first commit

tests/models/uvdoc/test_image_processing_uvdoc.py

@@ -0,0 +1,144 @@
+# 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.
+
+import unittest
+
+import torch
+
+from transformers.testing_utils import require_torch, require_vision
+
+from ...test_image_processing_common import ImageProcessingTestMixin, prepare_image_inputs
+
+
+class UVDocImageProcessingTester:
+    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=False,
+    ):
+        size = size if size is not None else {"height": 18, "width": 18}
+        self.parent = parent
+        self.batch_size = batch_size
+        self.num_channels = num_channels
+        self.image_size = image_size
+        self.min_resolution = min_resolution
+        self.max_resolution = max_resolution
+        self.do_resize = do_resize
+        self.size = size
+        self.do_normalize = do_normalize
+
+    def prepare_image_processor_dict(self):
+        return {
+            "do_normalize": self.do_normalize,
+            "do_resize": self.do_resize,
+            "size": self.size,
+        }
+
+    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 UVDocImageProcessingTest(ImageProcessingTestMixin, unittest.TestCase):
+    def setUp(self):
+        super().setUp()
+        self.image_processor_tester = UVDocImageProcessingTester(self)
+
+    @property
+    def image_processor_dict(self):
+        return self.image_processor_tester.prepare_image_processor_dict()
+
+    @unittest.skip("UVDoc image processors doesn't support 4 channel images")
+    def test_call_numpy_4_channels(self):
+        pass
+
+    def test_post_process_document_rectification(self):
+        for image_processing_class in self.image_processing_classes.values():
+            image_processor = image_processing_class(**self.image_processor_dict)
+
+            batch_size = 2
+            height, width = 32, 48
+            pred_height, pred_width = 16, 24
+
+            # Create identity grid in normalized coords [-1, 1] for grid_sample
+            y_coords = torch.linspace(-1, 1, pred_height)
+            x_coords = torch.linspace(-1, 1, pred_width)
+            grid_y, grid_x = torch.meshgrid(y_coords, x_coords, indexing="ij")
+            prediction = torch.stack([grid_x, grid_y], dim=0).unsqueeze(0).expand(batch_size, -1, -1, -1)
+
+            # Original images as list of tensors (C, H, W) each
+            original_images = [torch.rand(3, height, width) for _ in range(batch_size)]
+
+            results = image_processor.post_process_document_rectification(prediction, original_images, scale=255.0)
+
+            self.assertEqual(len(results), batch_size)
+            for i, result in enumerate(results):
+                self.assertIn("images", result)
+                images = result["images"]
+                self.assertEqual(images.shape, (height, width, 3))
+                self.assertEqual(images.dtype, torch.uint8)
+                self.assertTrue(torch.all(images >= 0) and torch.all(images <= 255))
+
+            # Test with custom scale
+            results_custom_scale = image_processor.post_process_document_rectification(
+                prediction, original_images, scale=1.0
+            )
+            for result in results_custom_scale:
+                self.assertTrue(torch.all(result["images"] <= 1))
+
+    def test_post_process_document_rectification_different_sizes(self):
+        """Test post-processing with original images of different sizes (list of tensors)."""
+        for image_processing_class in self.image_processing_classes.values():
+            image_processor = image_processing_class(**self.image_processor_dict)
+
+            # Create predictions for 2 images (model output size is fixed)
+            pred_height, pred_width = 16, 24
+            y_coords = torch.linspace(-1, 1, pred_height)
+            x_coords = torch.linspace(-1, 1, pred_width)
+            grid_y, grid_x = torch.meshgrid(y_coords, x_coords, indexing="ij")
+            prediction = torch.stack([grid_x, grid_y], dim=0).unsqueeze(0).expand(2, -1, -1, -1)
+
+            # Original images with different sizes: (32, 48) and (64, 96)
+            original_images = [
+                torch.rand(3, 32, 48),
+                torch.rand(3, 64, 96),
+            ]
+
+            results = image_processor.post_process_document_rectification(prediction, original_images, scale=255.0)
+
+            self.assertEqual(len(results), 2)
+            self.assertEqual(results[0]["images"].shape, (32, 48, 3))
+            self.assertEqual(results[1]["images"].shape, (64, 96, 3))
+            for result in results:
+                self.assertEqual(result["images"].dtype, torch.uint8)
+                self.assertTrue(torch.all(result["images"] >= 0) and torch.all(result["images"] <= 255))

tests/models/uvdoc/test_modeling_uvdoc.py

@@ -0,0 +1,346 @@
+# 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 UVDoc model."""
