[WIP] add EnCodec model (#23655)

* boilerplate stuff

* messing around with the feature extractor

* fix feature extractor

* unit tests for feature extractor

* rename speech to audio

* quick-and-dirty import of Meta's code

* import weights (sort of)

* cleaning up

* more cleaning up

* move encoder/decoder args into config

* cleanup model

* rename EnCodec -> Encodec

* RVQ parameters in config

* add slow test

* add lstm init and test_init

* Add save & load

* finish EncodecModel

* remove decoder_input_values as they are ont used anywhere (not removed from doc yet)

* fix test feature extraction model name

* Add better slow test

* Fix tests

* some fixup and cleaning

* Improve further

* cleaning up quantizer

* fix up conversion script

* test don't pass, _encode_fram does not work

* update tests with output per encode and decode

* more cleanup

* rename _codebook

* remove old config cruft

* ratios & hop_length

* use ModuleList instead of Sequential

* clean up resnet block

* update types

* update tests

* fixup

* quick cleanup

* fix padding

* more styl,ing

* add patrick feedback

* fix copies

* fixup

* fix lstm

* fix shape issues

* fixup

* rename conv layers

* fixup

* fix decoding

* small conv refactoring

* remove norm_params

* simplify conv layers

* rename conv layers

* stuff

* Clean up

* Add padding logic

use padding mask

small conv refactoring

remove norm_params

simplify conv layers

rename conv layers

stuff

add batched test

update

Clean up

merge and update for padding

fix padding

fixup

* clean up more

* clean up more

* More clean ups

* cleanup convolutions

* typo

* fix typos

* fixup

* build PR doc?

* start refactoring docstring

* fix don't pad when no strid and chunk

* update docstring

* update docstring

* nits

* update going to lunch

* update config and model

* fix broken testse (becaue of the config changes)

* fix scale computation

* fixu[

* only return dict if speciefied or if config returns it

* remove todos

* update defaults in config

* update conversion script

* fix doctest

* more docstring + fixup

* nits on batched_tests

* more nits

* Apply suggestions from code review

Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* update basxed on review

* fix update

* updaet tests

* Apply suggestions from code review

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* fixup

* add overlap and chunl_length_s

* cleanup feature extraction

* teste edge cases truncation and padding

* correct processor values

* update config encodec, nits

* fix tests

* fixup

* fix 24Hz test

* elle tests are green

* fix fixup

* Apply suggestions from code review

* revert readme changes

* fixup

* add example

* use facebook checkpoints

* fix typo

* no pipeline tests

* use slef.pad everywhere we can

* Apply suggestions from code review

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* update based on review

* update

* update mdx

* fix bug and tests

* fixup

* fix doctest

* remove comment

* more nits

* add more coverage for `test_truncation_and_padding`

* fixup

* add last test

* fix text

* nits

* Update tests/models/encodec/test_modeling_encodec.py

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* take care of the last comments

* typo

* fix test

* nits

* fixup

* Update src/transformers/models/encodec/feature_extraction_encodec.py

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

---------

Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
Co-authored-by: arthur.zucker@gmail.com <arthur.zucker@gmail.com>
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

README.md

@@ -346,6 +346,7 @@ Current number of checkpoints: ![](https://img.shields.io/endpoint?url=https://h
 1. **[EfficientFormer](https://huggingface.co/docs/transformers/model_doc/efficientformer)** (from Snap Research) released with the paper [EfficientFormer: Vision Transformers at MobileNetSpeed](https://arxiv.org/abs/2206.01191) by Yanyu Li, Geng Yuan, Yang Wen, Ju Hu, Georgios Evangelidis, Sergey Tulyakov, Yanzhi Wang, Jian Ren.
 1. **[EfficientNet](https://huggingface.co/docs/transformers/model_doc/efficientnet)** (from Google Brain) released with the paper [EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks](https://arxiv.org/abs/1905.11946) by Mingxing Tan, Quoc V. Le.
 1. **[ELECTRA](https://huggingface.co/docs/transformers/model_doc/electra)** (from Google Research/Stanford University) released with the paper [ELECTRA: Pre-training text encoders as discriminators rather than generators](https://arxiv.org/abs/2003.10555) by Kevin Clark, Minh-Thang Luong, Quoc V. Le, Christopher D. Manning.
+1. **[EnCodec](https://huggingface.co/docs/transformers/main/model_doc/encodec)** (from Meta AI) released with the paper [High Fidelity Neural Audio Compression](https://arxiv.org/abs/2210.13438) by Alexandre Défossez, Jade Copet, Gabriel Synnaeve, Yossi Adi.
 1. **[EncoderDecoder](https://huggingface.co/docs/transformers/model_doc/encoder-decoder)** (from Google Research) released with the paper [Leveraging Pre-trained Checkpoints for Sequence Generation Tasks](https://arxiv.org/abs/1907.12461) by Sascha Rothe, Shashi Narayan, Aliaksei Severyn.
 1. **[ERNIE](https://huggingface.co/docs/transformers/model_doc/ernie)** (from Baidu) released with the paper [ERNIE: Enhanced Representation through Knowledge Integration](https://arxiv.org/abs/1904.09223) by Yu Sun, Shuohuan Wang, Yukun Li, Shikun Feng, Xuyi Chen, Han Zhang, Xin Tian, Danxiang Zhu, Hao Tian, Hua Wu.
 1. **[ErnieM](https://huggingface.co/docs/transformers/model_doc/ernie_m)** (from Baidu) released with the paper [ERNIE-M: Enhanced Multilingual Representation by Aligning Cross-lingual Semantics with Monolingual Corpora](https://arxiv.org/abs/2012.15674) by Xuan Ouyang, Shuohuan Wang, Chao Pang, Yu Sun, Hao Tian, Hua Wu, Haifeng Wang.

