Add ALIGN to transformers (#21741)

Adds the ALIGN model to transformers. ALIGN is introduced in "Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision" by Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, Tom Duerig.

README.md

@@ -271,6 +271,7 @@ Current number of checkpoints: ![](https://img.shields.io/endpoint?url=https://h
 🤗 Transformers currently provides the following architectures (see [here](https://huggingface.co/docs/transformers/model_summary) for a high-level summary of each them):
 
 1. **[ALBERT](https://huggingface.co/docs/transformers/model_doc/albert)** (from Google Research and the Toyota Technological Institute at Chicago) released with the paper [ALBERT: A Lite BERT for Self-supervised Learning of Language Representations](https://arxiv.org/abs/1909.11942), by Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, Radu Soricut.
+1. **[ALIGN](https://huggingface.co/docs/transformers/main/model_doc/align)** (from Google Research) released with the paper [Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision](https://arxiv.org/abs/2102.05918) by Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, Tom Duerig.
 1. **[AltCLIP](https://huggingface.co/docs/transformers/model_doc/altclip)** (from BAAI) released with the paper [AltCLIP: Altering the Language Encoder in CLIP for Extended Language Capabilities](https://arxiv.org/abs/2211.06679) by Chen, Zhongzhi and Liu, Guang and Zhang, Bo-Wen and Ye, Fulong and Yang, Qinghong and Wu, Ledell.
 1. **[Audio Spectrogram Transformer](https://huggingface.co/docs/transformers/model_doc/audio-spectrogram-transformer)** (from MIT) released with the paper [AST: Audio Spectrogram Transformer](https://arxiv.org/abs/2104.01778) by Yuan Gong, Yu-An Chung, James Glass.
 1. **[BART](https://huggingface.co/docs/transformers/model_doc/bart)** (from Facebook) released with the paper [BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension](https://arxiv.org/abs/1910.13461) by Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov and Luke Zettlemoyer.

README_es.md

@@ -264,6 +264,7 @@ Número actual de puntos de control: ![](https://img.shields.io/endpoint?url=htt
 🤗 Transformers actualmente proporciona las siguientes arquitecturas (ver [aquí](https://huggingface.co/docs/transformers/model_summary) para un resumen de alto nivel de cada uno de ellas.):
 
 1. **[ALBERT](https://huggingface.co/docs/transformers/model_doc/albert)** (from Google Research and the Toyota Technological Institute at Chicago) released with the paper [ALBERT: A Lite BERT for Self-supervised Learning of Language Representations](https://arxiv.org/abs/1909.11942), by Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, Radu Soricut.
+1. **[ALIGN](https://huggingface.co/docs/transformers/main/model_doc/align)** (from Google Research) released with the paper [Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision](https://arxiv.org/abs/2102.05918) by Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, Tom Duerig.
 1. **[AltCLIP](https://huggingface.co/docs/transformers/model_doc/altclip)** (from BAAI) released with the paper [AltCLIP: Altering the Language Encoder in CLIP for Extended Language Capabilities](https://arxiv.org/abs/2211.06679) by Chen, Zhongzhi and Liu, Guang and Zhang, Bo-Wen and Ye, Fulong and Yang, Qinghong and Wu, Ledell.
 1. **[Audio Spectrogram Transformer](https://huggingface.co/docs/transformers/model_doc/audio-spectrogram-transformer)** (from MIT) released with the paper [AST: Audio Spectrogram Transformer](https://arxiv.org/abs/2104.01778) by Yuan Gong, Yu-An Chung, James Glass.
 1. **[BART](https://huggingface.co/docs/transformers/model_doc/bart)** (from Facebook) released with the paper [BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension](https://arxiv.org/abs/1910.13461) by Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov and Luke Zettlemoyer.

README_hd.md

@@ -236,6 +236,7 @@ conda install -c huggingface transformers
 🤗 ट्रांसफॉर्मर वर्तमान में निम्नलिखित आर्किटेक्चर का समर्थन करते हैं (मॉडल के अवलोकन के लिए [यहां] देखें (https://huggingface.co/docs/transformers/model_summary))：
 
 1. **[ALBERT](https://huggingface.co/docs/transformers/model_doc/albert)** (Google Research and the Toyota Technological Institute at Chicago) साथ थीसिस [ALBERT: A Lite BERT for Self-supervised भाषा प्रतिनिधित्व सीखना](https://arxiv.org/abs/1909.11942), झेंझोंग लैन, मिंगदा चेन, सेबेस्टियन गुडमैन, केविन गिम्पेल, पीयूष शर्मा, राडू सोरिकट
+1. **[ALIGN](https://huggingface.co/docs/transformers/main/model_doc/align)** (Google Research से) Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, Tom Duerig. द्वाराअनुसंधान पत्र [Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision](https://arxiv.org/abs/2102.05918) के साथ जारी किया गया
 1. **[AltCLIP](https://huggingface.co/docs/transformers/model_doc/altclip)** (from BAAI) released with the paper [AltCLIP: Altering the Language Encoder in CLIP for Extended Language Capabilities](https://arxiv.org/abs/2211.06679) by Chen, Zhongzhi and Liu, Guang and Zhang, Bo-Wen and Ye, Fulong and Yang, Qinghong and Wu, Ledell.
 1. **[Audio Spectrogram Transformer](https://huggingface.co/docs/transformers/model_doc/audio-spectrogram-transformer)** (from MIT) released with the paper [AST: Audio Spectrogram Transformer](https://arxiv.org/abs/2104.01778) by Yuan Gong, Yu-An Chung, James Glass.
 1. **[BART](https://huggingface.co/docs/transformers/model_doc/bart)** (फेसबुक) साथ थीसिस [बार्ट: प्राकृतिक भाषा निर्माण, अनुवाद के लिए अनुक्रम-से-अनुक्रम पूर्व प्रशिक्षण , और समझ] (https://arxiv.org/pdf/1910.13461.pdf) पर निर्भर माइक लुईस, यिनहान लियू, नमन गोयल, मार्जन ग़ज़विनिनेजाद, अब्देलरहमान मोहम्मद, ओमर लेवी, वेस स्टोयानोव और ल्यूक ज़ेटलमॉयर

README_ja.md

@@ -298,6 +298,7 @@ Flax、PyTorch、TensorFlowをcondaでインストールする方法は、それ
 🤗Transformersは現在、以下のアーキテクチャを提供しています（それぞれのハイレベルな要約は[こちら](https://huggingface.co/docs/transformers/model_summary)を参照してください）:
 
 1. **[ALBERT](https://huggingface.co/docs/transformers/model_doc/albert)** (Google Research and the Toyota Technological Institute at Chicago から) Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, Radu Soricut から公開された研究論文: [ALBERT: A Lite BERT for Self-supervised Learning of Language Representations](https://arxiv.org/abs/1909.11942)
+1. **[ALIGN](https://huggingface.co/docs/transformers/main/model_doc/align)** (Google Research から) Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, Tom Duerig. から公開された研究論文 [Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision](https://arxiv.org/abs/2102.05918)
 1. **[AltCLIP](https://huggingface.co/docs/transformers/model_doc/altclip)** (BAAI から) Chen, Zhongzhi and Liu, Guang and Zhang, Bo-Wen and Ye, Fulong and Yang, Qinghong and Wu, Ledell から公開された研究論文: [AltCLIP: Altering the Language Encoder in CLIP for Extended Language Capabilities](https://arxiv.org/abs/2211.06679)
 1. **[Audio Spectrogram Transformer](https://huggingface.co/docs/transformers/model_doc/audio-spectrogram-transformer)** (MIT から) Yuan Gong, Yu-An Chung, James Glass から公開された研究論文: [AST: Audio Spectrogram Transformer](https://arxiv.org/abs/2104.01778)
 1. **[BART](https://huggingface.co/docs/transformers/model_doc/bart)** (Facebook から) Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov and Luke Zettlemoyer から公開された研究論文: [BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension](https://arxiv.org/abs/1910.13461)

README_ko.md

@@ -213,6 +213,7 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
 🤗 Transformers는 다음 모델들을 제공합니다 (각 모델의 요약은 [여기](https://huggingface.co/docs/transformers/model_summary)서 확인하세요):
 
 1. **[ALBERT](https://huggingface.co/docs/transformers/model_doc/albert)** (from Google Research and the Toyota Technological Institute at Chicago) released with the paper [ALBERT: A Lite BERT for Self-supervised Learning of Language Representations](https://arxiv.org/abs/1909.11942), by Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, Radu Soricut.
+1. **[ALIGN](https://huggingface.co/docs/transformers/main/model_doc/align)** (Google Research 에서 제공)은 Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, Tom Duerig.의 [Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision](https://arxiv.org/abs/2102.05918)논문과 함께 발표했습니다.
 1. **[AltCLIP](https://huggingface.co/docs/transformers/model_doc/altclip)** (from BAAI) released with the paper [AltCLIP: Altering the Language Encoder in CLIP for Extended Language Capabilities](https://arxiv.org/abs/2211.06679) by Chen, Zhongzhi and Liu, Guang and Zhang, Bo-Wen and Ye, Fulong and Yang, Qinghong and Wu, Ledell.
 1. **[Audio Spectrogram Transformer](https://huggingface.co/docs/transformers/model_doc/audio-spectrogram-transformer)** (from MIT) released with the paper [AST: Audio Spectrogram Transformer](https://arxiv.org/abs/2104.01778) by Yuan Gong, Yu-An Chung, James Glass.
 1. **[BART](https://huggingface.co/docs/transformers/model_doc/bart)** (from Facebook) released with the paper [BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension](https://arxiv.org/pdf/1910.13461.pdf) by Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov and Luke Zettlemoyer.

README_zh-hans.md

@@ -237,6 +237,7 @@ conda install -c huggingface transformers
 🤗 Transformers 目前支持如下的架构（模型概述请阅[这里](https://huggingface.co/docs/transformers/model_summary)）：
 
 1. **[ALBERT](https://huggingface.co/docs/transformers/model_doc/albert)** (来自 Google Research and the Toyota Technological Institute at Chicago) 伴随论文 [ALBERT: A Lite BERT for Self-supervised Learning of Language Representations](https://arxiv.org/abs/1909.11942), 由 Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, Radu Soricut 发布。
+1. **[ALIGN](https://huggingface.co/docs/transformers/main/model_doc/align)** (来自 Google Research) 伴随论文 [Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision](https://arxiv.org/abs/2102.05918) 由 Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, Tom Duerig 发布。
 1. **[AltCLIP](https://huggingface.co/docs/transformers/model_doc/altclip)** (来自 BAAI) 伴随论文 [AltCLIP: Altering the Language Encoder in CLIP for Extended Language Capabilities](https://arxiv.org/abs/2211.06679) 由 Chen, Zhongzhi and Liu, Guang and Zhang, Bo-Wen and Ye, Fulong and Yang, Qinghong and Wu, Ledell 发布。
 1. **[Audio Spectrogram Transformer](https://huggingface.co/docs/transformers/model_doc/audio-spectrogram-transformer)** (来自 MIT) 伴随论文 [AST: Audio Spectrogram Transformer](https://arxiv.org/abs/2104.01778) 由 Yuan Gong, Yu-An Chung, James Glass 发布。
 1. **[BART](https://huggingface.co/docs/transformers/model_doc/bart)** (来自 Facebook) 伴随论文 [BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension](https://arxiv.org/pdf/1910.13461.pdf) 由 Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov and Luke Zettlemoyer 发布。

README_zh-hant.md

@@ -249,6 +249,7 @@ conda install -c huggingface transformers
 🤗 Transformers 目前支援以下的架構（模型概覽請參閱[這裡](https://huggingface.co/docs/transformers/model_summary)）：
 
 1. **[ALBERT](https://huggingface.co/docs/transformers/model_doc/albert)** (from Google Research and the Toyota Technological Institute at Chicago) released with the paper [ALBERT: A Lite BERT for Self-supervised Learning of Language Representations](https://arxiv.org/abs/1909.11942), by Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, Radu Soricut.
+1. **[ALIGN](https://huggingface.co/docs/transformers/main/model_doc/align)** (from Google Research) released with the paper [Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision](https://arxiv.org/abs/2102.05918) by Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, Tom Duerig.
 1. **[AltCLIP](https://huggingface.co/docs/transformers/model_doc/altclip)** (from BAAI) released with the paper [AltCLIP: Altering the Language Encoder in CLIP for Extended Language Capabilities](https://arxiv.org/abs/2211.06679) by Chen, Zhongzhi and Liu, Guang and Zhang, Bo-Wen and Ye, Fulong and Yang, Qinghong and Wu, Ledell.
 1. **[Audio Spectrogram Transformer](https://huggingface.co/docs/transformers/model_doc/audio-spectrogram-transformer)** (from MIT) released with the paper [AST: Audio Spectrogram Transformer](https://arxiv.org/abs/2104.01778) by Yuan Gong, Yu-An Chung, James Glass.
 1. **[BART](https://huggingface.co/docs/transformers/model_doc/bart)** (from Facebook) released with the paper [BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension](https://arxiv.org/pdf/1910.13461.pdf) by Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov and Luke Zettlemoyer.