+
+import inspect
+import unittest
+
+from parameterized import parameterized
+
+from transformers import (
+    AutoModel,
+    UVDocBackbone,
+    UVDocBackboneConfig,
+    UVDocConfig,
+    UVDocImageProcessor,
+    UVDocModel,
+    is_torch_available,
+)
+from transformers.image_utils import load_image
+from transformers.testing_utils import (
+    require_torch,
+    require_torch_accelerator,
+    require_vision,
+    slow,
+    torch_device,
+)
+
+from ...test_backbone_common import BackboneTesterMixin
+from ...test_configuration_common import ConfigTester
+from ...test_modeling_common import ModelTesterMixin, floats_tensor
+from ...test_processing_common import url_to_local_path
+
+
+if is_torch_available():
+    import torch
+
+
+class UVDocModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=3,
+        image_size=128,
+        num_channels=3,
+        is_training=False,
+        kernel_size=5,
+        bridge_connector=(32, 32),
+        out_point_positions2D=((32, 8), (8, 2)),
+        padding_mode="reflect",
+        hidden_act="prelu",
+        num_hidden_layers=2,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.num_channels = num_channels
+        self.image_size = image_size
+        self.is_training = is_training
+        self.kernel_size = kernel_size
+        self.bridge_connector = bridge_connector
+        self.out_point_positions2D = out_point_positions2D
+        self.padding_mode = padding_mode
+        self.hidden_act = hidden_act
+        self.num_hidden_layers = num_hidden_layers
+        # For test_hidden_states_output: UVDoc outputs spatial hidden states (B, C, H, W)
+        # with shape[-2:] = (8, 8) for image_size=128
+        self.seq_length = 8
+        self.hidden_size = 8
+
+    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 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) -> UVDocConfig:
+        resnet_configs = (
+            (
+                (8, 8, 1, False),
+                (8, 8, 3, False),
+                (8, 8, 3, False),
+            ),
+            (
+                (8, 16, 1, True),
+                (16, 16, 3, False),
+                (16, 16, 3, False),
+                (16, 16, 3, False),
+            ),
+            (
+                (16, 32, 1, True),
+                (32, 32, 3, False),
+                (32, 32, 3, False),
+                (32, 32, 3, False),
+                (32, 32, 3, False),
+                (32, 32, 3, False),
+            ),
+        )
+
+        stage_configs = (
+            ((32, 1),),
+            ((32, 2),),
+        )
+        backbone_config = {
+            "model_type": "uvdoc_backbone",
+            "kernel_size": self.kernel_size,
+            "resnet_configs": resnet_configs,
+            "stage_configs": stage_configs,
+            "out_features": ["stage1", "stage2"],
+            "out_indices": [1, 2],
+            "resnet_head": ((3, 8), (8, 8)),
+        }
+
+        return UVDocConfig(
+            kernel_size=self.kernel_size,
+            padding_mode=self.padding_mode,
+            hidden_act=self.hidden_act,
+            backbone_config=backbone_config,
+            bridge_connector=self.bridge_connector,
+            out_point_positions2D=self.out_point_positions2D,
+        )
+
+    def create_and_check_uvdoc_document_rectification(self, config, pixel_values):
+        model = UVDocModel(config=config)
+        model.to(torch_device)
+        model.eval()
+
+        result = model(pixel_values)
+
+        self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, 2, 8, 8))
+
+
+class UVDocBackboneTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=3,
+        image_size=128,
+        num_channels=3,
+        is_training=False,
+        kernel_size=5,
+        resnet_head=((3, 8), (8, 8)),
+        out_features=["stage1", "stage2"],
+        out_indices=[1, 2],
+        num_hidden_layers=2,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.num_channels = num_channels
+        self.image_size = image_size
+        self.is_training = is_training
+        self.kernel_size = kernel_size
+        self.resnet_head = resnet_head
+        self.out_features = out_features
+        self.out_indices = out_indices
+        self.num_hidden_layers = num_hidden_layers
+
+    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 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) -> UVDocBackbone:
+        resnet_configs = (
+            (
