first commit

tests/models/pp_doclayout_v2/test_image_processing_pp_doclayout_v2.py

@@ -0,0 +1,87 @@
+# Copyright 2026 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
+
+from ...test_image_processing_common import ImageProcessingTestMixin, prepare_image_inputs
+
+
+class PPDocLayoutV2ImageProcessingTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=7,
+        num_channels=3,
+        min_resolution=30,
+        max_resolution=400,
+        do_resize=True,
+        size=None,
+        do_normalize=True,
+        image_mean=[0.0, 0.0, 0.0],
+        image_std=[1.0, 1.0, 1.0],
+    ):
+        size = size if size is not None else {"height": 40, "width": 40}
+        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.size = size
+        self.do_normalize = do_normalize
+        self.image_mean = image_mean
+        self.image_std = image_std
+
+    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 PPDocLayoutV2ImageProcessingTest(ImageProcessingTestMixin, unittest.TestCase):
+    def setUp(self):
+        super().setUp()
+        self.image_processor_tester = PPDocLayoutV2ImageProcessingTester(self)
+
+    @property
+    def image_processor_dict(self):
+        return self.image_processor_tester.prepare_image_processor_dict()
+
+    @unittest.skip(
+        reason="PPDocLayoutV2 uses antialias=False which is not supported for 4-channel images consistently"
+    )
+    def test_call_numpy_4_channels(self):
+        pass

tests/models/pp_doclayout_v2/test_modeling_pp_doclayout_v2.py

@@ -0,0 +1,658 @@
+# coding = utf-8
+# Copyright 2026 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 PP-DocLayoutV2 model."""
+
+import inspect
+import math
+import tempfile
+import unittest
+
+from parameterized import parameterized
+
+from transformers import (
+    PPDocLayoutV2Config,
+    PPDocLayoutV2ForObjectDetection,
+    PPDocLayoutV2ImageProcessor,
+    is_torch_available,
+)
+from transformers.image_utils import load_image
+from transformers.testing_utils import (
+    cleanup,
+    require_torch,
+    require_torch_accelerator,
+    require_vision,
+    slow,
+    torch_device,
+)
+
+from ...test_configuration_common import ConfigTester
+from ...test_modeling_common import ModelTesterMixin, floats_tensor
+from ...test_pipeline_mixin import PipelineTesterMixin
+from ...test_processing_common import url_to_local_path
+
+
+if is_torch_available():
+    import torch
+
+
+class PPDocLayoutV2ModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=3,
+        is_training=False,
+        n_targets=3,
+        num_labels=25,
+        initializer_range=0.01,
+        layer_norm_eps=1e-5,
+        batch_norm_eps=1e-5,
+        # backbone
+        backbone_config=None,
+        # encoder HybridEncoder
+        encoder_hidden_dim=32,
+        encoder_in_channels=[32, 32, 32],
+        feat_strides=[8, 16, 32],
+        encoder_layers=1,
+        encoder_ffn_dim=8,
+        encoder_attention_heads=2,
+        dropout=0.0,
+        activation_dropout=0.0,
+        encode_proj_layers=[2],
+        positional_encoding_temperature=10000,
+        encoder_activation_function="gelu",
+        activation_function="silu",
+        eval_size=None,
+        normalize_before=False,
+        # decoder PPDocLayoutV2Transformer
+        d_model=32,
+        num_queries=30,
+        decoder_in_channels=[32, 32, 32],
+        decoder_ffn_dim=8,
+        num_feature_levels=3,
+        decoder_n_points=4,
+        decoder_layers=2,
+        decoder_attention_heads=2,
+        decoder_activation_function="relu",
+        attention_dropout=0.0,
+        num_denoising=0,
+        label_noise_ratio=0.5,
+        box_noise_scale=1.0,
+        learn_initial_query=False,
+        anchor_image_size=None,
+        image_size=128,
+        disable_custom_kernels=True,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.num_channels = 3
+        self.is_training = is_training
+        self.n_targets = n_targets
+        self.initializer_range = initializer_range
+        self.layer_norm_eps = layer_norm_eps
+        self.batch_norm_eps = batch_norm_eps
+        self.backbone_config = backbone_config
+        self.encoder_hidden_dim = encoder_hidden_dim
+        self.encoder_in_channels = encoder_in_channels