docs/source/de/index.mdx

@@ -52,6 +52,7 @@ Die Bibliothek enthält derzeit JAX-, PyTorch- und TensorFlow-Implementierungen,
 <!--This list is updated automatically from the README with _make fix-copies_. Do not update manually! -->
 
 1. **[ALBERT](model_doc/albert)** (from Google Research and the Toyota Technological Institute at Chicago) released with the paper [ALBERT: A Lite BERT for Self-supervised Learning of Language Representations](https://arxiv.org/abs/1909.11942), by Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, Radu Soricut.
+1. **[ALIGN](model_doc/align)** (from Google Research) released with the paper [Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision](https://arxiv.org/abs/2102.05918) by Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, Tom Duerig.
 1. **[BART](model_doc/bart)** (from Facebook) released with the paper [BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension](https://arxiv.org/abs/1910.13461) by Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov and Luke Zettlemoyer.
 1. **[BARThez](model_doc/barthez)** (from École polytechnique) released with the paper [BARThez: a Skilled Pretrained French Sequence-to-Sequence Model](https://arxiv.org/abs/2010.12321) by Moussa Kamal Eddine, Antoine J.-P. Tixier, Michalis Vazirgiannis.
 1. **[BARTpho](model_doc/bartpho)** (from VinAI Research) released with the paper [BARTpho: Pre-trained Sequence-to-Sequence Models for Vietnamese](https://arxiv.org/abs/2109.09701) by Nguyen Luong Tran, Duong Minh Le and Dat Quoc Nguyen.

docs/source/en/_toctree.yml

@@ -540,6 +540,8 @@
       title: Audio models
     - isExpanded: false
       sections:
+      - local: model_doc/align
+        title: ALIGN
       - local: model_doc/altclip
         title: AltCLIP
       - local: model_doc/blip

docs/source/en/index.mdx

@@ -50,6 +50,7 @@ The documentation is organized into five sections:
 <!--This list is updated automatically from the README with _make fix-copies_. Do not update manually! -->
 
 1. **[ALBERT](model_doc/albert)** (from Google Research and the Toyota Technological Institute at Chicago) released with the paper [ALBERT: A Lite BERT for Self-supervised Learning of Language Representations](https://arxiv.org/abs/1909.11942), by Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, Radu Soricut.
+1. **[ALIGN](model_doc/align)** (from Google Research) released with the paper [Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision](https://arxiv.org/abs/2102.05918) by Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, Tom Duerig.
 1. **[AltCLIP](model_doc/altclip)** (from BAAI) released with the paper [AltCLIP: Altering the Language Encoder in CLIP for Extended Language Capabilities](https://arxiv.org/abs/2211.06679) by Chen, Zhongzhi and Liu, Guang and Zhang, Bo-Wen and Ye, Fulong and Yang, Qinghong and Wu, Ledell.
 1. **[Audio Spectrogram Transformer](model_doc/audio-spectrogram-transformer)** (from MIT) released with the paper [AST: Audio Spectrogram Transformer](https://arxiv.org/abs/2104.01778) by Yuan Gong, Yu-An Chung, James Glass.
 1. **[BART](model_doc/bart)** (from Facebook) released with the paper [BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension](https://arxiv.org/abs/1910.13461) by Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov and Luke Zettlemoyer.
@@ -247,6 +248,7 @@ Flax), PyTorch, and/or TensorFlow.
 |             Model             | Tokenizer slow | Tokenizer fast | PyTorch support | TensorFlow support | Flax Support |
 |:-----------------------------:|:--------------:|:--------------:|:---------------:|:------------------:|:------------:|
 |            ALBERT             |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
+|             ALIGN             |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |            AltCLIP            |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 | Audio Spectrogram Transformer |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |             BART              |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |

docs/source/en/model_doc/align.mdx

+<!--Copyright 2023 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.
+-->
+
+# ALIGN
+
+## Overview
+
+The ALIGN model was proposed in [Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision](https://arxiv.org/abs/2102.05918) by Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, Tom Duerig. ALIGN features a dual-encoder architecture with [EfficientNet](efficientnet) as its vision encoder and [BERT](bert) as its text encoder, and learns to align visual and text representations with contrastive learning. Unlike previous work, ALIGN leverages a massive noisy dataset and shows that the scale of the corpus can be used to achieve SOTA representations with a simple recipe.
+
+The abstract from the paper is the following:
+
+*Pre-trained representations are becoming crucial for many NLP and perception tasks. While representation learning in NLP has transitioned to training on raw text without human annotations, visual and vision-language representations still rely heavily on curated training datasets that are expensive or require expert knowledge. For vision applications, representations are mostly learned using datasets with explicit class labels such as ImageNet or OpenImages. For vision-language, popular datasets like Conceptual Captions, MSCOCO, or CLIP all involve a non-trivial data collection (and cleaning) process. This costly curation process limits the size of datasets and hence hinders the scaling of trained models. In this paper, we leverage a noisy dataset of over one billion image alt-text pairs, obtained without expensive filtering or post-processing steps in the Conceptual Captions dataset. A simple dual-encoder architecture learns to align visual and language representations of the image and text pairs using a contrastive loss. We show that the scale of our corpus can make up for its noise and leads to state-of-the-art representations even with such a simple learning scheme. Our visual representation achieves strong performance when transferred to classification tasks such as ImageNet and VTAB. The aligned visual and language representations enables zero-shot image classification and also set new state-of-the-art results on Flickr30K and MSCOCO image-text retrieval benchmarks, even when compared with more sophisticated cross-attention models. The representations also enable cross-modality search with complex text and text + image queries.*
+
+This model was contributed by [Alara Dirik](https://huggingface.co/adirik).
+The original code is not released, this implementation is based on the Kakao Brain implementation based on the original paper.
+
+
+## AlignConfig
+
+[[autodoc]] AlignConfig
+    - from_text_vision_configs
+
+## AlignTextConfig
+
+[[autodoc]] AlignTextConfig
+
+## AlignVisionConfig
+
+[[autodoc]] AlignVisionConfig
+
+## AlignProcessor
+
+[[autodoc]] AlignProcessor
+
+## AlignModel
+
+[[autodoc]] AlignModel
+    - forward
+    - get_text_features
+    - get_image_features
+
+## AlignTextModel
+
+[[autodoc]] AlignTextModel
+    - forward
+
+## AlignVisionModel
+
+[[autodoc]] AlignVisionModel
+    - forward

docs/source/es/index.mdx

@@ -46,6 +46,7 @@ La biblioteca actualmente contiene implementaciones de JAX, PyTorch y TensorFlow
 <!--This list is updated automatically from the README with _make fix-copies_. Do not update manually! -->
 
 1. **[ALBERT](model_doc/albert)** (de Google Research y el Instituto Tecnológico de Toyota en Chicago) publicado con el paper [ALBERT: A Lite BERT for Self-supervised Learning of Language Representations](https://arxiv.org/abs/1909.11942), por Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, Radu Soricut.
+1. **[ALIGN](model_doc/align)** (de Google Research) publicado con el paper [Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision](https://arxiv.org/abs/2102.05918) por Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, Tom Duerig.
 1. **[BART](model_doc/bart)** (de Facebook) publicado con el paper [BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension](https://arxiv.org/abs/1910.13461) por Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov y Luke Zettlemoyer.
 1. **[BARThez](model_doc/barthez)** (de École polytechnique) publicado con el paper [BARThez: a Skilled Pretrained French Sequence-to-Sequence Model](https://arxiv.org/abs/2010.12321) por Moussa Kamal Eddine, Antoine J.-P. Tixier, Michalis Vazirgiannis.
 1. **[BARTpho](model_doc/bartpho)** (de VinAI Research) publicado con el paper [BARTpho: Pre-trained Sequence-to-Sequence Models for Vietnamese](https://arxiv.org/abs/2109.09701) por Nguyen Luong Tran, Duong Minh Le y Dat Quoc Nguyen.

docs/source/fr/index.mdx

@@ -50,6 +50,7 @@ La documentation est organisée en 5 parties:
 <!--This list is updated automatically from the README with _make fix-copies_. Do not update manually! -->
 
 1. **[ALBERT](model_doc/albert)** (from Google Research and the Toyota Technological Institute at Chicago) released with the paper [ALBERT: A Lite BERT for Self-supervised Learning of Language Representations](https://arxiv.org/abs/1909.11942), by Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, Radu Soricut.
+1. **[ALIGN](model_doc/align)** (from Google Research) released with the paper [Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision](https://arxiv.org/abs/2102.05918) by Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, Tom Duerig.
 1. **[AltCLIP](model_doc/altclip)** (from BAAI) released with the paper [AltCLIP: Altering the Language Encoder in CLIP for Extended Language Capabilities](https://arxiv.org/abs/2211.06679) by Chen, Zhongzhi and Liu, Guang and Zhang, Bo-Wen and Ye, Fulong and Yang, Qinghong and Wu, Ledell.
 1. **[Audio Spectrogram Transformer](model_doc/audio-spectrogram-transformer)** (from MIT) released with the paper [AST: Audio Spectrogram Transformer](https://arxiv.org/abs/2104.01778) by Yuan Gong, Yu-An Chung, James Glass.
 1. **[BART](model_doc/bart)** (from Facebook) released with the paper [BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension](https://arxiv.org/abs/1910.13461) by Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov and Luke Zettlemoyer.

docs/source/it/index.mdx

@@ -51,6 +51,7 @@ La libreria attualmente contiene implementazioni in JAX, PyTorch e TensorFlow, p
 <!--This list is updated automatically from the README with _make fix-copies_. Do not update manually! -->
 
 1. **[ALBERT](model_doc/albert)** (da Google Research e l'Istituto Tecnologico di Chicago) rilasciato con il paper [ALBERT: A Lite BERT for Self-supervised Learning of Language Representations](https://arxiv.org/abs/1909.11942), da Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, Radu Soricut.
+1. **[ALIGN](model_doc/align)** (from Google Research) rilasciato con il paper [Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision](https://arxiv.org/abs/2102.05918) da Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, Tom Duerig.
 1. **[BART](model_doc/bart)** (da Facebook) rilasciato con il paper [BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension](https://arxiv.org/abs/1910.13461) da Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov e Luke Zettlemoyer.
 1. **[BARThez](model_doc/barthez)** (da politecnico di École) rilasciato con il paper [BARThez: a Skilled Pretrained French Sequence-to-Sequence Model](https://arxiv.org/abs/2010.12321) da Moussa Kamal Eddine, Antoine J.-P. Tixier, Michalis Vazirgiannis.
 1. **[BARTpho](model_doc/bartpho)** (da VinAI Research) rilasciato con il paper [BARTpho: Pre-trained Sequence-to-Sequence Models for Vietnamese](https://arxiv.org/abs/2109.09701) da Nguyen Luong Tran, Duong Minh Le e Dat Quoc Nguyen.

src/transformers/__init__.py

@@ -123,6 +123,13 @@ _import_structure = {
     "models": [],
     # Models
     "models.albert": ["ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP", "AlbertConfig"],
+    "models.align": [
+        "ALIGN_PRETRAINED_CONFIG_ARCHIVE_MAP",
+        "AlignConfig",
+        "AlignProcessor",
+        "AlignTextConfig",
+        "AlignVisionConfig",
+    ],
     "models.altclip": [
         "ALTCLIP_PRETRAINED_CONFIG_ARCHIVE_MAP",
         "AltCLIPConfig",
@@ -983,6 +990,15 @@ else:
             "load_tf_weights_in_albert",
         ]
     )
+    _import_structure["models.align"].extend(
+        [
+            "ALIGN_PRETRAINED_MODEL_ARCHIVE_LIST",
+            "AlignModel",
+            "AlignPreTrainedModel",
+            "AlignTextModel",
+            "AlignVisionModel",
+        ]
+    )
     _import_structure["models.altclip"].extend(
         [
             "ALTCLIP_PRETRAINED_MODEL_ARCHIVE_LIST",
@@ -3710,6 +3726,13 @@ if TYPE_CHECKING:
         load_tf2_weights_in_pytorch_model,
     )
     from .models.albert import ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, AlbertConfig
+    from .models.align import (
+        ALIGN_PRETRAINED_CONFIG_ARCHIVE_MAP,
+        AlignConfig,
+        AlignProcessor,
+        AlignTextConfig,
+        AlignVisionConfig,
+    )
     from .models.altclip import (
         ALTCLIP_PRETRAINED_CONFIG_ARCHIVE_MAP,
         AltCLIPConfig,
@@ -4474,6 +4497,13 @@ if TYPE_CHECKING:
             AlbertPreTrainedModel,
             load_tf_weights_in_albert,
         )
+        from .models.align import (
+            ALIGN_PRETRAINED_MODEL_ARCHIVE_LIST,
+            AlignModel,
+            AlignPreTrainedModel,
+            AlignTextModel,
+            AlignVisionModel,
+        )
         from .models.altclip import (
             ALTCLIP_PRETRAINED_MODEL_ARCHIVE_LIST,
             AltCLIPModel,

