first commit

tests/models/eomt/test_image_processing_eomt.py

@@ -0,0 +1,323 @@
+# Copyright 2025 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 EoMT Image Processor."""
+
+import unittest
+
+import numpy as np
+from datasets import load_dataset
+
+from transformers.image_utils import load_image
+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
+from ...test_processing_common import url_to_local_path
+
+
+if is_torch_available():
+    import torch
+
+if is_vision_available():
+    from PIL import Image
+
+    from transformers.models.eomt.modeling_eomt import EomtForUniversalSegmentationOutput
+
+
+class EomtImageProcessingTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=7,
+        num_channels=3,
+        min_resolution=30,
+        max_resolution=400,
+        size=None,
+        do_resize=True,
+        do_pad=True,
+        do_normalize=True,
+        image_mean=[0.5, 0.5, 0.5],
+        image_std=[0.5, 0.5, 0.5],
+        num_labels=10,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.num_channels = num_channels
+        self.min_resolution = min_resolution
+        self.max_resolution = max_resolution
+        self.do_resize = do_resize
+        self.do_pad = do_pad
+        self.size = size if size is not None else {"shortest_edge": 18, "longest_edge": 18}
+        self.do_normalize = do_normalize
+        self.image_mean = image_mean
+        self.image_std = image_std
+        # for the post_process_functions
+        self.batch_size = 2
+        self.num_queries = 3
+        self.num_classes = 2
+        self.height = 18
+        self.width = 18
+        self.num_labels = num_labels
+
+    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,
+            "do_pad": self.do_pad,
+            "num_labels": self.num_labels,
+        }
+
+    def prepare_fake_eomt_outputs(self, batch_size, patch_offsets=None):
+        return EomtForUniversalSegmentationOutput(
+            masks_queries_logits=torch.randn((batch_size, self.num_queries, self.height, self.width)),
+            class_queries_logits=torch.randn((batch_size, self.num_queries, self.num_classes + 1)),
+            patch_offsets=patch_offsets,
+        )
+
+    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,
+        )
+
+
+def prepare_semantic_single_inputs():
+    ds = load_dataset("hf-internal-testing/fixtures_ade20k", split="test")
+    example = ds[0]
+    return example["image"], example["map"]
+
+
+def prepare_semantic_batch_inputs():
+    ds = load_dataset("hf-internal-testing/fixtures_ade20k", split="test")
+    return list(ds["image"][:2]), list(ds["map"][:2])
+
+
+@require_torch
+@require_vision
+class EomtImageProcessingTest(ImageProcessingTestMixin, unittest.TestCase):
+    def setUp(self):
+        super().setUp()
+        self.image_processor_tester = EomtImageProcessingTester(self)
+        self.model_id = "tue-mps/coco_panoptic_eomt_large_640"
+
+    @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, "image_mean"))
+            self.assertTrue(hasattr(image_processing, "image_std"))
+            self.assertTrue(hasattr(image_processing, "do_normalize"))
+            self.assertTrue(hasattr(image_processing, "do_resize"))
+            self.assertTrue(hasattr(image_processing, "size"))
+            self.assertTrue(hasattr(image_processing, "do_rescale"))
+            self.assertTrue(hasattr(image_processing, "rescale_factor"))
+            self.assertTrue(hasattr(image_processing, "resample"))
+
+    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, {"shortest_edge": 18, "longest_edge": 18})
+
+            image_processor = image_processing_class.from_dict(self.image_processor_dict, size=42)
+            self.assertEqual(image_processor.size, {"shortest_edge": 42})
+
+    def test_call_numpy(self):
+        for image_processing_class in self.image_processing_classes.values():
+            # Initialize image_processing
+            image_processing = image_processing_class(**self.image_processor_dict)
+            # create random numpy tensors
+            image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=True, numpify=True)
+            for image in image_inputs:
