first commit

tests/models/glpn/test_image_processing_glpn.py

@@ -0,0 +1,222 @@
+# Copyright 2022 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
+
+import numpy as np
+
+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():
+    from PIL import Image
+
+
+class GLPNImageProcessingTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=7,
+        num_channels=3,
+        image_size=18,
+        min_resolution=30,
+        max_resolution=400,
+        do_resize=True,
+        size_divisor=32,
+        do_rescale=True,
+    ):
+        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_divisor = size_divisor
+        self.do_rescale = do_rescale
+
+    def prepare_image_processor_dict(self):
+        return {
+            "do_resize": self.do_resize,
+            "size_divisor": self.size_divisor,
+            "do_rescale": self.do_rescale,
+        }
+
+    def expected_output_image_shape(self, images):
+        if isinstance(images[0], Image.Image):
+            width, height = images[0].size
+        elif isinstance(images[0], np.ndarray):
+            height, width = images[0].shape[0], images[0].shape[1]
+        else:
+            height, width = images[0].shape[1], images[0].shape[2]
+
+        height = height // self.size_divisor * self.size_divisor
+        width = width // self.size_divisor * self.size_divisor
+
+        return self.num_channels, height, 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,
+            size_divisor=self.size_divisor,
+            equal_resolution=equal_resolution,
+            numpify=numpify,
+            torchify=torchify,
+        )
+
+    def prepare_depth_outputs(self):
+        if not is_torch_available():
+            return None
+        depth_tensors = prepare_image_inputs(
+            batch_size=self.batch_size,
+            num_channels=1,
+            min_resolution=self.min_resolution,
+            max_resolution=self.max_resolution,
+            equal_resolution=True,
+            torchify=True,
+        )
+        depth_tensors = [depth_tensor.squeeze(0) for depth_tensor in depth_tensors]
+        stacked_depth_tensors = torch.stack(depth_tensors, dim=0)
+        return type("DepthOutput", (), {"predicted_depth": stacked_depth_tensors})
+
+
+@require_torch
+@require_vision
+class GLPNImageProcessingTest(ImageProcessingTestMixin, unittest.TestCase):
+    def setUp(self):
+        super().setUp()
+        self.image_processor_tester = GLPNImageProcessingTester(self)
+        self.image_processor_dict = 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, "do_resize"))
+            self.assertTrue(hasattr(image_processing, "size_divisor"))
+            self.assertTrue(hasattr(image_processing, "resample"))
+            self.assertTrue(hasattr(image_processing, "do_rescale"))
+
+    def test_call_pil(self):
+        # Initialize image_processing
+        for image_processing_class in self.image_processing_classes.values():
+            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 (GLPNImageProcessor doesn't support batching)
+            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)
+            self.assertTrue(tuple(encoded_images.shape) == (1, *expected_output_image_shape))
+
+    def test_call_numpy(self):
+        # Initialize image_processing
+        for image_processing_class in self.image_processing_classes.values():
+            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 (GLPNImageProcessor doesn't support batching)
+            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)
+            self.assertTrue(tuple(encoded_images.shape) == (1, *expected_output_image_shape))
+
+    def test_call_pytorch(self):
+        # Initialize image_processing
+        for image_processing_class in self.image_processing_classes.values():
+            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 (GLPNImageProcessor doesn't support batching)
+            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)
+            self.assertTrue(tuple(encoded_images.shape) == (1, *expected_output_image_shape))
+
+    def test_call_numpy_4_channels(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_processing_class.num_channels = 4
+            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 (GLPNImageProcessor doesn't support batching)
+            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)
+            self.assertTrue(tuple(encoded_images.shape) == (1, *expected_output_image_shape))
+            image_processing_class.num_channels = 3
+
+    # Override as GLPN image processors don't support heterogeneous batching (use equal_resolution=True)
+    @require_vision
+    @require_torch
+    def test_backends_equivalence_batched(self):
+        if len(self.image_processing_classes) < 2:
+            self.skipTest(reason="Skipping backends equivalence test as there are less than 2 backends")
+
+        dummy_images = self.image_processor_tester.prepare_image_inputs(equal_resolution=True, torchify=True)
+
+        # Create processors for each backend
+        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, return_tensors="pt")
+
+        # Compare all backends to the first one (reference backend)
+        backend_names = list(encodings.keys())
+        reference_backend = backend_names[0]
+        reference_encoding = encodings[reference_backend].pixel_values
+        for backend_name in backend_names[1:]:
+            self._assert_tensors_equivalence(reference_encoding, encodings[backend_name].pixel_values)
+
+    @require_vision
+    @require_torch
+    def test_backends_equivalence_post_process_depth(self):
+        """Check that all backends produce equivalent post-processed depth maps."""
