Move test model folders (#17034)

* move test model folders (TODO: fix imports and others)

* fix (potentially partially) imports (in model test modules)

* fix (potentially partially) imports (in tokenization test modules)

* fix (potentially partially) imports (in feature extraction test modules)

* fix import utils.test_modeling_tf_core

* fix path ../fixtures/

* fix imports about generation.test_generation_flax_utils

* fix more imports

* fix fixture path

* fix get_test_dir

* update module_to_test_file

* fix get_tests_dir from wrong transformers.utils

* update config.yml (CircleCI)

* fix style

* remove missing imports

* update new model script

* update check_repo

* update SPECIAL_MODULE_TO_TEST_MAP

* fix style

* add __init__

* update self-scheduled

* fix add_new_model scripts

* check one way to get location back

* python setup.py build install

* fix import in test auto

* update self-scheduled.yml

* update slack notification script

* Add comments about artifact names

* fix for yolos

Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>

tests/models/clip/test_feature_extraction_clip.py

@@ -0,0 +1,229 @@
+# coding=utf-8
+# Copyright 2021 HuggingFace Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import unittest
+
+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
+
+
+if is_torch_available():
+    import torch
+
+if is_vision_available():
+    from PIL import Image
+
+    from transformers import CLIPFeatureExtractor
+
+
+class CLIPFeatureExtractionTester(unittest.TestCase):
+    def __init__(
+        self,
+        parent,
+        batch_size=7,
+        num_channels=3,
+        image_size=18,
+        min_resolution=30,
+        max_resolution=400,
+        do_resize=True,
+        size=20,
+        do_center_crop=True,
+        crop_size=18,
+        do_normalize=True,
+        image_mean=[0.48145466, 0.4578275, 0.40821073],
+        image_std=[0.26862954, 0.26130258, 0.27577711],
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.num_channels = num_channels
+        self.image_size = image_size
+        self.min_resolution = min_resolution
+        self.max_resolution = max_resolution
+        self.do_resize = do_resize
+        self.size = size
+        self.do_center_crop = do_center_crop
+        self.crop_size = crop_size
+        self.do_normalize = do_normalize
+        self.image_mean = image_mean
+        self.image_std = image_std
+
+    def prepare_feat_extract_dict(self):
+        return {
+            "do_resize": self.do_resize,
+            "size": self.size,
+            "do_center_crop": self.do_center_crop,
+            "crop_size": self.crop_size,
+            "do_normalize": self.do_normalize,
+            "image_mean": self.image_mean,
+            "image_std": self.image_std,
+        }
+
+    def prepare_inputs(self, 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.
+        """
+
+        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(self.batch_size):
+                image_inputs.append(
+                    np.random.randint(
+                        255, size=(self.num_channels, self.max_resolution, self.max_resolution), dtype=np.uint8
+                    )
+                )
+        else:
+            image_inputs = []
+            for i in range(self.batch_size):
+                width, height = np.random.choice(np.arange(self.min_resolution, self.max_resolution), 2)
+                image_inputs.append(np.random.randint(255, size=(self.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]
+
+        if torchify:
+            image_inputs = [torch.from_numpy(x) for x in image_inputs]
+
+        return image_inputs
+
+
+@require_torch
+@require_vision
+class CLIPFeatureExtractionTest(FeatureExtractionSavingTestMixin, unittest.TestCase):
+
+    feature_extraction_class = CLIPFeatureExtractor if is_vision_available() else None
+
+    def setUp(self):
+        self.feature_extract_tester = CLIPFeatureExtractionTester(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, "do_resize"))
+        self.assertTrue(hasattr(feature_extractor, "size"))
+        self.assertTrue(hasattr(feature_extractor, "do_center_crop"))
+        self.assertTrue(hasattr(feature_extractor, "center_crop"))
+        self.assertTrue(hasattr(feature_extractor, "do_normalize"))
+        self.assertTrue(hasattr(feature_extractor, "image_mean"))
+        self.assertTrue(hasattr(feature_extractor, "image_std"))
+
+    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 images
+        image_inputs = self.feature_extract_tester.prepare_inputs(equal_resolution=False)
+        for image in image_inputs:
+            self.assertIsInstance(image, Image.Image)
+
+        # Test not batched input
+        encoded_images = feature_extractor(image_inputs[0], return_tensors="pt").pixel_values
+        self.assertEqual(
+            encoded_images.shape,
+            (
+                1,
+                self.feature_extract_tester.num_channels,
+                self.feature_extract_tester.crop_size,
+                self.feature_extract_tester.crop_size,
+            ),
+        )
+
+        # Test batched
+        encoded_images = feature_extractor(image_inputs, return_tensors="pt").pixel_values
+        self.assertEqual(
+            encoded_images.shape,
+            (
+                self.feature_extract_tester.batch_size,
+                self.feature_extract_tester.num_channels,
+                self.feature_extract_tester.crop_size,
+                self.feature_extract_tester.crop_size,
+            ),
+        )
+
+    def test_call_numpy(self):
+        # Initialize feature_extractor
+        feature_extractor = self.feature_extraction_class(**self.feat_extract_dict)
+        # create random numpy tensors
+        image_inputs = self.feature_extract_tester.prepare_inputs(equal_resolution=False, numpify=True)
+        for image in image_inputs:
+            self.assertIsInstance(image, np.ndarray)
+
+        # Test not batched input
+        encoded_images = feature_extractor(image_inputs[0], return_tensors="pt").pixel_values
+        self.assertEqual(
+            encoded_images.shape,
+            (
+                1,
+                self.feature_extract_tester.num_channels,
+                self.feature_extract_tester.crop_size,
+                self.feature_extract_tester.crop_size,
+            ),
+        )
+
+        # Test batched
+        encoded_images = feature_extractor(image_inputs, return_tensors="pt").pixel_values
+        self.assertEqual(
+            encoded_images.shape,
+            (
+                self.feature_extract_tester.batch_size,
+                self.feature_extract_tester.num_channels,
+                self.feature_extract_tester.crop_size,
+                self.feature_extract_tester.crop_size,
+            ),
+        )
+
+    def test_call_pytorch(self):
+        # Initialize feature_extractor
+        feature_extractor = self.feature_extraction_class(**self.feat_extract_dict)
+        # create random PyTorch tensors
+        image_inputs = self.feature_extract_tester.prepare_inputs(equal_resolution=False, torchify=True)
+        for image in image_inputs:
+            self.assertIsInstance(image, torch.Tensor)
+
+        # Test not batched input
+        encoded_images = feature_extractor(image_inputs[0], return_tensors="pt").pixel_values
+        self.assertEqual(
+            encoded_images.shape,
+            (
+                1,
+                self.feature_extract_tester.num_channels,
+                self.feature_extract_tester.crop_size,
+                self.feature_extract_tester.crop_size,
+            ),
+        )
+
+        # Test batched
+        encoded_images = feature_extractor(image_inputs, return_tensors="pt").pixel_values
+        self.assertEqual(
+            encoded_images.shape,
+            (
+                self.feature_extract_tester.batch_size,
+                self.feature_extract_tester.num_channels,
+                self.feature_extract_tester.crop_size,
+                self.feature_extract_tester.crop_size,
+            ),
+        )

