first commit

tests/models/owlv2/test_image_processing_owlv2.py

@@ -0,0 +1,170 @@
+# 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
+
+from transformers.testing_utils import require_torch, require_vision, slow
+from transformers.utils import is_torch_available, is_vision_available
+
+from ...test_image_processing_common import ImageProcessingTestMixin, prepare_image_inputs
+
+
+if is_vision_available():
+    from PIL import Image
+
+    from transformers import AutoProcessor, Owlv2ForObjectDetection
+
+if is_torch_available():
+    import torch
+
+
+class Owlv2ImageProcessingTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=7,
+        num_channels=3,
+        image_size=18,
+        min_resolution=30,
+        max_resolution=400,
+        do_resize=True,
+        size=None,
+        do_normalize=True,
+        image_mean=[0.48145466, 0.4578275, 0.40821073],
+        image_std=[0.26862954, 0.26130258, 0.27577711],
+        do_convert_rgb=True,
+    ):
+        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 if size is not None else {"height": 18, "width": 18}
+        self.do_normalize = do_normalize
+        self.image_mean = image_mean
+        self.image_std = image_std
+        self.do_convert_rgb = do_convert_rgb
+
+    def prepare_image_processor_dict(self):
+        return {
+            "do_resize": self.do_resize,
+            "size": self.size,
+            "do_normalize": self.do_normalize,
+            "image_mean": self.image_mean,
+            "image_std": self.image_std,
+        }
+
+    def expected_output_image_shape(self, images):
+        return self.num_channels, self.size["height"], self.size["width"]
+
+    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 Owlv2ImageProcessingTest(ImageProcessingTestMixin, unittest.TestCase):
+    def setUp(self):
+        super().setUp()
+        self.image_processor_tester = Owlv2ImageProcessingTester(self)
+
+    @property
+    def image_processor_dict(self):
+        return self.image_processor_tester.prepare_image_processor_dict()
+
+    def test_image_processor_properties(self):
+        for image_processing_class in self.image_processing_classes.values():
+            image_processing = image_processing_class(**self.image_processor_dict)
+            self.assertTrue(hasattr(image_processing, "do_resize"))
+            self.assertTrue(hasattr(image_processing, "size"))
+            self.assertTrue(hasattr(image_processing, "do_normalize"))
+            self.assertTrue(hasattr(image_processing, "image_mean"))
+            self.assertTrue(hasattr(image_processing, "image_std"))
+
+    def test_image_processor_from_dict_with_kwargs(self):
+        for image_processing_class in self.image_processing_classes.values():
+            image_processor = image_processing_class.from_dict(self.image_processor_dict)
+            self.assertEqual(image_processor.size, {"height": 18, "width": 18})
+
+            image_processor = image_processing_class.from_dict(
+                self.image_processor_dict, size={"height": 42, "width": 42}
+            )
+            self.assertEqual(image_processor.size, {"height": 42, "width": 42})
+
+    @slow
+    def test_image_processor_integration_test(self):
+        for image_processing_class in self.image_processing_classes.values():
+            processor = image_processing_class()
+
+            image = Image.open("./tests/fixtures/tests_samples/COCO/000000039769.png")
+            pixel_values = processor(image, return_tensors="pt").pixel_values
+
+            mean_value = round(pixel_values.mean().item(), 4)
+            self.assertEqual(mean_value, -0.2303)
+
+    @slow
+    def test_image_processor_integration_test_resize(self):
+        for backend_name in self.image_processing_classes.keys():
+            checkpoint = "google/owlv2-base-patch16-ensemble"
+            processor = AutoProcessor.from_pretrained(checkpoint, backend=backend_name)
+            model = Owlv2ForObjectDetection.from_pretrained(checkpoint)
+
+            image = Image.open("./tests/fixtures/tests_samples/COCO/000000039769.png")
+            text = ["cat"]
+            target_size = image.size[::-1]
+            expected_boxes = torch.tensor(
+                [
+                    [341.66656494140625, 23.38756561279297, 642.321044921875, 371.3482971191406],
+                    [6.753320693969727, 51.96149826049805, 326.61810302734375, 473.12982177734375],
+                ]
+            )
+
+            # single image
+            inputs = processor(text=[text], images=[image], return_tensors="pt")
+            with torch.no_grad():
+                outputs = model(**inputs)
+
+            results = processor.post_process_grounded_object_detection(
+                outputs, threshold=0.2, target_sizes=[target_size]
+            )[0]
+
+            boxes = results["boxes"]
+            torch.testing.assert_close(boxes, expected_boxes, atol=1e-1, rtol=1e-1)
+
+            # batch of images
+            inputs = processor(text=[text, text], images=[image, image], return_tensors="pt")
+            with torch.no_grad():
+                outputs = model(**inputs)
+            results = processor.post_process_grounded_object_detection(
+                outputs, threshold=0.2, target_sizes=[target_size, target_size]
+            )
+
+            for result in results:
+                boxes = result["boxes"]
+                torch.testing.assert_close(boxes, expected_boxes, atol=1e-1, rtol=1e-1)
+
+    @unittest.skip(reason="OWLv2 doesn't treat 4 channel PIL and numpy consistently yet")  # FIXME Amy
+    def test_call_numpy_4_channels(self):
+        pass

tests/models/owlv2/test_modeling_owlv2.py

@@ -0,0 +1,887 @@
+# 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 Owlv2 model."""