+        if len(self.image_processing_classes) < 2:
+            self.skipTest(reason="Skipping backends equivalence test as there are less than 2 backends")
+
+        outputs = self.image_processor_tester.prepare_depth_outputs()
+        target_sizes = [(240, 320)] * self.image_processor_tester.batch_size
+
+        # Create processors and run post-processing for each backend
+        processed = {}
+        for backend_name, image_processing_class in self.image_processing_classes.items():
+            image_processor = image_processing_class(**self.image_processor_dict)
+            processed[backend_name] = image_processor.post_process_depth_estimation(outputs, target_sizes=target_sizes)
+
+        # Compare all backends to the first one (reference backend)
+        backend_names = list(processed.keys())
+        reference_backend = backend_names[0]
+        for backend_name in backend_names[1:]:
+            for pred_ref, pred_other in zip(processed[reference_backend], processed[backend_name]):
+                depth_ref = pred_ref["predicted_depth"].float()
+                depth_other = pred_other["predicted_depth"].float()
+                self._assert_tensors_equivalence(depth_ref, depth_other)

tests/models/glpn/test_modeling_glpn.py

@@ -0,0 +1,346 @@
+# Copyright 2022 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 GLPN model."""
+
+import unittest
+
+from transformers import is_torch_available, is_vision_available
+from transformers.testing_utils import require_torch, require_vision, slow, torch_device
+
+from ...test_configuration_common import ConfigTester
+from ...test_modeling_common import ModelTesterMixin, floats_tensor, ids_tensor
+from ...test_pipeline_mixin import PipelineTesterMixin
+
+
+if is_torch_available():
+    import torch
+
+    from transformers import GLPNConfig, GLPNForDepthEstimation, GLPNModel
+    from transformers.models.auto.modeling_auto import MODEL_MAPPING_NAMES
+
+
+if is_vision_available():
+    from PIL import Image
+
+    from transformers import GLPNImageProcessorPil
+
+
+class GLPNConfigTester(ConfigTester):
+    def create_and_test_config_common_properties(self):
+        config = self.config_class(**self.inputs_dict)
+        self.parent.assertTrue(hasattr(config, "hidden_sizes"))
+        self.parent.assertTrue(hasattr(config, "num_attention_heads"))
+        self.parent.assertTrue(hasattr(config, "num_encoder_blocks"))
+
+
+class GLPNModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=13,
+        image_size=64,
+        num_channels=3,
+        num_encoder_blocks=4,
+        depths=[2, 2, 2, 2],
+        sr_ratios=[8, 4, 2, 1],
+        hidden_sizes=[16, 32, 64, 128],
+        downsampling_rates=[1, 4, 8, 16],
+        num_attention_heads=[1, 2, 4, 8],
+        is_training=True,
+        use_labels=True,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        initializer_range=0.02,
+        decoder_hidden_size=16,
+        num_labels=3,
+        scope=None,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.image_size = image_size
+        self.num_channels = num_channels
+        self.num_encoder_blocks = num_encoder_blocks
+        self.sr_ratios = sr_ratios
+        self.depths = depths
+        self.hidden_sizes = hidden_sizes
+        self.downsampling_rates = downsampling_rates
+        self.num_attention_heads = num_attention_heads
+        self.is_training = is_training
+        self.use_labels = use_labels
+        self.hidden_act = hidden_act
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.initializer_range = initializer_range
+        self.decoder_hidden_size = decoder_hidden_size
+        self.num_labels = num_labels
+        self.scope = scope
+
+    def prepare_config_and_inputs(self):
+        pixel_values = floats_tensor([self.batch_size, self.num_channels, self.image_size, self.image_size])
+
+        labels = None
+        if self.use_labels:
+            labels = ids_tensor([self.batch_size, self.image_size, self.image_size], self.num_labels)
+
