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

* Remove DPTInterpolate in favor of nn.Upsample

* Add comments

* Apply suggestions from code review

* Apply suggestions from code review

* Add proposed design

* Update design

* Add DPTReassembleLayer

* Add DPTFeatureFusionStage

* Apply more suggestions from code review

* Apply suggestions from code review

* Apply suggestions from code review

* Fix rebase

* Update in_index and out_indices

* Fix conversion script

* Fix code quality

* Add model to toctree and use DepthEstimatorOutput

* Fix rebase

* Fix code examples

* Improve code

* Fix copied from statements

* Apply suggestions from code review

* Remove compute_loss method

* Apply suggestions from code review

* Fix documentation tests file

* Remove test.py file

* Improve doc example
Co-authored-by: Niels Rogge <nielsrogge@Nielss-MacBook-Pro.local>
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
Co-authored-by: Niels Rogge <nielsrogge@nielss-mbp.home>

README.md

@@ -261,6 +261,7 @@ Current number of checkpoints: ![](https://img.shields.io/endpoint?url=https://h
 1. **[DPR](https://huggingface.co/docs/transformers/model_doc/dpr)** (from Facebook) released with the paper [Dense Passage Retrieval
 for Open-Domain Question Answering](https://arxiv.org/abs/2004.04906) by Vladimir Karpukhin, Barlas Oğuz, Sewon
 Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih.
+1. **[DPT](https://huggingface.co/docs/transformers/master/model_doc/dpt)** (from Intel Labs) released with the paper [Vision Transformers for Dense Prediction](https://arxiv.org/abs/2103.13413) by René Ranftl, Alexey Bochkovskiy, Vladlen Koltun.
 1. **[EncoderDecoder](https://huggingface.co/docs/transformers/model_doc/encoder-decoder)** (from Google Research) released with the paper [Leveraging Pre-trained Checkpoints for Sequence Generation Tasks](https://arxiv.org/abs/1907.12461) by Sascha Rothe, Shashi Narayan, Aliaksei Severyn.
 1. **[ELECTRA](https://huggingface.co/docs/transformers/model_doc/electra)** (from Google Research/Stanford University) released with the paper [ELECTRA: Pre-training text encoders as discriminators rather than generators](https://arxiv.org/abs/2003.10555) by Kevin Clark, Minh-Thang Luong, Quoc V. Le, Christopher D. Manning.
 1. **[FlauBERT](https://huggingface.co/docs/transformers/model_doc/flaubert)** (from CNRS) released with the paper [FlauBERT: Unsupervised Language Model Pre-training for French](https://arxiv.org/abs/1912.05372) by Hang Le, Loïc Vial, Jibril Frej, Vincent Segonne, Maximin Coavoux, Benjamin Lecouteux, Alexandre Allauzen, Benoît Crabbé, Laurent Besacier, Didier Schwab.

README_ko.md

@@ -240,6 +240,7 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
 1. **[DistilBERT](https://huggingface.co/docs/transformers/model_doc/distilbert)** (from HuggingFace), released together with the paper [DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter](https://arxiv.org/abs/1910.01108) by Victor Sanh, Lysandre Debut and Thomas Wolf. The same method has been applied to compress GPT2 into [DistilGPT2](https://github.com/huggingface/transformers/tree/main/examples/distillation), RoBERTa into [DistilRoBERTa](https://github.com/huggingface/transformers/tree/main/examples/distillation), Multilingual BERT into [DistilmBERT](https://github.com/huggingface/transformers/tree/main/examples/distillation) and a German version of DistilBERT.
 1. **[DiT](https://huggingface.co/docs/transformers/model_doc/dit)** (from Microsoft Research) released with the paper [DiT: Self-supervised Pre-training for Document Image Transformer](https://arxiv.org/abs/2203.02378) by Junlong Li, Yiheng Xu, Tengchao Lv, Lei Cui, Cha Zhang, Furu Wei.
 1. **[DPR](https://huggingface.co/docs/transformers/model_doc/dpr)** (from Facebook) released with the paper [Dense Passage Retrieval for Open-Domain Question Answering](https://arxiv.org/abs/2004.04906) by Vladimir Karpukhin, Barlas Oğuz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih.
+1. **[DPT](https://huggingface.co/docs/transformers/master/model_doc/dpt)** (from Intel Labs) released with the paper [Vision Transformers for Dense Prediction](https://arxiv.org/abs/2103.13413) by René Ranftl, Alexey Bochkovskiy, Vladlen Koltun.
 1. **[ELECTRA](https://huggingface.co/docs/transformers/model_doc/electra)** (from Google Research/Stanford University) released with the paper [ELECTRA: Pre-training text encoders as discriminators rather than generators](https://arxiv.org/abs/2003.10555) by Kevin Clark, Minh-Thang Luong, Quoc V. Le, Christopher D. Manning.
 1. **[EncoderDecoder](https://huggingface.co/docs/transformers/model_doc/encoder-decoder)** (from Google Research) released with the paper [Leveraging Pre-trained Checkpoints for Sequence Generation Tasks](https://arxiv.org/abs/1907.12461) by Sascha Rothe, Shashi Narayan, Aliaksei Severyn.
 1. **[FlauBERT](https://huggingface.co/docs/transformers/model_doc/flaubert)** (from CNRS) released with the paper [FlauBERT: Unsupervised Language Model Pre-training for French](https://arxiv.org/abs/1912.05372) by Hang Le, Loïc Vial, Jibril Frej, Vincent Segonne, Maximin Coavoux, Benjamin Lecouteux, Alexandre Allauzen, Benoît Crabbé, Laurent Besacier, Didier Schwab.

README_zh-hans.md

@@ -264,6 +264,7 @@ conda install -c huggingface transformers
 1. **[DistilBERT](https://huggingface.co/docs/transformers/model_doc/distilbert)** (来自 HuggingFace), 伴随论文 [DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter](https://arxiv.org/abs/1910.01108) 由 Victor Sanh, Lysandre Debut and Thomas Wolf 发布。 同样的方法也应用于压缩 GPT-2 到 [DistilGPT2](https://github.com/huggingface/transformers/tree/main/examples/distillation), RoBERTa 到 [DistilRoBERTa](https://github.com/huggingface/transformers/tree/main/examples/distillation), Multilingual BERT 到 [DistilmBERT](https://github.com/huggingface/transformers/tree/main/examples/distillation) 和德语版 DistilBERT。
 1. **[DiT](https://huggingface.co/docs/transformers/model_doc/dit)** (来自 Microsoft Research) 伴随论文 [DiT: Self-supervised Pre-training for Document Image Transformer](https://arxiv.org/abs/2203.02378) 由 Junlong Li, Yiheng Xu, Tengchao Lv, Lei Cui, Cha Zhang, Furu Wei 发布。
 1. **[DPR](https://huggingface.co/docs/transformers/model_doc/dpr)** (来自 Facebook) 伴随论文 [Dense Passage Retrieval for Open-Domain Question Answering](https://arxiv.org/abs/2004.04906) 由 Vladimir Karpukhin, Barlas Oğuz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih 发布。
+1. **[DPT](https://huggingface.co/docs/transformers/master/model_doc/dpt)** (来自 Intel Labs) 伴随论文 [Vision Transformers for Dense Prediction](https://arxiv.org/abs/2103.13413) 由 René Ranftl, Alexey Bochkovskiy, Vladlen Koltun 发布。
 1. **[ELECTRA](https://huggingface.co/docs/transformers/model_doc/electra)** (来自 Google Research/Stanford University) 伴随论文 [ELECTRA: Pre-training text encoders as discriminators rather than generators](https://arxiv.org/abs/2003.10555) 由 Kevin Clark, Minh-Thang Luong, Quoc V. Le, Christopher D. Manning 发布。
 1. **[EncoderDecoder](https://huggingface.co/docs/transformers/model_doc/encoder-decoder)** (来自 Google Research) 伴随论文 [Leveraging Pre-trained Checkpoints for Sequence Generation Tasks](https://arxiv.org/abs/1907.12461) 由 Sascha Rothe, Shashi Narayan, Aliaksei Severyn 发布。
 1. **[FlauBERT](https://huggingface.co/docs/transformers/model_doc/flaubert)** (来自 CNRS) 伴随论文 [FlauBERT: Unsupervised Language Model Pre-training for French](https://arxiv.org/abs/1912.05372) 由 Hang Le, Loïc Vial, Jibril Frej, Vincent Segonne, Maximin Coavoux, Benjamin Lecouteux, Alexandre Allauzen, Benoît Crabbé, Laurent Besacier, Didier Schwab 发布。

