CLIP (#11445)

* begin second draft

* fix import, style

* add loss

* fix embeds, logits_scale, and projection

* fix imports

* add conversion script

* add feature_extractor and processor

* style

* add tests for tokenizer, extractor and processor

* add vision model tests

* add weight init

* add more tests

* fix save_load  test

* model output, dosstrings, causal mask

* config doc

* add clip model tests

* return dict

* bigin integration test

* add integration tests

* fix-copies

* fix init

* Clip => CLIP

* fix module name

* docs

* fix doc

* output_dim => projection_dim

* fix checkpoint names

* remoe fast tokenizer file

* fix conversion script

* fix tests, quality

* put causal mask on device

* Apply suggestions from code review
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* fix attribute test

* style

* address sylvains comments

* style

* fix docstrings

* add qucik_gelu in activations, docstrings

* clean-up attention test

* fix act fun

* fix config

* fix torchscript tests

* even batch_size

* remove comment

* fix ouput tu_tuple

* fix save load tests

* fix add tokens test

* add fast tokenizer

* update copyright

* new processor API

* fix docs

* docstrings

* docs

* fix doc

* fix doc

* fix tokenizer

* fix import in doc example

* Apply suggestions from code review
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* check types of config

* valhalla => openai

* load image using url

* fix test

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

tests/test_feature_extraction_clip.py

+# 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.file_utils import is_torch_available, is_vision_available
+from transformers.testing_utils import require_torch, require_vision
+
+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/test_modeling_clip.py

+# coding=utf-8
+# Copyright 2021 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Testing suite for the PyTorch CLIP model. """
+
+
+import inspect
+import os
+import tempfile
+import unittest
+
+import requests
+from transformers.file_utils import is_torch_available, is_vision_available
+from transformers.testing_utils import require_torch, require_vision, slow, torch_device
+
+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 transformers import CLIPConfig, CLIPModel, CLIPTextConfig, CLIPTextModel, CLIPVisionConfig, 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
+
+
+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 = 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 create_and_check_model(self, config, pixel_values):
+        model = CLIPVisionModel(config=config)
+        model.to(torch_device)
+        model.eval()
+        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_torchscript = 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(), (torch.nn.Module))
+            x = model.get_output_embeddings()
+            self.assertTrue(x is None or isinstance(x, torch.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])
+
+        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 create_and_check_model(self, config, input_ids, input_mask):
+        model = CLIPTextModel(config=config)
+        model.to(torch_device)
+        model.eval()
+        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 = CLIPConfig.from_text_vision_configs(text_config, vision_config, projection_dim=64)
+
+        return config, input_ids, attention_mask, pixel_values
+
+    def create_and_check_model(self, config, input_ids, attention_mask, pixel_values):
+        model = CLIPModel(config).to(torch_device).eval()
+        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
+
+    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)
+
+    @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
+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
+        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([[24.5056, 18.8076]]).to(torch_device)
+
+        self.assertTrue(torch.allclose(outputs.logits_per_image, expected_logits, atol=1e-3))

tests/test_processor_clip.py

+# 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
+from transformers.file_utils import FEATURE_EXTRACTOR_NAME, is_vision_available
+from transformers.models.clip.tokenization_clip import VOCAB_FILES_NAMES
+from transformers.testing_utils import require_vision
+
+
+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", "low</w>", "er</w>", "lowest</w>", "newer</w>", "wider", "<unk>", "<|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_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 = self.get_tokenizer()
+        feature_extractor = self.get_feature_extractor()
+
+        processor = CLIPProcessor(tokenizer=tokenizer, feature_extractor=feature_extractor)
+
+        processor.save_pretrained(self.tmpdirname)
+        processor = CLIPProcessor.from_pretrained(self.tmpdirname)
+
+        self.assertEqual(processor.tokenizer.get_vocab(), tokenizer.get_vocab())
+        self.assertIsInstance(processor.tokenizer, CLIPTokenizer)
+
+        self.assertEqual(processor.feature_extractor.to_json_string(), feature_extractor.to_json_string())
+        self.assertIsInstance(processor.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, CLIPTokenizer)
+
+        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/test_tokenization_clip.py

