Add VideoMAE (#17821)

* First draft

* Add VideoMAEForVideoClassification

* Improve conversion script

* Add VideoMAEForPreTraining

* Add VideoMAEFeatureExtractor

* Improve VideoMAEFeatureExtractor

* Improve docs

* Add first draft of model tests

* Improve VideoMAEForPreTraining

* Fix base_model_prefix

* Make model take pixel_values of shape (B, T, C, H, W)

* Add loss computation of VideoMAEForPreTraining

* Improve tests

* Improve model testsé

* Make all tests pass

* Add VideoMAE to main README

* Add tests for VideoMAEFeatureExtractor

* Add integration test

* Improve conversion script

* Rename patch embedding class

* Remove VideoMAELayer from init

* Update design of patch embeddings

* Improve comments

* Improve conversion script

* Improve conversion script

* Add conversion of pretrained model

* Add loss verification of pretrained model

* Add loss verification of unnormalized targets

* Add integration test for pretraining model

* Apply suggestions from code review

* Fix bug to make feature extractor resize only shorter edge

* Address more comments

* Improve normalization of videos

* Add doc examples

* Move constants to dedicated script

* Remove scripts

* Transfer checkpoints, fix docs

* Update script

* Update image mean and std

* Fix doc tests

* Set return_tensors to NumPy by default

* Revert the previous change
Co-authored-by: Niels Rogge <nielsrogge@Nielss-MacBook-Pro.local>

src/transformers/utils/__init__.py

@@ -22,6 +22,7 @@
 from packaging import version
 
 from .. import __version__
+from .constants import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD, IMAGENET_STANDARD_MEAN, IMAGENET_STANDARD_STD
 from .doc import (
     add_code_sample_docstrings,
     add_end_docstrings,

src/transformers/utils/constants.py

+IMAGENET_DEFAULT_MEAN = [0.485, 0.456, 0.406]
+IMAGENET_DEFAULT_STD = [0.229, 0.224, 0.225]
+IMAGENET_STANDARD_MEAN = [0.5, 0.5, 0.5]
+IMAGENET_STANDARD_STD = [0.5, 0.5, 0.5]

src/transformers/utils/dummy_pt_objects.py

@@ -406,6 +406,9 @@ MODEL_FOR_TABLE_QUESTION_ANSWERING_MAPPING = None
 MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING = None
 
 
+MODEL_FOR_VIDEO_CLASSIFICATION_MAPPING = None
+
+
 MODEL_FOR_VISION_2_SEQ_MAPPING = None
 
 
@@ -572,6 +575,13 @@ class AutoModelForTokenClassification(metaclass=DummyObject):
         requires_backends(self, ["torch"])
 
 
+class AutoModelForVideoClassification(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
 class AutoModelForVision2Seq(metaclass=DummyObject):
     _backends = ["torch"]
 
@@ -4753,6 +4763,37 @@ class VanPreTrainedModel(metaclass=DummyObject):
         requires_backends(self, ["torch"])
 
 
+VIDEOMAE_PRETRAINED_MODEL_ARCHIVE_LIST = None
+
+
+class VideoMAEForPreTraining(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class VideoMAEForVideoClassification(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class VideoMAEModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class VideoMAEPreTrainedModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
 VILT_PRETRAINED_MODEL_ARCHIVE_LIST = None
 
 

src/transformers/utils/dummy_vision_objects.py

@@ -150,6 +150,13 @@ class SegformerFeatureExtractor(metaclass=DummyObject):
         requires_backends(self, ["vision"])
 
 
+class VideoMAEFeatureExtractor(metaclass=DummyObject):
+    _backends = ["vision"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["vision"])
+
+
 class ViltFeatureExtractor(metaclass=DummyObject):
     _backends = ["vision"]
 

