init

transformers/tests/models/beit/test_image_processing_beit.py

@@ -0,0 +1,313 @@
+# Copyright 2021 HuggingFace Inc.
+#
+# 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 datasets import load_dataset
+
+from transformers.testing_utils import require_torch, require_vision
+from transformers.utils import is_torch_available, is_torchvision_available, is_vision_available
+
+from ...test_image_processing_common import ImageProcessingTestMixin, prepare_image_inputs
+
+
+if is_torch_available():
+    import torch
+
+if is_vision_available():
+    from transformers import BeitImageProcessor
+
+    if is_torchvision_available():
+        from transformers import BeitImageProcessorFast
+
+
+class BeitImageProcessingTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=7,
+        num_channels=3,
+        image_size=18,
+        min_resolution=30,
+        max_resolution=400,
+        do_resize=True,
+        size=None,
+        do_center_crop=True,
+        crop_size=None,
+        do_normalize=True,
+        image_mean=[0.5, 0.5, 0.5],
+        image_std=[0.5, 0.5, 0.5],
+        do_reduce_labels=False,
+    ):
+        size = size if size is not None else {"height": 20, "width": 20}
+        crop_size = crop_size if crop_size is not None else {"height": 18, "width": 18}
+        self.parent = parent
+        self.batch_size = batch_size
+        self.num_channels = num_channels
+        self.image_size = image_size
+        self.min_resolution = min_resolution
+        self.max_resolution = max_resolution
+        self.do_resize = do_resize
+        self.size = size
+        self.do_center_crop = do_center_crop
+        self.crop_size = crop_size
+        self.do_normalize = do_normalize
+        self.image_mean = image_mean
+        self.image_std = image_std
+        self.do_reduce_labels = do_reduce_labels
+
+    def prepare_image_processor_dict(self):
+        return {
+            "do_resize": self.do_resize,
+            "size": self.size,
+            "do_center_crop": self.do_center_crop,
+            "crop_size": self.crop_size,
+            "do_normalize": self.do_normalize,
+            "image_mean": self.image_mean,
+            "image_std": self.image_std,
+            "do_reduce_labels": self.do_reduce_labels,
+        }
+
+    def expected_output_image_shape(self, images):
+        return self.num_channels, self.crop_size["height"], self.crop_size["width"]
+
+    def prepare_image_inputs(self, equal_resolution=False, numpify=False, torchify=False):
+        return prepare_image_inputs(
+            batch_size=self.batch_size,
+            num_channels=self.num_channels,
+            min_resolution=self.min_resolution,
+            max_resolution=self.max_resolution,
+            equal_resolution=equal_resolution,
+            numpify=numpify,
+            torchify=torchify,
+        )
+
+
+def prepare_semantic_single_inputs():
+    ds = load_dataset("hf-internal-testing/fixtures_ade20k", split="test")
+    example = ds[0]
+    return example["image"], example["map"]
+
+
+def prepare_semantic_batch_inputs():
+    ds = load_dataset("hf-internal-testing/fixtures_ade20k", split="test")
+    return list(ds["image"][:2]), list(ds["map"][:2])
+
+
+@require_torch
+@require_vision
+class BeitImageProcessingTest(ImageProcessingTestMixin, unittest.TestCase):
+    image_processing_class = BeitImageProcessor if is_vision_available() else None
+    fast_image_processing_class = BeitImageProcessorFast if is_torchvision_available() else None
+
+    def setUp(self):
+        super().setUp()
+        self.image_processor_tester = BeitImageProcessingTester(self)
+
+    @property
+    def image_processor_dict(self):
+        return self.image_processor_tester.prepare_image_processor_dict()
+
+    def test_image_processor_properties(self):
+        for image_processing_class in self.image_processor_list:
+            image_processing = image_processing_class(**self.image_processor_dict)
+            self.assertTrue(hasattr(image_processing, "do_resize"))
+            self.assertTrue(hasattr(image_processing, "size"))
+            self.assertTrue(hasattr(image_processing, "do_center_crop"))
+            self.assertTrue(hasattr(image_processing, "center_crop"))
+            self.assertTrue(hasattr(image_processing, "do_normalize"))
+            self.assertTrue(hasattr(image_processing, "image_mean"))
+            self.assertTrue(hasattr(image_processing, "image_std"))
+            self.assertTrue(hasattr(image_processing, "do_reduce_labels"))
+
+    def test_image_processor_from_dict_with_kwargs(self):