src/transformers/models/__init__.py

@@ -14,6 +14,7 @@
 
 from . import (
     albert,
+    align,
     altclip,
     audio_spectrogram_transformer,
     auto,

src/transformers/models/align/__init__.py

+# Copyright 2023 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,
+)
+
+
+_import_structure = {
+    "configuration_align": [
+        "ALIGN_PRETRAINED_CONFIG_ARCHIVE_MAP",
+        "AlignConfig",
+        "AlignTextConfig",
+        "AlignVisionConfig",
+    ],
+    "processing_align": ["AlignProcessor"],
+}
+
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_align"] = [
+        "ALIGN_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "AlignModel",
+        "AlignPreTrainedModel",
+        "AlignTextModel",
+        "AlignVisionModel",
+    ]
+
+if TYPE_CHECKING:
+    from .configuration_align import (
+        ALIGN_PRETRAINED_CONFIG_ARCHIVE_MAP,
+        AlignConfig,
+        AlignTextConfig,
+        AlignVisionConfig,
+    )
+    from .processing_align import AlignProcessor
+
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_align import (
+            ALIGN_PRETRAINED_MODEL_ARCHIVE_LIST,
+            AlignModel,
+            AlignPreTrainedModel,
+            AlignTextModel,
+            AlignVisionModel,
+        )
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)

