Add Data2Vec (#15507)

* Add data2vec model cloned from roberta

* Add checkpoint conversion script

* Fix copies

* Update docs

* Add checkpoint conversion script

* Remove fairseq data2vec_text script and fix format

* Add comment on where to get data2vec_text.py

* Remove mock implementation cheat.py and fix style

* Fix copies

* Remove TF and Flax classes from init

* Add back copy from fairseq data2vec_text.py and fix style

* Update model name in docs/source/index.mdx to be CamelCase

* Revert model name in table to lower-case to get check_table test to pass

* Update src/transformers/models/data2vec/__init__.py
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* Update src/transformers/models/data2vec/convert_data2vec_original_pytorch_checkpoint_to_pytorch.py
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* Update src/transformers/models/data2vec/modeling_data2vec.py
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* Update src/transformers/models/data2vec/modeling_data2vec.py
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* Update src/transformers/models/data2vec/modeling_data2vec.py
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* Update src/transformers/models/data2vec/modeling_data2vec.py
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* Update docs/source/model_doc/data2vec.mdx
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update docs/source/model_doc/data2vec.mdx
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/auto/configuration_auto.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/data2vec/configuration_data2vec.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/data2vec/modeling_data2vec.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/data2vec/modeling_data2vec.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/data2vec/modeling_data2vec.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update tests/test_modeling_data2vec.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/data2vec/configuration_data2vec.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/data2vec/modeling_data2vec.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update documentation

* Copy-paste Data2VecConfig from BertConfig

* Update config checkpoint to point to edugp/data2vec-nlp-base. Fix style and repo-consistency

* Update config special tokens to match RoBERTa

* Split multiple assertions and add individual error messages

* Rename Data2VecModel to Data2VecForTextModel

* Add Data2Vec to _toctree.yml

* Rename Data2VecEmbeddings to Data2VecForTextEmbeddings

* Add initial Data2VecForAudio model (unfinished). Only matching fairseq's implementation up to the feature encoder (before positional encoding).

* finish audio model

* finish audio file

* Update names and fix style, quality and repo consistency

* Remove Data2VecAudioForPretraining. Add tests for Data2VecAudio, mimicking the Wav2Vec2 test suite. Fix bias initilization in positional conv layers. Move back configurations for audio and text to separate files.

* add inputs to logits to data2vec'

* correct autio models

* correct config auto

* correct tok auto

* Update utils/tests_fetcher.py

* delete unnecessary files

* delete unnecessary files

* further renaming

* make all tests pass

* finish

* remove useless test file

* Update tests/test_modeling_common.py

* Update utils/check_repo.py
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* Update src/transformers/models/data2vec/modeling_data2vec_text.py
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* Fix copies

* Update docs

* Remove fairseq data2vec_text script and fix format

* Add comment on where to get data2vec_text.py

* Remove mock implementation cheat.py and fix style

* Fix copies

* Remove TF and Flax classes from init

* Add back copy from fairseq data2vec_text.py and fix style

* Update model name in docs/source/index.mdx to be CamelCase

* Revert model name in table to lower-case to get check_table test to pass

* Update documentation

* Update src/transformers/models/data2vec/__init__.py
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* Update src/transformers/models/data2vec/convert_data2vec_original_pytorch_checkpoint_to_pytorch.py
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* Update src/transformers/models/data2vec/modeling_data2vec.py
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* Update src/transformers/models/data2vec/modeling_data2vec.py
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* Update src/transformers/models/data2vec/modeling_data2vec.py
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* Update src/transformers/models/data2vec/modeling_data2vec.py
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* Update src/transformers/models/auto/configuration_auto.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/data2vec/configuration_data2vec.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/data2vec/modeling_data2vec.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/data2vec/modeling_data2vec.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/data2vec/modeling_data2vec.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update tests/test_modeling_data2vec.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/data2vec/configuration_data2vec.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/data2vec/modeling_data2vec.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Copy-paste Data2VecConfig from BertConfig

* Update config checkpoint to point to edugp/data2vec-nlp-base. Fix style and repo-consistency

* Update config special tokens to match RoBERTa

* Split multiple assertions and add individual error messages

* Rename Data2VecModel to Data2VecForTextModel

* Add Data2Vec to _toctree.yml

* Rename Data2VecEmbeddings to Data2VecForTextEmbeddings

* Add initial Data2VecForAudio model (unfinished). Only matching fairseq's implementation up to the feature encoder (before positional encoding).

* finish audio model

* finish audio file

* add inputs to logits to data2vec'

* Update names and fix style, quality and repo consistency

* Remove Data2VecAudioForPretraining. Add tests for Data2VecAudio, mimicking the Wav2Vec2 test suite. Fix bias initilization in positional conv layers. Move back configurations for audio and text to separate files.

* correct autio models

* correct config auto

* correct tok auto

* delete unnecessary files

* delete unnecessary files

* Update utils/tests_fetcher.py

* further renaming

* make all tests pass

* finish

* remove useless test file

* Update tests/test_modeling_common.py

* Update utils/check_repo.py
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* Update src/transformers/models/data2vec/modeling_data2vec_text.py
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* Move data2vec tests to new structure

* Fix test imports for text tests

* Remove fairseq files

* Change paper link to arxiv

* Modify Data2Vec documentation to reflect that the encoder is not shared across the audio and text models in the current implementation.

* Update text model checkpoint to be facebook/data2vec-text-base

* Add 'Copy from' statements and update paper links and docs

* fix copy from statements

* improve copied from

* correct more copied from statements

* finish copied from stuff

* make style

* add model to README

* add to master
Co-authored-by: Eduardo Gonzalez Ponferrada <eduardo@ferrumhealth.com>
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

README.md

@@ -249,6 +249,7 @@ Current number of checkpoints: ![](https://img.shields.io/endpoint?url=https://h
 1. **[ConvBERT](https://huggingface.co/docs/transformers/model_doc/convbert)** (from YituTech) released with the paper [ConvBERT: Improving BERT with Span-based Dynamic Convolution](https://arxiv.org/abs/2008.02496) by Zihang Jiang, Weihao Yu, Daquan Zhou, Yunpeng Chen, Jiashi Feng, Shuicheng Yan.
 1. **[CPM](https://huggingface.co/docs/transformers/model_doc/cpm)** (from Tsinghua University) released with the paper [CPM: A Large-scale Generative Chinese Pre-trained Language Model](https://arxiv.org/abs/2012.00413) by Zhengyan Zhang, Xu Han, Hao Zhou, Pei Ke, Yuxian Gu, Deming Ye, Yujia Qin, Yusheng Su, Haozhe Ji, Jian Guan, Fanchao Qi, Xiaozhi Wang, Yanan Zheng, Guoyang Zeng, Huanqi Cao, Shengqi Chen, Daixuan Li, Zhenbo Sun, Zhiyuan Liu, Minlie Huang, Wentao Han, Jie Tang, Juanzi Li, Xiaoyan Zhu, Maosong Sun.
 1. **[CTRL](https://huggingface.co/docs/transformers/model_doc/ctrl)** (from Salesforce) released with the paper [CTRL: A Conditional Transformer Language Model for Controllable Generation](https://arxiv.org/abs/1909.05858) by Nitish Shirish Keskar*, Bryan McCann*, Lav R. Varshney, Caiming Xiong and Richard Socher.
+1. **[Data2Vec](https://huggingface.co/docs/transformers/master/model_doc/data2vec)** (from Facebook) released with the paper [Data2Vec:  A General Framework for Self-supervised Learning in Speech, Vision and Language](https://arxiv.org/abs/2202.03555) by Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, Michael Auli.
 1. **[DeBERTa](https://huggingface.co/docs/transformers/model_doc/deberta)** (from Microsoft) released with the paper [DeBERTa: Decoding-enhanced BERT with Disentangled Attention](https://arxiv.org/abs/2006.03654) by Pengcheng He, Xiaodong Liu, Jianfeng Gao, Weizhu Chen.
 1. **[DeBERTa-v2](https://huggingface.co/docs/transformers/model_doc/deberta-v2)** (from Microsoft) released with the paper [DeBERTa: Decoding-enhanced BERT with Disentangled Attention](https://arxiv.org/abs/2006.03654) by Pengcheng He, Xiaodong Liu, Jianfeng Gao, Weizhu Chen.
 1. **[DeiT](https://huggingface.co/docs/transformers/model_doc/deit)** (from Facebook) released with the paper [Training data-efficient image transformers & distillation through attention](https://arxiv.org/abs/2012.12877) by Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, Hervé Jégou.

