Add Vision Transformer and ViTFeatureExtractor (#10950)

* Squash all commits into one

* Update ViTFeatureExtractor to use image_utils instead of torchvision

* Remove torchvision and add Pillow

* Small docs improvement

* Address most comments by @sgugger

* Fix tests

* Clean up conversion script

* Pooler first draft

* Fix quality

* Improve conversion script

* Make style and quality

* Make fix-copies

* Minor docs improvements

* Should use fix-copies instead of manual handling

* Revert "Should use fix-copies instead of manual handling"

This reverts commit fd4e591bce4496d41406425c82606a8fdaf8a50b.

* Place ViT in alphabetical order
Co-authored-by: Lysandre <lysandre.debut@reseau.eseo.fr>
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

tests/test_feature_extraction_vit.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 ViTFeatureExtractor
+
+
+class ViTFeatureExtractionTester(unittest.TestCase):
+    def __init__(
+        self,
+        parent,
+        batch_size=7,
+        num_channels=3,
+        image_size=18,
+        min_resolution=30,
+        max_resolution=400,
+        image_mean=[0.5, 0.5, 0.5],
+        image_std=[0.5, 0.5, 0.5],
+        do_normalize=True,
+        do_resize=True,
+        size=18,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.num_channels = num_channels
+        self.image_size = image_size
+        self.min_resolution = min_resolution
+        self.max_resolution = max_resolution
+        self.image_mean = image_mean
+        self.image_std = image_std
+        self.do_normalize = do_normalize
+        self.do_resize = do_resize
+        self.size = size
+
+    def prepare_feat_extract_dict(self):
+        return {
+            "image_mean": self.image_mean,
+            "image_std": self.image_std,
+            "do_normalize": self.do_normalize,
+            "do_resize": self.do_resize,
+            "size": self.size,
+        }
+
+    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 ViTFeatureExtractionTest(FeatureExtractionSavingTestMixin, unittest.TestCase):
+
+    feature_extraction_class = ViTFeatureExtractor if is_vision_available() else None
+
+    def setUp(self):
+        self.feature_extract_tester = ViTFeatureExtractionTester(self)
+
+    @property
+    def feat_extract_dict(self):
+        return self.feature_extract_tester.prepare_feat_extract_dict()
+
+    def test_feat_extract_properties(self):
+        feature_extractor = self.feature_extraction_class(**self.feat_extract_dict)
+        self.assertTrue(hasattr(feature_extractor, "image_mean"))
+        self.assertTrue(hasattr(feature_extractor, "image_std"))
+        self.assertTrue(hasattr(feature_extractor, "do_normalize"))
+        self.assertTrue(hasattr(feature_extractor, "do_resize"))
+        self.assertTrue(hasattr(feature_extractor, "size"))
+
+    def test_batch_feature(self):
+        pass
+
+    def test_call_pil(self):
+        # Initialize feature_extractor
+        feature_extractor = self.feature_extraction_class(**self.feat_extract_dict)
+        # create random PIL 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.size,
+                self.feature_extract_tester.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.size,
+                self.feature_extract_tester.size,
+            ),
+        )
+
+    def test_call_numpy(self):
+        # Initialize feature_extractor
+        feature_extractor = self.feature_extraction_class(**self.feat_extract_dict)
+        # create random numpy tensors
+        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.size,
+                self.feature_extract_tester.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.size,
+                self.feature_extract_tester.size,
+            ),
+        )
+
+    def test_call_pytorch(self):
+        # Initialize feature_extractor
+        feature_extractor = self.feature_extraction_class(**self.feat_extract_dict)
+        # create random PyTorch tensors
+        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.size,
+                self.feature_extract_tester.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.size,
+                self.feature_extract_tester.size,
+            ),
+        )

tests/test_image_utils.py

@@ -264,7 +264,9 @@ class ImageFeatureExtractionTester(unittest.TestCase):
 