+        config = self.get_config()
+        return config, pixel_values, labels
+
+    def get_config(self):
+        return GLPNConfig(
+            image_size=self.image_size,
+            num_channels=self.num_channels,
+            num_encoder_blocks=self.num_encoder_blocks,
+            depths=self.depths,
+            hidden_sizes=self.hidden_sizes,
+            num_attention_heads=self.num_attention_heads,
+            hidden_act=self.hidden_act,
+            hidden_dropout_prob=self.hidden_dropout_prob,
+            attention_probs_dropout_prob=self.attention_probs_dropout_prob,
+            initializer_range=self.initializer_range,
+            decoder_hidden_size=self.decoder_hidden_size,
+        )
+
+    def create_and_check_model(self, config, pixel_values, labels):
+        model = GLPNModel(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(pixel_values)
+        expected_height = expected_width = self.image_size // (self.downsampling_rates[-1] * 2)
+        self.parent.assertEqual(
+            result.last_hidden_state.shape, (self.batch_size, self.hidden_sizes[-1], expected_height, expected_width)
+        )
+
+    def create_and_check_for_depth_estimation(self, config, pixel_values, labels):
+        config.num_labels = self.num_labels
+        model = GLPNForDepthEstimation(config)
+        model.to(torch_device)
+        model.eval()
+        result = model(pixel_values)
+        self.parent.assertEqual(result.predicted_depth.shape, (self.batch_size, self.image_size, self.image_size))
+        result = model(pixel_values, labels=labels)
+        self.parent.assertEqual(result.predicted_depth.shape, (self.batch_size, self.image_size, self.image_size))
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        config, pixel_values, labels = config_and_inputs
+        inputs_dict = {"pixel_values": pixel_values}
+        return config, inputs_dict
+
+
+@require_torch
+class GLPNModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
+    all_model_classes = (GLPNModel, GLPNForDepthEstimation) if is_torch_available() else ()
+    pipeline_model_mapping = (
+        {"depth-estimation": GLPNForDepthEstimation, "image-feature-extraction": GLPNModel}
+        if is_torch_available()
+        else {}
+    )
+
+    test_resize_embeddings = False
+
+    def setUp(self):
+        self.model_tester = GLPNModelTester(self)
+        self.config_tester = GLPNConfigTester(self, config_class=GLPNConfig)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_model(*config_and_inputs)
+
+    def test_batching_equivalence(self, atol=3e-4, rtol=3e-4):
+        super().test_batching_equivalence(atol=atol, rtol=rtol)
+
+    def test_for_depth_estimation(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_depth_estimation(*config_and_inputs)
+
+    @unittest.skip(reason="GLPN does not use inputs_embeds")
+    def test_inputs_embeds(self):
+        pass
+
+    @unittest.skip(reason="GLPN does not have get_input_embeddings method and get_output_embeddings methods")
+    def test_model_get_set_embeddings(self):
+        pass
+
+    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.attentions
+
+            expected_num_attentions = sum(self.model_tester.depths)
+            self.assertEqual(len(attentions), expected_num_attentions)
+
+            # 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.attentions
+
+            self.assertEqual(len(attentions), expected_num_attentions)
+
+            # verify the first attentions (first block, first layer)
+            expected_seq_len = (self.model_tester.image_size // 4) ** 2
+            expected_reduced_seq_len = (self.model_tester.image_size // (4 * self.model_tester.sr_ratios[0])) ** 2
+            self.assertListEqual(
+                list(attentions[0].shape[-3:]),
+                [self.model_tester.num_attention_heads[0], expected_seq_len, expected_reduced_seq_len],
+            )
+
+            # verify the last attentions (last block, last layer)
+            expected_seq_len = (self.model_tester.image_size // 32) ** 2
