[feat] Add FLAVA model (#16654)

* [WIP] Add FLAVA model

This PR aims to add [FLAVA](ihttps://arxiv.org/abs/2112.04482) model to the transformers repo.

Following checklist delineates the list of things to be done for this PR
to be complete:

[x] Flava init
[x] Flava base models
[x] Flava layers
[x] Flava Configs
[x] Flava encoders
[x] Flava pretraining models
[ ] Flava classification/retrieval models (To be added in a separate PR)
[x] Documentation updates 
[x] Imports updates 
[x] Argstring updates
[x] Flava pretrained checkpoints 
[x] Flava tests
[x] Flava processors 
[x] Sanity check
[x] Lint

README.md

@@ -265,6 +265,7 @@ Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih.
 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.
+1. **[FLAVA](https://huggingface.co/docs/transformers/main/model_doc/flava)** (from Facebook AI) released with the paper [FLAVA: A Foundational Language And Vision Alignment Model](https://arxiv.org/abs/2112.04482) by Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela.
 1. **[FNet](https://huggingface.co/docs/transformers/model_doc/fnet)** (from Google Research) released with the paper [FNet: Mixing Tokens with Fourier Transforms](https://arxiv.org/abs/2105.03824) by James Lee-Thorp, Joshua Ainslie, Ilya Eckstein, Santiago Ontanon.
 1. **[Funnel Transformer](https://huggingface.co/docs/transformers/model_doc/funnel)** (from CMU/Google Brain) released with the paper [Funnel-Transformer: Filtering out Sequential Redundancy for Efficient Language Processing](https://arxiv.org/abs/2006.03236) by Zihang Dai, Guokun Lai, Yiming Yang, Quoc V. Le.
 1. **[GLPN](https://huggingface.co/docs/transformers/main/model_doc/glpn)** (from KAIST) released with the paper [Global-Local Path Networks for Monocular Depth Estimation with Vertical CutDepth](https://arxiv.org/abs/2201.07436) by Doyeon Kim, Woonghyun Ga, Pyungwhan Ahn, Donggyu Joo, Sehwan Chun, Junmo Kim.

README_ko.md

@@ -244,6 +244,7 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
 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.
+1. **[FLAVA](https://huggingface.co/docs/transformers/main/model_doc/flava)** (from Facebook AI) released with the paper [FLAVA: A Foundational Language And Vision Alignment Model](https://arxiv.org/abs/2112.04482) by Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela.
 1. **[FNet](https://huggingface.co/docs/transformers/model_doc/fnet)** (from Google Research) released with the paper [FNet: Mixing Tokens with Fourier Transforms](https://arxiv.org/abs/2105.03824) by James Lee-Thorp, Joshua Ainslie, Ilya Eckstein, Santiago Ontanon.
 1. **[Funnel Transformer](https://huggingface.co/docs/transformers/model_doc/funnel)** (from CMU/Google Brain) released with the paper [Funnel-Transformer: Filtering out Sequential Redundancy for Efficient Language Processing](https://arxiv.org/abs/2006.03236) by Zihang Dai, Guokun Lai, Yiming Yang, Quoc V. Le.
 1. **[GLPN](https://huggingface.co/docs/transformers/main/model_doc/glpn)** (from KAIST) released with the paper [Global-Local Path Networks for Monocular Depth Estimation with Vertical CutDepth](https://arxiv.org/abs/2201.07436) by Doyeon Kim, Woonghyun Ga, Pyungwhan Ahn, Donggyu Joo, Sehwan Chun, Junmo Kim.

README_zh-hans.md

@@ -268,6 +268,7 @@ conda install -c huggingface transformers
 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。
+1. **[FLAVA](https://huggingface.co/docs/transformers/main/model_doc/flava)** (来自 Facebook AI) 伴随论文 [FLAVA: A Foundational Language And Vision Alignment Model](https://arxiv.org/abs/2112.04482) 由 Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela 发布。
 1. **[FNet](https://huggingface.co/docs/transformers/model_doc/fnet)** (来自 Google Research) 伴随论文 [FNet: Mixing Tokens with Fourier Transforms](https://arxiv.org/abs/2105.03824) 由 James Lee-Thorp, Joshua Ainslie, Ilya Eckstein, Santiago Ontanon 发布。
 1. **[Funnel Transformer](https://huggingface.co/docs/transformers/model_doc/funnel)** (来自 CMU/Google Brain) 伴随论文 [Funnel-Transformer: Filtering out Sequential Redundancy for Efficient Language Processing](https://arxiv.org/abs/2006.03236) 由 Zihang Dai, Guokun Lai, Yiming Yang, Quoc V. Le 发布。
 1. **[GLPN](https://huggingface.co/docs/transformers/main/model_doc/glpn)** (来自 KAIST) 伴随论文 [Global-Local Path Networks for Monocular Depth Estimation with Vertical CutDepth](https://arxiv.org/abs/2201.07436) 由 Doyeon Kim, Woonghyun Ga, Pyungwhan Ahn, Donggyu Joo, Sehwan Chun, Junmo Kim 发布。

README_zh-hant.md

@@ -280,6 +280,7 @@ conda install -c huggingface transformers
 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.
+1. **[FLAVA](https://huggingface.co/docs/transformers/main/model_doc/flava)** (from Facebook AI) released with the paper [FLAVA: A Foundational Language And Vision Alignment Model](https://arxiv.org/abs/2112.04482) by Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela.
 1. **[FNet](https://huggingface.co/docs/transformers/model_doc/fnet)** (from Google Research) released with the paper [FNet: Mixing Tokens with Fourier Transforms](https://arxiv.org/abs/2105.03824) by James Lee-Thorp, Joshua Ainslie, Ilya Eckstein, Santiago Ontanon.
 1. **[Funnel Transformer](https://huggingface.co/docs/transformers/model_doc/funnel)** (from CMU/Google Brain) released with the paper [Funnel-Transformer: Filtering out Sequential Redundancy for Efficient Language Processing](https://arxiv.org/abs/2006.03236) by Zihang Dai, Guokun Lai, Yiming Yang, Quoc V. Le.
 1. **[GLPN](https://huggingface.co/docs/transformers/main/model_doc/glpn)** (from KAIST) released with the paper [Global-Local Path Networks for Monocular Depth Estimation with Vertical CutDepth](https://arxiv.org/abs/2201.07436) by Doyeon Kim, Woonghyun Ga, Pyungwhan Ahn, Donggyu Joo, Sehwan Chun, Junmo Kim.