         # During the conversion rescale and channel first will be applied.
         expected = array.transpose(2, 0, 1).astype(np.float32) / 255.0
-        expected = (expected - np.array(mean)[:, None, None]) / np.array(std)[:, None, None]
+        np_mean = np.array(mean).astype(np.float32)[:, None, None]
+        np_std = np.array(std).astype(np.float32)[:, None, None]
+        expected = (expected - np_mean) / np_std
         self.assertTrue(np.array_equal(normalized_image, expected))
 
     def test_normalize_array(self):

tests/test_modeling_common.py

@@ -34,6 +34,7 @@ if is_torch_available():
     from transformers import (
         BERT_PRETRAINED_MODEL_ARCHIVE_LIST,
         MODEL_FOR_CAUSAL_LM_MAPPING,
+        MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING,
         MODEL_FOR_MASKED_LM_MAPPING,
         MODEL_FOR_MULTIPLE_CHOICE_MAPPING,
         MODEL_FOR_NEXT_SENTENCE_PREDICTION_MAPPING,
@@ -99,6 +100,7 @@ class ModelTesterMixin:
             elif model_class in [
                 *MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING.values(),
                 *MODEL_FOR_NEXT_SENTENCE_PREDICTION_MAPPING.values(),
+                *MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING.values(),
             ]:
                 inputs_dict["labels"] = torch.zeros(
                     self.model_tester.batch_size, dtype=torch.long, device=torch_device