+# 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_tokenizers
+
+from .test_tokenization_common import TokenizerTesterMixin
+
+
+@require_tokenizers
+class CLIPTokenizationTest(TokenizerTesterMixin, unittest.TestCase):
+
+    tokenizer_class = CLIPTokenizer
+    rust_tokenizer_class = CLIPTokenizerFast
+    from_pretrained_kwargs = {"add_prefix_space": True}
+    test_seq2seq = False
+
+    def setUp(self):
+        super().setUp()
+
+        # fmt: off
+        vocab = ["l", "o", "w", "e", "r", "s", "t", "i", "d", "n", "lo", "low</w>", "er</w>", "lowest</w>", "newer</w>", "wider", "<unk>", "<|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, add_prefix_space=True)
+        self.assertListEqual(tokens, bpe_tokens)
+
+        input_tokens = tokens + [tokenizer.unk_token]
+        input_bpe_tokens = [10, 2, 12, 9, 3, 2, 12, 16]
+        self.assertListEqual(tokenizer.convert_tokens_to_ids(input_tokens), input_bpe_tokens)
+
+    def test_rust_and_python_full_tokenizers(self):
+        if not self.test_rust_tokenizer:
+            return
+
+        tokenizer = self.get_tokenizer()
+        rust_tokenizer = self.get_rust_tokenizer(add_prefix_space=True)
+
+        sequence = "lower newer"
+
+        # Testing tokenization
+        tokens = tokenizer.tokenize(sequence, add_prefix_space=True)
+        rust_tokens = rust_tokenizer.tokenize(sequence)
+        self.assertListEqual(tokens, rust_tokens)
+
+        # Testing conversion to ids without special tokens
+        ids = tokenizer.encode(sequence, add_special_tokens=False, add_prefix_space=True)
+        rust_ids = rust_tokenizer.encode(sequence, add_special_tokens=False)
+        self.assertListEqual(ids, rust_ids)
+
+        # Testing conversion to ids with special tokens
+        rust_tokenizer = self.get_rust_tokenizer(add_prefix_space=True)
+        ids = tokenizer.encode(sequence, add_prefix_space=True)
+        rust_ids = rust_tokenizer.encode(sequence)
+        self.assertListEqual(ids, rust_ids)
+
+        # Testing the unknown token
+        input_tokens = tokens + [rust_tokenizer.unk_token]
+        input_bpe_tokens = [10, 2, 12, 9, 3, 2, 12, 16]
+        self.assertListEqual(rust_tokenizer.convert_tokens_to_ids(input_tokens), input_bpe_tokens)
+
+    def test_pretokenized_inputs(self, *args, **kwargs):
+        # It's very difficult to mix/test pretokenization with byte-level
+        # And get both CLIP and Roberta to work at the same time (mostly an issue of adding a space before the string)
+        pass
+
+    def test_padding(self, max_length=15):
+        for tokenizer, pretrained_name, kwargs in self.tokenizers_list:
+            with self.subTest(f"{tokenizer.__class__.__name__} ({pretrained_name})"):
+                tokenizer_r = self.rust_tokenizer_class.from_pretrained(pretrained_name, **kwargs)
+
+                # Simple input
+                s = "This is a simple input"
+                s2 = ["This is a simple input 1", "This is a simple input 2"]
+                p = ("This is a simple input", "This is a pair")
+                p2 = [
+                    ("This is a simple input 1", "This is a simple input 2"),
+                    ("This is a simple pair 1", "This is a simple pair 2"),
+                ]
+
+                # Simple input tests
+                self.assertRaises(ValueError, tokenizer_r.encode, s, max_length=max_length, padding="max_length")
+
+                # Simple input
+                self.assertRaises(ValueError, tokenizer_r.encode_plus, s, max_length=max_length, padding="max_length")
+
+                # Simple input
+                self.assertRaises(
+                    ValueError,
+                    tokenizer_r.batch_encode_plus,