+
+import inspect
+import tempfile
+import unittest
+
+import numpy as np
+import requests
+from parameterized import parameterized
+
+from transformers import Owlv2Config, Owlv2TextConfig, Owlv2VisionConfig
+from transformers.testing_utils import (
+    require_torch,
+    require_torch_accelerator,
+    require_torch_fp16,
+    require_vision,
+    slow,
+    torch_device,
+)
+from transformers.utils import is_torch_available, is_vision_available
+
+from ...test_configuration_common import ConfigTester
+from ...test_modeling_common import (
+    TEST_EAGER_MATCHES_SDPA_INFERENCE_PARAMETERIZATION,
+    ModelTesterMixin,
+    floats_tensor,
+    ids_tensor,
+    random_attention_mask,
+)
+from ...test_pipeline_mixin import PipelineTesterMixin
+
+
+if is_torch_available():
+    import torch
+    from torch import nn
+
+    from transformers import Owlv2ForObjectDetection, Owlv2Model, Owlv2TextModel, Owlv2VisionModel
+
+
+if is_vision_available():
+    from PIL import Image
+
+    from transformers import OwlViTImageProcessorPil, OwlViTProcessor
+
+
+# Copied from tests.models.owlvit.test_modeling_owlvit.OwlViTVisionModelTester with OwlViT->Owlv2
+class Owlv2VisionModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=12,
+        image_size=32,
+        patch_size=2,
+        num_channels=3,
+        is_training=True,
+        hidden_size=32,
+        num_hidden_layers=2,
+        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
+
+        # in ViT, the seq length equals the number of patches + 1 (we add 1 for the [CLS] token)
+        num_patches = (image_size // patch_size) ** 2
+        self.seq_length = num_patches + 1
+
+    def prepare_config_and_inputs(self):
+        pixel_values = floats_tensor([self.batch_size, self.num_channels, self.image_size, self.image_size])
+        config = self.get_config()
+
+        return config, pixel_values
+
+    def get_config(self):
+        return Owlv2VisionConfig(
+            image_size=self.image_size,
+            patch_size=self.patch_size,
+            num_channels=self.num_channels,
+            hidden_size=self.hidden_size,
+            num_hidden_layers=self.num_hidden_layers,
+            num_attention_heads=self.num_attention_heads,
+            intermediate_size=self.intermediate_size,
+            dropout=self.dropout,
+            attention_dropout=self.attention_dropout,
+            initializer_range=self.initializer_range,
+        )
+
+    def create_and_check_model(self, config, pixel_values):
+        model = Owlv2VisionModel(config=config).to(torch_device)
+        model.eval()
+
+        pixel_values = pixel_values.to(torch.float32)
+
+        with torch.no_grad():
+            result = model(pixel_values)
+        # expected sequence length = num_patches + 1 (we add 1 for the [CLS] token)
+        num_patches = (self.image_size // self.patch_size) ** 2
+        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
+# Copied from tests.models.owlvit.test_modeling_owlvit.OwlViTVisionModelTest with OwlViT->Owlv2, OWL-ViT->OwlV2, OWLVIT->OWLV2, owlvit-base-patch32->owlv2-base-patch16-ensemble
+class Owlv2VisionModelTest(ModelTesterMixin, unittest.TestCase):
+    """
+    Here we also overwrite some of the tests of test_modeling_common.py, as OWLV2 does not use input_ids, inputs_embeds,
+    attention_mask and seq_length.
