[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

@@ -0,0 +1,96 @@
+<!--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

@@ -0,0 +1,99 @@
+# 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/configuration_flava.py

@@ -0,0 +1,646 @@
+# 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.
+""" FLAVA model configurations"""
+
+import copy
+import os
+from typing import Any, Dict, Union
+
+from ...configuration_utils import PretrainedConfig
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+FLAVA_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "facebook/flava-full": "https://huggingface.co/facebook/flava-full/resolve/main/config.json",
+}
+
+
+class FlavaImageConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`FlavaImageModel`]. It is used to instantiate an
+    FLAVA 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 FLAVA
+    [facebook/flava-full](https://huggingface.co/facebook/flava-full) 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 224):
+            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.
+        mask_token (`bool`, *optional*, defaults to `True`):
+            Whether to use a mask token or not. Used in MIM (Masked Image Modeling) loss for FLAVA.
+        vocab_size (`int`, *optional*, defaults to 8192):
+            Vocabulary size of the [`FlavaImageCodebook`] used in conjunction with [`FlavaImageModel`] for MIM (Masked
+            Image Modeling) loss for FLAVA.
+
+    Example:
+
+    ```python
+    >>> from transformers import FlavaImageModel, FlavaImageConfig
+
+    >>> # Initializing a FlavaImageModel with  style configuration
+    >>> configuration = FlavaImageConfig()
+
+    >>> # Initializing a FlavaImageModel model from the  style configuration
+    >>> model = FlavaImageModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "flava_image_model"
+
+    def __init__(
+        self,
+        hidden_size: int = 768,
+        num_hidden_layers: int = 12,
+        num_attention_heads: int = 12,
+        intermediate_size: int = 3072,
+        hidden_act: int = "gelu",
+        hidden_dropout_prob: float = 0.0,
+        attention_probs_dropout_prob: float = 0.0,
+        initializer_range: float = 0.02,
+        layer_norm_eps: float = 1e-12,
+        image_size: int = 224,
+        patch_size: int = 16,
+        num_channels: int = 3,
+        qkv_bias: bool = True,
+        mask_token: bool = True,
+        vocab_size: int = 8192,
+        **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.mask_token = mask_token
+        self.vocab_size = vocab_size
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
+
+        config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
+
+        # get the image config dict if we are loading from FlavaConfig
+        if config_dict.get("model_type") == "flava":
+            config_dict = config_dict["image_config"]
+
+        if "model_type" in config_dict and hasattr(cls, "model_type") and config_dict["model_type"] != cls.model_type:
+            logger.warning(
+                f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
+                f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
+            )
+
+        return cls.from_dict(config_dict, **kwargs)
+
+
+class FlavaTextConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`FlavaTextModel`]. It is used to instantiate an
+    FLAVA 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 FLAVA
+    [facebook/flava-full](https://huggingface.co/facebook/flava-full) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 30522):
+            Vocabulary size of the BERT model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`FlavaTextModel`].
+        type_vocab_size (`int`, *optional*, defaults to 2):
+            The vocabulary size of the `token_type_ids` passed when calling [`FlavaTextModel`]. Note that even though
+            text encoder allows `token_type_ids`'s value as 2, for text-only pretraining and fine-tuning, only 1 is
+            used similar to RoBERTa.
+        max_position_embeddings (`int`, *optional*, defaults to 512):
+            The maximum sequence length that this model might ever be used with. Typically set this to something large
+            just in case (e.g., 512 or 1024 or 2048). For VL, max_length passed to model is 77.
+        position_embedding_type (`str`, *optional*, defaults to `"absolute"`):
+            Type of position embedding. Choose one of `"absolute"`, `"relative_key"`, `"relative_key_query"`. For
+            positional embeddings use `"absolute"`. For more information on `"relative_key"`, please refer to
+            [Self-Attention with Relative Position Representations (Shaw et al.)](https://arxiv.org/abs/1803.02155).
+            For more information on `"relative_key_query"`, please refer to *Method 4* in [Improve Transformer Models
+            with Better Relative Position Embeddings (Huang et al.)](https://arxiv.org/abs/2009.13658).
+        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 224):
+            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.
+
+    Example:
+
+    ```python
+    >>> from transformers import FlavaTextModel, FlavaTextConfig
+
+    >>> # Initializing a FlavaTextModel with  style configuration
+    >>> configuration = FlavaTextConfig()
+
+    >>> # Initializing a FlavaTextConfig from the  style configuration
+    >>> model = FlavaTextModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+    model_type = "flava_text_model"
+
+    def __init__(
+        self,
+        vocab_size: int = 30522,
+        type_vocab_size: int = 2,
+        max_position_embeddings: int = 512,
+        position_embedding_type: str = "absolute",
+        hidden_size: int = 768,
+        num_hidden_layers: int = 12,
+        num_attention_heads: int = 12,
+        intermediate_size: int = 3072,
+        hidden_act: str = "gelu",
+        hidden_dropout_prob: float = 0.0,
+        attention_probs_dropout_prob: float = 0.0,
+        initializer_range: float = 0.02,
+        layer_norm_eps: float = 1e-12,
+        pad_token_id: int = 0,
+        qkv_bias: bool = True,
+        **kwargs
+    ):
+        super().__init__(**kwargs)
+
+        self.vocab_size = vocab_size
+        self.type_vocab_size = type_vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.position_embedding_type = position_embedding_type
+        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.qkv_bias = qkv_bias
+        self.pad_token_id = pad_token_id
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
+
+        config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
+
+        # get the text config dict if we are loading from FlavaConfig
+        if config_dict.get("model_type") == "flava":
+            config_dict = config_dict["text_config"]
+
+        if "model_type" in config_dict and hasattr(cls, "model_type") and config_dict["model_type"] != cls.model_type:
+            logger.warning(
+                f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
