first commit

tests/models/pp_ocrv5_server_det/test_image_processing_pp_ocrv5_server_det.py

@@ -0,0 +1,194 @@
+# 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 numpy as np
+
+from transformers import is_torch_available, is_vision_available
+from transformers.testing_utils import require_torch, require_vision
+
+from ...test_image_processing_common import ImageProcessingTestMixin, prepare_image_inputs
+
+
+if is_vision_available():
+    from PIL import Image
+
+
+if is_torch_available():
+    import torch
+
+
+class PPOCRV5ServerDetImageProcessingTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=7,
+        num_channels=3,
+        image_size=18,
+        min_resolution=10,
+        max_resolution=400,
+        limit_side_len=960,
+        limit_type="max",
+        max_side_limit=4000,
+        do_resize=True,
+        size=None,
+        do_rescale=True,
+        rescale_factor=1 / 255,
+        do_normalize=True,
+        image_mean=[0.485, 0.456, 0.406],
+        image_std=[0.229, 0.224, 0.225],
+    ):
+        size = size if size is not None else {"height": 512, "width": 512}
+        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.limit_side_len = limit_side_len
+        self.limit_type = limit_type
+        self.max_side_limit = max_side_limit
+        self.do_resize = do_resize
+        self.size = size
+        self.do_rescale = do_rescale
+        self.rescale_factor = rescale_factor
+        self.do_normalize = do_normalize
+        self.image_mean = image_mean
+        self.image_std = image_std
+
+    def prepare_image_processor_dict(self):
+        return {
+            "image_mean": self.image_mean,
+            "image_std": self.image_std,
+            "do_normalize": self.do_normalize,
+            "do_resize": self.do_resize,
+            "size": self.size,
+            "keep_aspect_ratio": False,
+            "do_pad": False,
+        }
+
+    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 get_expected_value(self, image_inputs):
+        image = image_inputs[0]
+
+        if isinstance(image, Image.Image):
+            width, height = image.size
+        elif isinstance(image, np.ndarray):
+            height, width = image.shape[0], image.shape[1]
+        else:
+            height, width = image.shape[1], image.shape[2]
+
+        if max(height, width) > self.limit_side_len:
+            ratio = float(self.limit_side_len) / max(height, width)
+        else:
+            ratio = 1.0
+
+        resize_height = int(height * ratio)
+        resize_width = int(width * ratio)
+
+        if self.max_side_limit is not None and max(resize_height, resize_width) > self.max_side_limit:
+            ratio = float(self.max_side_limit) / max(resize_height, resize_width)
+            resize_height = int(resize_height * ratio)
+            resize_width = int(resize_width * ratio)
+
+        resize_height = max(int(round(resize_height / 32) * 32), 32)
+        resize_width = max(int(round(resize_width / 32) * 32), 32)
+
+        if resize_height == height and resize_width == width:
+            return resize_height, resize_width
+
+        if resize_width <= 0 or resize_height <= 0:
+            return None, None
+
+        return resize_height, resize_width
+
+    def expected_output_image_shape(self, images):
+        height, width = self.get_expected_value(images)
+        return self.num_channels, height, width
+
+
+@require_torch
+@require_vision
+class PPOCRV5ServerDetImageProcessingTest(ImageProcessingTestMixin, unittest.TestCase):
+    def setUp(self):
+        super().setUp()
+        self.image_processor_tester = PPOCRV5ServerDetImageProcessingTester(self)
+
+    @property
+    def image_processor_dict(self):
+        return self.image_processor_tester.prepare_image_processor_dict()
+
+    # PPOCRV5ServerDet can’t stack the images into a batch because the image processor resizes them adaptively, leading to inconsistent output sizes."
+    # Skip Test batched
+    def test_call_pytorch(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 PyTorch tensors
+            image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=False, torchify=True)
+
+            for image in image_inputs:
+                self.assertIsInstance(image, torch.Tensor)
+
+            # Test not batched input
+            encoded_images = image_processing(image_inputs[0], return_tensors="pt").pixel_values
+            expected_output_image_shape = self.image_processor_tester.expected_output_image_shape([image_inputs[0]])
+            self.assertEqual(tuple(encoded_images.shape), (1, *expected_output_image_shape))
+
+    # PPOCRV5ServerDet can’t stack the images into a batch because the image processor resizes them adaptively, leading to inconsistent output sizes.
+    # Skip Test batched
+    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=False, 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 = self.image_processor_tester.expected_output_image_shape([image_inputs[0]])
+            self.assertEqual(tuple(encoded_images.shape), (1, *expected_output_image_shape))
+
+    # PPOCRV5ServerDet can’t stack the images into a batch because the image processor resizes them adaptively, leading to inconsistent output sizes.