README_es.md

@@ -321,6 +321,7 @@ Número actual de puntos de control: ![](https://img.shields.io/endpoint?url=htt
 1. **[EfficientFormer](https://huggingface.co/docs/transformers/model_doc/efficientformer)** (from Snap Research) released with the paper [EfficientFormer: Vision Transformers at MobileNetSpeed](https://arxiv.org/abs/2206.01191) by Yanyu Li, Geng Yuan, Yang Wen, Ju Hu, Georgios Evangelidis, Sergey Tulyakov, Yanzhi Wang, Jian Ren.
 1. **[EfficientNet](https://huggingface.co/docs/transformers/model_doc/efficientnet)** (from Google Brain) released with the paper [EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks](https://arxiv.org/abs/1905.11946) by Mingxing Tan, Quoc V. Le.
 1. **[ELECTRA](https://huggingface.co/docs/transformers/model_doc/electra)** (from Google Research/Stanford University) released with the paper [ELECTRA: Pre-training text encoders as discriminators rather than generators](https://arxiv.org/abs/2003.10555) by Kevin Clark, Minh-Thang Luong, Quoc V. Le, Christopher D. Manning.
+1. **[EnCodec](https://huggingface.co/docs/transformers/main/model_doc/encodec)** (from Meta AI) released with the paper [High Fidelity Neural Audio Compression](https://arxiv.org/abs/2210.13438) by Alexandre Défossez, Jade Copet, Gabriel Synnaeve, Yossi Adi.
 1. **[EncoderDecoder](https://huggingface.co/docs/transformers/model_doc/encoder-decoder)** (from Google Research) released with the paper [Leveraging Pre-trained Checkpoints for Sequence Generation Tasks](https://arxiv.org/abs/1907.12461) by Sascha Rothe, Shashi Narayan, Aliaksei Severyn.
 1. **[ERNIE](https://huggingface.co/docs/transformers/model_doc/ernie)** (from Baidu) released with the paper [ERNIE: Enhanced Representation through Knowledge Integration](https://arxiv.org/abs/1904.09223) by Yu Sun, Shuohuan Wang, Yukun Li, Shikun Feng, Xuyi Chen, Han Zhang, Xin Tian, Danxiang Zhu, Hao Tian, Hua Wu.
 1. **[ErnieM](https://huggingface.co/docs/transformers/model_doc/ernie_m)** (from Baidu) released with the paper [ERNIE-M: Enhanced Multilingual Representation by Aligning Cross-lingual Semantics with Monolingual Corpora](https://arxiv.org/abs/2012.15674) by Xuan Ouyang, Shuohuan Wang, Chao Pang, Yu Sun, Hao Tian, Hua Wu, Haifeng Wang.