docs/source/en/_toctree.yml

@@ -216,6 +216,8 @@
       title: Encoder Decoder Models
     - local: model_doc/flaubert
       title: FlauBERT
+    - local: model_doc/flava
+      title: FLAVA
     - local: model_doc/fnet
       title: FNet
     - local: model_doc/fsmt

docs/source/en/index.mdx

@@ -86,6 +86,7 @@ The library currently contains JAX, PyTorch and TensorFlow implementations, pret
 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.
+1. **[FLAVA](model_doc/flava)** (from Facebook AI) released with the paper [FLAVA: A Foundational Language And Vision Alignment Model](https://arxiv.org/abs/2112.04482) by Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela.
 1. **[FNet](model_doc/fnet)** (from Google Research) released with the paper [FNet: Mixing Tokens with Fourier Transforms](https://arxiv.org/abs/2105.03824) by James Lee-Thorp, Joshua Ainslie, Ilya Eckstein, Santiago Ontanon.
 1. **[Funnel Transformer](model_doc/funnel)** (from CMU/Google Brain) released with the paper [Funnel-Transformer: Filtering out Sequential Redundancy for Efficient Language Processing](https://arxiv.org/abs/2006.03236) by Zihang Dai, Guokun Lai, Yiming Yang, Quoc V. Le.
 1. **[GLPN](model_doc/glpn)** (from KAIST) released with the paper [Global-Local Path Networks for Monocular Depth Estimation with Vertical CutDepth](https://arxiv.org/abs/2201.07436) by Doyeon Kim, Woonghyun Ga, Pyungwhan Ahn, Donggyu Joo, Sehwan Chun, Junmo Kim.
@@ -204,6 +205,7 @@ Flax), PyTorch, and/or TensorFlow.
 |       Encoder decoder       |       ❌       |       ❌       |       ✅        |         ✅         |      ✅      |
 | FairSeq Machine-Translation |       ✅       |       ❌       |       ✅        |         ❌         |      ❌      |
 |          FlauBERT           |       ✅       |       ❌       |       ✅        |         ✅         |      ❌      |
+|            Flava            |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |            FNet             |       ✅       |       ✅       |       ✅        |         ❌         |      ❌      |
 |     Funnel Transformer      |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
 |            GLPN             |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |

docs/source/en/model_doc/flava.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.
+-->
+
+# FLAVA
+
+## Overview
+
+The FLAVA model was proposed in [FLAVA: A Foundational Language And Vision Alignment Model](https://arxiv.org/abs/2112.04482) by Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela and is accepted at CVPR 2022.
+
+The paper aims at creating a single unified foundation model which can work across vision, language 
+as well as vision-and-language multimodal tasks.
+
+The abstract from the paper is the following:
+
+*State-of-the-art vision and vision-and-language models rely on large-scale visio-linguistic pretraining for obtaining good performance on a variety 
+of downstream tasks. Generally, such models are often either cross-modal (contrastive) or multi-modal 
+(with earlier fusion) but not both; and they often only target specific modalities or tasks. A promising 
+direction would be to use a single holistic universal model, as a "foundation", that targets all modalities 
+at once -- a true vision and language foundation model should be good at vision tasks, language tasks, and 
+cross- and multi-modal vision and language tasks. We introduce FLAVA as such a model and demonstrate 
+impressive performance on a wide range of 35 tasks spanning these target modalities.*
+
+
+This model was contributed by [aps](https://huggingface.co/aps). The original code can be found [here](https://github.com/facebookresearch/multimodal/tree/main/examples/flava).
+
+
+## FlavaConfig
+
+[[autodoc]] FlavaConfig
+
+## FlavaTextConfig
+
+[[autodoc]] FlavaTextConfig
+
+## FlavaImageConfig
+
+[[autodoc]] FlavaImageConfig
+
+## FlavaMultimodalConfig
+
+[[autodoc]] FlavaMultimodalConfig
+
+## FlavaImageCodebookConfig
+
+[[autodoc]] FlavaImageCodebookConfig
+
+## FlavaProcessor
+
+[[autodoc]] FlavaProcessor
+
+## FlavaFeatureExtractor
+
+[[autodoc]] FlavaFeatureExtractor
+
+## FlavaForPreTraining
+
+[[autodoc]] FlavaForPreTraining
+    - forward
+
+## FlavaModel
+
+[[autodoc]] FlavaModel
+    - forward
+    - get_text_features
+    - get_image_features
+
+## FlavaImageCodebook
+
+[[autodoc]] FlavaImageCodebook
+    - forward
+    - get_codebook_indices
+    - get_codebook_probs
+
+## FlavaTextModel
+
+[[autodoc]] FlavaTextModel
+    - forward
+
+## FlavaImageModel
+
+[[autodoc]] FlavaImageModel
+    - forward
+
+## FlavaMultimodalModel
+
+[[autodoc]] FlavaMultimodalModel
+    - forward