README_zh-hant.md

@@ -276,6 +276,7 @@ conda install -c huggingface transformers
 1. **[DistilBERT](https://huggingface.co/docs/transformers/model_doc/distilbert)** (from HuggingFace), released together with the paper [DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter](https://arxiv.org/abs/1910.01108) by Victor Sanh, Lysandre Debut and Thomas Wolf. The same method has been applied to compress GPT2 into [DistilGPT2](https://github.com/huggingface/transformers/tree/main/examples/distillation), RoBERTa into [DistilRoBERTa](https://github.com/huggingface/transformers/tree/main/examples/distillation), Multilingual BERT into [DistilmBERT](https://github.com/huggingface/transformers/tree/main/examples/distillation) and a German version of DistilBERT.
 1. **[DiT](https://huggingface.co/docs/transformers/model_doc/dit)** (from Microsoft Research) released with the paper [DiT: Self-supervised Pre-training for Document Image Transformer](https://arxiv.org/abs/2203.02378) by Junlong Li, Yiheng Xu, Tengchao Lv, Lei Cui, Cha Zhang, Furu Wei.
 1. **[DPR](https://huggingface.co/docs/transformers/model_doc/dpr)** (from Facebook) released with the paper [Dense Passage Retrieval for Open-Domain Question Answering](https://arxiv.org/abs/2004.04906) by Vladimir Karpukhin, Barlas Oğuz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih.
+1. **[DPT](https://huggingface.co/docs/transformers/master/model_doc/dpt)** (from Intel Labs) released with the paper [Vision Transformers for Dense Prediction](https://arxiv.org/abs/2103.13413) by René Ranftl, Alexey Bochkovskiy, Vladlen Koltun.
 1. **[ELECTRA](https://huggingface.co/docs/transformers/model_doc/electra)** (from Google Research/Stanford University) released with the paper [ELECTRA: Pre-training text encoders as discriminators rather than generators](https://arxiv.org/abs/2003.10555) by Kevin Clark, Minh-Thang Luong, Quoc V. Le, Christopher D. Manning.
 1. **[EncoderDecoder](https://huggingface.co/docs/transformers/model_doc/encoder-decoder)** (from Google Research) released with the paper [Leveraging Pre-trained Checkpoints for Sequence Generation Tasks](https://arxiv.org/abs/1907.12461) by Sascha Rothe, Shashi Narayan, Aliaksei Severyn.
 1. **[FlauBERT](https://huggingface.co/docs/transformers/model_doc/flaubert)** (from CNRS) released with the paper [FlauBERT: Unsupervised Language Model Pre-training for French](https://arxiv.org/abs/1912.05372) by Hang Le, Loïc Vial, Jibril Frej, Vincent Segonne, Maximin Coavoux, Benjamin Lecouteux, Alexandre Allauzen, Benoît Crabbé, Laurent Besacier, Didier Schwab.

docs/source/_toctree.yml

@@ -206,6 +206,8 @@
       title: DiT
     - local: model_doc/dpr
       title: DPR
+    - local: model_doc/dpt
+      title: DPT
     - local: model_doc/electra
       title: ELECTRA
     - local: model_doc/encoder-decoder

docs/source/index.mdx

@@ -82,6 +82,7 @@ The library currently contains JAX, PyTorch and TensorFlow implementations, pret
 1. **[DialoGPT](model_doc/dialogpt)** (from Microsoft Research) released with the paper [DialoGPT: Large-Scale Generative Pre-training for Conversational Response Generation](https://arxiv.org/abs/1911.00536) by Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing Liu, Bill Dolan.
 1. **[DistilBERT](model_doc/distilbert)** (from HuggingFace), released together with the paper [DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter](https://arxiv.org/abs/1910.01108) by Victor Sanh, Lysandre Debut and Thomas Wolf. The same method has been applied to compress GPT2 into [DistilGPT2](https://github.com/huggingface/transformers/tree/main/examples/research_projects/distillation), RoBERTa into [DistilRoBERTa](https://github.com/huggingface/transformers/tree/main/examples/research_projects/distillation), Multilingual BERT into [DistilmBERT](https://github.com/huggingface/transformers/tree/main/examples/research_projects/distillation) and a German version of DistilBERT.
 1. **[DPR](model_doc/dpr)** (from Facebook) released with the paper [Dense Passage Retrieval for Open-Domain Question Answering](https://arxiv.org/abs/2004.04906) by Vladimir Karpukhin, Barlas Oğuz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih.
+1. **[DPT](master/model_doc/dpt)** (from Intel Labs) released with the paper [Vision Transformers for Dense Prediction](https://arxiv.org/abs/2103.13413) by René Ranftl, Alexey Bochkovskiy, Vladlen Koltun.
 1. **[EncoderDecoder](model_doc/encoder-decoder)** (from Google Research) released with the paper [Leveraging Pre-trained Checkpoints for Sequence Generation Tasks](https://arxiv.org/abs/1907.12461) by Sascha Rothe, Shashi Narayan, Aliaksei Severyn.
 1. **[ELECTRA](model_doc/electra)** (from Google Research/Stanford University) released with the paper [ELECTRA: Pre-training text encoders as discriminators rather than generators](https://arxiv.org/abs/2003.10555) by Kevin Clark, Minh-Thang Luong, Quoc V. Le, Christopher D. Manning.
 1. **[FlauBERT](model_doc/flaubert)** (from CNRS) released with the paper [FlauBERT: Unsupervised Language Model Pre-training for French](https://arxiv.org/abs/1912.05372) by Hang Le, Loïc Vial, Jibril Frej, Vincent Segonne, Maximin Coavoux, Benjamin Lecouteux, Alexandre Allauzen, Benoît Crabbé, Laurent Besacier, Didier Schwab.
@@ -194,6 +195,7 @@ Flax), PyTorch, and/or TensorFlow.
 |            DETR             |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |         DistilBERT          |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
 |             DPR             |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
+|             DPT             |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |           ELECTRA           |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
 |       Encoder decoder       |       ❌       |       ❌       |       ✅        |         ✅         |      ✅      |
 | FairSeq Machine-Translation |       ✅       |       ❌       |       ✅        |         ❌         |      ❌      |

