first commit

tests/models/sam3/test_image_processing_sam3.py

@@ -0,0 +1,209 @@
+# Copyright 2025 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 datasets import load_dataset
+
+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
+
+
+if is_torch_available():
+    import torch
+
+if is_vision_available():
+    pass
+
+
+class Sam3ImageProcessingTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=7,
+        num_channels=3,
+        min_resolution=30,
+        max_resolution=400,
+        mask_size=None,
+        do_resize=True,
+        size=None,
+        do_normalize=True,
+        image_mean=[0.5, 0.5, 0.5],
+        image_std=[0.5, 0.5, 0.5],
+    ):
+        size = size if size is not None else {"height": 20, "width": 20}
+        mask_size = mask_size if mask_size is not None else {"height": 12, "width": 12}
+        self.parent = parent
+        self.batch_size = batch_size
+        self.num_channels = num_channels
+        self.min_resolution = min_resolution
+        self.max_resolution = max_resolution
+        self.mask_size = mask_size
+        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 {
+            "image_mean": self.image_mean,
+            "image_std": self.image_std,
+            "do_normalize": self.do_normalize,
+            "do_resize": self.do_resize,
+            "size": self.size,
+            "mask_size": self.mask_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,
+        )
+
+
+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 Sam3ImageProcessingTest(ImageProcessingTestMixin, unittest.TestCase):
+    def setUp(self):
+        super().setUp()
+        self.image_processor_tester = Sam3ImageProcessingTester(self)
+
+    @property
+    def image_processor_dict(self):
+        return self.image_processor_tester.prepare_image_processor_dict()
+
+    def test_call_segmentation_maps(self):
+        for image_processing_class in self.image_processing_classes.values():
+            image_processor = image_processing_class(**self.image_processor_dict)
+            image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=False, torchify=True)
+            maps = []
+            for image in image_inputs:
+                self.assertIsInstance(image, torch.Tensor)
+                maps.append(torch.zeros(image.shape[-2:]).long())
+
+            # Test not batched input
+            encoding = image_processor(image_inputs[0], maps[0], return_tensors="pt")
+            self.assertEqual(
+                encoding["pixel_values"].shape,
+                (
+                    1,
+                    self.image_processor_tester.num_channels,
+                    self.image_processor_tester.size["height"],
+                    self.image_processor_tester.size["width"],
+                ),
+            )
+            self.assertEqual(
+                encoding["labels"].shape,
+                (
+                    1,
+                    self.image_processor_tester.mask_size["height"],
+                    self.image_processor_tester.mask_size["width"],
+                ),
+            )
+            self.assertEqual(encoding["labels"].dtype, torch.long)
+            self.assertTrue(encoding["labels"].min().item() >= 0)
+            self.assertTrue(encoding["labels"].max().item() <= 255)
+
+            # Test batched
+            encoding = image_processor(image_inputs, maps, return_tensors="pt")
+            self.assertEqual(
+                encoding["pixel_values"].shape,
+                (
+                    self.image_processor_tester.batch_size,
+                    self.image_processor_tester.num_channels,
+                    self.image_processor_tester.size["height"],
+                    self.image_processor_tester.size["width"],
+                ),
+            )
+            self.assertEqual(
+                encoding["labels"].shape,
+                (
+                    self.image_processor_tester.batch_size,
+                    self.image_processor_tester.mask_size["height"],
+                    self.image_processor_tester.mask_size["width"],
+                ),
+            )
+            self.assertEqual(encoding["labels"].dtype, torch.long)
+            self.assertTrue(encoding["labels"].min().item() >= 0)
+            self.assertTrue(encoding["labels"].max().item() <= 255)
+
+            # Test not batched input (PIL images) with segmentation maps from dataset
+            image, segmentation_map = prepare_semantic_single_inputs()
+            encoding = image_processor(image, segmentation_map, return_tensors="pt")
+            self.assertEqual(
+                encoding["pixel_values"].shape,
+                (
+                    1,
+                    self.image_processor_tester.num_channels,
+                    self.image_processor_tester.size["height"],
+                    self.image_processor_tester.size["width"],
+                ),
+            )
+            self.assertEqual(
+                encoding["labels"].shape,
+                (
+                    1,
+                    self.image_processor_tester.mask_size["height"],
+                    self.image_processor_tester.mask_size["width"],
+                ),
+            )
+            self.assertEqual(encoding["labels"].dtype, torch.long)
+            self.assertTrue(encoding["labels"].min().item() >= 0)
+            self.assertTrue(encoding["labels"].max().item() <= 255)
+
+            # Test batched input (PIL images)
+            images, segmentation_maps = prepare_semantic_batch_inputs()
+            encoding = image_processor(images, segmentation_maps, return_tensors="pt")
+            self.assertEqual(
+                encoding["pixel_values"].shape,
+                (
+                    2,
+                    self.image_processor_tester.num_channels,
+                    self.image_processor_tester.size["height"],
+                    self.image_processor_tester.size["width"],
+                ),
+            )
+            self.assertEqual(
+                encoding["labels"].shape,
+                (
+                    2,
+                    self.image_processor_tester.mask_size["height"],
+                    self.image_processor_tester.mask_size["width"],
+                ),
+            )
+            self.assertEqual(encoding["labels"].dtype, torch.long)
+            self.assertTrue(encoding["labels"].min().item() >= 0)
+            self.assertTrue(encoding["labels"].max().item() <= 255)

