Add DPT (#15991)

* First draft

* More improvements

* Add fusion blocks

* Make conversion script work for dpt_large

* Make conversion script work

* Improve implementation

* Improve conversion script

* Add DPTForSemanticSegmentation

* Make conversion work for semantic segmentation

* Add tests

* Remove print statements

* First draft

* Redesign neck

* Improve tests

* Improve implementation some more

* Make neck output list of tensors

* Improve neck and feature extractor

* Fix integration tests

* Make more tests pass

* Make all tests pass

* Add missing config archive map

* Add in_index attribute to make heads accept list of tensors

* Apply suggestions from code review

* Apply suggestions from code review
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Apply some more suggestions

* Add copied from statements

* Remove assert

* Apply suggestions from code review

* Apply suggestions from code review...

tests/dpt/test_modeling_dpt.py

+# coding=utf-8
+# 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 DPT model. """
+
+
+import inspect
+import unittest
+
+from transformers import DPTConfig
+from transformers.file_utils import is_torch_available, is_vision_available
+from transformers.models.auto import get_values
+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
+
+
+if is_torch_available():
+    import torch
+    from torch import nn
+
+    from transformers import MODEL_MAPPING, DPTForDepthEstimation, DPTForSemanticSegmentation, DPTModel
+    from transformers.models.dpt.modeling_dpt import DPT_PRETRAINED_MODEL_ARCHIVE_LIST
+
+
+if is_vision_available():
+    from PIL import Image
+
+    from transformers import DPTFeatureExtractor
+
+
+class DPTModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=2,
+        image_size=32,
+        patch_size=16,
+        num_channels=3,
+        is_training=True,
+        use_labels=True,
+        hidden_size=32,
+        num_hidden_layers=4,
+        backbone_out_indices=[0, 1, 2, 3],
+        num_attention_heads=4,
+        intermediate_size=37,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        initializer_range=0.02,
+        num_labels=3,
+        scope=None,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.image_size = image_size
+        self.patch_size = patch_size
+        self.num_channels = num_channels
+        self.is_training = is_training
+        self.use_labels = use_labels
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.backbone_out_indices = backbone_out_indices
+        self.num_attention_heads = num_attention_heads
+        self.intermediate_size = intermediate_size
+        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.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 DPTConfig(
+            image_size=self.image_size,
+            patch_size=self.patch_size,
+            num_channels=self.num_channels,
+            hidden_size=self.hidden_size,
+            num_hidden_layers=self.num_hidden_layers,
+            backbone_out_indices=self.backbone_out_indices,
+            num_attention_heads=self.num_attention_heads,
+            intermediate_size=self.intermediate_size,
+            hidden_act=self.hidden_act,
+            hidden_dropout_prob=self.hidden_dropout_prob,
+            attention_probs_dropout_prob=self.attention_probs_dropout_prob,
+            is_decoder=False,
+            initializer_range=self.initializer_range,
+        )
+
+    def create_and_check_model(self, config, pixel_values, labels):
+        model = DPTModel(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(pixel_values)
+        # expected sequence length = num_patches + 1 (we add 1 for the [CLS] token)
+        num_patches = (config.image_size // config.patch_size) ** 2
+        self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, num_patches + 1, self.hidden_size))
+
+    def create_and_check_for_depth_estimation(self, config, pixel_values, labels):
+        config.num_labels = self.num_labels
+        model = DPTForDepthEstimation(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))
+
+    def create_and_check_for_semantic_segmentation(self, config, pixel_values, labels):
+        config.num_labels = self.num_labels
+        model = DPTForSemanticSegmentation(config)
+        model.to(torch_device)
+        model.eval()
+        result = model(pixel_values, labels=labels)
+        self.parent.assertEqual(
+            result.logits.shape, (self.batch_size, self.num_labels, 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 DPTModelTest(ModelTesterMixin, unittest.TestCase):
+    """
+    Here we also overwrite some of the tests of test_modeling_common.py, as DPT does not use input_ids, inputs_embeds,
+    attention_mask and seq_length.
+    """
+
+    all_model_classes = (DPTModel, DPTForDepthEstimation, DPTForSemanticSegmentation) if is_torch_available() else ()
+
+    test_pruning = False
+    test_torchscript = False
+    test_resize_embeddings = False
+    test_head_masking = False
+
+    def setUp(self):
+        self.model_tester = DPTModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=DPTConfig, has_text_modality=False, hidden_size=37)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_inputs_embeds(self):
+        # DPT does not use inputs_embeds
+        pass
+
+    def test_model_common_attributes(self):
+        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            model = model_class(config)
+            self.assertIsInstance(model.get_input_embeddings(), (nn.Module))
+            x = model.get_output_embeddings()