+    """
+
+    all_model_classes = (Owlv2VisionModel,) if is_torch_available() else ()
+
+    test_resize_embeddings = False
+
+    def setUp(self):
+        self.model_tester = Owlv2VisionModelTester(self)
+        self.config_tester = ConfigTester(
+            self, config_class=Owlv2VisionConfig, has_text_modality=False, hidden_size=32
+        )
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    @unittest.skip(reason="OWLV2 does not use inputs_embeds")
+    def test_inputs_embeds(self):
+        pass
+
+    def test_model_get_set_embeddings(self):
+        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            model = model_class(config)
+            self.assertIsInstance(model.get_input_embeddings(), (nn.Module))
+            x = model.get_output_embeddings()
+            self.assertTrue(x is None or isinstance(x, nn.Linear))
+
+    def test_forward_signature(self):
+        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            model = model_class(config)
+            signature = inspect.signature(model.forward)
+            # signature.parameters is an OrderedDict => so arg_names order is deterministic
+            arg_names = [*signature.parameters.keys()]
+
+            expected_arg_names = ["pixel_values"]
+            self.assertListEqual(arg_names[:1], expected_arg_names)
+
+    def test_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_model(*config_and_inputs)
+
+    @unittest.skip(reason="This module does not support standalone training")
+    def test_training(self):
+        pass
+
+    @unittest.skip(reason="This module does not support standalone training")
+    def test_training_gradient_checkpointing(self):
+        pass
+
+    @unittest.skip(reason="This module does not support standalone training")
+    def test_training_gradient_checkpointing_use_reentrant_false(self):
+        pass
+
+    @unittest.skip(reason="This module does not support standalone training")
+    def test_training_gradient_checkpointing_use_reentrant_true(self):
+        pass
+
+    @slow
+    def test_model_from_pretrained(self):
+        model_name = "google/owlv2-base-patch16-ensemble"
+        model = Owlv2VisionModel.from_pretrained(model_name)
+        self.assertIsNotNone(model)
+
+
+# Copied from tests.models.owlvit.test_modeling_owlvit.OwlViTTextModelTester with OwlViT->Owlv2
+class Owlv2TextModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=12,
+        num_queries=4,
+        seq_length=16,
+        is_training=True,
+        use_input_mask=True,
+        use_labels=True,
+        vocab_size=99,
+        hidden_size=64,
+        num_hidden_layers=12,
+        num_attention_heads=4,
+        intermediate_size=37,
+        dropout=0.1,
+        attention_dropout=0.1,
+        max_position_embeddings=16,
+        initializer_range=0.02,
+        scope=None,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.num_queries = num_queries
+        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.num_queries, self.seq_length], self.vocab_size)
+        input_mask = None
+
+        if self.use_input_mask:
+            input_mask = random_attention_mask([self.batch_size * self.num_queries, self.seq_length])
+
+        if input_mask is not None:
+            num_text, seq_length = input_mask.shape
+
+            rnd_start_indices = np.random.randint(1, seq_length - 1, size=(num_text,))
+            for idx, start_index in enumerate(rnd_start_indices):
+                input_mask[idx, :start_index] = 1
+                input_mask[idx, start_index:] = 0
+
+        config = self.get_config()
+
+        return config, input_ids, input_mask
+
+    def get_config(self):
+        return Owlv2TextConfig(
+            vocab_size=self.vocab_size,
+            hidden_size=self.hidden_size,
+            num_hidden_layers=self.num_hidden_layers,
+            num_attention_heads=self.num_attention_heads,
+            intermediate_size=self.intermediate_size,
+            dropout=self.dropout,
+            attention_dropout=self.attention_dropout,
+            max_position_embeddings=self.max_position_embeddings,
+            initializer_range=self.initializer_range,
+        )
+
+    def create_and_check_model(self, config, input_ids, input_mask):
+        model = Owlv2TextModel(config=config).to(torch_device)
+        model.eval()
+        with torch.no_grad():
+            result = model(input_ids=input_ids, attention_mask=input_mask)
+
+        self.parent.assertEqual(
+            result.last_hidden_state.shape, (self.batch_size * self.num_queries, self.seq_length, self.hidden_size)
+        )
+        self.parent.assertEqual(result.pooler_output.shape, (self.batch_size * self.num_queries, 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
+# Copied from tests.models.owlvit.test_modeling_owlvit.OwlViTTextModelTest with OwlViT->Owlv2, OWL-ViT->OwlV2, OWLVIT->OWLV2, owlvit-base-patch32->owlv2-base-patch16-ensemble
+class Owlv2TextModelTest(ModelTesterMixin, unittest.TestCase):
+    all_model_classes = (Owlv2TextModel,) if is_torch_available() else ()
+
+    def setUp(self):