src/transformers/__init__.py

@@ -198,6 +198,14 @@ _import_structure = {
     "models.electra": ["ELECTRA_PRETRAINED_CONFIG_ARCHIVE_MAP", "ElectraConfig", "ElectraTokenizer"],
     "models.encoder_decoder": ["EncoderDecoderConfig"],
     "models.flaubert": ["FLAUBERT_PRETRAINED_CONFIG_ARCHIVE_MAP", "FlaubertConfig", "FlaubertTokenizer"],
+    "models.flava": [
+        "FLAVA_PRETRAINED_CONFIG_ARCHIVE_MAP",
+        "FlavaConfig",
+        "FlavaImageCodebookConfig",
+        "FlavaImageConfig",
+        "FlavaMultimodalConfig",
+        "FlavaTextConfig",
+    ],
     "models.fnet": ["FNET_PRETRAINED_CONFIG_ARCHIVE_MAP", "FNetConfig"],
     "models.fsmt": ["FSMT_PRETRAINED_CONFIG_ARCHIVE_MAP", "FSMTConfig", "FSMTTokenizer"],
     "models.funnel": ["FUNNEL_PRETRAINED_CONFIG_ARCHIVE_MAP", "FunnelConfig", "FunnelTokenizer"],
@@ -568,6 +576,7 @@ else:
     _import_structure["models.deit"].append("DeiTFeatureExtractor")
     _import_structure["models.detr"].append("DetrFeatureExtractor")
     _import_structure["models.dpt"].append("DPTFeatureExtractor")
+    _import_structure["models.flava"].extend(["FlavaFeatureExtractor", "FlavaProcessor"])
     _import_structure["models.glpn"].append("GLPNFeatureExtractor")
     _import_structure["models.imagegpt"].append("ImageGPTFeatureExtractor")
     _import_structure["models.layoutlmv2"].append("LayoutLMv2FeatureExtractor")
@@ -1038,6 +1047,18 @@ else:
             "FlaubertWithLMHeadModel",
         ]
     )
+    _import_structure["models.flava"].extend(
+        [
+            "FLAVA_PRETRAINED_MODEL_ARCHIVE_LIST",
+            "FlavaForPreTraining",
+            "FlavaImageCodebook",
+            "FlavaImageModel",
+            "FlavaModel",
+            "FlavaMultimodalModel",
+            "FlavaPreTrainedModel",
+            "FlavaTextModel",
+        ]
+    )
     _import_structure["models.fnet"].extend(
         [
             "FNET_PRETRAINED_MODEL_ARCHIVE_LIST",
@@ -2654,6 +2675,14 @@ if TYPE_CHECKING:
     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
+    from .models.flava import (
+        FLAVA_PRETRAINED_CONFIG_ARCHIVE_MAP,
+        FlavaConfig,
+        FlavaImageCodebookConfig,
+        FlavaImageConfig,
+        FlavaMultimodalConfig,
+        FlavaTextConfig,
+    )
     from .models.fnet import FNET_PRETRAINED_CONFIG_ARCHIVE_MAP, FNetConfig
     from .models.fsmt import FSMT_PRETRAINED_CONFIG_ARCHIVE_MAP, FSMTConfig, FSMTTokenizer
     from .models.funnel import FUNNEL_PRETRAINED_CONFIG_ARCHIVE_MAP, FunnelConfig, FunnelTokenizer
@@ -2974,6 +3003,7 @@ if TYPE_CHECKING:
         from .models.deit import DeiTFeatureExtractor
         from .models.detr import DetrFeatureExtractor
         from .models.dpt import DPTFeatureExtractor
+        from .models.flava import FlavaFeatureExtractor, FlavaProcessor
         from .models.glpn import GLPNFeatureExtractor
         from .models.imagegpt import ImageGPTFeatureExtractor
         from .models.layoutlmv2 import LayoutLMv2FeatureExtractor, LayoutLMv2Processor
@@ -3372,6 +3402,16 @@ if TYPE_CHECKING:
             FlaubertModel,
             FlaubertWithLMHeadModel,
         )
+        from .models.flava import (
+            FLAVA_PRETRAINED_MODEL_ARCHIVE_LIST,
+            FlavaForPreTraining,
+            FlavaImageCodebook,
+            FlavaImageModel,
+            FlavaModel,
+            FlavaMultimodalModel,
+            FlavaPreTrainedModel,
+            FlavaTextModel,
+        )
         from .models.fnet import (
             FNET_PRETRAINED_MODEL_ARCHIVE_LIST,
             FNetForMaskedLM,

src/transformers/models/__init__.py

@@ -54,6 +54,7 @@ from . import (
     electra,
     encoder_decoder,
     flaubert,
+    flava,
     fnet,
     fsmt,
     funnel,

src/transformers/models/auto/configuration_auto.py

@@ -66,6 +66,7 @@ CONFIG_MAPPING_NAMES = OrderedDict(
         ("canine", "CanineConfig"),
         ("roformer", "RoFormerConfig"),
         ("clip", "CLIPConfig"),
+        ("flava", "FlavaConfig"),
         ("bigbird_pegasus", "BigBirdPegasusConfig"),
         ("deit", "DeiTConfig"),
         ("luke", "LukeConfig"),
@@ -171,6 +172,7 @@ CONFIG_ARCHIVE_MAP_MAPPING_NAMES = OrderedDict(
         ("canine", "CANINE_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("roformer", "ROFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("clip", "CLIP_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("flava", "FLAVA_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("bigbird_pegasus", "BIGBIRD_PEGASUS_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("deit", "DEIT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("luke", "LUKE_PRETRAINED_CONFIG_ARCHIVE_MAP"),
@@ -268,6 +270,7 @@ MODEL_NAMES_MAPPING = OrderedDict(
         ("canine", "Canine"),
         ("roformer", "RoFormer"),
         ("clip", "CLIP"),
+        ("flava", "Flava"),
         ("bigbird_pegasus", "BigBirdPegasus"),
         ("deit", "DeiT"),
         ("luke", "LUKE"),