tests/models/clip/test_modeling_clip.py

@@ -0,0 +1,767 @@
+# coding=utf-8
+# Copyright 2021 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Testing suite for the PyTorch CLIP model. """
+
+
+import inspect
+import os
+import tempfile
+import unittest
+
+import numpy as np
+
+import requests
+import transformers
+from transformers import CLIPConfig, CLIPTextConfig, CLIPVisionConfig
+from transformers.testing_utils import (
+    is_flax_available,
+    is_pt_flax_cross_test,
+    require_torch,
+    require_vision,
+    slow,
+    torch_device,
+)
+from transformers.utils import is_torch_available, is_vision_available
+
+from ...test_configuration_common import ConfigTester
+from ...test_modeling_common import (
+    ModelTesterMixin,
+    _config_zero_init,
+    floats_tensor,
+    ids_tensor,
+    random_attention_mask,
+)
+
+
+if is_torch_available():
+    import torch
+    from torch import nn
+
+    from transformers import CLIPModel, CLIPTextModel, CLIPVisionModel
+    from transformers.models.clip.modeling_clip import CLIP_PRETRAINED_MODEL_ARCHIVE_LIST
+
+
+if is_vision_available():
+    from PIL import Image
+
+    from transformers import CLIPProcessor
+
+
+if is_flax_available():
+    import jax.numpy as jnp
+    from transformers.modeling_flax_pytorch_utils import (
+        convert_pytorch_state_dict_to_flax,
+        load_flax_weights_in_pytorch_model,
+    )
+
+
+class CLIPVisionModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=12,
+        image_size=30,
+        patch_size=2,
+        num_channels=3,
+        is_training=True,
+        hidden_size=32,
+        num_hidden_layers=5,
+        num_attention_heads=4,
+        intermediate_size=37,
+        dropout=0.1,
+        attention_dropout=0.1,
+        initializer_range=0.02,
+        scope=None,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.image_size = image_size
+        self.patch_size = patch_size
+        self.num_channels = num_channels
+        self.is_training = is_training
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.intermediate_size = intermediate_size
+        self.dropout = dropout
+        self.attention_dropout = attention_dropout
+        self.initializer_range = initializer_range
+        self.scope = scope
+
+    def prepare_config_and_inputs(self):
+        pixel_values = floats_tensor([self.batch_size, self.num_channels, self.image_size, self.image_size])
+        config = self.get_config()
+
+        return config, pixel_values
+
+    def get_config(self):
+        return CLIPVisionConfig(
+            image_size=self.image_size,
+            patch_size=self.patch_size,
+            num_channels=self.num_channels,
+            hidden_size=self.hidden_size,
+            num_hidden_layers=self.num_hidden_layers,
+            num_attention_heads=self.num_attention_heads,
+            intermediate_size=self.intermediate_size,
+            dropout=self.dropout,
+            attention_dropout=self.attention_dropout,
+            initializer_range=self.initializer_range,
+        )
+
+    def create_and_check_model(self, config, pixel_values):
+        model = CLIPVisionModel(config=config)
+        model.to(torch_device)
+        model.eval()
+        with torch.no_grad():
+            result = model(pixel_values)
+        # expected sequence length = num_patches + 1 (we add 1 for the [CLS] token)
+        image_size = (self.image_size, self.image_size)
+        patch_size = (self.patch_size, self.patch_size)
+        num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
+        self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, num_patches + 1, self.hidden_size))
+        self.parent.assertEqual(result.pooler_output.shape, (self.batch_size, self.hidden_size))
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        config, pixel_values = config_and_inputs
+        inputs_dict = {"pixel_values": pixel_values}
+        return config, inputs_dict
+
+
+@require_torch
+class CLIPVisionModelTest(ModelTesterMixin, unittest.TestCase):
+    """
+    Here we also overwrite some of the tests of test_modeling_common.py, as CLIP does not use input_ids, inputs_embeds,
+    attention_mask and seq_length.
+    """
+
+    all_model_classes = (CLIPVisionModel,) if is_torch_available() else ()
+
+    test_pruning = False
+    test_resize_embeddings = False
+    test_head_masking = False
+
+    def setUp(self):
+        self.model_tester = CLIPVisionModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=CLIPVisionConfig, has_text_modality=False, hidden_size=37)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_inputs_embeds(self):
+        # CLIP does not use inputs_embeds
+        pass
+
+    def test_model_common_attributes(self):
+        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            model = model_class(config)
+            self.assertIsInstance(model.get_input_embeddings(), (nn.Module))
+            x = model.get_output_embeddings()
+            self.assertTrue(x is None or isinstance(x, nn.Linear))
+
+    def test_forward_signature(self):
+        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            model = model_class(config)
+            signature = inspect.signature(model.forward)
+            # signature.parameters is an OrderedDict => so arg_names order is deterministic
+            arg_names = [*signature.parameters.keys()]
+
+            expected_arg_names = ["pixel_values"]
+            self.assertListEqual(arg_names[:1], expected_arg_names)
+
+    def test_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_model(*config_and_inputs)
+
+    def test_attention_outputs(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+        config.return_dict = True
+
+        # in CLIP, the seq_len equals the number of patches + 1 (we add 1 for the [CLS] token)
+        image_size = (self.model_tester.image_size, self.model_tester.image_size)
+        patch_size = (self.model_tester.patch_size, self.model_tester.patch_size)
+        num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
+        seq_len = num_patches + 1
+
+        for model_class in self.all_model_classes:
+            inputs_dict["output_attentions"] = True
+            inputs_dict["output_hidden_states"] = False
+            config.return_dict = True
+            model = model_class(config)
+            model.to(torch_device)
+            model.eval()
+            with torch.no_grad():
+                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
+            attentions = outputs.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)
+
+            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))
+
+            added_hidden_states = 1
+            self.assertEqual(out_len + added_hidden_states, 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.encoder_hidden_states if config.is_encoder_decoder else outputs.hidden_states
+
+            expected_num_layers = getattr(
+                self.model_tester, "expected_num_hidden_layers", self.model_tester.num_hidden_layers + 1
+            )
+            self.assertEqual(len(hidden_states), expected_num_layers)
+
+            # CLIP has a different seq_length
+            image_size = (self.model_tester.image_size, self.model_tester.image_size)
+            patch_size = (self.model_tester.patch_size, self.model_tester.patch_size)
+            num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
+            seq_length = num_patches + 1
+
+            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)
+
+    def test_training(self):
+        pass
+
+    def test_training_gradient_checkpointing(self):
+        pass
+
+    # skip this test as CLIPVisionModel has no base class and is
+    # not available in MODEL_MAPPING
+    def test_save_load_fast_init_from_base(self):
+        pass
+
+    # skip this test as CLIPVisionModel has no base class and is
+    # not available in MODEL_MAPPING
+    def test_save_load_fast_init_to_base(self):
+        pass
+
+    @slow
+    def test_model_from_pretrained(self):
+        for model_name in CLIP_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
+            model = CLIPVisionModel.from_pretrained(model_name)
+            self.assertIsNotNone(model)
+
+
+class CLIPTextModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=12,
+        seq_length=7,
+        is_training=True,
+        use_input_mask=True,
+        use_labels=True,
+        vocab_size=99,
+        hidden_size=32,
+        num_hidden_layers=5,
+        num_attention_heads=4,
+        intermediate_size=37,
+        dropout=0.1,
+        attention_dropout=0.1,
+        max_position_embeddings=512,
+        initializer_range=0.02,
+        scope=None,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.seq_length = seq_length
+        self.is_training = is_training
+        self.use_input_mask = use_input_mask
+        self.use_labels = use_labels
+        self.vocab_size = vocab_size
+        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.dropout = dropout
+        self.attention_dropout = attention_dropout
+        self.max_position_embeddings = max_position_embeddings
+        self.initializer_range = initializer_range
+        self.scope = scope
+
+    def prepare_config_and_inputs(self):
+        input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
+
+        input_mask = None
+        if self.use_input_mask:
+            input_mask = random_attention_mask([self.batch_size, self.seq_length])
+
+        if input_mask is not None:
+            batch_size, seq_length = input_mask.shape
+            rnd_start_indices = np.random.randint(1, seq_length - 1, size=(batch_size,))
+            for batch_idx, start_index in enumerate(rnd_start_indices):
+                input_mask[batch_idx, :start_index] = 1
+                input_mask[batch_idx, start_index:] = 0
+
+        config = self.get_config()
+
+        return config, input_ids, input_mask
+
+    def get_config(self):
+        return CLIPTextConfig(
+            vocab_size=self.vocab_size,
+            hidden_size=self.hidden_size,
+            num_hidden_layers=self.num_hidden_layers,
+            num_attention_heads=self.num_attention_heads,
+            intermediate_size=self.intermediate_size,
+            dropout=self.dropout,
+            attention_dropout=self.attention_dropout,
+            max_position_embeddings=self.max_position_embeddings,
+            initializer_range=self.initializer_range,
+        )
+
+    def create_and_check_model(self, config, input_ids, input_mask):
+        model = CLIPTextModel(config=config)
+        model.to(torch_device)
+        model.eval()
+        with torch.no_grad():
+            result = model(input_ids, attention_mask=input_mask)
+            result = model(input_ids)
+        self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
+        self.parent.assertEqual(result.pooler_output.shape, (self.batch_size, self.hidden_size))
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        config, input_ids, input_mask = config_and_inputs
+        inputs_dict = {"input_ids": input_ids, "attention_mask": input_mask}
+        return config, inputs_dict
+
+
+@require_torch
+class CLIPTextModelTest(ModelTesterMixin, unittest.TestCase):
+
+    all_model_classes = (CLIPTextModel,) if is_torch_available() else ()
+    test_pruning = False
+    test_head_masking = False
+
+    def setUp(self):
+        self.model_tester = CLIPTextModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=CLIPTextConfig, hidden_size=37)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    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_training(self):
+        pass
+
+    def test_training_gradient_checkpointing(self):
+        pass
+
+    def test_inputs_embeds(self):
+        # CLIP does not use inputs_embeds
+        pass
+
+    # skip this test as CLIPTextModel has no base class and is
+    # not available in MODEL_MAPPING
+    def test_save_load_fast_init_from_base(self):
+        pass
+
+    # skip this test as CLIPTextModel has no base class and is
+    # not available in MODEL_MAPPING
+    def test_save_load_fast_init_to_base(self):
+        pass
+
+    @slow
+    def test_model_from_pretrained(self):
+        for model_name in CLIP_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
+            model = CLIPTextModel.from_pretrained(model_name)
+            self.assertIsNotNone(model)
+
+
+class CLIPModelTester:
+    def __init__(self, parent, is_training=True):
+        self.parent = parent
+        self.text_model_tester = CLIPTextModelTester(parent)
+        self.vision_model_tester = CLIPVisionModelTester(parent)
+        self.is_training = is_training
+
+    def prepare_config_and_inputs(self):
+        text_config, input_ids, attention_mask = self.text_model_tester.prepare_config_and_inputs()
+        vision_config, pixel_values = self.vision_model_tester.prepare_config_and_inputs()
+
+        config = self.get_config()
+
+        return config, input_ids, attention_mask, pixel_values
+
+    def get_config(self):
+        return CLIPConfig.from_text_vision_configs(