docs/source/model_doc/dpt.mdx

+<!--Copyright 2022 The HuggingFace 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.
+-->
+
+# DPT
+
+## Overview
+
+The DPT model was proposed in [Vision Transformers for Dense Prediction](https://arxiv.org/abs/2103.13413) by René Ranftl, Alexey Bochkovskiy, Vladlen Koltun.
+DPT is a model that leverages the [Vision Transformer (ViT)](vit) as backbone for dense prediction tasks like semantic segmentation and depth estimation.
+
+The abstract from the paper is the following:
+
+*We introduce dense vision transformers, an architecture that leverages vision transformers in place of convolutional networks as a backbone for dense prediction tasks. We assemble tokens from various stages of the vision transformer into image-like representations at various resolutions and progressively combine them into full-resolution predictions using a convolutional decoder. The transformer backbone processes representations at a constant and relatively high resolution and has a global receptive field at every stage. These properties allow the dense vision transformer to provide finer-grained and more globally coherent predictions when compared to fully-convolutional networks. Our experiments show that this architecture yields substantial improvements on dense prediction tasks, especially when a large amount of training data is available. For monocular depth estimation, we observe an improvement of up to 28% in relative performance when compared to a state-of-the-art fully-convolutional network. When applied to semantic segmentation, dense vision transformers set a new state of the art on ADE20K with 49.02% mIoU. We further show that the architecture can be fine-tuned on smaller datasets such as NYUv2, KITTI, and Pascal Context where it also sets the new state of the art.*
+
+<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/dpt_architecture.jpg"
+alt="drawing" width="600"/> 
+
+<small> DPT architecture. Taken from the <a href="https://arxiv.org/abs/2103.13413" target="_blank">original paper</a>. </small>
+
+This model was contributed by [nielsr](https://huggingface.co/nielsr). The original code can be found [here](https://github.com/isl-org/DPT).
+
+## DPTConfig
+
+[[autodoc]] DPTConfig
+
+
+## DPTFeatureExtractor
+
+[[autodoc]] DPTFeatureExtractor
+    - __call__
+
+
+## DPTModel
+
+[[autodoc]] DPTModel
+    - forward
+
+
+## DPTForDepthEstimation
+
+[[autodoc]] DPTForDepthEstimation
+    - forward
+
+
+## DPTForSemanticSegmentation
+
+[[autodoc]] DPTForSemanticSegmentation
+    - forward
\ No newline at end of file

src/transformers/__init__.py

@@ -187,6 +187,7 @@ _import_structure = {
         "DPRReaderOutput",
         "DPRReaderTokenizer",
     ],
+    "models.dpt": ["DPT_PRETRAINED_CONFIG_ARCHIVE_MAP", "DPTConfig"],
     "models.electra": ["ELECTRA_PRETRAINED_CONFIG_ARCHIVE_MAP", "ElectraConfig", "ElectraTokenizer"],
     "models.encoder_decoder": ["EncoderDecoderConfig"],
     "models.flaubert": ["FLAUBERT_PRETRAINED_CONFIG_ARCHIVE_MAP", "FlaubertConfig", "FlaubertTokenizer"],
@@ -528,6 +529,7 @@ if is_vision_available():
     _import_structure["models.convnext"].append("ConvNextFeatureExtractor")
     _import_structure["models.deit"].append("DeiTFeatureExtractor")
     _import_structure["models.detr"].append("DetrFeatureExtractor")
+    _import_structure["models.dpt"].append("DPTFeatureExtractor")
     _import_structure["models.glpn"].append("GLPNFeatureExtractor")
     _import_structure["models.imagegpt"].append("ImageGPTFeatureExtractor")
     _import_structure["models.layoutlmv2"].append("LayoutLMv2FeatureExtractor")
@@ -947,6 +949,15 @@ if is_torch_available():
             "DPRReader",
         ]
     )
+    _import_structure["models.dpt"].extend(
+        [
+            "DPT_PRETRAINED_MODEL_ARCHIVE_LIST",
+            "DPTForDepthEstimation",
+            "DPTForSemanticSegmentation",
+            "DPTModel",
+            "DPTPreTrainedModel",
+        ]
+    )
     _import_structure["models.electra"].extend(
         [
             "ELECTRA_PRETRAINED_MODEL_ARCHIVE_LIST",
@@ -2543,6 +2554,7 @@ if TYPE_CHECKING:
         DPRReaderOutput,
         DPRReaderTokenizer,
     )
+    from .models.dpt import DPT_PRETRAINED_CONFIG_ARCHIVE_MAP, DPTConfig
     from .models.electra import ELECTRA_PRETRAINED_CONFIG_ARCHIVE_MAP, ElectraConfig, ElectraTokenizer
     from .models.encoder_decoder import EncoderDecoderConfig
     from .models.flaubert import FLAUBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, FlaubertConfig, FlaubertTokenizer
@@ -2838,6 +2850,7 @@ if TYPE_CHECKING:
         from .models.convnext import ConvNextFeatureExtractor
         from .models.deit import DeiTFeatureExtractor
         from .models.detr import DetrFeatureExtractor
+        from .models.dpt import DPTFeatureExtractor
         from .models.glpn import GLPNFeatureExtractor
         from .models.imagegpt import ImageGPTFeatureExtractor
         from .models.layoutlmv2 import LayoutLMv2FeatureExtractor, LayoutLMv2Processor
@@ -2877,7 +2890,6 @@ if TYPE_CHECKING:
         from .utils.dummy_scatter_objects import *
 
     if is_torch_available():
-
         # Benchmarks
         from .benchmark.benchmark import PyTorchBenchmark
         from .benchmark.benchmark_args import PyTorchBenchmarkArguments
@@ -3188,6 +3200,13 @@ if TYPE_CHECKING:
             DPRQuestionEncoder,
             DPRReader,
         )
+        from .models.dpt import (
+            DPT_PRETRAINED_MODEL_ARCHIVE_LIST,
+            DPTForDepthEstimation,
+            DPTForSemanticSegmentation,
+            DPTModel,
+            DPTPreTrainedModel,
+        )
         from .models.electra import (
             ELECTRA_PRETRAINED_MODEL_ARCHIVE_LIST,
             ElectraForCausalLM,

src/transformers/models/__init__.py

@@ -50,6 +50,7 @@ from . import (
     distilbert,
     dit,
     dpr,
+    dpt,
     electra,
     encoder_decoder,
     flaubert,

