Add Vision Transformer and ViTFeatureExtractor

.circleci/config.yml

@@ -79,8 +79,8 @@ jobs:
                       - v0.4-{{ checksum "setup.py" }}
             - run: sudo apt-get -y update && sudo apt-get install -y libsndfile1-dev
             - run: pip install --upgrade pip
-            - run: pip install .[sklearn,tf-cpu,torch,testing,sentencepiece,speech]
-            - run: pip install tapas torch-scatter -f https://pytorch-geometric.com/whl/torch-1.8.0+cpu.html
+            - run: pip install .[sklearn,tf-cpu,torch,testing,sentencepiece,speech,vision]
+            - run: pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-1.8.0+cpu.html
             - save_cache:
                 key: v0.4-{{ checksum "setup.py" }}
                 paths:
@@ -107,8 +107,8 @@ jobs:
                       - v0.4-{{ checksum "setup.py" }}
             - run: sudo apt-get -y update && sudo apt-get install -y libsndfile1-dev
             - run: pip install --upgrade pip
-            - run: pip install .[sklearn,flax,torch,testing,sentencepiece,speech]
-            - run: pip install tapas torch-scatter -f https://pytorch-geometric.com/whl/torch-1.8.0+cpu.html
+            - run: pip install .[sklearn,flax,torch,testing,sentencepiece,speech,vision]
+            - run: pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-1.8.0+cpu.html
             - save_cache:
                 key: v0.4-{{ checksum "setup.py" }}
                 paths:
@@ -135,8 +135,8 @@ jobs:
                       - v0.4-{{ checksum "setup.py" }}
             - run: sudo apt-get -y update && sudo apt-get install -y libsndfile1-dev
             - run: pip install --upgrade pip
-            - run: pip install .[sklearn,torch,testing,sentencepiece,speech]
-            - run: pip install tapas torch-scatter -f https://pytorch-geometric.com/whl/torch-1.8.0+cpu.html
+            - run: pip install .[sklearn,torch,testing,sentencepiece,speech,vision]
+            - run: pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-1.8.0+cpu.html
             - save_cache:
                   key: v0.4-torch-{{ checksum "setup.py" }}
                   paths:
@@ -215,8 +215,8 @@ jobs:
                       - v0.4-{{ checksum "setup.py" }}
             - run: sudo apt-get -y update && sudo apt-get install -y libsndfile1-dev
             - run: pip install --upgrade pip
-            - run: pip install .[sklearn,torch,testing,sentencepiece,speech]
-            - run: pip install tapas torch-scatter -f https://pytorch-geometric.com/whl/torch-1.8.0+cpu.html
+            - run: pip install .[sklearn,torch,testing,sentencepiece,speech,vision]
+            - run: pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-1.8.0+cpu.html
             - save_cache:
                   key: v0.4-torch-{{ checksum "setup.py" }}
                   paths:

README.md

@@ -232,6 +232,7 @@ Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih.
 1. **[T5](https://huggingface.co/transformers/model_doc/t5.html)** (from Google AI) released with the paper [Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer](https://arxiv.org/abs/1910.10683) by Colin Raffel and Noam Shazeer and Adam Roberts and Katherine Lee and Sharan Narang and Michael Matena and Yanqi Zhou and Wei Li and Peter J. Liu.
 1. **[TAPAS](https://huggingface.co/transformers/model_doc/tapas.html)** (from Google AI) released with the paper [TAPAS: Weakly Supervised Table Parsing via Pre-training](https://arxiv.org/abs/2004.02349) by Jonathan Herzig, Paweł Krzysztof Nowak, Thomas Müller, Francesco Piccinno and Julian Martin Eisenschlos.
 1. **[Transformer-XL](https://huggingface.co/transformers/model_doc/transformerxl.html)** (from Google/CMU) released with the paper [Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context](https://arxiv.org/abs/1901.02860) by Zihang Dai*, Zhilin Yang*, Yiming Yang, Jaime Carbonell, Quoc V. Le, Ruslan Salakhutdinov.
+1. **[Vision Transformer (ViT)](https://huggingface.co/transformers/model_doc/vit.html)** (from Google AI) released with the paper [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) by Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby.
 1. **[Wav2Vec2](https://huggingface.co/transformers/model_doc/wav2vec2.html)** (from Facebook AI) released with the paper [wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations](https://arxiv.org/abs/2006.11477) by Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, Michael Auli.
 1. **[XLM](https://huggingface.co/transformers/model_doc/xlm.html)** (from Facebook) released together with the paper [Cross-lingual Language Model Pretraining](https://arxiv.org/abs/1901.07291) by Guillaume Lample and Alexis Conneau.
 1. **[XLM-ProphetNet](https://huggingface.co/transformers/model_doc/xlmprophetnet.html)** (from Microsoft Research) released with the paper [ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training](https://arxiv.org/abs/2001.04063) by Yu Yan, Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou.