README_ko.md

@@ -230,6 +230,7 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
 1. **[ConvNeXT](https://huggingface.co/docs/transformers/master/model_doc/convnext)** (from Facebook AI) released with the paper [A ConvNet for the 2020s](https://arxiv.org/abs/2201.03545) by Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, Saining Xie.
 1. **[CPM](https://huggingface.co/docs/transformers/model_doc/cpm)** (from Tsinghua University) released with the paper [CPM: A Large-scale Generative Chinese Pre-trained Language Model](https://arxiv.org/abs/2012.00413) by Zhengyan Zhang, Xu Han, Hao Zhou, Pei Ke, Yuxian Gu, Deming Ye, Yujia Qin, Yusheng Su, Haozhe Ji, Jian Guan, Fanchao Qi, Xiaozhi Wang, Yanan Zheng, Guoyang Zeng, Huanqi Cao, Shengqi Chen, Daixuan Li, Zhenbo Sun, Zhiyuan Liu, Minlie Huang, Wentao Han, Jie Tang, Juanzi Li, Xiaoyan Zhu, Maosong Sun.
 1. **[CTRL](https://huggingface.co/docs/transformers/model_doc/ctrl)** (from Salesforce) released with the paper [CTRL: A Conditional Transformer Language Model for Controllable Generation](https://arxiv.org/abs/1909.05858) by Nitish Shirish Keskar*, Bryan McCann*, Lav R. Varshney, Caiming Xiong and Richard Socher.
+1. **[Data2Vec](https://huggingface.co/docs/transformers/master/model_doc/data2vec)** (from Facebook) released with the paper [Data2Vec:  A General Framework for Self-supervised Learning in Speech, Vision and Language](https://arxiv.org/abs/2202.03555) by Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, Michael Auli.
 1. **[DeBERTa](https://huggingface.co/docs/transformers/model_doc/deberta)** (from Microsoft) released with the paper [DeBERTa: Decoding-enhanced BERT with Disentangled Attention](https://arxiv.org/abs/2006.03654) by Pengcheng He, Xiaodong Liu, Jianfeng Gao, Weizhu Chen.
 1. **[DeBERTa-v2](https://huggingface.co/docs/transformers/model_doc/deberta-v2)** (from Microsoft) released with the paper [DeBERTa: Decoding-enhanced BERT with Disentangled Attention](https://arxiv.org/abs/2006.03654) by Pengcheng He, Xiaodong Liu, Jianfeng Gao, Weizhu Chen.
 1. **[DeiT](https://huggingface.co/docs/transformers/model_doc/deit)** (from Facebook) released with the paper [Training data-efficient image transformers & distillation through attention](https://arxiv.org/abs/2012.12877) by Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, Hervé Jégou.

README_zh-hans.md

@@ -254,6 +254,7 @@ conda install -c huggingface transformers
 1. **[ConvNeXT](https://huggingface.co/docs/transformers/master/model_doc/convnext)** (来自 Facebook AI) 伴随论文 [A ConvNet for the 2020s](https://arxiv.org/abs/2201.03545) 由 Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, Saining Xie 发布。
 1. **[CPM](https://huggingface.co/docs/transformers/model_doc/cpm)** (来自 Tsinghua University) 伴随论文 [CPM: A Large-scale Generative Chinese Pre-trained Language Model](https://arxiv.org/abs/2012.00413) 由 Zhengyan Zhang, Xu Han, Hao Zhou, Pei Ke, Yuxian Gu, Deming Ye, Yujia Qin, Yusheng Su, Haozhe Ji, Jian Guan, Fanchao Qi, Xiaozhi Wang, Yanan Zheng, Guoyang Zeng, Huanqi Cao, Shengqi Chen, Daixuan Li, Zhenbo Sun, Zhiyuan Liu, Minlie Huang, Wentao Han, Jie Tang, Juanzi Li, Xiaoyan Zhu, Maosong Sun 发布。
 1. **[CTRL](https://huggingface.co/docs/transformers/model_doc/ctrl)** (来自 Salesforce) 伴随论文 [CTRL: A Conditional Transformer Language Model for Controllable Generation](https://arxiv.org/abs/1909.05858) 由 Nitish Shirish Keskar*, Bryan McCann*, Lav R. Varshney, Caiming Xiong and Richard Socher 发布。
+1. **[Data2Vec](https://huggingface.co/docs/transformers/master/model_doc/data2vec)** (来自 Facebook) 伴随论文 [Data2Vec:  A General Framework for Self-supervised Learning in Speech, Vision and Language](https://arxiv.org/abs/2202.03555) 由 Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, Michael Auli 发布。
 1. **[DeBERTa](https://huggingface.co/docs/transformers/model_doc/deberta)** (来自 Microsoft) 伴随论文 [DeBERTa: Decoding-enhanced BERT with Disentangled Attention](https://arxiv.org/abs/2006.03654) 由 Pengcheng He, Xiaodong Liu, Jianfeng Gao, Weizhu Chen 发布。
 1. **[DeBERTa-v2](https://huggingface.co/docs/transformers/model_doc/deberta-v2)** (来自 Microsoft) 伴随论文 [DeBERTa: Decoding-enhanced BERT with Disentangled Attention](https://arxiv.org/abs/2006.03654) 由 Pengcheng He, Xiaodong Liu, Jianfeng Gao, Weizhu Chen 发布。
 1. **[DeiT](https://huggingface.co/docs/transformers/model_doc/deit)** (来自 Facebook) 伴随论文 [Training data-efficient image transformers & distillation through attention](https://arxiv.org/abs/2012.12877) 由 Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, Hervé Jégou 发布。

README_zh-hant.md

@@ -266,6 +266,7 @@ conda install -c huggingface transformers
 1. **[ConvNeXT](https://huggingface.co/docs/transformers/master/model_doc/convnext)** (from Facebook AI) released with the paper [A ConvNet for the 2020s](https://arxiv.org/abs/2201.03545) by Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, Saining Xie.
 1. **[CPM](https://huggingface.co/docs/transformers/model_doc/cpm)** (from Tsinghua University) released with the paper [CPM: A Large-scale Generative Chinese Pre-trained Language Model](https://arxiv.org/abs/2012.00413) by Zhengyan Zhang, Xu Han, Hao Zhou, Pei Ke, Yuxian Gu, Deming Ye, Yujia Qin, Yusheng Su, Haozhe Ji, Jian Guan, Fanchao Qi, Xiaozhi Wang, Yanan Zheng, Guoyang Zeng, Huanqi Cao, Shengqi Chen, Daixuan Li, Zhenbo Sun, Zhiyuan Liu, Minlie Huang, Wentao Han, Jie Tang, Juanzi Li, Xiaoyan Zhu, Maosong Sun.
 1. **[CTRL](https://huggingface.co/docs/transformers/model_doc/ctrl)** (from Salesforce) released with the paper [CTRL: A Conditional Transformer Language Model for Controllable Generation](https://arxiv.org/abs/1909.05858) by Nitish Shirish Keskar*, Bryan McCann*, Lav R. Varshney, Caiming Xiong and Richard Socher.
+1. **[Data2Vec](https://huggingface.co/docs/transformers/master/model_doc/data2vec)** (from Facebook) released with the paper [Data2Vec:  A General Framework for Self-supervised Learning in Speech, Vision and Language](https://arxiv.org/abs/2202.03555) by Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, Michael Auli.
 1. **[DeBERTa](https://huggingface.co/docs/transformers/model_doc/deberta)** (from Microsoft) released with the paper [DeBERTa: Decoding-enhanced BERT with Disentangled Attention](https://arxiv.org/abs/2006.03654) by Pengcheng He, Xiaodong Liu, Jianfeng Gao, Weizhu Chen.
 1. **[DeBERTa-v2](https://huggingface.co/docs/transformers/model_doc/deberta-v2)** (from Microsoft) released with the paper [DeBERTa: Decoding-enhanced BERT with Disentangled Attention](https://arxiv.org/abs/2006.03654) by Pengcheng He, Xiaodong Liu, Jianfeng Gao, Weizhu Chen.
 1. **[DeiT](https://huggingface.co/docs/transformers/model_doc/deit)** (from Facebook) released with the paper [Training data-efficient image transformers & distillation through attention](https://arxiv.org/abs/2012.12877) by Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, Hervé Jégou.

docs/source/_toctree.yml

@@ -178,6 +178,8 @@
       title: CPM
     - local: model_doc/ctrl
       title: CTRL
+    - local: model_doc/data2vec
+      title: Data2Vec
     - local: model_doc/deberta
       title: DeBERTa
     - local: model_doc/deberta-v2

docs/source/index.mdx

@@ -75,6 +75,7 @@ conversion utilities for the following models.
 1. **[ConvBERT](model_doc/convbert)** (from YituTech) released with the paper [ConvBERT: Improving BERT with Span-based Dynamic Convolution](https://arxiv.org/abs/2008.02496) by Zihang Jiang, Weihao Yu, Daquan Zhou, Yunpeng Chen, Jiashi Feng, Shuicheng Yan.
 1. **[CPM](model_doc/cpm)** (from Tsinghua University) released with the paper [CPM: A Large-scale Generative Chinese Pre-trained Language Model](https://arxiv.org/abs/2012.00413) by Zhengyan Zhang, Xu Han, Hao Zhou, Pei Ke, Yuxian Gu, Deming Ye, Yujia Qin, Yusheng Su, Haozhe Ji, Jian Guan, Fanchao Qi, Xiaozhi Wang, Yanan Zheng, Guoyang Zeng, Huanqi Cao, Shengqi Chen, Daixuan Li, Zhenbo Sun, Zhiyuan Liu, Minlie Huang, Wentao Han, Jie Tang, Juanzi Li, Xiaoyan Zhu, Maosong Sun.
 1. **[CTRL](model_doc/ctrl)** (from Salesforce) released with the paper [CTRL: A Conditional Transformer Language Model for Controllable Generation](https://arxiv.org/abs/1909.05858) by Nitish Shirish Keskar*, Bryan McCann*, Lav R. Varshney, Caiming Xiong and Richard Socher.
+1. **[Data2Vec](model_doc/data2vec)** (from Facebook) released with the paper [Data2Vec:  A General Framework for Self-supervised Learning in Speech, Vision and Language](https://arxiv.org/abs/2202.03555) by Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, Michael Auli.
 1. **[DeBERTa](model_doc/deberta)** (from Microsoft) released with the paper [DeBERTa: Decoding-enhanced BERT with Disentangled Attention](https://arxiv.org/abs/2006.03654) by Pengcheng He, Xiaodong Liu, Jianfeng Gao, Weizhu Chen.
 1. **[DeBERTa-v2](model_doc/deberta-v2)** (from Microsoft) released with the paper [DeBERTa: Decoding-enhanced BERT with Disentangled Attention](https://arxiv.org/abs/2006.03654) by Pengcheng He, Xiaodong Liu, Jianfeng Gao, Weizhu Chen.
 1. **[DeiT](model_doc/deit)** (from Facebook) released with the paper [Training data-efficient image transformers & distillation through attention](https://arxiv.org/abs/2012.12877) by Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, Hervé Jégou.
@@ -181,6 +182,8 @@ Flax), PyTorch, and/or TensorFlow.
 |          ConvBERT           |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
 |          ConvNext           |       ❌       |       ❌       |       ✅        |         ✅         |      ❌      |
 |            CTRL             |       ✅       |       ❌       |       ✅        |         ✅         |      ❌      |
+|        Data2VecAudio        |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
+|        Data2VecText         |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |           DeBERTa           |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
 |         DeBERTa-v2          |       ✅       |       ❌       |       ✅        |         ✅         |      ❌      |
 |            DeiT             |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |

docs/source/model_doc/data2vec.mdx

+<!--Copyright 2022 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# Data2Vec
+
+## Overview
+
+The Data2Vec model was proposed in [data2vec: A General Framework for Self-supervised Learning in Speech, Vision and Language](https://arxiv.org/pdf/2202.03555) by Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu and Michael Auli.
+Data2Vec proposes a unified framework for self-supervised learning across different data modalities - text, audio and images.
+Importantly, predicted targets for pre-training are contextualized latent representations of the inputs, rather than modality-specific, context-independent targets.
+
+The abstract from the paper is the following:
+
+*While the general idea of self-supervised learning is identical across modalities, the actual algorithms and
+objectives differ widely because they were developed with a single modality in mind. To get us closer to general
+self-supervised learning, we present data2vec, a framework that uses the same learning method for either speech,
+NLP or computer vision. The core idea is to predict latent representations of the full input data based on a
+masked view of the input in a selfdistillation setup using a standard Transformer architecture.
+Instead of predicting modality-specific targets such as words, visual tokens or units of human speech which
+are local in nature, data2vec predicts contextualized latent representations that contain information from
+the entire input. Experiments on the major benchmarks of speech recognition, image classification, and
+natural language understanding demonstrate a new state of the art or competitive performance to predominant approaches.
+Models and code are available at www.github.com/pytorch/fairseq/tree/master/examples/data2vec.*
+
+Tips:
+
+- Both Data2VecAudio and Data2VecText have been trained using the same self-supervised learning method.
+  In the case of Data2VecAudio, preprocessing is identical to [`RobertaModel`], including tokenization.
+
+This model was contributed by [edugp](https://huggingface.co/edugp).
+The original code can be found [here](https://github.com/pytorch/fairseq/tree/main/examples/data2vec).
+
+
+## Data2VecTextConfig
+
+[[autodoc]] Data2VecTextConfig
+
+## Data2VecAudioConfig
+
+[[autodoc]] Data2VecAudioConfig
+
+## Data2VecAudioModel
+
+[[autodoc]] Data2VecAudioModel
+    - forward
+
+
+## Data2VecAudioForAudioFrameClassification
+
+[[autodoc]] Data2VecAudioForAudioFrameClassification
+    - forward
+
+## Data2VecAudioForCTC
+
+[[autodoc]] Data2VecAudioForCTC
+    - forward
+
+## Data2VecAudioForSequenceClassification
+
+[[autodoc]] Data2VecAudioForSequenceClassification
+    - forward
+
+## Data2VecAudioForXVector
+
+[[autodoc]] Data2VecAudioForXVector
+    - forward
+
+## Data2VecTextModel
+
+[[autodoc]] Data2VecTextModel
+    - forward
+
+## Data2VecTextForCausalLM
+
+[[autodoc]] Data2VecTextForCausalLM
+    - forward
+
+## Data2VecTextForMaskedLM
+
+[[autodoc]] Data2VecTextForMaskedLM
+    - forward
+
+## Data2VecTextForSequenceClassification
+
+[[autodoc]] Data2VecTextForSequenceClassification
+    - forward
+
+## Data2VecTextForMultipleChoice
+
+[[autodoc]] Data2VecTextForMultipleChoice
+    - forward
+
+## Data2VecTextForTokenClassification
+
+[[autodoc]] Data2VecTextForTokenClassification
+    - forward
+
+## Data2VecTextForQuestionAnswering
+
+[[autodoc]] Data2VecTextForQuestionAnswering
+    - forward

docs/source/serialization.mdx

@@ -49,6 +49,7 @@ Ready-made configurations include the following architectures:
 - BART
 - BERT
 - CamemBERT
+- Data2VecText
 - DistilBERT
 - ELECTRA
 - GPT Neo

src/transformers/__init__.py

@@ -204,6 +204,7 @@ _import_structure = {
     "models.convnext": ["CONVNEXT_PRETRAINED_CONFIG_ARCHIVE_MAP", "ConvNextConfig"],
     "models.cpm": ["CpmTokenizer"],
     "models.ctrl": ["CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP", "CTRLConfig", "CTRLTokenizer"],
+    "models.data2vec": ["DATA2VEC_TEXT_PRETRAINED_CONFIG_ARCHIVE_MAP", "Data2VecAudioConfig", "Data2VecTextConfig"],
     "models.deberta": ["DEBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP", "DebertaConfig", "DebertaTokenizer"],
     "models.deberta_v2": ["DEBERTA_V2_PRETRAINED_CONFIG_ARCHIVE_MAP", "DebertaV2Config"],
     "models.deit": ["DEIT_PRETRAINED_CONFIG_ARCHIVE_MAP", "DeiTConfig"],
@@ -668,6 +669,7 @@ if is_torch_available():
     _import_structure["models.auto"].extend(
         [
             "MODEL_FOR_AUDIO_CLASSIFICATION_MAPPING",
+            "MODEL_FOR_AUDIO_XVECTOR_MAPPING",
             "MODEL_FOR_CAUSAL_IMAGE_MODELING_MAPPING",
             "MODEL_FOR_CAUSAL_LM_MAPPING",
             "MODEL_FOR_CTC_MAPPING",
@@ -871,6 +873,26 @@ if is_torch_available():
             "CTRLPreTrainedModel",
         ]
     )
+    _import_structure["models.data2vec"].extend(
+        [
+            "DATA2VEC_AUDIO_PRETRAINED_MODEL_ARCHIVE_LIST",
+            "DATA2VEC_TEXT_PRETRAINED_MODEL_ARCHIVE_LIST",
+            "Data2VecAudioForAudioFrameClassification",
+            "Data2VecAudioForCTC",
+            "Data2VecAudioForSequenceClassification",
+            "Data2VecAudioForXVector",
+            "Data2VecAudioModel",
+            "Data2VecAudioPreTrainedModel",
+            "Data2VecTextForCausalLM",
+            "Data2VecTextForMaskedLM",
+            "Data2VecTextForMultipleChoice",
+            "Data2VecTextForQuestionAnswering",
+            "Data2VecTextForSequenceClassification",
+            "Data2VecTextForTokenClassification",
+            "Data2VecTextModel",
+            "Data2VecTextPreTrainedModel",
+        ]
+    )
     _import_structure["models.deberta"].extend(
         [
             "DEBERTA_PRETRAINED_MODEL_ARCHIVE_LIST",
@@ -2468,6 +2490,7 @@ if TYPE_CHECKING:
     from .models.convnext import CONVNEXT_PRETRAINED_CONFIG_ARCHIVE_MAP, ConvNextConfig
     from .models.cpm import CpmTokenizer
     from .models.ctrl import CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP, CTRLConfig, CTRLTokenizer
+    from .models.data2vec import DATA2VEC_TEXT_PRETRAINED_CONFIG_ARCHIVE_MAP, Data2VecAudioConfig, Data2VecTextConfig
     from .models.deberta import DEBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP, DebertaConfig, DebertaTokenizer
     from .models.deberta_v2 import DEBERTA_V2_PRETRAINED_CONFIG_ARCHIVE_MAP, DebertaV2Config
     from .models.deit import DEIT_PRETRAINED_CONFIG_ARCHIVE_MAP, DeiTConfig
@@ -2847,6 +2870,7 @@ if TYPE_CHECKING:
         )
         from .models.auto import (
             MODEL_FOR_AUDIO_CLASSIFICATION_MAPPING,
+            MODEL_FOR_AUDIO_XVECTOR_MAPPING,
             MODEL_FOR_CAUSAL_IMAGE_MODELING_MAPPING,
             MODEL_FOR_CAUSAL_LM_MAPPING,
             MODEL_FOR_CTC_MAPPING,
@@ -3021,6 +3045,24 @@ if TYPE_CHECKING:
             CTRLModel,
             CTRLPreTrainedModel,
         )
+        from .models.data2vec import (
+            DATA2VEC_AUDIO_PRETRAINED_MODEL_ARCHIVE_LIST,
+            DATA2VEC_TEXT_PRETRAINED_MODEL_ARCHIVE_LIST,
+            Data2VecAudioForAudioFrameClassification,
+            Data2VecAudioForCTC,
+            Data2VecAudioForSequenceClassification,
+            Data2VecAudioForXVector,
+            Data2VecAudioModel,
+            Data2VecAudioPreTrainedModel,
+            Data2VecTextForCausalLM,
+            Data2VecTextForMaskedLM,
+            Data2VecTextForMultipleChoice,
+            Data2VecTextForQuestionAnswering,
+            Data2VecTextForSequenceClassification,
+            Data2VecTextForTokenClassification,
+            Data2VecTextModel,
+            Data2VecTextPreTrainedModel,
+        )
         from .models.deberta import (
             DEBERTA_PRETRAINED_MODEL_ARCHIVE_LIST,
             DebertaForMaskedLM,

src/transformers/models/__init__.py

@@ -40,6 +40,7 @@ from . import (
     convnext,
     cpm,
     ctrl,
+    data2vec,
     deberta,
     deberta_v2,
     deit,

src/transformers/models/auto/__init__.py

@@ -32,6 +32,7 @@ _import_structure = {
 if is_torch_available():
     _import_structure["modeling_auto"] = [
         "MODEL_FOR_AUDIO_CLASSIFICATION_MAPPING",
+        "MODEL_FOR_AUDIO_XVECTOR_MAPPING",
         "MODEL_FOR_CAUSAL_IMAGE_MODELING_MAPPING",
         "MODEL_FOR_CAUSAL_LM_MAPPING",
         "MODEL_FOR_CTC_MAPPING",
@@ -150,6 +151,7 @@ if TYPE_CHECKING:
     if is_torch_available():
         from .modeling_auto import (
             MODEL_FOR_AUDIO_CLASSIFICATION_MAPPING,
+            MODEL_FOR_AUDIO_XVECTOR_MAPPING,
             MODEL_FOR_CAUSAL_IMAGE_MODELING_MAPPING,
             MODEL_FOR_CAUSAL_LM_MAPPING,
             MODEL_FOR_CTC_MAPPING,

src/transformers/models/auto/configuration_auto.py

@@ -121,6 +121,8 @@ CONFIG_MAPPING_NAMES = OrderedDict(
         ("unispeech-sat", "UniSpeechSatConfig"),
         ("unispeech", "UniSpeechConfig"),
         ("wavlm", "WavLMConfig"),
+        ("data2vec-audio", "Data2VecAudioConfig"),
+        ("data2vec-text", "Data2VecTextConfig"),
     ]
 )
 