+        self.feat_strides = feat_strides
+        self.num_labels = num_labels
+        self.encoder_layers = encoder_layers
+        self.encoder_ffn_dim = encoder_ffn_dim
+        self.encoder_attention_heads = encoder_attention_heads
+        self.dropout = dropout
+        self.activation_dropout = activation_dropout
+        self.encode_proj_layers = encode_proj_layers
+        self.positional_encoding_temperature = positional_encoding_temperature
+        self.encoder_activation_function = encoder_activation_function
+        self.activation_function = activation_function
+        self.eval_size = eval_size
+        self.normalize_before = normalize_before
+        self.d_model = d_model
+        self.num_queries = num_queries
+        self.decoder_in_channels = decoder_in_channels
+        self.decoder_ffn_dim = decoder_ffn_dim
+        self.num_feature_levels = num_feature_levels
+        self.decoder_n_points = decoder_n_points
+        self.decoder_layers = decoder_layers
+        self.decoder_attention_heads = decoder_attention_heads
+        self.decoder_activation_function = decoder_activation_function
+        self.attention_dropout = attention_dropout
+        self.num_denoising = num_denoising
+        self.label_noise_ratio = label_noise_ratio
+        self.box_noise_scale = box_noise_scale
+        self.learn_initial_query = learn_initial_query
+        self.anchor_image_size = anchor_image_size
+        self.image_size = image_size
+        self.disable_custom_kernels = disable_custom_kernels
+
+        self.encoder_seq_length = math.ceil(self.image_size / 32) * math.ceil(self.image_size / 32)
+
+    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):
+        hidden_sizes = [10, 20, 30, 40]
+        backbone_config = {
+            "model_type": "hgnet_v2",
+            "arch": "L",
+            "return_idx": [1, 2, 3],
+            "freeze_stem_only": True,
+            "freeze_at": 0,
+            "freeze_norm": True,
+            "hidden_sizes": [32, 32, 32, 32],
+            "stem_channels": [3, 32, 32],
+            "stage_in_channels": [32, 32, 32, 32],
+            "stage_mid_channels": [32, 32, 32, 32],
+            "stage_out_channels": [32, 32, 32, 32],
+            "lr_mult_list": [0, 0.05, 0.05, 0.05, 0.05],
+            "out_features": ["stage2", "stage3", "stage4"],
+        }
+        reading_order_config = {
+            "hidden_size": 16,
+            "num_attention_heads": 8,
+            "attention_probs_dropout_prob": 0.1,
+            "has_relative_attention_bias": False,
+            "has_spatial_attention_bias": True,
+            "layer_norm_eps": 1e-5,
+            "hidden_dropout_prob": 0.1,
+            "intermediate_size": 16,
+            "hidden_act": "gelu",
+            "num_hidden_layers": 6,
+            "rel_pos_bins": 32,
+            "max_rel_pos": 128,
+            "rel_2d_pos_bins": 64,
+            "max_rel_2d_pos": 256,
+            "num_labels": 510,
+            "max_position_embeddings": 514,
+            "max_2d_position_embeddings": 512,
+            "type_vocab_size": 1,
+            "vocab_size": 4,
+            "start_token_id": 0,
+            "pad_token_id": 1,
+            "end_token_id": 2,
+            "pred_token_id": 3,
+            "coordinate_size": 171,
+            "shape_size": 170,
+            "num_classes": 20,
+            "relation_bias_embed_dim": 16,
+            "relation_bias_temperature": 10000,
+            "relation_bias_scale": 100,
+            "relative_head_num": 1,
+            "relative_head_size": 64,
+        }
+        id2label = {
+            0: "abstract",
+            1: "algorithm",
+            2: "aside_text",
+            3: "chart",
+            4: "content",
+            5: "formula",
+            6: "doc_title",
+            7: "figure_title",
+            8: "footer",
+            9: "footer",
+            10: "footnote",
+            11: "formula_number",
+            12: "header",
+            13: "header",
+            14: "image",
+            15: "formula",
+            16: "number",
+            17: "paragraph_title",
+            18: "reference",
+            19: "reference_content",
+            20: "seal",
+            21: "table",
+            22: "text",