src/transformers/models/auto/feature_extraction_auto.py

@@ -47,6 +47,7 @@ FEATURE_EXTRACTOR_MAPPING_NAMES = OrderedDict(
         ("detr", "DetrFeatureExtractor"),
         ("layoutlmv2", "LayoutLMv2FeatureExtractor"),
         ("clip", "CLIPFeatureExtractor"),
+        ("flava", "FlavaFeatureExtractor"),
         ("perceiver", "PerceiverFeatureExtractor"),
         ("swin", "ViTFeatureExtractor"),
         ("vit_mae", "ViTFeatureExtractor"),

src/transformers/models/auto/modeling_auto.py

@@ -62,6 +62,7 @@ MODEL_MAPPING_NAMES = OrderedDict(
         ("canine", "CanineModel"),
         ("roformer", "RoFormerModel"),
         ("clip", "CLIPModel"),
+        ("flava", "FlavaModel"),
         ("bigbird_pegasus", "BigBirdPegasusModel"),
         ("deit", "DeiTModel"),
         ("luke", "LukeModel"),
@@ -131,6 +132,7 @@ MODEL_MAPPING_NAMES = OrderedDict(
 MODEL_FOR_PRETRAINING_MAPPING_NAMES = OrderedDict(
     [
         # Model for pre-training mapping
+        ("flava", "FlavaForPreTraining"),
         ("vit_mae", "ViTMAEForPreTraining"),
         ("fnet", "FNetForPreTraining"),
         ("visual_bert", "VisualBertForPreTraining"),

src/transformers/models/auto/processing_auto.py

@@ -38,6 +38,7 @@ logger = logging.get_logger(__name__)
 PROCESSOR_MAPPING_NAMES = OrderedDict(
     [
         ("clip", "CLIPProcessor"),
+        ("flava", "FLAVAProcessor"),
         ("layoutlmv2", "LayoutLMv2Processor"),
         ("layoutxlm", "LayoutXLMProcessor"),
         ("speech_to_text", "Speech2TextProcessor"),

src/transformers/models/flava/__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 Meta Platforms authors and 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 ...utils import OptionalDependencyNotAvailable, _LazyModule, is_torch_available, is_vision_available
+
+
+_import_structure = {
+    "configuration_flava": [
+        "FLAVA_PRETRAINED_CONFIG_ARCHIVE_MAP",
+        "FlavaConfig",
+        "FlavaImageCodebookConfig",
+        "FlavaImageConfig",
+        "FlavaMultimodalConfig",
+        "FlavaTextConfig",
+    ],
+}
+
+try:
+    if not is_vision_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["feature_extraction_flava"] = ["FlavaFeatureExtractor"]
+    _import_structure["processing_flava"] = ["FlavaProcessor"]
+
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_flava"] = [
+        "FLAVA_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "FlavaForPreTraining",
+        "FlavaImageCodebook",
+        "FlavaImageModel",
+        "FlavaModel",
+        "FlavaMultimodalModel",
+        "FlavaPreTrainedModel",
+        "FlavaTextModel",
+    ]
+
+if TYPE_CHECKING:
+    from .configuration_flava import (
+        FLAVA_PRETRAINED_CONFIG_ARCHIVE_MAP,
+        FlavaConfig,
+        FlavaImageCodebookConfig,
+        FlavaImageConfig,
+        FlavaMultimodalConfig,
+        FlavaTextConfig,
+    )
+
+    try:
+        if not is_vision_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .feature_extraction_flava import FlavaFeatureExtractor
+        from .processing_flava import FlavaProcessor
+
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_flava import (
+            FLAVA_PRETRAINED_MODEL_ARCHIVE_LIST,
+            FlavaForPreTraining,
+            FlavaImageCodebook,
+            FlavaImageModel,
+            FlavaModel,
+            FlavaMultimodalModel,
+            FlavaPreTrainedModel,
+            FlavaTextModel,
+        )
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)

src/transformers/models/flava/convert_dalle_to_flava_codebook.py

+# coding=utf-8
+# Copyright 2022 Meta Platforms authors and 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.
+
+import argparse
+import os
+
+import torch
+
+from transformers import FlavaImageCodebook, FlavaImageCodebookConfig
+
+
+def rreplace(s, old, new, occurrence):
+    li = s.rsplit(old, occurrence)
+    return new.join(li)
+
+
+def count_parameters(state_dict):
+    # encoder.embeddings are double copied in original FLAVA
+    return sum(param.float().sum() if "encoder.embeddings" not in key else 0 for key, param in state_dict.items())
+
+
+def upgrade_state_dict(state_dict):
+    upgrade = {}
+
+    group_keys = ["group_1", "group_2", "group_3", "group_4"]
+    for key, value in state_dict.items():
+        for group_key in group_keys:
+            if group_key in key:
+                key = key.replace(f"{group_key}.", f"{group_key}.group.")
+
+        if "res_path" in key:
+            key = key.replace("res_path.", "res_path.path.")
+
+        if key.endswith(".w"):
+            key = rreplace(key, ".w", ".weight", 1)
+        if key.endswith(".b"):
+            key = rreplace(key, ".b", ".bias", 1)
+
+        upgrade[key] = value.float()
+
+    return upgrade
+
+
+@torch.no_grad()
+def convert_dalle_checkpoint(checkpoint_path, pytorch_dump_folder_path, config_path=None, save_checkpoint=True):
+    """
+    Copy/paste/tweak model's weights to transformers design.
+    """
+    from dall_e import Encoder
+
+    encoder = Encoder()
+    if os.path.exists(checkpoint_path):
+        ckpt = torch.load(checkpoint_path)
+    else:
+        ckpt = torch.hub.load_state_dict_from_url(checkpoint_path)
+
+    if isinstance(ckpt, Encoder):
+        ckpt = ckpt.state_dict()
+    encoder.load_state_dict(ckpt)
+
+    if config_path is not None:
+        config = FlavaImageCodebookConfig.from_pretrained(config_path)
+    else:
+        config = FlavaImageCodebookConfig()
+
+    hf_model = FlavaImageCodebook(config).eval()
+    state_dict = encoder.state_dict()
+
+    hf_state_dict = upgrade_state_dict(state_dict)
+    hf_model.load_state_dict(hf_state_dict)
+    hf_state_dict = hf_model.state_dict()
+    hf_count = count_parameters(hf_state_dict)
+    state_dict_count = count_parameters(state_dict)
+
+    assert torch.allclose(hf_count, state_dict_count, atol=1e-3)
+
+    if save_checkpoint:
+        hf_model.save_pretrained(pytorch_dump_folder_path)
+    else:
+        return hf_state_dict
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--pytorch_dump_folder_path", default=None, type=str, help="Path to the output PyTorch model.")
+    parser.add_argument("--checkpoint_path", default=None, type=str, help="Path to flava checkpoint")
+    parser.add_argument("--config_path", default=None, type=str, help="Path to hf config.json of model to convert")
+    args = parser.parse_args()
+
+    convert_dalle_checkpoint(args.checkpoint_path, args.pytorch_dump_folder_path, args.config_path)