tests/models/videomae/test_feature_extraction_videomae.py

+# coding=utf-8
+# Copyright 2022 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
+
+import numpy as np
+
+from transformers.testing_utils import require_torch, require_vision
+from transformers.utils import is_torch_available, is_vision_available
+
+from ...test_feature_extraction_common import FeatureExtractionSavingTestMixin, prepare_video_inputs
+
+
+if is_torch_available():
+    import torch
+
+if is_vision_available():
+    from PIL import Image
+
+    from transformers import VideoMAEFeatureExtractor
+
+
+class VideoMAEFeatureExtractionTester(unittest.TestCase):
+    def __init__(
+        self,
+        parent,
+        batch_size=7,
+        num_channels=3,
+        num_frames=10,
+        image_size=18,
+        min_resolution=30,
+        max_resolution=400,
+        do_resize=True,
+        size=18,
+        do_normalize=True,
+        image_mean=[0.5, 0.5, 0.5],
+        image_std=[0.5, 0.5, 0.5],
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.num_channels = num_channels
+        self.num_frames = num_frames
+        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_normalize = do_normalize
+        self.image_mean = image_mean
+        self.image_std = image_std
+
+    def prepare_feat_extract_dict(self):
+        return {
+            "image_mean": self.image_mean,
+            "image_std": self.image_std,
+            "do_normalize": self.do_normalize,
+            "do_resize": self.do_resize,
+            "size": self.size,
+        }
+
+
+@require_torch
+@require_vision
+class VideoMAEFeatureExtractionTest(FeatureExtractionSavingTestMixin, unittest.TestCase):
+
+    feature_extraction_class = VideoMAEFeatureExtractor if is_vision_available() else None
+
+    def setUp(self):
+        self.feature_extract_tester = VideoMAEFeatureExtractionTester(self)
+
+    @property
+    def feat_extract_dict(self):
+        return self.feature_extract_tester.prepare_feat_extract_dict()
+
+    def test_feat_extract_properties(self):
+        feature_extractor = self.feature_extraction_class(**self.feat_extract_dict)
+        self.assertTrue(hasattr(feature_extractor, "image_mean"))
+        self.assertTrue(hasattr(feature_extractor, "image_std"))
+        self.assertTrue(hasattr(feature_extractor, "do_normalize"))
+        self.assertTrue(hasattr(feature_extractor, "do_resize"))
+        self.assertTrue(hasattr(feature_extractor, "size"))
+
+    def test_batch_feature(self):
+        pass
+
+    def test_call_pil(self):
+        # Initialize feature_extractor
+        feature_extractor = self.feature_extraction_class(**self.feat_extract_dict)
+        # create random PIL videos
+        video_inputs = prepare_video_inputs(self.feature_extract_tester, equal_resolution=False)
+        for video in video_inputs:
+            self.assertIsInstance(video, list)
+            self.assertIsInstance(video[0], Image.Image)
+
+        # Test not batched input
+        encoded_videos = feature_extractor(video_inputs[0], return_tensors="pt").pixel_values
+        self.assertEqual(
+            encoded_videos.shape,
+            (
+                1,
+                self.feature_extract_tester.num_frames,
+                self.feature_extract_tester.num_channels,
+                self.feature_extract_tester.size,
+                self.feature_extract_tester.size,
+            ),
+        )
+
+        # Test batched
+        encoded_videos = feature_extractor(video_inputs, return_tensors="pt").pixel_values
+        self.assertEqual(
+            encoded_videos.shape,
+            (
+                self.feature_extract_tester.batch_size,
+                self.feature_extract_tester.num_frames,
+                self.feature_extract_tester.num_channels,
+                self.feature_extract_tester.size,
+                self.feature_extract_tester.size,
+            ),
+        )
+
+    def test_call_numpy(self):
+        # Initialize feature_extractor
+        feature_extractor = self.feature_extraction_class(**self.feat_extract_dict)
+        # create random numpy tensors
+        video_inputs = prepare_video_inputs(self.feature_extract_tester, equal_resolution=False, numpify=True)
+        for video in video_inputs:
+            self.assertIsInstance(video, list)
+            self.assertIsInstance(video[0], np.ndarray)
+
+        # Test not batched input
+        encoded_videos = feature_extractor(video_inputs[0], return_tensors="pt").pixel_values
+        self.assertEqual(
+            encoded_videos.shape,
+            (
+                1,
+                self.feature_extract_tester.num_frames,
+                self.feature_extract_tester.num_channels,
+                self.feature_extract_tester.size,
+                self.feature_extract_tester.size,
+            ),
+        )
+
+        # Test batched
+        encoded_videos = feature_extractor(video_inputs, return_tensors="pt").pixel_values
+        self.assertEqual(
+            encoded_videos.shape,
+            (
+                self.feature_extract_tester.batch_size,
+                self.feature_extract_tester.num_frames,
+                self.feature_extract_tester.num_channels,
+                self.feature_extract_tester.size,
+                self.feature_extract_tester.size,
+            ),
+        )
+
+    def test_call_pytorch(self):
+        # Initialize feature_extractor
+        feature_extractor = self.feature_extraction_class(**self.feat_extract_dict)
+        # create random PyTorch tensors
+        video_inputs = prepare_video_inputs(self.feature_extract_tester, equal_resolution=False, torchify=True)
+        for video in video_inputs:
+            self.assertIsInstance(video, list)
+            self.assertIsInstance(video[0], torch.Tensor)
+
+        # Test not batched input
+        encoded_videos = feature_extractor(video_inputs[0], return_tensors="pt").pixel_values
+        self.assertEqual(
+            encoded_videos.shape,
+            (
+                1,
+                self.feature_extract_tester.num_frames,
+                self.feature_extract_tester.num_channels,
+                self.feature_extract_tester.size,
+                self.feature_extract_tester.size,
+            ),
+        )
+
+        # Test batched
+        encoded_videos = feature_extractor(video_inputs, return_tensors="pt").pixel_values
+        self.assertEqual(
+            encoded_videos.shape,
+            (
+                self.feature_extract_tester.batch_size,
+                self.feature_extract_tester.num_frames,
+                self.feature_extract_tester.num_channels,
+                self.feature_extract_tester.size,
+                self.feature_extract_tester.size,
+            ),
+        )