+        self.model_tester = Owlv2TextModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=Owlv2TextConfig, hidden_size=32)
+
+    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)
+
+    @unittest.skip(reason="This module does not support standalone training")
+    def test_training(self):
+        pass
+
+    @unittest.skip(reason="This module does not support standalone training")
+    def test_training_gradient_checkpointing(self):
+        pass
+
+    @unittest.skip(reason="This module does not support standalone training")
+    def test_training_gradient_checkpointing_use_reentrant_false(self):
+        pass
+
+    @unittest.skip(reason="This module does not support standalone training")
+    def test_training_gradient_checkpointing_use_reentrant_true(self):
+        pass
+
+    @unittest.skip(reason="OWLV2 does not use inputs_embeds")
+    def test_inputs_embeds(self):
+        pass
+
+    @slow
+    def test_model_from_pretrained(self):
+        model_name = "google/owlv2-base-patch16-ensemble"
+        model = Owlv2TextModel.from_pretrained(model_name)
+        self.assertIsNotNone(model)
+
+
+class Owlv2ModelTester:
+    def __init__(self, parent, text_kwargs=None, vision_kwargs=None, is_training=True):
+        if text_kwargs is None:
+            text_kwargs = {}
+        if vision_kwargs is None:
+            vision_kwargs = {}
+
+        self.parent = parent
+        self.text_model_tester = Owlv2TextModelTester(parent, **text_kwargs)
+        self.vision_model_tester = Owlv2VisionModelTester(parent, **vision_kwargs)
+        self.is_training = is_training
+        self.text_config = self.text_model_tester.get_config().to_dict()
+        self.vision_config = self.vision_model_tester.get_config().to_dict()
+        self.batch_size = self.text_model_tester.batch_size  # need bs for batching_equivalence test
+
+    def prepare_config_and_inputs(self):
+        text_config, input_ids, attention_mask = self.text_model_tester.prepare_config_and_inputs()
+        vision_config, pixel_values = self.vision_model_tester.prepare_config_and_inputs()
+        config = self.get_config()
+        return config, input_ids, attention_mask, pixel_values
+
+    def get_config(self):
+        return Owlv2Config(
+            text_config=self.text_config,
+            vision_config=self.vision_config,
+            projection_dim=64,
+        )
+
+    def create_and_check_model(self, config, input_ids, attention_mask, pixel_values):
+        model = Owlv2Model(config).to(torch_device).eval()
+
+        with torch.no_grad():
+            result = model(
+                input_ids=input_ids,
+                pixel_values=pixel_values,
+                attention_mask=attention_mask,
+            )
+
+        image_logits_size = (
+            self.vision_model_tester.batch_size,
+            self.text_model_tester.batch_size * self.text_model_tester.num_queries,
+        )
+        text_logits_size = (
+            self.text_model_tester.batch_size * self.text_model_tester.num_queries,
+            self.vision_model_tester.batch_size,
+        )
+        self.parent.assertEqual(result.logits_per_image.shape, image_logits_size)
+        self.parent.assertEqual(result.logits_per_text.shape, text_logits_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 = {
+            "pixel_values": pixel_values,
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+            "return_loss": False,
+        }
+        return config, inputs_dict
+
+
+@require_torch
+# Copied from tests.models.owlvit.test_modeling_owlvit.OwlViTModelTest with OwlViT->Owlv2, OWL-ViT->OwlV2, OWLVIT->OWLV2, owlvit-base-patch32->owlv2-base-patch16-ensemble
+class Owlv2ModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
+    all_model_classes = (Owlv2Model,) if is_torch_available() else ()
+    pipeline_model_mapping = (
+        {
+            "feature-extraction": Owlv2Model,
+            "zero-shot-object-detection": Owlv2ForObjectDetection,
+        }
+        if is_torch_available()
+        else {}
+    )
+
+    test_resize_embeddings = False
+    test_attention_outputs = False
+    additional_model_inputs = ["pixel_values"]
+    _is_composite = True
+
+    def setUp(self):
+        self.model_tester = Owlv2ModelTester(self)
+        common_properties = ["projection_dim", "logit_scale_init_value"]
+        self.config_tester = ConfigTester(
+            self, config_class=Owlv2Config, has_text_modality=False, common_properties=common_properties
+        )
+
+    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)
+
+    @unittest.skip(reason="Hidden_states is tested in individual model tests")
+    def test_hidden_states_output(self):
+        pass
+
+    @unittest.skip(reason="Inputs_embeds is tested in individual model tests")
+    def test_inputs_embeds(self):
+        pass
+
+    @unittest.skip(reason="Retain_grad is tested in individual model tests")
+    def test_retain_grad_hidden_states_attentions(self):
+        pass
+
+    @unittest.skip(reason="Owlv2Model does not have input/output embeddings")
+    def test_model_get_set_embeddings(self):
+        pass
+