src/transformers/models/auto/configuration_auto.py

@@ -29,6 +29,7 @@ logger = logging.get_logger(__name__)
 CONFIG_MAPPING_NAMES = OrderedDict(
     [
         # Add configs here
+        ("dpt", "DPTConfig"),
         ("decision_transformer", "DecisionTransformerConfig"),
         ("glpn", "GLPNConfig"),
         ("maskformer", "MaskFormerConfig"),
@@ -134,6 +135,7 @@ CONFIG_MAPPING_NAMES = OrderedDict(
 CONFIG_ARCHIVE_MAP_MAPPING_NAMES = OrderedDict(
     [
         # Add archive maps here
+        ("dpt", "DPT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("glpn", "GLPN_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("maskformer", "MASKFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("poolformer", "POOLFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP"),
@@ -224,6 +226,7 @@ CONFIG_ARCHIVE_MAP_MAPPING_NAMES = OrderedDict(
 MODEL_NAMES_MAPPING = OrderedDict(
     [
         # Add full (and cased) model names here
+        ("dpt", "DPT"),
         ("decision_transformer", "Decision Transformer"),
         ("glpn", "GLPN"),
         ("maskformer", "MaskFormer"),

src/transformers/models/auto/modeling_auto.py

@@ -28,6 +28,7 @@ logger = logging.get_logger(__name__)
 MODEL_MAPPING_NAMES = OrderedDict(
     [
         # Base model mapping
+        ("dpt", "DPTModel"),
         ("decision_transformer", "DecisionTransformerModel"),
         ("glpn", "GLPNModel"),
         ("maskformer", "MaskFormerModel"),
@@ -319,6 +320,7 @@ MODEL_FOR_SEMANTIC_SEGMENTATION_MAPPING_NAMES = OrderedDict(
         # Model for Semantic Segmentation mapping
         ("beit", "BeitForSemanticSegmentation"),
         ("segformer", "SegformerForSemanticSegmentation"),
+        ("dpt", "DPTForSemanticSegmentation"),
     ]
 )
 

src/transformers/models/dpt/__init__.py

+# flake8: noqa
+# There's no way to ignore "F401 '...' imported but unused" warnings in this
+# module, but to preserve other warnings. So, don't check this module at all.
+
+# Copyright 2022 The HuggingFace 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.
+from typing import TYPE_CHECKING
+
+from ...file_utils import _LazyModule, is_tokenizers_available, is_torch_available, is_vision_available
+
+
+_import_structure = {
+    "configuration_dpt": ["DPT_PRETRAINED_CONFIG_ARCHIVE_MAP", "DPTConfig"],
+}
+
+if is_vision_available():
+    _import_structure["feature_extraction_dpt"] = ["DPTFeatureExtractor"]
+
+if is_torch_available():
+    _import_structure["modeling_dpt"] = [
+        "DPT_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "DPTForDepthEstimation",
+        "DPTForSemanticSegmentation",
+        "DPTModel",
+        "DPTPreTrainedModel",
+    ]
+
+
+if TYPE_CHECKING:
+    from .configuration_dpt import DPT_PRETRAINED_CONFIG_ARCHIVE_MAP, DPTConfig
+
+    if is_vision_available():
+        from .feature_extraction_dpt import DPTFeatureExtractor
+
+    if is_torch_available():
+        from .modeling_dpt import (
+            DPT_PRETRAINED_MODEL_ARCHIVE_LIST,
+            DPTForDepthEstimation,
+            DPTForSemanticSegmentation,
+            DPTModel,
+            DPTPreTrainedModel,
+        )
+
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)