README_hd.md

@@ -293,6 +293,7 @@ conda install -c huggingface transformers
 1. **[EfficientFormer](https://huggingface.co/docs/transformers/model_doc/efficientformer)** (from Snap Research) released with the paper [EfficientFormer: Vision Transformers at MobileNetSpeed](https://arxiv.org/abs/2206.01191) by Yanyu Li, Geng Yuan, Yang Wen, Ju Hu, Georgios Evangelidis, Sergey Tulyakov, Yanzhi Wang, Jian Ren.
 1. **[EfficientNet](https://huggingface.co/docs/transformers/model_doc/efficientnet)** (from Google Brain) released with the paper [EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks](https://arxiv.org/abs/1905.11946) by Mingxing Tan, Quoc V. Le.
 1. **[ELECTRA](https://huggingface.co/docs/transformers/model_doc/electra)** (Google रिसर्च/स्टैनफोर्ड यूनिवर्सिटी से) साथ में दिया गया पेपर [इलेक्ट्रा: जेनरेटर के बजाय भेदभाव करने वाले के रूप में टेक्स्ट एन्कोडर्स का पूर्व-प्रशिक्षण] (https://arxiv.org/abs/2003.10555) केविन क्लार्क, मिन्ह-थांग लुओंग, क्वोक वी. ले, क्रिस्टोफर डी. मैनिंग द्वारा पोस्ट किया गया।
+1. **[EnCodec](https://huggingface.co/docs/transformers/main/model_doc/encodec)** (Meta AI से) Alexandre Défossez, Jade Copet, Gabriel Synnaeve, Yossi Adi. द्वाराअनुसंधान पत्र [High Fidelity Neural Audio Compression](https://arxiv.org/abs/2210.13438) के साथ जारी किया गया
 1. **[EncoderDecoder](https://huggingface.co/docs/transformers/model_doc/encoder-decoder)** (Google रिसर्च से) साथ में दिया गया पेपर [सीक्वेंस जेनरेशन टास्क के लिए प्री-ट्रेंड चेकपॉइंट का इस्तेमाल करना](https:/ /arxiv.org/abs/1907.12461) साशा रोठे, शशि नारायण, अलियाक्सि सेवेरिन द्वारा।
 1. **[ERNIE](https://huggingface.co/docs/transformers/model_doc/ernie)**(Baidu से) साथ देने वाला पेपर [ERNIE: एन्हांस्ड रिप्रेजेंटेशन थ्रू नॉलेज इंटीग्रेशन](https://arxiv.org/abs/1904.09223) यू सन, शुओहुआन वांग, युकुन ली, शिकुन फेंग, ज़ुई चेन, हान झांग, शिन तियान, डैनक्सियांग झू, हाओ तियान, हुआ वू द्वारा पोस्ट किया गया।
 1. **[ErnieM](https://huggingface.co/docs/transformers/model_doc/ernie_m)** (Baidu से) Xuan Ouyang, Shuohuan Wang, Chao Pang, Yu Sun, Hao Tian, Hua Wu, Haifeng Wang. द्वाराअनुसंधान पत्र [ERNIE-M: Enhanced Multilingual Representation by Aligning Cross-lingual Semantics with Monolingual Corpora](https://arxiv.org/abs/2012.15674) के साथ जारी किया गया

README_ja.md

@@ -355,6 +355,7 @@ Flax、PyTorch、TensorFlowをcondaでインストールする方法は、それ
 1. **[EfficientFormer](https://huggingface.co/docs/transformers/model_doc/efficientformer)** (Snap Research から) Yanyu Li, Geng Yuan, Yang Wen, Ju Hu, Georgios Evangelidis, Sergey Tulyakov, Yanzhi Wang, Jian Ren. から公開された研究論文 [EfficientFormer: Vision Transformers at MobileNetSpeed](https://arxiv.org/abs/2206.01191)
 1. **[EfficientNet](https://huggingface.co/docs/transformers/model_doc/efficientnet)** (from Google Brain) released with the paper [EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks](https://arxiv.org/abs/1905.11946) by Mingxing Tan, Quoc V. Le.
 1. **[ELECTRA](https://huggingface.co/docs/transformers/model_doc/electra)** (Google Research/Stanford University から) Kevin Clark, Minh-Thang Luong, Quoc V. Le, Christopher D. Manning から公開された研究論文: [ELECTRA: Pre-training text encoders as discriminators rather than generators](https://arxiv.org/abs/2003.10555)
+1. **[EnCodec](https://huggingface.co/docs/transformers/main/model_doc/encodec)** (Meta AI から) Alexandre Défossez, Jade Copet, Gabriel Synnaeve, Yossi Adi. から公開された研究論文 [High Fidelity Neural Audio Compression](https://arxiv.org/abs/2210.13438)
 1. **[EncoderDecoder](https://huggingface.co/docs/transformers/model_doc/encoder-decoder)** (Google Research から) Sascha Rothe, Shashi Narayan, Aliaksei Severyn から公開された研究論文: [Leveraging Pre-trained Checkpoints for Sequence Generation Tasks](https://arxiv.org/abs/1907.12461)
 1. **[ERNIE](https://huggingface.co/docs/transformers/model_doc/ernie)** (Baidu から) Yu Sun, Shuohuan Wang, Yukun Li, Shikun Feng, Xuyi Chen, Han Zhang, Xin Tian, Danxiang Zhu, Hao Tian, Hua Wu から公開された研究論文: [ERNIE: Enhanced Representation through Knowledge Integration](https://arxiv.org/abs/1904.09223)
 1. **[ErnieM](https://huggingface.co/docs/transformers/model_doc/ernie_m)** (Baidu から) Xuan Ouyang, Shuohuan Wang, Chao Pang, Yu Sun, Hao Tian, Hua Wu, Haifeng Wang. から公開された研究論文 [ERNIE-M: Enhanced Multilingual Representation by Aligning Cross-lingual Semantics with Monolingual Corpora](https://arxiv.org/abs/2012.15674)