docs/source/index.rst

@@ -205,22 +205,26 @@ and conversion utilities for the following models:
 41. :doc:`Transformer-XL <model_doc/transformerxl>` (from Google/CMU) released with the paper `Transformer-XL:
     Attentive Language Models Beyond a Fixed-Length Context <https://arxiv.org/abs/1901.02860>`__ by Zihang Dai*,
     Zhilin Yang*, Yiming Yang, Jaime Carbonell, Quoc V. Le, Ruslan Salakhutdinov.
-42. :doc:`Wav2Vec2 <model_doc/wav2vec2>` (from Facebook AI) released with the paper `wav2vec 2.0: A Framework for
+42. :doc:`Vision Transformer (ViT) <model_doc/vit>` (from Google AI) released with the paper `An Image is Worth 16x16
+    Words: Transformers for Image Recognition at Scale <https://arxiv.org/abs/2010.11929>`__ by Alexey Dosovitskiy,
+    Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias
+    Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby.
+43. :doc:`Wav2Vec2 <model_doc/wav2vec2>` (from Facebook AI) released with the paper `wav2vec 2.0: A Framework for
     Self-Supervised Learning of Speech Representations <https://arxiv.org/abs/2006.11477>`__ by Alexei Baevski, Henry
     Zhou, Abdelrahman Mohamed, Michael Auli.
-43. :doc:`XLM <model_doc/xlm>` (from Facebook) released together with the paper `Cross-lingual Language Model
+44. :doc:`XLM <model_doc/xlm>` (from Facebook) released together with the paper `Cross-lingual Language Model
     Pretraining <https://arxiv.org/abs/1901.07291>`__ by Guillaume Lample and Alexis Conneau.
-44. :doc:`XLM-ProphetNet <model_doc/xlmprophetnet>` (from Microsoft Research) released with the paper `ProphetNet:
+45. :doc:`XLM-ProphetNet <model_doc/xlmprophetnet>` (from Microsoft Research) released with the paper `ProphetNet:
     Predicting Future N-gram for Sequence-to-Sequence Pre-training <https://arxiv.org/abs/2001.04063>`__ by Yu Yan,
     Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou.
-45. :doc:`XLM-RoBERTa <model_doc/xlmroberta>` (from Facebook AI), released together with the paper `Unsupervised
+46. :doc:`XLM-RoBERTa <model_doc/xlmroberta>` (from Facebook AI), released together with the paper `Unsupervised
     Cross-lingual Representation Learning at Scale <https://arxiv.org/abs/1911.02116>`__ by Alexis Conneau*, Kartikay
     Khandelwal*, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke
     Zettlemoyer and Veselin Stoyanov.
-46. :doc:`XLNet <model_doc/xlnet>` (from Google/CMU) released with the paper `​XLNet: Generalized Autoregressive
+47. :doc:`XLNet <model_doc/xlnet>` (from Google/CMU) released with the paper `​XLNet: Generalized Autoregressive
     Pretraining for Language Understanding <https://arxiv.org/abs/1906.08237>`__ by Zhilin Yang*, Zihang Dai*, Yiming
     Yang, Jaime Carbonell, Ruslan Salakhutdinov, Quoc V. Le.
-47. :doc:`XLSR-Wav2Vec2 <model_doc/xlsr_wav2vec2>` (from Facebook AI) released with the paper `Unsupervised
+48. :doc:`XLSR-Wav2Vec2 <model_doc/xlsr_wav2vec2>` (from Facebook AI) released with the paper `Unsupervised
     Cross-Lingual Representation Learning For Speech Recognition <https://arxiv.org/abs/2006.13979>`__ by Alexis
     Conneau, Alexei Baevski, Ronan Collobert, Abdelrahman Mohamed, Michael Auli.
 