src/transformers/models/dpt/configuration_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.
+""" DPT model configuration"""
+
+from ...configuration_utils import PretrainedConfig
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+DPT_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "Intel/dpt-large": "https://huggingface.co/Intel/dpt-large/resolve/main/config.json",
+    # See all DPT models at https://huggingface.co/models?filter=dpt
+}
+
+
+class DPTConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`DPTModel`]. It is used to instantiate an DPT
+    model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
+    defaults will yield a similar configuration to that of the DPT
+    [Intel/dpt-large](https://huggingface.co/Intel/dpt-large) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+
+    Args:
+        hidden_size (`int`, *optional*, defaults to 768):
+            Dimensionality of the encoder layers and the pooler layer.
+        num_hidden_layers (`int`, *optional*, defaults to 12):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 12):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        intermediate_size (`int`, *optional*, defaults to 3072):
+            Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
+        hidden_act (`str` or `function`, *optional*, defaults to `"gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"selu"` and `"gelu_new"` are supported.
+        hidden_dropout_prob (`float`, *optional*, defaults to 0.1):
+            The dropout probabilitiy for all fully connected layers in the embeddings, encoder, and pooler.
+        attention_probs_dropout_prob (`float`, *optional*, defaults to 0.1):
+            The dropout ratio for the attention probabilities.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        layer_norm_eps (`float`, *optional*, defaults to 1e-12):
+            The epsilon used by the layer normalization layers.
+        image_size (`int`, *optional*, defaults to 384):
+            The size (resolution) of each image.
+        patch_size (`int`, *optional*, defaults to 16):
+            The size (resolution) of each patch.
+        num_channels (`int`, *optional*, defaults to 3):
+            The number of input channels.
+        qkv_bias (`bool`, *optional*, defaults to `True`):
+            Whether to add a bias to the queries, keys and values.
+        backbone_out_indices (`List[int]`, *optional*, defaults to `[2, 5, 8, 11]`):
+            Indices of the intermediate hidden states to use from backbone.
+        readout_type (`str`, *optional*, defaults to `"project"`):
+            The readout type to use when processing the readout token (CLS token) of the intermediate hidden states of
+            the ViT backbone. Can be one of [`"ignore"`, `"add"`, `"project"`].
+
+            - "ignore" simply ignores the CLS token.
+            - "add" passes the information from the CLS token to all other tokens by adding the representations.
+            - "project" passes information to the other tokens by concatenating the readout to all other tokens before
+              projecting the
+            representation to the original feature dimension D using a linear layer followed by a GELU non-linearity.
+        reassemble_factors (`List[int]`, *optional*, defaults to `[4, 2, 1, 0.5]`):
+            The up/downsampling factors of the reassemble layers.
+        neck_hidden_sizes (`List[str]`, *optional*, defaults to [96, 192, 384, 768]):
+            The hidden sizes to project to for the feature maps of the backbone.
+        fusion_hidden_size (`int`, *optional*, defaults to 256):
+            The number of channels before fusion.
+        head_in_index (`int`, *optional*, defaults to -1):
+            The index of the features to use in the heads.
+        use_batch_norm_in_fusion_residual (`bool`, *optional*, defaults to `False`):
+            Whether to use batch normalization in the pre-activate residual units of the fusion blocks.
+        use_auxiliary_head (`bool`, *optional*, defaults to `True`):
+            Whether to use an auxiliary head during training.
+        auxiliary_loss_weight (`float`, *optional*, defaults to 0.4):
+            Weight of the cross-entropy loss of the auxiliary head.
+        semantic_loss_ignore_index (`int`, *optional*, defaults to 255):
+            The index that is ignored by the loss function of the semantic segmentation model.
+        semantic_classifier_dropout (`float`, *optional*, defaults to 0.1):
+            The dropout ratio for the semantic classification head.
+
+    Example:
+
+    ```python
+    >>> from transformers import DPTModel, DPTConfig
+
+    >>> # Initializing a DPT dpt-large style configuration
+    >>> configuration = DPTConfig()
+
+    >>> # Initializing a model from the dpt-large style configuration
+    >>> model = DPTModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+    model_type = "dpt"
+
+    def __init__(
+        self,
+        hidden_size=768,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        intermediate_size=3072,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.0,
+        attention_probs_dropout_prob=0.0,
+        initializer_range=0.02,
+        layer_norm_eps=1e-12,
+        image_size=384,
+        patch_size=16,
+        num_channels=3,
+        qkv_bias=True,
+        backbone_out_indices=[2, 5, 8, 11],
+        readout_type="project",
+        reassemble_factors=[4, 2, 1, 0.5],
+        neck_hidden_sizes=[96, 192, 384, 768],
+        fusion_hidden_size=256,
+        head_in_index=-1,
+        use_batch_norm_in_fusion_residual=False,
+        use_auxiliary_head=True,
+        auxiliary_loss_weight=0.4,
+        semantic_loss_ignore_index=255,
+        semantic_classifier_dropout=0.1,
+        **kwargs
+    ):
+        super().__init__(**kwargs)
+
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        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.layer_norm_eps = layer_norm_eps
+        self.image_size = image_size
+        self.patch_size = patch_size
+        self.num_channels = num_channels
+        self.qkv_bias = qkv_bias
+        self.backbone_out_indices = backbone_out_indices
+        if readout_type not in ["ignore", "add", "project"]:
+            raise ValueError("Readout_type must be one of ['ignore', 'add', 'project']")
+        self.readout_type = readout_type
+        self.reassemble_factors = reassemble_factors
+        self.neck_hidden_sizes = neck_hidden_sizes
+        self.fusion_hidden_size = fusion_hidden_size
+        self.head_in_index = head_in_index
+        self.use_batch_norm_in_fusion_residual = use_batch_norm_in_fusion_residual
+        # auxiliary head attributes (semantic segmentation)
+        self.use_auxiliary_head = use_auxiliary_head
+        self.auxiliary_loss_weight = auxiliary_loss_weight
+        self.semantic_loss_ignore_index = semantic_loss_ignore_index
+        self.semantic_classifier_dropout = semantic_classifier_dropout

src/transformers/models/dpt/convert_dpt_to_pytorch.py

+# coding=utf-8
+# Copyright 2022 The HuggingFace Inc. team.
+#
+# 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.
+"""Convert DPT checkpoints from the original repository. URL: https://github.com/isl-org/DPT"""
+
+
+import argparse
+import json
+from pathlib import Path
+
+import torch
+from PIL import Image
+
+import requests
+from huggingface_hub import cached_download, hf_hub_url
+from transformers import DPTConfig, DPTFeatureExtractor, DPTForDepthEstimation, DPTForSemanticSegmentation
+from transformers.utils import logging
+
+
+logging.set_verbosity_info()
+logger = logging.get_logger(__name__)
+
+
+def get_dpt_config(checkpoint_url):
+    config = DPTConfig()
+
+    if "large" in checkpoint_url:
+        config.hidden_size = 1024
+        config.intermediate_size = 4096
+        config.num_hidden_layers = 24
+        config.num_attention_heads = 16
+        config.backbone_out_indices = [5, 11, 17, 23]
+        config.neck_hidden_sizes = [256, 512, 1024, 1024]
+        expected_shape = (1, 384, 384)
+
+    if "ade" in checkpoint_url:
+        config.use_batch_norm_in_fusion_residual = True
+
+        config.num_labels = 150
+        repo_id = "datasets/huggingface/label-files"
+        filename = "ade20k-id2label.json"