+            self.text_model_tester.get_config(), self.vision_model_tester.get_config(), projection_dim=64
+        )
+
+    def create_and_check_model(self, config, input_ids, attention_mask, pixel_values):
+        model = CLIPModel(config).to(torch_device).eval()
+        with torch.no_grad():
+            result = model(input_ids, pixel_values, attention_mask)
+        self.parent.assertEqual(
+            result.logits_per_image.shape, (self.vision_model_tester.batch_size, self.text_model_tester.batch_size)
+        )
+        self.parent.assertEqual(
+            result.logits_per_text.shape, (self.text_model_tester.batch_size, self.vision_model_tester.batch_size)
+        )
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        config, input_ids, attention_mask, pixel_values = config_and_inputs
+        inputs_dict = {
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+            "pixel_values": pixel_values,
+            "return_loss": True,
+        }
+        return config, inputs_dict
+
+
+@require_torch
+class CLIPModelTest(ModelTesterMixin, unittest.TestCase):
+    all_model_classes = (CLIPModel,) if is_torch_available() else ()
+    test_head_masking = False
+    test_pruning = False
+    test_resize_embeddings = False
+    test_attention_outputs = False
+
+    def setUp(self):
+        self.model_tester = CLIPModelTester(self)
+
+    def test_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_model(*config_and_inputs)
+
+    # hidden_states are tested in individual model tests
+    def test_hidden_states_output(self):
+        pass
+
+    # input_embeds are tested in individual model tests
+    def test_inputs_embeds(self):
+        pass
+
+    # tested in individual model tests
+    def test_retain_grad_hidden_states_attentions(self):
+        pass
+
+    # CLIPModel does not have input/output embeddings
+    def test_model_common_attributes(self):
+        pass
+
+    # override as the `logit_scale` parameter initilization is different for CLIP
+    def test_initialization(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        configs_no_init = _config_zero_init(config)
+        for model_class in self.all_model_classes:
+            model = model_class(config=configs_no_init)
+            for name, param in model.named_parameters():
+                if param.requires_grad:
+                    # check if `logit_scale` is initilized as per the original implementation
+                    if name == "logit_scale":
+                        self.assertAlmostEqual(
+                            param.data.item(),
+                            np.log(1 / 0.07),
+                            delta=1e-3,
+                            msg=f"Parameter {name} of model {model_class} seems not properly initialized",
+                        )
+                    else:
+                        self.assertIn(
+                            ((param.data.mean() * 1e9).round() / 1e9).item(),
+                            [0.0, 1.0],
+                            msg=f"Parameter {name} of model {model_class} seems not properly initialized",
+                        )
+
+    def _create_and_check_torchscript(self, config, inputs_dict):
+        if not self.test_torchscript:
+            return
+
+        configs_no_init = _config_zero_init(config)  # To be sure we have no Nan
+        configs_no_init.torchscript = True
+        configs_no_init.return_dict = False
+        for model_class in self.all_model_classes:
+            model = model_class(config=configs_no_init)
+            model.to(torch_device)
+            model.eval()
+
+            try:
+                input_ids = inputs_dict["input_ids"]
+                pixel_values = inputs_dict["pixel_values"]  # CLIP needs pixel_values
+                traced_model = torch.jit.trace(model, (input_ids, pixel_values))
+            except RuntimeError:
+                self.fail("Couldn't trace module.")
+
+            with tempfile.TemporaryDirectory() as tmp_dir_name:
+                pt_file_name = os.path.join(tmp_dir_name, "traced_model.pt")
+
+                try:
+                    torch.jit.save(traced_model, pt_file_name)
+                except Exception:
+                    self.fail("Couldn't save module.")
+
+                try:
+                    loaded_model = torch.jit.load(pt_file_name)
+                except Exception:
+                    self.fail("Couldn't load module.")
+
+            model.to(torch_device)
+            model.eval()
+
+            loaded_model.to(torch_device)
+            loaded_model.eval()
+
+            model_state_dict = model.state_dict()
+            loaded_model_state_dict = loaded_model.state_dict()
+
+            self.assertEqual(set(model_state_dict.keys()), set(loaded_model_state_dict.keys()))
+
+            models_equal = True
+            for layer_name, p1 in model_state_dict.items():
+                p2 = loaded_model_state_dict[layer_name]
+                if p1.data.ne(p2.data).sum() > 0:
+                    models_equal = False
+
+            self.assertTrue(models_equal)
+
+    def test_load_vision_text_config(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        # Save CLIPConfig and check if we can load CLIPVisionConfig from it
+        with tempfile.TemporaryDirectory() as tmp_dir_name:
+            config.save_pretrained(tmp_dir_name)
+            vision_config = CLIPVisionConfig.from_pretrained(tmp_dir_name)
+            self.assertDictEqual(config.vision_config.to_dict(), vision_config.to_dict())
+
+        # Save CLIPConfig and check if we can load CLIPTextConfig from it
+        with tempfile.TemporaryDirectory() as tmp_dir_name:
+            config.save_pretrained(tmp_dir_name)
+            text_config = CLIPTextConfig.from_pretrained(tmp_dir_name)
+            self.assertDictEqual(config.text_config.to_dict(), text_config.to_dict())
+
+    # overwrite from common since FlaxCLIPModel returns nested output
+    # which is not supported in the common test
+    @is_pt_flax_cross_test
+    def test_equivalence_pt_to_flax(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            with self.subTest(model_class.__name__):
+
+                # load PyTorch class
+                pt_model = model_class(config).eval()
+                # Flax models don't use the `use_cache` option and cache is not returned as a default.
+                # So we disable `use_cache` here for PyTorch model.
+                pt_model.config.use_cache = False
+
+                fx_model_class_name = "Flax" + model_class.__name__
+
+                if not hasattr(transformers, fx_model_class_name):
+                    return
+
+                fx_model_class = getattr(transformers, fx_model_class_name)
+
+                # load Flax class
+                fx_model = fx_model_class(config, dtype=jnp.float32)
+                # make sure only flax inputs are forward that actually exist in function args
+                fx_input_keys = inspect.signature(fx_model.__call__).parameters.keys()
+
+                # prepare inputs
+                pt_inputs = self._prepare_for_class(inputs_dict, model_class)
+
+                # remove function args that don't exist in Flax
+                pt_inputs = {k: v for k, v in pt_inputs.items() if k in fx_input_keys}
+
+                fx_state = convert_pytorch_state_dict_to_flax(pt_model.state_dict(), fx_model)
+                fx_model.params = fx_state
+
+                with torch.no_grad():
+                    pt_outputs = pt_model(**pt_inputs).to_tuple()
+
+                # convert inputs to Flax
+                fx_inputs = {k: np.array(v) for k, v in pt_inputs.items() if torch.is_tensor(v)}
+                fx_outputs = fx_model(**fx_inputs).to_tuple()
+                self.assertEqual(len(fx_outputs), len(pt_outputs), "Output lengths differ between Flax and PyTorch")
+                for fx_output, pt_output in zip(fx_outputs[:4], pt_outputs[:4]):
+                    self.assert_almost_equals(fx_output, pt_output.numpy(), 4e-2)
+
+                with tempfile.TemporaryDirectory() as tmpdirname:
+                    pt_model.save_pretrained(tmpdirname)
+                    fx_model_loaded = fx_model_class.from_pretrained(tmpdirname, from_pt=True)
+
+                fx_outputs_loaded = fx_model_loaded(**fx_inputs).to_tuple()
+                self.assertEqual(
+                    len(fx_outputs_loaded), len(pt_outputs), "Output lengths differ between Flax and PyTorch"
+                )
+                for fx_output_loaded, pt_output in zip(fx_outputs_loaded[:4], pt_outputs[:4]):
+                    self.assert_almost_equals(fx_output_loaded, pt_output.numpy(), 4e-2)
+
+    # overwrite from common since FlaxCLIPModel returns nested output
+    # which is not supported in the common test
+    @is_pt_flax_cross_test
+    def test_equivalence_flax_to_pt(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            with self.subTest(model_class.__name__):
+                # load corresponding PyTorch class
+                pt_model = model_class(config).eval()
+
+                # So we disable `use_cache` here for PyTorch model.
+                pt_model.config.use_cache = False
+
+                fx_model_class_name = "Flax" + model_class.__name__
+
+                if not hasattr(transformers, fx_model_class_name):
+                    # no flax model exists for this class
+                    return
+
+                fx_model_class = getattr(transformers, fx_model_class_name)
+
+                # load Flax class
+                fx_model = fx_model_class(config, dtype=jnp.float32)
+                # make sure only flax inputs are forward that actually exist in function args
+                fx_input_keys = inspect.signature(fx_model.__call__).parameters.keys()
+
+                pt_model = load_flax_weights_in_pytorch_model(pt_model, fx_model.params)
+
+                # make sure weights are tied in PyTorch
+                pt_model.tie_weights()
+
+                # prepare inputs
+                pt_inputs = self._prepare_for_class(inputs_dict, model_class)
+
+                # remove function args that don't exist in Flax
+                pt_inputs = {k: v for k, v in pt_inputs.items() if k in fx_input_keys}
+
+                with torch.no_grad():
+                    pt_outputs = pt_model(**pt_inputs).to_tuple()
+
+                fx_inputs = {k: np.array(v) for k, v in pt_inputs.items() if torch.is_tensor(v)}
+
+                fx_outputs = fx_model(**fx_inputs).to_tuple()
+                self.assertEqual(len(fx_outputs), len(pt_outputs), "Output lengths differ between Flax and PyTorch")
+
+                for fx_output, pt_output in zip(fx_outputs[:4], pt_outputs[:4]):
+                    self.assert_almost_equals(fx_output, pt_output.numpy(), 4e-2)
+
+                with tempfile.TemporaryDirectory() as tmpdirname:
+                    fx_model.save_pretrained(tmpdirname)
+                    pt_model_loaded = model_class.from_pretrained(tmpdirname, from_flax=True)
+
+                with torch.no_grad():
+                    pt_outputs_loaded = pt_model_loaded(**pt_inputs).to_tuple()
+
+                self.assertEqual(
+                    len(fx_outputs), len(pt_outputs_loaded), "Output lengths differ between Flax and PyTorch"
+                )
+                for fx_output, pt_output in zip(fx_outputs[:4], pt_outputs_loaded[:4]):
+                    self.assert_almost_equals(fx_output, pt_output.numpy(), 4e-2)
+
+    @slow
+    def test_model_from_pretrained(self):
+        for model_name in CLIP_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
+            model = CLIPModel.from_pretrained(model_name)
+            self.assertIsNotNone(model)
+
+
+# We will verify our results on an image of cute cats
+def prepare_img():
+    url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+    im = Image.open(requests.get(url, stream=True).raw)
+    return im
+
+
+@require_vision
+@require_torch
+class CLIPModelIntegrationTest(unittest.TestCase):
+    @slow
+    def test_inference(self):
+        model_name = "openai/clip-vit-base-patch32"
+        model = CLIPModel.from_pretrained(model_name).to(torch_device)
+        processor = CLIPProcessor.from_pretrained(model_name)
+
+        image = prepare_img()
+        inputs = processor(
+            text=["a photo of a cat", "a photo of a dog"], images=image, padding=True, return_tensors="pt"
+        ).to(torch_device)
+
+        # forward pass
+        with torch.no_grad():
+            outputs = model(**inputs)
+
+        # verify the logits
+        self.assertEqual(
+            outputs.logits_per_image.shape,
+            torch.Size((inputs.pixel_values.shape[0], inputs.input_ids.shape[0])),
+        )
+        self.assertEqual(
+            outputs.logits_per_text.shape,
+            torch.Size((inputs.input_ids.shape[0], inputs.pixel_values.shape[0])),
+        )
+
+        expected_logits = torch.tensor([[24.5701, 19.3049]], device=torch_device)
+
+        self.assertTrue(torch.allclose(outputs.logits_per_image, expected_logits, atol=1e-3))