+    def test_load_vision_text_config(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        # Save Owlv2Config and check if we can load Owlv2VisionConfig from it
+        with tempfile.TemporaryDirectory() as tmp_dir_name:
+            config.save_pretrained(tmp_dir_name)
+            vision_config = Owlv2VisionConfig.from_pretrained(tmp_dir_name)
+            self.assertDictEqual(config.vision_config.to_dict(), vision_config.to_dict())
+
+        # Save Owlv2Config and check if we can load Owlv2TextConfig from it
+        with tempfile.TemporaryDirectory() as tmp_dir_name:
+            config.save_pretrained(tmp_dir_name)
+            text_config = Owlv2TextConfig.from_pretrained(tmp_dir_name)
+            self.assertDictEqual(config.text_config.to_dict(), text_config.to_dict())
+
+    @slow
+    def test_model_from_pretrained(self):
+        model_name = "google/owlv2-base-patch16-ensemble"
+        model = Owlv2Model.from_pretrained(model_name)
+        self.assertIsNotNone(model)
+
+
+# Copied from tests.models.owlvit.test_modeling_owlvit.OwlViTForObjectDetectionTester with OwlViT->Owlv2, OWL-ViT->OwlV2, OWLVIT->OWLV2
+class Owlv2ForObjectDetectionTester:
+    def __init__(self, parent, is_training=True):
+        self.parent = parent
+        self.text_model_tester = Owlv2TextModelTester(parent)
+        self.vision_model_tester = Owlv2VisionModelTester(parent)
+        self.is_training = is_training
+        self.text_config = self.text_model_tester.get_config().to_dict()
+        self.vision_config = self.vision_model_tester.get_config().to_dict()
+        self.batch_size = self.text_model_tester.batch_size  # need bs for batching_equivalence test
+
+    def prepare_config_and_inputs(self):
+        text_config, input_ids, attention_mask = self.text_model_tester.prepare_config_and_inputs()
+        vision_config, pixel_values = self.vision_model_tester.prepare_config_and_inputs()
+        config = self.get_config()
+        return config, pixel_values, input_ids, attention_mask
+
+    def get_config(self):
+        return Owlv2Config(
+            text_config=self.text_config,
+            vision_config=self.vision_config,
+            projection_dim=64,
+        )
+
+    def create_and_check_model(self, config, pixel_values, input_ids, attention_mask):
+        model = Owlv2ForObjectDetection(config).to(torch_device).eval()
+        with torch.no_grad():
+            result = model(
+                pixel_values=pixel_values,
+                input_ids=input_ids,
+                attention_mask=attention_mask,
+                return_dict=True,
+            )
+
+        pred_boxes_size = (
+            self.vision_model_tester.batch_size,
+            (self.vision_model_tester.image_size // self.vision_model_tester.patch_size) ** 2,
+            4,
+        )
+        pred_logits_size = (
+            self.vision_model_tester.batch_size,
+            (self.vision_model_tester.image_size // self.vision_model_tester.patch_size) ** 2,
+            4,
+        )
+        pred_class_embeds_size = (
+            self.vision_model_tester.batch_size,
+            (self.vision_model_tester.image_size // self.vision_model_tester.patch_size) ** 2,
+            self.text_model_tester.hidden_size,
+        )
+        self.parent.assertEqual(result.pred_boxes.shape, pred_boxes_size)
+        self.parent.assertEqual(result.logits.shape, pred_logits_size)
+        self.parent.assertEqual(result.class_embeds.shape, pred_class_embeds_size)
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        config, pixel_values, input_ids, attention_mask = config_and_inputs
+        inputs_dict = {
+            "pixel_values": pixel_values,
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+        }
+        return config, inputs_dict
+
+
+@require_torch
+# Copied from tests.models.owlvit.test_modeling_owlvit.OwlViTForObjectDetectionTest with OwlViT->Owlv2, OWL-ViT->OwlV2, OWLVIT->OWLV2, owlvit-base-patch32->owlv2-base-patch16-ensemble
+class Owlv2ForObjectDetectionTest(ModelTesterMixin, unittest.TestCase):
+    all_model_classes = (Owlv2ForObjectDetection,) if is_torch_available() else ()
+
+    test_resize_embeddings = False
+    test_attention_outputs = False
+
+    additional_model_inputs = ["pixel_values", "attention_mask"]
+
+    def setUp(self):
+        self.model_tester = Owlv2ForObjectDetectionTester(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)
+
+    @unittest.skip(reason="Hidden_states is tested in individual model tests")
+    def test_hidden_states_output(self):
+        pass
+
+    @parameterized.expand(TEST_EAGER_MATCHES_SDPA_INFERENCE_PARAMETERIZATION)
+    def test_eager_matches_sdpa_inference(self, *args):
+        self.skipTest("Owlv2ObjectDetectionOutput has no top-level hidden_states; SDPA tested in sub-models")
+
+    @unittest.skip(reason="Inputs_embeds is tested in individual model tests")
+    def test_inputs_embeds(self):
+        pass
+
+    @unittest.skip(reason="Retain_grad is tested in individual model tests")
+    def test_retain_grad_hidden_states_attentions(self):
+        pass
+
+    @unittest.skip(reason="Owlv2Model does not have input/output embeddings")