src/transformers/models/align/configuration_align.py

+# coding=utf-8
+# Copyright 2023 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" ALIGN model configuration"""
+
+import copy
+import os
+from typing import TYPE_CHECKING, List, Union
+
+
+if TYPE_CHECKING:
+    pass
+
+from ...configuration_utils import PretrainedConfig
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+ALIGN_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "kakaobrain/align-base": "https://huggingface.co/kakaobrain/align-base/resolve/main/config.json",
+}
+
+
+class AlignTextConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`AlignTextModel`]. It is used to instantiate a
+    ALIGN text encoder according to the specified arguments, defining the model architecture. Instantiating a
+    configuration with the defaults will yield a similar configuration to that of the text encoder of the ALIGN
+    [kakaobrain/align-base](https://huggingface.co/kakaobrain/align-base) architecture. The default values here are
+    copied from BERT.
+
+    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 Align Text model. Defines the number of different tokens that can be represented by
+            the `inputs_ids` passed when calling [`AlignTextModel`].
+        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" (often named feed-forward) layer in the Transformer encoder.
+        hidden_act (`str` or `Callable`, *optional*, defaults to `"gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"silu"` and `"gelu_new"` are supported.
+        hidden_dropout_prob (`float`, *optional*, defaults to 0.1):
+            The dropout probability 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.
+        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).
+        type_vocab_size (`int`, *optional*, defaults to 2):
+            The vocabulary size of the `token_type_ids` passed when calling [`AlignTextModel`].
+        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.
+        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).
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models). Only
+            relevant if `config.is_decoder=True`.
+        pad_token_id (`int`, *optional*, defaults to 0)
+            Padding token id.
+
+    Example:
+
+    ```python
+    >>> from transformers import AlignTextConfig, AlignTextModel
+
+    >>> # Initializing a AlignTextConfig with kakaobrain/align-base style configuration
+    >>> configuration = AlignTextConfig()
+
+    >>> # Initializing a AlignTextModel (with random weights) from the kakaobrain/align-base style configuration
+    >>> model = AlignTextModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+    model_type = "align_text_model"
+
+    def __init__(
+        self,
+        vocab_size=30522,
+        hidden_size=768,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        intermediate_size=3072,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        max_position_embeddings=512,
+        type_vocab_size=2,
+        initializer_range=0.02,
+        layer_norm_eps=1e-12,
+        pad_token_id=0,
+        position_embedding_type="absolute",
+        use_cache=True,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.hidden_act = hidden_act
+        self.intermediate_size = intermediate_size
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.max_position_embeddings = max_position_embeddings
+        self.type_vocab_size = type_vocab_size
+        self.initializer_range = initializer_range
+        self.layer_norm_eps = layer_norm_eps
+        self.position_embedding_type = position_embedding_type
+        self.use_cache = use_cache
+        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 AlignConfig
+        if config_dict.get("model_type") == "align":
+            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 AlignVisionConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`AlignVisionModel`]. It is used to instantiate a
+    ALIGN vision encoder according to the specified arguments, defining the model architecture. Instantiating a
+    configuration with the defaults will yield a similar configuration to that of the vision encoder of the ALIGN
+    [kakaobrain/align-base](https://huggingface.co/kakaobrain/align-base) architecture. The default values are copied
+    from EfficientNet (efficientnet-b7)
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        num_channels (`int`, *optional*, defaults to 3):
+            The number of input channels.
+        image_size (`int`, *optional*, defaults to 600):
+            The input image size.
+        width_coefficient (`float`, *optional*, defaults to 2.0):
+            Scaling coefficient for network width at each stage.
+        depth_coefficient (`float`, *optional*, defaults to 3.1):
+            Scaling coefficient for network depth at each stage.
+        depth_divisor `int`, *optional*, defaults to 8):
+            A unit of network width.
+        kernel_sizes (`List[int]`, *optional*, defaults to `[3, 3, 5, 3, 5, 5, 3]`):
+            List of kernel sizes to be used in each block.
+        in_channels (`List[int]`, *optional*, defaults to `[32, 16, 24, 40, 80, 112, 192]`):
+            List of input channel sizes to be used in each block for convolutional layers.
+        out_channels (`List[int]`, *optional*, defaults to `[16, 24, 40, 80, 112, 192, 320]`):
+            List of output channel sizes to be used in each block for convolutional layers.
+        depthwise_padding (`List[int]`, *optional*, defaults to `[]`):
+            List of block indices with square padding.
+        strides: (`List[int]`, *optional*, defaults to `[1, 2, 2, 2, 1, 2, 1]`):
+            List of stride sizes to be used in each block for convolutional layers.
+        num_block_repeats (`List[int]`, *optional*, defaults to `[1, 2, 2, 3, 3, 4, 1]`):
+            List of the number of times each block is to repeated.
+        expand_ratios (`List[int]`, *optional*, defaults to `[1, 6, 6, 6, 6, 6, 6]`):
+            List of scaling coefficient of each block.
+        squeeze_expansion_ratio (`float`, *optional*, defaults to 0.25):
+            Squeeze expansion ratio.
+        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
+            The non-linear activation function (function or string) in each block. If string, `"gelu"`, `"relu"`,
+            `"selu", `"gelu_new"`, `"silu"` and `"mish"` are supported.
+        hiddem_dim (`int`, *optional*, defaults to 1280):
+            The hidden dimension of the layer before the classification head.
+        pooling_type (`str` or `function`, *optional*, defaults to `"mean"`):
+            Type of final pooling to be applied before the dense classification head. Available options are [`"mean"`,
+            `"max"`]
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        batch_norm_eps (`float`, *optional*, defaults to 1e-3):
+            The epsilon used by the batch normalization layers.
+        batch_norm_momentum (`float`, *optional*, defaults to 0.99):
+            The momentum used by the batch normalization layers.
+        dropout_rate (`float`, *optional*, defaults to 0.5):
+            The dropout rate to be applied before final classifier layer.
+        drop_connect_rate (`float`, *optional*, defaults to 0.2):
+            The drop rate for skip connections.
+
+    Example:
+
+    ```python
+    >>> from transformers import AlignVisionConfig, AlignVisionModel
+
+    >>> # Initializing a AlignVisionConfig with kakaobrain/align-base style configuration
+    >>> configuration = AlignVisionConfig()
+
+    >>> # Initializing a AlignVisionModel (with random weights) from the kakaobrain/align-base style configuration
+    >>> model = AlignVisionModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "align_vision_model"
+
+    def __init__(
+        self,
+        num_channels: int = 3,
+        image_size: int = 600,
+        width_coefficient: float = 2.0,
+        depth_coefficient: float = 3.1,
+        depth_divisor: int = 8,
+        kernel_sizes: List[int] = [3, 3, 5, 3, 5, 5, 3],
+        in_channels: List[int] = [32, 16, 24, 40, 80, 112, 192],
+        out_channels: List[int] = [16, 24, 40, 80, 112, 192, 320],
+        depthwise_padding: List[int] = [],
+        strides: List[int] = [1, 2, 2, 2, 1, 2, 1],
+        num_block_repeats: List[int] = [1, 2, 2, 3, 3, 4, 1],
+        expand_ratios: List[int] = [1, 6, 6, 6, 6, 6, 6],
+        squeeze_expansion_ratio: float = 0.25,
+        hidden_act: str = "swish",
+        hidden_dim: int = 2560,
+        pooling_type: str = "mean",
+        initializer_range: float = 0.02,
+        batch_norm_eps: float = 0.001,
+        batch_norm_momentum: float = 0.99,
+        dropout_rate: float = 0.5,
+        drop_connect_rate: float = 0.2,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        self.num_channels = num_channels
+        self.image_size = image_size
+        self.width_coefficient = width_coefficient
+        self.depth_coefficient = depth_coefficient
+        self.depth_divisor = depth_divisor
+        self.kernel_sizes = kernel_sizes
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.depthwise_padding = depthwise_padding
+        self.strides = strides
+        self.num_block_repeats = num_block_repeats
+        self.expand_ratios = expand_ratios
+        self.squeeze_expansion_ratio = squeeze_expansion_ratio
+        self.hidden_act = hidden_act
+        self.hidden_dim = hidden_dim
+        self.pooling_type = pooling_type
+        self.initializer_range = initializer_range
+        self.batch_norm_eps = batch_norm_eps
+        self.batch_norm_momentum = batch_norm_momentum
+        self.dropout_rate = dropout_rate
+        self.drop_connect_rate = drop_connect_rate
+        self.num_hidden_layers = sum(num_block_repeats) * 4
+
+    @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 vision config dict if we are loading from AlignConfig
+        if config_dict.get("model_type") == "align":
+            config_dict = config_dict["vision_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 AlignConfig(PretrainedConfig):
+    r"""
+    [`AlignConfig`] is the configuration class to store the configuration of a [`AlignModel`]. It is used to
+    instantiate a ALIGN model according to the specified arguments, defining the text model and vision model configs.
+    Instantiating a configuration with the defaults will yield a similar configuration to that of the ALIGN
+    [kakaobrain/align-base](https://huggingface.co/kakaobrain/align-base) 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`, *optional*):
+            Dictionary of configuration options used to initialize [`AlignTextConfig`].
+        vision_config (`dict`, *optional*):
+            Dictionary of configuration options used to initialize [`AlignVisionConfig`].
+        projection_dim (`int`, *optional*, defaults to 512):
+            Dimentionality of text and vision projection layers.
+        temperature_init_value (`float`, *optional*, defaults to 1.0):
+            The inital value of the *temperature* paramter. Default is used as per the original ALIGN implementation.
+        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 AlignConfig, AlignModel
+
+    >>> # Initializing a AlignConfig with kakaobrain/align-base style configuration
+    >>> configuration = AlignConfig()
+
+    >>> # Initializing a AlignModel (with random weights) from the kakaobrain/align-base style configuration
+    >>> model = AlignModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+
+    >>> # We can also initialize a AlignConfig from a AlignTextConfig and a AlignVisionConfig
+    >>> from transformers import AlignTextConfig, AlignVisionConfig
+
+    >>> # Initializing ALIGN Text and Vision configurations
+    >>> config_text = AlignTextConfig()
+    >>> config_vision = AlignVisionConfig()
+
+    >>> config = AlignConfig.from_text_vision_configs(config_text, config_vision)
+    ```"""
+
+    model_type = "align"
+    is_composition = True
+
+    def __init__(
+        self,
+        text_config=None,
+        vision_config=None,
+        projection_dim=640,
+        temperature_init_value=1.0,
+        initializer_range=0.02,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        if text_config is None:
+            text_config = {}
+            logger.info("text_config is None. Initializing the AlignTextConfig with default values.")
+
+        if vision_config is None:
+            vision_config = {}
+            logger.info("vision_config is None. Initializing the AlignVisionConfig with default values.")
+
+        self.text_config = AlignTextConfig(**text_config)
+        self.vision_config = AlignVisionConfig(**vision_config)
+
+        self.projection_dim = projection_dim
+        self.temperature_init_value = temperature_init_value
+        self.initializer_range = initializer_range
+
+    @classmethod
+    def from_text_vision_configs(cls, text_config: AlignTextConfig, vision_config: AlignVisionConfig, **kwargs):
+        r"""
+        Instantiate a [`AlignConfig`] (or a derived class) from align text model configuration and align vision model
+        configuration.
+
+        Returns:
+            [`AlignConfig`]: An instance of a configuration object
+        """
+
+        return cls(text_config=text_config.to_dict(), vision_config=vision_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["text_config"] = self.text_config.to_dict()
+        output["vision_config"] = self.vision_config.to_dict()
+        output["model_type"] = self.__class__.model_type
+        return output

src/transformers/models/align/convert_align_tf_to_hf.py

+# coding=utf-8
+# Copyright 2023 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Convert ALIGN checkpoints from the original repository."""
+
+import argparse
+import os
+
+import align
+import numpy as np
+import requests
+import tensorflow as tf
+import torch
+from PIL import Image
+from tokenizer import Tokenizer
+
+from transformers import (
+    AlignConfig,
+    AlignModel,
+    AlignProcessor,
+    BertConfig,
+    BertTokenizer,
+    EfficientNetConfig,
+    EfficientNetImageProcessor,
+)
+from transformers.utils import logging
+
+
+logging.set_verbosity_info()
+logger = logging.get_logger(__name__)
+
+
+def preprocess(image):
+    image = tf.image.resize(image, (346, 346))
+    image = tf.image.crop_to_bounding_box(image, (346 - 289) // 2, (346 - 289) // 2, 289, 289)
+    return image
+
+
+def get_align_config():
+    vision_config = EfficientNetConfig.from_pretrained("google/efficientnet-b7")
+    vision_config.image_size = 289
+    vision_config.hidden_dim = 640
+    vision_config.id2label = {"0": "LABEL_0", "1": "LABEL_1"}
+    vision_config.label2id = {"LABEL_0": 0, "LABEL_1": 1}
+    vision_config.depthwise_padding = []
+
+    text_config = BertConfig()
+    config = AlignConfig.from_text_vision_configs(
+        text_config=text_config, vision_config=vision_config, projection_dim=640
+    )
+    return config
+
+
+# We will verify our results on an image of cute cats
+def prepare_img():
+    url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+    im = Image.open(requests.get(url, stream=True).raw)
+    return im
+
+
+def get_processor():
+    image_processor = EfficientNetImageProcessor(
+        do_center_crop=True,
+        rescale_factor=1 / 127.5,
+        rescale_offset=True,
+        do_normalize=False,
+        include_top=False,
+        resample=Image.BILINEAR,
+    )
+    tokenizer = BertTokenizer.from_pretrained("bert-base-uncased")
+    tokenizer.model_max_length = 64
+    processor = AlignProcessor(image_processor=image_processor, tokenizer=tokenizer)
+    return processor
+
+
+# here we list all keys to be renamed (original name on the left, our name on the right)
+def rename_keys(original_param_names):
+    # EfficientNet image encoder
+    block_names = [v.split("_")[0].split("block")[1] for v in original_param_names if v.startswith("block")]
+    block_names = list(set(block_names))
+    block_names = sorted(block_names)
+    num_blocks = len(block_names)
+    block_name_mapping = {b: str(i) for b, i in zip(block_names, range(num_blocks))}
+
+    rename_keys = []
+    rename_keys.append(("stem_conv/kernel:0", "embeddings.convolution.weight"))
+    rename_keys.append(("stem_bn/gamma:0", "embeddings.batchnorm.weight"))
+    rename_keys.append(("stem_bn/beta:0", "embeddings.batchnorm.bias"))
+    rename_keys.append(("stem_bn/moving_mean:0", "embeddings.batchnorm.running_mean"))
+    rename_keys.append(("stem_bn/moving_variance:0", "embeddings.batchnorm.running_var"))
+
+    for b in block_names:
+        hf_b = block_name_mapping[b]
+        rename_keys.append((f"block{b}_expand_conv/kernel:0", f"encoder.blocks.{hf_b}.expansion.expand_conv.weight"))
+        rename_keys.append((f"block{b}_expand_bn/gamma:0", f"encoder.blocks.{hf_b}.expansion.expand_bn.weight"))
+        rename_keys.append((f"block{b}_expand_bn/beta:0", f"encoder.blocks.{hf_b}.expansion.expand_bn.bias"))
+        rename_keys.append(
+            (f"block{b}_expand_bn/moving_mean:0", f"encoder.blocks.{hf_b}.expansion.expand_bn.running_mean")
+        )