+        id2label = json.load(open(cached_download(hf_hub_url(repo_id, filename)), "r"))
+        id2label = {int(k): v for k, v in id2label.items()}
+        config.id2label = id2label
+        config.label2id = {v: k for k, v in id2label.items()}
+        expected_shape = [1, 150, 480, 480]
+
+    return config, expected_shape
+
+
+def remove_ignore_keys_(state_dict):
+    ignore_keys = ["pretrained.model.head.weight", "pretrained.model.head.bias"]
+    for k in ignore_keys:
+        state_dict.pop(k, None)
+
+
+def rename_key(name):
+    if (
+        "pretrained.model" in name
+        and "cls_token" not in name
+        and "pos_embed" not in name
+        and "patch_embed" not in name
+    ):
+        name = name.replace("pretrained.model", "dpt.encoder")
+    if "pretrained.model" in name:
+        name = name.replace("pretrained.model", "dpt.embeddings")
+    if "patch_embed" in name:
+        name = name.replace("patch_embed", "patch_embeddings")
+    if "pos_embed" in name:
+        name = name.replace("pos_embed", "position_embeddings")
+    if "attn.proj" in name:
+        name = name.replace("attn.proj", "attention.output.dense")
+    if "proj" in name and "project" not in name:
+        name = name.replace("proj", "projection")
+    if "blocks" in name:
+        name = name.replace("blocks", "layer")
+    if "mlp.fc1" in name:
+        name = name.replace("mlp.fc1", "intermediate.dense")
+    if "mlp.fc2" in name:
+        name = name.replace("mlp.fc2", "output.dense")
+    if "norm1" in name:
+        name = name.replace("norm1", "layernorm_before")
+    if "norm2" in name:
+        name = name.replace("norm2", "layernorm_after")
+    if "scratch.output_conv" in name:
+        name = name.replace("scratch.output_conv", "head")
+    if "scratch" in name:
+        name = name.replace("scratch", "neck")
+    if "layer1_rn" in name:
+        name = name.replace("layer1_rn", "convs.0")
+    if "layer2_rn" in name:
+        name = name.replace("layer2_rn", "convs.1")
+    if "layer3_rn" in name:
+        name = name.replace("layer3_rn", "convs.2")
+    if "layer4_rn" in name:
+        name = name.replace("layer4_rn", "convs.3")
+    if "refinenet" in name:
+        layer_idx = int(name[len("neck.refinenet") : len("neck.refinenet") + 1])
+        # tricky here: we need to map 4 to 0, 3 to 1, 2 to 2 and 1 to 3
+        name = name.replace(f"refinenet{layer_idx}", f"fusion_stage.layers.{abs(layer_idx-4)}")
+    if "out_conv" in name:
+        name = name.replace("out_conv", "projection")
+    if "resConfUnit1" in name:
+        name = name.replace("resConfUnit1", "residual_layer1")
+    if "resConfUnit2" in name:
+        name = name.replace("resConfUnit2", "residual_layer2")
+    if "conv1" in name:
+        name = name.replace("conv1", "convolution1")
+    if "conv2" in name:
+        name = name.replace("conv2", "convolution2")
+    # readout blocks
+    if "pretrained.act_postprocess1.0.project.0" in name:
+        name = name.replace("pretrained.act_postprocess1.0.project.0", "neck.reassemble_stage.readout_projects.0.0")
+    if "pretrained.act_postprocess2.0.project.0" in name:
+        name = name.replace("pretrained.act_postprocess2.0.project.0", "neck.reassemble_stage.readout_projects.1.0")
+    if "pretrained.act_postprocess3.0.project.0" in name:
+        name = name.replace("pretrained.act_postprocess3.0.project.0", "neck.reassemble_stage.readout_projects.2.0")
+    if "pretrained.act_postprocess4.0.project.0" in name:
+        name = name.replace("pretrained.act_postprocess4.0.project.0", "neck.reassemble_stage.readout_projects.3.0")
+    # resize blocks
+    if "pretrained.act_postprocess1.3" in name:
+        name = name.replace("pretrained.act_postprocess1.3", "neck.reassemble_stage.layers.0.projection")
+    if "pretrained.act_postprocess1.4" in name:
+        name = name.replace("pretrained.act_postprocess1.4", "neck.reassemble_stage.layers.0.resize")
+    if "pretrained.act_postprocess2.3" in name:
+        name = name.replace("pretrained.act_postprocess2.3", "neck.reassemble_stage.layers.1.projection")
+    if "pretrained.act_postprocess2.4" in name:
+        name = name.replace("pretrained.act_postprocess2.4", "neck.reassemble_stage.layers.1.resize")
+    if "pretrained.act_postprocess3.3" in name:
+        name = name.replace("pretrained.act_postprocess3.3", "neck.reassemble_stage.layers.2.projection")
+    if "pretrained.act_postprocess4.3" in name:
+        name = name.replace("pretrained.act_postprocess4.3", "neck.reassemble_stage.layers.3.projection")
+    if "pretrained.act_postprocess4.4" in name:
+        name = name.replace("pretrained.act_postprocess4.4", "neck.reassemble_stage.layers.3.resize")
+    if "pretrained" in name:
+        name = name.replace("pretrained", "dpt")
+    if "bn" in name:
+        name = name.replace("bn", "batch_norm")
+    if "head" in name:
+        name = name.replace("head", "head.head")
+    if "encoder.norm" in name:
+        name = name.replace("encoder.norm", "layernorm")
+    if "auxlayer" in name:
+        name = name.replace("auxlayer", "auxiliary_head.head")
+
+    return name
+
+
+# we split up the matrix of each encoder layer into queries, keys and values
+def read_in_q_k_v(state_dict, config):
+    for i in range(config.num_hidden_layers):
+        # read in weights + bias of input projection layer (in timm, this is a single matrix + bias)
+        in_proj_weight = state_dict.pop(f"dpt.encoder.layer.{i}.attn.qkv.weight")
+        in_proj_bias = state_dict.pop(f"dpt.encoder.layer.{i}.attn.qkv.bias")
+        # next, add query, keys and values (in that order) to the state dict
+        state_dict[f"dpt.encoder.layer.{i}.attention.attention.query.weight"] = in_proj_weight[: config.hidden_size, :]
+        state_dict[f"dpt.encoder.layer.{i}.attention.attention.query.bias"] = in_proj_bias[: config.hidden_size]
+        state_dict[f"dpt.encoder.layer.{i}.attention.attention.key.weight"] = in_proj_weight[
+            config.hidden_size : config.hidden_size * 2, :
+        ]
+        state_dict[f"dpt.encoder.layer.{i}.attention.attention.key.bias"] = in_proj_bias[
+            config.hidden_size : config.hidden_size * 2
+        ]
+        state_dict[f"dpt.encoder.layer.{i}.attention.attention.value.weight"] = in_proj_weight[
+            -config.hidden_size :, :
+        ]
+        state_dict[f"dpt.encoder.layer.{i}.attention.attention.value.bias"] = in_proj_bias[-config.hidden_size :]
+
+
+# We will verify our results on an image of cute cats
+def prepare_img():
+    url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+    im = Image.open(requests.get(url, stream=True).raw)
+    return im
+
+
+@torch.no_grad()
+def convert_dpt_checkpoint(checkpoint_url, pytorch_dump_folder_path, push_to_hub, model_name):
+    """
+    Copy/paste/tweak model's weights to our DPT structure.
+    """
+
+    # define DPT configuration based on URL
+    config, expected_shape = get_dpt_config(checkpoint_url)
+    # load original state_dict from URL
+    state_dict = torch.hub.load_state_dict_from_url(checkpoint_url, map_location="cpu")
+    # remove certain keys
+    remove_ignore_keys_(state_dict)
+    # rename keys
+    for key in state_dict.copy().keys():
+        val = state_dict.pop(key)
+        state_dict[rename_key(key)] = val
+    # read in qkv matrices
+    read_in_q_k_v(state_dict, config)
+
+    # load HuggingFace model
+    model = DPTForSemanticSegmentation(config) if "ade" in checkpoint_url else DPTForDepthEstimation(config)
+    model.load_state_dict(state_dict)
+    model.eval()
+
+    # Check outputs on an image
+    size = 480 if "ade" in checkpoint_url else 384
+    feature_extractor = DPTFeatureExtractor(size=size)
+
+    image = prepare_img()
+    encoding = feature_extractor(image, return_tensors="pt")
+
+    # forward pass
+    outputs = model(**encoding).logits if "ade" in checkpoint_url else model(**encoding).predicted_depth
+
+    # Assert logits
+    expected_slice = torch.tensor([[6.3199, 6.3629, 6.4148], [6.3850, 6.3615, 6.4166], [6.3519, 6.3176, 6.3575]])
+    if "ade" in checkpoint_url:
+        expected_slice = torch.tensor([[4.0480, 4.2420, 4.4360], [4.3124, 4.5693, 4.8261], [4.5768, 4.8965, 5.2163]])
+    assert outputs.shape == torch.Size(expected_shape)
+    assert (
+        torch.allclose(outputs[0, 0, :3, :3], expected_slice, atol=1e-4)