+    def test_model_get_set_embeddings(self):
+        pass
+
+    @unittest.skip(reason="Test_forward_signature is tested in individual model tests")
+    def test_forward_signature(self):
+        pass
+
+    @unittest.skip(reason="This module does not support standalone training")
+    def test_training(self):
+        pass
+
+    @unittest.skip(reason="This module does not support standalone training")
+    def test_training_gradient_checkpointing(self):
+        pass
+
+    @unittest.skip(reason="This module does not support standalone training")
+    def test_training_gradient_checkpointing_use_reentrant_false(self):
+        pass
+
+    @unittest.skip(reason="This module does not support standalone training")
+    def test_training_gradient_checkpointing_use_reentrant_true(self):
+        pass
+
+    @slow
+    def test_model_from_pretrained(self):
+        model_name = "google/owlv2-base-patch16-ensemble"
+        model = Owlv2ForObjectDetection.from_pretrained(model_name)
+        self.assertIsNotNone(model)
+
+
+# We will verify our results on an image of cute cats
+def prepare_img():
+    url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+    im = Image.open(requests.get(url, stream=True).raw)
+    return im
+
+
+@require_vision
+@require_torch
+class Owlv2ModelIntegrationTest(unittest.TestCase):
+    @slow
+    def test_inference(self):
+        model_name = "google/owlv2-base-patch16"
+        model = Owlv2Model.from_pretrained(model_name).to(torch_device)
+        image_processor = OwlViTImageProcessorPil.from_pretrained(model_name)
+        processor = OwlViTProcessor.from_pretrained(model_name, image_processor=image_processor)
+
+        image = prepare_img()
+        inputs = processor(
+            text=[["a photo of a cat", "a photo of a dog"]],
+            images=image,
+            max_length=16,
+            padding="max_length",
+            return_tensors="pt",
+        ).to(torch_device)
+
+        # forward pass
+        with torch.no_grad():
+            outputs = model(**inputs)
+
+        # verify the logits
+        self.assertEqual(
+            outputs.logits_per_image.shape,
+            torch.Size((inputs.pixel_values.shape[0], inputs.input_ids.shape[0])),
+        )
+        self.assertEqual(
+            outputs.logits_per_text.shape,
+            torch.Size((inputs.input_ids.shape[0], inputs.pixel_values.shape[0])),
+        )
+        expected_logits = torch.tensor([[-6.2229, -8.2601]], device=torch_device)
+        torch.testing.assert_close(outputs.logits_per_image, expected_logits, rtol=1e-3, atol=1e-3)
+
+    @slow
+    def test_inference_interpolate_pos_encoding(self):
+        model_name = "google/owlv2-base-patch16"
+        model = Owlv2Model.from_pretrained(model_name).to(torch_device)
+        image_processor = OwlViTImageProcessorPil.from_pretrained(model_name)
+        processor = OwlViTProcessor.from_pretrained(model_name, image_processor=image_processor)
+        processor.image_processor.size = {"height": 1024, "width": 1024}
+
+        image = prepare_img()
+        inputs = processor(
+            text=[["a photo of a cat", "a photo of a dog"]],
+            images=image,
+            max_length=16,
+            padding="max_length",
+            return_tensors="pt",
+        ).to(torch_device)
+
+        # forward pass
+        with torch.no_grad():
+            outputs = model(**inputs, interpolate_pos_encoding=True)
+
+        # 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([[-6.2520, -8.2970]], device=torch_device)
+        torch.testing.assert_close(outputs.logits_per_image, expected_logits, rtol=1e-3, atol=1e-3)
+        expected_shape = torch.Size((1, 4097, 768))
+        self.assertEqual(outputs.vision_model_output.last_hidden_state.shape, expected_shape)
+
+        # Owlv2ForObjectDetection part.
+        model = Owlv2ForObjectDetection.from_pretrained(model_name).to(torch_device)
+        processor.image_processor.size = {"height": 1024, "width": 1024}
+
+        with torch.no_grad():
+            outputs = model(**inputs, interpolate_pos_encoding=True)
+
+        num_queries = int((inputs.pixel_values.shape[-1] / model.config.vision_config.patch_size) ** 2)
+        self.assertEqual(outputs.pred_boxes.shape, torch.Size((1, num_queries, 4)))
+        expected_slice_boxes = torch.tensor(
+            [[0.2407, 0.0553, 0.4636], [0.1082, 0.0494, 0.1861], [0.2459, 0.0527, 0.4398]]
+        ).to(torch_device)
+        torch.testing.assert_close(outputs.pred_boxes[0, :3, :3], expected_slice_boxes, rtol=1e-4, atol=1e-4)
+
+        model = Owlv2ForObjectDetection.from_pretrained(model_name).to(torch_device)
+        query_image = prepare_img()
+        inputs = processor(
+            images=image,
+            query_images=query_image,
+            max_length=16,
+            padding="max_length",
+            return_tensors="pt",
+        ).to(torch_device)
+
+        with torch.no_grad():
+            outputs = model.image_guided_detection(**inputs, interpolate_pos_encoding=True)
+
+        # No need to check the logits, we just check inference runs fine.