src/transformers/models/flava/convert_flava_original_pytorch_to_hf.py

+# coding=utf-8
+# Copyright 2022 Meta Platforms authors and 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.
+
+import argparse
+import os
+
+import torch
+
+from transformers import FlavaConfig, FlavaForPreTraining
+from transformers.models.flava.convert_dalle_to_flava_codebook import convert_dalle_checkpoint
+
+
+def count_parameters(state_dict):
+    # encoder.embeddings are double copied in original FLAVA
+    return sum(param.float().sum() if "encoder.embeddings" not in key else 0 for key, param in state_dict.items())
+
+
+def upgrade_state_dict(state_dict, codebook_state_dict):
+    upgrade = {}
+
+    for key, value in state_dict.items():
+        if "text_encoder.embeddings" in key or "image_encoder.embeddings" in key:
+            continue
+
+        key = key.replace("heads.cmd.mim_head.cls.predictions", "mmm_image_head")
+        key = key.replace("heads.cmd.mlm_head.cls.predictions", "mmm_text_head")
+        key = key.replace("heads.cmd.itm_head.cls", "itm_head")
+        key = key.replace("heads.cmd.itm_head.pooler", "itm_head.pooler")
+        key = key.replace("heads.cmd.clip_head.logit_scale", "flava.logit_scale")
+        key = key.replace("heads.fairseq_mlm.cls.predictions", "mlm_head")
+        key = key.replace("heads.imagenet.mim_head.cls.predictions", "mim_head")
+        key = key.replace("mm_text_projection", "flava.text_to_mm_projection")
+        key = key.replace("mm_image_projection", "flava.image_to_mm_projection")
+        key = key.replace("image_encoder.module", "flava.image_model")
+        key = key.replace("text_encoder.module", "flava.text_model")
+        key = key.replace("mm_encoder.module.encoder.cls_token", "flava.multimodal_model.cls_token")
+        key = key.replace("mm_encoder.module", "flava.multimodal_model")
+        key = key.replace("text_projection", "flava.text_projection")
+        key = key.replace("image_projection", "flava.image_projection")
+
+        upgrade[key] = value.float()
+
+    for key, value in codebook_state_dict.items():
+        upgrade[f"image_codebook.{key}"] = value
+
+    return upgrade
+
+
+@torch.no_grad()
+def convert_flava_checkpoint(checkpoint_path, codebook_path, pytorch_dump_folder_path, config_path=None):
+    """
+    Copy/paste/tweak model's weights to transformers design.
+    """
+    if config_path is not None:
+        config = FlavaConfig.from_pretrained(config_path)
+    else:
+        config = FlavaConfig()
+
+    hf_model = FlavaForPreTraining(config).eval()
+
+    codebook_state_dict = convert_dalle_checkpoint(codebook_path, None, save_checkpoint=False)
+
+    if os.path.exists(checkpoint_path):
+        state_dict = torch.load(checkpoint_path, map_location="cpu")
+    else:
+        state_dict = torch.hub.load_state_dict_from_url(checkpoint_path, map_location="cpu")
+
+    hf_state_dict = upgrade_state_dict(state_dict, codebook_state_dict)
+    hf_model.load_state_dict(hf_state_dict)
+    hf_state_dict = hf_model.state_dict()
+    hf_count = count_parameters(hf_state_dict)
+    state_dict_count = count_parameters(state_dict) + count_parameters(codebook_state_dict)
+
+    assert torch.allclose(hf_count, state_dict_count, atol=1e-3)
+
+    hf_model.save_pretrained(pytorch_dump_folder_path)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--pytorch_dump_folder_path", default=None, type=str, help="Path to the output PyTorch model.")
+    parser.add_argument("--checkpoint_path", default=None, type=str, help="Path to flava checkpoint")
+    parser.add_argument("--codebook_path", default=None, type=str, help="Path to flava codebook checkpoint")
+    parser.add_argument("--config_path", default=None, type=str, help="Path to hf config.json of model to convert")
+    args = parser.parse_args()
+
+    convert_flava_checkpoint(args.checkpoint_path, args.codebook_path, args.pytorch_dump_folder_path, args.config_path)