+    # Skip Test batched
+    def test_call_pil(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 PIL images
+            image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=False)
+            for image in image_inputs:
+                self.assertIsInstance(image, Image.Image)
+
+            # Test not batched input
+            encoded_images = image_processing(image_inputs[0], return_tensors="pt").pixel_values
+            expected_output_image_shape = self.image_processor_tester.expected_output_image_shape([image_inputs[0]])
+            self.assertEqual(tuple(encoded_images.shape), (1, *expected_output_image_shape))
+
+    @unittest.skip(reason="PPOCRV5ServerDetImageProcessor does not support 4 channel images yet")
+    def test_call_numpy_4_channels():
+        pass

tests/models/pp_ocrv5_server_det/test_modeling_pp_ocrv5_server_det.py

@@ -0,0 +1,311 @@
+# 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-OCRV5ServerDet model."""
+
+import inspect
+import unittest
+
+from parameterized import parameterized
+
+from transformers import (
+    PPOCRV5ServerDetConfig,
+    PPOCRV5ServerDetForObjectDetection,
+    PPOCRV5ServerDetImageProcessor,
+    PPOCRV5ServerDetModel,
+    is_torch_available,
+    is_vision_available,
+)
+from transformers.image_utils import load_image
+from transformers.testing_utils import (
+    require_cv2,
+    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 PPOCRV5ServerDetModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=3,
+        image_size=128,
+        num_channels=3,
+        num_stages=5,
+        is_training=False,
+        scale=1.0,
+        divisor=16,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.num_channels = num_channels
+        self.image_size = image_size
+        self.is_training = is_training
+        self.num_stages = num_stages
+        self.scale = scale
+        self.divisor = divisor
+
+    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) -> PPOCRV5ServerDetConfig:
+        # Minimal backbone config for fast tests (< 1M params for test_model_is_small)
+        backbone_config = {
+            "model_type": "hgnet_v2",
+            "out_features": ["stage1", "stage2", "stage3", "stage4"],
+            "out_indices": [1, 2, 3, 4],
+            "depths": [1, 1, 1, 1],
+            "hidden_sizes": [16, 32, 64, 128],
+            "stage_in_channels": [16, 16, 32, 64],
+            "stage_mid_channels": [16, 16, 32, 64],
+            "stage_out_channels": [16, 32, 64, 128],
+            "stage_num_blocks": [1, 1, 1, 1],
+            "stage_downsample": [False, True, True, True],
+            "stage_light_block": [False, False, True, True],
+            "stage_kernel_size": [3, 3, 5, 5],
+            "stage_numb_of_layers": [1, 1, 1, 1],
+            "stem_channels": [3, 8, 16],
+            "embedding_size": 16,
+        }
+
+        intraclass_block_config = {
+            "reduce_channel": [1, 1, 0],
+            "return_channel": [1, 1, 0],
+            "vertical_long_to_small_conv_longratio": [[7, 1], [1, 1], [3, 0]],
+            "vertical_long_to_small_conv_midratio": [[5, 1], [1, 1], [2, 0]],
+            "vertical_long_to_small_conv_shortratio": [[3, 1], [1, 1], [1, 0]],
+            "horizontal_small_to_long_conv_longratio": [[1, 7], [1, 1], [0, 3]],
+            "horizontal_small_to_long_conv_midratio": [[1, 5], [1, 1], [0, 2]],
+            "horizontal_small_to_long_conv_shortratio": [[1, 3], [1, 1], [0, 1]],
+            "symmetric_conv_long_longratio": [[7, 7], [1, 1], [3, 3]],
+            "symmetric_conv_long_midratio": [[5, 5], [1, 1], [2, 2]],
+            "symmetric_conv_long_shortratio": [[3, 3], [1, 1], [1, 1]],
+        }
+
+        config = PPOCRV5ServerDetConfig(
+            backbone_config=backbone_config,
+            interpolate_mode="nearest",
+            neck_out_channels=32,
+            reduce_factor=2,
+            intraclass_block_number=4,
+            intraclass_block_config=intraclass_block_config,
+            mode="large",
+            scale_factor=2,
+            scale_factor_list=[1, 2, 4, 8],
+            hidden_act="relu",
+            kernel_list=[3, 2, 2],
+        )
+
+        return config
+
+    def create_and_check_pp_ocrv5_server_det_object_detection(self, config, pixel_values):
+        model = PPOCRV5ServerDetForObjectDetection(config=config)
+        model.to(torch_device)
+        model.eval()
+
+        result = model(pixel_values)
+
+        self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, 1, self.image_size, self.image_size))
+
+
+@require_torch
+class PPOCRV5ServerDetModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
+    all_model_classes = (PPOCRV5ServerDetModel, PPOCRV5ServerDetForObjectDetection) if is_torch_available() else ()
+    pipeline_model_mapping = {"object-detection": PPOCRV5ServerDetForObjectDetection} if is_torch_available() else {}
+
+    test_resize_embeddings = False
+    has_attentions = False
+
+    def setUp(self):
+        self.model_tester = PPOCRV5ServerDetModelTester(
+            self,
+            batch_size=3,
+            is_training=False,