+                self.assertIsInstance(image, np.ndarray)
+
+            # Test not batched input
+            encoded_images = image_processing(image_inputs[0], return_tensors="pt").pixel_values
+            expected_output_image_shape = (1, 3, 18, 18)
+            self.assertEqual(tuple(encoded_images.shape), expected_output_image_shape)
+
+            # Test batched
+            encoded_images = image_processing(image_inputs, return_tensors="pt").pixel_values
+            expected_output_image_shape = (2, 3, 18, 18)
+            self.assertEqual(tuple(encoded_images.shape), expected_output_image_shape)
+
+    @unittest.skip(reason="Not supported")
+    def test_call_numpy_4_channels(self):
+        pass
+
+    def test_call_pil(self):
+        for image_processing_class in self.image_processing_classes.values():
+            image_processing = image_processing_class(**self.image_processor_dict)
+            image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=True)
+            for image in image_inputs:
+                self.assertIsInstance(image, Image.Image)
+
+            # Test Non batched input
+            encoded_images = image_processing(image_inputs[0], return_tensors="pt").pixel_values
+            expected_output_image_shape = (1, 3, 18, 18)
+            self.assertEqual(tuple(encoded_images.shape), expected_output_image_shape)
+
+            # Test batched
+            encoded_images = image_processing(image_inputs, return_tensors="pt").pixel_values
+            expected_output_image_shape = (2, 3, 18, 18)
+            self.assertEqual(tuple(encoded_images.shape), expected_output_image_shape)
+
+    def test_call_pytorch(self):
+        for image_processing_class in self.image_processing_classes.values():
+            image_processing = image_processing_class(**self.image_processor_dict)
+            image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=True, torchify=True)
+
+            for image in image_inputs:
+                self.assertIsInstance(image, torch.Tensor)
+
+            encoded_images = image_processing(image_inputs[0], return_tensors="pt").pixel_values
+            expected_output_image_shape = (1, 3, 18, 18)
+            self.assertEqual(tuple(encoded_images.shape), expected_output_image_shape)
+
+            encoded_images = image_processing(image_inputs, return_tensors="pt").pixel_values
+            expected_output_image_shape = (2, 3, 18, 18)
+            self.assertEqual(tuple(encoded_images.shape), expected_output_image_shape)
+
+    def test_backends_equivalence(self):
+        """Test equivalence across backends including segmentation maps."""
+        if len(self.image_processing_classes) < 2:
+            self.skipTest(reason="Skipping backends equivalence test as there are less than 2 backends")
+
+        dummy_image, dummy_map = prepare_semantic_single_inputs()
+
+        encodings = {}
+        for backend_name, image_processing_class in self.image_processing_classes.items():
+            image_processor = image_processing_class(**self.image_processor_dict)
+            encodings[backend_name] = image_processor(dummy_image, segmentation_maps=dummy_map, return_tensors="pt")
+
+        backend_names = list(encodings.keys())
+        reference_backend = backend_names[0]
+        reference_pixel_values = encodings[reference_backend].pixel_values
+        reference_mask_labels = encodings[reference_backend].mask_labels
+
+        for backend_name in backend_names[1:]:
+            self._assert_tensors_equivalence(
+                reference_pixel_values, encodings[backend_name].pixel_values, atol=1e-1, mean_atol=1e-3
+            )
+
+            # Check whether 99.9% of mask_labels values match or not.
+            match_ratio = (reference_mask_labels[0] == encodings[backend_name].mask_labels[0]).float().mean().item()
+            self.assertGreaterEqual(
+                match_ratio,
+                0.999,
+                f"Mask labels do not match between {reference_backend} and {backend_name} image processors.",
+            )
+
+    def test_slow_fast_equivalence_batched(self):
+        """Test batched equivalence across backends including segmentation maps."""