tests/models/clip/test_modeling_flax_clip.py

@@ -0,0 +1,587 @@
+import inspect
+import tempfile
+import unittest
+
+import numpy as np
+
+import transformers
+from transformers import CLIPConfig, CLIPTextConfig, CLIPVisionConfig, is_flax_available, is_torch_available
+from transformers.testing_utils import is_pt_flax_cross_test, require_flax, slow
+
+from ...test_modeling_flax_common import FlaxModelTesterMixin, floats_tensor, ids_tensor, random_attention_mask
+
+
+if is_flax_available():
+    import jax
+    import jax.numpy as jnp
+    from transformers.modeling_flax_pytorch_utils import (
+        convert_pytorch_state_dict_to_flax,
+        load_flax_weights_in_pytorch_model,
+    )
+    from transformers.models.clip.modeling_flax_clip import FlaxCLIPModel, FlaxCLIPTextModel, FlaxCLIPVisionModel
+
+if is_torch_available():
+    import torch
+
+
+class FlaxCLIPVisionModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=12,
+        image_size=30,
+        patch_size=2,
+        num_channels=3,
+        is_training=True,
+        hidden_size=32,
+        num_hidden_layers=5,
+        num_attention_heads=4,
+        intermediate_size=37,
+        dropout=0.1,
+        attention_dropout=0.1,
+        initializer_range=0.02,
+        scope=None,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.image_size = image_size
+        self.patch_size = patch_size
+        self.num_channels = num_channels
+        self.is_training = is_training
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.intermediate_size = intermediate_size
+        self.dropout = dropout
+        self.attention_dropout = attention_dropout
+        self.initializer_range = initializer_range
+        self.scope = scope
+
+    def prepare_config_and_inputs(self):
+        pixel_values = floats_tensor([self.batch_size, self.num_channels, self.image_size, self.image_size])
+        config = CLIPVisionConfig(
+            image_size=self.image_size,
+            patch_size=self.patch_size,
+            num_channels=self.num_channels,
+            hidden_size=self.hidden_size,
+            num_hidden_layers=self.num_hidden_layers,
+            num_attention_heads=self.num_attention_heads,
+            intermediate_size=self.intermediate_size,
+            dropout=self.dropout,
+            attention_dropout=self.attention_dropout,
+            initializer_range=self.initializer_range,
+        )
+
+        return config, pixel_values
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        config, pixel_values = config_and_inputs
+        inputs_dict = {"pixel_values": pixel_values}
+        return config, inputs_dict
+
+
+@require_flax
+class FlaxCLIPVisionModelTest(FlaxModelTesterMixin, unittest.TestCase):
+    """
+    Here we also overwrite some of the tests of test_modeling_common.py, as CLIP does not use input_ids, inputs_embeds,
+    attention_mask and seq_length.
+    """
+
+    all_model_classes = (FlaxCLIPVisionModel,) if is_flax_available() else ()
+
+    def setUp(self):
+        self.model_tester = FlaxCLIPVisionModelTester(self)
+
+    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.__call__)
+            # 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_jit_compilation(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            with self.subTest(model_class.__name__):
+                prepared_inputs_dict = self._prepare_for_class(inputs_dict, model_class)
+                model = model_class(config)
+
+                @jax.jit
+                def model_jitted(pixel_values, **kwargs):
+                    return model(pixel_values=pixel_values, **kwargs).to_tuple()
+
+                with self.subTest("JIT Enabled"):
+                    jitted_outputs = model_jitted(**prepared_inputs_dict)
+
+                with self.subTest("JIT Disabled"):
+                    with jax.disable_jit():
+                        outputs = model_jitted(**prepared_inputs_dict)
+
+                self.assertEqual(len(outputs), len(jitted_outputs))
+                for jitted_output, output in zip(jitted_outputs, outputs):
+                    self.assertEqual(jitted_output.shape, output.shape)
+
+    def test_hidden_states_output(self):
+        def check_hidden_states_output(inputs_dict, config, model_class):
+            model = model_class(config)
+
+            outputs = model(**self._prepare_for_class(inputs_dict, model_class))
+            hidden_states = outputs.hidden_states
+
+            self.assertEqual(len(hidden_states), self.model_tester.num_hidden_layers + 1)
+
+            # CLIP has a different seq_length
+            image_size = (self.model_tester.image_size, self.model_tester.image_size)
+            patch_size = (self.model_tester.patch_size, self.model_tester.patch_size)
+            num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
+            seq_length = num_patches + 1
+
+            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)
+
+    def test_attention_outputs(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+        config.return_dict = True
+
+        # in CLIP, the seq_len equals the number of patches + 1 (we add 1 for the [CLS] token)
+        image_size = (self.model_tester.image_size, self.model_tester.image_size)
+        patch_size = (self.model_tester.patch_size, self.model_tester.patch_size)
+        num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
+        seq_length = num_patches + 1
+
+        for model_class in self.all_model_classes:
+            inputs_dict["output_attentions"] = True
+            inputs_dict["output_hidden_states"] = False
+            model = model_class(config)
+            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)
+            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_length, seq_length],
+            )
+            out_len = len(outputs)
+
+            # Check attention is always last and order is fine
+            inputs_dict["output_attentions"] = True
+            inputs_dict["output_hidden_states"] = True
+            model = model_class(config)
+            outputs = model(**self._prepare_for_class(inputs_dict, model_class))
+
+            added_hidden_states = 1
+            self.assertEqual(out_len + added_hidden_states, len(outputs))
+
+            self_attentions = outputs.encoder_attentions if config.is_encoder_decoder else outputs.attentions
+            self.assertEqual(len(self_attentions), self.model_tester.num_hidden_layers)
+
+            self.assertListEqual(
+                list(self_attentions[0].shape[-3:]),
+                [self.model_tester.num_attention_heads, seq_length, seq_length],
+            )
+
+    # FlaxCLIPVisionModel does not have any base model
+    def test_save_load_from_base(self):
+        pass
+
+    # FlaxCLIPVisionModel does not have any base model
+    def test_save_load_to_base(self):
+        pass
+
+    # FlaxCLIPVisionModel does not have any base model
+    @is_pt_flax_cross_test
+    def test_save_load_from_base_pt(self):
+        pass
+
+    # FlaxCLIPVisionModel does not have any base model
+    @is_pt_flax_cross_test
+    def test_save_load_to_base_pt(self):
+        pass
+
+    # FlaxCLIPVisionModel does not have any base model
+    @is_pt_flax_cross_test
+    def test_save_load_bf16_to_base_pt(self):
+        pass
+
+    @slow
+    def test_model_from_pretrained(self):
+        for model_class_name in self.all_model_classes:
+            model = model_class_name.from_pretrained("openai/clip-vit-base-patch32", from_pt=True)
+            outputs = model(np.ones((1, 3, 224, 224)))
+            self.assertIsNotNone(outputs)
+
+
+class FlaxCLIPTextModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=12,
+        seq_length=7,
+        is_training=True,
+        use_input_mask=True,
+        use_labels=True,
+        vocab_size=99,
+        hidden_size=32,
+        num_hidden_layers=5,
+        num_attention_heads=4,
+        intermediate_size=37,
+        dropout=0.1,
+        attention_dropout=0.1,
+        max_position_embeddings=512,
+        initializer_range=0.02,
+        scope=None,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.seq_length = seq_length
+        self.is_training = is_training
+        self.use_input_mask = use_input_mask
+        self.use_labels = use_labels
+        self.vocab_size = vocab_size
+        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.dropout = dropout
+        self.attention_dropout = attention_dropout
+        self.max_position_embeddings = max_position_embeddings
+        self.initializer_range = initializer_range
+        self.scope = scope
+
+    def prepare_config_and_inputs(self):
+        input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
+
+        input_mask = None
+        if self.use_input_mask:
+            input_mask = random_attention_mask([self.batch_size, self.seq_length])
+
+        if input_mask is not None:
+            batch_size, seq_length = input_mask.shape
+            rnd_start_indices = np.random.randint(1, seq_length - 1, size=(batch_size,))
+            for batch_idx, start_index in enumerate(rnd_start_indices):
+                input_mask[batch_idx, :start_index] = 1
+                input_mask[batch_idx, start_index:] = 0
+
+        config = CLIPTextConfig(
+            vocab_size=self.vocab_size,
+            hidden_size=self.hidden_size,
+            num_hidden_layers=self.num_hidden_layers,
+            num_attention_heads=self.num_attention_heads,
+            intermediate_size=self.intermediate_size,
+            dropout=self.dropout,
+            attention_dropout=self.attention_dropout,
+            max_position_embeddings=self.max_position_embeddings,
+            initializer_range=self.initializer_range,
+        )
+
+        return config, input_ids, input_mask
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        config, input_ids, input_mask = config_and_inputs
+        inputs_dict = {"input_ids": input_ids, "attention_mask": input_mask}
+        return config, inputs_dict
+
+
+@require_flax
+class FlaxCLIPTextModelTest(FlaxModelTesterMixin, unittest.TestCase):
+    all_model_classes = (FlaxCLIPTextModel,) if is_flax_available() else ()
+
+    def setUp(self):
+        self.model_tester = FlaxCLIPTextModelTester(self)
+
+    # FlaxCLIPTextModel does not have any base model
+    def test_save_load_from_base(self):
+        pass
+
+    # FlaxCLIPVisionModel does not have any base model
+    def test_save_load_to_base(self):
+        pass
+
+    # FlaxCLIPVisionModel does not have any base model
+    @is_pt_flax_cross_test
+    def test_save_load_from_base_pt(self):
+        pass
+
+    # FlaxCLIPVisionModel does not have any base model
+    @is_pt_flax_cross_test
+    def test_save_load_to_base_pt(self):
+        pass
+
+    # FlaxCLIPVisionModel does not have any base model
+    @is_pt_flax_cross_test