+                f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
+            )
+
+        return cls.from_dict(config_dict, **kwargs)
+
+
+class FlavaMultimodalConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`FlavaMultimodalModel`]. It is used to instantiate
+    an FLAVA 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 FLAVA
+    [facebook/flava-full](https://huggingface.co/facebook/flava-full) 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.
+        qkv_bias (`bool`, *optional*, defaults to `True`):
+            Whether to add a bias to the queries, keys and values.
+        use_cls_token (`bool`, *optional*, defaults to `True`):
+            Whether to use an extra CLS token for multimodal settings. Usually needed by the FLAVA model.
+
+
+    Example:
+
+    ```python
+    >>> from transformers import FlavaMultimodalModel, FlavaMultimodalConfig
+
+    >>> # Initializing a FlavaMultimodalModel with  style configuration
+    >>> configuration = FlavaMultimodalConfig()
+
+    >>> # Initializing a FlavaMultimodalModel model from the  style configuration
+    >>> model = FlavaMultimodalModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "flava_multimodal_model"
+
+    def __init__(
+        self,
+        hidden_size: int = 768,
+        num_hidden_layers: int = 6,
+        num_attention_heads: int = 12,
+        intermediate_size: int = 3072,
+        hidden_act: int = "gelu",
+        hidden_dropout_prob: int = 0.0,
+        attention_probs_dropout_prob: int = 0.0,
+        initializer_range: float = 0.02,
+        layer_norm_eps: float = 1e-12,
+        qkv_bias: bool = True,
+        use_cls_token: bool = True,
+        **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.qkv_bias = qkv_bias
+        self.use_cls_token = use_cls_token
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
+        config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
+
+        # get the multimodal config dict if we are loading from FlavaConfig
+        if config_dict.get("model_type") == "flava":
+            config_dict = config_dict["multimodal_config"]
+
+        if "model_type" in config_dict and hasattr(cls, "model_type") and config_dict["model_type"] != cls.model_type:
+            logger.warning(
+                f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
+                f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
+            )
+
+        return cls.from_dict(config_dict, **kwargs)
+
+
+class FlavaImageCodebookConfig(PretrainedConfig):
+    model_type = "flava_image_codebook"
+
+    r"""
+    [`FlavaImageCodebookConfig`] is the configuration class to store the configuration of a [`FlavaImageCodebook`]. It
+    is used to instantiate an FLAVA 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 FLAVA
+    [facebook/flava-image-codebook](https://huggingface.co/facebook/flava-image-codebook) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        num_groups (`int`, defaults to 4):
+            Number of groups to be created. This parameter as of now doesn't affect the model and is used for some
+            internal calculation and estimations.
+        input_channels (`int`, defaults to 3):
+            Number of channels in the image to be passed.
+        num_blocks_per_group (`int`, defaults to 2):
+            Number of conv-based blocks per group.
+        hidden_size (`int`, defaults to 256):
+            Size of hidden dim for the blocks.
+        vocab_size (`int`, defaults to 8192):
+            Size of the output vocabulary for the codebook.
+        freeze (`bool`, defaults to `True`):
+            Whether to freeze the weights of the model.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        kwargs (*optional*):
+            Dictionary of keyword arguments.
+
+    Example:
+
+    ```python
+    >>> from transformers import FlavaImageCodebook, FlavaImageCodebookConfig
+
+    >>> # Initializing a FlavaImageCodebook with style configuration
+    >>> configuration = FlavaImageCodebookConfig()
+
+    >>> # Initializing a FlavaImageCodebook model from the style configuration
+    >>> model = FlavaImageCodebook(configuration)
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```
+    """
+
+    def __init__(
+        self,
+        num_groups: int = 4,
+        input_channels: int = 3,
+        num_blocks_per_group: int = 2,
+        hidden_size: int = 256,
+        vocab_size: int = 8192,
+        freeze: int = True,
+        initializer_range: float = 0.02,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.num_groups = num_groups
+        self.input_channels = input_channels
+        self.num_blocks_per_group = num_blocks_per_group
+        self.hidden_size = hidden_size
+        self.vocab_size = vocab_size
+        self.freeze = freeze
+        self.initializer_range = initializer_range
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
+
+        config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
+
+        # get the image codebook config dict if we are loading from FlavaConfig
+        if config_dict.get("model_type") == "flava":
+            config_dict = config_dict["image_codebook_config"]
+
+        if "model_type" in config_dict and hasattr(cls, "model_type") and config_dict["model_type"] != cls.model_type:
+            logger.warning(
+                f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
+                f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
+            )
+
+        return cls.from_dict(config_dict, **kwargs)
+
+
+class FlavaConfig(PretrainedConfig):
+    r"""
+    [`FlavaConfig`] is the configuration class to store the configuration of a [`FlavaModel`]. It is used to
+    instantiate FLAVA model according to the specified arguments, defining the text model, image model, image codebook
+    and multimodal model configs. Instantiating a configuration with the defaults will yield a similar configuration to
+    that of the FLAVA [facebook/flava-full](https://huggingface.co/facebook/flava-full) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        text_config_dict (`dict`, *optional*):
+            Dictionary of configuration options used to initialize [`FlavaTextConfig`].
+        image_config_dict (`dict`, *optional*):
+            Dictionary of configuration options used to initialize [`FlavaImageConfig`].
+        multimodal_config_dict (`dict`, *optional*):
+            Dictionary of configuration options used to initialize [`FlavaMultimodalConfig`].
+        hidden_size (`int`, *optional*, defaults to 768):
+            Dimensionality of the encoder layers and the pooler layer.
+        layer_norm_eps (`float`, *optional*, defaults to 1e-12):
+            The epsilon used by the layer normalization layers.
+        projection_dim (`int`, *optional*, defaults to 512):
+            Dimentionality of text and image projection layers.
+        logit_scale_init_value (`float`, *optional*, defaults to 2.6592):