+            image_size=128,
+        )
+        self.model_tester.parent = self
+        self.config_tester = ConfigTester(
+            self,
+            config_class=PPOCRV5ServerDetConfig,
+            has_text_modality=False,
+            common_properties=[],
+        )
+
+    def test_config(self):
+        # Skip create_and_test_config_with_num_labels: PP-OCRV5 has fixed single class (text)
+        self.config_tester.create_and_test_config_common_properties()
+        self.config_tester.create_and_test_config_to_json_string()
+        self.config_tester.create_and_test_config_to_json_file()
+        self.config_tester.create_and_test_config_from_and_save_pretrained()
+        self.config_tester.create_and_test_config_from_and_save_pretrained_subfolder()
+        self.config_tester.create_and_test_config_from_and_save_pretrained_composite()
+        self.config_tester.check_config_can_be_init_without_params()
+        self.config_tester.check_config_arguments_init()
+        self.config_tester.create_and_test_config_from_pretrained_custom_kwargs()
+
+    def test_pp_ocrv5_server_det_object_detection(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_pp_ocrv5_server_det_object_detection(*config_and_inputs)
+
+    @unittest.skip(reason="PPOCRV5ServerDet does not support input and output embeddings")
+    def test_model_get_set_embeddings(self):
+        pass
+
+    def test_hidden_states_output(self):
+        # PP-OCRV5 uses HGNetV2 backbone: hidden_states = (embedding, stage1, ..., stageN) = num_stages + 1
+        config = self.model_tester.get_config()
+        num_expected_hidden_states = len(config.backbone_config.depths) + 1
+
+        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.hidden_states
+            self.assertIsNotNone(hidden_states)
+            self.assertEqual(len(hidden_states), num_expected_hidden_states)
+
+            # First hidden state (embedding output) is 4x downsampled
+            self.assertListEqual(
+                list(hidden_states[0].shape[-2:]),
+                [self.model_tester.image_size // 4, self.model_tester.image_size // 4],
+            )
+
+        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.copy(), config, model_class)
+
+            # Check that output_hidden_states also works via config (including backbone subconfig)
+            del inputs_dict["output_hidden_states"]
+            config.output_hidden_states = True
+            if config.backbone_config is not None:
+                config.backbone_config.output_hidden_states = True
+            check_hidden_states_output(inputs_dict.copy(), config, model_class)
+
+    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))
+
+
+@require_torch
+@require_vision
+@require_cv2
+@slow
+class PPOCRV5ServerDetModelIntegrationTest(unittest.TestCase):
+    def setUp(self):
+        model_path = "PaddlePaddle/PP-OCRv5_server_det_safetensors"
+        self.model = PPOCRV5ServerDetForObjectDetection.from_pretrained(model_path).to(torch_device)
+        self.image_processor = (
+            PPOCRV5ServerDetImageProcessor.from_pretrained(model_path) if is_vision_available() else None
+        )
+        img_url = url_to_local_path(
+            "https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/general_ocr_001.png"
+        )
+        self.image = load_image(img_url)
+
+    def test_inference_object_detection_head(self):
+        inputs = self.image_processor(images=self.image, return_tensors="pt").to(torch_device)
+        bs, c, h, w = inputs["pixel_values"].shape
+
+        with torch.no_grad():
+            outputs = self.model(**inputs)
+
+        results = self.image_processor.post_process_object_detection(outputs, target_sizes=inputs["target_sizes"])
+
+        expected_shape_logits = torch.Size((bs, c // 3, h, w))
+
+        expected_logits = torch.tensor(
+            [
+                [0.0004, 0.0003, 0.0002],
+                [0.0003, 0.0002, 0.0002],
+                [0.0006, 0.0003, 0.0003],
+            ],
+            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_shape_boxes = torch.Size((4, 4, 2))
+        expected_boxes = torch.tensor(
+            [
+                [[76, 550], [399, 538], [400, 575], [77, 587]],
+                [[14, 505], [517, 484], [519, 532], [16, 553]],
+                [[193, 452], [401, 443], [403, 483], [195, 492]],
+                [[32, 406], [488, 384], [491, 434], [34, 456]],
+            ],
+            dtype=torch.short,
+            device=torch_device,
+        )
+
+        self.assertEqual(results[0]["boxes"].shape, expected_shape_boxes)
+        torch.testing.assert_close(results[0]["boxes"], expected_boxes, rtol=2e-2, atol=2e-2)
+
+        expected_scores = torch.tensor([0.9023, 0.8941, 0.8937, 0.8781], device=torch_device)
+        self.assertEqual(results[0]["scores"].shape, (4,))
+        torch.testing.assert_close(results[0]["scores"], expected_scores, rtol=2e-2, atol=2e-2)
+
+        self.assertEqual(results[0]["labels"].shape, (4,))
+        self.assertTrue((results[0]["labels"] == 0).all())  # Single class: text