@@ -177,6 +179,8 @@ CONFIG_ARCHIVE_MAP_MAPPING_NAMES = OrderedDict(
         ("xlnet", "XLNET_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("xlm", "XLM_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("roberta", "ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("data2vec-text", "DATA2VEC_TEXT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("data2vec-audio", "DATA2VEC_AUDIO_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("distilbert", "DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("albert", "ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("camembert", "CAMEMBERT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
@@ -321,10 +325,14 @@ MODEL_NAMES_MAPPING = OrderedDict(
         ("xlsr_wav2vec2", "XLSR-Wav2Vec2"),
         ("mluke", "mLUKE"),
         ("layoutxlm", "LayoutXLM"),
+        ("data2vec-audio", "Data2VecAudio"),
+        ("data2vec-text", "Data2VecText"),
     ]
 )
 
-SPECIAL_MODEL_TYPE_TO_MODULE_NAME = OrderedDict([("openai-gpt", "openai")])
+SPECIAL_MODEL_TYPE_TO_MODULE_NAME = OrderedDict(
+    [("openai-gpt", "openai"), ("data2vec-audio", "data2vec"), ("data2vec-text", "data2vec")]
+)
 
 
 def model_type_to_module_name(key):

src/transformers/models/auto/modeling_auto.py

@@ -83,6 +83,8 @@ MODEL_MAPPING_NAMES = OrderedDict(
         ("bart", "BartModel"),
         ("longformer", "LongformerModel"),
         ("roberta", "RobertaModel"),
+        ("data2vec-text", "Data2VecTextModel"),
+        ("data2vec-audio", "Data2VecAudioModel"),
         ("layoutlm", "LayoutLMModel"),
         ("squeezebert", "SqueezeBertModel"),
         ("bert", "BertModel"),
@@ -133,6 +135,7 @@ MODEL_FOR_PRETRAINING_MAPPING_NAMES = OrderedDict(
         ("fsmt", "FSMTForConditionalGeneration"),
         ("longformer", "LongformerForMaskedLM"),
         ("roberta", "RobertaForMaskedLM"),
+        ("data2vec-text", "Data2VecTextForMaskedLM"),
         ("squeezebert", "SqueezeBertForMaskedLM"),
         ("bert", "BertForPreTraining"),
         ("big_bird", "BigBirdForPreTraining"),
@@ -191,6 +194,7 @@ MODEL_WITH_LM_HEAD_MAPPING_NAMES = OrderedDict(
         ("bart", "BartForConditionalGeneration"),
         ("longformer", "LongformerForMaskedLM"),
         ("roberta", "RobertaForMaskedLM"),
+        ("data2vec-text", "Data2VecTextForMaskedLM"),
         ("squeezebert", "SqueezeBertForMaskedLM"),
         ("bert", "BertForMaskedLM"),
         ("openai-gpt", "OpenAIGPTLMHeadModel"),
@@ -251,6 +255,7 @@ MODEL_FOR_CAUSAL_LM_MAPPING_NAMES = OrderedDict(
         ("blenderbot-small", "BlenderbotSmallForCausalLM"),
         ("megatron-bert", "MegatronBertForCausalLM"),
         ("speech_to_text_2", "Speech2Text2ForCausalLM"),
+        ("data2vec-text", "Data2VecTextForCausalLM"),
     ]
 )
 
@@ -337,6 +342,7 @@ MODEL_FOR_MASKED_LM_MAPPING_NAMES = OrderedDict(
         ("xlm-roberta", "XLMRobertaForMaskedLM"),
         ("longformer", "LongformerForMaskedLM"),
         ("roberta", "RobertaForMaskedLM"),
+        ("data2vec-text", "Data2VecTextForMaskedLM"),
         ("squeezebert", "SqueezeBertForMaskedLM"),
         ("bert", "BertForMaskedLM"),
         ("megatron-bert", "MegatronBertForMaskedLM"),
@@ -417,6 +423,7 @@ MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
         ("bart", "BartForSequenceClassification"),
         ("longformer", "LongformerForSequenceClassification"),
         ("roberta", "RobertaForSequenceClassification"),
+        ("data2vec-text", "Data2VecTextForSequenceClassification"),
         ("squeezebert", "SqueezeBertForSequenceClassification"),
         ("layoutlm", "LayoutLMForSequenceClassification"),
         ("bert", "BertForSequenceClassification"),
@@ -482,6 +489,7 @@ MODEL_FOR_QUESTION_ANSWERING_MAPPING_NAMES = OrderedDict(
         ("deberta-v2", "DebertaV2ForQuestionAnswering"),
         ("ibert", "IBertForQuestionAnswering"),
         ("splinter", "SplinterForQuestionAnswering"),
+        ("data2vec-text", "Data2VecTextForQuestionAnswering"),
     ]
 )
 
@@ -527,6 +535,7 @@ MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
         ("deberta-v2", "DebertaV2ForTokenClassification"),
         ("gpt2", "GPT2ForTokenClassification"),
         ("ibert", "IBertForTokenClassification"),
+        ("data2vec-text", "Data2VecTextForTokenClassification"),
     ]
 )
 
@@ -548,6 +557,7 @@ MODEL_FOR_MULTIPLE_CHOICE_MAPPING_NAMES = OrderedDict(
         ("xlm-roberta", "XLMRobertaForMultipleChoice"),
         ("longformer", "LongformerForMultipleChoice"),
         ("roberta", "RobertaForMultipleChoice"),
+        ("data2vec-text", "Data2VecTextForMultipleChoice"),
         ("squeezebert", "SqueezeBertForMultipleChoice"),
         ("bert", "BertForMultipleChoice"),
         ("distilbert", "DistilBertForMultipleChoice"),
@@ -583,6 +593,7 @@ MODEL_FOR_AUDIO_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
         ("sew", "SEWForSequenceClassification"),
         ("sew-d", "SEWDForSequenceClassification"),
         ("wavlm", "WavLMForSequenceClassification"),
+        ("data2vec-audio", "Data2VecAudioForSequenceClassification"),
     ]
 )
 
@@ -596,6 +607,7 @@ MODEL_FOR_CTC_MAPPING_NAMES = OrderedDict(
         ("sew", "SEWForCTC"),
         ("sew-d", "SEWDForCTC"),
         ("wavlm", "WavLMForCTC"),
+        ("data2vec-audio", "Data2VecAudioForCTC"),
     ]
 )
 
@@ -605,6 +617,7 @@ MODEL_FOR_AUDIO_FRAME_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
         ("wav2vec2", "Wav2Vec2ForAudioFrameClassification"),
         ("unispeech-sat", "UniSpeechSatForAudioFrameClassification"),
         ("wavlm", "WavLMForAudioFrameClassification"),
+        ("data2vec-audio", "Data2VecAudioForAudioFrameClassification"),
     ]
 )
 
@@ -614,6 +627,7 @@ MODEL_FOR_AUDIO_XVECTOR_MAPPING_NAMES = OrderedDict(
         ("wav2vec2", "Wav2Vec2ForXVector"),
         ("unispeech-sat", "UniSpeechSatForXVector"),
         ("wavlm", "WavLMForXVector"),
+        ("data2vec-audio", "Data2VecAudioForXVector"),
     ]
 )
 