+        if len(self.image_processing_classes) < 2:
+            self.skipTest(reason="Skipping backends equivalence test as there are less than 2 backends")
+
+        if hasattr(self.image_processor_tester, "do_center_crop") and self.image_processor_tester.do_center_crop:
+            self.skipTest(
+                reason="Skipping as do_center_crop is True and center_crop functions are not equivalent for fast and slow processors"
+            )
+
+        dummy_images, dummy_maps = prepare_semantic_batch_inputs()
+
+        encodings = {}
+        for backend_name, image_processing_class in self.image_processing_classes.items():
+            image_processor = image_processing_class(**self.image_processor_dict)
+            encodings[backend_name] = image_processor(dummy_images, segmentation_maps=dummy_maps, return_tensors="pt")
+
+        backend_names = list(encodings.keys())
+        reference_backend = backend_names[0]
+        reference_pixel_values = encodings[reference_backend].pixel_values
+        reference_mask_labels = encodings[reference_backend].mask_labels
+
+        for backend_name in backend_names[1:]:
+            self._assert_tensors_equivalence(
+                reference_pixel_values, encodings[backend_name].pixel_values, atol=1e-1, mean_atol=1e-3
+            )
+
+            for idx in range(len(dummy_maps)):
+                match_ratio = (
+                    (reference_mask_labels[idx] == encodings[backend_name].mask_labels[idx]).float().mean().item()
+                )
+                self.assertGreaterEqual(
+                    match_ratio,
+                    0.999,
+                    f"Mask labels do not match between {reference_backend} and {backend_name} image processors.",
+                )
+
+    def test_post_process_semantic_segmentation(self):
+        for image_processing_class in self.image_processing_classes.values():
+            processor = image_processing_class(**self.image_processor_dict)
+            # Set longest_edge to None to test for semantic segmentatiom.
+            processor.size = {"shortest_edge": 18, "longest_edge": None}
+            image = load_image(url_to_local_path("http://images.cocodataset.org/val2017/000000039769.jpg"))
+
+            inputs = processor(images=image, do_split_image=True, return_tensors="pt")
+            patch_offsets = inputs["patch_offsets"]
+
+            target_sizes = [image.size[::-1]]
+
+            # For semantic segmentation, the BS of output is 2 coz, two patches are created for the image.
+            outputs = self.image_processor_tester.prepare_fake_eomt_outputs(
+                inputs["pixel_values"].shape[0], patch_offsets
+            )
+            segmentation = processor.post_process_semantic_segmentation(outputs, target_sizes)
+
+            self.assertEqual(segmentation[0].shape, (image.height, image.width))
+
+    def test_post_process_panoptic_segmentation(self):
+        for image_processing_class in self.image_processing_classes.values():
+            processor = image_processing_class(**self.image_processor_dict)
+            image = load_image(url_to_local_path("http://images.cocodataset.org/val2017/000000039769.jpg"))
+
+            original_sizes = [image.size[::-1], image.size[::-1]]
+
+            # lets test for batched input of 2
+            outputs = self.image_processor_tester.prepare_fake_eomt_outputs(2)
+            segmentation = processor.post_process_panoptic_segmentation(outputs, original_sizes)
+
+            self.assertTrue(len(segmentation) == 2)
+            for el in segmentation:
+                self.assertTrue("segmentation" in el)
+                self.assertTrue("segments_info" in el)
+                self.assertEqual(type(el["segments_info"]), list)
+                self.assertEqual(el["segmentation"].shape, (image.height, image.width))
+
+    def test_post_process_instance_segmentation(self):
+        for image_processing_class in self.image_processing_classes.values():
+            processor = image_processing_class(**self.image_processor_dict)
+            image = load_image(url_to_local_path("http://images.cocodataset.org/val2017/000000039769.jpg"))
+
+            original_sizes = [image.size[::-1], image.size[::-1]]
+
+            # lets test for batched input of 2
+            outputs = self.image_processor_tester.prepare_fake_eomt_outputs(2)
+            segmentation = processor.post_process_instance_segmentation(outputs, original_sizes)
+
+            self.assertTrue(len(segmentation) == 2)
+            for el in segmentation:
+                self.assertTrue("segmentation" in el)
+                self.assertTrue("segments_info" in el)
+                self.assertEqual(type(el["segments_info"]), list)
+                self.assertEqual(el["segmentation"].shape, (image.height, image.width))

tests/models/eomt/test_modeling_eomt.py

@@ -0,0 +1,377 @@
+# Copyright 2025 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 EoMT model."""