README_ko.md

@@ -270,6 +270,7 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
 1. **[EfficientFormer](https://huggingface.co/docs/transformers/model_doc/efficientformer)** (from Snap Research) released with the paper [EfficientFormer: Vision Transformers at MobileNetSpeed](https://arxiv.org/abs/2206.01191) by Yanyu Li, Geng Yuan, Yang Wen, Ju Hu, Georgios Evangelidis, Sergey Tulyakov, Yanzhi Wang, Jian Ren.
 1. **[EfficientNet](https://huggingface.co/docs/transformers/model_doc/efficientnet)** (from Google Brain) released with the paper [EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks](https://arxiv.org/abs/1905.11946) by Mingxing Tan, Quoc V. Le.
 1. **[ELECTRA](https://huggingface.co/docs/transformers/model_doc/electra)** (Google Research/Stanford University 에서) Kevin Clark, Minh-Thang Luong, Quoc V. Le, Christopher D. Manning 의 [ELECTRA: Pre-training text encoders as discriminators rather than generators](https://arxiv.org/abs/2003.10555) 논문과 함께 발표했습니다.
+1. **[EnCodec](https://huggingface.co/docs/transformers/main/model_doc/encodec)** (Meta AI 에서 제공)은 Alexandre Défossez, Jade Copet, Gabriel Synnaeve, Yossi Adi.의 [High Fidelity Neural Audio Compression](https://arxiv.org/abs/2210.13438)논문과 함께 발표했습니다.
 1. **[EncoderDecoder](https://huggingface.co/docs/transformers/model_doc/encoder-decoder)** (Google Research 에서) Sascha Rothe, Shashi Narayan, Aliaksei Severyn 의 [Leveraging Pre-trained Checkpoints for Sequence Generation Tasks](https://arxiv.org/abs/1907.12461) 논문과 함께 발표했습니다.
 1. **[ERNIE](https://huggingface.co/docs/transformers/model_doc/ernie)** (Baidu 에서) Yu Sun, Shuohuan Wang, Yukun Li, Shikun Feng, Xuyi Chen, Han Zhang, Xin Tian, Danxiang Zhu, Hao Tian, Hua Wu 의 [ERNIE: Enhanced Representation through Knowledge Integration](https://arxiv.org/abs/1904.09223) 논문과 함께 발표했습니다.
 1. **[ErnieM](https://huggingface.co/docs/transformers/model_doc/ernie_m)** (Baidu 에서 제공)은 Xuan Ouyang, Shuohuan Wang, Chao Pang, Yu Sun, Hao Tian, Hua Wu, Haifeng Wang.의 [ERNIE-M: Enhanced Multilingual Representation by Aligning Cross-lingual Semantics with Monolingual Corpora](https://arxiv.org/abs/2012.15674)논문과 함께 발표했습니다.

README_zh-hans.md

@@ -294,6 +294,7 @@ conda install -c huggingface transformers
 1. **[EfficientFormer](https://huggingface.co/docs/transformers/model_doc/efficientformer)** (来自 Snap Research) 伴随论文 [EfficientFormer: Vision Transformers at MobileNetSpeed](https://arxiv.org/abs/2206.01191) 由 Yanyu Li, Geng Yuan, Yang Wen, Ju Hu, Georgios Evangelidis, Sergey Tulyakov, Yanzhi Wang, Jian Ren 发布。
 1. **[EfficientNet](https://huggingface.co/docs/transformers/model_doc/efficientnet)** (from Google Brain) released with the paper [EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks](https://arxiv.org/abs/1905.11946) by Mingxing Tan, Quoc V. Le.
 1. **[ELECTRA](https://huggingface.co/docs/transformers/model_doc/electra)** (来自 Google Research/Stanford University) 伴随论文 [ELECTRA: Pre-training text encoders as discriminators rather than generators](https://arxiv.org/abs/2003.10555) 由 Kevin Clark, Minh-Thang Luong, Quoc V. Le, Christopher D. Manning 发布。
+1. **[EnCodec](https://huggingface.co/docs/transformers/main/model_doc/encodec)** (来自 Meta AI) 伴随论文 [High Fidelity Neural Audio Compression](https://arxiv.org/abs/2210.13438) 由 Alexandre Défossez, Jade Copet, Gabriel Synnaeve, Yossi Adi 发布。
 1. **[EncoderDecoder](https://huggingface.co/docs/transformers/model_doc/encoder-decoder)** (来自 Google Research) 伴随论文 [Leveraging Pre-trained Checkpoints for Sequence Generation Tasks](https://arxiv.org/abs/1907.12461) 由 Sascha Rothe, Shashi Narayan, Aliaksei Severyn 发布。
 1. **[ERNIE](https://huggingface.co/docs/transformers/model_doc/ernie)** (来自 Baidu) 伴随论文 [ERNIE: Enhanced Representation through Knowledge Integration](https://arxiv.org/abs/1904.09223) by Yu Sun, Shuohuan Wang, Yukun Li, Shikun Feng, Xuyi Chen, Han Zhang, Xin Tian, Danxiang Zhu, Hao Tian, Hua Wu 发布。
 1. **[ErnieM](https://huggingface.co/docs/transformers/model_doc/ernie_m)** (来自 Baidu) 伴随论文 [ERNIE-M: Enhanced Multilingual Representation by Aligning Cross-lingual Semantics with Monolingual Corpora](https://arxiv.org/abs/2012.15674) 由 Xuan Ouyang, Shuohuan Wang, Chao Pang, Yu Sun, Hao Tian, Hua Wu, Haifeng Wang 发布。