src/transformers/models/data2vec/__init__.py

+# flake8: noqa
+# There's no way to ignore "F401 '...' imported but unused" warnings in this
+# module, but to preserve other warnings. So, don't check this module at all.
+
+# Copyright 2022 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import TYPE_CHECKING
+
+from ...file_utils import _LazyModule, is_torch_available
+
+
+_import_structure = {
+    "configuration_data2vec_audio": [
+        "DATA2VEC_AUDIO_PRETRAINED_CONFIG_ARCHIVE_MAP",
+        "Data2VecAudioConfig",
+    ],
+    "configuration_data2vec_text": [
+        "DATA2VEC_TEXT_PRETRAINED_CONFIG_ARCHIVE_MAP",
+        "Data2VecTextConfig",
+        "Data2VecTextOnnxConfig",
+    ],
+}
+
+if is_torch_available():
+    _import_structure["modeling_data2vec_audio"] = [
+        "DATA2VEC_AUDIO_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "Data2VecAudioForAudioFrameClassification",
+        "Data2VecAudioForCTC",
+        "Data2VecAudioForSequenceClassification",
+        "Data2VecAudioForXVector",
+        "Data2VecAudioModel",
+        "Data2VecAudioPreTrainedModel",
+    ]
+    _import_structure["modeling_data2vec_text"] = [
+        "DATA2VEC_TEXT_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "Data2VecTextForCausalLM",
+        "Data2VecTextForMaskedLM",
+        "Data2VecTextForMultipleChoice",
+        "Data2VecTextForQuestionAnswering",
+        "Data2VecTextForSequenceClassification",
+        "Data2VecTextForTokenClassification",
+        "Data2VecTextModel",
+        "Data2VecTextPreTrainedModel",
+    ]
+
+if TYPE_CHECKING:
+    from .configuration_data2vec_audio import DATA2VEC_AUDIO_PRETRAINED_CONFIG_ARCHIVE_MAP, Data2VecAudioConfig
+    from .configuration_data2vec_text import (
+        DATA2VEC_TEXT_PRETRAINED_CONFIG_ARCHIVE_MAP,
+        Data2VecTextConfig,
+        Data2VecTextOnnxConfig,
+    )
+
+    if is_torch_available():
+        from .modeling_data2vec_audio import (
+            DATA2VEC_AUDIO_PRETRAINED_MODEL_ARCHIVE_LIST,
+            Data2VecAudioForAudioFrameClassification,
+            Data2VecAudioForCTC,
+            Data2VecAudioForSequenceClassification,
+            Data2VecAudioForXVector,
+            Data2VecAudioModel,
+            Data2VecAudioPreTrainedModel,
+        )
+        from .modeling_data2vec_text import (
+            DATA2VEC_TEXT_PRETRAINED_MODEL_ARCHIVE_LIST,
+            Data2VecTextForCausalLM,
+            Data2VecTextForMaskedLM,
+            Data2VecTextForMultipleChoice,
+            Data2VecTextForQuestionAnswering,
+            Data2VecTextForSequenceClassification,
+            Data2VecTextForTokenClassification,
+            Data2VecTextModel,
+            Data2VecTextPreTrainedModel,
+        )
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)