+            The inital value of the *logit_scale* paramter. Default is used as per the original FLAVA/CLIP
+            implementation.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        ce_ignore_index (`int`, *optional*, defaults to -100):
+            Cross entropy index to ignore.
+        mim_weight (`float`, *optional*, defaults to 1.0):
+            Weight to be assigned to MIM (Masked Image Modeling) unimodal loss
+        mlm_weight (`float`, *optional*, defaults to 1.0):
+            Weight to be assigned to MLM (Masked Language Modeling) unimodal loss
+        global_contrastive_weight (`float`, *optional*, defaults to 1.0):
+            Weight to be assigned to global contrastive cross-alignment loss.
+        itm_weight (`float`, *optional*, defaults to 1.0):
+            Weight to be assigned to image-text matching multimodal loss.
+        mmm_image_weight (`float`, *optional*, defaults to 1.0):
+            Weight to be assigned to MMM loss's image part.
+        mmm_text_weight (`float`, *optional*, defaults to 1.0):
+            Weight to be assigned to MMM loss's text part.
+        global_backprop_contrastive (`bool`, *optional*, defaults to `True`):
+            Whether to use global backpropgation through all workers in contrastive loss.
+        skip_unmasked_multimodal_encoder (`bool`, *optional*, defaults to `True`):
+            Whether to skip running unmasked multimodal encoder whose outputs are not used by FLAVA losses.
+        return_loss (`bool`, *optional*, defaults to `True`):
+            Whether to return loss or not
+
+        kwargs (*optional*):
+            Dictionary of keyword arguments.
+
+    Example:
+
+    ```python
+    >>> from transformers import FlavaModel, FlavaForPreTraining, FlavaConfig
+
+    >>> # Initializing a FlavaConfig with style configuration
+    >>> configuration = FlavaConfig()
+
+    >>> # Initializing a FlavaModel and FlavaForPreTraining model from the style configuration
+    >>> model = FlavaModel(configuration)
+    >>> model_pre = FlavaForPreTraining(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    >>> configuration_pre = model_pre.config
+    ```
+    """
+
+    model_type = "flava"
+    is_composition = True
+
+    def __init__(
+        self,
+        image_config_dict: Dict[str, Any] = None,
+        text_config_dict: Dict[str, Any] = None,
+        multimodal_config_dict: Dict[str, Any] = None,
+        image_codebook_config_dict: Dict[str, Any] = None,
+        hidden_size: int = 768,
+        layer_norm_eps: float = 1e-12,
+        projection_dim: int = 768,
+        init_codebook: bool = True,
+        logit_scale_init_value: float = 2.6592,
+        initializer_range: float = 0.02,
+        ce_ignore_index: int = -100,
+        mim_weight: float = 1.0,
+        mlm_weight: float = 1.0,
+        global_contrastive_weight: float = 1.0,
+        itm_weight: float = 1.0,
+        mmm_image_weight: float = 1.0,
+        mmm_text_weight: float = 1.0,
+        global_backprop_contrastive: bool = True,
+        skip_unmasked_multimodal_encoder: bool = True,
+        return_loss: bool = True,
+        **kwargs
+    ):
+        super().__init__(**kwargs)
+
+        if image_config_dict is None:
+            image_config_dict = {}
+            logger.info("image_config_dict is None. initializing the FlavaImageConfig with default values.")
+
+        if text_config_dict is None:
+            text_config_dict = {}
+            logger.info("text_config_dict is None. Initializing the FlavaTextConfig with default values.")
+
+        if multimodal_config_dict is None:
+            multimodal_config_dict = {}
+            logger.info("multimodal_config_dict is None. initializing the FlavaMultimodalConfig with default values.")
+
+        if image_codebook_config_dict is None:
+            image_codebook_config_dict = {}
+            logger.info(
+                "image_codebook_config_dict is None. initializing the FlavaImageCodebookConfig with default values."
+            )
+
+        self.image_config_dict = image_config_dict
+        self.text_config_dict = text_config_dict
+        self.multimodal_config_dict = multimodal_config_dict
+        self.image_codebook_config_dict = image_codebook_config_dict
+
+        self.image_config = FlavaImageConfig(**self.image_config_dict)
+        self.text_config = FlavaTextConfig(**self.text_config_dict)
+        self.multimodal_config = FlavaMultimodalConfig(**self.multimodal_config_dict)
+        self.image_codebook_config = FlavaImageCodebookConfig(**self.image_codebook_config_dict)
+        self.projection_dim = projection_dim
+        self.init_codebook = init_codebook
+
+        self.hidden_size = hidden_size
+        self.layer_norm_eps = layer_norm_eps
+        self.initializer_range = initializer_range
+        self.logit_scale_init_value = logit_scale_init_value
+        self.initializer_factor = 1.0
+        self.ce_ignore_index = ce_ignore_index
+        self.mim_weight = mim_weight
+        self.mlm_weight = mlm_weight
+        self.global_contrastive_weight = global_contrastive_weight
+        self.itm_weight = itm_weight
+        self.mmm_image_weight = mmm_image_weight
+        self.mmm_text_weight = mmm_text_weight
+        self.global_backprop_contrastive = global_backprop_contrastive
+        self.skip_unmasked_multimodal_encoder = skip_unmasked_multimodal_encoder
+        self.return_loss = return_loss
+
+    @classmethod
+    def from_configs(
+        cls,
+        image_config: FlavaImageConfig,
+        text_config: FlavaTextConfig,
+        multimodal_config: FlavaMultimodalConfig,
+        image_codebook_config: FlavaImageCodebookConfig,
+        **kwargs
+    ):
+        r"""
+        Instantiate a [`FlavaConfig`] (or a derived class) from flava text model configuration, flava image model
+        configuration, flava multimodal model and flava codebook model configuration.
+
+        Returns:
+            [`FlavaConfig`]: An instance of a configuration object
+        """
+
+        return cls(
+            image_config_dict=image_config.to_dict(),
+            text_config_dict=text_config.to_dict(),
+            multimodal_config_dict=multimodal_config.to_dict(),
+            image_codebook_config_dict=image_codebook_config.to_dict(),
+            **kwargs,
+        )
+
+    def to_dict(self):
+        """
+        Serializes this instance to a Python dictionary. Override the default [`~PretrainedConfig.to_dict`].
+
+        Returns:
+            `Dict[str, any]`: Dictionary of all the attributes that make up this configuration instance,
+        """
+        output = copy.deepcopy(self.__dict__)
+        output["image_config"] = self.image_config.to_dict()
+        output["text_config"] = self.text_config.to_dict()
+        output["multimodal_config"] = self.multimodal_config.to_dict()
+        output["image_codebook_config"] = self.image_codebook_config.to_dict()
+        output["model_type"] = self.__class__.model_type
+        return output