+            self.assertTrue(x is None or isinstance(x, nn.Linear))
+
+    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)
+            # signature.parameters is an OrderedDict => so arg_names order is deterministic
+            arg_names = [*signature.parameters.keys()]
+
+            expected_arg_names = ["pixel_values"]
+            self.assertListEqual(arg_names[:1], expected_arg_names)
+
+    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_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)
+
+    def test_for_semantic_segmentation(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_semantic_segmentation(*config_and_inputs)
+
+    def test_attention_outputs(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+        config.return_dict = True
+
+        # in DPT, the seq_len equals the number of patches + 1 (we add 1 for the [CLS] token)
+        num_patches = (config.image_size // config.patch_size) ** 2
+        seq_len = num_patches + 1
+
+        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(config)
+            model.to(torch_device)
+            model.eval()
+            with torch.no_grad():
+                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
+
+            self.assertEqual(len(outputs.attentions), self.model_tester.num_hidden_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.attentions
+            self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)
+
+            self.assertListEqual(
+                list(attentions[0].shape[-3:]),
+                [self.model_tester.num_attention_heads, seq_len, 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), self.model_tester.num_hidden_layers)
+            self.assertListEqual(
+                list(self_attentions[0].shape[-3:]),
+                [self.model_tester.num_attention_heads, seq_len, 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 = getattr(
+                self.model_tester, "expected_num_hidden_layers", self.model_tester.num_hidden_layers + 1
+            )
+            self.assertEqual(len(hidden_states), expected_num_layers)
+
+            # DPT has a different seq_length
+            num_patches = (config.image_size // config.patch_size) ** 2
+            seq_len = num_patches + 1
+
+            self.assertListEqual(
+                list(hidden_states[0].shape[-2:]),
+                [seq_len, self.model_tester.hidden_size],
+            )
+
+        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):
+        for model_class in self.all_model_classes:
+            if model_class.__name__ == "DPTForDepthEstimation":
+                continue
+
+            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+            config.return_dict = True
+
+            if model_class in get_values(MODEL_MAPPING):
+                continue
+
+            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()
+
+    def test_training_gradient_checkpointing(self):
+        for model_class in self.all_model_classes:
+            if model_class.__name__ == "DPTForDepthEstimation":
+                continue
+
+            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+            config.use_cache = False
+            config.return_dict = True
+
+            if model_class in get_values(MODEL_MAPPING) or not model_class.supports_gradient_checkpointing:
+                continue
+            model = model_class(config)
+            model.to(torch_device)
+            model.gradient_checkpointing_enable()
+            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):
+        for model_name in DPT_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
+            model = DPTModel.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 DPTModelIntegrationTest(unittest.TestCase):
+    def test_inference_depth_estimation(self):
+        feature_extractor = DPTFeatureExtractor.from_pretrained("Intel/dpt-large")
+        model = DPTForDepthEstimation.from_pretrained("Intel/dpt-large").to(torch_device)
+
+        image = prepare_img()
+        inputs = feature_extractor(images=image, return_tensors="pt").to(torch_device)
+
+        # forward pass
+        with torch.no_grad():
+            outputs = model(**inputs)
+            predicted_depth = outputs.predicted_depth
+
+        # verify the predicted depth
+        expected_shape = torch.Size((1, 384, 384))
+        self.assertEqual(predicted_depth.shape, expected_shape)
+
+        expected_slice = torch.tensor(
+            [[6.3199, 6.3629, 6.4148], [6.3850, 6.3615, 6.4166], [6.3519, 6.3176, 6.3575]]
+        ).to(torch_device)
+
+        self.assertTrue(torch.allclose(outputs.predicted_depth[0, :3, :3], expected_slice, atol=1e-4))
+
+    def test_inference_semantic_segmentation(self):
+        feature_extractor = DPTFeatureExtractor.from_pretrained("Intel/dpt-large-ade")
+        model = DPTForSemanticSegmentation.from_pretrained("Intel/dpt-large-ade").to(torch_device)
+
+        image = prepare_img()
+        inputs = feature_extractor(images=image, return_tensors="pt").to(torch_device)
+
+        # forward pass
+        with torch.no_grad():
+            outputs = model(**inputs)
+
+        # verify the logits
+        expected_shape = torch.Size((1, 150, 480, 480))
+        self.assertEqual(outputs.logits.shape, expected_shape)
+
+        expected_slice = torch.tensor(
+            [[4.0480, 4.2420, 4.4360], [4.3124, 4.5693, 4.8261], [4.5768, 4.8965, 5.2163]]
+        ).to(torch_device)
+
+        self.assertTrue(torch.allclose(outputs.logits[0, 0, :3, :3], expected_slice, atol=1e-4))