tests/models/videomae/test_modeling_videomae.py

+# coding=utf-8
+# Copyright 2022 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 VideoMAE model. """
+
+
+import copy
+import inspect
+import unittest
+
+import numpy as np
+
+from huggingface_hub import hf_hub_download
+from transformers import VideoMAEConfig
+from transformers.models.auto import get_values
+from transformers.testing_utils import require_torch, require_vision, slow, torch_device
+from transformers.utils import cached_property, is_torch_available, is_vision_available
+
+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_VIDEO_CLASSIFICATION_MAPPING,
+        VideoMAEForPreTraining,
+        VideoMAEForVideoClassification,
+        VideoMAEModel,
+    )
+    from transformers.models.videomae.modeling_videomae import VIDEOMAE_PRETRAINED_MODEL_ARCHIVE_LIST
+
+
+if is_vision_available():
+    from transformers import VideoMAEFeatureExtractor
+
+
+class VideoMAEModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=13,
+        image_size=10,
+        num_channels=3,
+        patch_size=2,
+        tubelet_size=2,
+        num_frames=2,
+        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,
+        mask_ratio=0.9,
+        scope=None,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.image_size = image_size
+        self.num_channels = num_channels
+        self.patch_size = patch_size
+        self.tubelet_size = tubelet_size
+        self.num_frames = num_frames
+        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.mask_ratio = mask_ratio
+        self.scope = scope
+
+        # in VideoMAE, the number of tokens equals num_frames/tubelet_size * num_patches per frame
+        self.num_patches_per_frame = (image_size // patch_size) ** 2
+        self.seq_length = (num_frames // tubelet_size) * self.num_patches_per_frame
+
+        # use this variable to define bool_masked_pos
+        self.num_masks = int(mask_ratio * self.seq_length)
+
+    def prepare_config_and_inputs(self):
+        pixel_values = floats_tensor(
+            [self.batch_size, self.num_frames, 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 VideoMAEConfig(
+            image_size=self.image_size,
+            patch_size=self.patch_size,
+            num_channels=self.num_channels,
+            num_frames=self.num_frames,
+            tubelet_size=self.tubelet_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,
+            is_decoder=False,
+            initializer_range=self.initializer_range,
+        )
+
+    def create_and_check_model(self, config, pixel_values, labels):
+        model = VideoMAEModel(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_for_pretraining(self, config, pixel_values, labels):
+        model = VideoMAEForPreTraining(config)
+        model.to(torch_device)
+        model.eval()
+        # important: each video needs to have the same number of masked patches
+        # hence we define a single mask, which we then repeat for each example in the batch
+        mask = torch.ones((self.num_masks,))
+        mask = torch.cat([mask, torch.zeros(self.seq_length - mask.size(0))])
+        bool_masked_pos = mask.expand(self.batch_size, -1).bool()
+
+        result = model(pixel_values, bool_masked_pos)
+        # model only returns predictions for masked patches
+        num_masked_patches = mask.sum().item()
+        decoder_num_labels = 3 * self.tubelet_size * self.patch_size**2
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, num_masked_patches, decoder_num_labels))
+
+    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 VideoMAEModelTest(ModelTesterMixin, unittest.TestCase):
+    """
+    Here we also overwrite some of the tests of test_modeling_common.py, as VideoMAE does not use input_ids, inputs_embeds,
+    attention_mask and seq_length.
+    """
+
+    all_model_classes = (
+        (VideoMAEModel, VideoMAEForPreTraining, VideoMAEForVideoClassification) 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 = VideoMAEModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=VideoMAEConfig, has_text_modality=False, hidden_size=37)
+
+    def _prepare_for_class(self, inputs_dict, model_class, return_labels=False):
+        inputs_dict = copy.deepcopy(inputs_dict)
+
+        if model_class == VideoMAEForPreTraining:
+            # important: each video needs to have the same number of masked patches
+            # hence we define a single mask, which we then repeat for each example in the batch
+            mask = torch.ones((self.model_tester.num_masks,))
+            mask = torch.cat([mask, torch.zeros(self.model_tester.seq_length - mask.size(0))])
+            bool_masked_pos = mask.expand(self.model_tester.batch_size, -1).bool()
+            inputs_dict["bool_masked_pos"] = bool_masked_pos.to(torch_device)
+
+        if return_labels:
+            if model_class in [
+                *get_values(MODEL_FOR_VIDEO_CLASSIFICATION_MAPPING),
+            ]:
+                inputs_dict["labels"] = torch.zeros(
+                    self.model_tester.batch_size, dtype=torch.long, device=torch_device
+                )
+
+        return inputs_dict
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    @unittest.skip(reason="VideoMAE does not use inputs_embeds")
+    def test_inputs_embeds(self):
+        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_for_pretraining(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_pretraining(*config_and_inputs)
+
+    @slow