src/transformers/models/flava/convert_dalle_to_flava_codebook.py

@@ -0,0 +1,102 @@
+# 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

@@ -0,0 +1,99 @@
+# 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

@@ -0,0 +1,351 @@
+# 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

@@ -0,0 +1,124 @@
+# 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)

src/transformers/utils/dummy_pt_objects.py

@@ -1787,6 +1787,58 @@ class FlaubertWithLMHeadModel(metaclass=DummyObject):
         requires_backends(self, ["torch"])
 
 
+FLAVA_PRETRAINED_MODEL_ARCHIVE_LIST = None
+
+
+class FlavaForPreTraining(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class FlavaImageCodebook(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class FlavaImageModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class FlavaModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class FlavaMultimodalModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class FlavaPreTrainedModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class FlavaTextModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
 FNET_PRETRAINED_MODEL_ARCHIVE_LIST = None
 
 

src/transformers/utils/dummy_vision_objects.py

@@ -59,6 +59,20 @@ class DPTFeatureExtractor(metaclass=DummyObject):
         requires_backends(self, ["vision"])
 
 
+class FlavaFeatureExtractor(metaclass=DummyObject):
+    _backends = ["vision"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["vision"])
+
+
+class FlavaProcessor(metaclass=DummyObject):
+    _backends = ["vision"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["vision"])
+
+
 class GLPNFeatureExtractor(metaclass=DummyObject):
     _backends = ["vision"]
 