+        num_queries = int((inputs.pixel_values.shape[-1] / model.config.vision_config.patch_size) ** 2)
+        self.assertEqual(outputs.target_pred_boxes.shape, torch.Size((1, num_queries, 4)))
+
+        # Deactivate interpolate_pos_encoding on same model, and use default image size.
+        # Verify the dynamic change caused by the activation/deactivation of interpolate_pos_encoding of variables: self.sqrt_num_patches, self.box_bias from (OwlViTForObjectDetection).
+        image_processor = OwlViTImageProcessorPil.from_pretrained(model_name)
+        processor = OwlViTProcessor.from_pretrained(model_name, image_processor=image_processor)
+
+        image = prepare_img()
+        inputs = processor(
+            text=[["a photo of a cat", "a photo of a dog"]],
+            images=image,
+            max_length=16,
+            padding="max_length",
+            return_tensors="pt",
+        ).to(torch_device)
+
+        with torch.no_grad():
+            outputs = model(**inputs, interpolate_pos_encoding=False)
+
+        num_queries = int((inputs.pixel_values.shape[-1] // model.config.vision_config.patch_size) ** 2)
+        self.assertEqual(outputs.pred_boxes.shape, torch.Size((1, num_queries, 4)))
+
+        expected_default_box_bias = torch.tensor(
+            [
+                [-4.0717, -4.0717, -4.0717, -4.0717],
+                [-3.3644, -4.0717, -4.0717, -4.0717],
+                [-2.9425, -4.0717, -4.0717, -4.0717],
+            ]
+        ).to(torch_device)
+
+        torch.testing.assert_close(model.box_bias[:3, :4], expected_default_box_bias, rtol=1e-4, atol=1e-4)
+
+        # Interpolate with any resolution size.
+        processor.image_processor.size = {"height": 1264, "width": 1024}
+
+        image = prepare_img()
+        inputs = processor(
+            text=[["a photo of a cat", "a photo of a dog"]],
+            images=image,
+            max_length=16,
+            padding="max_length",
+            return_tensors="pt",
+        ).to(torch_device)
+
+        with torch.no_grad():
+            outputs = model(**inputs, interpolate_pos_encoding=True)
+
+        num_queries = int(
+            (inputs.pixel_values.shape[-2] // model.config.vision_config.patch_size)
+            * (inputs.pixel_values.shape[-1] // model.config.vision_config.patch_size)
+        )
+        self.assertEqual(outputs.pred_boxes.shape, torch.Size((1, num_queries, 4)))
+        expected_slice_boxes = torch.tensor(
+            [[0.2438, 0.0945, 0.4675], [0.1361, 0.0431, 0.2406], [0.2465, 0.0428, 0.4429]]
+        ).to(torch_device)
+        torch.testing.assert_close(outputs.pred_boxes[0, :3, :3], expected_slice_boxes, rtol=1e-4, atol=1e-4)
+
+        query_image = prepare_img()
+        inputs = processor(
+            images=image,
+            query_images=query_image,
+            max_length=16,
+            padding="max_length",
+            return_tensors="pt",
+        ).to(torch_device)
+
+        with torch.no_grad():
+            outputs = model.image_guided_detection(**inputs, interpolate_pos_encoding=True)
+
+        # No need to check the logits, we just check inference runs fine.