+
+import unittest
+
+import requests
+
+from transformers import AutoImageProcessor, EomtConfig, EomtForUniversalSegmentation, pipeline
+from transformers.testing_utils import require_torch, require_torch_accelerator, require_torch_fp16, 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
+
+
+if is_vision_available():
+    from PIL import Image
+
+
+class EomtForUniversalSegmentationTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=2,
+        is_training=True,
+        image_size=40,
+        patch_size=2,
+        num_queries=5,
+        num_register_tokens=19,
+        num_labels=4,
+        hidden_size=8,
+        num_attention_heads=2,
+        num_hidden_layers=2,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.is_training = is_training
+        self.num_queries = num_queries
+        self.image_size = image_size
+        self.patch_size = patch_size
+        self.num_labels = num_labels
+        self.hidden_size = hidden_size
+        self.num_attention_heads = num_attention_heads
+        self.num_hidden_layers = num_hidden_layers
+        self.num_register_tokens = num_register_tokens
+
+        num_patches = (image_size // patch_size) ** 2
+        self.seq_length = num_patches + 1 + self.num_register_tokens
+
+    def get_config(self):
+        config = {
+            "image_size": self.image_size,
+            "patch_size": self.patch_size,
+            "num_labels": self.num_labels,
+            "hidden_size": self.hidden_size,
+            "num_attention_heads": self.num_attention_heads,
+            "num_hidden_layers": self.num_hidden_layers,
+            "num_register_tokens": self.num_register_tokens,
+            "num_queries": self.num_queries,
+            "num_blocks": 1,
+        }
+        return EomtConfig(**config)
+
+    def prepare_config_and_inputs(self):
+        pixel_values = floats_tensor([self.batch_size, 3, self.image_size, self.image_size]).to(torch_device)
+
+        mask_labels = (
+            torch.rand([self.batch_size, self.num_labels, self.image_size, self.image_size], device=torch_device) > 0.5
+        ).float()
+        class_labels = (torch.rand((self.batch_size, self.num_labels), device=torch_device) > 0.5).long()
+
+        config = self.get_config()
+        return config, pixel_values, mask_labels, class_labels
+
+    def prepare_config_and_inputs_for_common(self):
+        config, pixel_values, mask_labels, class_labels = self.prepare_config_and_inputs()
+        inputs_dict = {"pixel_values": pixel_values}
+        return config, inputs_dict
+
+    def prepare_config_and_inputs_for_training(self):
+        config, pixel_values, mask_labels, class_labels = self.prepare_config_and_inputs()
+        inputs_dict = {"pixel_values": pixel_values, "mask_labels": mask_labels, "class_labels": class_labels}
+        return config, inputs_dict
+
+
+@require_torch
+class EomtForUniversalSegmentationTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
+    all_model_classes = (EomtForUniversalSegmentation,) if is_torch_available() else ()
+    pipeline_model_mapping = {"image-segmentation": EomtForUniversalSegmentation} if is_torch_available() else {}
+    is_encoder_decoder = False
+
+    test_missing_keys = False
+    test_torch_exportable = False
+
+    def setUp(self):
+        self.model_tester = EomtForUniversalSegmentationTester(self)
+        self.config_tester = ConfigTester(self, config_class=EomtConfig, has_text_modality=False)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_model_with_labels(self):
+        size = (self.model_tester.image_size,) * 2
+        inputs = {
+            "pixel_values": torch.randn((2, 3, *size), device=torch_device),
+            "mask_labels": torch.randn((2, 10, *size), device=torch_device),
+            "class_labels": torch.zeros(2, 10, device=torch_device).long(),
+        }
+        config = self.model_tester.get_config()
+
+        model = EomtForUniversalSegmentation(config).to(torch_device)
+        outputs = model(**inputs)
+        self.assertTrue(outputs.loss is not None)
+
+    @unittest.skip(reason="EoMT does not use inputs_embeds")
+    def test_inputs_embeds(self):
+        pass
+
+    @unittest.skip(reason="EoMT does not have a get_input_embeddings method")
+    def test_model_get_set_embeddings(self):
+        pass
+
+    @unittest.skip(reason="EoMT is not a generative model")
+    def test_generate_without_input_ids(self):
+        pass
+
+    @unittest.skip(reason="EoMT does not use token embeddings")
+    def test_resize_tokens_embeddings(self):
+        pass
+
+    def test_training(self):
+        # We override this test because EoMT requires `mask_labels` and `class_labels` for training,
+        # which are not standard labels that `_prepare_for_class` can generate. We can't include
+        # these labels in `prepare_config_and_inputs_for_common` because that would break determinism
+        # tests (the Hungarian matching in the loss computation is non-deterministic).