src/transformers/models/data2vec/configuration_data2vec_audio.py

+# coding=utf-8
+# Copyright 2022 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Data2VecText configuration"""
+
+import math
+
+from ...configuration_utils import PretrainedConfig
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+DATA2VEC_AUDIO_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "facebook/data2vec-base-960h": "https://huggingface.co/facebook/data2vec-audio-base-960h/resolve/main/config.json",
+    # See all Data2VecAudio models at https://huggingface.co/models?filter=data2vec-audio
+}
+
+
+class Data2VecAudioConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`Data2VecAudioModel`]. It is used to instantiate
+    an Data2VecAudio model according to the specified arguments, defining the model architecture. Instantiating a
+    configuration with the defaults will yield a similar configuration to that of the Data2VecAudio
+    [facebook/data2vec-audio-base-960h](https://huggingface.co/facebook/data2vec-audio-base-960h) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 32):
+            Vocabulary size of the Data2VecAudio model. Defines the number of different tokens that can be represented
+            by the `inputs_ids` passed when calling [`Data2VecAudioModel`] or [`TFData2VecAudioModel`]. Vocabulary size
+            of the model. Defines the different tokens that can be represented by the *inputs_ids* passed to the
+            forward method of [`Data2VecAudioModel`].
+        hidden_size (`int`, *optional*, defaults to 768):
+            Dimensionality of the encoder layers and the pooler layer.
+        num_hidden_layers (`int`, *optional*, defaults to 12):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 12):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        intermediate_size (`int`, *optional*, defaults to 3072):
+            Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
+        hidden_act (`str` or `function`, *optional*, defaults to `"gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"selu"` and `"gelu_new"` are supported.
+        hidden_dropout (`float`, *optional*, defaults to 0.1):
+            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
+        attention_dropout (`float`, *optional*, defaults to 0.1):
+            The dropout ratio for the attention probabilities.
+        final_dropout (`float`, *optional*, defaults to 0.1):
+            The dropout probability for the final projection layer of [`Data2VecAudioForCTC`].
+        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.
+        feat_proj_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout probability for output of the feature encoder.
+        feat_extract_activation (`str, `optional`, defaults to `"gelu"`):
+            The non-linear activation function (function or string) in the 1D convolutional layers of the feature
+            extractor. If string, `"gelu"`, `"relu"`, `"selu"` and `"gelu_new"` are supported.
+        conv_dim (`Tuple[int]`, *optional*, defaults to `(512, 512, 512, 512, 512, 512, 512)`):
+            A tuple of integers defining the number of input and output channels of each 1D convolutional layer in the
+            feature encoder. The length of *conv_dim* defines the number of 1D convolutional layers.
+        conv_stride (`Tuple[int]`, *optional*, defaults to `(5, 2, 2, 2, 2, 2, 2)`):
+            A tuple of integers defining the stride of each 1D convolutional layer in the feature encoder. The length
+            of *conv_stride* defines the number of convolutional layers and has to match the length of *conv_dim*.
+        conv_kernel (`Tuple[int]`, *optional*, defaults to `(10, 3, 3, 3, 3, 3, 3)`):
+            A tuple of integers defining the kernel size of each 1D convolutional layer in the feature encoder. The
+            length of *conv_kernel* defines the number of convolutional layers and has to match the length of
+            *conv_dim*.
+        conv_bias (`bool`, *optional*, defaults to `False`):
+            Whether the 1D convolutional layers have a bias.
+        num_conv_pos_embeddings (`int`, *optional*, defaults to 128):
+            Number of convolutional positional embeddings. Defines the kernel size of 1D convolutional positional
+            embeddings layer.
+        num_conv_pos_embedding_groups (`int`, *optional*, defaults to 16):
+            Number of groups of 1D convolutional positional embeddings layer.
+        mask_time_prob (`float`, *optional*, defaults to 0.05):
+            Percentage (between 0 and 1) of all feature vectors along the time axis which will be masked. The masking
+            procecure generates ''mask_time_prob*len(time_axis)/mask_time_length'' independent masks over the axis. If
+            reasoning from the propability of each feature vector to be chosen as the start of the vector span to be
+            masked, *mask_time_prob* should be `prob_vector_start*mask_time_length`. Note that overlap may decrease the
+        mask_time_length (`int`, *optional*, defaults to 10):
+            Length of vector span along the time axis.
+        mask_time_min_masks (`int`, *optional*, defaults to 2),:
+            The minimum number of masks of length `mask_feature_length` generated along the time axis, each time step,
+            irrespectively of `mask_feature_prob`. Only relevant if ''mask_time_prob*len(time_axis)/mask_time_length <
+            mask_time_min_masks''
+        mask_feature_prob (`float`, *optional*, defaults to 0.0):
+            Percentage (between 0 and 1) of all feature vectors along the feature axis which will be masked. The
+            masking procecure generates ''mask_feature_prob*len(feature_axis)/mask_time_length'' independent masks over
+            the axis. If reasoning from the propability of each feature vector to be chosen as the start of the vector
+            span to be masked, *mask_feature_prob* should be `prob_vector_start*mask_feature_length`. Note that overlap
+            may decrease the actual percentage of masked vectors. This is only relevant if `apply_spec_augment is
+            True`.
+        mask_feature_length (`int`, *optional*, defaults to 10):
+            Length of vector span along the feature axis.
+        mask_feature_min_masks (`int`, *optional*, defaults to 0),:
+            The minimum number of masks of length `mask_feature_length` generated along the feature axis, each time
+            step, irrespectively of `mask_feature_prob`. Only relevant if
+            ''mask_feature_prob*len(feature_axis)/mask_feature_length < mask_feature_min_masks''
+        ctc_loss_reduction (`str`, *optional*, defaults to `"sum"`):
+            Specifies the reduction to apply to the output of `torch.nn.CTCLoss`. Only relevant when training an
+            instance of [`Data2VecAudioForCTC`].
+        ctc_zero_infinity (`bool`, *optional*, defaults to `False`):
+            Whether to zero infinite losses and the associated gradients of `torch.nn.CTCLoss`. Infinite losses mainly
+            occur when the inputs are too short to be aligned to the targets. Only relevant when training an instance
+            of [`Data2VecAudioForCTC`].
+        use_weighted_layer_sum (`bool`, *optional*, defaults to `False`):
+            Whether to use a weighted average of layer outputs with learned weights. Only relevant when using an
+            instance of [`Data2VecAudioForSequenceClassification`].
+        classifier_proj_size (`int`, *optional*, defaults to 256):
+            Dimensionality of the projection before token mean-pooling for classification.
+        tdnn_dim (`Tuple[int]`, *optional*, defaults to `(512, 512, 512, 512, 1500)`):
+            A tuple of integers defining the number of output channels of each 1D convolutional layer in the *TDNN*
+            module of the *XVector* model. The length of *tdnn_dim* defines the number of *TDNN* layers.
+        tdnn_kernel (`Tuple[int]`, *optional*, defaults to `(5, 3, 3, 1, 1)`):
+            A tuple of integers defining the kernel size of each 1D convolutional layer in the *TDNN* module of the
+            *XVector* model. The length of *tdnn_kernel* has to match the length of *tdnn_dim*.
+        tdnn_dilation (`Tuple[int]`, *optional*, defaults to `(1, 2, 3, 1, 1)`):
+            A tuple of integers defining the dilation factor of each 1D convolutional layer in *TDNN* module of the
+            *XVector* model. The length of *tdnn_dilation* has to match the length of *tdnn_dim*.
+        xvector_output_dim (`int`, *optional*, defaults to 512):
+            Dimensionality of the *XVector* embedding vectors.
+        add_adapter (`bool`, *optional*, defaults to `False`):
+            Whether a convolutional network should be stacked on top of the Data2VecAudio Encoder. Can be very useful
+            for warm-starting Data2VecAudio for SpeechEncoderDecoder models.
+        adapter_kernel_size (`int`, *optional*, defaults to 3):
+            Kernel size of the convolutional layers in the adapter network. Only relevant if `add_adapter is True`.
+        adapter_stride (`int`, *optional*, defaults to 2):
+            Stride of the convolutional layers in the adapter network. Only relevant if `add_adapter is True`.
+        num_adapter_layers (`int`, *optional*, defaults to 3):
+            Number of convolutional layers that should be used in the adapter network. Only relevant if `add_adapter is
+            True`.
+        output_hidden_size (`int`, *optional*):
+            Dimensionality of the encoder output layer. If not defined, this defaults to *hidden-size*. Only relevant
+            if `add_adapter is True`.
+
+    Example:
+
+    ```python
+    >>> from transformers import Data2VecAudioModel, Data2VecAudioConfig
+
+    >>> # Initializing a Data2VecAudio facebook/data2vec-audio-base-960h style configuration
+    >>> configuration = Data2VecAudioConfig()
+
+    >>> # Initializing a model from the facebook/data2vec-audio-base-960h style configuration
+    >>> model = Data2VecAudioModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+    model_type = "data2vec-audio"
+
+    def __init__(
+        self,
+        vocab_size=32,
+        hidden_size=768,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        intermediate_size=3072,
+        hidden_act="gelu",
+        hidden_dropout=0.1,
+        activation_dropout=0.1,
+        attention_dropout=0.1,
+        feat_proj_dropout=0.0,
+        final_dropout=0.1,
+        layerdrop=0.1,
+        initializer_range=0.02,
+        layer_norm_eps=1e-5,
+        feat_extract_activation="gelu",
+        conv_dim=(512, 512, 512, 512, 512, 512, 512),
+        conv_stride=(5, 2, 2, 2, 2, 2, 2),
+        conv_kernel=(10, 3, 3, 3, 3, 2, 2),
+        conv_bias=False,
+        num_conv_pos_embedding_groups=16,
+        conv_pos_kernel_size=19,
+        num_conv_pos_embeddings=5,
+        mask_time_prob=0.05,
+        mask_time_length=10,
+        mask_time_min_masks=2,
+        mask_feature_prob=0.0,
+        mask_feature_length=10,
+        mask_feature_min_masks=0,
+        ctc_loss_reduction="sum",
+        ctc_zero_infinity=False,
+        use_weighted_layer_sum=False,
+        classifier_proj_size=256,
+        tdnn_dim=(512, 512, 512, 512, 1500),
+        tdnn_kernel=(5, 3, 3, 1, 1),
+        tdnn_dilation=(1, 2, 3, 1, 1),
+        xvector_output_dim=512,
+        pad_token_id=0,
+        bos_token_id=1,
+        eos_token_id=2,
+        add_adapter=False,
+        adapter_kernel_size=3,
+        adapter_stride=2,
+        num_adapter_layers=3,
+        output_hidden_size=None,
+        **kwargs
+    ):
+        super().__init__(**kwargs, pad_token_id=pad_token_id, bos_token_id=bos_token_id, eos_token_id=eos_token_id)
+        self.hidden_size = hidden_size
+        self.feat_extract_activation = feat_extract_activation
+        self.conv_dim = list(conv_dim)
+        self.conv_stride = list(conv_stride)
+        self.conv_kernel = list(conv_kernel)
+        self.conv_bias = conv_bias
+        self.num_conv_pos_embeddings = num_conv_pos_embeddings
+        self.num_conv_pos_embedding_groups = num_conv_pos_embedding_groups
+        self.conv_pos_kernel_size = conv_pos_kernel_size
+        self.num_feat_extract_layers = len(self.conv_dim)
+        self.num_hidden_layers = num_hidden_layers
+        self.intermediate_size = intermediate_size
+        self.hidden_act = hidden_act
+        self.num_attention_heads = num_attention_heads
+        self.hidden_dropout = hidden_dropout
+        self.attention_dropout = attention_dropout
+        self.activation_dropout = activation_dropout
+        self.feat_proj_dropout = feat_proj_dropout
+        self.final_dropout = final_dropout
+        self.layerdrop = layerdrop
+        self.layer_norm_eps = layer_norm_eps
+        self.initializer_range = initializer_range
+        self.vocab_size = vocab_size
+        self.use_weighted_layer_sum = use_weighted_layer_sum
+
+        if (
+            (len(self.conv_stride) != self.num_feat_extract_layers)
+            or (len(self.conv_kernel) != self.num_feat_extract_layers)
+            or (len(self.conv_dim) != self.num_feat_extract_layers)
+        ):
+            raise ValueError(
+                "Configuration for convolutional layers is incorrect. "
+                "It is required that `len(config.conv_dim)` == `len(config.conv_stride)` == `len(config.conv_kernel)`, "
+                f"but is `len(config.conv_dim) = {len(self.conv_dim)}`, `len(config.conv_stride) "
+                f"= {len(self.conv_stride)}`, `len(config.conv_kernel) = {len(self.conv_kernel)}`."
+            )
+
+        # fine-tuning config parameters for SpecAugment: https://arxiv.org/abs/1904.08779
+        self.mask_time_prob = mask_time_prob
+        self.mask_time_length = mask_time_length
+        self.mask_time_min_masks = mask_time_min_masks
+        self.mask_feature_prob = mask_feature_prob
+        self.mask_feature_length = mask_feature_length
+        self.mask_feature_min_masks = mask_feature_min_masks
+
+        # ctc loss
+        self.ctc_loss_reduction = ctc_loss_reduction
+        self.ctc_zero_infinity = ctc_zero_infinity
+
+        # adapter
+        self.add_adapter = add_adapter
+        self.adapter_kernel_size = adapter_kernel_size
+        self.adapter_stride = adapter_stride
+        self.num_adapter_layers = num_adapter_layers
+        self.output_hidden_size = output_hidden_size or hidden_size
+
+        # SequenceClassification-specific parameter. Feel free to ignore for other classes.
+        self.classifier_proj_size = classifier_proj_size
+
+        # XVector-specific parameters. Feel free to ignore for other classes.
+        self.tdnn_dim = list(tdnn_dim)
+        self.tdnn_kernel = list(tdnn_kernel)
+        self.tdnn_dilation = list(tdnn_dilation)
+        self.xvector_output_dim = xvector_output_dim
+
+    @property
+    def inputs_to_logits_ratio(self):
+        return math.prod(self.conv_stride)

src/transformers/models/data2vec/configuration_data2vec_text.py

+# coding=utf-8
+# Copyright 2022 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Data2VecText configuration"""
+from collections import OrderedDict
+from typing import Mapping
+
+from ...configuration_utils import PretrainedConfig
+from ...onnx import OnnxConfig
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+DATA2VEC_TEXT_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "facebook/data2vec-text-base": "https://huggingface.co/data2vec/resolve/main/config.json",
+}
+
+
+class Data2VecTextConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`Data2VecTextModel`] and [`Data2VecTextModel`]. It
+    is used to instantiate a Data2VecText model according to the specified arguments, defining the model architecture.
+    Instantiating a configuration with the defaults will yield a similar configuration to that of the Data2VecText
+    [facebook/data2vec-text-base](https://huggingface.co/facebook/data2vec-text-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:
+        vocab_size (`int`, *optional*, defaults to 30522):
+            Vocabulary size of the DATA2VEC model. Defines the number of different tokens that can be represented by
+            the `inputs_ids` passed when calling [`Data2VecModel`].
+        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 [`Data2VecModel`].
+        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`.
+        classifier_dropout (`float`, *optional*):
+            The dropout ratio for the classification head.
+
+    Examples:
+
+    ```python
+    >>> from transformers import Data2VecTextModel, Data2VecTextConfig
+
+    >>> # Initializing a Data2VecText facebook/data2vec-text-base style configuration
+    >>> configuration = Data2VecTextConfig()
+
+    >>> # Initializing a model from the facebook/data2vec-text-base style configuration
+    >>> model = Data2VecTextModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+    model_type = "data2vec-text"
+
+    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=1,
+        bos_token_id=0,
+        eos_token_id=2,
+        position_embedding_type="absolute",
+        use_cache=True,
+        classifier_dropout=None,
+        **kwargs
+    ):
+        super().__init__(pad_token_id=pad_token_id, bos_token_id=bos_token_id, eos_token_id=eos_token_id, **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.classifier_dropout = classifier_dropout
+
+
+class Data2VecTextOnnxConfig(OnnxConfig):
+    @property
+    def inputs(self) -> Mapping[str, Mapping[int, str]]:
+        return OrderedDict(
+            [
+                ("input_ids", {0: "batch", 1: "sequence"}),
+                ("attention_mask", {0: "batch", 1: "sequence"}),
+            ]
+        )

src/transformers/models/data2vec/convert_data2vec_audio_original_pytorch_checkpoint_to_pytorch.py