+        for image_processing_class in self.image_processor_list:
+            image_processor = image_processing_class.from_dict(self.image_processor_dict)
+            self.assertEqual(image_processor.size, {"height": 20, "width": 20})
+            self.assertEqual(image_processor.crop_size, {"height": 18, "width": 18})
+            self.assertEqual(image_processor.do_reduce_labels, False)
+
+            image_processor = image_processing_class.from_dict(
+                self.image_processor_dict, size=42, crop_size=84, do_reduce_labels=True
+            )
+            self.assertEqual(image_processor.size, {"height": 42, "width": 42})
+            self.assertEqual(image_processor.crop_size, {"height": 84, "width": 84})
+            self.assertEqual(image_processor.do_reduce_labels, True)
+
+    def test_call_segmentation_maps(self):
+        for image_processing_class in self.image_processor_list:
+            # Initialize image_processing
+            image_processing = image_processing_class(**self.image_processor_dict)
+            # create random PyTorch tensors
+            image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=False, torchify=True)
+            maps = []
+            for image in image_inputs:
+                self.assertIsInstance(image, torch.Tensor)
+                maps.append(torch.zeros(image.shape[-2:]).long())
+
+            # Test not batched input
+            encoding = image_processing(image_inputs[0], maps[0], return_tensors="pt")
+            self.assertEqual(
+                encoding["pixel_values"].shape,
+                (
+                    1,
+                    self.image_processor_tester.num_channels,
+                    self.image_processor_tester.crop_size["height"],
+                    self.image_processor_tester.crop_size["width"],
+                ),
+            )
+            self.assertEqual(
+                encoding["labels"].shape,
+                (
+                    1,
+                    self.image_processor_tester.crop_size["height"],
+                    self.image_processor_tester.crop_size["width"],
+                ),
+            )
+            self.assertEqual(encoding["labels"].dtype, torch.long)
+            self.assertTrue(encoding["labels"].min().item() >= 0)
+            self.assertTrue(encoding["labels"].max().item() <= 255)
+
+            # Test batched
+            encoding = image_processing(image_inputs, maps, return_tensors="pt")
+            self.assertEqual(
+                encoding["pixel_values"].shape,
+                (
+                    self.image_processor_tester.batch_size,
+                    self.image_processor_tester.num_channels,
+                    self.image_processor_tester.crop_size["height"],
+                    self.image_processor_tester.crop_size["width"],
+                ),
+            )
+            self.assertEqual(
+                encoding["labels"].shape,
+                (
+                    self.image_processor_tester.batch_size,
+                    self.image_processor_tester.crop_size["height"],
+                    self.image_processor_tester.crop_size["width"],
+                ),
+            )
+            self.assertEqual(encoding["labels"].dtype, torch.long)
+            self.assertTrue(encoding["labels"].min().item() >= 0)
+            self.assertTrue(encoding["labels"].max().item() <= 255)
+
+            # Test not batched input (PIL images)
+            image, segmentation_map = prepare_semantic_single_inputs()
+
+            encoding = image_processing(image, segmentation_map, return_tensors="pt")
+            self.assertEqual(
+                encoding["pixel_values"].shape,
+                (
+                    1,
+                    self.image_processor_tester.num_channels,
+                    self.image_processor_tester.crop_size["height"],
+                    self.image_processor_tester.crop_size["width"],
+                ),
+            )
+            self.assertEqual(
+                encoding["labels"].shape,
+                (
+                    1,
+                    self.image_processor_tester.crop_size["height"],
+                    self.image_processor_tester.crop_size["width"],
+                ),
+            )
+            self.assertEqual(encoding["labels"].dtype, torch.long)
+            self.assertTrue(encoding["labels"].min().item() >= 0)
+            self.assertTrue(encoding["labels"].max().item() <= 255)
+
+            # Test batched input (PIL images)
+            images, segmentation_maps = prepare_semantic_batch_inputs()
