Add TVLT (#20725)

* Update image_processing_tvlt.py

* Update modeling_tvlt.py

* Update

* Update modeling_tvlt.py

* Create tvlt.mdx

* Update configuration_tvlt.py

* Update modeling_tvlt.py

* Update test_modeling_tvlt.py

* Update modeling_tvlt.py

* Update modeling_tvlt.py

* Update image_processing_tvlt.py

* Update feature_extraction_tvlt.py

* Update tvlt models

* Update tests

* Update

* Update

* Update tests

* Update README_ko.md

* Update README_ja.md

* Update README_ko.md

* Update README_zh-hans.md

* Update docs/source/en/model_doc/tvlt.mdx
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update docs/source/en/model_doc/tvlt.mdx
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/tvlt/configuration_tvlt.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/tvlt/configuration_tvlt.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/tvlt/configuration_tvlt.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/tvlt/configuration_tvlt.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update tvlt.mdx

* Update modeling_tvlt.py

* Update configuration_tvlt.py

* Update modeling_tvlt.py

* Update modeling_tvlt.py

* Update modeling_tvlt.py

* Update modeling_tvlt.py

* Add files via upload

* Update model

* Update modeling_tvlt.py

* Update tvlt models

* Update src/transformers/models/tvlt/__init__.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/models/tvlt/__init__.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/models/tvlt/image_processing_tvlt.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/models/tvlt/image_processing_tvlt.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Add files via upload

* Add files via upload

* Delete modeling_tvlt.py

* Delete feature_extraction_tvlt.py

* Delete configuration_tvlt.py

* Delete image_processing_tvlt.py

* Delete processing_tvlt.py

* Update tvlt

* Update src/transformers/models/tvlt/configuration_tvlt.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/tvlt/image_processing_tvlt.py
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* Update src/transformers/models/tvlt/configuration_tvlt.py
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* Update tests/models/tvlt/test_modeling_tvlt.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update tests/models/tvlt/test_modeling_tvlt.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/tvlt/modeling_tvlt.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/tvlt/modeling_tvlt.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update README.md
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update README_es.md

* Update README_hd.md

* Update README_ja.md

* Update README_ko.md

* Update README_zh-hans.md

* Update README_zh-hant.md

* Update index.mdx

* Update tvlt.mdx

* Update tvlt.mdx

* Update configuration_tvlt.py

* Update src/transformers/models/tvlt/image_processing_tvlt.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/tvlt/image_processing_tvlt.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/tvlt/image_processing_tvlt.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/tvlt/image_processing_tvlt.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/tvlt/modeling_tvlt.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update modeling_tvlt.py

* Add files via upload

* Update tvlt.mdx

* Update modeling_auto.py

* Add files via upload

* Add files via upload

* Update dummy_pt_objects.py

* Update __init__.py

* Update feature_extraction_tvlt.py

* Update feature_extraction_tvlt.py

* Update image_processing_tvlt.py

* Update modeling_auto.py

* Update test_feature_extraction_tvlt.py

* Update test_processor_tvlt.py

* Update test_feature_extraction_tvlt.py

* Add files via upload

* Update test_image_processor_tvlt.py

* Update tests/models/tvlt/test_processor_tvlt.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/models/tvlt/modeling_tvlt.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/tvlt/processing_tvlt.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/models/tvlt/modeling_tvlt.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/models/tvlt/modeling_tvlt.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update tests/models/tvlt/test_image_processor_tvlt.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/models/tvlt/modeling_tvlt.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update tests/models/tvlt/test_image_processor_tvlt.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/models/tvlt/modeling_tvlt.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update tests/models/tvlt/test_image_processor_tvlt.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update tests/models/tvlt/test_image_processor_tvlt.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update tests/models/tvlt/test_modeling_tvlt.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/models/tvlt/modeling_tvlt.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/tvlt/modeling_tvlt.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/tvlt/configuration_tvlt.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update tests/models/tvlt/test_feature_extraction_tvlt.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/models/tvlt/processing_tvlt.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update docs/source/en/model_doc/tvlt.mdx
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/models/tvlt/configuration_tvlt.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/models/tvlt/configuration_tvlt.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/models/tvlt/configuration_tvlt.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/models/tvlt/configuration_tvlt.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/models/tvlt/configuration_tvlt.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/models/tvlt/configuration_tvlt.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/models/tvlt/configuration_tvlt.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/models/tvlt/configuration_tvlt.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/models/tvlt/configuration_tvlt.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/models/tvlt/configuration_tvlt.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/models/tvlt/feature_extraction_tvlt.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/models/tvlt/configuration_tvlt.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/models/tvlt/feature_extraction_tvlt.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/models/tvlt/feature_extraction_tvlt.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/models/tvlt/feature_extraction_tvlt.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update feature_extraction_tvlt.py

* Update feature_extraction_tvlt.py

* Update src/transformers/models/tvlt/image_processing_tvlt.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/models/tvlt/image_processing_tvlt.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update image_processing_tvlt.py

* Update src/transformers/models/tvlt/modeling_tvlt.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update modeling_tvlt.py

* Update modeling_tvlt.py

* Update modeling_tvlt.py

* Update test_image_processor_tvlt.py

* Update tests/models/tvlt/test_modeling_tvlt.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update tests/models/tvlt/test_modeling_tvlt.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update tests/models/tvlt/test_modeling_tvlt.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update tests/models/tvlt/test_modeling_tvlt.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update tests/models/tvlt/test_modeling_tvlt.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update tests/models/tvlt/test_modeling_tvlt.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update tests/models/tvlt/test_modeling_tvlt.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update tests/models/tvlt/test_modeling_tvlt.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update tests/models/tvlt/test_modeling_tvlt.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Add files via upload

* Add files via upload

* Update modeling_tvlt.py

* Update modeling_tvlt.py

* Update modeling_tvlt.py

* Add files via upload

* Update docs/source/en/model_doc/tvlt.mdx
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update image_processing_tvlt.py