+        if not self.model_tester.is_training:
+            self.skipTest(reason="ModelTester is not configured to run training tests")
+
+        for model_class in self.all_model_classes:
+            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_training()
+            config.return_dict = True
+
+            model = model_class(config)
+            model.to(torch_device)
+            model.train()
+            inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
+            loss = model(**inputs).loss
+            loss.backward()
+
+
+@require_torch
+class EomtForUniversalSegmentationIntegrationTest(unittest.TestCase):
+    def setUp(self):
+        self.model_id = "tue-mps/coco_panoptic_eomt_large_640"
+
+    @slow
+    def test_inference(self):
+        model = EomtForUniversalSegmentation.from_pretrained(self.model_id, device_map="auto")
+        processor = AutoImageProcessor.from_pretrained(self.model_id)
+
+        image = Image.open(requests.get("http://images.cocodataset.org/val2017/000000039769.jpg", stream=True).raw)
+
+        inputs = processor(images=image, return_tensors="pt").to(model.device)
+
+        with torch.inference_mode():
+            outputs = model(**inputs)
+
+        self.assertTrue(outputs.class_queries_logits.shape == (1, 200, 134))
+        self.assertTrue(outputs.masks_queries_logits.shape == (1, 200, 160, 160))
+
+        # fmt: off
+        EXPECTED_SLICE = torch.tensor([
+            [ 13.2540,   8.9279,   8.6631,  12.3760,  10.1429],
+            [ -3.4815, -36.4630, -45.5604, -46.8404, -37.5099],
+            [ -6.8689, -44.4206, -62.7591, -59.2928, -47.7035],
+            [ -2.9380, -42.0659, -57.4382, -55.1537, -43.5142],
+            [ -8.4387, -38.5275, -53.1383, -47.0064, -38.9667],
+        ]).to(model.device)
+        # fmt: on
+
+        output_slice = outputs.masks_queries_logits[0, 0, :5, :5]
+        torch.testing.assert_close(output_slice, EXPECTED_SLICE, rtol=1e-2, atol=1e-2)
+
+        # fmt: off
+        EXPECTED_SLICE = torch.tensor([
+            [-0.6977, -6.4907, -4.1178, -6.5554, -6.6529],
+            [-0.3650, -6.6560, -4.0143, -6.5776, -6.5879],
+            [-0.8820, -6.7175, -3.5334, -6.8569, -6.2415],
+            [ 0.4502, -5.3911, -3.0232, -5.9411, -6.3243],
+            [ 0.3157, -5.6321, -2.6716, -5.5740, -5.5607],
+        ]).to(model.device)
+        # fmt: on
+
+        output_slice = outputs.class_queries_logits[0, :5, :5]
+        torch.testing.assert_close(output_slice, EXPECTED_SLICE, rtol=1e-2, atol=1e-2)
+
+    @require_torch_accelerator
+    @require_torch_fp16
+    @slow
+    def test_inference_fp16(self):
+        model = EomtForUniversalSegmentation.from_pretrained(self.model_id, dtype=torch.float16, device_map="auto")
+        processor = AutoImageProcessor.from_pretrained(self.model_id)
+
+        image = Image.open(requests.get("http://images.cocodataset.org/val2017/000000039769.jpg", stream=True).raw)
+
+        inputs = processor(images=image, return_tensors="pt").to(model.device)
+
+        with torch.inference_mode():
+            outputs = model(**inputs)
+
+        self.assertTrue(outputs.class_queries_logits.shape == (1, 200, 134))
+        self.assertTrue(outputs.masks_queries_logits.shape == (1, 200, 160, 160))
+
+    @slow
+    def test_semantic_segmentation_inference(self):
+        model_id = "tue-mps/ade20k_semantic_eomt_large_512"
+        model = EomtForUniversalSegmentation.from_pretrained(model_id, device_map="auto")