+        if "ade" in checkpoint_url
+        else torch.allclose(outputs[0, :3, :3], expected_slice)
+    )
+
+    Path(pytorch_dump_folder_path).mkdir(exist_ok=True)
+    print(f"Saving model to {pytorch_dump_folder_path}")
+    model.save_pretrained(pytorch_dump_folder_path)
+    print(f"Saving feature extractor to {pytorch_dump_folder_path}")
+    feature_extractor.save_pretrained(pytorch_dump_folder_path)
+
+    if push_to_hub:
+        print("Pushing model to hub...")
+        model.push_to_hub(
+            repo_path_or_name=Path(pytorch_dump_folder_path, model_name),
+            organization="nielsr",
+            commit_message="Add model",
+            use_temp_dir=True,
+        )
+        feature_extractor.push_to_hub(
+            repo_path_or_name=Path(pytorch_dump_folder_path, model_name),
+            organization="nielsr",
+            commit_message="Add feature extractor",
+            use_temp_dir=True,
+        )
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    # Required parameters
+    parser.add_argument(
+        "--checkpoint_url",
+        default="https://github.com/intel-isl/DPT/releases/download/1_0/dpt_large-midas-2f21e586.pt",
+        type=str,
+        help="URL of the original DPT checkpoint you'd like to convert.",
+    )
+    parser.add_argument(
+        "--pytorch_dump_folder_path",
+        default=None,
+        type=str,
+        required=True,
+        help="Path to the output PyTorch model directory.",
+    )
+    parser.add_argument(
+        "--push_to_hub",
+        action="store_true",
+    )
+    parser.add_argument(
+        "--model_name",
+        default="dpt-large",
+        type=str,
+        help="Name of the model, in case you're pushing to the hub.",
+    )
+
+    args = parser.parse_args()
+    convert_dpt_checkpoint(args.checkpoint_url, args.pytorch_dump_folder_path, args.push_to_hub, args.model_name)

src/transformers/models/dpt/feature_extraction_dpt.py

+# coding=utf-8
+# Copyright 2021 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.
+"""Feature extractor class for DPT."""
+
+from typing import Optional, Union
+
+import numpy as np
+from PIL import Image
+
+from ...feature_extraction_utils import BatchFeature, FeatureExtractionMixin
+from ...file_utils import TensorType
+from ...image_utils import (
+    IMAGENET_STANDARD_MEAN,
+    IMAGENET_STANDARD_STD,
+    ImageFeatureExtractionMixin,
+    ImageInput,
+    is_torch_tensor,
+)
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class DPTFeatureExtractor(FeatureExtractionMixin, ImageFeatureExtractionMixin):
+    r"""
+    Constructs a DPT feature extractor.
+
+    This feature extractor inherits from [`FeatureExtractionMixin`] which contains most of the main methods. Users
+    should refer to this superclass for more information regarding those methods.
+
+    Args:
+        do_resize (`bool`, *optional*, defaults to `True`):
+            Whether to resize the input to a certain `size`.
+        size ('int' or `Tuple(int)`, *optional*, defaults to 384):
+            Resize the input to the given size. If a tuple is provided, it should be (width, height). If only an
+            integer is provided, then the input will be resized to (size, size). Only has an effect if `do_resize` is
+            set to `True`.
+        ensure_multiple_of (`int`, *optional*, defaults to 1):
+            Ensure that the input is resized to a multiple of this value. Only has an effect if `do_resize` is set to
+            `True`.
+        keep_aspect_ratio (`bool`, *optional*, defaults to `False`):
+            Whether to keep the aspect ratio of the input. Only has an effect if `do_resize` is set to `True`.
+        resample (`int`, *optional*, defaults to `PIL.Image.BILINEAR`):
+            An optional resampling filter. This can be one of `PIL.Image.NEAREST`, `PIL.Image.BOX`,
+            `PIL.Image.BILINEAR`, `PIL.Image.HAMMING`, `PIL.Image.BICUBIC` or `PIL.Image.LANCZOS`. Only has an effect
+            if `do_resize` is set to `True`.
+        do_normalize (`bool`, *optional*, defaults to `True`):
+            Whether or not to normalize the input with mean and standard deviation.
+        image_mean (`List[int]`, defaults to `[0.5, 0.5, 0.5]`):
+            The sequence of means for each channel, to be used when normalizing images.
+        image_std (`List[int]`, defaults to `[0.5, 0.5, 0.5]`):
+            The sequence of standard deviations for each channel, to be used when normalizing images.
+    """
+
+    model_input_names = ["pixel_values"]
+
+    def __init__(
+        self,
+        do_resize=True,
+        size=384,
+        keep_aspect_ratio=False,
+        ensure_multiple_of=1,
+        resample=Image.BILINEAR,
+        do_normalize=True,
+        image_mean=None,
+        image_std=None,
+        **kwargs
+    ):
+        super().__init__(**kwargs)
+        self.do_resize = do_resize
+        self.size = size
+        self.keep_aspect_ratio = keep_aspect_ratio
+        self.ensure_multiple_of = ensure_multiple_of
+        self.resample = resample
+        self.do_normalize = do_normalize
+        self.image_mean = image_mean if image_mean is not None else IMAGENET_STANDARD_MEAN
+        self.image_std = image_std if image_std is not None else IMAGENET_STANDARD_STD
+
+    def constrain_to_multiple_of(self, size, min_val=0, max_val=None):
+        y = (np.round(size / self.ensure_multiple_of) * self.ensure_multiple_of).astype(int)
+
+        if max_val is not None and y > max_val:
+            y = (np.floor(size / self.ensure_multiple_of) * self.ensure_multiple_of).astype(int)
+
+        if y < min_val:
+            y = (np.ceil(size / self.ensure_multiple_of) * self.ensure_multiple_of).astype(int)
+
+        return y
+
+    def update_size(self, image):
+        image = self.to_pil_image(image)
+        width, height = image.size
+
+        size = self.size
+
+        if isinstance(size, list):
+            size = tuple(size)
+
+        if isinstance(size, int) or len(size) == 1:
+            size = (size, size)
+
+        # determine new width and height
+        scale_width = size[0] / width
+        scale_height = size[1] / height
+
+        if self.keep_aspect_ratio:
+            # scale as least as possbile
+            if abs(1 - scale_width) < abs(1 - scale_height):
+                # fit width
+                scale_height = scale_width
+            else:
+                # fit height
+                scale_width = scale_height
+        else:
+            new_width = self.constrain_to_multiple_of(scale_width * width)
+            new_height = self.constrain_to_multiple_of(scale_height * height)
+
+        return (new_width, new_height)
+
+    def __call__(
+        self, images: ImageInput, return_tensors: Optional[Union[str, TensorType]] = None, **kwargs
+    ) -> BatchFeature:
+        """
+        Main method to prepare for the model one or several image(s).
+
+        <Tip warning={true}>
+
+        NumPy arrays and PyTorch tensors are converted to PIL images when resizing, so the most efficient is to pass
+        PIL images.
+
+        </Tip>
+
+        Args:
+            images (`PIL.Image.Image`, `np.ndarray`, `torch.Tensor`, `List[PIL.Image.Image]`, `List[np.ndarray]`, `List[torch.Tensor]`):
+                The image or batch of images to be prepared. Each image can be a PIL image, NumPy array or PyTorch
+                tensor. In case of a NumPy array/PyTorch tensor, each image should be of shape (C, H, W), where C is a
+                number of channels, H and W are image height and width.
+
+            return_tensors (`str` or [`~file_utils.TensorType`], *optional*, defaults to `'np'`):
+                If set, will return tensors of a particular framework. Acceptable values are:
+
+                - `'tf'`: Return TensorFlow `tf.constant` objects.
+                - `'pt'`: Return PyTorch `torch.Tensor` objects.
+                - `'np'`: Return NumPy `np.ndarray` objects.
+                - `'jax'`: Return JAX `jnp.ndarray` objects.
+
+        Returns:
+            [`BatchFeature`]: A [`BatchFeature`] with the following fields:
+
+            - **pixel_values** -- Pixel values to be fed to a model, of shape (batch_size, num_channels, height,
+              width).
+        """