* Add files via upload

* Add files via upload

* Update tvlt.mdx

* Update docs/source/en/model_doc/tvlt.mdx
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update docs/source/en/model_doc/tvlt.mdx
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/tvlt/image_processing_tvlt.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/tvlt/image_processing_tvlt.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update docs/source/en/model_doc/tvlt.mdx
Co-authored-by: Sanchit Gandhi <93869735+sanchit-gandhi@users.noreply.github.com>

* Update docs/source/en/model_doc/tvlt.mdx
Co-authored-by: Sanchit Gandhi <93869735+sanchit-gandhi@users.noreply.github.com>

* Update src/transformers/models/tvlt/configuration_tvlt.py
Co-authored-by: Sanchit Gandhi <93869735+sanchit-gandhi@users.noreply.github.com>

* Add files via upload

* Add files via upload

* Add files via upload

* Add files via upload

* Update modeling_auto.py

* Update tvlt.mdx

* Update dummy_pt_objects.py

* Update feature_extraction_tvlt.py

* Update modeling_tvlt.py

* Update test_feature_extraction_tvlt.py

* Update test_image_processor_tvlt.py

* Update test_feature_extraction_tvlt.py

* Update modeling_tvlt.py

* Update dummy_pt_objects.py

* Update dummy_speech_objects.py

* Add files via upload

* Update README_hd.md

* Update modeling_tvlt.py

* Update modeling_tvlt.py

* Update modeling_tvlt.py

* Update modeling_tvlt.py

* Update modeling_tvlt.py

* Update modeling_tvlt.py

* Update test_modeling_tvlt.py

* Update src/transformers/models/tvlt/configuration_tvlt.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/models/tvlt/feature_extraction_tvlt.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/models/tvlt/image_processing_tvlt.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/models/tvlt/image_processing_tvlt.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/models/tvlt/modeling_tvlt.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/models/tvlt/image_processing_tvlt.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/models/tvlt/modeling_tvlt.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update MAE processing

* Update modeling_tvlt.py

* Update modeling_tvlt.py

* Update modeling

* Update style

* Update src/transformers/models/tvlt/modeling_tvlt.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/models/tvlt/modeling_tvlt.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update check_repo.py

* Update tvlt.mdx

* Update __init__.py

* Update tests

* Update tvlt models

* Update configuration_tvlt.py

* Update configuration_tvlt.py

* Update image_processing_tvlt.py

* Update dummy_pt_objects.py

* Add files via upload

* Update test_modeling_tvlt.py

* Update test_feature_extraction_tvlt.py

* Update test_feature_extraction_tvlt.py

* Update test_feature_extraction_tvlt.py

* Update test_feature_extraction_tvlt.py

* Update test_feature_extraction_tvlt.py

* Update test_feature_extraction_tvlt.py

---------
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>
Co-authored-by: Sanchit Gandhi <93869735+sanchit-gandhi@users.noreply.github.com>

src/transformers/utils/dummy_pt_objects.py

@@ -6068,6 +6068,37 @@ class TrOCRPreTrainedModel(metaclass=DummyObject):
         requires_backends(self, ["torch"])
 
 
+TVLT_PRETRAINED_MODEL_ARCHIVE_LIST = None
+
+
+class TvltForAudioVisualClassification(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class TvltForPreTraining(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class TvltModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class TvltPreTrainedModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
 UNISPEECH_PRETRAINED_MODEL_ARCHIVE_LIST = None
 
 

src/transformers/utils/dummy_speech_objects.py

@@ -28,3 +28,10 @@ class SpeechT5FeatureExtractor(metaclass=DummyObject):
 
     def __init__(self, *args, **kwargs):
         requires_backends(self, ["speech"])
+
+
+class TvltFeatureExtractor(metaclass=DummyObject):
+    _backends = ["speech"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["speech"])

src/transformers/utils/dummy_vision_objects.py

@@ -415,6 +415,13 @@ class Swin2SRImageProcessor(metaclass=DummyObject):
         requires_backends(self, ["vision"])
 
 
+class TvltImageProcessor(metaclass=DummyObject):
+    _backends = ["vision"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["vision"])
+
+
 class VideoMAEFeatureExtractor(metaclass=DummyObject):
     _backends = ["vision"]
 