tests/models/flava/test_feature_extraction_flava.py

@@ -0,0 +1,347 @@
+# coding=utf-8
+# Copyright 2022 Meta Platforms authors and 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 random
+import unittest
+
+import numpy as np
+
+from transformers.testing_utils import require_torch, require_vision
+from transformers.utils import is_torch_available, is_vision_available
+
+from ...test_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 FlavaFeatureExtractor
+    from transformers.models.flava.feature_extraction_flava import (
+        FLAVA_CODEBOOK_MEAN,
+        FLAVA_CODEBOOK_STD,
+        FLAVA_IMAGE_MEAN,
+        FLAVA_IMAGE_STD,
+    )
+else:
+    FLAVA_IMAGE_MEAN = FLAVA_IMAGE_STD = FLAVA_CODEBOOK_MEAN = FLAVA_CODEBOOK_STD = None
+
+
+class FlavaFeatureExtractionTester(unittest.TestCase):
+    def __init__(
+        self,
+        parent,
+        batch_size=7,
+        num_channels=3,
+        min_resolution=30,
+        max_resolution=400,
+        do_resize=True,
+        size=224,
+        do_center_crop=True,
+        crop_size=224,
+        resample=None,
+        do_normalize=True,
+        image_mean=FLAVA_IMAGE_MEAN,
+        image_std=FLAVA_IMAGE_STD,
+        input_size_patches=14,
+        total_mask_patches=75,
+        mask_group_max_patches=None,
+        mask_group_min_patches=16,
+        mask_group_min_aspect_ratio=0.3,
+        mask_group_max_aspect_ratio=None,
+        codebook_do_resize=True,
+        codebook_size=112,
+        codebook_resample=None,
+        codebook_do_center_crop=True,
+        codebook_crop_size=112,
+        codebook_do_map_pixels=True,
+        codebook_do_normalize=True,
+        codebook_image_mean=FLAVA_CODEBOOK_MEAN,
+        codebook_image_std=FLAVA_CODEBOOK_STD,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.num_channels = num_channels
+        self.do_resize = do_resize
+        self.min_resolution = min_resolution
+        self.max_resolution = max_resolution
+        self.size = size
+        self.resample = resample if resample is not None else Image.BICUBIC
+        self.do_normalize = do_normalize
+        self.image_mean = image_mean
+        self.image_std = image_std
+        self.do_center_crop = do_center_crop
+        self.crop_size = crop_size
+
+        self.input_size_patches = input_size_patches
+        self.total_mask_patches = total_mask_patches
+        self.mask_group_max_patches = mask_group_max_patches
+        self.mask_group_min_patches = mask_group_min_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 if codebook_resample is not None else Image.LANCZOS
+        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
+
+    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,
+            "resample": self.resample,
+            "do_center_crop": self.do_center_crop,
+            "crop_size": self.crop_size,
+            "input_size_patches": self.input_size_patches,
+            "total_mask_patches": self.total_mask_patches,
+            "mask_group_max_patches": self.mask_group_max_patches,
+            "mask_group_min_patches": self.mask_group_min_patches,
+            "mask_group_min_aspect_ratio": self.mask_group_min_aspect_ratio,
+            "mask_group_max_aspect_ratio": self.mask_group_min_aspect_ratio,
+            "codebook_do_resize": self.codebook_do_resize,
+            "codebook_size": self.codebook_size,
+            "codebook_resample": self.codebook_resample,
+            "codebook_do_center_crop": self.codebook_do_center_crop,
+            "codebook_crop_size": self.codebook_crop_size,
+            "codebook_do_map_pixels": self.codebook_do_map_pixels,
+            "codebook_do_normalize": self.codebook_do_normalize,
+            "codebook_image_mean": self.codebook_image_mean,
+            "codebook_image_std": self.codebook_image_std,
+        }
+
+    def get_expected_image_size(self):
+        return (self.size, self.size) if not isinstance(self.size, tuple) else self.size
+
+    def get_expected_mask_size(self):
+        return (
+            (self.input_size_patches, self.input_size_patches)
+            if not isinstance(self.input_size_patches, tuple)
+            else self.input_size_patches
+        )
+
+    def get_expected_codebook_image_size(self):
+        if not isinstance(self.codebook_size, tuple):
+            return (self.codebook_size, self.codebook_size)
+        else:
+            return self.codebook_size
+
+
+@require_torch
+@require_vision
+class FlavaFeatureExtractionTest(FeatureExtractionSavingTestMixin, unittest.TestCase):
+
+    feature_extraction_class = FlavaFeatureExtractor if is_vision_available() else None
+    maxDiff = None
+
+    def setUp(self):
+        self.feature_extract_tester = FlavaFeatureExtractionTester(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, "resample"))
+        self.assertTrue(hasattr(feature_extractor, "crop_size"))
+        self.assertTrue(hasattr(feature_extractor, "do_center_crop"))
+        self.assertTrue(hasattr(feature_extractor, "masking_generator"))
+        self.assertTrue(hasattr(feature_extractor, "codebook_do_resize"))
+        self.assertTrue(hasattr(feature_extractor, "codebook_size"))
+        self.assertTrue(hasattr(feature_extractor, "codebook_resample"))
+        self.assertTrue(hasattr(feature_extractor, "codebook_do_center_crop"))
+        self.assertTrue(hasattr(feature_extractor, "codebook_crop_size"))
+        self.assertTrue(hasattr(feature_extractor, "codebook_do_map_pixels"))
+        self.assertTrue(hasattr(feature_extractor, "codebook_do_normalize"))
+        self.assertTrue(hasattr(feature_extractor, "codebook_image_mean"))
+        self.assertTrue(hasattr(feature_extractor, "codebook_image_std"))
+
+    def test_batch_feature(self):
+        pass
+