+    def test_model_from_pretrained(self):
+        for model_name in VIDEOMAE_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
+            model = VideoMAEModel.from_pretrained(model_name)
+            self.assertIsNotNone(model)
+
+    def test_attention_outputs(self):
+        if not self.has_attentions:
+            pass
+
+        else:
+            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+            config.return_dict = True
+
+            for model_class in self.all_model_classes:
+                num_visible_patches = self.model_tester.seq_length - self.model_tester.num_masks
+                seq_len = (
+                    num_visible_patches if model_class == VideoMAEForPreTraining else self.model_tester.seq_length
+                )
+
+                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.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.attentions
+                self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)
+
+                self.assertListEqual(
+                    list(attentions[0].shape[-3:]),
+                    [self.model_tester.num_attention_heads, seq_len, seq_len],
+                )
+                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))
+
+                self.assertEqual(out_len + 1, len(outputs))
+
+                self_attentions = outputs.attentions
+
+                self.assertEqual(len(self_attentions), self.model_tester.num_hidden_layers)
+                self.assertListEqual(
+                    list(self_attentions[0].shape[-3:]),
+                    [self.model_tester.num_attention_heads, seq_len, seq_len],
+                )
+
+    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.hidden_states
+            expected_num_layers = self.model_tester.num_hidden_layers + 1
+            self.assertEqual(len(hidden_states), expected_num_layers)
+
+            num_visible_patches = self.model_tester.seq_length - self.model_tester.num_masks
+            seq_length = num_visible_patches if model_class == VideoMAEForPreTraining else self.model_tester.seq_length
+
+            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)
+
+
+# We will verify our results on a video of eating spaghetti
+# Frame indices used: [164 168 172 176 181 185 189 193 198 202 206 210 215 219 223 227]
+def prepare_video():
+    file = hf_hub_download(repo_id="datasets/hf-internal-testing/spaghetti-video", filename="eating_spaghetti.npy")
+    video = np.load(file)
+    return list(video)
+
+
+@require_torch
+@require_vision
+class VideoMAEModelIntegrationTest(unittest.TestCase):
+    @cached_property
+    def default_feature_extractor(self):
+        # logits were tested with a different mean and std, so we use the same here
+        return (
+            VideoMAEFeatureExtractor(image_mean=[0.5, 0.5, 0.5], image_std=[0.5, 0.5, 0.5])
+            if is_vision_available()
+            else None
+        )
+
+    @slow
+    def test_inference_for_video_classification(self):
+        model = VideoMAEForVideoClassification.from_pretrained("MCG-NJU/videomae-base-finetuned-kinetics").to(
+            torch_device
+        )
+
+        feature_extractor = self.default_feature_extractor
+        video = prepare_video()
+        inputs = feature_extractor(video, return_tensors="pt").to(torch_device)
+
+        # forward pass
+        with torch.no_grad():
+            outputs = model(**inputs)
+
+        # verify the logits
+        expected_shape = torch.Size((1, 400))
+        self.assertEqual(outputs.logits.shape, expected_shape)
+
+        expected_slice = torch.tensor([0.3669, -0.0688, -0.2421]).to(torch_device)
+
+        self.assertTrue(torch.allclose(outputs.logits[0, :3], expected_slice, atol=1e-4))
+
+    @slow
+    def test_inference_for_pretraining(self):
+        model = VideoMAEForPreTraining.from_pretrained("MCG-NJU/videomae-base-short").to(torch_device)
+
+        feature_extractor = self.default_feature_extractor
+        video = prepare_video()
+        inputs = feature_extractor(video, return_tensors="pt").to(torch_device)
+
+        # add boolean mask, indicating which patches to mask
+        local_path = hf_hub_download(repo_id="hf-internal-testing/bool-masked-pos", filename="bool_masked_pos.pt")
+        inputs["bool_masked_pos"] = torch.load(local_path)
+
+        # forward pass
+        with torch.no_grad():
+            outputs = model(**inputs)
+
+        # verify the logits
+        expected_shape = torch.Size([1, 1408, 1536])
+        expected_slice = torch.tensor(
+            [[0.7994, 0.9612, 0.8508], [0.7401, 0.8958, 0.8302], [0.5862, 0.7468, 0.7325]], device=torch_device
+        )
+        self.assertEqual(outputs.logits.shape, expected_shape)
+        self.assertTrue(torch.allclose(outputs.logits[0, :3, :3], expected_slice, atol=1e-4))
+
+        # verify the loss (`config.norm_pix_loss` = `True`)
+        expected_loss = torch.tensor([0.5142], device=torch_device)
+        self.assertTrue(torch.allclose(outputs.loss, expected_loss, atol=1e-4))
+
+        # verify the loss (`config.norm_pix_loss` = `False`)
+        model = VideoMAEForPreTraining.from_pretrained("MCG-NJU/videomae-base-short", norm_pix_loss=False).to(
+            torch_device
+        )
+
+        with torch.no_grad():
+            outputs = model(**inputs)
+
+        expected_loss = torch.tensor(torch.tensor([0.6469]), device=torch_device)
+        self.assertTrue(torch.allclose(outputs.loss, expected_loss, atol=1e-4))