tests/models/tvlt/test_feature_extraction_tvlt.py

+# coding=utf-8
+# Copyright 2023 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.
+""" Testing suite for the TVLT feature extraction. """
+
+import itertools
+import os
+import random
+import tempfile
+import unittest
+
+import numpy as np
+
+from transformers import is_datasets_available, is_speech_available
+from transformers.testing_utils import check_json_file_has_correct_format, require_torch, require_torchaudio
+from transformers.utils.import_utils import is_torch_available
+
+from ...test_sequence_feature_extraction_common import SequenceFeatureExtractionTestMixin
+
+
+if is_torch_available():
+    import torch
+
+if is_datasets_available():
+    from datasets import load_dataset
+
+if is_speech_available():
+    from transformers import TvltFeatureExtractor
+
+global_rng = random.Random()
+
+
+def floats_list(shape, scale=1.0, rng=None, name=None):
+    """Creates a random float32 tensor"""
+    if rng is None:
+        rng = global_rng
+
+    values = []
+    for batch_idx in range(shape[0]):
+        values.append([])
+        for _ in range(shape[1]):
+            values[-1].append(rng.random() * scale)
+
+    return values
+
+
+class TvltFeatureExtractionTester(unittest.TestCase):
+    def __init__(
+        self,
+        parent,
+        batch_size=7,
+        min_seq_length=400,
+        max_seq_length=2000,
+        spectrogram_length=2048,
+        feature_size=128,
+        num_audio_channels=1,
+        hop_length=512,
+        chunk_length=30,
+        sampling_rate=44100,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.min_seq_length = min_seq_length
+        self.max_seq_length = max_seq_length
+        self.seq_length_diff = (self.max_seq_length - self.min_seq_length) // (self.batch_size - 1)
+        self.spectrogram_length = spectrogram_length
+        self.feature_size = feature_size
+        self.num_audio_channels = num_audio_channels
+        self.hop_length = hop_length
+        self.chunk_length = chunk_length
+        self.sampling_rate = sampling_rate
+
+    def prepare_feat_extract_dict(self):
+        return {
+            "spectrogram_length": self.spectrogram_length,
+            "feature_size": self.feature_size,
+            "num_audio_channels": self.num_audio_channels,
+            "hop_length": self.hop_length,
+            "chunk_length": self.chunk_length,
+            "sampling_rate": self.sampling_rate,
+        }
+
+    def prepare_inputs_for_common(self, equal_length=False, numpify=False):
+        def _flatten(list_of_lists):
+            return list(itertools.chain(*list_of_lists))
+
+        if equal_length:
+            speech_inputs = [floats_list((self.max_seq_length, self.feature_size)) for _ in range(self.batch_size)]
+        else:
+            # make sure that inputs increase in size
+            speech_inputs = [
+                floats_list((x, self.feature_size))
+                for x in range(self.min_seq_length, self.max_seq_length, self.seq_length_diff)
+            ]
+        if numpify:
+            speech_inputs = [np.asarray(x) for x in speech_inputs]
+        return speech_inputs
+
+
+@require_torch
+@require_torchaudio
+class TvltFeatureExtractionTest(SequenceFeatureExtractionTestMixin, unittest.TestCase):
+    feature_extraction_class = TvltFeatureExtractor if is_speech_available() else None
+
+    def setUp(self):
+        self.feat_extract_tester = TvltFeatureExtractionTester(self)
+
+    def test_feat_extract_properties(self):
+        feature_extractor = self.feature_extraction_class(**self.feat_extract_dict)
+        self.assertTrue(hasattr(feature_extractor, "spectrogram_length"))
+        self.assertTrue(hasattr(feature_extractor, "feature_size"))
+        self.assertTrue(hasattr(feature_extractor, "num_audio_channels"))
+        self.assertTrue(hasattr(feature_extractor, "hop_length"))
+        self.assertTrue(hasattr(feature_extractor, "chunk_length"))
+        self.assertTrue(hasattr(feature_extractor, "sampling_rate"))
+
+    def test_feat_extract_from_and_save_pretrained(self):
+        feat_extract_first = self.feature_extraction_class(**self.feat_extract_dict)
+
+        with tempfile.TemporaryDirectory() as tmpdirname:
+            saved_file = feat_extract_first.save_pretrained(tmpdirname)[0]
+            check_json_file_has_correct_format(saved_file)
+            feat_extract_second = self.feature_extraction_class.from_pretrained(tmpdirname)
+
+        dict_first = feat_extract_first.to_dict()
+        dict_second = feat_extract_second.to_dict()
+        mel_1 = dict_first.pop("mel_filters")
+        mel_2 = dict_second.pop("mel_filters")
+        self.assertTrue(np.allclose(mel_1, mel_2))
+        self.assertEqual(dict_first, dict_second)
+
+    def test_feat_extract_to_json_file(self):
+        feat_extract_first = self.feature_extraction_class(**self.feat_extract_dict)
+
+        with tempfile.TemporaryDirectory() as tmpdirname:
+            json_file_path = os.path.join(tmpdirname, "feat_extract.json")
+            feat_extract_first.to_json_file(json_file_path)
+            feat_extract_second = self.feature_extraction_class.from_json_file(json_file_path)
+
+        dict_first = feat_extract_first.to_dict()
+        dict_second = feat_extract_second.to_dict()
+        mel_1 = dict_first.pop("mel_filters")
+        mel_2 = dict_second.pop("mel_filters")
+        self.assertTrue(np.allclose(mel_1, mel_2))
+        self.assertEqual(dict_first, dict_second)
+
+    def test_call(self):
+        # Initialize feature_extractor
+        feature_extractor = self.feature_extraction_class(**self.feat_extract_dict)
+
+        # create three inputs of length 800, 1000, and 1200
+        speech_inputs = [floats_list((1, x))[0] for x in range(8000, 14000, 20000)]
+        np_speech_inputs = [np.asarray(speech_input) for speech_input in speech_inputs]
+
+        # Test not batched input
+        encoded_audios = feature_extractor(np_speech_inputs[0], return_tensors="np", sampling_rate=44100).audio_values
+
+        self.assertTrue(encoded_audios.ndim == 4)
+        self.assertTrue(encoded_audios.shape[-1] == feature_extractor.feature_size)
+        self.assertTrue(encoded_audios.shape[-2] <= feature_extractor.spectrogram_length)
+        self.assertTrue(encoded_audios.shape[-3] == feature_extractor.num_channels)
+
+        # Test batched
+        encoded_audios = feature_extractor(np_speech_inputs, return_tensors="np", sampling_rate=44100).audio_values
+
+        self.assertTrue(encoded_audios.ndim == 4)
+        self.assertTrue(encoded_audios.shape[-1] == feature_extractor.feature_size)
+        self.assertTrue(encoded_audios.shape[-2] <= feature_extractor.spectrogram_length)
+        self.assertTrue(encoded_audios.shape[-3] == feature_extractor.num_channels)
+
+        # Test audio masking
+        encoded_audios = feature_extractor(
+            np_speech_inputs, return_tensors="np", sampling_rate=44100, mask_audio=True
+        ).audio_values
+
+        self.assertTrue(encoded_audios.ndim == 4)
+        self.assertTrue(encoded_audios.shape[-1] == feature_extractor.feature_size)
+        self.assertTrue(encoded_audios.shape[-2] <= feature_extractor.spectrogram_length)
+        self.assertTrue(encoded_audios.shape[-3] == feature_extractor.num_channels)
+
+    def _load_datasamples(self, num_samples):
+        ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
+        # automatic decoding with librispeech
+        speech_samples = ds.sort("id").select(range(num_samples))[:num_samples]["audio"]
+
+        return [x["array"] for x in speech_samples]
+
+    def test_integration(self):
+        input_speech = self._load_datasamples(1)
+        feaure_extractor = TvltFeatureExtractor()
+        audio_values = feaure_extractor(input_speech, return_tensors="pt").audio_values
+
+        self.assertTrue(audio_values.shape, [1, 1, 192, 128])
+
+        expected_slice = torch.tensor([[-0.3032, -0.2708], [-0.4434, -0.4007]])
+        self.assertTrue(torch.allclose(audio_values[0, 0, :2, :2], expected_slice, atol=1e-4))