+            23: "text",
+            24: "vision_footnote",
+        }
+        num_labels = 25
+        class_thresholds = [
+            0.5,
+            0.5,
+            0.5,
+            0.5,
+            0.5,
+            0.4,
+            0.4,
+            0.5,
+            0.5,
+            0.5,
+            0.5,
+            0.5,
+            0.5,
+            0.5,
+            0.5,
+            0.4,
+            0.5,
+            0.4,
+            0.5,
+            0.5,
+            0.45,
+            0.5,
+            0.4,
+            0.4,
+            0.5,
+        ]
+        class_order = [4, 2, 14, 1, 5, 7, 8, 6, 11, 11, 9, 13, 10, 10, 1, 2, 3, 0, 2, 2, 12, 1, 2, 15, 6]
+        return PPDocLayoutV2Config(
+            backbone_config=backbone_config,
+            reading_order_config=reading_order_config,
+            encoder_hidden_dim=self.encoder_hidden_dim,
+            encoder_in_channels=hidden_sizes[1:],
+            feat_strides=self.feat_strides,
+            encoder_layers=self.encoder_layers,
+            encoder_ffn_dim=self.encoder_ffn_dim,
+            encoder_attention_heads=self.encoder_attention_heads,
+            dropout=self.dropout,
+            activation_dropout=self.activation_dropout,
+            encode_proj_layers=self.encode_proj_layers,
+            positional_encoding_temperature=self.positional_encoding_temperature,
+            encoder_activation_function=self.encoder_activation_function,
+            activation_function=self.activation_function,
+            eval_size=self.eval_size,
+            normalize_before=self.normalize_before,
+            d_model=self.d_model,
+            num_queries=self.num_queries,
+            decoder_in_channels=self.decoder_in_channels,
+            decoder_ffn_dim=self.decoder_ffn_dim,
+            num_feature_levels=self.num_feature_levels,
+            decoder_n_points=self.decoder_n_points,
+            decoder_layers=self.decoder_layers,
+            decoder_attention_heads=self.decoder_attention_heads,
+            decoder_activation_function=self.decoder_activation_function,
+            attention_dropout=self.attention_dropout,
+            num_denoising=self.num_denoising,
+            label_noise_ratio=self.label_noise_ratio,
+            box_noise_scale=self.box_noise_scale,
+            learn_initial_query=self.learn_initial_query,
+            anchor_image_size=self.anchor_image_size,
+            image_size=self.image_size,
+            disable_custom_kernels=self.disable_custom_kernels,
+            id2label=id2label,
+            num_labels=num_labels,
+            class_thresholds=class_thresholds,
+            class_order=class_order,
+        )
+
+    def prepare_config_and_inputs_for_common(self):
+        config, pixel_values = self.prepare_config_and_inputs()
+        inputs_dict = {"pixel_values": pixel_values}
+        return config, inputs_dict
+
+    def create_and_check_pp_doclayout_v2_object_detection_head_model(self, config, pixel_values):
+        model = PPDocLayoutV2ForObjectDetection(config=config)
+        model.to(torch_device)
+        model.eval()
+
+        result = model(pixel_values)
+
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_queries, self.num_labels))
+        self.parent.assertEqual(result.pred_boxes.shape, (self.batch_size, self.num_queries, 4))
+
+
+@require_torch
+class PPDocLayoutV2ModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
+    all_model_classes = (PPDocLayoutV2ForObjectDetection,) if is_torch_available() else ()
+    pipeline_model_mapping = {"object-detection": PPDocLayoutV2ForObjectDetection} if is_torch_available() else {}
+    is_encoder_decoder = True
+
+    test_missing_keys = False
+    test_inputs_embeds = False
+    test_resize_embeddings = False
+
+    def setUp(self):
+        self.model_tester = PPDocLayoutV2ModelTester(self)
+        self.config_tester = ConfigTester(
+            self,
+            config_class=PPDocLayoutV2Config,
+            has_text_modality=False,
+        )
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_pp_doclayout_v2_object_detection_head_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_pp_doclayout_v2_object_detection_head_model(*config_and_inputs)
+