tests/test_feature_extraction_common.py

@@ -48,47 +48,89 @@ SAMPLE_FEATURE_EXTRACTION_CONFIG_DIR = get_tests_dir("fixtures")
 def prepare_image_inputs(feature_extract_tester, equal_resolution=False, numpify=False, torchify=False):
     """This function prepares a list of PIL images, or a list of numpy arrays if one specifies numpify=True,
     or a list of PyTorch tensors if one specifies torchify=True.
+
+    One can specify whether the images are of the same resolution or not.
     """
 
     assert not (numpify and torchify), "You cannot specify both numpy and PyTorch tensors at the same time"
 
-    if equal_resolution:
     image_inputs = []
     for i in range(feature_extract_tester.batch_size):
+        if equal_resolution:
+            width = height = feature_extract_tester.max_resolution
+        else:
+            # To avoid getting image width/height 0
+            min_resolution = feature_extract_tester.min_resolution
+            if getattr(feature_extract_tester, "size_divisor", None):
+                # If `size_divisor` is defined, the image needs to have width/size >= `size_divisor`
+                min_resolution = max(feature_extract_tester.size_divisor, min_resolution)
+            width, height = np.random.choice(np.arange(min_resolution, feature_extract_tester.max_resolution), 2)
         image_inputs.append(
             np.random.randint(
                 255,
                 size=(
                     feature_extract_tester.num_channels,
-                        feature_extract_tester.max_resolution,
-                        feature_extract_tester.max_resolution,
+                    width,
+                    height,
                 ),
                 dtype=np.uint8,
             )
         )
-    else:
-        image_inputs = []
 