tests/test_modeling_common.py

@@ -79,6 +79,7 @@ if is_torch_available():
         MODEL_FOR_MULTIPLE_CHOICE_MAPPING,
         MODEL_FOR_NEXT_SENTENCE_PREDICTION_MAPPING,
         MODEL_FOR_QUESTION_ANSWERING_MAPPING,
+        MODEL_FOR_SEMANTIC_SEGMENTATION_MAPPING,
         MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING,
         MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING,
         MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING,
@@ -177,6 +178,11 @@ class ModelTesterMixin:
                 inputs_dict["bool_masked_pos"] = torch.zeros(
                     (self.model_tester.batch_size, num_patches**2), dtype=torch.long, device=torch_device
                 )
+            elif model_class in get_values(MODEL_FOR_SEMANTIC_SEGMENTATION_MAPPING):
+                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()
 
         return inputs_dict
 

utils/check_repo.py

@@ -117,6 +117,7 @@ TEST_FILES_WITH_NO_COMMON_TESTS = [
 # should **not** be the rule.
 IGNORE_NON_AUTO_CONFIGURED = PRIVATE_MODELS.copy() + [
     # models to ignore for model xxx mapping
+    "DPTForDepthEstimation",
     "DecisionTransformerGPT2Model",
     "GLPNForDepthEstimation",
     "ViltForQuestionAnswering",

utils/documentation_tests.txt

@@ -16,6 +16,7 @@ src/transformers/models/blenderbot_small/modeling_blenderbot_small.py
 src/transformers/models/convnext/modeling_convnext.py
 src/transformers/models/data2vec/modeling_data2vec_audio.py
 src/transformers/models/deit/modeling_deit.py
+src/transformers/models/dpt/modeling_dpt.py
 src/transformers/models/glpn/modeling_glpn.py
 src/transformers/models/hubert/modeling_hubert.py
 src/transformers/models/marian/modeling_marian.py
@@ -48,4 +49,4 @@ src/transformers/models/vit_mae/modeling_vit_mae.py
 src/transformers/models/wav2vec2/modeling_wav2vec2.py
 src/transformers/models/wav2vec2/tokenization_wav2vec2.py
 src/transformers/models/wav2vec2_with_lm/processing_wav2vec2_with_lm.py
-src/transformers/models/wavlm/modeling_wavlm.py
+src/transformers/models/wavlm/modeling_wavlm.py
\ No newline at end of file