README_zh-hant.md

@@ -306,6 +306,7 @@ conda install -c huggingface transformers
 1. **[EfficientFormer](https://huggingface.co/docs/transformers/model_doc/efficientformer)** (from Snap Research) released with the paper [EfficientFormer: Vision Transformers at MobileNetSpeed](https://arxiv.org/abs/2206.01191) by Yanyu Li, Geng Yuan, Yang Wen, Ju Hu, Georgios Evangelidis, Sergey Tulyakov, Yanzhi Wang, Jian Ren.
 1. **[EfficientNet](https://huggingface.co/docs/transformers/model_doc/efficientnet)** (from Google Brain) released with the paper [EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks](https://arxiv.org/abs/1905.11946) by Mingxing Tan, Quoc V. Le.
 1. **[ELECTRA](https://huggingface.co/docs/transformers/model_doc/electra)** (from Google Research/Stanford University) released with the paper [ELECTRA: Pre-training text encoders as discriminators rather than generators](https://arxiv.org/abs/2003.10555) by Kevin Clark, Minh-Thang Luong, Quoc V. Le, Christopher D. Manning.
+1. **[EnCodec](https://huggingface.co/docs/transformers/main/model_doc/encodec)** (from Meta AI) released with the paper [High Fidelity Neural Audio Compression](https://arxiv.org/abs/2210.13438) by Alexandre Défossez, Jade Copet, Gabriel Synnaeve, Yossi Adi.
 1. **[EncoderDecoder](https://huggingface.co/docs/transformers/model_doc/encoder-decoder)** (from Google Research) released with the paper [Leveraging Pre-trained Checkpoints for Sequence Generation Tasks](https://arxiv.org/abs/1907.12461) by Sascha Rothe, Shashi Narayan, Aliaksei Severyn.
 1. **[ERNIE](https://huggingface.co/docs/transformers/model_doc/ernie)** (from Baidu) released with the paper [ERNIE: Enhanced Representation through Knowledge Integration](https://arxiv.org/abs/1904.09223) by Yu Sun, Shuohuan Wang, Yukun Li, Shikun Feng, Xuyi Chen, Han Zhang, Xin Tian, Danxiang Zhu, Hao Tian, Hua Wu.
 1. **[ErnieM](https://huggingface.co/docs/transformers/model_doc/ernie_m)** (from Baidu) released with the paper [ERNIE-M: Enhanced Multilingual Representation by Aligning Cross-lingual Semantics with Monolingual Corpora](https://arxiv.org/abs/2012.15674) by Xuan Ouyang, Shuohuan Wang, Chao Pang, Yu Sun, Hao Tian, Hua Wu, Haifeng Wang.

docs/source/en/_toctree.yml

@@ -541,6 +541,8 @@
         title: Audio Spectrogram Transformer
       - local: model_doc/clap
         title: CLAP
+      - local: model_doc/encodec
+        title: EnCodec
       - local: model_doc/hubert
         title: Hubert
       - local: model_doc/mctct

docs/source/en/index.mdx

@@ -107,6 +107,7 @@ The documentation is organized into five sections:
 1. **[EfficientFormer](model_doc/efficientformer)** (from Snap Research) released with the paper [EfficientFormer: Vision Transformers at MobileNetSpeed](https://arxiv.org/abs/2206.01191) by Yanyu Li, Geng Yuan, Yang Wen, Ju Hu, Georgios Evangelidis, Sergey Tulyakov, Yanzhi Wang, Jian Ren.
 1. **[EfficientNet](model_doc/efficientnet)** (from Google Brain) released with the paper [EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks](https://arxiv.org/abs/1905.11946) by Mingxing Tan, Quoc V. Le.
 1. **[ELECTRA](model_doc/electra)** (from Google Research/Stanford University) released with the paper [ELECTRA: Pre-training text encoders as discriminators rather than generators](https://arxiv.org/abs/2003.10555) by Kevin Clark, Minh-Thang Luong, Quoc V. Le, Christopher D. Manning.
+1. **[EnCodec](model_doc/encodec)** (from Meta AI) released with the paper [High Fidelity Neural Audio Compression](https://arxiv.org/abs/2210.13438) by Alexandre Défossez, Jade Copet, Gabriel Synnaeve, Yossi Adi.
 1. **[EncoderDecoder](model_doc/encoder-decoder)** (from Google Research) released with the paper [Leveraging Pre-trained Checkpoints for Sequence Generation Tasks](https://arxiv.org/abs/1907.12461) by Sascha Rothe, Shashi Narayan, Aliaksei Severyn.
 1. **[ERNIE](model_doc/ernie)** (from Baidu) released with the paper [ERNIE: Enhanced Representation through Knowledge Integration](https://arxiv.org/abs/1904.09223) by Yu Sun, Shuohuan Wang, Yukun Li, Shikun Feng, Xuyi Chen, Han Zhang, Xin Tian, Danxiang Zhu, Hao Tian, Hua Wu.
 1. **[ErnieM](model_doc/ernie_m)** (from Baidu) released with the paper [ERNIE-M: Enhanced Multilingual Representation by Aligning Cross-lingual Semantics with Monolingual Corpora](https://arxiv.org/abs/2012.15674) by Xuan Ouyang, Shuohuan Wang, Chao Pang, Yu Sun, Hao Tian, Hua Wu, Haifeng Wang.
@@ -318,6 +319,7 @@ Flax), PyTorch, and/or TensorFlow.
 |        EfficientFormer        |       ❌       |       ❌       |       ✅        |         ✅         |      ❌      |
 |         EfficientNet          |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |            ELECTRA            |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
+|            EnCodec            |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |        Encoder decoder        |       ❌       |       ❌       |       ✅        |         ✅         |      ✅      |
 |             ERNIE             |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |            ErnieM             |       ✅       |       ❌       |       ✅        |         ❌         |      ❌      |