+    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")
+
+        # Test no bool masked pos
+        self.assertFalse("bool_masked_pos" in encoded_images)
+
+        expected_height, expected_width = self.feature_extract_tester.get_expected_image_size()
+
+        self.assertEqual(
+            encoded_images.pixel_values.shape,
+            (1, self.feature_extract_tester.num_channels, expected_height, expected_width),
+        )
+
+        # Test batched
+        encoded_images = feature_extractor(image_inputs, return_tensors="pt")
+        expected_height, expected_width = self.feature_extract_tester.get_expected_image_size()
+
+        # Test no bool masked pos
+        self.assertFalse("bool_masked_pos" in encoded_images)
+
+        self.assertEqual(
+            encoded_images.pixel_values.shape,
+            (
+                self.feature_extract_tester.batch_size,
+                self.feature_extract_tester.num_channels,
+                expected_height,
+                expected_width,
+            ),
+        )
+
+    def _test_call_framework(self, instance_class, prepare_kwargs):
+        # Initialize feature_extractor
+        feature_extractor = self.feature_extraction_class(**self.feat_extract_dict)
+        # create random tensors
+        image_inputs = prepare_image_inputs(self.feature_extract_tester, equal_resolution=False, **prepare_kwargs)
+        for image in image_inputs:
+            self.assertIsInstance(image, instance_class)
+
+        # Test not batched input
+        encoded_images = feature_extractor(image_inputs[0], return_tensors="pt")
+
+        expected_height, expected_width = self.feature_extract_tester.get_expected_image_size()
+        self.assertEqual(
+            encoded_images.pixel_values.shape,
+            (1, self.feature_extract_tester.num_channels, expected_height, expected_width),
+        )
+
+        encoded_images = feature_extractor(image_inputs, return_image_mask=True, return_tensors="pt")
+
+        expected_height, expected_width = self.feature_extract_tester.get_expected_image_size()
+        self.assertEqual(
+            encoded_images.pixel_values.shape,
+            (
+                self.feature_extract_tester.batch_size,
+                self.feature_extract_tester.num_channels,
+                expected_height,
+                expected_width,
+            ),
+        )
+
+        expected_height, expected_width = self.feature_extract_tester.get_expected_mask_size()
+        self.assertEqual(
+            encoded_images.bool_masked_pos.shape,
+            (
+                self.feature_extract_tester.batch_size,
+                expected_height,
+                expected_width,
+            ),
+        )
+
+        # Test batched
+        encoded_images = feature_extractor(image_inputs, return_tensors="pt").pixel_values
+
+        expected_height, expected_width = self.feature_extract_tester.get_expected_image_size()
+        self.assertEqual(
+            encoded_images.shape,
+            (
+                self.feature_extract_tester.batch_size,
+                self.feature_extract_tester.num_channels,
+                expected_height,
+                expected_width,
+            ),
+        )
+
+        # Test masking
+        encoded_images = feature_extractor(image_inputs, return_image_mask=True, return_tensors="pt")
+
+        expected_height, expected_width = self.feature_extract_tester.get_expected_image_size()
+        self.assertEqual(
+            encoded_images.pixel_values.shape,
+            (
+                self.feature_extract_tester.batch_size,
+                self.feature_extract_tester.num_channels,
+                expected_height,
+                expected_width,
+            ),
+        )
+
+        expected_height, expected_width = self.feature_extract_tester.get_expected_mask_size()
+        self.assertEqual(
+            encoded_images.bool_masked_pos.shape,
+            (
+                self.feature_extract_tester.batch_size,
+                expected_height,
+                expected_width,
+            ),
+        )
+
+    def test_call_numpy(self):
+        self._test_call_framework(np.ndarray, prepare_kwargs={"numpify": True})
+
+    def test_call_pytorch(self):
+        self._test_call_framework(torch.Tensor, prepare_kwargs={"torchify": True})
+
+    def test_masking(self):
+        # Initialize feature_extractor
+        random.seed(1234)
+        feature_extractor = self.feature_extraction_class(**self.feat_extract_dict)
+        image_inputs = prepare_image_inputs(self.feature_extract_tester, equal_resolution=False, torchify=True)
+
+        # Test not batched input
+        encoded_images = feature_extractor(image_inputs[0], return_image_mask=True, return_tensors="pt")
+        self.assertEqual(encoded_images.bool_masked_pos.sum().item(), 75)
+
+    def test_codebook_pixels(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_codebook_pixels=True, return_tensors="pt")
+        expected_height, expected_width = self.feature_extract_tester.get_expected_codebook_image_size()
+        self.assertEqual(
+            encoded_images.codebook_pixel_values.shape,
+            (1, self.feature_extract_tester.num_channels, expected_height, expected_width),
+        )
+
+        # Test batched
+        encoded_images = feature_extractor(image_inputs, return_codebook_pixels=True, return_tensors="pt")
+        expected_height, expected_width = self.feature_extract_tester.get_expected_codebook_image_size()
+        self.assertEqual(
+            encoded_images.codebook_pixel_values.shape,
+            (
+                self.feature_extract_tester.batch_size,
+                self.feature_extract_tester.num_channels,
+                expected_height,
+                expected_width,
+            ),
+        )

