Add ViTMatte (#25843)

* First draft

* Simplify image processor

* Fix rebase

* Address comments

* Address more comments

* Address more comments

* Address more comments

* Address more comments

* Improve pad_image

* Add tests

* Update integration test

* Fix image processor tests

* Fix model tests

* Convert checkpoints

* Fix doc tests

* Remove file

* Apply suggestions

* Address comments

* Fix typing hint

* Add batch_norm_eps

* Address comments

* Fix style

src/transformers/utils/dummy_vision_objects.py

@@ -513,6 +513,13 @@ class ViTHybridImageProcessor(metaclass=DummyObject):
         requires_backends(self, ["vision"])
 
 
+class VitMatteImageProcessor(metaclass=DummyObject):
+    _backends = ["vision"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["vision"])
+
+
 class VivitImageProcessor(metaclass=DummyObject):
     _backends = ["vision"]
 

tests/models/vitmatte/test_image_processing_vitmatte.py

+# coding=utf-8
+# Copyright 2023 HuggingFace Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import unittest
+
+import numpy as np
+
+from transformers.testing_utils import require_torch, require_vision
+from transformers.utils import is_torch_available, is_vision_available
+
+from ...test_image_processing_common import ImageProcessingTestMixin, prepare_image_inputs
+
+
+if is_torch_available():
+    import torch
+
+
+if is_vision_available():
+    from PIL import Image
+
+    from transformers import VitMatteImageProcessor
+
+
+class VitMatteImageProcessingTester(unittest.TestCase):
+    def __init__(
+        self,
+        parent,
+        batch_size=7,
+        num_channels=3,
+        image_size=18,
+        min_resolution=30,
+        max_resolution=400,
+        do_rescale=True,
+        rescale_factor=0.5,
+        do_pad=True,
+        size_divisibility=10,
+        do_normalize=True,
+        image_mean=[0.5, 0.5, 0.5],
+        image_std=[0.5, 0.5, 0.5],
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.num_channels = num_channels
+        self.image_size = image_size
+        self.min_resolution = min_resolution
+        self.max_resolution = max_resolution
+        self.do_rescale = do_rescale
+        self.rescale_factor = rescale_factor
+        self.do_pad = do_pad
+        self.size_divisibility = size_divisibility
+        self.do_normalize = do_normalize
+        self.image_mean = image_mean
+        self.image_std = image_std
+
+    def prepare_image_processor_dict(self):
+        return {
+            "image_mean": self.image_mean,
+            "image_std": self.image_std,
+            "do_normalize": self.do_normalize,
+            "do_rescale": self.do_rescale,
+            "rescale_factor": self.rescale_factor,
+            "do_pad": self.do_pad,
+            "size_divisibility": self.size_divisibility,
+        }
+
+    def prepare_image_inputs(self, equal_resolution=False, numpify=False, torchify=False):
+        return prepare_image_inputs(
+            batch_size=self.batch_size,
+            num_channels=self.num_channels,
+            min_resolution=self.min_resolution,
+            max_resolution=self.max_resolution,
+            equal_resolution=equal_resolution,
+            numpify=numpify,
+            torchify=torchify,
+        )
+
+
+@require_torch
+@require_vision
+class VitMatteImageProcessingTest(ImageProcessingTestMixin, unittest.TestCase):
+    image_processing_class = VitMatteImageProcessor if is_vision_available() else None
+
+    def setUp(self):
+        self.image_processor_tester = VitMatteImageProcessingTester(self)
+
+    @property
+    def image_processor_dict(self):
+        return self.image_processor_tester.prepare_image_processor_dict()
+
+    def test_image_processor_properties(self):
+        image_processing = self.image_processing_class(**self.image_processor_dict)
+        self.assertTrue(hasattr(image_processing, "image_mean"))
+        self.assertTrue(hasattr(image_processing, "image_std"))
+        self.assertTrue(hasattr(image_processing, "do_normalize"))
+        self.assertTrue(hasattr(image_processing, "do_rescale"))
+        self.assertTrue(hasattr(image_processing, "rescale_factor"))
+        self.assertTrue(hasattr(image_processing, "do_pad"))
+        self.assertTrue(hasattr(image_processing, "size_divisibility"))