tests/models/tvlt/test_image_processor_tvlt.py

+# coding=utf-8
+# Copyright 2023 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.
+""" Testing suite for the TVLT image processor. """
+
+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_image_processing_common import ImageProcessingSavingTestMixin
+
+
+if is_torch_available():
+    import torch
+
+if is_vision_available():
+    from PIL import Image
+
+    from transformers import TvltImageProcessor
+
+
+def prepare_video(image_processor_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(image_processor_tester.num_frames):
+        video.append(np.random.randint(255, size=(image_processor_tester.num_channels, width, height), dtype=np.uint8))
+
+    if not numpify and not torchify:
+        # PIL expects the channel dimension as last dimension
+        video = [Image.fromarray(np.moveaxis(frame, 0, -1)) for frame in video]
+
+    if torchify:
+        video = [torch.from_numpy(frame) for frame in video]
+
+    return video
+
+
+def prepare_video_inputs(image_processor_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(image_processor_tester.batch_size):
+        if equal_resolution:
+            width = height = image_processor_tester.max_resolution
+        else:
+            width, height = np.random.choice(
+                np.arange(image_processor_tester.min_resolution, image_processor_tester.max_resolution), 2
+            )
+            video = prepare_video(
+                image_processor_tester=image_processor_tester,
+                width=width,
+                height=height,
+                numpify=numpify,
+                torchify=torchify,
+            )
+        video_inputs.append(video)
+
+    return video_inputs
+
+
+class TvltImageProcessorTester(unittest.TestCase):
+    def __init__(
+        self,
+        parent,
+        batch_size=7,
+        num_channels=3,
+        num_frames=4,
+        image_size=18,
+        min_resolution=30,
+        max_resolution=400,
+        do_resize=True,
+        size=None,
+        do_normalize=True,
+        image_mean=[0.5, 0.5, 0.5],
+        image_std=[0.5, 0.5, 0.5],
+        do_center_crop=True,
+        crop_size=None,
+    ):
+        size = size if size is not None else {"shortest_edge": 18}
+        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.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
+        self.do_center_crop = do_center_crop
+        self.crop_size = crop_size
+
+    def prepare_image_processor_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,
+            "do_center_crop": self.do_center_crop,
+            "crop_size": self.crop_size,
+        }
+
+
+@require_torch
+@require_vision
+class TvltImageProcessorTest(ImageProcessingSavingTestMixin, unittest.TestCase):
+    image_processing_class = TvltImageProcessor if is_vision_available() else None
+
+    def setUp(self):
+        self.image_processor_tester = TvltImageProcessorTester(self)
+
+    @property
+    def image_processor_dict(self):
+        return self.image_processor_tester.prepare_image_processor_dict()
+
+    def test_image_processor_properties(self):
+        image_processor = self.image_processing_class(**self.image_processor_dict)
+        self.assertTrue(hasattr(image_processor, "image_mean"))
+        self.assertTrue(hasattr(image_processor, "image_std"))
+        self.assertTrue(hasattr(image_processor, "do_normalize"))
+        self.assertTrue(hasattr(image_processor, "do_resize"))
+        self.assertTrue(hasattr(image_processor, "do_center_crop"))
+        self.assertTrue(hasattr(image_processor, "size"))
+
+    def test_call_pil(self):
+        # Initialize image_processor
+        image_processor = self.image_processing_class(**self.image_processor_dict)
+        # create random PIL videos
+        video_inputs = prepare_video_inputs(self.image_processor_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 = image_processor(video_inputs[0], return_tensors="pt").pixel_values
+        self.assertEqual(
+            encoded_videos.shape,
+            (
+                1,
+                self.image_processor_tester.num_frames,
+                self.image_processor_tester.num_channels,
+                self.image_processor_tester.crop_size["height"],
+                self.image_processor_tester.crop_size["width"],
+            ),
+        )
+
+        # Test batched
+        encoded_videos = image_processor(video_inputs, return_tensors="pt").pixel_values
+        self.assertEqual(
+            encoded_videos.shape,
+            (
+                self.image_processor_tester.batch_size,
+                self.image_processor_tester.num_frames,
+                self.image_processor_tester.num_channels,
+                self.image_processor_tester.crop_size["height"],
+                self.image_processor_tester.crop_size["width"],
+            ),
+        )
+
+    def test_call_numpy(self):
+        # Initialize image_processor
+        image_processor = self.image_processing_class(**self.image_processor_dict)
+        # create random numpy tensors
+        video_inputs = prepare_video_inputs(self.image_processor_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 = image_processor(video_inputs[0], return_tensors="pt").pixel_values
+        self.assertEqual(
+            encoded_videos.shape,
+            (
+                1,
+                self.image_processor_tester.num_frames,
+                self.image_processor_tester.num_channels,
+                self.image_processor_tester.crop_size["height"],
+                self.image_processor_tester.crop_size["width"],
+            ),
+        )
+
+        # Test batched
+        encoded_videos = image_processor(video_inputs, return_tensors="pt").pixel_values
+        self.assertEqual(
+            encoded_videos.shape,
+            (
+                self.image_processor_tester.batch_size,
+                self.image_processor_tester.num_frames,
+                self.image_processor_tester.num_channels,
+                self.image_processor_tester.crop_size["height"],
+                self.image_processor_tester.crop_size["width"],
+            ),
+        )
+
+    def test_call_pytorch(self):
+        # Initialize image_processor
+        image_processor = self.image_processing_class(**self.image_processor_dict)
+        # create random PyTorch tensors
+        video_inputs = prepare_video_inputs(self.image_processor_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 = image_processor(video_inputs[0], return_tensors="pt").pixel_values
+        self.assertEqual(
+            encoded_videos.shape,
+            (
+                1,
+                self.image_processor_tester.num_frames,
+                self.image_processor_tester.num_channels,
+                self.image_processor_tester.crop_size["height"],
+                self.image_processor_tester.crop_size["width"],
+            ),
+        )
+
+        # Test batched
+        encoded_videos = image_processor(video_inputs, return_tensors="pt").pixel_values
+        self.assertEqual(
+            encoded_videos.shape,
+            (
+                self.image_processor_tester.batch_size,
+                self.image_processor_tester.num_frames,
+                self.image_processor_tester.num_channels,
+                self.image_processor_tester.crop_size["height"],
+                self.image_processor_tester.crop_size["width"],
+            ),
+        )

