnings.catch_warnings(record=True) as warning_list:
                        loss = model(**inputs).loss
                    for w in warning_list:
                        if "Using a target size that is different to the input size" in str(w.message):
                            raise ValueError(
                                f"Something is going wrong in the regression problem: intercepted {w.message}"
                            )

                    loss.backward()

    def test_for_image_classification(self):
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
        self.model_tester.create_and_check_for_image_classification(*config_and_inputs)

    @slow
    def test_model_from_pretrained(self):
        for model_name in DEIT_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
            model = DeiTModel.from_pretrained(model_name)
            self.assertIsNotNone(model)


# We will verify our results on an image of cute cats
def prepare_img():
    image = Image.open("./tests/fixtures/tests_samples/COCO/000000039769.png")
    return image


@require_torch
@require_vision
class DeiTModelIntegrationTest(unittest.TestCase):
    @cached_property
    def default_feature_extractor(self):
        return (
            DeiTFeatureExtractor.from_pretrained("facebook/deit-base-distilled-patch16-224")
            if is_vision_available()
            else None
        )

    @slow
    def test_inference_image_classification_head(self):
        model = DeiTForImageClassificationWithTeacher.from_pretrained("facebook/deit-base-distilled-patch16-224").to(
            torch_device
        )

        feature_extractor = self.default_feature_extractor
        image = prepare_img()
        inputs = feature_extractor(images=image, return_tensors="pt").to(torch_device)

        # forward pass
        outputs = model(**inputs)

        # verify the logits
        expected_shape = torch.Size((1, 1000))
        self.assertEqual(outputs.logits.shape, expected_shape)

        expected_slice = torch.tensor([-1.0266, 0.1912, -1.2861]).to(torch_device)

        self.assertTrue(torch.allclose(outputs.logits[0, :3], expected_slice, atol=1e-4))