-        # To avoid getting image width/height 0
-        min_resolution = feature_extract_tester.min_resolution
-        if getattr(feature_extract_tester, "size_divisor", None):
-            # If `size_divisor` is defined, the image needs to have width/size >= `size_divisor`
-            min_resolution = max(feature_extract_tester.size_divisor, min_resolution)
+    if not numpify and not torchify:
+        # PIL expects the channel dimension as last dimension
+        image_inputs = [Image.fromarray(np.moveaxis(image, 0, -1)) for image in image_inputs]
 
-        for i in range(feature_extract_tester.batch_size):
-            width, height = np.random.choice(np.arange(min_resolution, feature_extract_tester.max_resolution), 2)
-            image_inputs.append(
-                np.random.randint(255, size=(feature_extract_tester.num_channels, width, height), dtype=np.uint8)
-            )
+    if torchify:
+        image_inputs = [torch.from_numpy(image) for image in image_inputs]
+
+    return image_inputs
+
+
+def prepare_video(feature_extract_tester, width=10, height=10, numpify=False, torchify=False):
+    """This function prepares a video as a list of PIL images/NumPy arrays/PyTorch tensors."""
+
+    video = []
+    for i in range(feature_extract_tester.num_frames):
+        video.append(np.random.randint(255, size=(feature_extract_tester.num_channels, width, height), dtype=np.uint8))
 
     if not numpify and not torchify:
         # PIL expects the channel dimension as last dimension
-        image_inputs = [Image.fromarray(np.moveaxis(x, 0, -1)) for x in image_inputs]
+        video = [Image.fromarray(np.moveaxis(frame, 0, -1)) for frame in video]
 
     if torchify:
-        image_inputs = [torch.from_numpy(x) for x in image_inputs]
+        video = [torch.from_numpy(frame) for frame in video]
 
-    return image_inputs
+    return video
+
+
+def prepare_video_inputs(feature_extract_tester, equal_resolution=False, numpify=False, torchify=False):
+    """This function prepares a batch of videos: a list of list of PIL images, or a list of list of numpy arrays if
+    one specifies numpify=True, or a list of list of PyTorch tensors if one specifies torchify=True.
+
+    One can specify whether the videos are of the same resolution or not.
+    """
+
+    assert not (numpify and torchify), "You cannot specify both numpy and PyTorch tensors at the same time"
+
+    video_inputs = []
+    for i in range(feature_extract_tester.batch_size):
+        if equal_resolution:
+            width = height = feature_extract_tester.max_resolution
+        else:
+            width, height = np.random.choice(
+                np.arange(feature_extract_tester.min_resolution, feature_extract_tester.max_resolution), 2
+            )
+            video = prepare_video(
+                feature_extract_tester=feature_extract_tester,
+                width=width,
+                height=height,
+                numpify=numpify,
+                torchify=torchify,
+            )
+        video_inputs.append(video)
+
+    return video_inputs
 
 
 class FeatureExtractionSavingTestMixin:

tests/test_modeling_common.py

@@ -99,6 +99,7 @@ if is_torch_available():
         MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING,
         MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING,
         MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING,
+        MODEL_FOR_VIDEO_CLASSIFICATION_MAPPING,
         MODEL_MAPPING,
         AdaptiveEmbedding,
         AutoModelForCausalLM,
@@ -182,6 +183,7 @@ class ModelTesterMixin:
                 *get_values(MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING),
                 *get_values(MODEL_FOR_NEXT_SENTENCE_PREDICTION_MAPPING),
                 *get_values(MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING),
+                *get_values(MODEL_FOR_VIDEO_CLASSIFICATION_MAPPING),
             ]:
                 inputs_dict["labels"] = torch.zeros(
                     self.model_tester.batch_size, dtype=torch.long, device=torch_device