@@ -319,6 +323,8 @@ TensorFlow and/or Flax.
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
 |       Transformer-XL        |       ✅       |       ❌       |       ✅        |         ✅         |      ❌      |
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
+|             ViT             |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
++-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
 |          Wav2Vec2           |       ✅       |       ❌       |       ✅        |         ❌         |      ❌      |
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
 |             XLM             |       ✅       |       ❌       |       ✅        |         ✅         |      ❌      |
@@ -449,6 +455,7 @@ TensorFlow and/or Flax.
     model_doc/t5
     model_doc/tapas
     model_doc/transformerxl
+    model_doc/vit
     model_doc/wav2vec2
     model_doc/xlm
     model_doc/xlmprophetnet

docs/source/model_doc/vit.rst

@@ -0,0 +1,102 @@
+.. 
+    Copyright 2020 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.
+
+Vision Transformer (ViT)
+-----------------------------------------------------------------------------------------------------------------------
+
+.. note::
+
+    This is a recently introduced model so the API hasn't been tested extensively. There may be some bugs or slight
+    breaking changes to fix it in the future. If you see something strange, file a `Github Issue
+    <https://github.com/huggingface/transformers/issues/new?assignees=&labels=&template=bug-report.md&title>`__.
+
+
+Overview
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+The Vision Transformer (ViT) model was proposed in `An Image is Worth 16x16 Words: Transformers for Image Recognition
+at Scale <https://arxiv.org/abs/2010.11929>`__ by Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk
+Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob
+Uszkoreit, Neil Houlsby. It's the first paper that successfully trains a Transformer encoder on ImageNet, attaining
+very good results compared to familiar convolutional architectures.
+
+
+The abstract from the paper is the following:
+
+*While the Transformer architecture has become the de-facto standard for natural language processing tasks, its
+applications to computer vision remain limited. In vision, attention is either applied in conjunction with
+convolutional networks, or used to replace certain components of convolutional networks while keeping their overall
+structure in place. We show that this reliance on CNNs is not necessary and a pure transformer applied directly to
+sequences of image patches can perform very well on image classification tasks. When pre-trained on large amounts of
+data and transferred to multiple mid-sized or small image recognition benchmarks (ImageNet, CIFAR-100, VTAB, etc.),
+Vision Transformer (ViT) attains excellent results compared to state-of-the-art convolutional networks while requiring
+substantially fewer computational resources to train.*
+
+Tips:
+
+- To feed images to the Transformer encoder, each image is split into a sequence of fixed-size non-overlapping patches,
+  which are then linearly embedded. A [CLS] token is added to serve as representation of an entire image, which can be
+  used for classification. The authors also add absolute position embeddings, and feed the resulting sequence of
+  vectors to a standard Transformer encoder.
+- The Vision Transformer was pre-trained using a resolution of 224x224. During fine-tuning, it is often beneficial to
+  use a higher resolution than pre-training `(Touvron et al., 2019) <https://arxiv.org/abs/1906.06423>`__, `(Kolesnikov
+  et al., 2020) <https://arxiv.org/abs/1912.11370>`__. The authors report the best results with a resolution of 384x384
+  during fine-tuning.
+- As the Vision Transformer expects each image to be of the same size (resolution), one can use
+  :class:`~transformers.ViTFeatureExtractor` to resize (or rescale) and normalize images for the model.
+- Both the patch resolution and image resolution used during pre-training or fine-tuning are reflected in the name of
+  each checkpoint. For example, :obj:`google/vit-base-patch16-224` refers to a base-sized architecture with patch
+  resolution of 16x16 and fine-tuning resolution of 224x224. All checkpoints can be found on the `hub
+  <https://huggingface.co/models?search=vit>`__.
+- The available checkpoints are either (1) pre-trained on `ImageNet-21k <http://www.image-net.org/>`__ (a collection of
+  14 million images and 21k classes) only, or (2) also fine-tuned on `ImageNet
+  <http://www.image-net.org/challenges/LSVRC/2012/>`__ (also referred to as ILSVRC 2012, a collection of 1.3 million
+  images and 1,000 classes).
+- The best results are obtained with supervised pre-training, which is not the case in NLP. The authors also performed
+  an experiment with a self-supervised pre-training objective, namely masked patched prediction (inspired by masked
+  language modeling). With this approach, the smaller ViT-B/16 model achieves 79.9% accuracy on ImageNet, a significant
+  improvement of 2% to training from scratch, but still 4% behind supervised pre-training.
+
+
+The original code (written in JAX) can be found `here <https://github.com/google-research/vision_transformer>`__.
+
+Note that we converted the weights from Ross Wightman's `timm library
+<https://github.com/rwightman/pytorch-image-models>`__, who already converted the weights from JAX to PyTorch. Credits
+go to him!
+
+
+ViTConfig
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.ViTConfig
+    :members:
+
+
+ViTFeatureExtractor
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.ViTFeatureExtractor
+    :members: __call__
+
+
+ViTModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.ViTModel
+    :members: forward
+
+
+ViTForImageClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.ViTForImageClassification
+    :members: forward