+        processor = AutoImageProcessor.from_pretrained(model_id)
+
+        image = Image.open(requests.get("http://images.cocodataset.org/val2017/000000039769.jpg", stream=True).raw)
+
+        inputs = processor(images=image, return_tensors="pt").to(model.device)
+
+        with torch.inference_mode():
+            outputs = model(**inputs)
+
+        self.assertTrue(outputs.class_queries_logits.shape == (2, 100, 151))
+        self.assertTrue(outputs.masks_queries_logits.shape == (2, 100, 128, 128))
+
+        preds = processor.post_process_semantic_segmentation(outputs, target_sizes=[(image.size[1], image.size[0])])[0]
+
+        self.assertTrue(preds.shape == (image.size[1], image.size[0]))
+
+        # fmt: off
+        EXPECTED_SLICE = torch.tensor([
+            [39, 39, 39, 39, 39, 39, 39, 39, 39, 39],
+            [39, 39, 39, 39, 39, 39, 39, 39, 39, 39],
+            [39, 39, 39, 39, 39, 39, 39, 39, 39, 39],
+            [39, 39, 39, 39, 39, 39, 39, 39, 39, 39],
+            [39, 39, 39, 39, 39, 39, 39, 39, 39, 39],
+            [39, 39, 39, 39, 39, 39, 39, 39, 39, 39],
+            [39, 39, 39, 39, 39, 39, 39, 39, 39, 39],
+            [39, 39, 39, 39, 39, 39, 39, 39, 39, 39],
+            [39, 39, 39, 39, 39, 39, 39, 39, 39, 39],
+            [39, 39, 39, 39, 39, 39, 39, 39, 39, 39]
+        ], device=model.device)
+        # fmt: on
+
+        output_slice = preds[:10, :10]
+        torch.testing.assert_close(output_slice, EXPECTED_SLICE, rtol=1e-2, atol=1e-2)
+
+    @slow
+    def test_panoptic_segmentation_inference(self):
+        model = EomtForUniversalSegmentation.from_pretrained(self.model_id, device_map="auto")
+        processor = AutoImageProcessor.from_pretrained(self.model_id)
+
+        image = Image.open(requests.get("http://images.cocodataset.org/val2017/000000039769.jpg", stream=True).raw)
+
+        inputs = processor(images=image, return_tensors="pt").to(model.device)
+
+        with torch.inference_mode():
+            outputs = model(**inputs)
+
+        self.assertTrue(outputs.class_queries_logits.shape == (1, 200, 134))
+        self.assertTrue(outputs.masks_queries_logits.shape == (1, 200, 160, 160))
+
+        preds = processor.post_process_panoptic_segmentation(outputs, target_sizes=[(image.size[1], image.size[0])])[0]
+        segmentation, segments_info = preds["segmentation"], preds["segments_info"]
+
+        # fmt: off
+        EXPECTED_SLICE = torch.tensor([
+            [-1, -1, -1, -1, -1, -1, -1, -1, -1, -1],
+            [-1, -1, -1, -1, -1, -1, -1, -1, -1, -1],
+            [-1, -1, -1, -1, -1, -1, -1, -1, -1, -1],
+            [-1, -1, -1, -1, -1,  2,  2,  2,  2,  2],
+            [-1, -1, -1,  2,  2,  2,  2,  2,  2,  2],
+            [ 2,  2,  2,  2,  2,  2,  2,  2,  2,  2],
+            [ 2,  2,  2,  2,  2,  2,  2,  2,  2,  2],
+            [ 2,  2,  2,  2,  2,  2,  2,  2,  2,  2],
+            [ 2,  2,  2,  2,  2,  2,  2,  2,  2,  2],
+            [ 2,  2,  2,  2,  2,  2,  2,  2,  2,  2]
+        ], device=model.device)
+
+        EXPECTED_SEGMENTS_INFO = [
+            {"id": 0, "label_id": 15, "score": 0.99935},
+            {"id": 1, "label_id": 15, "score": 0.998688},
+            {"id": 2, "label_id": 57, "score": 0.954325},
+            {"id": 3, "label_id": 65, "score": 0.997285},
+            {"id": 4, "label_id": 65, "score": 0.99711}
+        ]
+        # fmt: on
+
+        output_slice = segmentation[:10, :10]