docs/source/en/model_doc/encodec.mdx

@@ -0,0 +1,59 @@
+<!--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.
+-->
+
+# EnCodec
+
+## Overview
+
+The EnCodec neural codec model was proposed in [High Fidelity Neural Audio Compression](https://arxiv.org/abs/2210.13438) by Alexandre Défossez, Jade Copet, Gabriel Synnaeve, Yossi Adi.
+
+The abstract from the paper is the following:
+
+*We introduce a state-of-the-art real-time, high-fidelity, audio codec leveraging neural networks. It consists in a streaming encoder-decoder architecture with quantized latent space trained in an end-to-end fashion. We simplify and speed-up the training by using a single multiscale spectrogram adversary that efficiently reduces artifacts and produce high-quality samples. We introduce a novel loss balancer mechanism to stabilize training: the weight of a loss now defines the fraction of the overall gradient it should represent, thus decoupling the choice of this hyper-parameter from the typical scale of the loss. Finally, we study how lightweight Transformer models can be used to further compress the obtained representation by up to 40%, while staying faster than real time. We provide a detailed description of the key design choices of the proposed model including: training objective, architectural changes and a study of various perceptual loss functions. We present an extensive subjective evaluation (MUSHRA tests) together with an ablation study for a range of bandwidths and audio domains, including speech, noisy-reverberant speech, and music. Our approach is superior to the baselines methods across all evaluated settings, considering both 24 kHz monophonic and 48 kHz stereophonic audio.*
+
+This model was contributed by [Matthijs](https://huggingface.co/Matthijs), [Patrick Von Platen](https://huggingface.co/patrickvonplaten) and [Arthur Zucker](https://huggingface.co/ArthurZ). 
+The original code can be found [here](https://github.com/facebookresearch/encodec).
+Here is a quick example of how to encode and decode an audio using this model:
+
+```python 
+>>> from datasets import load_dataset, Audio
+>>> from transformers import EncodecModel, AutoProcessor
+>>> librispeech_dummy = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
+
+>>> model = EncodecModel.from_pretrained("facebook/encodec_24khz")
+>>> processor = AutoProcessor.from_pretrained("facebook/encodec_24khz")
+>>> librispeech_dummy = librispeech_dummy.cast_column("audio", Audio(sampling_rate=processor.sampling_rate))
+>>> audio_sample = librispeech_dummy[-1]["audio"]["array"]
+>>> inputs = processor(raw_audio=audio_sample, sampling_rate=processor.sampling_rate, return_tensors="pt")
+
+>>> encoder_outputs = model.encode(inputs["input_values"], inputs["padding_mask"])
+>>> audio_values = model.decode(encoder_outputs.audio_codes, encoder_outputs.audio_scales, inputs["padding_mask"])[0]
+>>> # or the equivalent with a forward pass
+>>> audio_values = model(inputs["input_values"], inputs["padding_mask"]).audio_values
+```
+
+
+## EncodecConfig
+
+[[autodoc]] EncodecConfig
+
+## EncodecFeatureExtractor
+
+[[autodoc]] EncodecFeatureExtractor
+    - __call__
+
+## EncodecModel
+
+[[autodoc]] EncodecModel
+    - decode
+    - encode
+    - forward