setup.py

@@ -105,6 +105,7 @@
     "onnxruntime>=1.4.0",
     "packaging",
     "parameterized",
+    "Pillow",
     "protobuf",
     "psutil",
     "pydantic",
@@ -225,6 +226,7 @@ def run(self):
 
 extras["serving"] = deps_list("pydantic", "uvicorn", "fastapi", "starlette")
 extras["speech"] = deps_list("soundfile", "torchaudio")
+extras["vision"] = deps_list("Pillow")
 
 extras["sentencepiece"] = deps_list("sentencepiece", "protobuf")
 extras["testing"] = (
@@ -235,7 +237,7 @@ def run(self):
 extras["docs"] = deps_list("recommonmark", "sphinx", "sphinx-markdown-tables", "sphinx-rtd-theme", "sphinx-copybutton")
 extras["quality"] = deps_list("black", "isort", "flake8")
 
-extras["all"] = extras["tf"] + extras["torch"] + extras["flax"] + extras["sentencepiece"] + extras["tokenizers"]
+extras["all"] = extras["tf"] + extras["torch"] + extras["flax"] + extras["sentencepiece"] + extras["tokenizers"] + extras["speech"] + extras["vision"]
 
 extras["dev"] = (
     extras["all"]

src/transformers/__init__.py

@@ -211,6 +211,7 @@
         "TransfoXLCorpus",
         "TransfoXLTokenizer",
     ],
+    "models.vit": ["VIT_PRETRAINED_CONFIG_ARCHIVE_MAP", "ViTConfig"],
     "models.wav2vec2": [
         "WAV_2_VEC_2_PRETRAINED_CONFIG_ARCHIVE_MAP",
         "Wav2Vec2Config",
@@ -297,7 +298,7 @@
         name for name in dir(dummy_sentencepiece_objects) if not name.startswith("_")
     ]
 