tests/test_modeling_vit.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 ViT model. """
+
+
+import inspect
+import unittest
+
+from transformers.file_utils import cached_property, 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, floats_tensor, ids_tensor
+
+
+if is_torch_available():
+    import torch
+
+    from transformers import ViTConfig, ViTForImageClassification, ViTModel
+    from transformers.models.vit.modeling_vit import VIT_PRETRAINED_MODEL_ARCHIVE_LIST, to_2tuple
+
+
+if is_vision_available():
+    from PIL import Image
+
+    from transformers import ViTFeatureExtractor
+
+
+class ViTModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=13,
+        image_size=30,
+        patch_size=2,
+        num_channels=3,
+        is_training=True,
+        use_labels=True,
+        hidden_size=32,
+        num_hidden_layers=5,
+        num_attention_heads=4,
+        intermediate_size=37,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        type_sequence_label_size=10,
+        initializer_range=0.02,
+        num_labels=3,
+        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.use_labels = use_labels
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.intermediate_size = intermediate_size
+        self.hidden_act = hidden_act
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.type_sequence_label_size = type_sequence_label_size
+        self.initializer_range = initializer_range
+        self.scope = scope
+
+    def prepare_config_and_inputs(self):
+        pixel_values = floats_tensor([self.batch_size, self.num_channels, self.image_size, self.image_size])
+
+        labels = None
+        if self.use_labels:
+            labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
+
+        config = ViTConfig(
+            image_size=self.image_size,
+            patch_size=self.patch_size,
+            num_channels=self.num_channels,
+            hidden_size=self.hidden_size,
+            num_hidden_layers=self.num_hidden_layers,
+            num_attention_heads=self.num_attention_heads,
+            intermediate_size=self.intermediate_size,
+            hidden_act=self.hidden_act,
+            hidden_dropout_prob=self.hidden_dropout_prob,
+            attention_probs_dropout_prob=self.attention_probs_dropout_prob,
+            is_decoder=False,
+            initializer_range=self.initializer_range,
+        )
+
+        return config, pixel_values, labels
+
+    def create_and_check_model(self, config, pixel_values, labels):
+        model = ViTModel(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 = to_2tuple(self.image_size)
+        patch_size = to_2tuple(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))
+
+    def create_and_check_for_image_classification(self, config, pixel_values, labels):
+        config.num_labels = self.type_sequence_label_size
+        model = ViTForImageClassification(config)
+        model.to(torch_device)
+        model.eval()
+        result = model(pixel_values, labels=labels)
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.type_sequence_label_size))
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        (
+            config,
+            pixel_values,
+            labels,
+        ) = config_and_inputs
+        inputs_dict = {"pixel_values": pixel_values}
+        return config, inputs_dict
+
+
+@require_torch
+class ViTModelTest(ModelTesterMixin, unittest.TestCase):
+    """
+    Here we also overwrite some of the tests of test_modeling_common.py, as ViT does not use input_ids, inputs_embeds,
+    attention_mask and seq_length.
+    """
+
+    all_model_classes = (
+        (
+            ViTModel,
+            ViTForImageClassification,
+        )
+        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 = ViTModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=ViTConfig, hidden_size=37)
+
+    def test_config(self):
+        config = self.config_tester.config_class(**self.config_tester.inputs_dict)
+        # we omit vocab_size since ViT does not use this
+        self.config_tester.parent.assertTrue(hasattr(config, "hidden_size"))
+        self.config_tester.parent.assertTrue(hasattr(config, "num_attention_heads"))
+        self.config_tester.parent.assertTrue(hasattr(config, "num_hidden_layers"))
+
+        self.config_tester.create_and_test_config_to_json_string()
+        self.config_tester.create_and_test_config_to_json_file()
+        self.config_tester.create_and_test_config_from_and_save_pretrained()
+        self.config_tester.create_and_test_config_with_num_labels()
+        self.config_tester.check_config_can_be_init_without_params()
+
+    def test_inputs_embeds(self):
+        # ViT 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 ViT, the seq_len equals the number of patches + 1 (we add 1 for the [CLS] token)
+        image_size = to_2tuple(self.model_tester.image_size)
+        patch_size = to_2tuple(self.model_tester.patch_size)
+        num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
+        seq_len = num_patches + 1
+        encoder_seq_length = getattr(self.model_tester, "encoder_seq_length", seq_len)
+        encoder_key_length = getattr(self.model_tester, "key_length", encoder_seq_length)
+        chunk_length = getattr(self.model_tester, "chunk_length", None)
+        if chunk_length is not None and hasattr(self.model_tester, "num_hashes"):
+            encoder_seq_length = encoder_seq_length * self.model_tester.num_hashes
+
+        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.encoder_attentions if config.is_encoder_decoder else 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.encoder_attentions if config.is_encoder_decoder else outputs.attentions
+            self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)
+
+            if chunk_length is not None:
+                self.assertListEqual(
+                    list(attentions[0].shape[-4:]),
+                    [self.model_tester.num_attention_heads, encoder_seq_length, chunk_length, encoder_key_length],
+                )
+            else:
+                self.assertListEqual(
+                    list(attentions[0].shape[-3:]),
+                    [self.model_tester.num_attention_heads, encoder_seq_length, encoder_key_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)
+            model.to(torch_device)
+            model.eval()
+            with torch.no_grad():
+                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
+
+            if hasattr(self.model_tester, "num_hidden_states_types"):
+                added_hidden_states = self.model_tester.num_hidden_states_types
+            elif self.is_encoder_decoder:
+                added_hidden_states = 2
+            else:
+                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)
+            if chunk_length is not None:
+                self.assertListEqual(
+                    list(self_attentions[0].shape[-4:]),
+                    [self.model_tester.num_attention_heads, encoder_seq_length, chunk_length, encoder_key_length],
+                )
+            else:
+                self.assertListEqual(
+                    list(self_attentions[0].shape[-3:]),
+                    [self.model_tester.num_attention_heads, encoder_seq_length, encoder_key_length],
+                )
+
+    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)
+
+            # ViT has a different seq_length
+            image_size = to_2tuple(self.model_tester.image_size)
+            patch_size = to_2tuple(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_for_image_classification(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_image_classification(*config_and_inputs)
+
+    @slow
+    def test_model_from_pretrained(self):
+        for model_name in VIT_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
+            model = ViTModel.from_pretrained(model_name)
+            self.assertIsNotNone(model)
+
+
+# We will verify our results on an image of cute cats
+def prepare_img():
+    image = Image.open("./tests/fixtures/tests_samples/COCO/cats.png")
+    return image
+
+
+@require_vision
+class ViTModelIntegrationTest(unittest.TestCase):
+    @cached_property
+    def default_feature_extractor(self):
+        return ViTFeatureExtractor.from_pretrained("google/vit-base-patch16-224") if is_vision_available() else None
+
+    @slow
+    def test_inference_image_classification_head(self):
+        model = ViTForImageClassification.from_pretrained("google/vit-base-patch16-224").to(torch_device)
+
+        feature_extractor = self.default_feature_extractor
+        image = prepare_img()
+        inputs = feature_extractor(images=image, return_tensors="pt").to(torch_device)
+
+        # forward pass
+        # currently failing
+        # see https://discuss.pytorch.org/t/runtimeerror-expected-object-of-scalar-type-double-but-got-scalar-type-float-for-argument-2-weight/38961/2
+        outputs = model(inputs["pixel_values"])
+        # outputs = model(**inputs)
+
+        # verify the logits
+        expected_shape = torch.Size((1, 1000))
+        self.assertEqual(outputs.logits.shape, expected_shape)
+
+        expected_slice = torch.tensor([-0.2744, 0.8215, -0.0836]).to(torch_device)
+
+        self.assertTrue(torch.allclose(outputs.logits[0, :3], expected_slice, atol=1e-4))