+    def test_save_load_bf16_to_base_pt(self):
+        pass
+
+    @slow
+    def test_model_from_pretrained(self):
+        for model_class_name in self.all_model_classes:
+            model = model_class_name.from_pretrained("openai/clip-vit-base-patch32", from_pt=True)
+            outputs = model(np.ones((1, 1)))
+            self.assertIsNotNone(outputs)
+
+
+class FlaxCLIPModelTester:
+    def __init__(self, parent, is_training=True):
+        self.parent = parent
+        self.text_model_tester = FlaxCLIPTextModelTester(parent)
+        self.vision_model_tester = FlaxCLIPVisionModelTester(parent)
+        self.is_training = is_training
+
+    def prepare_config_and_inputs(self):
+        text_config, input_ids, attention_mask = self.text_model_tester.prepare_config_and_inputs()
+        vision_config, pixel_values = self.vision_model_tester.prepare_config_and_inputs()
+
+        config = CLIPConfig.from_text_vision_configs(text_config, vision_config, projection_dim=64)
+
+        return config, input_ids, attention_mask, pixel_values
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        config, input_ids, attention_mask, pixel_values = config_and_inputs
+        inputs_dict = {
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+            "pixel_values": pixel_values,
+        }
+        return config, inputs_dict
+
+
+@require_flax
+class FlaxCLIPModelTest(FlaxModelTesterMixin, unittest.TestCase):
+    all_model_classes = (FlaxCLIPModel,) if is_flax_available() else ()
+    test_attention_outputs = False
+
+    def setUp(self):
+        self.model_tester = FlaxCLIPModelTester(self)
+
+    # hidden_states are tested in individual model tests
+    def test_hidden_states_output(self):
+        pass
+
+    def test_jit_compilation(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            with self.subTest(model_class.__name__):
+                prepared_inputs_dict = self._prepare_for_class(inputs_dict, model_class)
+                model = model_class(config)
+
+                @jax.jit
+                def model_jitted(input_ids, pixel_values, **kwargs):
+                    return model(input_ids=input_ids, pixel_values=pixel_values, **kwargs).to_tuple()
+
+                with self.subTest("JIT Enabled"):
+                    jitted_outputs = model_jitted(**prepared_inputs_dict)
+
+                with self.subTest("JIT Disabled"):
+                    with jax.disable_jit():
+                        outputs = model_jitted(**prepared_inputs_dict)
+
+                self.assertEqual(len(outputs), len(jitted_outputs))
+                for jitted_output, output in zip(jitted_outputs[:4], outputs[:4]):
+                    self.assertEqual(jitted_output.shape, output.shape)
+
+    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.__call__)
+            # signature.parameters is an OrderedDict => so arg_names order is deterministic
+            arg_names = [*signature.parameters.keys()]
+
+            expected_arg_names = ["input_ids", "pixel_values", "attention_mask", "position_ids"]
+            self.assertListEqual(arg_names[:4], expected_arg_names)
+
+    def test_get_image_features(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+        model = FlaxCLIPModel(config)
+
+        @jax.jit
+        def model_jitted(pixel_values):
+            return model.get_image_features(pixel_values=pixel_values)
+
+        with self.subTest("JIT Enabled"):
+            jitted_output = model_jitted(inputs_dict["pixel_values"])
+
+        with self.subTest("JIT Disabled"):
+            with jax.disable_jit():
+                output = model_jitted(inputs_dict["pixel_values"])
+
+        self.assertEqual(jitted_output.shape, output.shape)
+        self.assertTrue(np.allclose(jitted_output, output, atol=1e-3))
+
+    def test_get_text_features(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+        model = FlaxCLIPModel(config)
+
+        @jax.jit
+        def model_jitted(input_ids, attention_mask, **kwargs):
+            return model.get_text_features(input_ids=input_ids, attention_mask=attention_mask)
+
+        with self.subTest("JIT Enabled"):
+            jitted_output = model_jitted(**inputs_dict)
+
+        with self.subTest("JIT Disabled"):
+            with jax.disable_jit():
+                output = model_jitted(**inputs_dict)
+
+        self.assertEqual(jitted_output.shape, output.shape)
+        self.assertTrue(np.allclose(jitted_output, output, atol=1e-3))
+
+    @slow
+    def test_model_from_pretrained(self):
+        for model_class_name in self.all_model_classes:
+            model = model_class_name.from_pretrained("openai/clip-vit-base-patch32", from_pt=True)
+            outputs = model(input_ids=np.ones((1, 1)), pixel_values=np.ones((1, 3, 224, 224)))
+            self.assertIsNotNone(outputs)
+
+    # overwrite from common since FlaxCLIPModel returns nested output
+    # which is not supported in the common test
+    @is_pt_flax_cross_test
+    def test_equivalence_pt_to_flax(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            with self.subTest(model_class.__name__):
+                # prepare inputs
+                prepared_inputs_dict = self._prepare_for_class(inputs_dict, model_class)
+                pt_inputs = {k: torch.tensor(v.tolist()) for k, v in prepared_inputs_dict.items()}
+
+                # load corresponding PyTorch class
+                pt_model_class_name = model_class.__name__[4:]  # Skip the "Flax" at the beginning
+                pt_model_class = getattr(transformers, pt_model_class_name)
+
+                pt_model = pt_model_class(config).eval()
+                fx_model = model_class(config, dtype=jnp.float32)
+
+                fx_state = convert_pytorch_state_dict_to_flax(pt_model.state_dict(), fx_model)
+                fx_model.params = fx_state
+
+                with torch.no_grad():
+                    pt_outputs = pt_model(**pt_inputs).to_tuple()
+
+                fx_outputs = fx_model(**prepared_inputs_dict).to_tuple()
+                self.assertEqual(len(fx_outputs), len(pt_outputs), "Output lengths differ between Flax and PyTorch")
+                for fx_output, pt_output in zip(fx_outputs[:4], pt_outputs[:4]):
+                    self.assert_almost_equals(fx_output, pt_output.numpy(), 4e-2)
+
+                with tempfile.TemporaryDirectory() as tmpdirname:
+                    pt_model.save_pretrained(tmpdirname)
+                    fx_model_loaded = model_class.from_pretrained(tmpdirname, from_pt=True)
+
+                fx_outputs_loaded = fx_model_loaded(**prepared_inputs_dict).to_tuple()
+                self.assertEqual(
+                    len(fx_outputs_loaded), len(pt_outputs), "Output lengths differ between Flax and PyTorch"
+                )
+                for fx_output_loaded, pt_output in zip(fx_outputs_loaded[:4], pt_outputs[:4]):
+                    self.assert_almost_equals(fx_output_loaded, pt_output.numpy(), 4e-2)
+
+    # overwrite from common since FlaxCLIPModel returns nested output
+    # which is not supported in the common test
+    @is_pt_flax_cross_test
+    def test_equivalence_flax_to_pt(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            with self.subTest(model_class.__name__):
+                # prepare inputs
+                prepared_inputs_dict = self._prepare_for_class(inputs_dict, model_class)
+                pt_inputs = {k: torch.tensor(v.tolist()) for k, v in prepared_inputs_dict.items()}
+
+                # load corresponding PyTorch class
+                pt_model_class_name = model_class.__name__[4:]  # Skip the "Flax" at the beginning
+                pt_model_class = getattr(transformers, pt_model_class_name)
+
+                pt_model = pt_model_class(config).eval()
+                fx_model = model_class(config, dtype=jnp.float32)
+
+                pt_model = load_flax_weights_in_pytorch_model(pt_model, fx_model.params)
+
+                # make sure weights are tied in PyTorch
+                pt_model.tie_weights()
+
+                with torch.no_grad():
+                    pt_outputs = pt_model(**pt_inputs).to_tuple()
+
+                fx_outputs = fx_model(**prepared_inputs_dict).to_tuple()
+                self.assertEqual(len(fx_outputs), len(pt_outputs), "Output lengths differ between Flax and PyTorch")
+                for fx_output, pt_output in zip(fx_outputs[:4], pt_outputs[:4]):
+                    self.assert_almost_equals(fx_output, pt_output.numpy(), 4e-2)
+
+                with tempfile.TemporaryDirectory() as tmpdirname:
+                    fx_model.save_pretrained(tmpdirname)
+                    pt_model_loaded = pt_model_class.from_pretrained(tmpdirname, from_flax=True)
+
+                with torch.no_grad():
+                    pt_outputs_loaded = pt_model_loaded(**pt_inputs).to_tuple()
+
+                self.assertEqual(
+                    len(fx_outputs), len(pt_outputs_loaded), "Output lengths differ between Flax and PyTorch"
+                )
+                for fx_output, pt_output in zip(fx_outputs[:4], pt_outputs_loaded[:4]):
+                    self.assert_almost_equals(fx_output, pt_output.numpy(), 4e-2)
+
+    # overwrite from common since FlaxCLIPModel returns nested output
+    # which is not supported in the common test
+    def test_from_pretrained_save_pretrained(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            if model_class.__name__ != "FlaxBertModel":
+                continue
+
+            with self.subTest(model_class.__name__):
+                model = model_class(config)
+
+                prepared_inputs_dict = self._prepare_for_class(inputs_dict, model_class)
+                outputs = model(**prepared_inputs_dict).to_tuple()
+
+                # verify that normal save_pretrained works as expected
+                with tempfile.TemporaryDirectory() as tmpdirname:
+                    model.save_pretrained(tmpdirname)
+                    model_loaded = model_class.from_pretrained(tmpdirname)
+
+                outputs_loaded = model_loaded(**prepared_inputs_dict).to_tuple()[:4]
+                for output_loaded, output in zip(outputs_loaded, outputs):
+                    self.assert_almost_equals(output_loaded, output, 1e-3)
+
+                # verify that save_pretrained for distributed training
+                # with `params=params` works as expected
+                with tempfile.TemporaryDirectory() as tmpdirname:
+                    model.save_pretrained(tmpdirname, params=model.params)
+                    model_loaded = model_class.from_pretrained(tmpdirname)
+
+                outputs_loaded = model_loaded(**prepared_inputs_dict).to_tuple()[:4]
+                for output_loaded, output in zip(outputs_loaded, outputs):
+                    self.assert_almost_equals(output_loaded, output, 1e-3)