tests/models/tvlt/test_modeling_tvlt.py

+# coding=utf-8
+# Copyright 2023 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 TVLT model. """
+
+import copy
+import inspect
+import unittest
+
+import numpy as np
+from huggingface_hub import hf_hub_download
+
+from transformers import (
+    TvltConfig,
+    is_datasets_available,
+    is_speech_available,
+    is_torch_available,
+    is_vision_available,
+)
+from transformers.testing_utils import require_torch, require_vision, slow, torch_device
+from transformers.utils import cached_property
+
+from ...test_configuration_common import ConfigTester
+from ...test_modeling_common import ModelTesterMixin, floats_tensor
+
+
+if is_torch_available():
+    import torch
+    import torch.nn as nn
+
+    from transformers import TvltForAudioVisualClassification, TvltForPreTraining, TvltModel
+    from transformers.models.tvlt.modeling_tvlt import TVLT_PRETRAINED_MODEL_ARCHIVE_LIST
+    from transformers.pytorch_utils import is_torch_greater_or_equal_than_1_10
+else:
+    is_torch_greater_or_equal_than_1_10 = False
+
+
+if is_datasets_available():
+    from datasets import load_dataset
+
+if is_vision_available():
+    from transformers import TvltImageProcessor
+
+if is_speech_available():
+    from transformers import TvltFeatureExtractor
+
+
+class TvltModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=2,
+        image_size=32,
+        spectrogram_length=32,
+        frequency_length=16,
+        image_patch_size=[2, 2],
+        audio_patch_size=[2, 2],
+        num_image_channels=3,
+        num_audio_channels=1,
+        num_frames=2,
+        hidden_size=128,
+        num_hidden_layers=12,
+        num_attention_heads=4,
+        intermediate_size=128,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.0,
+        attention_probs_dropout_prob=0.0,
+        initializer_range=0.02,
+        layer_norm_eps=1e-12,
+        qkv_bias=True,
+        use_mean_pooling=True,
+        decoder_num_attention_heads=4,
+        decoder_hidden_size=64,
+        decoder_num_hidden_layers=2,
+        decoder_intermediate_size=128,
+        image_mask_ratio=0.75,
+        audio_mask_ratio=0.15,
+        audio_mask_type="frame-level",
+        task_matching=True,
+        task_mae=True,
+        num_labels=1,
+        is_training=True,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.image_size = image_size
+        self.spectrogram_length = spectrogram_length
+        self.frequency_length = frequency_length
+        self.image_patch_size = image_patch_size
+        self.audio_patch_size = audio_patch_size
+        self.num_image_channels = num_image_channels
+        self.num_audio_channels = num_audio_channels
+        self.num_frames = num_frames
+
+        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.initializer_range = initializer_range
+        self.layer_norm_eps = layer_norm_eps
+        self.qkv_bias = qkv_bias
+        self.use_mean_pooling = use_mean_pooling
+
+        self.decoder_num_attention_heads = decoder_num_attention_heads
+        self.decoder_hidden_size = decoder_hidden_size
+        self.decoder_num_hidden_layers = decoder_num_hidden_layers
+        self.decoder_intermediate_size = decoder_intermediate_size
+        self.image_mask_ratio = image_mask_ratio
+        self.audio_mask_ratio = audio_mask_ratio
+
+        self.task_matching = task_matching
+        self.task_mae = task_mae
+        self.num_labels = num_labels
+
+        self.expected_pixel_seq_len = (self.image_size // self.image_patch_size[0]) ** 2 * self.num_frames
+        self.expected_audio_seq_len = (self.spectrogram_length // self.audio_patch_size[0]) * (
+            self.frequency_length // self.audio_patch_size[1]
+        )
+        # we set the expected sequence length (which is used in several tests)
+        # this is equal to the seq length of number of image/video patches + number of audio patches
+        self.expected_seq_len = self.expected_pixel_seq_len + self.expected_audio_seq_len + 1
+
+        self.image_mae_output_dim = image_patch_size[0] ** 2 * num_image_channels
+        self.audio_mae_output_dim = audio_patch_size[0] * audio_patch_size[1] * num_audio_channels
+        self.is_training = is_training
+
+    def prepare_config_and_inputs(self):
+        pixel_values = floats_tensor(
+            [self.batch_size, self.num_frames, self.num_image_channels, self.image_size, self.image_size]
+        )
+        audio_values = floats_tensor(
+            [self.batch_size, self.num_audio_channels, self.spectrogram_length, self.frequency_length]
+        )
+
+        pixel_mask = floats_tensor([self.batch_size, self.expected_pixel_seq_len])
+        audio_mask = floats_tensor([self.batch_size, self.expected_audio_seq_len])
+
+        config = self.get_config()
+
+        return (config, pixel_values, audio_values, pixel_mask, audio_mask)
+
+    def prepare_config_and_inputs_for_pretraining(self):
+        pixel_values = floats_tensor(
+            [self.batch_size, self.num_frames, self.num_image_channels, self.image_size, self.image_size]
+        )
+        audio_values = floats_tensor(
+            [self.batch_size, self.num_audio_channels, self.spectrogram_length, self.frequency_length]
+        )
+
+        pixel_mask = floats_tensor([self.batch_size, self.expected_pixel_seq_len])
+        audio_mask = floats_tensor([self.batch_size, self.expected_audio_seq_len])
+
+        pixel_values_mixed = floats_tensor(
+            [self.batch_size, self.num_frames, self.num_image_channels, self.image_size, self.image_size]
+        )
+        pixel_mask_mixed = floats_tensor([self.batch_size, self.expected_pixel_seq_len])
+        labels = floats_tensor([self.batch_size])
+        config = self.get_config()
+
+        return (
+            config,
+            pixel_values,
+            audio_values,
+            pixel_mask,
+            audio_mask,
+            pixel_values_mixed,
+            pixel_mask_mixed,
+            labels,
+        )
+
+    def get_config(self):
+        return TvltConfig(
+            image_size=self.image_size,
+            spectrogram_length=self.spectrogram_length,
+            frequency_length=self.frequency_length,
+            image_patch_size=self.image_patch_size,
+            audio_patch_size=self.audio_patch_size,
+            num_image_channels=self.num_image_channels,
+            num_audio_channels=self.num_audio_channels,