src/transformers/models/flava/feature_extraction_flava.py

+# coding=utf-8
+# Copyright 2022 Meta Platforms authors and 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.
+"""Feature extractor class for FLAVA."""
+
+import math
+import random
+from functools import lru_cache
+from typing import Any, List, Optional, Tuple, Union
+
+import numpy as np
+from PIL import Image
+
+from ...feature_extraction_utils import BatchFeature, FeatureExtractionMixin
+from ...image_utils import ImageFeatureExtractionMixin, is_torch_tensor
+from ...utils import TensorType, logging
+
+
+logger = logging.get_logger(__name__)
+
+
+# These values are taken from CLIP
+FLAVA_IMAGE_MEAN = [0.48145466, 0.4578275, 0.40821073]
+FLAVA_IMAGE_STD = [0.26862954, 0.26130258, 0.27577711]
+FLAVA_CODEBOOK_MEAN = [0.0, 0.0, 0.0]
+FLAVA_CODEBOOK_STD = [1.0, 1.0, 1.0]
+LOGIT_LAPLACE_EPS: float = 0.1
+
+
+# Inspired from https://github.com/microsoft/unilm/blob/master/beit/masking_generator.py
+class FlavaMaskingGenerator:
+    def __init__(
+        self,
+        input_size: Union[int, Tuple[int, int]] = 14,
+        total_mask_patches: int = 75,
+        mask_group_max_patches: Optional[int] = None,
+        mask_group_min_patches: int = 16,
+        mask_group_min_aspect_ratio: Optional[float] = 0.3,
+        mask_group_max_aspect_ratio: float = None,
+    ):
+        if not isinstance(input_size, tuple):
+            input_size = (input_size,) * 2
+        self.height, self.width = input_size
+
+        self.num_patches = self.height * self.width
+        self.total_mask_patches = total_mask_patches
+
+        self.mask_group_min_patches = mask_group_min_patches
+        self.mask_group_max_patches = total_mask_patches if mask_group_max_patches is None else mask_group_max_patches
+
+        mask_group_max_aspect_ratio = mask_group_max_aspect_ratio or 1 / mask_group_min_aspect_ratio
+        self.log_aspect_ratio = (math.log(mask_group_min_aspect_ratio), math.log(mask_group_max_aspect_ratio))
+
+    def __repr__(self):
+        repr_str = "MaskingGenerator(%d, %d -> [%d ~ %d], max = %d, %.3f ~ %.3f)" % (
+            self.height,
+            self.width,
+            self.mask_group_min_patches,
+            self.mask_group_max_patches,
+            self.total_mask_patches,
+            self.log_aspect_ratio[0],
+            self.log_aspect_ratio[1],
+        )
+        return repr_str
+
+    def get_shape(self):
+        return self.height, self.width
+
+    def _mask(self, mask, max_mask_patches):
+        delta = 0
+        for _attempt in range(10):
+            target_area = random.uniform(self.mask_group_min_patches, max_mask_patches)
+            aspect_ratio = math.exp(random.uniform(*self.log_aspect_ratio))
+            height = int(round(math.sqrt(target_area * aspect_ratio)))
+            width = int(round(math.sqrt(target_area / aspect_ratio)))
+            if width < self.width and height < self.height:
+                top = random.randint(0, self.height - height)
+                left = random.randint(0, self.width - width)
+
+                num_masked = mask[top : top + height, left : left + width].sum()
+                # Overlap
+                if 0 < height * width - num_masked <= max_mask_patches:
+                    for i in range(top, top + height):
+                        for j in range(left, left + width):
+                            if mask[i, j] == 0:
+                                mask[i, j] = 1
+                                delta += 1
+
+                if delta > 0:
+                    break
+        return delta
+
+    def __call__(self):
+        mask = np.zeros(shape=self.get_shape(), dtype=int)
+        mask_count = 0
+        while mask_count < self.total_mask_patches:
+            max_mask_patches = self.total_mask_patches - mask_count
+            max_mask_patches = min(max_mask_patches, self.mask_group_max_patches)
+
+            delta = self._mask(mask, max_mask_patches)
+            if delta == 0:
+                break
+            else:
+                mask_count += delta
+
+        return mask
+
+
+class FlavaFeatureExtractor(FeatureExtractionMixin, ImageFeatureExtractionMixin):
+    r"""
+    Constructs a FLAVA 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`, *optional*, defaults to 224):
+            Resize the input to the given size. Only has an effect if `do_resize` is set to `True`.
+        resample (`int`, *optional*, defaults to `PIL.Image.BICUBIC`):
+            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
+        do_center_crop (`bool`, *optional*, defaults to `True`):
+            Whether to crop the input at the center. If the input size is smaller than `crop_size` along any edge, the
+            image is padded with 0's and then center cropped.
+        crop_size (`int`, *optional*, defaults to 224):
+            Desired output size when applying center-cropping. Only has an effect if `do_center_crop` is set to `True`.
+        do_normalize (`bool`, *optional*, defaults to `True`):
+            Whether or not to normalize the input with `image_mean` and `image_std`.
+        image_mean (`Tuple[float, float, float]`, *optional*, defaults to `[0.485, 0.456, 0.406]`):
+            The sequence of means for each channel, to be used when normalizing images.
+        image_std (`Tuple[float, float, float]`, *optional*, defaults to `[0.229, 0.224, 0.225]`):
+            The sequence of standard deviations for each channel, to be used when normalizing images.
+        input_size_patches (`int`, *optional*, defaults to 14):
+            Number of patches in the image in height and width direction. 14x14 = 196 total patches.
+        total_mask_patches (`int`, *optional*, defaults to 75):
+            Total number of patches that should be masked.
+        mask_group_min_patches (`int`, *optional*, defaults to 16):
+            Minimum number of patches that should be masked.
+        mask_group_max_patches (`int`, *optional*, defaults to None):
+            Maximum number of patches that should be masked.
+        mask_group_min_aspect_ratio (`float`, *optional*, defaults to 0.3):
+            Minimum aspect ratio of the mask window.
+        mask_group_max_aspect_ratio (`float`, *optional*, defaults to None):
+            Maximum aspect ratio of the mask window
+        codebook_do_resize (`bool`, *optional*, defaults to `True`):
+            Whether to resize the input for codebook to a certain `codebook_size`.
+        codebook_size (`int`, *optional*, defaults to 224):
+            Resize the input for codebook to the given size. Only has an effect if `codebook_do_resize` is set to
+            `True`.
+        codebook_resample (`int`, *optional*, defaults to `PIL.Image.BICUBIC`):
+            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
+        codebook_do_center_crop (`bool`, *optional*, defaults to `True`):
+            Whether to crop the input for codebook at the center. If the input size is smaller than
+            `codebook_crop_size` along any edge, the image is padded with 0's and then center cropped.