-# tokenziers-backed objects
+# tokenizers-backed objects
 if is_tokenizers_available():
     # Fast tokenizers
     _import_structure["models.convbert"].append("ConvBertTokenizerFast")
@@ -346,6 +347,7 @@
 # Vision-specific objects
 if is_vision_available():
     _import_structure["image_utils"] = ["ImageFeatureExtractionMixin"]
+    _import_structure["models.vit"].append("ViTFeatureExtractor")
 else:
     from .utils import dummy_vision_objects
 
@@ -423,6 +425,7 @@
     _import_structure["models.auto"].extend(
         [
             "MODEL_FOR_CAUSAL_LM_MAPPING",
+            "MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING",
             "MODEL_FOR_MASKED_LM_MAPPING",
             "MODEL_FOR_MULTIPLE_CHOICE_MAPPING",
             "MODEL_FOR_NEXT_SENTENCE_PREDICTION_MAPPING",
@@ -839,6 +842,14 @@
             "load_tf_weights_in_transfo_xl",
         ]
     )
+    _import_structure["models.vit"].extend(
+        [
+            "VIT_PRETRAINED_MODEL_ARCHIVE_LIST",
+            "ViTForImageClassification",
+            "ViTModel",
+            "ViTPreTrainedModel",
+        ]
+    )
     _import_structure["models.wav2vec2"].extend(
         [
             "WAV_2_VEC_2_PRETRAINED_MODEL_ARCHIVE_LIST",
@@ -1271,7 +1282,6 @@
         name for name in dir(dummy_flax_objects) if not name.startswith("_")
     ]
 
-
 # Direct imports for type-checking
 if TYPE_CHECKING:
     # Configuration
@@ -1437,6 +1447,7 @@
         TransfoXLCorpus,
         TransfoXLTokenizer,
     )
+    from .models.vit import VIT_PRETRAINED_CONFIG_ARCHIVE_MAP, ViTConfig
     from .models.wav2vec2 import (
         WAV_2_VEC_2_PRETRAINED_CONFIG_ARCHIVE_MAP,
         Wav2Vec2Config,
@@ -1559,6 +1570,7 @@
 
     if is_vision_available():
         from .image_utils import ImageFeatureExtractionMixin
+        from .models.vit import ViTFeatureExtractor
     else:
         from .utils.dummy_vision_objects import *
 
@@ -1624,6 +1636,7 @@
         )
         from .models.auto import (
             MODEL_FOR_CAUSAL_LM_MAPPING,
+            MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING,
             MODEL_FOR_MASKED_LM_MAPPING,
             MODEL_FOR_MULTIPLE_CHOICE_MAPPING,
             MODEL_FOR_NEXT_SENTENCE_PREDICTION_MAPPING,
@@ -1962,6 +1975,12 @@
             TransfoXLPreTrainedModel,
             load_tf_weights_in_transfo_xl,
         )
+        from .models.vit import (
+            VIT_PRETRAINED_MODEL_ARCHIVE_LIST,
+            ViTForImageClassification,
+            ViTModel,
+            ViTPreTrainedModel,
+        )
         from .models.wav2vec2 import (
             WAV_2_VEC_2_PRETRAINED_MODEL_ARCHIVE_LIST,
             Wav2Vec2ForCTC,
@@ -2327,6 +2346,7 @@
         # Import the same objects as dummies to get them in the namespace.
         # They will raise an import error if the user tries to instantiate / use them.
         from .utils.dummy_flax_objects import *
+
 else:
     import importlib
     import os

src/transformers/dependency_versions_table.py

@@ -24,6 +24,7 @@
     "onnxruntime": "onnxruntime>=1.4.0",
     "packaging": "packaging",
     "parameterized": "parameterized",
+    "Pillow": "Pillow",
     "protobuf": "protobuf",
     "psutil": "psutil",
     "pydantic": "pydantic",