src/transformers/__init__.py

@@ -281,6 +281,10 @@ _import_structure = {
     "models.efficientformer": ["EFFICIENTFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP", "EfficientFormerConfig"],
     "models.efficientnet": ["EFFICIENTNET_PRETRAINED_CONFIG_ARCHIVE_MAP", "EfficientNetConfig"],
     "models.electra": ["ELECTRA_PRETRAINED_CONFIG_ARCHIVE_MAP", "ElectraConfig", "ElectraTokenizer"],
+    "models.encodec": [
+        "ENCODEC_PRETRAINED_CONFIG_ARCHIVE_MAP",
+        "EncodecConfig",
+    ],
     "models.encoder_decoder": ["EncoderDecoderConfig"],
     "models.ernie": [
         "ERNIE_PRETRAINED_CONFIG_ARCHIVE_MAP",
@@ -825,6 +829,7 @@ except OptionalDependencyNotAvailable:
     ]
 else:
     _import_structure["models.audio_spectrogram_transformer"].append("ASTFeatureExtractor")
+    _import_structure["models.encodec"].append("EncodecFeatureExtractor")
     _import_structure["models.mctct"].append("MCTCTFeatureExtractor")
     _import_structure["models.speech_to_text"].append("Speech2TextFeatureExtractor")
     _import_structure["models.speecht5"].append("SpeechT5FeatureExtractor")
@@ -1568,6 +1573,13 @@ else:
             "load_tf_weights_in_electra",
         ]
     )
+    _import_structure["models.encodec"].extend(
+        [
+            "ENCODEC_PRETRAINED_MODEL_ARCHIVE_LIST",
+            "EncodecModel",
+            "EncodecPreTrainedModel",
+        ]
+    )
     _import_structure["models.encoder_decoder"].append("EncoderDecoderModel")
     _import_structure["models.ernie"].extend(
         [
@@ -4100,6 +4112,10 @@ if TYPE_CHECKING:
     from .models.efficientformer import EFFICIENTFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP, EfficientFormerConfig
     from .models.efficientnet import EFFICIENTNET_PRETRAINED_CONFIG_ARCHIVE_MAP, EfficientNetConfig
     from .models.electra import ELECTRA_PRETRAINED_CONFIG_ARCHIVE_MAP, ElectraConfig, ElectraTokenizer
+    from .models.encodec import (
+        ENCODEC_PRETRAINED_CONFIG_ARCHIVE_MAP,
+        EncodecConfig,
+    )
     from .models.encoder_decoder import EncoderDecoderConfig
     from .models.ernie import ERNIE_PRETRAINED_CONFIG_ARCHIVE_MAP, ErnieConfig
     from .models.ernie_m import ERNIE_M_PRETRAINED_CONFIG_ARCHIVE_MAP, ErnieMConfig
@@ -4598,6 +4614,7 @@ if TYPE_CHECKING:
         from .utils.dummy_speech_objects import *
     else:
         from .models.audio_spectrogram_transformer import ASTFeatureExtractor
+        from .models.encodec import EncodecFeatureExtractor
         from .models.mctct import MCTCTFeatureExtractor
         from .models.speech_to_text import Speech2TextFeatureExtractor
         from .models.speecht5 import SpeechT5FeatureExtractor
@@ -5210,6 +5227,11 @@ if TYPE_CHECKING:
             ElectraPreTrainedModel,
             load_tf_weights_in_electra,
         )
+        from .models.encodec import (
+            ENCODEC_PRETRAINED_MODEL_ARCHIVE_LIST,
+            EncodecModel,
+            EncodecPreTrainedModel,
+        )
         from .models.encoder_decoder import EncoderDecoderModel
         from .models.ernie import (
             ERNIE_PRETRAINED_MODEL_ARCHIVE_LIST,

src/transformers/models/__init__.py

@@ -72,6 +72,7 @@ from . import (
     efficientformer,
     efficientnet,
     electra,
+    encodec,
     encoder_decoder,
     ernie,
     ernie_m,

src/transformers/models/auto/configuration_auto.py

@@ -80,6 +80,7 @@ CONFIG_MAPPING_NAMES = OrderedDict(
         ("efficientformer", "EfficientFormerConfig"),
         ("efficientnet", "EfficientNetConfig"),
         ("electra", "ElectraConfig"),
+        ("encodec", "EncodecConfig"),
         ("encoder-decoder", "EncoderDecoderConfig"),
         ("ernie", "ErnieConfig"),
         ("ernie_m", "ErnieMConfig"),
@@ -273,6 +274,7 @@ CONFIG_ARCHIVE_MAP_MAPPING_NAMES = OrderedDict(
         ("efficientformer", "EFFICIENTFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("efficientnet", "EFFICIENTNET_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("electra", "ELECTRA_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("encodec", "ENCODEC_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("ernie", "ERNIE_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("ernie_m", "ERNIE_M_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("esm", "ESM_PRETRAINED_CONFIG_ARCHIVE_MAP"),
@@ -461,6 +463,7 @@ MODEL_NAMES_MAPPING = OrderedDict(
         ("efficientformer", "EfficientFormer"),
         ("efficientnet", "EfficientNet"),
         ("electra", "ELECTRA"),
+        ("encodec", "EnCodec"),
         ("encoder-decoder", "Encoder decoder"),
         ("ernie", "ERNIE"),
         ("ernie_m", "ErnieM"),

src/transformers/models/auto/feature_extraction_auto.py

@@ -54,6 +54,7 @@ FEATURE_EXTRACTOR_MAPPING_NAMES = OrderedDict(
         ("dinat", "ViTFeatureExtractor"),
         ("donut-swin", "DonutFeatureExtractor"),
         ("dpt", "DPTFeatureExtractor"),
+        ("encodec", "EncodecFeatureExtractor"),
         ("flava", "FlavaFeatureExtractor"),
         ("glpn", "GLPNFeatureExtractor"),
         ("groupvit", "CLIPFeatureExtractor"),