+                (8, 8, 1, False),
+                (8, 8, 3, False),
+                (8, 8, 3, False),
+            ),
+            (
+                (8, 16, 1, True),
+                (16, 16, 3, False),
+                (16, 16, 3, False),
+                (16, 16, 3, False),
+            ),
+            (
+                (16, 32, 1, True),
+                (32, 32, 3, False),
+                (32, 32, 3, False),
+                (32, 32, 3, False),
+                (32, 32, 3, False),
+                (32, 32, 3, False),
+            ),
+        )
+
+        stage_configs = (
+            ((32, 1),),
+            ((32, 2),),
+        )
+
+        return UVDocBackboneConfig(
+            kernel_size=self.kernel_size,
+            resnet_head=self.resnet_head,
+            resnet_configs=resnet_configs,
+            stage_configs=stage_configs,
+            out_features=self.out_features,
+            out_indices=self.out_indices,
+        )
+
+
+@require_torch
+class UVDocBackboneTest(BackboneTesterMixin, unittest.TestCase):
+    all_model_classes = (UVDocBackbone,) if is_torch_available() else ()
+    has_attentions = False
+    config_class = UVDocBackboneConfig
+
+    def setUp(self):
+        self.model_tester = UVDocBackboneTester(self)
+        self.config_tester = ConfigTester(
+            self,
+            config_class=UVDocBackboneConfig,
+            has_text_modality=False,
+            common_properties=[],
+        )
+
+
+@require_torch
+class UVDocModelTest(ModelTesterMixin, unittest.TestCase):
+    all_model_classes = (UVDocModel,) if is_torch_available() else ()
+
+    has_attentions = False
+    test_resize_embeddings = False
+
+    def setUp(self):
+        self.model_tester = UVDocModelTester(self)
+        self.config_tester = ConfigTester(
+            self,
+            config_class=UVDocConfig,
+            has_text_modality=False,
+            common_properties=[],
+        )
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_uvdoc_document_rectification(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_uvdoc_document_rectification(*config_and_inputs)
+
+    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():
+                if tensor.dtype == torch.float32:
+                    inputs_dict[key] = tensor.to(dtype)
+            with torch.no_grad():
+                _ = model(**self._prepare_for_class(inputs_dict, model_class))
+
+    @unittest.skip(reason="UVDoc does not support input and output embeddings")
+    def test_model_get_set_embeddings(self):
+        pass
+
+    @unittest.skip(reason="UVDoc does not support training")
+    def test_retain_grad_hidden_states_attentions(self):
+        pass
+
+
+@require_torch
+@require_vision
+@slow
+class UVDocModelIntegrationTest(unittest.TestCase):
+    def setUp(self):
+        model_path = "PaddlePaddle/UVDoc_safetensors"
+        self.model = AutoModel.from_pretrained(model_path).to(torch_device)
+        self.image_processor = UVDocImageProcessor()
+        img_url = url_to_local_path("https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/doc_test.jpg")
+        self.image = load_image(img_url)
+
+    def test_inference_document_rectification(self):
+        inputs = self.image_processor(images=self.image, return_tensors="pt").to(torch_device)
+        bs = inputs["pixel_values"].shape[0]
+
+        with torch.no_grad():
+            outputs = self.model(**inputs)
+
+        results = self.image_processor.post_process_document_rectification(
+            outputs.last_hidden_state, inputs["original_images"]
+        )
+
+        expected_shape_logits = torch.Size((bs, 2, 45, 31))
+        expected_logits = torch.tensor(
+            [
+                [-0.7635, -0.7251, -0.6819],
+                [-0.7643, -0.7250, -0.6814],
+                [-0.7647, -0.7252, -0.6816],
+            ],
+            device=torch_device,
+        )
+
+        self.assertEqual(outputs.last_hidden_state.shape, expected_shape_logits)
+        torch.testing.assert_close(outputs.last_hidden_state[0, 0, :3, :3], expected_logits, rtol=2e-4, atol=2e-4)
+
+        expected_images = torch.tensor(
+            [
+                [131, 130, 128],
+                [131, 129, 127],
+                [130, 129, 127],
+            ],
+            device=torch_device,
+            dtype=torch.uint8,
+        )
+        torch.testing.assert_close(results[0]["images"][:3, :3, 0], expected_images, rtol=2e-4, atol=2e-4)