+        rename_keys.append(
+            (f"block{b}_expand_bn/moving_variance:0", f"encoder.blocks.{hf_b}.expansion.expand_bn.running_var")
+        )
+        rename_keys.append(
+            (f"block{b}_dwconv/depthwise_kernel:0", f"encoder.blocks.{hf_b}.depthwise_conv.depthwise_conv.weight")
+        )
+        rename_keys.append((f"block{b}_bn/gamma:0", f"encoder.blocks.{hf_b}.depthwise_conv.depthwise_norm.weight"))
+        rename_keys.append((f"block{b}_bn/beta:0", f"encoder.blocks.{hf_b}.depthwise_conv.depthwise_norm.bias"))
+        rename_keys.append(
+            (f"block{b}_bn/moving_mean:0", f"encoder.blocks.{hf_b}.depthwise_conv.depthwise_norm.running_mean")
+        )
+        rename_keys.append(
+            (f"block{b}_bn/moving_variance:0", f"encoder.blocks.{hf_b}.depthwise_conv.depthwise_norm.running_var")
+        )
+
+        rename_keys.append((f"block{b}_se_reduce/kernel:0", f"encoder.blocks.{hf_b}.squeeze_excite.reduce.weight"))
+        rename_keys.append((f"block{b}_se_reduce/bias:0", f"encoder.blocks.{hf_b}.squeeze_excite.reduce.bias"))
+        rename_keys.append((f"block{b}_se_expand/kernel:0", f"encoder.blocks.{hf_b}.squeeze_excite.expand.weight"))
+        rename_keys.append((f"block{b}_se_expand/bias:0", f"encoder.blocks.{hf_b}.squeeze_excite.expand.bias"))
+        rename_keys.append(
+            (f"block{b}_project_conv/kernel:0", f"encoder.blocks.{hf_b}.projection.project_conv.weight")
+        )
+        rename_keys.append((f"block{b}_project_bn/gamma:0", f"encoder.blocks.{hf_b}.projection.project_bn.weight"))
+        rename_keys.append((f"block{b}_project_bn/beta:0", f"encoder.blocks.{hf_b}.projection.project_bn.bias"))
+        rename_keys.append(
+            (f"block{b}_project_bn/moving_mean:0", f"encoder.blocks.{hf_b}.projection.project_bn.running_mean")
+        )
+        rename_keys.append(
+            (f"block{b}_project_bn/moving_variance:0", f"encoder.blocks.{hf_b}.projection.project_bn.running_var")
+        )
+
+    key_mapping = {}
+    for item in rename_keys:
+        if item[0] in original_param_names:
+            key_mapping[item[0]] = "vision_model." + item[1]
+
+    # BERT text encoder
+    rename_keys = []
+    old = "tf_bert_model/bert"
+    new = "text_model"
+    for i in range(12):
+        rename_keys.append(
+            (
+                f"{old}/encoder/layer_._{i}/attention/self/query/kernel:0",
+                f"{new}.encoder.layer.{i}.attention.self.query.weight",
+            )
+        )
+        rename_keys.append(
+            (
+                f"{old}/encoder/layer_._{i}/attention/self/query/bias:0",
+                f"{new}.encoder.layer.{i}.attention.self.query.bias",
+            )
+        )
+        rename_keys.append(
+            (
+                f"{old}/encoder/layer_._{i}/attention/self/key/kernel:0",
+                f"{new}.encoder.layer.{i}.attention.self.key.weight",
+            )
+        )
+        rename_keys.append(
+            (
+                f"{old}/encoder/layer_._{i}/attention/self/key/bias:0",
+                f"{new}.encoder.layer.{i}.attention.self.key.bias",
+            )
+        )
+        rename_keys.append(
+            (
+                f"{old}/encoder/layer_._{i}/attention/self/value/kernel:0",
+                f"{new}.encoder.layer.{i}.attention.self.value.weight",
+            )
+        )
+        rename_keys.append(
+            (
+                f"{old}/encoder/layer_._{i}/attention/self/value/bias:0",
+                f"{new}.encoder.layer.{i}.attention.self.value.bias",
+            )
+        )
+        rename_keys.append(
+            (
+                f"{old}/encoder/layer_._{i}/attention/output/dense/kernel:0",
+                f"{new}.encoder.layer.{i}.attention.output.dense.weight",
+            )
+        )
+        rename_keys.append(
+            (
+                f"{old}/encoder/layer_._{i}/attention/output/dense/bias:0",
+                f"{new}.encoder.layer.{i}.attention.output.dense.bias",
+            )
+        )
+        rename_keys.append(
+            (
+                f"{old}/encoder/layer_._{i}/attention/output/LayerNorm/gamma:0",
+                f"{new}.encoder.layer.{i}.attention.output.LayerNorm.weight",
+            )
+        )
+        rename_keys.append(
+            (
+                f"{old}/encoder/layer_._{i}/attention/output/LayerNorm/beta:0",
+                f"{new}.encoder.layer.{i}.attention.output.LayerNorm.bias",
+            )
+        )
+        rename_keys.append(
+            (
+                f"{old}/encoder/layer_._{i}/intermediate/dense/kernel:0",
+                f"{new}.encoder.layer.{i}.intermediate.dense.weight",
+            )
+        )
+        rename_keys.append(
+            (
+                f"{old}/encoder/layer_._{i}/intermediate/dense/bias:0",
+                f"{new}.encoder.layer.{i}.intermediate.dense.bias",
+            )
+        )
+        rename_keys.append(
+            (f"{old}/encoder/layer_._{i}/output/dense/kernel:0", f"{new}.encoder.layer.{i}.output.dense.weight")
+        )
+        rename_keys.append(
+            (f"{old}/encoder/layer_._{i}/output/dense/bias:0", f"{new}.encoder.layer.{i}.output.dense.bias")
+        )
+        rename_keys.append(
+            (f"{old}/encoder/layer_._{i}/output/LayerNorm/gamma:0", f"{new}.encoder.layer.{i}.output.LayerNorm.weight")
+        )
+        rename_keys.append(
+            (f"{old}/encoder/layer_._{i}/output/LayerNorm/beta:0", f"{new}.encoder.layer.{i}.output.LayerNorm.bias")
+        )
+
+    rename_keys.append((f"{old}/embeddings/word_embeddings/weight:0", f"{new}.embeddings.word_embeddings.weight"))
+    rename_keys.append(
+        (f"{old}/embeddings/position_embeddings/embeddings:0", f"{new}.embeddings.position_embeddings.weight")
+    )
+    rename_keys.append(
+        (f"{old}/embeddings/token_type_embeddings/embeddings:0", f"{new}.embeddings.token_type_embeddings.weight")
+    )
+    rename_keys.append((f"{old}/embeddings/LayerNorm/gamma:0", f"{new}.embeddings.LayerNorm.weight"))
+    rename_keys.append((f"{old}/embeddings/LayerNorm/beta:0", f"{new}.embeddings.LayerNorm.bias"))
+
+    rename_keys.append((f"{old}/pooler/dense/kernel:0", f"{new}.pooler.dense.weight"))
+    rename_keys.append((f"{old}/pooler/dense/bias:0", f"{new}.pooler.dense.bias"))
+    rename_keys.append(("dense/kernel:0", "text_projection.weight"))
+    rename_keys.append(("dense/bias:0", "text_projection.bias"))
+    rename_keys.append(("dense/bias:0", "text_projection.bias"))
+    rename_keys.append(("temperature:0", "temperature"))
+
+    for item in rename_keys:
+        if item[0] in original_param_names:
+            key_mapping[item[0]] = item[1]
+    return key_mapping
+
+
+def replace_params(hf_params, tf_params, key_mapping):
+    list(hf_params.keys())
+
+    for key, value in tf_params.items():
+        if key not in key_mapping:
+            continue
+
+        hf_key = key_mapping[key]
+        if "_conv" in key and "kernel" in key:
+            new_hf_value = torch.from_numpy(value).permute(3, 2, 0, 1)
+        elif "embeddings" in key:
+            new_hf_value = torch.from_numpy(value)
+        elif "depthwise_kernel" in key:
+            new_hf_value = torch.from_numpy(value).permute(2, 3, 0, 1)
+        elif "kernel" in key:
+            new_hf_value = torch.from_numpy(np.transpose(value))
+        elif "temperature" in key:
+            new_hf_value = value
+        elif "bn/gamma" or "bn/beta" in key:
+            new_hf_value = torch.from_numpy(np.transpose(value)).squeeze()
+        else:
+            new_hf_value = torch.from_numpy(value)
+
+        # Replace HF parameters with original TF model parameters
+        hf_params[hf_key].copy_(new_hf_value)
+
+
+@torch.no_grad()
+def convert_align_checkpoint(checkpoint_path, pytorch_dump_folder_path, save_model, push_to_hub):
+    """
+    Copy/paste/tweak model's weights to our ALIGN structure.
+    """
+    # Load original model
+    seq_length = 64
+    tok = Tokenizer(seq_length)
+    original_model = align.Align("efficientnet-b7", "bert-base", 640, seq_length, tok.get_vocab_size())
+    original_model.compile()
+    original_model.load_weights(checkpoint_path)
+
+    tf_params = original_model.trainable_variables
+    tf_non_train_params = original_model.non_trainable_variables
+    tf_params = {param.name: param.numpy() for param in tf_params}
+    for param in tf_non_train_params:
+        tf_params[param.name] = param.numpy()
+    tf_param_names = list(tf_params.keys())
+
+    # Load HuggingFace model
+    config = get_align_config()
+    hf_model = AlignModel(config).eval()
+    hf_params = hf_model.state_dict()
+
+    # Create src-to-dst parameter name mapping dictionary
+    print("Converting parameters...")
+    key_mapping = rename_keys(tf_param_names)
+    replace_params(hf_params, tf_params, key_mapping)
+
+    # Initialize processor
+    processor = get_processor()
+    inputs = processor(
+        images=prepare_img(), text="A picture of a cat", padding="max_length", max_length=64, return_tensors="pt"
+    )
+
+    # HF model inference
+    hf_model.eval()
+    with torch.no_grad():
+        outputs = hf_model(**inputs)
+
+    hf_image_features = outputs.image_embeds.detach().numpy()
+    hf_text_features = outputs.text_embeds.detach().numpy()
+
+    # Original model inference
+    original_model.trainable = False
+    tf_image_processor = EfficientNetImageProcessor(
+        do_center_crop=True,
+        do_rescale=False,
+        do_normalize=False,
+        include_top=False,
+        resample=Image.BILINEAR,
+    )
+    image = tf_image_processor(images=prepare_img(), return_tensors="tf", data_format="channels_last")["pixel_values"]
+    text = tok(tf.constant(["A picture of a cat"]))
+
+    image_features = original_model.image_encoder(image, training=False)
+    text_features = original_model.text_encoder(text, training=False)
+
+    image_features = tf.nn.l2_normalize(image_features, axis=-1)
+    text_features = tf.nn.l2_normalize(text_features, axis=-1)
+
+    # Check whether original and HF model outputs match  -> np.allclose
+    assert np.allclose(image_features, hf_image_features, atol=1e-3), "The predicted image features are not the same."
+    assert np.allclose(text_features, hf_text_features, atol=1e-3), "The predicted text features are not the same."
+    print("Model outputs match!")
+
+    if save_model:
+        # Create folder to save model
+        if not os.path.isdir(pytorch_dump_folder_path):
+            os.mkdir(pytorch_dump_folder_path)
+        # Save converted model and feature extractor
+        hf_model.save_pretrained(pytorch_dump_folder_path)
+        processor.save_pretrained(pytorch_dump_folder_path)
+
+    if push_to_hub:
+        # Push model and feature extractor to hub
+        print("Pushing converted ALIGN to the hub...")
+        processor.push_to_hub("align-base")
+        hf_model.push_to_hub("align-base")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    # Required parameters
+    parser.add_argument(
+        "--checkpoint_path",
+        default="./weights/model-weights",
+        type=str,
+        help="Path to the pretrained TF ALIGN checkpoint.",
+    )
+    parser.add_argument(
+        "--pytorch_dump_folder_path",
+        default="hf_model",
+        type=str,
+        help="Path to the output PyTorch model directory.",
+    )
+    parser.add_argument("--save_model", action="store_true", help="Save model to local")
+    parser.add_argument("--push_to_hub", action="store_true", help="Push model and feature extractor to the hub")
+
+    args = parser.parse_args()
+    convert_align_checkpoint(args.checkpoint_path, args.pytorch_dump_folder_path, args.save_model, args.push_to_hub)

src/transformers/models/align/modeling_align.py

+# coding=utf-8
+# Copyright 2023 The Google Research Team 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.
+""" PyTorch ALIGN model."""
+
+import math
+from dataclasses import dataclass
+from typing import Any, Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+
+from ...activations import ACT2FN
+from ...modeling_outputs import (
+    BaseModelOutputWithNoAttention,
+    BaseModelOutputWithPastAndCrossAttentions,
+    BaseModelOutputWithPoolingAndCrossAttentions,
+    BaseModelOutputWithPoolingAndNoAttention,
+)
+from ...modeling_utils import PreTrainedModel
+from ...pytorch_utils import apply_chunking_to_forward, find_pruneable_heads_and_indices, prune_linear_layer
+from ...utils import (
+    ModelOutput,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    logging,
+    replace_return_docstrings,
+)
+from .configuration_align import AlignConfig, AlignTextConfig, AlignVisionConfig
+
+
+logger = logging.get_logger(__name__)
+
+_CHECKPOINT_FOR_DOC = "kakaobrain/align-base"
+_CONFIG_FOR_DOC = "AlignConfig"
+
+
+ALIGN_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "kakaobrain/align-base",
+    # See all ALIGN models at https://huggingface.co/models?filter=align
+]
+
+
+ALIGN_START_DOCSTRING = r"""
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+
+    Parameters:
+        config ([`AlignConfig`]): Model configuration class with all the parameters of the model.
+            Initializing with a config file does not load the weights associated with the model, only the
+            configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+ALIGN_TEXT_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+            it.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.max_position_embeddings - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        token_type_ids (`torch.LongTensor` of shape `({0})`, *optional*):
+            Segment token indices to indicate first and second portions of the inputs. Indices are selected in `[0,
+            1]`:
+
+            - 0 corresponds to a *sentence A* token,
+            - 1 corresponds to a *sentence B* token.
+
+            [What are token type IDs?](../glossary#token-type-ids)
+        head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
+            Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        inputs_embeds (`torch.FloatTensor` of shape `({0}, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
+            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
+            model's internal embedding lookup matrix.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+ALIGN_VISION_INPUTS_DOCSTRING = r"""
+    Args:
+        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            Pixel values. Padding will be ignored by default should you provide it. Pixel values can be obtained using
+            [`AutoImageProcessor`]. See [`CLIPImageProcessor.__call__`] for details.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+ALIGN_INPUTS_DOCSTRING = r"""
+    Args:
+       input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+            it.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.max_position_embeddings - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        token_type_ids (`torch.LongTensor` of shape `({0})`, *optional*):
+            Segment token indices to indicate first and second portions of the inputs. Indices are selected in `[0,
+            1]`:
+
+            - 0 corresponds to a *sentence A* token,
+            - 1 corresponds to a *sentence B* token.
+
+            [What are token type IDs?](../glossary#token-type-ids)
+        head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
+            Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        inputs_embeds (`torch.FloatTensor` of shape `({0}, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
+            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
+            model's internal embedding lookup matrix.
+        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            Pixel values. Padding will be ignored by default should you provide it. Pixel values can be obtained using
+            [`AutoImageProcessor`]. See [`CLIPImageProcessor.__call__`] for details.
+        return_loss (`bool`, *optional*):
+            Whether or not to return the contrastive loss.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+@dataclass
+class AlignVisionModelOutput(ModelOutput):
+    """
+    Base class for vision model's outputs that also contains image embeddings of the pooling of the last hidden states.
+
+    Args:
+        image_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim)` *optional* returned when model is initialized with `with_projection=True`):
+            The image embeddings obtained by applying the projection layer to the pooler_output.
+        last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
+            Sequence of hidden-states at the output of the last layer of the model.
+        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
+            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
+    """
+
+    image_embeds: Optional[torch.FloatTensor] = None
+    last_hidden_state: torch.FloatTensor = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+
+
+@dataclass
+class AlignTextModelOutput(ModelOutput):
+    """