+# coding=utf-8
+# Copyright 2021 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 Wav2Vec2 checkpoint."""
+
+
+import argparse
+import os
+from functools import reduce
+
+import fairseq
+import torch
+from datasets import load_dataset
+
+from transformers import Wav2Vec2Processor, logging
+from transformers.models.data2vec.configuration_data2vec_audio import Data2VecAudioConfig
+
+# Copied from https://github.com/pytorch/fairseq/blob/main/examples/data2vec/models/data2vec_audio.py
+from transformers.models.data2vec.data2vec_audio import Data2VecAudioModel as Dummy  # noqa: F401
+from transformers.models.data2vec.modeling_data2vec_audio import Data2VecAudioForCTC, Data2VecAudioModel
+
+
+logging.set_verbosity_info()
+logger = logging.get_logger(__name__)
+
+MAPPING = {
+    "post_extract_proj": "feature_projection.projection",
+    "models.0.layer_norm": "feature_projection.layer_norm",
+    "self_attn.k_proj": "encoder.layers.*.attention.k_proj",
+    "self_attn.v_proj": "encoder.layers.*.attention.v_proj",
+    "self_attn.q_proj": "encoder.layers.*.attention.q_proj",
+    "self_attn.out_proj": "encoder.layers.*.attention.out_proj",
+    "self_attn_layer_norm": "encoder.layers.*.layer_norm",
+    "fc1": "encoder.layers.*.feed_forward.intermediate_dense",
+    "fc2": "encoder.layers.*.feed_forward.output_dense",
+    "final_layer_norm": "encoder.layers.*.final_layer_norm",
+    "encoder.layer_norm": "encoder.layer_norm",
+    "w2v_model.layer_norm": "feature_projection.layer_norm",
+    "w2v_encoder.proj": "lm_head",
+    "mask_emb": "masked_spec_embed",
+}
+TOP_LEVEL_KEYS = [
+    "lm_head",
+]
+
+
+def set_recursively(hf_pointer, key, value, full_name, weight_type):
+    for attribute in key.split("."):
+        hf_pointer = getattr(hf_pointer, attribute)
+
+    if weight_type is not None:
+        hf_shape = getattr(hf_pointer, weight_type).shape
+    else:
+        hf_shape = hf_pointer.shape
+
+    if hf_shape != value.shape:
+        raise ValueError(
+            f"Shape of hf {key + '.' + weight_type if weight_type is not None else ''} is {hf_shape}, but should be {value.shape} for {full_name}"
+        )
+
+    if weight_type == "weight":
+        hf_pointer.weight.data = value
+    elif weight_type == "weight_g":
+        hf_pointer.weight_g.data = value
+    elif weight_type == "weight_v":
+        hf_pointer.weight_v.data = value
+    elif weight_type == "bias":
+        hf_pointer.bias.data = value
+    else:
+        hf_pointer.data = value
+
+    logger.info(f"{key + '.' + weight_type if weight_type is not None else ''} was initialized from {full_name}.")
+
+
+def recursively_load_weights(fairseq_model, hf_model, is_headless):
+    unused_weights = []
+    fairseq_dict = fairseq_model.state_dict()
+
+    if not is_headless:
+        feature_extractor = hf_model.data2vec_audio.feature_extractor
+        pos_conv_embedding = hf_model.data2vec_audio.encoder.pos_conv_embed
+
+    else:
+        feature_extractor = hf_model.feature_extractor
+        pos_conv_embedding = hf_model.encoder.pos_conv_embed
+
+    for name, value in fairseq_dict.items():
+        is_used = False
+        if "conv_layers" in name:
+            load_conv_layer(
+                name,
+                value,
+                feature_extractor,
+                unused_weights,
+            )
+            is_used = True
+        elif "pos_conv" in name:
+            load_pos_conv_layer(
+                name,
+                value,
+                pos_conv_embedding,
+                unused_weights,
+            )
+            is_used = True
+        else:
+            for key, mapped_key in MAPPING.items():
+                if not is_headless:
+                    mapped_key = "data2vec_audio." + mapped_key if mapped_key not in TOP_LEVEL_KEYS else mapped_key
+                if key in name or key.split("w2v_model.")[-1] == name.split(".")[0]:
+                    is_used = True
+                    if "*" in mapped_key:
+                        layer_index = name.split(key)[0].split(".")[-2]
+                        mapped_key = mapped_key.replace("*", layer_index)
+                    if "weight_g" in name:
+                        weight_type = "weight_g"
+                    elif "weight_v" in name:
+                        weight_type = "weight_v"
+                    elif "bias" in name:
+                        weight_type = "bias"
+                    elif "weight" in name:
+                        # TODO: don't match quantizer.weight_proj
+                        weight_type = "weight"
+                    else:
+                        weight_type = None
+                    set_recursively(hf_model, mapped_key, value, name, weight_type)
+                continue
+        if not is_used:
+            unused_weights.append(name)
+
+    logger.warning(f"Unused weights: {unused_weights}")
+
+
+def access_by_string(module, path):
+    names = path.split(".")
+    return reduce(getattr, names, module)
+
+
+def set_weights(full_name, module, fsq_value, hf_weight_path):
+    hf_weight = access_by_string(module, hf_weight_path)
+    hf_value = hf_weight.data
+
+    if fsq_value.shape != hf_value.shape:
+        raise ValueError(f"{full_name} has size {fsq_value.shape}, but {hf_value.shape} was found.")
+    hf_weight.data = fsq_value
+    logger.info(f"{full_name} was correctly initialized from {hf_weight_path}.")
+
+
+def load_conv_layer(full_name, value, feature_extractor, unused_weights):
+    name = full_name.split("conv_layers.")[-1]
+    items = name.split(".")
+    layer_id = int(items[0])
+    type_id = int(items[1])
+
+    weight_type = name.split(".")[-1]
+    if type_id == 0:
+        layer_type = "conv"
+    elif type_id == 2:
+        layer_type = "layer_norm"
+    else:
+        unused_weights.append(full_name)
+        return
+
+    set_weights(full_name, feature_extractor, value, f"conv_layers.{layer_id}.{layer_type}.{weight_type}")
+
+
+def load_pos_conv_layer(full_name, value, pos_conv_embeddings, unused_weights):
+    name = full_name.split("pos_conv.")[-1]
+    items = name.split(".")
+    layer_id = int(items[0])
+    type_id = int(items[1])
+
+    weight_type = name.split(".")[-1]
+    if type_id != 0:
+        unused_weights.append(full_name)
+        return
+    else:
+        layer_type = "conv"
+
+    set_weights(full_name, pos_conv_embeddings, value, f"layers.{layer_id}.{layer_type}.{weight_type}")
+
+
+@torch.no_grad()
+def convert_wav2vec2_checkpoint(
+    checkpoint_path, pytorch_dump_folder_path, config_path=None, dict_path=None, is_finetuned=True
+):
+    """
+    Copy/paste/tweak model's weights to transformers design.
+    """
+    if config_path is not None:
+        config = Data2VecAudioConfig.from_pretrained(config_path)
+    else:
+        config = Data2VecAudioConfig()
+
+    if not is_finetuned:
+        # Modify final_proj layer name
+        hf_wav2vec = Data2VecAudioModel(config)
+        data2vec_checkpoint_dir = os.path.dirname(checkpoint_path)
+
+        state_dict = torch.load(checkpoint_path)
+        state_dict["model"]["final_proj.weight"] = state_dict["model"].pop("final_proj.0.weight")
+        state_dict["model"]["final_proj.bias"] = state_dict["model"].pop("final_proj.0.bias")
+        converted_ckpt = os.path.join(data2vec_checkpoint_dir, "converted.pt")
+        torch.save(state_dict, converted_ckpt)
+    else:
+        hf_wav2vec = Data2VecAudioForCTC(config)
+        converted_ckpt = checkpoint_path
+
+    def load_data2vec(path):
+        model, _, _ = fairseq.checkpoint_utils.load_model_ensemble_and_task([path])
+        return model[0].eval()
+
+    model = load_data2vec(converted_ckpt)
+
+    recursively_load_weights(model, hf_wav2vec, not is_finetuned)
+
+    processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-large-lv60")
+
+    ds = load_dataset("patrickvonplaten/librispeech_asr_dummy", "clean", split="validation")
+    input_audio = [x["array"] for x in ds[:4]["audio"]]
+
+    inputs = processor(input_audio, return_tensors="pt", padding=True)
+
+    input_values = inputs.input_values
+    attention_mask = inputs.attention_mask
+    #    input_values = inputs.input_values[:, :-1]
+    #    attention_mask = inputs.attention_mask[:, :-1]
+
+    hf_wav2vec.eval()
+    model.eval()
+    if is_finetuned:
+        their_output = model(source=input_values, padding_mask=(1 - attention_mask), mask=False, features_only=True)[
+            "encoder_out"
+        ].transpose(0, 1)
+        our_output = hf_wav2vec(input_values, attention_mask=attention_mask)["logits"]
+
+        pred_ids = torch.argmax(our_output, dim=-1)
+        output_string = processor.batch_decode(pred_ids)
+
+        print(f"Expected Output: {ds[:4]['text']}, Pred: {output_string}")
+    else:
+        their_output = model(source=input_values, padding_mask=(1 - attention_mask), mask=False, features_only=True)[
+            "layer_results"
+        ][-1][0].transpose(0, 1)
+        our_output = hf_wav2vec(input_values, attention_mask=attention_mask)["last_hidden_state"]
+
+    print(our_output.shape, their_output.shape)
+    max_absolute_diff = torch.max(torch.abs(our_output - their_output)).item()
+    print(f"max_absolute_diff = {max_absolute_diff}")  # ~ 1e-7
+    success = torch.allclose(our_output, their_output, atol=1e-3)
+    print("Do both models output the same tensors?", "🔥" if success else "💩")
+    if not success:
+        raise Exception("Something went wRoNg")
+
+    hf_wav2vec.save_pretrained(pytorch_dump_folder_path)
+
+    if is_finetuned:
+        processor.save_pretrained(pytorch_dump_folder_path)
+    else:
+        processor.feature_extractor.save_pretrained(pytorch_dump_folder_path)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--pytorch_dump_folder_path", default=None, type=str, help="Path to the output PyTorch model.")
+    parser.add_argument("--checkpoint_path", default=None, type=str, help="Path to fairseq checkpoint")
+    parser.add_argument("--dict_path", default=None, type=str, help="Path to dict of fine-tuned model")
+    parser.add_argument("--config_path", default=None, type=str, help="Path to hf config.json of model to convert")
+    parser.add_argument(
+        "--not_finetuned", action="store_true", help="Whether the model to convert is a fine-tuned model or not"
+    )
+    args = parser.parse_args()
+    convert_wav2vec2_checkpoint(
+        args.checkpoint_path, args.pytorch_dump_folder_path, args.config_path, args.dict_path, not args.not_finetuned
+    )

src/transformers/models/data2vec/convert_data2vec_text_original_pytorch_checkpoint_to_pytorch.py