+        # Input type checking for clearer error
+        valid_images = False
+
+        # Check that images has a valid type
+        if isinstance(images, (Image.Image, np.ndarray)) or is_torch_tensor(images):
+            valid_images = True
+        elif isinstance(images, (list, tuple)):
+            if len(images) == 0 or isinstance(images[0], (Image.Image, np.ndarray)) or is_torch_tensor(images[0]):
+                valid_images = True
+
+        if not valid_images:
+            raise ValueError(
+                "Images must of type `PIL.Image.Image`, `np.ndarray` or `torch.Tensor` (single example), "
+                "`List[PIL.Image.Image]`, `List[np.ndarray]` or `List[torch.Tensor]` (batch of examples)."
+            )
+
+        is_batched = bool(
+            isinstance(images, (list, tuple))
+            and (isinstance(images[0], (Image.Image, np.ndarray)) or is_torch_tensor(images[0]))
+        )
+
+        if not is_batched:
+            images = [images]
+
+        # transformations (resizing + normalization)
+        if self.do_resize and self.size is not None:
+            for idx, image in enumerate(images):
+                size = self.update_size(image)
+                images[idx] = self.resize(image, size=size, resample=self.resample)
+        if self.do_normalize:
+            images = [self.normalize(image=image, mean=self.image_mean, std=self.image_std) for image in images]
+
+        # return as BatchFeature
+        data = {"pixel_values": images}
+        encoded_inputs = BatchFeature(data=data, tensor_type=return_tensors)
+
+        return encoded_inputs

src/transformers/utils/dummy_pt_objects.py

@@ -1604,6 +1604,37 @@ class DPRReader(metaclass=DummyObject):
         requires_backends(self, ["torch"])
 
 
+DPT_PRETRAINED_MODEL_ARCHIVE_LIST = None
+
+
+class DPTForDepthEstimation(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class DPTForSemanticSegmentation(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class DPTModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class DPTPreTrainedModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
 ELECTRA_PRETRAINED_MODEL_ARCHIVE_LIST = None
 
 

src/transformers/utils/dummy_vision_objects.py

@@ -52,6 +52,13 @@ class DetrFeatureExtractor(metaclass=DummyObject):
         requires_backends(self, ["vision"])
 
 
+class DPTFeatureExtractor(metaclass=DummyObject):
+    _backends = ["vision"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["vision"])
+
+
 class GLPNFeatureExtractor(metaclass=DummyObject):
     _backends = ["vision"]
 

tests/dpt/test_feature_extraction_dpt.py

+# coding=utf-8
+# 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.file_utils import is_torch_available, is_vision_available
+from transformers.testing_utils import require_torch, require_vision
+
+from ..test_feature_extraction_common import FeatureExtractionSavingTestMixin, prepare_image_inputs
+
+
+if is_torch_available():
+    import torch
+
+if is_vision_available():
+    from PIL import Image
+
+    from transformers import DPTFeatureExtractor
+
+
+class DPTFeatureExtractionTester(unittest.TestCase):
+    def __init__(
+        self,
+        parent,
+        batch_size=7,
+        num_channels=3,
+        image_size=18,
+        min_resolution=30,
+        max_resolution=400,
+        do_resize=True,
+        size=18,
+        do_normalize=True,
+        image_mean=[0.5, 0.5, 0.5],
+        image_std=[0.5, 0.5, 0.5],
+    ):
+        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
+        self.image_mean = image_mean
+        self.image_std = image_std
+
+    def prepare_feat_extract_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,
+        }
+
+
+@require_torch
+@require_vision
+class DPTFeatureExtractionTest(FeatureExtractionSavingTestMixin, unittest.TestCase):
+
+    feature_extraction_class = DPTFeatureExtractor if is_vision_available() else None
+
+    def setUp(self):
+        self.feature_extract_tester = DPTFeatureExtractionTester(self)
+
+    @property
+    def feat_extract_dict(self):
+        return self.feature_extract_tester.prepare_feat_extract_dict()
+
+    def test_feat_extract_properties(self):
+        feature_extractor = self.feature_extraction_class(**self.feat_extract_dict)
+        self.assertTrue(hasattr(feature_extractor, "image_mean"))
+        self.assertTrue(hasattr(feature_extractor, "image_std"))
+        self.assertTrue(hasattr(feature_extractor, "do_normalize"))
+        self.assertTrue(hasattr(feature_extractor, "do_resize"))
+        self.assertTrue(hasattr(feature_extractor, "size"))
+
+    def test_call_pil(self):
+        # Initialize feature_extractor
+        feature_extractor = self.feature_extraction_class(**self.feat_extract_dict)
+        # create random PIL images
+        image_inputs = prepare_image_inputs(self.feature_extract_tester, equal_resolution=False)
+        for image in image_inputs:
+            self.assertIsInstance(image, Image.Image)
+
+        # Test not batched input
+        encoded_images = feature_extractor(image_inputs[0], return_tensors="pt").pixel_values
+        self.assertEqual(
+            encoded_images.shape,
+            (
+                1,
+                self.feature_extract_tester.num_channels,
+                self.feature_extract_tester.size,
+                self.feature_extract_tester.size,
+            ),
+        )
+
+        # Test batched
+        encoded_images = feature_extractor(image_inputs, return_tensors="pt").pixel_values
+        self.assertEqual(
+            encoded_images.shape,
+            (
+                self.feature_extract_tester.batch_size,
+                self.feature_extract_tester.num_channels,
+                self.feature_extract_tester.size,
+                self.feature_extract_tester.size,
+            ),
+        )
+
+    def test_call_numpy(self):
+        # Initialize feature_extractor
+        feature_extractor = self.feature_extraction_class(**self.feat_extract_dict)
+        # create random numpy tensors
+        image_inputs = prepare_image_inputs(self.feature_extract_tester, equal_resolution=False, numpify=True)
+        for image in image_inputs:
+            self.assertIsInstance(image, np.ndarray)
+
+        # Test not batched input
+        encoded_images = feature_extractor(image_inputs[0], return_tensors="pt").pixel_values
+        self.assertEqual(
+            encoded_images.shape,
+            (
+                1,
+                self.feature_extract_tester.num_channels,
+                self.feature_extract_tester.size,
+                self.feature_extract_tester.size,
+            ),
+        )
+
+        # Test batched
+        encoded_images = feature_extractor(image_inputs, return_tensors="pt").pixel_values
+        self.assertEqual(
+            encoded_images.shape,
+            (
+                self.feature_extract_tester.batch_size,
+                self.feature_extract_tester.num_channels,
+                self.feature_extract_tester.size,
+                self.feature_extract_tester.size,
+            ),
+        )
+
+    def test_call_pytorch(self):
+        # Initialize feature_extractor
+        feature_extractor = self.feature_extraction_class(**self.feat_extract_dict)
+        # create random PyTorch tensors
+        image_inputs = prepare_image_inputs(self.feature_extract_tester, equal_resolution=False, torchify=True)
+        for image in image_inputs:
+            self.assertIsInstance(image, torch.Tensor)
+
+        # Test not batched input
+        encoded_images = feature_extractor(image_inputs[0], return_tensors="pt").pixel_values
+        self.assertEqual(
+            encoded_images.shape,
+            (
+                1,
+                self.feature_extract_tester.num_channels,
+                self.feature_extract_tester.size,
+                self.feature_extract_tester.size,
+            ),
+        )
+
+        # Test batched
+        encoded_images = feature_extractor(image_inputs, return_tensors="pt").pixel_values
+        self.assertEqual(
+            encoded_images.shape,
+            (
+                self.feature_extract_tester.batch_size,
+                self.feature_extract_tester.num_channels,
+                self.feature_extract_tester.size,
+                self.feature_extract_tester.size,
+            ),
+        )