+
+            encoding = image_processing(images, segmentation_maps, return_tensors="pt")
+            self.assertEqual(
+                encoding["pixel_values"].shape,
+                (
+                    2,
+                    self.image_processor_tester.num_channels,
+                    self.image_processor_tester.crop_size["height"],
+                    self.image_processor_tester.crop_size["width"],
+                ),
+            )
+            self.assertEqual(
+                encoding["labels"].shape,
+                (
+                    2,
+                    self.image_processor_tester.crop_size["height"],
+                    self.image_processor_tester.crop_size["width"],
+                ),
+            )
+            self.assertEqual(encoding["labels"].dtype, torch.long)
+            self.assertTrue(encoding["labels"].min().item() >= 0)
+            self.assertTrue(encoding["labels"].max().item() <= 255)
+
+    def test_reduce_labels(self):
+        for image_processing_class in self.image_processor_list:
+            # Initialize image_processing
+            image_processing = image_processing_class(**self.image_processor_dict)
+
+            # ADE20k has 150 classes, and the background is included, so labels should be between 0 and 150
+            image, map = prepare_semantic_single_inputs()
+            encoding = image_processing(image, map, return_tensors="pt")
+            self.assertTrue(encoding["labels"].min().item() >= 0)
+            self.assertTrue(encoding["labels"].max().item() <= 150)
+
+            image_processing.do_reduce_labels = True
+            encoding = image_processing(image, map, return_tensors="pt")
+            self.assertTrue(encoding["labels"].min().item() >= 0)
+            self.assertTrue(encoding["labels"].max().item() <= 255)
+
+    def test_slow_fast_equivalence(self):
+        if not self.test_slow_image_processor or not self.test_fast_image_processor:
+            self.skipTest(reason="Skipping slow/fast equivalence test")
+
+        if self.image_processing_class is None or self.fast_image_processing_class is None:
+            self.skipTest(reason="Skipping slow/fast equivalence test as one of the image processors is not defined")
+
+        dummy_image, dummy_map = prepare_semantic_single_inputs()
+
+        image_processor_slow = self.image_processing_class(**self.image_processor_dict)
+        image_processor_fast = self.fast_image_processing_class(**self.image_processor_dict)
+
+        image_encoding_slow = image_processor_slow(dummy_image, segmentation_maps=dummy_map, return_tensors="pt")
+        image_encoding_fast = image_processor_fast(dummy_image, segmentation_maps=dummy_map, return_tensors="pt")
+
+        self._assert_slow_fast_tensors_equivalence(image_encoding_slow.pixel_values, image_encoding_fast.pixel_values)
+        self._assert_slow_fast_tensors_equivalence(
+            image_encoding_slow.labels.float(), image_encoding_fast.labels.float()
+        )
+
+    def test_slow_fast_equivalence_batched(self):
+        if not self.test_slow_image_processor or not self.test_fast_image_processor:
+            self.skipTest(reason="Skipping slow/fast equivalence test")
+
+        if self.image_processing_class is None or self.fast_image_processing_class is None:
+            self.skipTest(reason="Skipping slow/fast equivalence test as one of the image processors is not defined")
+
+        if hasattr(self.image_processor_tester, "do_center_crop") and self.image_processor_tester.do_center_crop:
+            self.skipTest(
+                reason="Skipping as do_center_crop is True and center_crop functions are not equivalent for fast and slow processors"
+            )
+
+        dummy_images, dummy_maps = prepare_semantic_batch_inputs()
+
+        image_processor_slow = self.image_processing_class(**self.image_processor_dict)
+        image_processor_fast = self.fast_image_processing_class(**self.image_processor_dict)
+
+        encoding_slow = image_processor_slow(dummy_images, segmentation_maps=dummy_maps, return_tensors="pt")
+        encoding_fast = image_processor_fast(dummy_images, segmentation_maps=dummy_maps, return_tensors="pt")
+
+        self._assert_slow_fast_tensors_equivalence(encoding_slow.pixel_values, encoding_fast.pixel_values)
+        self._assert_slow_fast_tensors_equivalence(encoding_slow.labels.float(), encoding_fast.labels.float())

transformers/tests/models/beit/test_modeling_beit.py

@@ -0,0 +1,582 @@
+# 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 BEiT model."""