+    Base class for text model's outputs that also contains a pooling of the last hidden states.
+
+    Args:
+        text_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim)` *optional* returned when model is initialized with `with_projection=True`):
+            The text embeddings obtained by applying the projection layer to the pooler_output.
+        last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
+            Sequence of hidden-states at the output of the last layer of the model.
+        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
+            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
+        attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
+            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
+            sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    text_embeds: Optional[torch.FloatTensor] = None
+    last_hidden_state: torch.FloatTensor = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+
+@dataclass
+class AlignOutput(ModelOutput):
+    """
+    Args:
+        loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `return_loss` is `True`):
+            Contrastive loss for image-text similarity.
+        logits_per_image:(`torch.FloatTensor` of shape `(image_batch_size, text_batch_size)`):
+            The scaled dot product scores between `image_embeds` and `text_embeds`. This represents the image-text
+            similarity scores.
+        logits_per_text:(`torch.FloatTensor` of shape `(text_batch_size, image_batch_size)`):
+            The scaled dot product scores between `text_embeds` and `image_embeds`. This represents the text-image
+            similarity scores.
+        text_embeds(`torch.FloatTensor` of shape `(batch_size, output_dim`):
+            The text embeddings obtained by applying the projection layer to the pooled output of [`AlignTextModel`].
+        image_embeds(`torch.FloatTensor` of shape `(batch_size, output_dim`):
+            The output of [`AlignVisionModel`].
+        text_model_output(`BaseModelOutputWithPoolingAndCrossAttentions`):
+            The output of the [`AlignTextModel`].
+        vision_model_output(`BaseModelOutputWithPoolingAndNoAttention`):
+            The output of the [`AlignVisionModel`].
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    logits_per_image: torch.FloatTensor = None
+    logits_per_text: torch.FloatTensor = None
+    text_embeds: torch.FloatTensor = None
+    image_embeds: torch.FloatTensor = None
+    text_model_output: BaseModelOutputWithPoolingAndCrossAttentions = None
+    vision_model_output: BaseModelOutputWithPoolingAndNoAttention = None
+
+    def to_tuple(self) -> Tuple[Any]:
+        return tuple(
+            self[k] if k not in ["text_model_output", "vision_model_output"] else getattr(self, k).to_tuple()
+            for k in self.keys()
+        )
+
+
+# contrastive loss function, adapted from
+# https://sachinruk.github.io/blog/pytorch/pytorch%20lightning/loss%20function/gpu/2021/03/07/CLIP.html
+def contrastive_loss(logits: torch.Tensor) -> torch.Tensor:
+    return nn.functional.cross_entropy(logits, torch.arange(len(logits), device=logits.device), label_smoothing=0.1)
+
+
+def align_loss(similarity: torch.Tensor) -> torch.Tensor:
+    caption_loss = contrastive_loss(similarity)
+    image_loss = contrastive_loss(similarity.t())
+    return (caption_loss + image_loss) / 2.0
+
+
+# Copied from transformers.models.efficientnet.modeling_efficientnet.round_filters with EfficientNet -> AlignVision
+def round_filters(config: AlignVisionConfig, num_channels: int):
+    r"""
+    Round number of filters based on depth multiplier.
+    """
+    divisor = config.depth_divisor
+    num_channels *= config.width_coefficient
+    new_dim = max(divisor, int(num_channels + divisor / 2) // divisor * divisor)
+
+    # Make sure that round down does not go down by more than 10%.
+    if new_dim < 0.9 * num_channels:
+        new_dim += divisor
+
+    return int(new_dim)
+
+
+# Copied from transformers.models.efficientnet.modeling_efficientnet.correct_pad
+def correct_pad(kernel_size: Union[int, Tuple], adjust: bool = True):
+    r"""
+    Utility function to get the tuple padding value for the depthwise convolution.
+
+    Args:
+        kernel_size (`int` or `tuple`):
+            Kernel size of the convolution layers.
+        adjust (`bool`, *optional*, defaults to `True`):
+            Adjusts padding value to apply to right and bottom sides of the input.
+    """
+    if isinstance(kernel_size, int):
+        kernel_size = (kernel_size, kernel_size)
+
+    correct = (kernel_size[0] // 2, kernel_size[1] // 2)
+    if adjust:
+        return (correct[1] - 1, correct[1], correct[0] - 1, correct[0])
+    else:
+        return (correct[1], correct[1], correct[0], correct[0])
+
+
+# Copied from transformers.models.efficientnet.modeling_efficientnet.EfficientNetEmbeddings with EfficientNet->AlignVision
+class AlignVisionEmbeddings(nn.Module):
+    r"""
+    A module that corresponds to the stem module of the original work.
+    """
+
+    def __init__(self, config: AlignVisionConfig):
+        super().__init__()
+
+        self.out_dim = round_filters(config, 32)
+        self.padding = nn.ZeroPad2d(padding=(0, 1, 0, 1))
+        self.convolution = nn.Conv2d(
+            config.num_channels, self.out_dim, kernel_size=3, stride=2, padding="valid", bias=False
+        )
+        self.batchnorm = nn.BatchNorm2d(self.out_dim, eps=config.batch_norm_eps, momentum=config.batch_norm_momentum)
+        self.activation = ACT2FN[config.hidden_act]
+
+    def forward(self, pixel_values: torch.Tensor) -> torch.Tensor:
+        features = self.padding(pixel_values)
+        features = self.convolution(features)
+        features = self.batchnorm(features)
+        features = self.activation(features)
+
+        return features
+
+
+# Copied from transformers.models.efficientnet.modeling_efficientnet.EfficientNetDepthwiseConv2d with EfficientNet->AlignVision
+class AlignVisionDepthwiseConv2d(nn.Conv2d):
+    def __init__(
+        self,
+        in_channels,
+        depth_multiplier=1,
+        kernel_size=3,
+        stride=1,
+        padding=0,
+        dilation=1,
+        bias=True,
+        padding_mode="zeros",
+    ):
+        out_channels = in_channels * depth_multiplier
+        super().__init__(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            kernel_size=kernel_size,
+            stride=stride,
+            padding=padding,
+            dilation=dilation,
+            groups=in_channels,
+            bias=bias,
+            padding_mode=padding_mode,
+        )
+
+
+# Copied from transformers.models.efficientnet.modeling_efficientnet.EfficientNetExpansionLayer with EfficientNet->AlignVision
+class AlignVisionExpansionLayer(nn.Module):
+    r"""
+    This corresponds to the expansion phase of each block in the original implementation.
+    """
+
+    def __init__(self, config: AlignVisionConfig, in_dim: int, out_dim: int, stride: int):
+        super().__init__()
+        self.expand_conv = nn.Conv2d(
+            in_channels=in_dim,
+            out_channels=out_dim,
+            kernel_size=1,
+            padding="same",
+            bias=False,
+        )
+        self.expand_bn = nn.BatchNorm2d(num_features=out_dim, eps=config.batch_norm_eps)
+        self.expand_act = ACT2FN[config.hidden_act]
+
+    def forward(self, hidden_states: torch.FloatTensor) -> torch.Tensor:
+        # Expand phase
+        hidden_states = self.expand_conv(hidden_states)
+        hidden_states = self.expand_bn(hidden_states)
+        hidden_states = self.expand_act(hidden_states)
+
+        return hidden_states
+
+
+# Copied from transformers.models.efficientnet.modeling_efficientnet.EfficientNetDepthwiseLayer with with EfficientNet->AlignVision
+class AlignVisionDepthwiseLayer(nn.Module):
+    r"""
+    This corresponds to the depthwise convolution phase of each block in the original implementation.
+    """
+
+    def __init__(
+        self,
+        config: AlignVisionConfig,
+        in_dim: int,
+        stride: int,
+        kernel_size: int,
+        adjust_padding: bool,
+    ):
+        super().__init__()
+        self.stride = stride
+        conv_pad = "valid" if self.stride == 2 else "same"
+        padding = correct_pad(kernel_size, adjust=adjust_padding)
+
+        self.depthwise_conv_pad = nn.ZeroPad2d(padding=padding)
+        self.depthwise_conv = AlignVisionDepthwiseConv2d(
+            in_dim, kernel_size=kernel_size, stride=stride, padding=conv_pad, bias=False
+        )
+        self.depthwise_norm = nn.BatchNorm2d(
+            num_features=in_dim, eps=config.batch_norm_eps, momentum=config.batch_norm_momentum
+        )
+        self.depthwise_act = ACT2FN[config.hidden_act]
+
+    def forward(self, hidden_states: torch.FloatTensor) -> torch.Tensor:
+        # Depthwise convolution
+        if self.stride == 2:
+            hidden_states = self.depthwise_conv_pad(hidden_states)
+
+        hidden_states = self.depthwise_conv(hidden_states)
+        hidden_states = self.depthwise_norm(hidden_states)
+        hidden_states = self.depthwise_act(hidden_states)
+
+        return hidden_states
+
+
+# Copied from transformers.models.efficientnet.modeling_efficientnet.EfficientNetSqueezeExciteLayer with with EfficientNet->AlignVision
+class AlignVisionSqueezeExciteLayer(nn.Module):
+    r"""
+    This corresponds to the Squeeze and Excitement phase of each block in the original implementation.
+    """
+
+    def __init__(self, config: AlignVisionConfig, in_dim: int, expand_dim: int, expand: bool = False):
+        super().__init__()
+        self.dim = expand_dim if expand else in_dim
+        self.dim_se = max(1, int(in_dim * config.squeeze_expansion_ratio))
+
+        self.squeeze = nn.AdaptiveAvgPool2d(output_size=1)
+        self.reduce = nn.Conv2d(
+            in_channels=self.dim,
+            out_channels=self.dim_se,
+            kernel_size=1,
+            padding="same",
+        )
+        self.expand = nn.Conv2d(
+            in_channels=self.dim_se,
+            out_channels=self.dim,
+            kernel_size=1,
+            padding="same",
+        )
+        self.act_reduce = ACT2FN[config.hidden_act]
+        self.act_expand = nn.Sigmoid()
+
+    def forward(self, hidden_states: torch.FloatTensor) -> torch.Tensor:
+        inputs = hidden_states
+        hidden_states = self.squeeze(hidden_states)
+        hidden_states = self.reduce(hidden_states)
+        hidden_states = self.act_reduce(hidden_states)
+
+        hidden_states = self.expand(hidden_states)
+        hidden_states = self.act_expand(hidden_states)
+        hidden_states = torch.mul(inputs, hidden_states)
+
+        return hidden_states
+
+
+class AlignVisionFinalBlockLayer(nn.Module):
+    r"""
+    This corresponds to the final phase of each block in the original implementation.
+    """
+
+    def __init__(
+        self, config: AlignVisionConfig, in_dim: int, out_dim: int, stride: int, drop_rate: float, id_skip: bool
+    ):
+        super().__init__()
+        self.apply_dropout = stride == 1 and not id_skip
+        self.project_conv = nn.Conv2d(
+            in_channels=in_dim,
+            out_channels=out_dim,
+            kernel_size=1,
+            padding="same",
+            bias=False,
+        )
+        self.project_bn = nn.BatchNorm2d(
+            num_features=out_dim, eps=config.batch_norm_eps, momentum=config.batch_norm_momentum
+        )
+        self.dropout = nn.Dropout(p=drop_rate)
+
+    def forward(self, embeddings: torch.FloatTensor, hidden_states: torch.FloatTensor) -> torch.Tensor:
+        hidden_states = self.project_conv(hidden_states)
+        hidden_states = self.project_bn(hidden_states)
+
+        if self.apply_dropout:
+            hidden_states = self.dropout(hidden_states)
+            hidden_states = hidden_states + embeddings
+
+        return hidden_states
+
+
+class AlignVisionBlock(nn.Module):
+    r"""
+    This corresponds to the block module of original the EfficientNet vision encoder implementation.
+
+    Args:
+        config ([`AlignVisionConfig`]):
+            Model configuration class.
+        in_dim (`int`):
+            Number of input channels.
+        out_dim (`int`):
+            Number of output channels.
+        stride (`int`):
+            Stride size to be used in convolution layers.
+        expand_ratio (`int`):
+            Expand ratio to set the output dimensions for the expansion and squeeze-excite layers.
+        kernel_size (`int`):
+            Kernel size for the depthwise convolution layer.
+        drop_rate (`float`):
+            Dropout rate to be used in the final phase of each block.
+        id_skip (`bool`):
+            Whether to apply dropout and sum the final hidden states with the input embeddings during the final phase
+            of each block. Set to `True` for the first block of each stage.
+        adjust_padding (`bool`):
+            Whether to apply padding to only right and bottom side of the input kernel before the depthwise convolution
+            operation, set to `True` for inputs with odd input sizes.
+    """
+
+    def __init__(
+        self,
+        config: AlignVisionConfig,
+        in_dim: int,
+        out_dim: int,
+        stride: int,
+        expand_ratio: int,
+        kernel_size: int,
+        drop_rate: float,
+        id_skip: bool,
+        adjust_padding: bool,
+    ):
+        super().__init__()
+        self.expand_ratio = expand_ratio
+        self.expand = True if self.expand_ratio != 1 else False
+        expand_in_dim = in_dim * expand_ratio
+
+        if self.expand:
+            self.expansion = AlignVisionExpansionLayer(
+                config=config, in_dim=in_dim, out_dim=expand_in_dim, stride=stride
+            )
+
+        self.depthwise_conv = AlignVisionDepthwiseLayer(
+            config=config,
+            in_dim=expand_in_dim if self.expand else in_dim,
+            stride=stride,
+            kernel_size=kernel_size,
+            adjust_padding=adjust_padding,
+        )
+        self.squeeze_excite = AlignVisionSqueezeExciteLayer(
+            config=config, in_dim=in_dim, expand_dim=expand_in_dim, expand=self.expand
+        )
+        self.projection = AlignVisionFinalBlockLayer(
+            config=config,
+            in_dim=expand_in_dim if self.expand else in_dim,
+            out_dim=out_dim,
+            stride=stride,
+            drop_rate=drop_rate,
+            id_skip=id_skip,
+        )
+
+    def forward(self, hidden_states: torch.FloatTensor) -> torch.Tensor:
+        embeddings = hidden_states
+        # Expansion and depthwise convolution phase
+        if self.expand_ratio != 1:
+            hidden_states = self.expansion(hidden_states)
+        hidden_states = self.depthwise_conv(hidden_states)
+
+        # Squeeze and excite phase
+        hidden_states = self.squeeze_excite(hidden_states)
+        hidden_states = self.projection(embeddings, hidden_states)
+        return hidden_states
+
+
+class AlignVisionEncoder(nn.Module):
+    r"""
+    Forward propogates the embeddings through each vision encoder (EfficientNet) block.
+
+    Args:
+        config ([`AlignVisionConfig`]):
+            Model configuration class.
+    """
+
+    def __init__(self, config: AlignVisionConfig):
+        super().__init__()
+        self.depth_coefficient = config.depth_coefficient
+
+        def round_repeats(repeats):
+            # Round number of block repeats based on depth multiplier.
+            return int(math.ceil(self.depth_coefficient * repeats))
+
+        num_base_blocks = len(config.in_channels)
+        num_blocks = sum(round_repeats(n) for n in config.num_block_repeats)
+
+        curr_block_num = 0
+        blocks = []
+        for i in range(num_base_blocks):
+            in_dim = round_filters(config, config.in_channels[i])
+            out_dim = round_filters(config, config.out_channels[i])
+            stride = config.strides[i]
+            kernel_size = config.kernel_sizes[i]
+            expand_ratio = config.expand_ratios[i]
+
+            for j in range(round_repeats(config.num_block_repeats[i])):
+                id_skip = True if j == 0 else False
+                stride = 1 if j > 0 else stride
+                in_dim = out_dim if j > 0 else in_dim
+                adjust_padding = False if curr_block_num in config.depthwise_padding else True
+                drop_rate = config.drop_connect_rate * curr_block_num / num_blocks
+
+                block = AlignVisionBlock(
+                    config=config,
+                    in_dim=in_dim,
+                    out_dim=out_dim,
+                    stride=stride,
+                    kernel_size=kernel_size,
+                    expand_ratio=expand_ratio,
+                    drop_rate=drop_rate,
+                    id_skip=id_skip,
+                    adjust_padding=adjust_padding,
+                )
+                blocks.append(block)
+                curr_block_num += 1
+
+        self.blocks = nn.ModuleList(blocks)
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        output_hidden_states: Optional[bool] = False,
+        return_dict: Optional[bool] = True,