+            num_frames=self.num_frames,
+            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,
+            initializer_range=self.initializer_range,
+            layer_norm_eps=self.layer_norm_eps,
+            qkv_bias=self.qkv_bias,
+            use_mean_pooling=self.use_mean_pooling,
+            decoder_num_attention_heads=self.decoder_num_attention_heads,
+            decoder_hidden_size=self.decoder_hidden_size,
+            decoder_num_hidden_layers=self.decoder_num_hidden_layers,
+            decoder_intermediate_size=self.decoder_intermediate_size,
+            image_mask_ratio=self.image_mask_ratio,
+            audio_mask_ratio=self.audio_mask_ratio,
+            task_matching=self.task_matching,
+            task_mae=self.task_mae,
+            num_labels=self.num_labels,
+        )
+
+    def create_and_check_model(self, config, pixel_values, audio_values, pixel_mask, audio_mask):
+        model = TvltModel(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(pixel_values, audio_values, pixel_mask=pixel_mask, audio_mask=audio_mask)
+        result = model(pixel_values, audio_values)
+        self.parent.assertEqual(
+            result.last_hidden_state.shape, (self.batch_size, self.expected_seq_len, self.hidden_size)
+        )
+
+    def create_and_check_for_audiovisual_classification(
+        self, config, pixel_values, audio_values, pixel_mask, audio_mask
+    ):
+        model = TvltForAudioVisualClassification(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(pixel_values, audio_values, pixel_mask=pixel_mask, audio_mask=audio_mask)
+        result = model(pixel_values, audio_values)
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_labels))
+
+    def create_and_check_for_pretraining(
+        self,
+        config,
+        pixel_values,
+        audio_values,
+        pixel_mask,
+        audio_mask,
+        pixel_values_mixed,
+        pixel_mask_mixed,
+        labels,
+    ):
+        model = TvltForPreTraining(config=config)
+        model.to(torch_device)
+        model.train()
+        result = model(
+            pixel_values,
+            audio_values,
+            pixel_mask,
+            audio_mask,
+            pixel_values_mixed=pixel_values_mixed,
+            pixel_mask_mixed=pixel_mask_mixed,
+            labels=labels,
+        )
+        self.parent.assertEqual(
+            result.pixel_logits.shape, (self.batch_size, self.expected_pixel_seq_len, self.image_mae_output_dim)
+        )
+        self.parent.assertEqual(
+            result.audio_logits.shape, (self.batch_size, self.expected_audio_seq_len, self.audio_mae_output_dim)
+        )
+        self.parent.assertEqual(result.matching_logits.shape, (self.batch_size, self.num_labels))
+
+    def create_and_check_for_pretraining_inference(
+        self,
+        config,
+        pixel_values,
+        audio_values,
+        pixel_mask,
+        audio_mask,
+        pixel_values_mixed,
+        pixel_mask_mixed,
+        labels,
+    ):
+        model = TvltForPreTraining(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(
+            pixel_values,
+            audio_values,
+            pixel_mask,
+            audio_mask,
+            pixel_values_mixed=pixel_values_mixed,
+            pixel_mask_mixed=pixel_mask_mixed,
+            labels=labels,
+        )
+        if result.pixel_logits is not None:
+            self.parent.assertEqual(
+                result.pixel_logits.shape, (self.batch_size, self.expected_pixel_seq_len, self.image_mae_output_dim)
+            )
+        if result.audio_logits is not None:
+            self.parent.assertEqual(
+                result.audio_logits.shape, (self.batch_size, self.expected_audio_seq_len, self.audio_mae_output_dim)
+            )
+        self.parent.assertEqual(result.matching_logits.shape, (self.batch_size, self.num_labels))
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        (config, pixel_values, audio_values, pixel_mask, audio_mask) = config_and_inputs
+        inputs_dict = {
+            "pixel_values": pixel_values,
+            "audio_values": audio_values,
+            "pixel_mask": pixel_mask,
+            "audio_mask": audio_mask,
+        }
+        return config, inputs_dict
+
+    def prepare_pixel_values(self):
+        return floats_tensor(
+            [self.batch_size, self.num_frames, self.num_image_channels, self.image_size, self.image_size]
+        )
+
+    def prepare_audio_values(self):
+        return floats_tensor(
+            [self.batch_size, self.num_audio_channels, self.spectrogram_length, self.frequency_length]
+        )
+
+
+@require_torch
+@unittest.skipIf(not is_torch_greater_or_equal_than_1_10, "TVLT is only available in torch v1.10+")
+class TvltModelTest(ModelTesterMixin, unittest.TestCase):
+    all_model_classes = (
+        (TvltModel, TvltForPreTraining, TvltForAudioVisualClassification) if is_torch_available() else ()
+    )
+
+    fx_compatible = False
+    test_pruning = False
+    test_headmasking = False
+    test_torchscript = False
+    test_resize_embeddings = False
+    main_input_name = "pixel_values"
+
+    # TvltForAudioVisualClassification and TvltForPreTraining require special treatment
+    def _prepare_for_class(self, inputs_dict, model_class, return_labels=True):
+        inputs_dict = copy.deepcopy(inputs_dict)
+
+        if return_labels:
+            if model_class.__name__ == "TvltForAudioVisualClassification":
+                inputs_dict["labels"] = torch.zeros(
+                    (self.model_tester.batch_size,), dtype=torch.long, device=torch_device
+                )
+            elif model_class.__name__ == "TvltForPreTraining":
+                inputs_dict["labels"] = torch.zeros(
+                    (self.model_tester.batch_size,), dtype=torch.float, device=torch_device
+                )
+                inputs_dict["pixel_values_mixed"] = torch.zeros(
+                    (
+                        self.model_tester.batch_size,
+                        self.model_tester.num_frames,
+                        self.model_tester.num_image_channels,
+                        self.model_tester.image_size,
+                        self.model_tester.image_size,
+                    ),
+                    dtype=torch.float,
+                    device=torch_device,
+                )
+                inputs_dict["pixel_mask_mixed"] = torch.zeros(