+
+import unittest
+from functools import cached_property
+
+import pytest
+from datasets import load_dataset
+
+from transformers import BeitConfig
+from transformers.testing_utils import (
+    require_torch,
+    require_torch_multi_gpu,
+    require_vision,
+    slow,
+    torch_device,
+)
+from transformers.utils import (
+    is_torch_available,
+    is_vision_available,
+)
+
+from ...test_backbone_common import BackboneTesterMixin
+from ...test_configuration_common import ConfigTester
+from ...test_modeling_common import ModelTesterMixin, _config_zero_init, floats_tensor, ids_tensor
+from ...test_pipeline_mixin import PipelineTesterMixin
+
+
+if is_torch_available():
+    import torch
+    from torch import nn
+
+    from transformers import (
+        BeitBackbone,
+        BeitForImageClassification,
+        BeitForMaskedImageModeling,
+        BeitForSemanticSegmentation,
+        BeitModel,
+    )
+    from transformers.models.auto.modeling_auto import MODEL_FOR_BACKBONE_MAPPING_NAMES, MODEL_MAPPING_NAMES
+
+
+if is_vision_available():
+    from PIL import Image
+
+    from transformers import BeitImageProcessor
+
+
+class BeitModelTester:
+    def __init__(
+        self,
+        parent,
+        vocab_size=100,
+        batch_size=13,
+        image_size=30,
+        patch_size=2,
+        num_channels=3,
+        is_training=True,
+        use_labels=True,
+        hidden_size=32,
+        num_hidden_layers=4,
+        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,
+        out_indices=[1, 2, 3, 4],
+        out_features=["stage1", "stage2", "stage3", "stage4"],
+        attn_implementation="eager",
+        mask_ratio=0.5,
+    ):
+        self.parent = parent
+        self.vocab_size = vocab_size
+        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.out_indices = out_indices
+        self.out_features = out_features
+        self.num_labels = num_labels
+
+        # in BeiT, the seq length equals the number of patches + 1 (we add 1 for the [CLS] token)
+        num_patches = (image_size // patch_size) ** 2
+        self.seq_length = num_patches + 1
+        self.mask_length = self.seq_length - 1
+        self.num_masks = int(mask_ratio * self.seq_length)
+        self.attn_implementation = attn_implementation
+
+    def prepare_config_and_inputs(self):
+        pixel_values = floats_tensor([self.batch_size, self.num_channels, self.image_size, self.image_size])
+
+        labels = None
+        pixel_labels = None
+        if self.use_labels:
+            labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
+            pixel_labels = ids_tensor([self.batch_size, self.image_size, self.image_size], self.num_labels)
+
+        config = self.get_config()
+
+        return config, pixel_values, labels, pixel_labels
+
+    def get_config(self):
+        return BeitConfig(
+            vocab_size=self.vocab_size,
+            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,
+            out_indices=self.out_indices,
+            out_features=self.out_features,
+            attn_implementation=self.attn_implementation,
+        )
+
+    def create_and_check_model(self, config, pixel_values, labels, pixel_labels):
+        model = BeitModel(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(pixel_values)
+        self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
+
+    def create_and_check_backbone(self, config, pixel_values, labels, pixel_labels):
+        model = BeitBackbone(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(pixel_values)
+
+        # verify hidden states
+        self.parent.assertEqual(len(result.feature_maps), len(config.out_features))
+        expected_height = expected_width = self.image_size // config.patch_size
+        self.parent.assertListEqual(
+            list(result.feature_maps[0].shape), [self.batch_size, self.hidden_size, expected_height, expected_width]
+        )
+
+        # verify channels
+        self.parent.assertEqual(len(model.channels), len(config.out_features))
+
+        # verify backbone works with out_features=None
+        config.out_features = None
+        model = BeitBackbone(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(pixel_values)
+
+        # verify feature maps
+        self.parent.assertEqual(len(result.feature_maps), 1)
+        self.parent.assertListEqual(
+            list(result.feature_maps[0].shape), [self.batch_size, self.hidden_size, expected_height, expected_width]
+        )
+
+        # verify channels
+        self.parent.assertEqual(len(model.channels), 1)
+
+    def create_and_check_for_masked_lm(self, config, pixel_values, labels, pixel_labels):