tests/models/flava/test_processor_flava.py

@@ -0,0 +1,234 @@
+# 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 json
+import os
+import random
+import shutil
+import tempfile
+import unittest
+
+import numpy as np
+import pytest
+
+from transformers import BertTokenizer, BertTokenizerFast
+from transformers.models.bert.tokenization_bert import VOCAB_FILES_NAMES
+from transformers.testing_utils import require_vision
+from transformers.utils import FEATURE_EXTRACTOR_NAME, is_vision_available
+
+
+if is_vision_available():
+    from PIL import Image
+
+    from transformers import FlavaFeatureExtractor, FlavaProcessor
+    from transformers.models.flava.feature_extraction_flava import (
+        FLAVA_CODEBOOK_MEAN,
+        FLAVA_CODEBOOK_STD,
+        FLAVA_IMAGE_MEAN,
+        FLAVA_IMAGE_STD,
+    )
+
+
+@require_vision
+class FlavaProcessorTest(unittest.TestCase):
+    def setUp(self):
+        self.tmpdirname = tempfile.mkdtemp()
+
+        # fmt: off
+        vocab_tokens = ["[UNK]", "[CLS]", "[SEP]", "[PAD]", "[MASK]", "want", "##want", "##ed", "wa", "un", "runn", "##ing", ",", "low", "lowest"]
+        # fmt: on
+        self.vocab_file = os.path.join(self.tmpdirname, VOCAB_FILES_NAMES["vocab_file"])
+
+        with open(self.vocab_file, "w", encoding="utf-8") as fp:
+            fp.write("".join([x + "\n" for x in vocab_tokens]))
+
+        feature_extractor_map = {
+            "image_mean": FLAVA_IMAGE_MEAN,
+            "image_std": FLAVA_IMAGE_STD,
+            "do_normalize": True,
+            "do_resize": True,
+            "size": 224,
+            "do_center_crop": True,
+            "crop_size": 224,
+            "input_size_patches": 14,
+            "total_mask_patches": 75,
+            "mask_group_max_patches": None,
+            "mask_group_min_patches": 16,
+            "mask_group_min_aspect_ratio": 0.3,
+            "mask_group_max_aspect_ratio": None,
+            "codebook_do_resize": True,
+            "codebook_size": 112,
+            "codebook_resample": None,
+            "codebook_do_center_crop": True,
+            "codebook_crop_size": 112,
+            "codebook_do_map_pixels": True,
+            "codebook_do_normalize": True,
+            "codebook_image_mean": FLAVA_CODEBOOK_MEAN,
+            "codebook_image_std": FLAVA_CODEBOOK_STD,
+        }
+
+        self.feature_extractor_file = os.path.join(self.tmpdirname, FEATURE_EXTRACTOR_NAME)
+        with open(self.feature_extractor_file, "w", encoding="utf-8") as fp:
+            json.dump(feature_extractor_map, fp)
+
+    def get_tokenizer(self, **kwargs):
+        return BertTokenizer.from_pretrained(self.tmpdirname, **kwargs)
+
+    def get_rust_tokenizer(self, **kwargs):
+        return BertTokenizerFast.from_pretrained(self.tmpdirname, **kwargs)
+
+    def get_feature_extractor(self, **kwargs):
+        return FlavaFeatureExtractor.from_pretrained(self.tmpdirname, **kwargs)
+
+    def tearDown(self):
+        shutil.rmtree(self.tmpdirname)
+
+    def prepare_image_inputs(self):
+        """This function prepares a list of PIL images, or a list of numpy arrays if one specifies numpify=True,
+        or a list of PyTorch tensors if one specifies torchify=True.
+        """
+
+        image_inputs = [np.random.randint(255, size=(3, 30, 400), dtype=np.uint8)]
+
+        image_inputs = [Image.fromarray(np.moveaxis(x, 0, -1)) for x in image_inputs]
+
+        return image_inputs
+
+    def test_save_load_pretrained_default(self):
+        tokenizer_slow = self.get_tokenizer()
+        tokenizer_fast = self.get_rust_tokenizer()
+        feature_extractor = self.get_feature_extractor()
+
+        processor_slow = FlavaProcessor(tokenizer=tokenizer_slow, feature_extractor=feature_extractor)
+        processor_slow.save_pretrained(self.tmpdirname)
+        processor_slow = FlavaProcessor.from_pretrained(self.tmpdirname, use_fast=False)
+
+        processor_fast = FlavaProcessor(tokenizer=tokenizer_fast, feature_extractor=feature_extractor)
+        processor_fast.save_pretrained(self.tmpdirname)
+        processor_fast = FlavaProcessor.from_pretrained(self.tmpdirname)
+
+        self.assertEqual(processor_slow.tokenizer.get_vocab(), tokenizer_slow.get_vocab())
+        self.assertEqual(processor_fast.tokenizer.get_vocab(), tokenizer_fast.get_vocab())
+        self.assertEqual(tokenizer_slow.get_vocab(), tokenizer_fast.get_vocab())
+        self.assertIsInstance(processor_slow.tokenizer, BertTokenizer)
+        self.assertIsInstance(processor_fast.tokenizer, BertTokenizerFast)
+
+        self.assertEqual(processor_slow.feature_extractor.to_json_string(), feature_extractor.to_json_string())
+        self.assertEqual(processor_fast.feature_extractor.to_json_string(), feature_extractor.to_json_string())