+    @unittest.skip(reason="PPDocLayoutV2 does not support input and output embeddings")
+    def test_model_get_set_embeddings(self):
+        pass
+
+    @unittest.skip(reason="PPDocLayoutV2 does not support input and output embeddings")
+    def test_model_common_attributes(self):
+        pass
+
+    @unittest.skip(reason="Feed forward chunking is not implemented")
+    def test_feed_forward_chunking(self):
+        pass
+
+    @unittest.skip(reason="PPDocLayoutV2 does not support training")
+    def test_retain_grad_hidden_states_attentions(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)
+            arg_names = [*signature.parameters.keys()]
+            expected_arg_names = ["pixel_values"]
+            self.assertListEqual(arg_names[:1], expected_arg_names)
+
+    @parameterized.expand(["float32", "float16", "bfloat16"])
+    @require_torch_accelerator
+    @slow
+    def test_inference_with_different_dtypes(self, dtype_str):
+        dtype = {
+            "float32": torch.float32,
+            "float16": torch.float16,
+            "bfloat16": torch.bfloat16,
+        }[dtype_str]
+
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            model = model_class(config)
+            model.to(torch_device).to(dtype)
+            model.eval()
+            for key, tensor in inputs_dict.items():
+                inputs_dict[key] = tensor.to(dtype)
+            with torch.no_grad():
+                _ = model(**self._prepare_for_class(inputs_dict, model_class))
+
+    @parameterized.expand(["float32", "float16", "bfloat16"])
+    @require_torch_accelerator
+    @slow
+    def test_inference_equivalence_for_static_and_dynamic_anchors(self, dtype_str):
+        dtype = {
+            "float32": torch.float32,
+            "float16": torch.float16,
+            "bfloat16": torch.bfloat16,
+        }[dtype_str]
+
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+        h, w = inputs_dict["pixel_values"].shape[-2:]
+
+        # convert inputs to the desired dtype
+        for key, tensor in inputs_dict.items():
+            inputs_dict[key] = tensor.to(dtype)
+
+        for model_class in self.all_model_classes:
+            with tempfile.TemporaryDirectory() as tmpdirname:
+                model_class(config).save_pretrained(tmpdirname)
+                model_static = model_class.from_pretrained(
+                    tmpdirname, anchor_image_size=[h, w], device_map=torch_device, dtype=dtype
+                ).eval()
+                model_dynamic = model_class.from_pretrained(
+                    tmpdirname, anchor_image_size=None, device_map=torch_device, dtype=dtype
+                ).eval()
+
+            self.assertIsNotNone(model_static.config.anchor_image_size)
+            self.assertIsNone(model_dynamic.config.anchor_image_size)
+
+            with torch.no_grad():
+                outputs_static = model_static(**self._prepare_for_class(inputs_dict, model_class))
+                outputs_dynamic = model_dynamic(**self._prepare_for_class(inputs_dict, model_class))
+
+            self.assertTrue(
+                torch.allclose(outputs_static.logits, outputs_dynamic.logits, rtol=1e-4, atol=1e-4),
+                f"Max diff: {(outputs_static.logits - outputs_dynamic.logits).abs().max()}",
+            )
+
+    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", len(self.model_tester.encoder_in_channels) - 1
+            )
+            self.assertEqual(len(hidden_states), expected_num_layers)
+
+            self.assertListEqual(
+                list(hidden_states[1].shape[-2:]),
+                [
+                    self.model_tester.image_size // self.model_tester.feat_strides[-1],
+                    self.model_tester.image_size // self.model_tester.feat_strides[-1],
+                ],
+            )
+
+            if config.is_encoder_decoder:
+                hidden_states = outputs.decoder_hidden_states
+
+                expected_num_layers = getattr(
+                    self.model_tester, "expected_num_hidden_layers", self.model_tester.decoder_layers + 1
+                )
+
+                self.assertIsInstance(hidden_states, (list, tuple))
+                self.assertEqual(len(hidden_states), expected_num_layers)
+
+                self.assertListEqual(