+        model = BeitForMaskedImageModeling(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(pixel_values)
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length - 1, self.vocab_size))
+
+    def create_and_check_for_image_classification(self, config, pixel_values, labels, pixel_labels):
+        config.num_labels = self.type_sequence_label_size
+        model = BeitForImageClassification(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))
+
+        # test greyscale images
+        config.num_channels = 1
+        model = BeitForImageClassification(config)
+        model.to(torch_device)
+        model.eval()
+
+        pixel_values = floats_tensor([self.batch_size, 1, self.image_size, self.image_size])
+        result = model(pixel_values, labels=labels)
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.type_sequence_label_size))
+
+    def create_and_check_for_semantic_segmentation(self, config, pixel_values, labels, pixel_labels):
+        config.num_labels = self.num_labels
+        model = BeitForSemanticSegmentation(config)
+        model.to(torch_device)
+        model.eval()
+        result = model(pixel_values)
+        self.parent.assertEqual(
+            result.logits.shape, (self.batch_size, self.num_labels, self.image_size * 2, self.image_size * 2)
+        )
+        result = model(pixel_values, labels=pixel_labels)
+        self.parent.assertEqual(
+            result.logits.shape, (self.batch_size, self.num_labels, self.image_size * 2, self.image_size * 2)
+        )
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        config, pixel_values, labels, pixel_labels = config_and_inputs
+        inputs_dict = {"pixel_values": pixel_values}
+        return config, inputs_dict
+
+
+@require_torch
+class BeitModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
+    """
+    Here we also overwrite some of the tests of test_modeling_common.py, as BEiT does not use input_ids, inputs_embeds,
+    attention_mask and seq_length.
+    """
+
+    all_model_classes = (
+        (
+            BeitModel,
+            BeitForImageClassification,
+            BeitForMaskedImageModeling,
+            BeitForSemanticSegmentation,
+            BeitBackbone,
+        )
+        if is_torch_available()
+        else ()
+    )
+    pipeline_model_mapping = (
+        {
+            "image-feature-extraction": BeitModel,
+            "image-classification": BeitForImageClassification,
+            "image-segmentation": BeitForSemanticSegmentation,
+        }
+        if is_torch_available()
+        else {}
+    )
+
+    test_pruning = False
+    test_resize_embeddings = False
+    test_head_masking = False
+    test_torch_exportable = True
+
+    def setUp(self):
+        self.model_tester = BeitModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=BeitConfig, has_text_modality=False, hidden_size=37)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    @unittest.skip(reason="BEiT does not use inputs_embeds")
+    def test_inputs_embeds(self):
+        pass
+
+    @require_torch_multi_gpu
+    @unittest.skip(reason="BEiT has some layers using `add_module` which doesn't work well with `nn.DataParallel`")
+    def test_multi_gpu_data_parallel_forward(self):
+        pass
+
+    @unittest.skip(reason="BEiT does not support feedforward chunking yet")
+    def test_feed_forward_chunking(self):
+        pass
+
+    @unittest.skip(reason="BEiT can't compile dynamic")
+    @pytest.mark.torch_compile_test
+    def test_sdpa_can_compile_dynamic(self):
+        pass
+
+    def test_model_get_set_embeddings(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_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_model(*config_and_inputs)
+
+    def test_backbone(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_backbone(*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_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)
+
+    def test_for_semantic_segmentation(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_semantic_segmentation(*config_and_inputs)
+
+    def test_training(self):
+        if not self.model_tester.is_training:
+            self.skipTest(reason="model_tester.is_training is set to False")
+
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+        config.return_dict = True
+
+        for model_class in self.all_model_classes:
+            # we don't test BeitForMaskedImageModeling
+            if model_class.__name__ in [
+                *MODEL_MAPPING_NAMES.values(),
+                *MODEL_FOR_BACKBONE_MAPPING_NAMES.values(),
+                "BeitForMaskedImageModeling",
+            ]:
+                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:
+            self.skipTest(reason="model_tester.is_training is set to False")
+
+        config.use_cache = False
+        config.return_dict = True
+
+        for model_class in self.all_model_classes:
+            # we don't test BeitForMaskedImageModeling
+            if (
+                model_class.__name__
+                in [
+                    *MODEL_MAPPING_NAMES.values(),
+                    *MODEL_FOR_BACKBONE_MAPPING_NAMES.values(),
+                    "BeitForMaskedImageModeling",
+                ]
+                or not model_class.supports_gradient_checkpointing
+            ):
+                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()
+
+    @unittest.skip(
+        reason="This architecture seem to not compute gradients properly when using GC, check: https://github.com/huggingface/transformers/pull/27124"
+    )
+    def test_training_gradient_checkpointing_use_reentrant(self):
+        pass
+
+    @unittest.skip(
+        reason="This architecture seem to not compute gradients properly when using GC, check: https://github.com/huggingface/transformers/pull/27124"
+    )
+    def test_training_gradient_checkpointing_use_reentrant_false(self):
+        pass
+
+    def test_initialization(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        configs_no_init = _config_zero_init(config)
+        for model_class in self.all_model_classes:
+            model = model_class(config=configs_no_init)
+            for name, param in model.named_parameters():
+                # we skip lambda parameters as these require special initial values
+                # determined by config.layer_scale_init_value
+                if "lambda" in name:
+                    continue
+                if param.requires_grad:
+                    self.assertIn(
+                        ((param.data.mean() * 1e9).round() / 1e9).item(),
+                        [0.0, 1.0],
+                        msg=f"Parameter {name} of model {model_class} seems not properly initialized",
+                    )
+
+    @slow
+    def test_model_from_pretrained(self):
+        model_name = "microsoft/beit-base-patch16-224"
+        model = BeitModel.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 BeitModelIntegrationTest(unittest.TestCase):
+    @cached_property
+    def default_image_processor(self):
+        return BeitImageProcessor.from_pretrained("microsoft/beit-base-patch16-224") if is_vision_available() else None
+
+    @slow
+    def test_inference_masked_image_modeling_head(self):
+        model = BeitForMaskedImageModeling.from_pretrained("microsoft/beit-base-patch16-224-pt22k").to(torch_device)
+
+        image_processor = self.default_image_processor
+        image = prepare_img()
+        pixel_values = image_processor(images=image, return_tensors="pt").pixel_values.to(torch_device)
+
+        # prepare bool_masked_pos
+        bool_masked_pos = torch.ones((1, 196), dtype=torch.bool).to(torch_device)
+
+        # forward pass
+        with torch.no_grad():
+            outputs = model(pixel_values=pixel_values, bool_masked_pos=bool_masked_pos)
+        logits = outputs.logits
+
+        # verify the logits
+        expected_shape = torch.Size((1, 196, 8192))
+        self.assertEqual(logits.shape, expected_shape)
+
+        expected_slice = torch.tensor(
+            [[-3.2437, 0.5072, -13.9174], [-3.2456, 0.4948, -13.9401], [-3.2033, 0.5121, -13.8550]]
+        ).to(torch_device)
+
+        torch.testing.assert_close(logits[bool_masked_pos][:3, :3], expected_slice, rtol=1e-2, atol=1e-2)
+
+    @slow
+    def test_inference_image_classification_head_imagenet_1k(self):
+        model = BeitForImageClassification.from_pretrained("microsoft/beit-base-patch16-224").to(torch_device)
+
+        image_processor = self.default_image_processor
+        image = prepare_img()
+        inputs = image_processor(images=image, return_tensors="pt").to(torch_device)
+
+        # forward pass
+        with torch.no_grad():
+            outputs = model(**inputs)
+        logits = outputs.logits
+
+        # verify the logits
+        expected_shape = torch.Size((1, 1000))
+        self.assertEqual(logits.shape, expected_shape)
+
+        expected_slice = torch.tensor([-1.2385, -1.0987, -1.0108]).to(torch_device)
+
+        torch.testing.assert_close(logits[0, :3], expected_slice, rtol=1e-4, atol=1e-4)
+
+        expected_class_idx = 281
+        self.assertEqual(logits.argmax(-1).item(), expected_class_idx)
+
+    @slow
+    def test_inference_image_classification_head_imagenet_22k(self):