+        codebook_crop_size (`int`, *optional*, defaults to 224):
+            Desired output size for codebook input when applying center-cropping. Only has an effect if
+            `codebook_do_center_crop` is set to `True`.
+        codebook_do_normalize (`bool`, *optional*, defaults to `True`):
+            Whether or not to normalize the input for codebook with `codebook_image_mean` and `codebook_image_std`.
+        codebook_image_mean (`Tuple[float, float, float]`, *optional*, defaults to `[0, 0, 0]`):
+            The sequence of means for each channel, to be used when normalizing images for codebook.
+        codebook_image_std (`Tuple[float, float, float]`, *optional*, defaults to `[0.5, 0.5, 0.5]`):
+            The sequence of standard deviations for each channel, to be used when normalizing images for codebook.
+
+    """
+
+    model_input_names = ["pixel_values"]
+
+    def __init__(
+        self,
+        do_resize: bool = True,
+        size: Union[int, Tuple[int, int]] = 224,
+        resample: int = Image.BICUBIC,
+        do_center_crop: bool = True,
+        crop_size: Union[int, Tuple[int, int]] = 224,
+        do_normalize: bool = True,
+        image_mean: Tuple[float, float, float] = FLAVA_IMAGE_MEAN,
+        image_std: Tuple[float, float, float] = FLAVA_IMAGE_STD,
+        # Mask related params
+        input_size_patches: int = 14,
+        total_mask_patches: int = 75,
+        mask_group_min_patches: int = 16,
+        mask_group_max_patches: Optional[int] = None,
+        mask_group_min_aspect_ratio: float = 0.3,
+        mask_group_max_aspect_ratio: Optional[float] = None,
+        # Codebook related params
+        codebook_do_resize: bool = True,
+        codebook_size: bool = 112,
+        codebook_resample: int = Image.LANCZOS,
+        codebook_do_center_crop: bool = True,
+        codebook_crop_size: int = 112,
+        codebook_do_map_pixels: bool = True,
+        codebook_do_normalize: bool = True,
+        codebook_image_mean: Tuple[float, float, float] = FLAVA_CODEBOOK_MEAN,
+        codebook_image_std: Tuple[float, float, float] = FLAVA_CODEBOOK_STD,
+        **kwargs: Any,
+    ):
+        super().__init__(**kwargs)
+        self.do_resize = do_resize
+        self.size = size
+        self.resample = resample
+        self.do_center_crop = do_center_crop
+        self.crop_size = crop_size
+        self.do_normalize = do_normalize
+        self.image_mean = image_mean
+        self.image_std = image_std
+
+        self.input_size_patches = input_size_patches
+        self.total_mask_patches = total_mask_patches
+        self.mask_group_min_patches = mask_group_min_patches
+        self.mask_group_max_patches = mask_group_max_patches
+        self.mask_group_min_aspect_ratio = mask_group_min_aspect_ratio
+        self.mask_group_max_aspect_ratio = mask_group_max_aspect_ratio
+
+        self.codebook_do_resize = codebook_do_resize
+        self.codebook_size = codebook_size
+        self.codebook_resample = codebook_resample
+        self.codebook_do_center_crop = codebook_do_center_crop
+        self.codebook_crop_size = codebook_crop_size
+        self.codebook_do_map_pixels = codebook_do_map_pixels
+        self.codebook_do_normalize = codebook_do_normalize
+        self.codebook_image_mean = codebook_image_mean
+        self.codebook_image_std = codebook_image_std
+
+    @property
+    @lru_cache()
+    def masking_generator(self):
+        return FlavaMaskingGenerator(
+            input_size=self.input_size_patches,
+            total_mask_patches=self.total_mask_patches,
+            mask_group_min_patches=self.mask_group_min_patches,
+            mask_group_max_patches=self.mask_group_max_patches,
+            mask_group_min_aspect_ratio=self.mask_group_min_aspect_ratio,
+            mask_group_max_aspect_ratio=self.mask_group_max_aspect_ratio,
+        )
+
+    def map_pixels(self, x):
+        return (1 - 2 * LOGIT_LAPLACE_EPS) * x + LOGIT_LAPLACE_EPS
+
+    def __call__(
+        self,
+        images: Union[
+            Image.Image, np.ndarray, "torch.Tensor", List[Image.Image], List[np.ndarray], List["torch.Tensor"]  # noqa
+        ],
+        return_image_mask: Optional[bool] = None,
+        return_codebook_pixels: Optional[bool] = None,
+        return_tensors: Optional[Union[str, TensorType]] = None,
+        **kwargs: Any
+    ) -> 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_image_mask (`bool`, *optional*, defaults to None):
+                If True, the processor will return `bool_masked_pos` suggesting masks for image's patch version.
+
+            return_codebook_pixels (`bool`, *optional*, defaults to None):
+                If True, the processor will return `codebook_pixel_values` providing image pixels to be used with the
+                default FLAVA codebook. Used in pretraining by Masked Image Modeling (MIM) loss.
+
+            return_tensors (`str` or [`~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.
+        """
+        # Input type checking for clearer error
+        if isinstance(images, (list, tuple)) and len(images) != 0:
+            self._ensure_format_supported(images[0])
+        else:
+            self._ensure_format_supported(images)
+
+        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]
+
+        images_for_codebook = images
+
+        # transformations (resizing + center cropping + normalization)
+        if self.do_resize and self.size is not None and self.resample is not None:
+            images = [self.resize(image=image, size=self.size, resample=self.resample) for image in images]
+        if self.do_center_crop and self.crop_size is not None:
+            images = [self.center_crop(image, self.crop_size) for image in images]
+        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}
+
+        if return_codebook_pixels:
+            images = images_for_codebook
+            if self.codebook_do_resize and self.codebook_size is not None and self.codebook_resample is not None:
+                images = [
+                    self.resize(image=image, size=self.codebook_size, resample=self.codebook_resample)
+                    for image in images
+                ]
+            if self.codebook_do_center_crop and self.codebook_crop_size is not None:
+                images = [self.center_crop(image, self.codebook_crop_size) for image in images]
+            if self.codebook_do_normalize:
+                images = [
+                    self.normalize(image=image, mean=self.codebook_image_mean, std=self.codebook_image_std)
+                    for image in images
+                ]
+            if self.codebook_do_map_pixels:
+                images = [self.map_pixels(image) for image in images]
+
+            data["codebook_pixel_values"] = images
+
+        if return_image_mask:
+            masks = [self.masking_generator() for _ in images]
+            data["bool_masked_pos"] = masks
+
+        encoded_inputs = BatchFeature(data=data, tensor_type=return_tensors)
+
+        return encoded_inputs