+                    s2,
+                    max_length=max_length,
+                    padding="max_length",
+                )
+
+                # Pair input
+                self.assertRaises(ValueError, tokenizer_r.encode, p, max_length=max_length, padding="max_length")
+
+                # Pair input
+                self.assertRaises(ValueError, tokenizer_r.encode_plus, p, max_length=max_length, padding="max_length")
+
+                # Pair input
+                self.assertRaises(
+                    ValueError,
+                    tokenizer_r.batch_encode_plus,
+                    p2,
+                    max_length=max_length,
+                    padding="max_length",
+                )
+
+    def test_add_tokens_tokenizer(self):
+        tokenizers = self.get_tokenizers(do_lower_case=False)
+        for tokenizer in tokenizers:
+            with self.subTest(f"{tokenizer.__class__.__name__}"):
+                vocab_size = tokenizer.vocab_size
+                all_size = len(tokenizer)
+
+                self.assertNotEqual(vocab_size, 0)
+
+                # We usually have added tokens from the start in tests because our vocab fixtures are
+                # smaller than the original vocabs - let's not assert this
+                # self.assertEqual(vocab_size, all_size)
+
+                new_toks = ["aaaaa bbbbbb", "cccccccccdddddddd"]
+                added_toks = tokenizer.add_tokens(new_toks)
+                vocab_size_2 = tokenizer.vocab_size
+                all_size_2 = len(tokenizer)
+
+                self.assertNotEqual(vocab_size_2, 0)
+                self.assertEqual(vocab_size, vocab_size_2)
+                self.assertEqual(added_toks, len(new_toks))
+                self.assertEqual(all_size_2, all_size + len(new_toks))
+
+                tokens = tokenizer.encode("aaaaa bbbbbb low cccccccccdddddddd l", add_special_tokens=False)
+
+                self.assertGreaterEqual(len(tokens), 4)
+                self.assertGreater(tokens[0], tokenizer.vocab_size - 1)
+                self.assertGreater(tokens[-2], tokenizer.vocab_size - 1)
+
+                new_toks_2 = {"eos_token": ">>>>|||<||<<|<<", "pad_token": "<<<<<|||>|>>>>|>"}
+                added_toks_2 = tokenizer.add_special_tokens(new_toks_2)
+                vocab_size_3 = tokenizer.vocab_size
+                all_size_3 = len(tokenizer)
+
+                self.assertNotEqual(vocab_size_3, 0)
+                self.assertEqual(vocab_size, vocab_size_3)
+                self.assertEqual(added_toks_2, len(new_toks_2))
+                self.assertEqual(all_size_3, all_size_2 + len(new_toks_2))
+
+                tokens = tokenizer.encode(
+                    ">>>>|||<||<<|<< aaaaabbbbbb low cccccccccdddddddd <<<<<|||>|>>>>|> l", add_special_tokens=False
+                )
+
+                self.assertGreaterEqual(len(tokens), 6)
+                self.assertGreater(tokens[0], tokenizer.vocab_size - 1)
+                self.assertGreater(tokens[0], tokens[1])
+                self.assertGreater(tokens[-2], tokenizer.vocab_size - 1)
+                self.assertGreater(tokens[-2], tokens[-3])
+                self.assertEqual(tokens[0], tokenizer.eos_token_id)
+                # padding is very hacky in CLIPTokenizer, pad_token_id is always 0
+                # so skip this check
+                # self.assertEqual(tokens[-2], tokenizer.pad_token_id)

utils/check_repo.py

@@ -91,6 +91,8 @@ TEST_FILES_WITH_NO_COMMON_TESTS = [
 # should **not** be the rule.
 IGNORE_NON_AUTO_CONFIGURED = [
     # models to ignore for model xxx mapping
+    "CLIPTextModel",
+    "CLIPVisionModel",
     "DPRReader",
     "DPRSpanPredictor",
     "FlaubertForQuestionAnswering",