tests/models/sam3/test_processing_sam3.py

@@ -0,0 +1,105 @@
+# Copyright 2025 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 import AutoTokenizer
+from transformers.models.sam3.processing_sam3 import Sam3Processor
+from transformers.testing_utils import require_torch, require_vision
+
+from ...test_processing_common import ProcessorTesterMixin
+
+
+@require_torch
+@require_vision
+class Sam3ProcessorTest(ProcessorTesterMixin, unittest.TestCase):
+    processor_class = Sam3Processor
+
+    @classmethod
+    def _setup_tokenizer(cls):
+        return AutoTokenizer.from_pretrained("openai/clip-vit-base-patch32", max_length=32, model_max_length=32)
+
+    # Sam3Processor has a custom non-standard __call__ signature (no chat template, extra
+    # prompting args like input_boxes). Skip mixin tests that assume a standard VLM interface.
+    def test_chat_template_save_loading(self):
+        self.skipTest("Sam3Processor does not use a chat template")
+
+    def test_model_input_names(self):
+        self.skipTest("Sam3Processor outputs extra keys (e.g. original_sizes) beyond model_input_names")
+
+    def test_tokenizer_defaults(self):
+        self.skipTest("Sam3Processor always pads tokenizer output to max_length=32")
+
+    def test_processor_text_has_no_visual(self):
+        self.skipTest("Sam3Processor has a custom interface, not a standard VLM text+image interface")
+
+    def test_processor_with_multiple_inputs(self):
+        self.skipTest("Sam3Processor has a custom interface, not a standard VLM text+image interface")
+
+    # --- Sam3-specific tests ---
+
+    def test_input_boxes_default_labels_mixed_batch(self):
+        # Regression test for https://github.com/huggingface/transformers/issues/45059:
+        # None entries should get pad label (-10), real entries should get positive label (1).
+        processor = self.get_processor()
+        images = self.prepare_image_inputs(batch_size=2)
+        inputs = processor(
+            images=images,
+            text=["cat", None],
+            input_boxes=[None, [[100, 100, 200, 200]]],
+            return_tensors="pt",
+        )
+        self.assertIn("input_boxes", inputs)
+        self.assertIn("input_boxes_labels", inputs)
+
+        # The None entry (index 0) should have label -10 (pad value)
+        self.assertEqual(inputs["input_boxes_labels"][0, 0].item(), -10)
+        # The real entry (index 1) should have label 1 (positive)
+        self.assertEqual(inputs["input_boxes_labels"][1, 0].item(), 1)
+
+    def test_input_boxes_default_labels_all_real(self):
+        processor = self.get_processor()
+        images = self.prepare_image_inputs(batch_size=2)
+        inputs = processor(
+            images=images,
+            text=["cat", "dog"],
+            input_boxes=[[[50, 50, 150, 150]], [[200, 200, 300, 300]]],
+            return_tensors="pt",
+        )
+        self.assertIn("input_boxes_labels", inputs)
+        self.assertTrue((inputs["input_boxes_labels"] == 1).all())
+
+    def test_no_input_boxes_omits_labels(self):
+        processor = self.get_processor()
+        images = self.prepare_image_inputs(batch_size=1)
+        inputs = processor(
+            images=images,
+            text=["cat"],
+            return_tensors="pt",
+        )
+        self.assertNotIn("input_boxes", inputs)
+        self.assertNotIn("input_boxes_labels", inputs)
+
+    def test_user_provided_labels_preserved(self):
+        processor = self.get_processor()
+        images = self.prepare_image_inputs(batch_size=2)
+        inputs = processor(
+            images=images,
+            text=["cat", "dog"],
+            input_boxes=[[[50, 50, 150, 150]], [[200, 200, 300, 300]]],
+            input_boxes_labels=[[1], [0]],
+            return_tensors="pt",
+        )
+        self.assertEqual(inputs["input_boxes_labels"][0, 0].item(), 1)
+        self.assertEqual(inputs["input_boxes_labels"][1, 0].item(), 0)