+# coding=utf-8
+# Copyright 2022 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Convert data2vec checkpoint."""
+
+
+import argparse
+import os
+import pathlib
+
+import fairseq
+import torch
+from fairseq.modules import TransformerSentenceEncoderLayer
+from packaging import version
+
+from transformers import Data2VecTextConfig, Data2VecTextForMaskedLM, Data2VecTextForSequenceClassification
+from transformers.models.bert.modeling_bert import (
+    BertIntermediate,
+    BertLayer,
+    BertOutput,
+    BertSelfAttention,
+    BertSelfOutput,
+)
+
+# IMPORTANT: In order for this script to run, please make sure to download the dictionary: `dict.txt` from wget https://dl.fbaipublicfiles.com/fairseq/models/roberta.large.tar.gz
+# File copied from https://github.com/pytorch/fairseq/blob/main/examples/data2vec/models/data2vec_text.py
+from transformers.models.data2vec.data2vec_text import Data2VecTextModel
+from transformers.utils import logging
+
+
+if version.parse(fairseq.__version__) < version.parse("0.9.0"):
+    raise Exception("requires fairseq >= 0.9.0")
+
+
+logging.set_verbosity_info()
+logger = logging.get_logger(__name__)
+
+SAMPLE_TEXT = "Hello world! cécé herlolip"
+
+
+def convert_data2vec_checkpoint_to_pytorch(
+    data2vec_checkpoint_path: str, pytorch_dump_folder_path: str, classification_head: bool
+):
+    """
+    Copy/paste/tweak data2vec's weights to our BERT structure.
+    """
+    data2vec_checkpoint_dir, data2vec_checkpoint_file_name = os.path.split(data2vec_checkpoint_path)
+    data2vec = Data2VecTextModel.from_pretrained(
+        data2vec_checkpoint_dir, checkpoint_file=data2vec_checkpoint_file_name
+    )
+    data2vec.eval()  # disable dropout
+    data2vec_model = data2vec.models[0]
+    data2vec_sent_encoder = data2vec_model.encoder.sentence_encoder
+    config = Data2VecTextConfig(
+        vocab_size=data2vec_sent_encoder.embed_tokens.num_embeddings,
+        hidden_size=data2vec_model.args.encoder_embed_dim,
+        num_hidden_layers=data2vec_model.args.encoder_layers,
+        num_attention_heads=data2vec_model.args.encoder_attention_heads,
+        intermediate_size=data2vec_model.args.encoder_ffn_embed_dim,
+        max_position_embeddings=514,
+        type_vocab_size=1,
+        layer_norm_eps=1e-5,  # PyTorch default used in fairseq
+    )
+    if classification_head:
+        config.num_labels = data2vec.model.classification_heads["mnli"].out_proj.weight.shape[0]
+    print("Our BERT config:", config)
+
+    model = Data2VecTextForSequenceClassification(config) if classification_head else Data2VecTextForMaskedLM(config)
+    model.eval()
+
+    # Now let's copy all the weights.
+    # Embeddings
+    model.data2vec_text.embeddings.word_embeddings.weight = data2vec_sent_encoder.embed_tokens.weight
+    model.data2vec_text.embeddings.position_embeddings.weight = data2vec_sent_encoder.embed_positions.weight
+    model.data2vec_text.embeddings.token_type_embeddings.weight.data = torch.zeros_like(
+        model.data2vec_text.embeddings.token_type_embeddings.weight
+    )  # just zero them out b/c data2vec doesn't use them.
+    model.data2vec_text.embeddings.LayerNorm.weight = data2vec_sent_encoder.layernorm_embedding.weight
+    model.data2vec_text.embeddings.LayerNorm.bias = data2vec_sent_encoder.layernorm_embedding.bias
+
+    for i in range(config.num_hidden_layers):
+        # Encoder: start of layer
+        layer: BertLayer = model.data2vec_text.encoder.layer[i]
+        data2vec_layer: TransformerSentenceEncoderLayer = data2vec_sent_encoder.layers[i]
+
+        # self attention
+        self_attn: BertSelfAttention = layer.attention.self
+        assert data2vec_layer.self_attn.k_proj.weight.data.shape == torch.Size(
+            (config.hidden_size, config.hidden_size)
+        ), f"Shape for data2vec_layer.self_attn.k_proj.weight.data should be {torch.Size((config.hidden_size, config.hidden_size))}"
+        assert data2vec_layer.self_attn.q_proj.weight.data.shape == torch.Size(
+            (config.hidden_size, config.hidden_size)
+        ), f"Shape for data2vec_layer.self_attn.q_proj.weight.data should be {torch.Size((config.hidden_size, config.hidden_size))}"
+        assert data2vec_layer.self_attn.v_proj.weight.data.shape == torch.Size(
+            (config.hidden_size, config.hidden_size)
+        ), f"Shape for data2vec_layer.self_attn.v_proj.weight.data should be {torch.Size((config.hidden_size, config.hidden_size))}"
+
+        self_attn.query.weight.data = data2vec_layer.self_attn.q_proj.weight
+        self_attn.query.bias.data = data2vec_layer.self_attn.q_proj.bias
+        self_attn.key.weight.data = data2vec_layer.self_attn.k_proj.weight
+        self_attn.key.bias.data = data2vec_layer.self_attn.k_proj.bias
+        self_attn.value.weight.data = data2vec_layer.self_attn.v_proj.weight
+        self_attn.value.bias.data = data2vec_layer.self_attn.v_proj.bias
+
+        # self-attention output
+        self_output: BertSelfOutput = layer.attention.output
+        assert (
+            self_output.dense.weight.shape == data2vec_layer.self_attn.out_proj.weight.shape
+        ), f"Shape for self_output.dense.weight should be {data2vec_layer.self_attn.out_proj.weight.shape}"
+        self_output.dense.weight = data2vec_layer.self_attn.out_proj.weight
+        self_output.dense.bias = data2vec_layer.self_attn.out_proj.bias
+        self_output.LayerNorm.weight = data2vec_layer.self_attn_layer_norm.weight
+        self_output.LayerNorm.bias = data2vec_layer.self_attn_layer_norm.bias
+
+        # intermediate
+        intermediate: BertIntermediate = layer.intermediate
+        assert (
+            intermediate.dense.weight.shape == data2vec_layer.fc1.weight.shape
+        ), f"Shape for intermediate.dense.weight should be {data2vec_layer.fc1.weight.shape}"
+        intermediate.dense.weight = data2vec_layer.fc1.weight
+        intermediate.dense.bias = data2vec_layer.fc1.bias
+
+        # output
+        bert_output: BertOutput = layer.output
+        assert (
+            bert_output.dense.weight.shape == data2vec_layer.fc2.weight.shape
+        ), f"Shape for bert_output.dense.weight should be {data2vec_layer.fc2.weight.shape}"
+        bert_output.dense.weight = data2vec_layer.fc2.weight
+        bert_output.dense.bias = data2vec_layer.fc2.bias
+        bert_output.LayerNorm.weight = data2vec_layer.final_layer_norm.weight
+        bert_output.LayerNorm.bias = data2vec_layer.final_layer_norm.bias
+        # end of layer
+
+    if classification_head:
+        model.classifier.dense.weight = data2vec.model.classification_heads["mnli"].dense.weight
+        model.classifier.dense.bias = data2vec.model.classification_heads["mnli"].dense.bias
+        model.classifier.out_proj.weight = data2vec.model.classification_heads["mnli"].out_proj.weight
+        model.classifier.out_proj.bias = data2vec.model.classification_heads["mnli"].out_proj.bias
+    else:
+        # LM Head
+        model.lm_head.dense.weight = data2vec_model.encoder.lm_head.dense.weight
+        model.lm_head.dense.bias = data2vec_model.encoder.lm_head.dense.bias
+        model.lm_head.layer_norm.weight = data2vec_model.encoder.lm_head.layer_norm.weight
+        model.lm_head.layer_norm.bias = data2vec_model.encoder.lm_head.layer_norm.bias
+        model.lm_head.decoder.weight = data2vec_model.encoder.lm_head.weight
+        model.lm_head.decoder.bias = data2vec_model.encoder.lm_head.bias
+
+    # Let's check that we get the same results.
+    input_ids: torch.Tensor = data2vec.encode(SAMPLE_TEXT).unsqueeze(0)  # batch of size 1
+
+    our_output = model(input_ids)[0]
+    if classification_head:
+        their_output = data2vec.model.classification_heads["mnli"](data2vec.extract_features(input_ids))
+    else:
+        their_output = data2vec_model(input_ids)[0]
+    print(our_output.shape, their_output.shape)
+    max_absolute_diff = torch.max(torch.abs(our_output - their_output)).item()
+    print(f"max_absolute_diff = {max_absolute_diff}")  # ~ 1e-7
+    success = torch.allclose(our_output, their_output, atol=1e-3)
+    print("Do both models output the same tensors?", "🔥" if success else "💩")
+    if not success:
+        raise Exception("Something went wRoNg")
+
+    pathlib.Path(pytorch_dump_folder_path).mkdir(parents=True, exist_ok=True)
+    print(f"Saving model to {pytorch_dump_folder_path}")
+    model.save_pretrained(pytorch_dump_folder_path)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    # Required parameters
+    parser.add_argument(
+        "--checkpoint_path", default=None, type=str, required=True, help="Path the official PyTorch dump."
+    )
+    parser.add_argument(
+        "--pytorch_dump_folder_path", default=None, type=str, required=True, help="Path to the output PyTorch model."
+    )
+    parser.add_argument(
+        "--classification_head", action="store_true", help="Whether to convert a final classification head."
+    )
+    args = parser.parse_args()
+    convert_data2vec_checkpoint_to_pytorch(
+        args.checkpoint_path, args.pytorch_dump_folder_path, args.classification_head
+    )