src/transformers/models/flava/processing_flava.py

+# coding=utf-8
+# Copyright 2022 Meta Platforms authors and 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.
+"""
+Image/Text processor class for FLAVA
+"""
+from typing import List, Optional, Union
+
+from ...image_utils import ImageInput
+from ...processing_utils import ProcessorMixin
+from ...tokenization_utils_base import BatchEncoding, PaddingStrategy, PreTokenizedInput, TextInput, TruncationStrategy
+from ...utils import TensorType
+
+
+class FlavaProcessor(ProcessorMixin):
+    r"""
+    Constructs a FLAVA processor which wraps a FLAVA feature extractor and a FLAVA tokenizer into a single processor.
+
+    [`FlavaProcessor`] offers all the functionalities of [`FlavaFeatureExtractor`] and [`BertTokenizerFast`]. See the
+    [`~FlavaProcessor.__call__`] and [`~FlavaProcessor.decode`] for more information.
+
+    Args:
+        feature_extractor ([`FlavaFeatureExtractor`]): The feature extractor is a required input.
+        tokenizer ([`BertTokenizerFast`]): The tokenizer is a required input.
+    """
+    feature_extractor_class = "FlavaFeatureExtractor"
+    tokenizer_class = ("BertTokenizer", "BertTokenizerFast")
+
+    def __init__(self, feature_extractor, tokenizer):
+        super().__init__(feature_extractor, tokenizer)
+        self.current_processor = self.feature_extractor
+
+    def __call__(
+        self,
+        images: Optional[ImageInput] = None,
+        text: Optional[Union[TextInput, PreTokenizedInput, List[TextInput], List[PreTokenizedInput]]] = None,
+        add_special_tokens: bool = True,
+        padding: Union[bool, str, PaddingStrategy] = False,
+        truncation: Union[bool, str, TruncationStrategy] = False,
+        max_length: Optional[int] = None,
+        stride: int = 0,
+        pad_to_multiple_of: Optional[int] = None,
+        return_image_mask: Optional[bool] = None,
+        return_codebook_pixels: Optional[bool] = None,
+        return_token_type_ids: Optional[bool] = None,
+        return_attention_mask: Optional[bool] = None,
+        return_overflowing_tokens: bool = False,
+        return_special_tokens_mask: bool = False,
+        return_offsets_mapping: bool = False,
+        return_length: bool = False,
+        verbose: bool = True,
+        return_tensors: Optional[Union[str, TensorType]] = None,
+        **kwargs
+    ):
+        """
+        This method uses [`FLAVAFeatureExtractor.__call__`] method to prepare image(s) for the model, and
+        [`BertTokenizerFast.__call__`] to prepare text for the model.
+
+        Please refer to the docstring of the above two methods for more information.
+        """
+
+        if text is None and images is None:
+            raise ValueError("You have to specify either text or images. Both cannot be none.")
+
+        if text is not None:
+            encoding = self.tokenizer(
+                text=text,
+                add_special_tokens=add_special_tokens,
+                padding=padding,
+                truncation=truncation,
+                max_length=max_length,
+                stride=stride,
+                pad_to_multiple_of=pad_to_multiple_of,
+                return_token_type_ids=return_token_type_ids,
+                return_attention_mask=return_attention_mask,
+                return_overflowing_tokens=return_overflowing_tokens,
+                return_special_tokens_mask=return_special_tokens_mask,
+                return_offsets_mapping=return_offsets_mapping,
+                return_length=return_length,
+                verbose=verbose,
+                return_tensors=return_tensors,
+                **kwargs,
+            )
+        if images is not None:
+            image_features = self.feature_extractor(
+                images,
+                return_image_mask=return_image_mask,
+                return_codebook_pixels=return_codebook_pixels,
+                return_tensors=return_tensors,
+                **kwargs,
+            )
+
+        if text is not None and images is not None:
+            encoding.update(image_features)
+            return encoding
+        elif text is not None:
+            return encoding
+        else:
+            return BatchEncoding(data=dict(**image_features), tensor_type=return_tensors)
+
+    def batch_decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to BertTokenizerFast's [`~PreTrainedTokenizer.batch_decode`]. Please
+        refer to the docstring of this method for more information.
+        """
+        return self.tokenizer.batch_decode(*args, **kwargs)
+
+    def decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to BertTokenizerFast's [`~PreTrainedTokenizer.decode`]. Please refer to
+        the docstring of this method for more information.
+        """
+        return self.tokenizer.decode(*args, **kwargs)