tests/models/clip/test_modeling_tf_clip.py

@@ -0,0 +1,541 @@
+# coding=utf-8
+# Copyright 2021 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Testing suite for the TensorFlow CLIP model. """
+
+
+import inspect
+import os
+import tempfile
+import unittest
+from importlib import import_module
+
+import requests
+from transformers import CLIPConfig, CLIPTextConfig, CLIPVisionConfig
+from transformers.testing_utils import require_tf, require_vision, slow
+from transformers.utils import is_tf_available, is_vision_available
+
+from ...test_configuration_common import ConfigTester
+from ...test_modeling_tf_common import TFModelTesterMixin, floats_tensor, ids_tensor, random_attention_mask
+
+
+if is_tf_available():
+    import tensorflow as tf
+
+    from transformers import TFCLIPModel, TFCLIPTextModel, TFCLIPVisionModel, TFSharedEmbeddings
+    from transformers.models.clip.modeling_tf_clip import TF_CLIP_PRETRAINED_MODEL_ARCHIVE_LIST
+
+
+if is_vision_available():
+    from PIL import Image
+
+    from transformers import CLIPProcessor
+
+
+class TFCLIPVisionModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=12,
+        image_size=30,
+        patch_size=2,
+        num_channels=3,
+        is_training=True,
+        hidden_size=32,
+        num_hidden_layers=5,
+        num_attention_heads=4,
+        intermediate_size=37,
+        dropout=0.1,
+        attention_dropout=0.1,
+        initializer_range=0.02,
+        scope=None,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.image_size = image_size
+        self.patch_size = patch_size
+        self.num_channels = num_channels
+        self.is_training = is_training
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.intermediate_size = intermediate_size
+        self.dropout = dropout
+        self.attention_dropout = attention_dropout
+        self.initializer_range = initializer_range
+        self.scope = scope
+
+    def prepare_config_and_inputs(self):
+        pixel_values = floats_tensor([self.batch_size, self.num_channels, self.image_size, self.image_size])
+        config = self.get_config()
+
+        return config, pixel_values
+
+    def get_config(self):
+        return CLIPVisionConfig(
+            image_size=self.image_size,
+            patch_size=self.patch_size,
+            num_channels=self.num_channels,
+            hidden_size=self.hidden_size,
+            num_hidden_layers=self.num_hidden_layers,
+            num_attention_heads=self.num_attention_heads,
+            intermediate_size=self.intermediate_size,
+            dropout=self.dropout,
+            attention_dropout=self.attention_dropout,
+            initializer_range=self.initializer_range,
+        )
+
+    def create_and_check_model(self, config, pixel_values):
+        model = TFCLIPVisionModel(config=config)
+        result = model(pixel_values, training=False)
+        # expected sequence length = num_patches + 1 (we add 1 for the [CLS] token)
+        image_size = (self.image_size, self.image_size)
+        patch_size = (self.patch_size, self.patch_size)
+        num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
+        self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, num_patches + 1, self.hidden_size))
+        self.parent.assertEqual(result.pooler_output.shape, (self.batch_size, self.hidden_size))
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        config, pixel_values = config_and_inputs
+        inputs_dict = {"pixel_values": pixel_values}
+        return config, inputs_dict
+
+
+@require_tf
+class TFCLIPVisionModelTest(TFModelTesterMixin, unittest.TestCase):
+    """
+    Here we also overwrite some of the tests of test_modeling_common.py, as CLIP does not use input_ids, inputs_embeds,
+    attention_mask and seq_length.
+    """
+
+    all_model_classes = (TFCLIPVisionModel,) if is_tf_available() else ()
+
+    test_pruning = False
+    test_resize_embeddings = False
+    test_head_masking = False
+    test_onnx = False
+
+    def setUp(self):
+        self.model_tester = TFCLIPVisionModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=CLIPVisionConfig, has_text_modality=False, hidden_size=37)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_inputs_embeds(self):
+        # CLIP does not use inputs_embeds
+        pass
+
+    def test_graph_mode_with_inputs_embeds(self):
+        # CLIP does not use inputs_embeds
+        pass
+
+    def test_model_common_attributes(self):
+        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            model = model_class(config)
+            self.assertIsInstance(model.get_input_embeddings(), (tf.keras.layers.Layer))
+            x = model.get_output_embeddings()
+            self.assertTrue(x is None or isinstance(x, tf.keras.layers.Layer))
+
+    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.call)
+            # signature.parameters is an OrderedDict => so arg_names order is deterministic
+            arg_names = [*signature.parameters.keys()]
+
+            expected_arg_names = ["pixel_values"]
+            self.assertListEqual(arg_names[:1], expected_arg_names)
+
+    def test_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_model(*config_and_inputs)
+
+    def test_attention_outputs(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+        config.return_dict = True
+
+        # in CLIP, the seq_len equals the number of patches + 1 (we add 1 for the [CLS] token)
+        image_size = (self.model_tester.image_size, self.model_tester.image_size)
+        patch_size = (self.model_tester.patch_size, self.model_tester.patch_size)
+        num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
+        seq_len = num_patches + 1
+
+        for model_class in self.all_model_classes:
+            inputs_dict["output_attentions"] = True
+            inputs_dict["output_hidden_states"] = False
+            config.return_dict = True
+            model = model_class(config)
+            outputs = model(**self._prepare_for_class(inputs_dict, model_class), training=False)
+            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)
+            outputs = model(**self._prepare_for_class(inputs_dict, model_class), training=False)
+            attentions = outputs.attentions
+            self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)
+
+            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)
+            outputs = model(**self._prepare_for_class(inputs_dict, model_class), training=False)
+
+            added_hidden_states = 1
+            self.assertEqual(out_len + added_hidden_states, 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)
+
+            outputs = model(**self._prepare_for_class(inputs_dict, model_class), training=False)
+
+            hidden_states = outputs.encoder_hidden_states if config.is_encoder_decoder else outputs.hidden_states
+
+            expected_num_layers = getattr(
+                self.model_tester, "expected_num_hidden_layers", self.model_tester.num_hidden_layers + 1
+            )
+            self.assertEqual(len(hidden_states), expected_num_layers)
+
+            # CLIP has a different seq_length
+            image_size = (self.model_tester.image_size, self.model_tester.image_size)
+            patch_size = (self.model_tester.patch_size, self.model_tester.patch_size)
+            num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
+            seq_length = num_patches + 1
+
+            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)
+
+    @slow
+    def test_model_from_pretrained(self):
+        for model_name in TF_CLIP_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
+            model = TFCLIPVisionModel.from_pretrained(model_name)
+            self.assertIsNotNone(model)
+
+
+class TFCLIPTextModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=12,
+        seq_length=7,
+        is_training=True,
+        use_input_mask=True,
+        use_labels=True,
+        vocab_size=99,
+        hidden_size=32,
+        num_hidden_layers=5,
+        num_attention_heads=4,
+        intermediate_size=37,
+        dropout=0.1,
+        attention_dropout=0.1,
+        max_position_embeddings=512,
+        initializer_range=0.02,
+        scope=None,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.seq_length = seq_length
+        self.is_training = is_training
+        self.use_input_mask = use_input_mask
+        self.use_labels = use_labels
+        self.vocab_size = vocab_size
+        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.dropout = dropout
+        self.attention_dropout = attention_dropout
+        self.max_position_embeddings = max_position_embeddings
+        self.initializer_range = initializer_range
+        self.scope = scope
+
+    def prepare_config_and_inputs(self):
+        input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
+
+        input_mask = None
+        if self.use_input_mask:
+            input_mask = random_attention_mask([self.batch_size, self.seq_length])
+            # make sure the first token has attention mask `1` to ensure that, after combining the causal mask, there
+            # is still at least one token being attended to for each batch.
+            # TODO: Change `random_attention_mask` in PT/TF/Flax common test file, after a discussion with the team.
+            input_mask = tf.concat(
+                [tf.ones_like(input_mask[:, :1], dtype=input_mask.dtype), input_mask[:, 1:]], axis=-1
+            )
+
+        config = self.get_config()
+
+        return config, input_ids, input_mask
+
+    def get_config(self):
+        return CLIPTextConfig(
+            vocab_size=self.vocab_size,
+            hidden_size=self.hidden_size,
+            num_hidden_layers=self.num_hidden_layers,
+            num_attention_heads=self.num_attention_heads,
+            intermediate_size=self.intermediate_size,
+            dropout=self.dropout,
+            attention_dropout=self.attention_dropout,
+            max_position_embeddings=self.max_position_embeddings,
+            initializer_range=self.initializer_range,
+        )
+
+    def create_and_check_model(self, config, input_ids, input_mask):
+        model = TFCLIPTextModel(config=config)
+        result = model(input_ids, attention_mask=input_mask, training=False)
+        result = model(input_ids, training=False)
+        self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
+        self.parent.assertEqual(result.pooler_output.shape, (self.batch_size, self.hidden_size))
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        config, input_ids, input_mask = config_and_inputs
+        inputs_dict = {"input_ids": input_ids, "attention_mask": input_mask}
+        return config, inputs_dict
+
+
+@require_tf
+class TFCLIPTextModelTest(TFModelTesterMixin, unittest.TestCase):
+
+    all_model_classes = (TFCLIPTextModel,) if is_tf_available() else ()
+    test_pruning = False
+    test_head_masking = False
+    test_onnx = False
+
+    def setUp(self):
+        self.model_tester = TFCLIPTextModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=CLIPTextConfig, hidden_size=37)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    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_inputs_embeds(self):
+        # CLIP does not use inputs_embeds
+        pass
+
+    @slow
+    def test_model_from_pretrained(self):
+        for model_name in TF_CLIP_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
+            model = TFCLIPTextModel.from_pretrained(model_name)
+            self.assertIsNotNone(model)
+
+
+class TFCLIPModelTester:
+    def __init__(self, parent, is_training=True):
+        self.parent = parent
+        self.text_model_tester = TFCLIPTextModelTester(parent)
+        self.vision_model_tester = TFCLIPVisionModelTester(parent)
+        self.is_training = is_training
+
+    def prepare_config_and_inputs(self):
+        text_config, input_ids, attention_mask = self.text_model_tester.prepare_config_and_inputs()
+        vision_config, pixel_values = self.vision_model_tester.prepare_config_and_inputs()
+
+        config = self.get_config()
+
+        return config, input_ids, attention_mask, pixel_values
+
+    def get_config(self):
+        return CLIPConfig.from_text_vision_configs(
+            self.text_model_tester.get_config(), self.vision_model_tester.get_config(), projection_dim=64
+        )
+
+    def create_and_check_model(self, config, input_ids, attention_mask, pixel_values):
+        model = TFCLIPModel(config)
+        result = model(input_ids, pixel_values, attention_mask, training=False)
+        self.parent.assertEqual(
+            result.logits_per_image.shape, (self.vision_model_tester.batch_size, self.text_model_tester.batch_size)
+        )
+        self.parent.assertEqual(
+            result.logits_per_text.shape, (self.text_model_tester.batch_size, self.vision_model_tester.batch_size)
+        )
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        config, input_ids, attention_mask, pixel_values = config_and_inputs
+        inputs_dict = {
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+            "pixel_values": pixel_values,
+            "return_loss": True,
+        }
+        return config, inputs_dict
+
+
+@require_tf
+class TFCLIPModelTest(TFModelTesterMixin, unittest.TestCase):
+    all_model_classes = (TFCLIPModel,) if is_tf_available() else ()
+    test_head_masking = False
+    test_pruning = False
+    test_resize_embeddings = False
+    test_attention_outputs = False
+    test_onnx = False
+
+    def setUp(self):
+        self.model_tester = TFCLIPModelTester(self)
+
+    def test_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_model(*config_and_inputs)
+
+    # hidden_states are tested in individual model tests
+    def test_hidden_states_output(self):
+        pass
+
+    # input_embeds are tested in individual model tests
+    def test_inputs_embeds(self):
+        pass
+
+    # CLIPModel does not have input/output embeddings
+    def test_model_common_attributes(self):
+        pass
+
+    # overwrite from common since `TFCLIPModelTester` set `return_loss` to `True` and causes the preparation of
+    # `symbolic_inputs` failed.
+    def test_keras_save_load(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        # remove `return_loss` to make code work
+        if self.__class__.__name__ == "TFCLIPModelTest":
+            inputs_dict.pop("return_loss", None)
+
+        tf_main_layer_classes = set(
+            module_member
+            for model_class in self.all_model_classes
+            for module in (import_module(model_class.__module__),)
+            for module_member_name in dir(module)
+            if module_member_name.endswith("MainLayer")
+            # This condition is required, since `modeling_tf_clip.py` has 3 classes whose names end with `MainLayer`.
+            and module_member_name[: -len("MainLayer")] == model_class.__name__[: -len("Model")]
+            for module_member in (getattr(module, module_member_name),)
+            if isinstance(module_member, type)
+            and tf.keras.layers.Layer in module_member.__bases__
+            and getattr(module_member, "_keras_serializable", False)
+        )
+        for main_layer_class in tf_main_layer_classes:
+            # T5MainLayer needs an embed_tokens parameter when called without the inputs_embeds parameter
+            if "T5" in main_layer_class.__name__:
+                # Take the same values than in TFT5ModelTester for this shared layer
+                shared = TFSharedEmbeddings(99, 32, name="shared")
+                config.use_cache = inputs_dict.pop("use_cache", None)
+                main_layer = main_layer_class(config, embed_tokens=shared)
+            else:
+                main_layer = main_layer_class(config)
+
+            symbolic_inputs = {
+                name: tf.keras.Input(tensor.shape[1:], dtype=tensor.dtype) for name, tensor in inputs_dict.items()
+            }
+
+            model = tf.keras.Model(symbolic_inputs, outputs=main_layer(symbolic_inputs))
+            outputs = model(inputs_dict)
+
+            with tempfile.TemporaryDirectory() as tmpdirname:
+                filepath = os.path.join(tmpdirname, "keras_model.h5")
+                model.save(filepath)
+                if "T5" in main_layer_class.__name__:
+                    model = tf.keras.models.load_model(
+                        filepath,
+                        custom_objects={
+                            main_layer_class.__name__: main_layer_class,
+                            "TFSharedEmbeddings": TFSharedEmbeddings,
+                        },
+                    )
+                else:
+                    model = tf.keras.models.load_model(
+                        filepath, custom_objects={main_layer_class.__name__: main_layer_class}
+                    )
+                assert isinstance(model, tf.keras.Model)
+                after_outputs = model(inputs_dict)
+                self.assert_outputs_same(after_outputs, outputs)
+
+    @slow
+    def test_model_from_pretrained(self):
+        for model_name in TF_CLIP_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
+            model = TFCLIPModel.from_pretrained(model_name)
+            self.assertIsNotNone(model)
+
+
+# We will verify our results on an image of cute cats
+def prepare_img():
+    url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+    im = Image.open(requests.get(url, stream=True).raw)
+    return im
+
+
+@require_vision
+@require_tf
+class TFCLIPModelIntegrationTest(unittest.TestCase):
+    @slow
+    def test_inference(self):
+        model_name = "openai/clip-vit-base-patch32"
+        model = TFCLIPModel.from_pretrained(model_name)
+        processor = CLIPProcessor.from_pretrained(model_name)
+
+        image = prepare_img()
+        inputs = processor(
+            text=["a photo of a cat", "a photo of a dog"], images=image, padding=True, return_tensors="tf"
+        )
+
+        outputs = model(**inputs, training=False)
+
+        # verify the logits
+        self.assertEqual(
+            outputs.logits_per_image.shape,
+            tf.TensorShape((inputs.pixel_values.shape[0], inputs.input_ids.shape[0])),
+        )
+        self.assertEqual(
+            outputs.logits_per_text.shape,
+            tf.TensorShape((inputs.input_ids.shape[0], inputs.pixel_values.shape[0])),
+        )
+
+        expected_logits = tf.constant([[24.5701, 19.3049]])
+
+        tf.debugging.assert_near(outputs.logits_per_image, expected_logits, atol=1e-3)