+                    list(hidden_states[0].shape[-2:]),
+                    [self.model_tester.num_queries, self.model_tester.d_model],
+                )
+
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            inputs_dict["output_hidden_states"] = True
+            check_hidden_states_output(inputs_dict, config, model_class)
+
+            # check that output_hidden_states also work using config
+            del inputs_dict["output_hidden_states"]
+            config.output_hidden_states = True
+
+            check_hidden_states_output(inputs_dict, config, model_class)
+
+    def test_attention_outputs(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+        config.return_dict = True
+
+        for model_class in self.all_model_classes:
+            inputs_dict["output_attentions"] = True
+            inputs_dict["output_hidden_states"] = False
+            config.return_dict = True
+            model = model_class._from_config(config, attn_implementation="eager")
+            config = model.config
+            model.to(torch_device)
+            model.eval()
+            with torch.no_grad():
+                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
+            attentions = outputs.encoder_attentions
+            self.assertEqual(len(attentions), self.model_tester.encoder_layers)
+
+            # check that output_attentions also work using config
+            del inputs_dict["output_attentions"]
+            config.output_attentions = True
+            model = model_class(config)
+            model.to(torch_device)
+            model.eval()
+            with torch.no_grad():
+                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
+            attentions = outputs.encoder_attentions
+            self.assertEqual(len(attentions), self.model_tester.encoder_layers)
+
+            self.assertListEqual(
+                list(attentions[0].shape[-3:]),
+                [
+                    self.model_tester.encoder_attention_heads,
+                    self.model_tester.encoder_seq_length,
+                    self.model_tester.encoder_seq_length,
+                ],
+            )
+            out_len = len(outputs)
+
+            correct_outlen = 13
+
+            # loss is at first position
+            if "labels" in inputs_dict:
+                correct_outlen += 1  # loss is added to beginning
+            # Object Detection model returns pred_logits and pred_boxes
+            if model_class.__name__ == "PPDocLayoutV2ForObjectDetection":
+                correct_outlen += 3
+
+            self.assertEqual(out_len, correct_outlen)
+
+            # decoder attentions
+            decoder_attentions = outputs.decoder_attentions
+            self.assertIsInstance(decoder_attentions, (list, tuple))
+            self.assertEqual(len(decoder_attentions), self.model_tester.decoder_layers)
+            self.assertListEqual(
+                list(decoder_attentions[0].shape[-3:]),
+                [
+                    self.model_tester.decoder_attention_heads,
+                    self.model_tester.num_queries,
+                    self.model_tester.num_queries,
+                ],
+            )
+
+            # cross attentions
+            cross_attentions = outputs.cross_attentions
+            self.assertIsInstance(cross_attentions, (list, tuple))
+            self.assertEqual(len(cross_attentions), self.model_tester.decoder_layers)
+            self.assertListEqual(
+                list(cross_attentions[0].shape[-3:]),
+                [
+                    self.model_tester.decoder_attention_heads,
+                    self.model_tester.num_feature_levels,
+                    self.model_tester.decoder_n_points,
+                ],
+            )
+
+            # Check attention is always last and order is fine
+            inputs_dict["output_attentions"] = True
+            inputs_dict["output_hidden_states"] = True
+            model = model_class(config)
+            model.to(torch_device)
+            model.eval()
+            with torch.no_grad():
+                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
+
+            if hasattr(self.model_tester, "num_hidden_states_types"):
+                added_hidden_states = self.model_tester.num_hidden_states_types