+                    (self.model_tester.batch_size, self.model_tester.expected_pixel_seq_len),
+                    dtype=torch.float,
+                    device=torch_device,
+                )
+
+        return inputs_dict
+
+    def setUp(self):
+        self.model_tester = TvltModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=TvltConfig, has_text_modality=False, hidden_size=37)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    @unittest.skip(reason="TVLT 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)
+            input_embeddings = model.get_input_embeddings()
+            self.assertIsInstance(input_embeddings, (tuple))
+            for embedding in input_embeddings:
+                self.assertIsInstance(embedding, (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", "audio_values"]
+            self.assertListEqual(arg_names[:2], 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_audiovisual_classification(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_audiovisual_classification(*config_and_inputs)
+
+    def test_for_pretraining(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs_for_pretraining()
+        self.model_tester.create_and_check_for_pretraining(*config_and_inputs)
+        self.model_tester.create_and_check_for_pretraining_inference(*config_and_inputs)
+
+    @slow
+    def test_model_from_pretrained(self):
+        for model_name in TVLT_PRETRAINED_MODEL_ARCHIVE_LIST:
+            model = TvltModel.from_pretrained(model_name)
+            self.assertIsNotNone(model)
+
+    def test_training(self):
+        if not self.model_tester.is_training:
+            return
+
+        for model_class in self.all_model_classes[1:]:
+            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+            config.return_dict = True
+
+            model = model_class(config)
+            model.to(torch_device)
+            model.train()
+            inputs = self._prepare_for_class(inputs_dict, model_class)
+            for k, v in inputs.items():
+                print(k, v.shape)
+            loss = model(**inputs).loss
+            loss.backward()
+
+    def test_training_gradient_checkpointing(self):
+        if not self.model_tester.is_training:
+            return
+
+        for model_class in self.all_model_classes[1:]:
+            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+            config.use_cache = False
+            config.return_dict = True
+
+            model = model_class(config)
+            model.to(torch_device)
+            model.gradient_checkpointing_enable()
+            model.train()
+            inputs = self._prepare_for_class(inputs_dict, model_class)
+            loss = model(**inputs).loss
+            loss.backward()
+
+    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[2:]:
+                seq_len = self.model_tester.expected_seq_len
+
+                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)
+
+            seq_length = self.model_tester.expected_seq_len
+
+            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[2:]:
+            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(num_frames=8):
+    file = hf_hub_download(
+        repo_id="hf-internal-testing/spaghetti-video", filename="eating_spaghetti.npy", repo_type="dataset"
+    )
+    video = np.load(file)[:num_frames]
+    return list(video)
+
+
+def prepare_audio(num_samples=1):
+    ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
+    # automatic decoding with librispeech
+    speech_samples = ds.sort("id").select(range(num_samples))[:num_samples]["audio"]
+    return [x["array"] for x in speech_samples]
+
+
+@require_torch
+@require_vision
+class TvltModelIntegrationTest(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 (
+            TvltImageProcessor() if is_vision_available() else None,
+            TvltFeatureExtractor(),
+        )
+
+    def test_inference_for_base_model(self):
+        model = TvltModel.from_pretrained("ZinengTang/tvlt-base").to(torch_device)
+
+        image_processor, audio_feature_extractor = self.default_feature_extractor
+        video = prepare_video()
+        audio = prepare_audio()
+        video_inputs = image_processor(video, return_tensors="pt").to(torch_device)
+        audio_inputs = audio_feature_extractor(audio, return_tensors="pt").to(torch_device)
+        inputs = dict()
+        inputs.update(video_inputs)
+        inputs.update(audio_inputs)
+
+        # forward pass
+        with torch.no_grad():
+            outputs = model(**inputs)
+
+        # verify the logits
+        expected_last_hidden_state_slice = torch.tensor([[-0.0186, -0.0691], [0.0242, -0.0398]])
+        self.assertTrue(
+            torch.allclose(outputs.last_hidden_state[:, :2, :2], expected_last_hidden_state_slice, atol=1e-4)
+        )
+
+    def test_inference_for_pretraining(self):
+        model = TvltForPreTraining.from_pretrained("ZinengTang/tvlt-base").to(torch_device)
+
+        image_processor, audio_feature_extractor = self.default_feature_extractor
+        video = prepare_video()
+        video_mixed = prepare_video()
+        audio = prepare_audio()
+        video_inputs = image_processor(video, return_tensors="pt", mask_pixel=True).to(torch_device)
+        video_mixed_inputs = image_processor(video_mixed, is_mixed=True, return_tensors="pt").to(torch_device)
+        audio_inputs = audio_feature_extractor(audio, return_tensors="pt", mask_audio=True).to(torch_device)
+        labels = torch.tensor([[0.0]], device=torch_device)
+        inputs = dict()
+        inputs.update(video_inputs)
+        inputs.update(video_mixed_inputs)
+        inputs.update(audio_inputs)
+        inputs.update({"labels": labels})
+
+        # forward pass
+        with torch.no_grad():
+            outputs = model(**inputs)
+
+        # verify the logits
+        expected_pixel_logits_shape = torch.Size([1, 1568, 768])
+        expected_audio_logits_shape = torch.Size([1, 96, 256])
+        expected_matching_logits_shape = torch.Size([1, 1])
+
+        if outputs.pixel_logits is not None:
+            self.assertEqual(outputs.pixel_logits.shape, expected_pixel_logits_shape)
+        if outputs.audio_logits is not None:
+            self.assertEqual(outputs.audio_logits.shape, expected_audio_logits_shape)
+        self.assertTrue(outputs.matching_logits.shape, expected_matching_logits_shape)