tests/models/clip/test_processor_clip.py

@@ -0,0 +1,189 @@
+# Copyright 2021 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 json
+import os
+import shutil
+import tempfile
+import unittest
+
+import numpy as np
+import pytest
+
+from transformers import CLIPTokenizer, CLIPTokenizerFast
+from transformers.models.clip.tokenization_clip import VOCAB_FILES_NAMES
+from transformers.testing_utils import require_vision
+from transformers.utils import FEATURE_EXTRACTOR_NAME, is_vision_available
+
+
+if is_vision_available():
+    from PIL import Image
+
+    from transformers import CLIPFeatureExtractor, CLIPProcessor
+
+
+@require_vision
+class CLIPProcessorTest(unittest.TestCase):
+    def setUp(self):
+        self.tmpdirname = tempfile.mkdtemp()
+
+        # fmt: off
+        vocab = ["l", "o", "w", "e", "r", "s", "t", "i", "d", "n", "lo", "l</w>", "w</w>", "r</w>", "t</w>", "low</w>", "er</w>", "lowest</w>", "newer</w>", "wider", "<unk>", "<|startoftext|>", "<|endoftext|>"]
+        # fmt: on
+        vocab_tokens = dict(zip(vocab, range(len(vocab))))
+        merges = ["#version: 0.2", "l o", "lo w</w>", "e r</w>", ""]
+        self.special_tokens_map = {"unk_token": "<unk>"}
+
+        self.vocab_file = os.path.join(self.tmpdirname, VOCAB_FILES_NAMES["vocab_file"])
+        self.merges_file = os.path.join(self.tmpdirname, VOCAB_FILES_NAMES["merges_file"])
+        with open(self.vocab_file, "w", encoding="utf-8") as fp:
+            fp.write(json.dumps(vocab_tokens) + "\n")
+        with open(self.merges_file, "w", encoding="utf-8") as fp:
+            fp.write("\n".join(merges))
+
+        feature_extractor_map = {
+            "do_resize": True,
+            "size": 20,
+            "do_center_crop": True,
+            "crop_size": 18,
+            "do_normalize": True,
+            "image_mean": [0.48145466, 0.4578275, 0.40821073],
+            "image_std": [0.26862954, 0.26130258, 0.27577711],
+        }
+        self.feature_extractor_file = os.path.join(self.tmpdirname, FEATURE_EXTRACTOR_NAME)
+        with open(self.feature_extractor_file, "w", encoding="utf-8") as fp:
+            json.dump(feature_extractor_map, fp)
+
+    def get_tokenizer(self, **kwargs):
+        return CLIPTokenizer.from_pretrained(self.tmpdirname, **kwargs)
+
+    def get_rust_tokenizer(self, **kwargs):
+        return CLIPTokenizerFast.from_pretrained(self.tmpdirname, **kwargs)
+
+    def get_feature_extractor(self, **kwargs):
+        return CLIPFeatureExtractor.from_pretrained(self.tmpdirname, **kwargs)
+
+    def tearDown(self):
+        shutil.rmtree(self.tmpdirname)
+
+    def prepare_image_inputs(self):
+        """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.
+        """
+
+        image_inputs = [np.random.randint(255, size=(3, 30, 400), dtype=np.uint8)]
+
+        image_inputs = [Image.fromarray(np.moveaxis(x, 0, -1)) for x in image_inputs]
+
+        return image_inputs
+
+    def test_save_load_pretrained_default(self):
+        tokenizer_slow = self.get_tokenizer()
+        tokenizer_fast = self.get_rust_tokenizer()
+        feature_extractor = self.get_feature_extractor()
+
+        processor_slow = CLIPProcessor(tokenizer=tokenizer_slow, feature_extractor=feature_extractor)
+        processor_slow.save_pretrained(self.tmpdirname)
+        processor_slow = CLIPProcessor.from_pretrained(self.tmpdirname, use_fast=False)
+
+        processor_fast = CLIPProcessor(tokenizer=tokenizer_fast, feature_extractor=feature_extractor)
+        processor_fast.save_pretrained(self.tmpdirname)
+        processor_fast = CLIPProcessor.from_pretrained(self.tmpdirname)
+
+        self.assertEqual(processor_slow.tokenizer.get_vocab(), tokenizer_slow.get_vocab())
+        self.assertEqual(processor_fast.tokenizer.get_vocab(), tokenizer_fast.get_vocab())
+        self.assertEqual(tokenizer_slow.get_vocab(), tokenizer_fast.get_vocab())
+        self.assertIsInstance(processor_slow.tokenizer, CLIPTokenizer)
+        self.assertIsInstance(processor_fast.tokenizer, CLIPTokenizerFast)
+
+        self.assertEqual(processor_slow.feature_extractor.to_json_string(), feature_extractor.to_json_string())
+        self.assertEqual(processor_fast.feature_extractor.to_json_string(), feature_extractor.to_json_string())
+        self.assertIsInstance(processor_slow.feature_extractor, CLIPFeatureExtractor)
+        self.assertIsInstance(processor_fast.feature_extractor, CLIPFeatureExtractor)
+
+    def test_save_load_pretrained_additional_features(self):
+        processor = CLIPProcessor(tokenizer=self.get_tokenizer(), feature_extractor=self.get_feature_extractor())
+        processor.save_pretrained(self.tmpdirname)
+
+        tokenizer_add_kwargs = self.get_tokenizer(bos_token="(BOS)", eos_token="(EOS)")
+        feature_extractor_add_kwargs = self.get_feature_extractor(do_normalize=False, padding_value=1.0)
+
+        processor = CLIPProcessor.from_pretrained(
+            self.tmpdirname, bos_token="(BOS)", eos_token="(EOS)", do_normalize=False, padding_value=1.0
+        )
+
+        self.assertEqual(processor.tokenizer.get_vocab(), tokenizer_add_kwargs.get_vocab())
+        self.assertIsInstance(processor.tokenizer, CLIPTokenizerFast)
+
+        self.assertEqual(processor.feature_extractor.to_json_string(), feature_extractor_add_kwargs.to_json_string())
+        self.assertIsInstance(processor.feature_extractor, CLIPFeatureExtractor)
+
+    def test_feature_extractor(self):
+        feature_extractor = self.get_feature_extractor()
+        tokenizer = self.get_tokenizer()
+
+        processor = CLIPProcessor(tokenizer=tokenizer, feature_extractor=feature_extractor)
+
+        image_input = self.prepare_image_inputs()
+
+        input_feat_extract = feature_extractor(image_input, return_tensors="np")
+        input_processor = processor(images=image_input, return_tensors="np")
+
+        for key in input_feat_extract.keys():
+            self.assertAlmostEqual(input_feat_extract[key].sum(), input_processor[key].sum(), delta=1e-2)
+
+    def test_tokenizer(self):
+        feature_extractor = self.get_feature_extractor()
+        tokenizer = self.get_tokenizer()
+
+        processor = CLIPProcessor(tokenizer=tokenizer, feature_extractor=feature_extractor)
+
+        input_str = "lower newer"
+
+        encoded_processor = processor(text=input_str)
+
+        encoded_tok = tokenizer(input_str)
+
+        for key in encoded_tok.keys():
+            self.assertListEqual(encoded_tok[key], encoded_processor[key])
+
+    def test_processor(self):
+        feature_extractor = self.get_feature_extractor()
+        tokenizer = self.get_tokenizer()
+
+        processor = CLIPProcessor(tokenizer=tokenizer, feature_extractor=feature_extractor)
+
+        input_str = "lower newer"
+        image_input = self.prepare_image_inputs()
+
+        inputs = processor(text=input_str, images=image_input)
+
+        self.assertListEqual(list(inputs.keys()), ["input_ids", "attention_mask", "pixel_values"])
+
+        # test if it raises when no input is passed
+        with pytest.raises(ValueError):
+            processor()
+
+    def test_tokenizer_decode(self):
+        feature_extractor = self.get_feature_extractor()
+        tokenizer = self.get_tokenizer()
+
+        processor = CLIPProcessor(tokenizer=tokenizer, feature_extractor=feature_extractor)
+
+        predicted_ids = [[1, 4, 5, 8, 1, 0, 8], [3, 4, 3, 1, 1, 8, 9]]
+
+        decoded_processor = processor.batch_decode(predicted_ids)
+        decoded_tok = tokenizer.batch_decode(predicted_ids)
+
+        self.assertListEqual(decoded_tok, decoded_processor)