+        torch.testing.assert_close(output_slice, EXPECTED_SLICE, rtol=1e-2, atol=1e-2)
+        for actual, expected in zip(segments_info, EXPECTED_SEGMENTS_INFO):
+            self.assertEqual(actual["id"], expected["id"])
+            self.assertEqual(actual["label_id"], expected["label_id"])
+            self.assertAlmostEqual(actual["score"], expected["score"], delta=1e-3)
+
+    @slow
+    def test_instance_segmentation_inference(self):
+        model_id = "tue-mps/coco_instance_eomt_large_640"
+        model = EomtForUniversalSegmentation.from_pretrained(model_id, device_map="auto")
+        processor = AutoImageProcessor.from_pretrained(model_id)
+
+        image = Image.open(requests.get("http://images.cocodataset.org/val2017/000000039769.jpg", stream=True).raw)
+
+        inputs = processor(images=image, return_tensors="pt").to(model.device)
+
+        with torch.inference_mode():
+            outputs = model(**inputs)
+
+        self.assertTrue(outputs.class_queries_logits.shape == (1, 200, 81))
+        self.assertTrue(outputs.masks_queries_logits.shape == (1, 200, 160, 160))
+
+        preds = processor.post_process_instance_segmentation(outputs, target_sizes=[(image.size[1], image.size[0])])[0]
+        segmentation, segments_info = preds["segmentation"], preds["segments_info"]
+
+        # fmt: off
+        EXPECTED_SLICE = torch.tensor([
+            [-1., -1., -1., -1., -1., -1., -1., -1., -1., -1.],
+            [-1., -1., -1., -1., -1., -1., -1., -1., -1., -1.],
+            [-1., -1., -1., -1., -1., -1., -1., -1., -1., -1.],
+            [-1., -1., -1.,  0.,  0.,  1.,  1.,  1.,  1.,  1.],
+            [ 0.,  0.,  1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.],
+            [ 1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.],
+            [ 1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.],
+            [ 1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.],
+            [ 1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.],
+            [ 1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.]
+        ], device=model.device)
+
+        EXPECTED_SEGMENTS_INFO = [
+            {'id': 0, 'label_id': 57, 'score': 0.871247},
+            {'id': 1, 'label_id': 57, 'score': 0.821225},
+            {'id': 2, 'label_id': 15, 'score': 0.976252},
+            {'id': 3, 'label_id': 65, 'score': 0.972960},
+            {'id': 4, 'label_id': 65, 'score': 0.981109},
+            {'id': 5, 'label_id': 15, 'score': 0.972689}
+        ]
+        # fmt: on
+
+        output_slice = segmentation[:10, :10]
+        torch.testing.assert_close(output_slice, EXPECTED_SLICE, rtol=1e-2, atol=1e-2)
+        for actual, expected in zip(segments_info, EXPECTED_SEGMENTS_INFO):
+            self.assertEqual(actual["id"], expected["id"])
+            self.assertEqual(actual["label_id"], expected["label_id"])
+            self.assertAlmostEqual(actual["score"], expected["score"], delta=1e-3)
+
+    @slow
+    def test_segmentation_pipeline(self):
+        image = Image.open(requests.get("http://images.cocodataset.org/val2017/000000039769.jpg", stream=True).raw)
+
+        pipe = pipeline(model=self.model_id, subtask="panoptic", device=torch_device)
+        output = pipe(image)
+
+        EXPECTED_OUTPUT_LABELS = ["cat", "cat", "couch", "remote", "remote"]
+
+        output_labels = [segment["label"] for segment in output]
+        self.assertEqual(output_labels, EXPECTED_OUTPUT_LABELS)