+        self.assertIsInstance(processor_slow.feature_extractor, FlavaFeatureExtractor)
+        self.assertIsInstance(processor_fast.feature_extractor, FlavaFeatureExtractor)
+
+    def test_save_load_pretrained_additional_features(self):
+        processor = FlavaProcessor(tokenizer=self.get_tokenizer(), feature_extractor=self.get_feature_extractor())
+        processor.save_pretrained(self.tmpdirname)
+
+        tokenizer_add_kwargs = self.get_tokenizer(bos_token="(BOS)", eos_token="(EOS)")
+        feature_extractor_add_kwargs = self.get_feature_extractor(do_normalize=False, padding_value=1.0)
+
+        processor = FlavaProcessor.from_pretrained(
+            self.tmpdirname, bos_token="(BOS)", eos_token="(EOS)", do_normalize=False, padding_value=1.0
+        )
+
+        self.assertEqual(processor.tokenizer.get_vocab(), tokenizer_add_kwargs.get_vocab())
+        self.assertIsInstance(processor.tokenizer, BertTokenizerFast)
+
+        self.assertEqual(processor.feature_extractor.to_json_string(), feature_extractor_add_kwargs.to_json_string())
+        self.assertIsInstance(processor.feature_extractor, FlavaFeatureExtractor)
+
+    def test_feature_extractor(self):
+        feature_extractor = self.get_feature_extractor()
+        tokenizer = self.get_tokenizer()
+
+        processor = FlavaProcessor(tokenizer=tokenizer, feature_extractor=feature_extractor)
+
+        image_input = self.prepare_image_inputs()
+
+        input_feat_extract = feature_extractor(image_input, return_tensors="np")
+        input_processor = processor(images=image_input, return_tensors="np")
+
+        for key in input_feat_extract.keys():
+            self.assertAlmostEqual(input_feat_extract[key].sum(), input_processor[key].sum(), delta=1e-2)
+
+        # With rest of the args
+        random.seed(1234)
+        input_feat_extract = feature_extractor(
+            image_input, return_image_mask=True, return_codebook_pixels=True, return_tensors="np"
+        )
+        random.seed(1234)
+        input_processor = processor(
+            images=image_input, return_image_mask=True, return_codebook_pixels=True, return_tensors="np"
+        )
+
+        for key in input_feat_extract.keys():
+            self.assertAlmostEqual(input_feat_extract[key].sum(), input_processor[key].sum(), delta=1e-2)
+
+    def test_tokenizer(self):
+        feature_extractor = self.get_feature_extractor()
+        tokenizer = self.get_tokenizer()
+
+        processor = FlavaProcessor(tokenizer=tokenizer, feature_extractor=feature_extractor)
+
+        input_str = "lower newer"
+
+        encoded_processor = processor(text=input_str)
+
+        encoded_tok = tokenizer(input_str)
+
+        for key in encoded_tok.keys():
+            self.assertListEqual(encoded_tok[key], encoded_processor[key])
+
+    def test_processor(self):
+        feature_extractor = self.get_feature_extractor()
+        tokenizer = self.get_tokenizer()
+
+        processor = FlavaProcessor(tokenizer=tokenizer, feature_extractor=feature_extractor)
+
+        input_str = "lower newer"
+        image_input = self.prepare_image_inputs()
+
+        inputs = processor(text=input_str, images=image_input)
+
+        self.assertListEqual(list(inputs.keys()), ["input_ids", "token_type_ids", "attention_mask", "pixel_values"])
+
+        # add extra args
+        inputs = processor(text=input_str, images=image_input, return_codebook_pixels=True, return_image_mask=True)
+
+        self.assertListEqual(
+            list(inputs.keys()),
+            [
+                "input_ids",
+                "token_type_ids",
+                "attention_mask",
+                "pixel_values",
+                "codebook_pixel_values",
+                "bool_masked_pos",
+            ],
+        )
+
+        # test if it raises when no input is passed
+        with pytest.raises(ValueError):
+            processor()
+
+    def test_tokenizer_decode(self):
+        feature_extractor = self.get_feature_extractor()
+        tokenizer = self.get_tokenizer()
+
+        processor = FlavaProcessor(tokenizer=tokenizer, feature_extractor=feature_extractor)
+
+        predicted_ids = [[1, 4, 5, 8, 1, 0, 8], [3, 4, 3, 1, 1, 8, 9]]
+
+        decoded_processor = processor.batch_decode(predicted_ids)
+        decoded_tok = tokenizer.batch_decode(predicted_ids)
+
+        self.assertListEqual(decoded_tok, decoded_processor)

utils/check_repo.py

@@ -146,6 +146,10 @@ IGNORE_NON_AUTO_CONFIGURED = PRIVATE_MODELS.copy() + [
     "DetrForSegmentation",
     "DPRReader",
     "FlaubertForQuestionAnswering",
+    "FlavaImageCodebook",
+    "FlavaTextModel",
+    "FlavaImageModel",
+    "FlavaMultimodalModel",
     "GPT2DoubleHeadsModel",
     "LukeForMaskedLM",
     "LukeForEntityClassification",