+        num_queries = int(
+            (inputs.pixel_values.shape[-2] // model.config.vision_config.patch_size)
+            * (inputs.pixel_values.shape[-1] // model.config.vision_config.patch_size)
+        )
+        self.assertEqual(outputs.target_pred_boxes.shape, torch.Size((1, num_queries, 4)))
+
+    @slow
+    def test_inference_object_detection(self):
+        model_name = "google/owlv2-base-patch16"
+        model = Owlv2ForObjectDetection.from_pretrained(model_name).to(torch_device)
+        image_processor = OwlViTImageProcessorPil.from_pretrained(model_name)
+        processor = OwlViTProcessor.from_pretrained(model_name, image_processor=image_processor)
+
+        image = prepare_img()
+        text_labels = [["a photo of a cat", "a photo of a dog"]]
+        inputs = processor(
+            text=text_labels,
+            images=image,
+            max_length=16,
+            padding="max_length",
+            return_tensors="pt",
+        ).to(torch_device)
+
+        with torch.no_grad():
+            outputs = model(**inputs)
+
+        num_queries = int((model.config.vision_config.image_size / model.config.vision_config.patch_size) ** 2)
+        self.assertEqual(outputs.pred_boxes.shape, torch.Size((1, num_queries, 4)))
+
+        expected_slice_logits = torch.tensor(
+            [[-21.413497, -21.612638], [-19.008193, -19.548841], [-20.958896, -21.382694]]
+        ).to(torch_device)
+        torch.testing.assert_close(outputs.logits[0, :3, :3], expected_slice_logits, rtol=1e-4, atol=1e-4)
+        expected_slice_boxes = torch.tensor(
+            [[0.241309, 0.051896, 0.453267], [0.139474, 0.045701, 0.250660], [0.233022, 0.050479, 0.427671]],
+        ).to(torch_device)
+        torch.testing.assert_close(outputs.pred_boxes[0, :3, :3], expected_slice_boxes, rtol=1e-4, atol=1e-4)
+        resulted_slice_boxes = outputs.pred_boxes[0, :3, :3]
+        max_diff = torch.max(torch.abs(resulted_slice_boxes - expected_slice_boxes)).item()
+        self.assertLess(max_diff, 3e-4)
+
+        # test post-processing
+        post_processed_output = processor.post_process_grounded_object_detection(outputs)
+        self.assertIsNone(post_processed_output[0]["text_labels"])
+
+        post_processed_output_with_text_labels = processor.post_process_grounded_object_detection(
+            outputs, text_labels=text_labels
+        )
+
+        objects_labels = post_processed_output_with_text_labels[0]["labels"].tolist()
+        self.assertListEqual(objects_labels, [0, 0])
+
+        objects_text_labels = post_processed_output_with_text_labels[0]["text_labels"]
+        self.assertIsNotNone(objects_text_labels)
+        self.assertListEqual(objects_text_labels, ["a photo of a cat", "a photo of a cat"])
+
+    @slow
+    def test_inference_one_shot_object_detection(self):
+        model_name = "google/owlv2-base-patch16"
+        model = Owlv2ForObjectDetection.from_pretrained(model_name).to(torch_device)
+        image_processor = OwlViTImageProcessorPil.from_pretrained(model_name)
+        processor = OwlViTProcessor.from_pretrained(model_name, image_processor=image_processor)
+
+        image = prepare_img()
+        query_image = prepare_img()
+        inputs = processor(
+            images=image,
+            query_images=query_image,
+            max_length=16,
+            padding="max_length",
+            return_tensors="pt",
+        ).to(torch_device)
+
+        with torch.no_grad():
+            outputs = model.image_guided_detection(**inputs)
+
+        num_queries = int((model.config.vision_config.image_size / model.config.vision_config.patch_size) ** 2)
+        self.assertEqual(outputs.target_pred_boxes.shape, torch.Size((1, num_queries, 4)))
+
+        expected_slice_boxes = torch.tensor(
+            [[0.2413, 0.0519, 0.4533], [0.1395, 0.0457, 0.2507], [0.2330, 0.0505, 0.4277]],
+        ).to(torch_device)
+        torch.testing.assert_close(outputs.target_pred_boxes[0, :3, :3], expected_slice_boxes, rtol=1e-4, atol=1e-4)
+
+    @slow
+    @require_torch_accelerator
+    @require_torch_fp16
+    def test_inference_one_shot_object_detection_fp16(self):
+        model_name = "google/owlv2-base-patch16"
+        model = Owlv2ForObjectDetection.from_pretrained(model_name, dtype=torch.float16).to(torch_device)
+
+        image_processor = OwlViTImageProcessorPil.from_pretrained(model_name)
+        processor = OwlViTProcessor.from_pretrained(model_name, image_processor=image_processor)
+
+        image = prepare_img()
+        query_image = prepare_img()
+        inputs = processor(
+            images=image,
+            query_images=query_image,
+            max_length=16,
+            padding="max_length",
+            return_tensors="pt",
+        ).to(torch_device)
+
+        with torch.no_grad():
+            outputs = model.image_guided_detection(**inputs)
+
+        # No need to check the logits, we just check inference runs fine.
+        num_queries = int((model.config.vision_config.image_size / model.config.vision_config.patch_size) ** 2)
+        self.assertEqual(outputs.target_pred_boxes.shape, torch.Size((1, num_queries, 4)))

tests/models/owlv2/test_processing_owlv2.py

@@ -0,0 +1,12 @@
+import unittest
+
+from transformers import Owlv2Processor
+from transformers.testing_utils import require_scipy
+
+from ...test_processing_common import ProcessorTesterMixin
+
+
+@require_scipy
+class Owlv2ProcessorTest(ProcessorTesterMixin, unittest.TestCase):
+    processor_class = Owlv2Processor
+    model_id = "google/owlv2-base-patch16-ensemble"