+            else:
+                # RTDetr should maintin encoder_hidden_states output
+                added_hidden_states = 2
+            self.assertEqual(out_len + added_hidden_states, len(outputs))
+
+            self_attentions = outputs.encoder_attentions
+
+            self.assertEqual(len(self_attentions), self.model_tester.encoder_layers)
+            self.assertListEqual(
+                list(self_attentions[0].shape[-3:]),
+                [
+                    self.model_tester.encoder_attention_heads,
+                    self.model_tester.encoder_seq_length,
+                    self.model_tester.encoder_seq_length,
+                ],
+            )
+
+
+@require_torch
+@require_vision
+@slow
+class PPDocLayoutV2ModelIntegrationTest(unittest.TestCase):
+    def setUp(self):
+        model_path = "PaddlePaddle/PP-DocLayoutV2_safetensors"
+        self.model = PPDocLayoutV2ForObjectDetection.from_pretrained(model_path).to(torch_device)
+        self.image_processor = PPDocLayoutV2ImageProcessor.from_pretrained(model_path)
+        img_url = url_to_local_path(
+            "https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/layout_demo.jpg"
+        )
+        self.image = load_image(img_url)
+
+    def tearDown(self):
+        cleanup(torch_device, gc_collect=True)
+
+    def test_inference_object_detection_head(self):
+        inputs = self.image_processor(images=self.image, return_tensors="pt").to(torch_device)
+
+        with torch.no_grad():
+            outputs = self.model(**inputs)
+
+        expected_shape_logits = torch.Size((1, 300, self.model.config.num_labels))
+        expected_logits = torch.tensor(
+            [[-3.6572, -4.4185, -4.3930], [-3.7213, -4.5011, -4.6771], [-3.8721, -4.4524, -4.4162]]
+        ).to(torch_device)
+        self.assertEqual(outputs.logits.shape, expected_shape_logits)
+        torch.testing.assert_close(outputs.logits[0, :3, :3], expected_logits, rtol=2e-2, atol=2e-2)
+
+        expected_shape_boxes = torch.Size((1, 300, 4))
+        expected_boxes = torch.tensor(
+            [[0.3709, 0.4911, 0.3358], [0.7263, 0.4419, 0.3394], [0.3724, 0.1793, 0.3392]]
+        ).to(torch_device)
+        self.assertEqual(outputs.pred_boxes.shape, expected_shape_boxes)
+        torch.testing.assert_close(outputs.pred_boxes[0, :3, :3], expected_boxes, rtol=2e-2, atol=2e-2)
+
+        expected_shape_order_logits = torch.Size((1, 300, 300))
+        expected_order_logits = torch.tensor(
+            [
+                [-10000.0000, 43.8388, -32.8785],
+                [-10000.0000, -10000.0000, -63.9118],
+                [-10000.0000, -10000.0000, -10000.0000],
+            ]
+        ).to(torch_device)
+        self.assertEqual(outputs.order_logits.shape, expected_shape_order_logits)
+        torch.testing.assert_close(outputs.order_logits[0, :3, :3], expected_order_logits, rtol=2e-2, atol=2e-2)
+
+        # verify postprocessing
+        results = self.image_processor.post_process_object_detection(
+            outputs, threshold=0.5, target_sizes=[self.image.size[::-1]]
+        )[0]
+
+        expected_scores = torch.tensor(
+            [
+                0.9878,
+                0.9675,
+                0.9882,
+                0.9852,
+                0.9828,
+                0.9843,
+                0.9700,
+                0.8182,
+                0.5148,
+                0.8273,
+                0.8718,
+                0.9494,
+                0.8733,
+                0.9266,
+            ]
+        ).to(torch_device)
+        torch.testing.assert_close(results["scores"], expected_scores, rtol=2e-2, atol=2e-2)
+
+        expected_labels = [22, 17, 22, 22, 22, 22, 22, 10, 22, 10, 22, 10, 16, 8]
+        self.assertSequenceEqual(results["labels"].tolist(), expected_labels)
+
+        expected_slice_boxes = torch.tensor(
+            [
+                [335.3923, 184.2622, 896.4918, 654.4847],
+                [337.1364, 683.4911, 869.4224, 798.2716],
+                [335.7133, 843.0425, 891.1711, 1454.1525],
+                [920.4213, 185.5302, 1476.3922, 464.2497],
+            ]
+        ).to(torch_device)
+        torch.testing.assert_close(results["boxes"][:4], expected_slice_boxes, rtol=2e-2, atol=2e-2)