+            expected_reduced_seq_len = (self.model_tester.image_size // (32 * self.model_tester.sr_ratios[-1])) ** 2
+            self.assertListEqual(
+                list(attentions[-1].shape[-3:]),
+                [self.model_tester.num_attention_heads[-1], expected_seq_len, expected_reduced_seq_len],
+            )
+            out_len = len(outputs)
+
+            # 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))
+
+            self.assertEqual(out_len + 1, len(outputs))
+
+            self_attentions = outputs.attentions
+
+            self.assertEqual(len(self_attentions), expected_num_attentions)
+            # verify the first attentions (first block, first layer)
+            expected_seq_len = (self.model_tester.image_size // 4) ** 2
+            expected_reduced_seq_len = (self.model_tester.image_size // (4 * self.model_tester.sr_ratios[0])) ** 2
+            self.assertListEqual(
+                list(self_attentions[0].shape[-3:]),
+                [self.model_tester.num_attention_heads[0], expected_seq_len, expected_reduced_seq_len],
+            )
+
+    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.hidden_states
+
+            expected_num_layers = self.model_tester.num_encoder_blocks
+            self.assertEqual(len(hidden_states), expected_num_layers)
+
+            # verify the first hidden states (first block)
+            self.assertListEqual(
+                list(hidden_states[0].shape[-3:]),
+                [
+                    self.model_tester.hidden_sizes[0],
+                    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, 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_training(self):
+        if not self.model_tester.is_training:
+            self.skipTest(reason="model_tester.is_training is set to False")
+
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+        config.return_dict = True
+
+        for model_class in self.all_model_classes:
+            if model_class.__name__ in MODEL_MAPPING_NAMES.values():
+                continue
+            # TODO: remove the following 3 lines once we have a MODEL_FOR_DEPTH_ESTIMATION_MAPPING
+            # this can then be incorporated into _prepare_for_class in test_modeling_common.py
+            if model_class.__name__ == "GLPNForDepthEstimation":
+                batch_size, num_channels, height, width = inputs_dict["pixel_values"].shape
+                inputs_dict["labels"] = torch.zeros(
+                    [self.model_tester.batch_size, height, width], device=torch_device
+                ).long()
+            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()
+
+    @slow
+    def test_model_from_pretrained(self):
+        model_name = "vinvino02/glpn-kitti"
+        model = GLPNModel.from_pretrained(model_name)
+        self.assertIsNotNone(model)
+
+
+# We will verify our results on an image of cute cats
+def prepare_img():
+    image = Image.open("./tests/fixtures/tests_samples/COCO/000000039769.png")
+    return image
+
+
+@require_torch
+@require_vision
+@slow
+class GLPNModelIntegrationTest(unittest.TestCase):
+    @slow
+    def test_inference_depth_estimation(self):
+        image_processor = GLPNImageProcessorPil.from_pretrained("vinvino02/glpn-kitti")
+        model = GLPNForDepthEstimation.from_pretrained("vinvino02/glpn-kitti").to(torch_device)
+
+        image = prepare_img()
+        inputs = image_processor(images=image, return_tensors="pt").to(torch_device)
+
+        # forward pass
+        with torch.no_grad():
+            outputs = model(**inputs)
+
+        # verify the predicted depth
+        expected_shape = torch.Size([1, 480, 640])
+        self.assertEqual(outputs.predicted_depth.shape, expected_shape)
+
+        expected_slice = torch.tensor(
+            [[3.4291, 2.7865, 2.5151], [3.2841, 2.7021, 2.3502], [3.1147, 2.4625, 2.2481]]
+        ).to(torch_device)
+
+        torch.testing.assert_close(outputs.predicted_depth[0, :3, :3], expected_slice, rtol=1e-4, atol=1e-4)