tests/models/clip/test_tokenization_clip.py

@@ -0,0 +1,186 @@
+# coding=utf-8
+# Copyright 2021 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 json
+import os
+import unittest
+
+from transformers import CLIPTokenizer, CLIPTokenizerFast
+from transformers.models.clip.tokenization_clip import VOCAB_FILES_NAMES
+from transformers.testing_utils import require_ftfy, require_tokenizers
+
+from ...test_tokenization_common import TokenizerTesterMixin
+
+
+@require_tokenizers
+class CLIPTokenizationTest(TokenizerTesterMixin, unittest.TestCase):
+
+    tokenizer_class = CLIPTokenizer
+    rust_tokenizer_class = CLIPTokenizerFast
+    test_rust_tokenizer = True
+    from_pretrained_kwargs = {}
+    test_seq2seq = False
+
+    def setUp(self):
+        super().setUp()
+
+        # fmt: off
+        vocab = ["l", "o", "w", "e", "r", "s", "t", "i", "d", "n", "lo", "l</w>", "w</w>", "r</w>", "t</w>", "low</w>", "er</w>", "lowest</w>", "newer</w>", "wider", "<unk>", "<|startoftext|>", "<|endoftext|>"]
+        # fmt: on
+        vocab_tokens = dict(zip(vocab, range(len(vocab))))
+        merges = ["#version: 0.2", "l o", "lo w</w>", "e r</w>"]
+        self.special_tokens_map = {"unk_token": "<unk>"}
+
+        self.vocab_file = os.path.join(self.tmpdirname, VOCAB_FILES_NAMES["vocab_file"])
+        self.merges_file = os.path.join(self.tmpdirname, VOCAB_FILES_NAMES["merges_file"])
+        with open(self.vocab_file, "w", encoding="utf-8") as fp:
+            fp.write(json.dumps(vocab_tokens) + "\n")
+        with open(self.merges_file, "w", encoding="utf-8") as fp:
+            fp.write("\n".join(merges))
+
+    def get_tokenizer(self, **kwargs):
+        kwargs.update(self.special_tokens_map)
+        return CLIPTokenizer.from_pretrained(self.tmpdirname, **kwargs)
+
+    def get_rust_tokenizer(self, **kwargs):
+        kwargs.update(self.special_tokens_map)
+        return CLIPTokenizerFast.from_pretrained(self.tmpdirname, **kwargs)
+
+    def get_input_output_texts(self, tokenizer):
+        input_text = "lower newer"
+        output_text = "lower newer"
+        return input_text, output_text
+
+    def test_full_tokenizer(self):
+        tokenizer = CLIPTokenizer(self.vocab_file, self.merges_file, **self.special_tokens_map)
+        text = "lower newer"
+        bpe_tokens = ["lo", "w", "er</w>", "n", "e", "w", "er</w>"]
+        tokens = tokenizer.tokenize(text)
+        self.assertListEqual(tokens, bpe_tokens)
+
+        input_tokens = tokens + [tokenizer.unk_token]
+        input_bpe_tokens = [10, 2, 16, 9, 3, 2, 16, 20]
+        self.assertListEqual(tokenizer.convert_tokens_to_ids(input_tokens), input_bpe_tokens)
+
+    @require_ftfy
+    def test_check_encoding_slow_fast(self):
+        for tokenizer, pretrained_name, kwargs in self.tokenizers_list:
+            with self.subTest(f"{tokenizer.__class__.__name__} ({pretrained_name})"):
+                tokenizer_s = self.tokenizer_class.from_pretrained(pretrained_name, **kwargs)
+                tokenizer_r = self.rust_tokenizer_class.from_pretrained(pretrained_name, **kwargs)
+
+                text = "A\n'll 11p223RF☆ho!!to?'d'd''d of a cat"
+                text_tokenized_s = tokenizer_s.tokenize(text)
+                text_tokenized_r = tokenizer_r.tokenize(text)
+
+                self.assertListEqual(text_tokenized_s, text_tokenized_r)
+
+                # Test that the tokenization is identical on an example containing a character (Latin Small Letter A
+                # with Tilde) encoded in 2 different ways
+                text = "xa\u0303y" + " " + "x\xe3y"
+                text_tokenized_s = tokenizer_s.tokenize(text)
+                text_tokenized_r = tokenizer_r.tokenize(text)
+
+                self.assertListEqual(text_tokenized_s, text_tokenized_r)
+
+                # Test that the tokenization is identical on unicode of space type
+                spaces_unicodes = [
+                    "\u0009",  # (horizontal tab, '\t')
+                    "\u000B",  # (vertical tab)
+                    "\u000C",  # (form feed)
+                    "\u0020",  # (space, ' ')
+                    "\u200E",  # (left-to-right mark):w
+                    "\u200F",  # (right-to-left mark)
+                ]
+                for unicode_seq in spaces_unicodes:
+                    text_tokenized_s = tokenizer_s.tokenize(unicode_seq)
+                    text_tokenized_r = tokenizer_r.tokenize(unicode_seq)
+
+                    self.assertListEqual(text_tokenized_s, text_tokenized_r)
+
+                # Test that the tokenization is identical on unicode of line break type
+                line_break_unicodes = [
+                    "\u000A",  # (line feed, '\n')
+                    "\r\n",  # (carriage return and line feed, '\r\n')
+                    "\u000D",  # (carriage return, '\r')
+                    "\r",  # (carriage return, '\r')
+                    "\u000D",  # (carriage return, '\r')
+                    "\u2028",  # (line separator)
+                    "\u2029",  # (paragraph separator)
+                    # "\u0085", # (next line)
+                ]
+
+                # The tokenization is not identical for the character "\u0085" (next line). The slow version transforms
+                # it into the Horizontal Ellipsis character "…" ("\u2026") while the fast version transforms it into a
+                # space (and thus into an empty list).
+
+                for unicode_seq in line_break_unicodes:
+                    text_tokenized_s = tokenizer_s.tokenize(unicode_seq)
+                    text_tokenized_r = tokenizer_r.tokenize(unicode_seq)
+
+                    self.assertListEqual(text_tokenized_s, text_tokenized_r)
+
+    def test_offsets_mapping_with_different_add_prefix_space_argument(self):
+        # Test which aims to verify that the offsets are well adapted to the argument `add_prefix_space`
+        for tokenizer, pretrained_name, kwargs in self.tokenizers_list:
+            with self.subTest(f"{tokenizer.__class__.__name__} ({pretrained_name})"):
+                text_of_1_token = "hello"  # `hello` is a token in the vocabulary of `pretrained_name`
+                text = f"{text_of_1_token} {text_of_1_token}"
+
+                tokenizer_r = self.rust_tokenizer_class.from_pretrained(
+                    pretrained_name,
+                    use_fast=True,
+                )
+                encoding = tokenizer_r(text, return_offsets_mapping=True, add_special_tokens=False)
+                self.assertEqual(encoding.offset_mapping[0], (0, len(text_of_1_token)))
+                self.assertEqual(
+                    encoding.offset_mapping[1],
+                    (len(text_of_1_token) + 1, len(text_of_1_token) + 1 + len(text_of_1_token)),
+                )
+
+                text = f" {text}"
+
+                tokenizer_r = self.rust_tokenizer_class.from_pretrained(
+                    pretrained_name,
+                    use_fast=True,
+                )
+                encoding = tokenizer_r(text, return_offsets_mapping=True, add_special_tokens=False)
+                self.assertEqual(encoding.offset_mapping[0], (1, 1 + len(text_of_1_token)))
+                self.assertEqual(
+                    encoding.offset_mapping[1],
+                    (1 + len(text_of_1_token) + 1, 1 + len(text_of_1_token) + 1 + len(text_of_1_token)),
+                )
+
+    def test_log_warning(self):
+        # Test related to the breaking change introduced in transformers v4.17.0
+        # We need to check that an error in raised when the user try to load a previous version of the tokenizer.
+        with self.assertRaises(ValueError) as context:
+            self.rust_tokenizer_class.from_pretrained("robot-test/old-clip-tokenizer")
+
+        self.assertTrue(
+            context.exception.args[0].startswith(
+                "The `backend_tokenizer` provided does not match the expected format."
+            )
+        )
+
+    @require_ftfy
+    def test_tokenization_python_rust_equals(self):
+        super().test_tokenization_python_rust_equals()
+
+    # overwrite common test
+    def test_added_tokens_do_lower_case(self):
+        # CLIP always lower cases letters
+        pass