tests/models/tvlt/test_processor_tvlt.py

+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import shutil
+import tempfile
+import unittest
+
+import numpy as np
+import pytest
+
+from transformers import is_speech_available, is_vision_available
+from transformers.testing_utils import require_torch
+
+
+if is_vision_available():
+    from transformers import TvltImageProcessor
+
+if is_speech_available():
+    from transformers import TvltFeatureExtractor
+
+from transformers import TvltProcessor
+
+
+@require_torch
+class TvltProcessorTest(unittest.TestCase):
+    def setUp(self):
+        self.checkpoint = "ZinengTang/tvlt-base"
+        self.tmpdirname = tempfile.mkdtemp()
+
+    def get_image_processor(self, **kwargs):
+        return TvltImageProcessor.from_pretrained(self.checkpoint, **kwargs)
+
+    def get_feature_extractor(self, **kwargs):
+        return TvltFeatureExtractor.from_pretrained(self.checkpoint, **kwargs)
+
+    def tearDown(self):
+        shutil.rmtree(self.tmpdirname)
+
+    def test_save_load_pretrained_default(self):
+        image_processor = self.get_image_processor()
+        feature_extractor = self.get_feature_extractor()
+
+        processor = TvltProcessor(image_processor=image_processor, feature_extractor=feature_extractor)
+        processor.save_pretrained(self.tmpdirname)
+        processor = TvltProcessor.from_pretrained(self.tmpdirname)
+
+        self.assertIsInstance(processor.feature_extractor, TvltFeatureExtractor)
+        self.assertIsInstance(processor.image_processor, TvltImageProcessor)
+
+    def test_feature_extractor(self):
+        image_processor = self.get_image_processor()
+        feature_extractor = self.get_feature_extractor()
+
+        processor = TvltProcessor(image_processor=image_processor, feature_extractor=feature_extractor)
+
+        audio = np.ones([12000])
+
+        audio_dict = feature_extractor(audio, return_tensors="np")
+        input_processor = processor(audio=audio, return_tensors="np")
+
+        for key in audio_dict.keys():
+            self.assertAlmostEqual(audio_dict[key].sum(), input_processor[key].sum(), delta=1e-2)
+
+    def test_image_processor(self):
+        image_processor = self.get_image_processor()
+        feature_extractor = self.get_feature_extractor()
+
+        processor = TvltProcessor(image_processor=image_processor, feature_extractor=feature_extractor)
+
+        images = np.ones([3, 224, 224])
+
+        image_dict = image_processor(images, return_tensors="np")
+        input_processor = processor(images=images, return_tensors="np")
+
+        for key in image_dict.keys():
+            self.assertAlmostEqual(image_dict[key].sum(), input_processor[key].sum(), delta=1e-2)
+
+    def test_processor(self):
+        image_processor = self.get_image_processor()
+        feature_extractor = self.get_feature_extractor()
+
+        processor = TvltProcessor(image_processor=image_processor, feature_extractor=feature_extractor)
+
+        audio = np.ones([12000])
+        images = np.ones([3, 224, 224])
+
+        inputs = processor(audio=audio, images=images)
+
+        self.assertListEqual(list(inputs.keys()), ["audio_values", "audio_mask", "pixel_values", "pixel_mask"])
+
+        # test if it raises when no input is passed
+        with pytest.raises(ValueError):
+            processor()
+
+    def test_model_input_names(self):
+        image_processor = self.get_image_processor()
+        feature_extractor = self.get_feature_extractor()
+
+        processor = TvltProcessor(image_processor=image_processor, feature_extractor=feature_extractor)
+
+        self.assertListEqual(
+            processor.model_input_names,
+            image_processor.model_input_names + feature_extractor.model_input_names,
+            msg="`processor` and `image_processor`+`feature_extractor` model input names do not match",
+        )

utils/check_repo.py

@@ -287,6 +287,7 @@ IGNORE_NON_AUTO_CONFIGURED = PRIVATE_MODELS.copy() + [
     "AltCLIPTextModel",
     "AltCLIPVisionModel",
     "AltRobertaModel",
+    "TvltForAudioVisualClassification",
     "SpeechT5ForSpeechToSpeech",
     "SpeechT5ForTextToSpeech",
     "SpeechT5HifiGan",