+
+    def test_call_numpy(self):
+        # Initialize image_processing
+        image_processing = self.image_processing_class(**self.image_processor_dict)
+        # create random numpy tensors
+        image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=False, numpify=True)
+        for image in image_inputs:
+            self.assertIsInstance(image, np.ndarray)
+
+        # Test not batched input (image processor does not support batched inputs)
+        image = image_inputs[0]
+        trimap = np.random.randint(0, 3, size=image.shape[:2])
+        encoded_images = image_processing(images=image, trimaps=trimap, return_tensors="pt").pixel_values
+
+        # Verify that width and height can be divided by size_divisibility
+        self.assertTrue(encoded_images.shape[-1] % self.image_processor_tester.size_divisibility == 0)
+        self.assertTrue(encoded_images.shape[-2] % self.image_processor_tester.size_divisibility == 0)
+
+    def test_call_pytorch(self):
+        # Initialize image_processing
+        image_processing = self.image_processing_class(**self.image_processor_dict)
+        # create random PyTorch tensors
+        image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=False, torchify=True)
+
+        for image in image_inputs:
+            self.assertIsInstance(image, torch.Tensor)
+
+        # Test not batched input (image processor does not support batched inputs)
+        image = image_inputs[0]
+        trimap = np.random.randint(0, 3, size=image.shape[:2])
+        encoded_images = image_processing(images=image, trimaps=trimap, return_tensors="pt").pixel_values
+
+        # Verify that width and height can be divided by size_divisibility
+        self.assertTrue(encoded_images.shape[-1] % self.image_processor_tester.size_divisibility == 0)
+        self.assertTrue(encoded_images.shape[-2] % self.image_processor_tester.size_divisibility == 0)
+
+    def test_call_pil(self):
+        # Initialize image_processing
+        image_processing = self.image_processing_class(**self.image_processor_dict)
+        # create random PIL images
+        image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=False)
+        for image in image_inputs:
+            self.assertIsInstance(image, Image.Image)
+
+        # Test not batched input (image processor does not support batched inputs)
+        image = image_inputs[0]
+        trimap = np.random.randint(0, 3, size=image.size[::-1])
+        encoded_images = image_processing(images=image, trimaps=trimap, return_tensors="pt").pixel_values
+
+        # Verify that width and height can be divided by size_divisibility
+        self.assertTrue(encoded_images.shape[-1] % self.image_processor_tester.size_divisibility == 0)
+        self.assertTrue(encoded_images.shape[-2] % self.image_processor_tester.size_divisibility == 0)
+
+    def test_call_numpy_4_channels(self):
+        # Test that can process images which have an arbitrary number of channels
+        # Initialize image_processing
+        image_processor = self.image_processing_class(**self.image_processor_dict)
+
+        # create random numpy tensors
+        self.image_processor_tester.num_channels = 4
+        image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=False, numpify=True)
+
+        # Test not batched input (image processor does not support batched inputs)
+        image = image_inputs[0]
+        trimap = np.random.randint(0, 3, size=image.shape[:2])
+        encoded_images = image_processor(
+            images=image,
+            trimaps=trimap,
+            input_data_format="channels_first",
+            image_mean=0,
+            image_std=1,
+            return_tensors="pt",
+        ).pixel_values
+
+        # Verify that width and height can be divided by size_divisibility
+        self.assertTrue(encoded_images.shape[-1] % self.image_processor_tester.size_divisibility == 0)
+        self.assertTrue(encoded_images.shape[-2] % self.image_processor_tester.size_divisibility == 0)
+
+    def test_padding(self):
+        image_processing = self.image_processing_class(**self.image_processor_dict)
+        image = np.random.randn(3, 249, 491)
+        images = image_processing.pad_image(image)
+        assert images.shape == (3, 256, 512)