+        model = BeitForImageClassification.from_pretrained("microsoft/beit-large-patch16-224-pt22k-ft22k").to(
+            torch_device
+        )
+
+        image_processor = self.default_image_processor
+        image = prepare_img()
+        inputs = image_processor(images=image, return_tensors="pt").to(torch_device)
+
+        # forward pass
+        with torch.no_grad():
+            outputs = model(**inputs)
+        logits = outputs.logits
+
+        # verify the logits
+        expected_shape = torch.Size((1, 21841))
+        self.assertEqual(logits.shape, expected_shape)
+
+        expected_slice = torch.tensor([1.6881, -0.2787, 0.5901]).to(torch_device)
+
+        torch.testing.assert_close(logits[0, :3], expected_slice, rtol=1e-4, atol=1e-4)
+
+        expected_class_idx = 2396
+        self.assertEqual(logits.argmax(-1).item(), expected_class_idx)
+
+    @slow
+    def test_inference_semantic_segmentation(self):
+        model = BeitForSemanticSegmentation.from_pretrained("microsoft/beit-base-finetuned-ade-640-640")
+        model = model.to(torch_device)
+
+        image_processor = BeitImageProcessor(do_resize=True, size=640, do_center_crop=False)
+
+        ds = load_dataset("hf-internal-testing/fixtures_ade20k", split="test")
+        image = ds[0]["image"].convert("RGB")
+        inputs = image_processor(images=image, return_tensors="pt").to(torch_device)
+
+        # forward pass
+        with torch.no_grad():
+            outputs = model(**inputs)
+        logits = outputs.logits
+
+        # verify the logits
+        expected_shape = torch.Size((1, 150, 160, 160))
+        self.assertEqual(logits.shape, expected_shape)
+
+        expected_slice = torch.tensor(
+            [
+                [[-4.8960, -2.3688, -3.0355], [-2.8479, -0.9836, -1.7418], [-2.9449, -1.3333, -2.1456]],
+                [[-5.8081, -3.4124, -4.1006], [-3.8561, -2.2081, -3.0323], [-3.8365, -2.4601, -3.3669]],
+                [[-0.0309, 3.9868, 4.0540], [2.9640, 4.6877, 4.9976], [3.2081, 4.7690, 4.9942]],
+            ],
+            device=torch_device,
+        )
+        torch.testing.assert_close(logits[0, :3, :3, :3], expected_slice, rtol=1e-4, atol=1e-4)
+
+    @slow
+    def test_post_processing_semantic_segmentation(self):
+        model = BeitForSemanticSegmentation.from_pretrained("microsoft/beit-base-finetuned-ade-640-640")
+        model = model.to(torch_device)
+
+        image_processor = BeitImageProcessor(do_resize=True, size=640, do_center_crop=False)
+
+        ds = load_dataset("hf-internal-testing/fixtures_ade20k", split="test")
+        image = ds[0]["image"].convert("RGB")
+        inputs = image_processor(images=image, return_tensors="pt").to(torch_device)
+
+        # forward pass
+        with torch.no_grad():
+            outputs = model(**inputs)
+
+        outputs.logits = outputs.logits.detach().cpu()
+
+        segmentation = image_processor.post_process_semantic_segmentation(outputs=outputs, target_sizes=[(500, 300)])
+        expected_shape = torch.Size((500, 300))
+        self.assertEqual(segmentation[0].shape, expected_shape)
+
+        segmentation = image_processor.post_process_semantic_segmentation(outputs=outputs)
+        expected_shape = torch.Size((160, 160))
+        self.assertEqual(segmentation[0].shape, expected_shape)
+
+    @slow
+    def test_inference_interpolate_pos_encoding(self):
+        model_name = "microsoft/beit-base-patch16-224-pt22k"
+        model = BeitModel.from_pretrained(model_name, **{"use_absolute_position_embeddings": True}).to(torch_device)
+
+        image = Image.open("./tests/fixtures/tests_samples/COCO/000000039769.png")
+        processor = BeitImageProcessor.from_pretrained(model_name)
+        inputs = processor(images=image, return_tensors="pt", size={"height": 480, "width": 480})
+        pixel_values = inputs.pixel_values.to(torch_device)
+
+        # with interpolate_pos_encoding being True the model should process the higher resolution image
+        # successfully and produce the expected output.
+        with torch.no_grad():
+            outputs = model(pixel_values, interpolate_pos_encoding=True)
+
+        # num_cls_tokens + (height / patch_size) * (width / patch_size)
+        # 1 + (480 / 16) * (480 / 16) = 1 + 30 * 30 = 901
+        expected_shape = torch.Size((1, 901, 768))
+        self.assertEqual(outputs.last_hidden_state.shape, expected_shape)
+
+
+@require_torch
+class BeitBackboneTest(unittest.TestCase, BackboneTesterMixin):
+    all_model_classes = (BeitBackbone,) if is_torch_available() else ()
+    config_class = BeitConfig
+
+    def setUp(self):
+        self.model_tester = BeitModelTester(self)