+    ) -> BaseModelOutputWithPoolingAndNoAttention:
+        all_hidden_states = (hidden_states,) if output_hidden_states else None
+
+        for block in self.blocks:
+            hidden_states = block(hidden_states)
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, all_hidden_states] if v is not None)
+
+        return BaseModelOutputWithNoAttention(
+            last_hidden_state=hidden_states,
+            hidden_states=all_hidden_states,
+        )
+
+
+# Copied from transformers.models.bert.modeling_bert.BertEmbeddings with Bert->AlignText
+class AlignTextEmbeddings(nn.Module):
+    """Construct the embeddings from word, position and token_type embeddings."""
+
+    def __init__(self, config):
+        super().__init__()
+        self.word_embeddings = nn.Embedding(config.vocab_size, config.hidden_size, padding_idx=config.pad_token_id)
+        self.position_embeddings = nn.Embedding(config.max_position_embeddings, config.hidden_size)
+        self.token_type_embeddings = nn.Embedding(config.type_vocab_size, config.hidden_size)
+
+        # self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load
+        # any TensorFlow checkpoint file
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        # position_ids (1, len position emb) is contiguous in memory and exported when serialized
+        self.position_embedding_type = getattr(config, "position_embedding_type", "absolute")
+        self.register_buffer("position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)))
+        self.register_buffer(
+            "token_type_ids", torch.zeros(self.position_ids.size(), dtype=torch.long), persistent=False
+        )
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        past_key_values_length: int = 0,
+    ) -> torch.Tensor:
+        if input_ids is not None:
+            input_shape = input_ids.size()
+        else:
+            input_shape = inputs_embeds.size()[:-1]
+
+        seq_length = input_shape[1]
+
+        if position_ids is None:
+            position_ids = self.position_ids[:, past_key_values_length : seq_length + past_key_values_length]
+
+        # Setting the token_type_ids to the registered buffer in constructor where it is all zeros, which usually occurs
+        # when its auto-generated, registered buffer helps users when tracing the model without passing token_type_ids, solves
+        # issue #5664
+        if token_type_ids is None:
+            if hasattr(self, "token_type_ids"):
+                buffered_token_type_ids = self.token_type_ids[:, :seq_length]
+                buffered_token_type_ids_expanded = buffered_token_type_ids.expand(input_shape[0], seq_length)
+                token_type_ids = buffered_token_type_ids_expanded
+            else:
+                token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=self.position_ids.device)
+
+        if inputs_embeds is None:
+            inputs_embeds = self.word_embeddings(input_ids)
+        token_type_embeddings = self.token_type_embeddings(token_type_ids)
+
+        embeddings = inputs_embeds + token_type_embeddings
+        if self.position_embedding_type == "absolute":
+            position_embeddings = self.position_embeddings(position_ids)
+            embeddings += position_embeddings
+        embeddings = self.LayerNorm(embeddings)
+        embeddings = self.dropout(embeddings)
+        return embeddings
+
+
+# Copied from transformers.models.bert.modeling_bert.BertSelfAttention with Bert->AlignText
+class AlignTextSelfAttention(nn.Module):
+    def __init__(self, config, position_embedding_type=None):
+        super().__init__()
+        if config.hidden_size % config.num_attention_heads != 0 and not hasattr(config, "embedding_size"):
+            raise ValueError(
+                f"The hidden size ({config.hidden_size}) is not a multiple of the number of attention "
+                f"heads ({config.num_attention_heads})"
+            )
+
+        self.num_attention_heads = config.num_attention_heads
+        self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
+        self.all_head_size = self.num_attention_heads * self.attention_head_size
+
+        self.query = nn.Linear(config.hidden_size, self.all_head_size)
+        self.key = nn.Linear(config.hidden_size, self.all_head_size)
+        self.value = nn.Linear(config.hidden_size, self.all_head_size)
+
+        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
+        self.position_embedding_type = position_embedding_type or getattr(
+            config, "position_embedding_type", "absolute"
+        )
+        if self.position_embedding_type == "relative_key" or self.position_embedding_type == "relative_key_query":
+            self.max_position_embeddings = config.max_position_embeddings
+            self.distance_embedding = nn.Embedding(2 * config.max_position_embeddings - 1, self.attention_head_size)
+
+        self.is_decoder = config.is_decoder
+
+    def transpose_for_scores(self, x: torch.Tensor) -> torch.Tensor:
+        new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size)
+        x = x.view(new_x_shape)
+        return x.permute(0, 2, 1, 3)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        past_key_value: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.Tensor]:
+        mixed_query_layer = self.query(hidden_states)
+
+        # If this is instantiated as a cross-attention module, the keys
+        # and values come from an encoder; the attention mask needs to be
+        # such that the encoder's padding tokens are not attended to.
+        is_cross_attention = encoder_hidden_states is not None
+
+        if is_cross_attention and past_key_value is not None:
+            # reuse k,v, cross_attentions
+            key_layer = past_key_value[0]
+            value_layer = past_key_value[1]
+            attention_mask = encoder_attention_mask
+        elif is_cross_attention:
+            key_layer = self.transpose_for_scores(self.key(encoder_hidden_states))
+            value_layer = self.transpose_for_scores(self.value(encoder_hidden_states))
+            attention_mask = encoder_attention_mask
+        elif past_key_value is not None:
+            key_layer = self.transpose_for_scores(self.key(hidden_states))
+            value_layer = self.transpose_for_scores(self.value(hidden_states))
+            key_layer = torch.cat([past_key_value[0], key_layer], dim=2)
+            value_layer = torch.cat([past_key_value[1], value_layer], dim=2)
+        else:
+            key_layer = self.transpose_for_scores(self.key(hidden_states))
+            value_layer = self.transpose_for_scores(self.value(hidden_states))
+
+        query_layer = self.transpose_for_scores(mixed_query_layer)
+
+        use_cache = past_key_value is not None
+        if self.is_decoder:
+            # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
+            # Further calls to cross_attention layer can then reuse all cross-attention
+            # key/value_states (first "if" case)
+            # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
+            # all previous decoder key/value_states. Further calls to uni-directional self-attention
+            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
+            # if encoder bi-directional self-attention `past_key_value` is always `None`
+            past_key_value = (key_layer, value_layer)
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
+
+        if self.position_embedding_type == "relative_key" or self.position_embedding_type == "relative_key_query":
+            query_length, key_length = query_layer.shape[2], key_layer.shape[2]
+            if use_cache:
+                position_ids_l = torch.tensor(key_length - 1, dtype=torch.long, device=hidden_states.device).view(
+                    -1, 1
+                )
+            else:
+                position_ids_l = torch.arange(query_length, dtype=torch.long, device=hidden_states.device).view(-1, 1)
+            position_ids_r = torch.arange(key_length, dtype=torch.long, device=hidden_states.device).view(1, -1)
+            distance = position_ids_l - position_ids_r
+
+            positional_embedding = self.distance_embedding(distance + self.max_position_embeddings - 1)
+            positional_embedding = positional_embedding.to(dtype=query_layer.dtype)  # fp16 compatibility
+
+            if self.position_embedding_type == "relative_key":
+                relative_position_scores = torch.einsum("bhld,lrd->bhlr", query_layer, positional_embedding)
+                attention_scores = attention_scores + relative_position_scores
+            elif self.position_embedding_type == "relative_key_query":
+                relative_position_scores_query = torch.einsum("bhld,lrd->bhlr", query_layer, positional_embedding)
+                relative_position_scores_key = torch.einsum("bhrd,lrd->bhlr", key_layer, positional_embedding)
+                attention_scores = attention_scores + relative_position_scores_query + relative_position_scores_key
+
+        attention_scores = attention_scores / math.sqrt(self.attention_head_size)
+        if attention_mask is not None:
+            # Apply the attention mask is (precomputed for all layers in AlignTextModel forward() function)
+            attention_scores = attention_scores + attention_mask
+
+        # Normalize the attention scores to probabilities.
+        attention_probs = nn.functional.softmax(attention_scores, dim=-1)
+
+        # This is actually dropping out entire tokens to attend to, which might
+        # seem a bit unusual, but is taken from the original Transformer paper.
+        attention_probs = self.dropout(attention_probs)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            attention_probs = attention_probs * head_mask
+
+        context_layer = torch.matmul(attention_probs, value_layer)
+
+        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
+        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
+        context_layer = context_layer.view(new_context_layer_shape)
+
+        outputs = (context_layer, attention_probs) if output_attentions else (context_layer,)
+
+        if self.is_decoder:
+            outputs = outputs + (past_key_value,)
+        return outputs
+
+
+# Copied from transformers.models.bert.modeling_bert.BertSelfOutput with Bert->AlignText
+class AlignTextSelfOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+# Copied from transformers.models.bert.modeling_bert.BertAttention with Bert->AlignText
+class AlignTextAttention(nn.Module):
+    def __init__(self, config, position_embedding_type=None):
+        super().__init__()
+        self.self = AlignTextSelfAttention(config, position_embedding_type=position_embedding_type)
+        self.output = AlignTextSelfOutput(config)
+        self.pruned_heads = set()
+
+    def prune_heads(self, heads):
+        if len(heads) == 0:
+            return
+        heads, index = find_pruneable_heads_and_indices(
+            heads, self.self.num_attention_heads, self.self.attention_head_size, self.pruned_heads
+        )
+
+        # Prune linear layers
+        self.self.query = prune_linear_layer(self.self.query, index)
+        self.self.key = prune_linear_layer(self.self.key, index)
+        self.self.value = prune_linear_layer(self.self.value, index)
+        self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)
+
+        # Update hyper params and store pruned heads
+        self.self.num_attention_heads = self.self.num_attention_heads - len(heads)
+        self.self.all_head_size = self.self.attention_head_size * self.self.num_attention_heads
+        self.pruned_heads = self.pruned_heads.union(heads)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        past_key_value: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.Tensor]:
+        self_outputs = self.self(
+            hidden_states,
+            attention_mask,
+            head_mask,
+            encoder_hidden_states,
+            encoder_attention_mask,
+            past_key_value,
+            output_attentions,
+        )
+        attention_output = self.output(self_outputs[0], hidden_states)
+        outputs = (attention_output,) + self_outputs[1:]  # add attentions if we output them
+        return outputs
+
+
+# Copied from transformers.models.bert.modeling_bert.BertIntermediate with Bert->AlignText
+class AlignTextIntermediate(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.intermediate_size)
+        if isinstance(config.hidden_act, str):
+            self.intermediate_act_fn = ACT2FN[config.hidden_act]
+        else:
+            self.intermediate_act_fn = config.hidden_act
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.intermediate_act_fn(hidden_states)
+        return hidden_states
+
+
+# Copied from transformers.models.bert.modeling_bert.BertOutput with Bert->AlignText
+class AlignTextOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+# Copied from transformers.models.bert.modeling_bert.BertLayer with Bert->AlignText
+class AlignTextLayer(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.chunk_size_feed_forward = config.chunk_size_feed_forward
+        self.seq_len_dim = 1
+        self.attention = AlignTextAttention(config)
+        self.is_decoder = config.is_decoder
+        self.add_cross_attention = config.add_cross_attention
+        if self.add_cross_attention:
+            if not self.is_decoder:
+                raise ValueError(f"{self} should be used as a decoder model if cross attention is added")
+            self.crossattention = AlignTextAttention(config, position_embedding_type="absolute")
+        self.intermediate = AlignTextIntermediate(config)
+        self.output = AlignTextOutput(config)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        past_key_value: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.Tensor]:
+        # decoder uni-directional self-attention cached key/values tuple is at positions 1,2
+        self_attn_past_key_value = past_key_value[:2] if past_key_value is not None else None
+        self_attention_outputs = self.attention(
+            hidden_states,
+            attention_mask,
+            head_mask,
+            output_attentions=output_attentions,
+            past_key_value=self_attn_past_key_value,
+        )
+        attention_output = self_attention_outputs[0]
+
+        # if decoder, the last output is tuple of self-attn cache
+        if self.is_decoder:
+            outputs = self_attention_outputs[1:-1]
+            present_key_value = self_attention_outputs[-1]
+        else:
+            outputs = self_attention_outputs[1:]  # add self attentions if we output attention weights
+
+        cross_attn_present_key_value = None
+        if self.is_decoder and encoder_hidden_states is not None:
+            if not hasattr(self, "crossattention"):
+                raise ValueError(
+                    f"If `encoder_hidden_states` are passed, {self} has to be instantiated with cross-attention layers"
+                    " by setting `config.add_cross_attention=True`"
+                )
+
+            # cross_attn cached key/values tuple is at positions 3,4 of past_key_value tuple
+            cross_attn_past_key_value = past_key_value[-2:] if past_key_value is not None else None
+            cross_attention_outputs = self.crossattention(
+                attention_output,
+                attention_mask,
+                head_mask,
+                encoder_hidden_states,
+                encoder_attention_mask,
+                cross_attn_past_key_value,
+                output_attentions,
+            )
+            attention_output = cross_attention_outputs[0]
+            outputs = outputs + cross_attention_outputs[1:-1]  # add cross attentions if we output attention weights
+
+            # add cross-attn cache to positions 3,4 of present_key_value tuple
+            cross_attn_present_key_value = cross_attention_outputs[-1]
+            present_key_value = present_key_value + cross_attn_present_key_value
+
+        layer_output = apply_chunking_to_forward(
+            self.feed_forward_chunk, self.chunk_size_feed_forward, self.seq_len_dim, attention_output
+        )
+        outputs = (layer_output,) + outputs
+
+        # if decoder, return the attn key/values as the last output
+        if self.is_decoder:
+            outputs = outputs + (present_key_value,)
+
+        return outputs
+
+    def feed_forward_chunk(self, attention_output):
+        intermediate_output = self.intermediate(attention_output)
+        layer_output = self.output(intermediate_output, attention_output)
+        return layer_output
+
+
+# Copied from transformers.models.bert.modeling_bert.BertEncoder with Bert->AlignText
+class AlignTextEncoder(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.layer = nn.ModuleList([AlignTextLayer(config) for _ in range(config.num_hidden_layers)])
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = False,
+        output_hidden_states: Optional[bool] = False,
+        return_dict: Optional[bool] = True,
+    ) -> Union[Tuple[torch.Tensor], BaseModelOutputWithPastAndCrossAttentions]:
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attentions = () if output_attentions else None
+        all_cross_attentions = () if output_attentions and self.config.add_cross_attention else None
+
+        next_decoder_cache = () if use_cache else None
+        for i, layer_module in enumerate(self.layer):
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            layer_head_mask = head_mask[i] if head_mask is not None else None
+            past_key_value = past_key_values[i] if past_key_values is not None else None
+
+            if self.gradient_checkpointing and self.training:
+                if use_cache:
+                    logger.warning_once(
+                        "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                    )
+                    use_cache = False
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs, past_key_value, output_attentions)
+
+                    return custom_forward
+
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(layer_module),
+                    hidden_states,
+                    attention_mask,
+                    layer_head_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                )
+            else:
+                layer_outputs = layer_module(
+                    hidden_states,
+                    attention_mask,
+                    layer_head_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    past_key_value,
+                    output_attentions,
+                )
+
+            hidden_states = layer_outputs[0]
+            if use_cache:
+                next_decoder_cache += (layer_outputs[-1],)
+            if output_attentions:
+                all_self_attentions = all_self_attentions + (layer_outputs[1],)
+                if self.config.add_cross_attention:
+                    all_cross_attentions = all_cross_attentions + (layer_outputs[2],)
+
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(
+                v
+                for v in [
+                    hidden_states,
+                    next_decoder_cache,
+                    all_hidden_states,
+                    all_self_attentions,
+                    all_cross_attentions,
+                ]
+                if v is not None
+            )
+        return BaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            past_key_values=next_decoder_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attentions,
+            cross_attentions=all_cross_attentions,
+        )
+
+
+# Copied from transformers.models.bert.modeling_bert.BertPooler with Bert -> AlignText
+class AlignTextPooler(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.activation = nn.Tanh()
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        # We "pool" the model by simply taking the hidden state corresponding
+        # to the first token.
+        first_token_tensor = hidden_states[:, 0]
+        pooled_output = self.dense(first_token_tensor)
+        pooled_output = self.activation(pooled_output)
+        return pooled_output
+
+
+class AlignPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = AlignConfig
+    base_model_prefix = "align"
+    supports_gradient_checkpointing = True
+    _keys_to_ignore_on_load_missing = [r"position_ids"]
+
+    def _init_weights(self, module):
+        """Initialize the weights"""
+        if isinstance(module, (nn.Linear, nn.Conv2d)):
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, AlignModel):
+            nn.init.xavier_uniform_(module.text_projection.weight)
+            module.text_projection.bias.data.zero_()
+            module.text_projection._is_hf_initialized = True
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+        if isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, (AlignTextModel, AlignVisionModel)):
+            module.gradient_checkpointing = value
+
+
+@add_start_docstrings(
+    """The text model from ALIGN without any head or projection on top.""",
+    ALIGN_START_DOCSTRING,
+)
+class AlignTextModel(AlignPreTrainedModel):
+    config_class = AlignTextConfig
+
+    def __init__(self, config: AlignTextConfig, add_pooling_layer: bool = True):
+        super().__init__(config)
+        self.config = config
+
+        self.embeddings = AlignTextEmbeddings(config)
+        self.encoder = AlignTextEncoder(config)
+
+        self.pooler = AlignTextPooler(config) if add_pooling_layer else None
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embeddings.word_embeddings
+
+    def set_input_embeddings(self, value):
+        self.embeddings.word_embeddings = value
+
+    @add_start_docstrings_to_model_forward(ALIGN_TEXT_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=BaseModelOutputWithPoolingAndCrossAttentions, config_class=AlignTextConfig)
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPoolingAndCrossAttentions]:
+        r"""
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> from transformers import AutoTokenizer, AlignTextModel
+
+        >>> model = AlignTextModel.from_pretrained("kakaobrain/align-base")
+        >>> tokenizer = AutoTokenizer.from_pretrained("kakaobrain/align-base")
+
+        >>> inputs = tokenizer(["a photo of a cat", "a photo of a dog"], padding=True, return_tensors="pt")
+
+        >>> outputs = model(**inputs)
+        >>> last_hidden_state = outputs.last_hidden_state
+        >>> pooled_output = outputs.pooler_output  # pooled (EOS token) states
+        ```"""
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        batch_size, seq_length = input_shape
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+
+        if attention_mask is None:
+            attention_mask = torch.ones(((batch_size, seq_length)), device=device)
+
+        if token_type_ids is None:
+            if hasattr(self.embeddings, "token_type_ids"):
+                buffered_token_type_ids = self.embeddings.token_type_ids[:, :seq_length]
+                buffered_token_type_ids_expanded = buffered_token_type_ids.expand(batch_size, seq_length)
+                token_type_ids = buffered_token_type_ids_expanded
+            else:
+                token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
+
+        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
+        # ourselves in which case we just need to make it broadcastable to all heads.
+        extended_attention_mask: torch.Tensor = self.get_extended_attention_mask(attention_mask, input_shape)
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
+        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
+        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
+
+        embedding_output = self.embeddings(
+            input_ids=input_ids,
+            position_ids=position_ids,
+            token_type_ids=token_type_ids,
+            inputs_embeds=inputs_embeds,
+        )
+        encoder_outputs = self.encoder(
+            embedding_output,
+            attention_mask=extended_attention_mask,
+            head_mask=head_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        sequence_output = encoder_outputs[0]
+        pooled_output = self.pooler(sequence_output) if self.pooler is not None else None
+
+        if not return_dict:
+            return (sequence_output, pooled_output) + encoder_outputs[1:]
+
+        return BaseModelOutputWithPoolingAndCrossAttentions(
+            last_hidden_state=sequence_output,
+            pooler_output=pooled_output,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+            cross_attentions=encoder_outputs.cross_attentions,
+        )
+
+
+@add_start_docstrings(
+    """The vision model from ALIGN without any head or projection on top.""",
+    ALIGN_START_DOCSTRING,
+)
+class AlignVisionModel(AlignPreTrainedModel):
+    config_class = AlignVisionConfig
+    main_input_name = "pixel_values"
+
+    def __init__(self, config: AlignVisionConfig):
+        super().__init__(config)
+        self.config = config
+        self.embeddings = AlignVisionEmbeddings(config)
+        self.encoder = AlignVisionEncoder(config)
+
+        # Final pooling layer
+        if config.pooling_type == "mean":
+            self.pooler = nn.AvgPool2d(config.hidden_dim, ceil_mode=True)
+        elif config.pooling_type == "max":
+            self.pooler = nn.MaxPool2d(config.hidden_dim, ceil_mode=True)
+        else:
+            raise ValueError(f"config.pooling must be one of ['mean', 'max'] got {config.pooling}")
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self) -> nn.Module:
+        return self.vision_model.embeddings.convolution
+
+    @add_start_docstrings_to_model_forward(ALIGN_VISION_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=BaseModelOutputWithPoolingAndNoAttention, config_class=AlignVisionConfig)
+    def forward(
+        self,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPoolingAndNoAttention]:
+        r"""
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> from PIL import Image
+        >>> import requests
+        >>> from transformers import AutoProcessor, AlignVisionModel
+
+        >>> model = AlignVisionModel.from_pretrained("kakaobrain/align-base")
+        >>> processor = AutoProcessor.from_pretrained("kakaobrain/align-base")
+
+        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        >>> image = Image.open(requests.get(url, stream=True).raw)
+
+        >>> inputs = processor(images=image, return_tensors="pt")
+
+        >>> outputs = model(**inputs)
+        >>> last_hidden_state = outputs.last_hidden_state
+        >>> pooled_output = outputs.pooler_output  # pooled CLS states
+        ```"""
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if pixel_values is None:
+            raise ValueError("You have to specify pixel_values")
+
+        embedding_output = self.embeddings(pixel_values)
+        encoder_outputs = self.encoder(
+            embedding_output,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        # Apply pooling
+        last_hidden_state = encoder_outputs[0]
+        pooled_output = self.pooler(last_hidden_state)
+        # Reshape (batch_size, projection_dim, 1 , 1) -> (batch_size, projection_dim)
+        pooled_output = pooled_output.reshape(pooled_output.shape[:2])
+
+        if not return_dict:
+            return (last_hidden_state, pooled_output) + encoder_outputs[1:]
+
+        return BaseModelOutputWithPoolingAndNoAttention(
+            last_hidden_state=last_hidden_state,
+            pooler_output=pooled_output,
+            hidden_states=encoder_outputs.hidden_states,
+        )
+
+
+@add_start_docstrings(ALIGN_START_DOCSTRING)
+class AlignModel(AlignPreTrainedModel):
+    config_class = AlignConfig
+
+    def __init__(self, config: AlignConfig):
+        super().__init__(config)
+
+        if not isinstance(config.text_config, AlignTextConfig):
+            raise ValueError(
+                "config.text_config is expected to be of type AlignTextConfig but is of type"
+                f" {type(config.text_config)}."
+            )
+
+        if not isinstance(config.vision_config, AlignVisionConfig):
+            raise ValueError(
+                "config.vision_config is expected to be of type AlignVisionConfig but is of type"
+                f" {type(config.vision_config)}."
+            )
+
+        text_config = config.text_config
+        vision_config = config.vision_config
+
+        self.projection_dim = config.projection_dim
+        self.text_embed_dim = text_config.hidden_size
+
+        self.text_model = AlignTextModel(text_config)
+        self.vision_model = AlignVisionModel(vision_config)
+
+        self.text_projection = nn.Linear(self.text_embed_dim, self.projection_dim)
+        self.temperature = nn.Parameter(torch.ones([]) * self.config.temperature_init_value)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @add_start_docstrings_to_model_forward(ALIGN_TEXT_INPUTS_DOCSTRING)
+    def get_text_features(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> torch.FloatTensor:
+        r"""
+        Returns:
+            text_features (`torch.FloatTensor` of shape `(batch_size, output_dim`): The text embeddings obtained by
+            applying the projection layer to the pooled output of [`AlignTextModel`].
+
+        Examples:
+
+        ```python
+        >>> from transformers import AutoTokenizer, AlignModel
+
+        >>> model = AlignModel.from_pretrained("kakaobrain/align-base")
+        >>> tokenizer = AutoTokenizer.from_pretrained("kakaobrain/align-base")
+
+        >>> inputs = tokenizer(["a photo of a cat", "a photo of a dog"], padding=True, return_tensors="pt")
+        >>> text_features = model.get_text_features(**inputs)
+        ```"""
+        # Use ALIGN model's config for some fields (if specified) instead of those of vision & text components.
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        text_outputs = self.text_model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        last_hidden_state = text_outputs[0][:, 0, :]
+        text_features = self.text_projection(last_hidden_state)
+
+        return text_features
+
+    @add_start_docstrings_to_model_forward(ALIGN_VISION_INPUTS_DOCSTRING)
+    def get_image_features(
+        self,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> torch.FloatTensor:
+        r"""
+        Returns:
+            image_features (`torch.FloatTensor` of shape `(batch_size, output_dim`): The image embeddings obtained by
+            applying the projection layer to the pooled output of [`AlignVisionModel`].
+
+        Examples:
+
+        ```python
+        >>> from PIL import Image
+        >>> import requests
+        >>> from transformers import AutoProcessor, AlignModel
+
+        >>> model = AlignModel.from_pretrained("kakaobrain/align-base")
+        >>> processor = AutoProcessor.from_pretrained("kakaobrain/align-base")
+
+        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        >>> image = Image.open(requests.get(url, stream=True).raw)
+
+        >>> inputs = processor(images=image, return_tensors="pt")
+
+        >>> image_features = model.get_image_features(**inputs)
+        ```"""
+        # Use ALIGN model's config for some fields (if specified) instead of those of vision & text components.
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        vision_outputs = self.vision_model(
+            pixel_values=pixel_values,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        image_features = vision_outputs[1]  # pooled_output
+
+        return image_features
+
+    @add_start_docstrings_to_model_forward(ALIGN_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=AlignOutput, config_class=AlignConfig)
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        return_loss: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, AlignOutput]:
+        r"""
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> from PIL import Image
+        >>> import requests
+        >>> from transformers import AutoProcessor, AlignModel
+
+        >>> model = AlignModel.from_pretrained("kakaobrain/align-base")
+        >>> processor = AutoProcessor.from_pretrained("kakaobrain/align-base")
+
+        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        >>> image = Image.open(requests.get(url, stream=True).raw)
+
+        >>> inputs = processor(
+        ...     text=["a photo of a cat", "a photo of a dog"], images=image, return_tensors="pt", padding=True
+        ... )
+
+        >>> outputs = model(**inputs)
+        >>> logits_per_image = outputs.logits_per_image  # this is the image-text similarity score
+        >>> probs = logits_per_image.softmax(dim=1)  # we can take the softmax to get the label probabilities
+        ```"""
+        # Use ALIGN model's config for some fields (if specified) instead of those of vision & text components.
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        vision_outputs = self.vision_model(
+            pixel_values=pixel_values,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        text_outputs = self.text_model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        image_embeds = vision_outputs[1]
+        text_embeds = text_outputs[0][:, 0, :]
+        text_embeds = self.text_projection(text_embeds)
+
+        # normalized features
+        image_embeds = image_embeds / image_embeds.norm(p=2, dim=-1, keepdim=True)
+        text_embeds = text_embeds / text_embeds.norm(p=2, dim=-1, keepdim=True)
+
+        # cosine similarity as logits
+        logits_per_text = torch.matmul(text_embeds, image_embeds.t()) / self.temperature
+        logits_per_image = logits_per_text.t()
+
+        loss = None
+        if return_loss:
+            loss = align_loss(logits_per_text)
+
+        if not return_dict:
+            output = (logits_per_image, logits_per_text, text_embeds, image_embeds, text_outputs, vision_outputs)
+            return ((loss,) + output) if loss is not None else output
+
+        return AlignOutput(
+            loss=loss,
+            logits_per_image=logits_per_image,
+            logits_per_text=logits_per_text,
+            text_embeds=text_embeds,
+            image_embeds=image_embeds,
+            text_model_output=text_outputs,
+            vision_model_output=vision_outputs,
+        )