tests/models/vitmatte/test_modeling_vitmatte.py

+# coding=utf-8
+# Copyright 2023 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Testing suite for the PyTorch VitMatte model. """
+
+
+import inspect
+import unittest
+
+from huggingface_hub import hf_hub_download
+
+from transformers import VitMatteConfig
+from transformers.testing_utils import (
+    require_torch,
+    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, floats_tensor
+from ...test_pipeline_mixin import PipelineTesterMixin
+
+
+if is_torch_available():
+    import torch
+
+    from transformers import VitDetConfig, VitMatteForImageMatting
+    from transformers.models.vitmatte.modeling_vitmatte import VITMATTE_PRETRAINED_MODEL_ARCHIVE_LIST
+
+
+if is_vision_available():
+    from PIL import Image
+
+    from transformers import VitMatteImageProcessor
+
+
+class VitMatteModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=13,
+        image_size=32,
+        patch_size=16,
+        num_channels=4,
+        is_training=True,
+        use_labels=False,
+        hidden_size=2,
+        num_hidden_layers=2,
+        num_attention_heads=2,
+        hidden_act="gelu",
+        type_sequence_label_size=10,
+        initializer_range=0.02,
+        scope=None,
+        out_features=["stage1"],
+        fusion_hidden_sizes=[128, 64, 32, 16],
+    ):
+        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.hidden_act = hidden_act
+        self.type_sequence_label_size = type_sequence_label_size
+        self.initializer_range = initializer_range
+        self.scope = scope
+        self.out_features = out_features
+        self.fusion_hidden_sizes = fusion_hidden_sizes
+
+        self.seq_length = (self.image_size // self.patch_size) ** 2
+
+    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:
+            raise NotImplementedError("Training is not yet supported")
+
+        config = self.get_config()
+
+        return config, pixel_values, labels
+
+    def get_backbone_config(self):
+        return VitDetConfig(
+            image_size=self.image_size,
+            patch_size=self.patch_size,
+            num_channels=self.num_channels,
+            num_hidden_layers=self.num_hidden_layers,
+            num_attention_heads=self.num_attention_heads,
+            hidden_size=self.hidden_size,
+            is_training=self.is_training,
+            hidden_act=self.hidden_act,
+            out_features=self.out_features,
+        )
+
+    def get_config(self):
+        return VitMatteConfig(
+            backbone_config=self.get_backbone_config(),
+            hidden_size=self.hidden_size,
+            fusion_hidden_sizes=self.fusion_hidden_sizes,
+        )
+
+    def create_and_check_model(self, config, pixel_values, labels):
+        model = VitMatteForImageMatting(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(pixel_values)
+        self.parent.assertEqual(result.alphas.shape, (self.batch_size, 1, self.image_size, self.image_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 VitMatteModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
+    """
+    Here we also overwrite some of the tests of test_modeling_common.py, as VitMatte does not use input_ids, inputs_embeds,
+    attention_mask and seq_length.
+    """
+
+    all_model_classes = (VitMatteForImageMatting,) if is_torch_available() else ()
+    pipeline_model_mapping = {}
+
+    fx_compatible = False
+    test_pruning = False
+    test_resize_embeddings = False
+    test_head_masking = False
+
+    def setUp(self):
+        self.model_tester = VitMatteModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=VitMatteConfig, has_text_modality=False, hidden_size=37)
+
+    def test_config(self):
+        self.create_and_test_config_common_properties()
+        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()
+        self.config_tester.check_config_arguments_init()
+
+    def create_and_test_config_common_properties(self):
+        return
+
+    @unittest.skip(reason="VitMatte does not use inputs_embeds")
+    def test_inputs_embeds(self):
+        pass
+
+    @unittest.skip(reason="Training is not yet supported")
+    def test_training(self):
+        pass
+
+    @unittest.skip(reason="Training is not yet supported")
+    def test_training_gradient_checkpointing(self):
+        pass
+
+    @unittest.skip(reason="ViTMatte does not support input and output embeddings")
+    def test_model_common_attributes(self):
+        pass
+
+    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)
+
+    @slow
+    def test_model_from_pretrained(self):
+        for model_name in VITMATTE_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
+            model = VitMatteForImageMatting.from_pretrained(model_name)
+            self.assertIsNotNone(model)
+
+    @unittest.skip(reason="ViTMatte does not support retaining gradient on attention logits")
+    def test_retain_grad_hidden_states_attentions(self):
+        pass
+
+    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)
+
+            self.assertListEqual(
+                list(hidden_states[0].shape[-2:]),
+                [2, 2],
+            )
+
+        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
+
+            print("Hello we're here")
+
+            check_hidden_states_output(inputs_dict, config, model_class)
+
+
+@require_torch
+class VitMatteModelIntegrationTest(unittest.TestCase):
+    @slow
+    def test_inference(self):
+        processor = VitMatteImageProcessor.from_pretrained("hustvl/vitmatte-small-composition-1k")
+        model = VitMatteForImageMatting.from_pretrained("hustvl/vitmatte-small-composition-1k").to(torch_device)
+
+        filepath = hf_hub_download(
+            repo_id="hf-internal-testing/image-matting-fixtures", filename="image.png", repo_type="dataset"
+        )
+        image = Image.open(filepath).convert("RGB")
+        filepath = hf_hub_download(
+            repo_id="hf-internal-testing/image-matting-fixtures", filename="trimap.png", repo_type="dataset"
+        )
+        trimap = Image.open(filepath).convert("L")
+
+        # prepare image + trimap for the model
+        inputs = processor(images=image, trimaps=trimap, return_tensors="pt").to(torch_device)
+
+        with torch.no_grad():
+            alphas = model(**inputs).alphas
+
+        expected_shape = torch.Size((1, 1, 640, 960))
+        self.assertEqual(alphas.shape, expected_shape)
+
+        expected_slice = torch.tensor(
+            [[0.9977, 0.9987, 0.9990], [0.9980, 0.9998, 0.9998], [0.9983, 0.9998, 0.9998]], device=torch_device
+        )
+        self.assertTrue(torch.allclose(alphas[0, 0, :3, :3], expected_slice, atol=1e-4))

utils/check_repo.py

@@ -277,6 +277,7 @@ IGNORE_NON_AUTO_CONFIGURED = PRIVATE_MODELS.copy() + [
     "SpeechT5ForSpeechToSpeech",
     "SpeechT5ForTextToSpeech",
     "SpeechT5HifiGan",
+    "VitMatteForImageMatting",
 ]
 
 # DO NOT edit this list!