src/transformers/models/auto/modeling_auto.py

@@ -79,6 +79,7 @@ MODEL_MAPPING_NAMES = OrderedDict(
         ("efficientformer", "EfficientFormerModel"),
         ("efficientnet", "EfficientNetModel"),
         ("electra", "ElectraModel"),
+        ("encodec", "EncodecModel"),
         ("ernie", "ErnieModel"),
         ("ernie_m", "ErnieMModel"),
         ("esm", "EsmModel"),

src/transformers/models/encodec/__init__.py

@@ -0,0 +1,65 @@
+# 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_encodec": [
+        "ENCODEC_PRETRAINED_CONFIG_ARCHIVE_MAP",
+        "EncodecConfig",
+    ],
+    "feature_extraction_encodec": ["EncodecFeatureExtractor"],
+}
+
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_encodec"] = [
+        "ENCODEC_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "EncodecModel",
+        "EncodecPreTrainedModel",
+    ]
+
+if TYPE_CHECKING:
+    from .configuration_encodec import (
+        ENCODEC_PRETRAINED_CONFIG_ARCHIVE_MAP,
+        EncodecConfig,
+    )
+    from .feature_extraction_encodec import EncodecFeatureExtractor
+
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_encodec import (
+            ENCODEC_PRETRAINED_MODEL_ARCHIVE_LIST,
+            EncodecModel,
+            EncodecPreTrainedModel,
+        )
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)

src/transformers/models/encodec/configuration_encodec.py

@@ -0,0 +1,189 @@
+# coding=utf-8
+# Copyright 2023 Meta Platforms, Inc. and affiliates, and 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.
+""" EnCodec model configuration"""
+
+
+import math
+from typing import Optional
+
+import numpy as np
+
+from ...configuration_utils import PretrainedConfig
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+ENCODEC_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "facebook/encodec_24khz": "https://huggingface.co/facebook/encodec_24khz/resolve/main/config.json",
+    "facebook/encodec_48khz": "https://huggingface.co/facebook/encodec_48khz/resolve/main/config.json",
+}
+
+
+class EncodecConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of an [`EncodecModel`]. It is used to instantiate a
+    Encodec 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
+    [facebook/encodec_24khz](https://huggingface.co/facebook/encodec_24khz) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        target_bandwidths (`List[float]`, *optional*, defaults to `[1.5, 3.0, 6.0, 12.0, 24.0]`):
+            The range of diffent bandwiths the model can encode audio with.
+        sampling_rate (`int`, *optional*, defaults to 24000):
+            The sampling rate at which the audio waveform should be digitalized expressed in hertz (Hz).
+        audio_channels (`int`, *optional*, defaults to 1):
+            Number of channels in the audio data. Either 1 for mono or 2 for stereo.
+        normalize (`bool`, *optional*, defaults to `False`):
+            Whether the audio shall be normalized when passed.
+        chunk_length_s (`float`, *optional*):
+            If defined the audio is pre-processed into chunks of lengths `chunk_length_s` and then encoded.
+        overlap (`float`, *optional*):
+            Defines the overlap between each chunk. It is used to compute the `chunk_stride` using the following
+            formulae : `int((1.0 - self.overlap) * self.chunk_length)`.
+        hidden_size (`int`, *optional*, defaults to 128):
+            Intermediate representation dimension.
+        num_filters (`int`, *optional*, defaults to 32):
+            Number of convolution kernels of first `EncodecConv1d` down sampling layer.
+        num_residual_layers (`int`,  *optional*, defaults to 1):
+            Number of residual layers.
+        upsampling_ratios (`Sequence[int]` , *optional*, defaults to `[8, 5, 4, 2]`):
+            Kernel size and stride ratios. The encoder uses downsampling ratios instead of upsampling ratios, hence it
+            will use the ratios in the reverse order to the ones specified here that must match the decoder order.
+        norm_type (`str`, *optional*, defaults to `"weight_norm"`):
+            Normalization method. Should be in `["weight_norm", "time_group_norm"]`
+        kernel_size (`int`, *optional*, defaults to 7):
+            Kernel size for the initial convolution.
+        last_kernel_size (`int`, *optional*, defaults to 7):
+            Kernel size for the last convolution layer.
+        residual_kernel_size (`int`, *optional*, defaults to 3):
+            Kernel size for the residual layers.
+        dilation_growth_rate (`int`, *optional*, defaults to 2):
+            How much to increase the dilation with each layer.
+        use_causal_conv (`bool`, *optional*, defaults to `True`):
+            Whether to use fully causal convolution.
+        pad_mode (`str`, *optional*, defaults to `"reflect"`):
+            Padding mode for the convolutions.
+        compress (`int`, *optional*, defaults to 2):
+            Reduced dimensionality in residual branches (from Demucs v3).
+        num_lstm_layers (`int`, *optional*, defaults to 2):
+            Number of LSTM layers at the end of the encoder.
+        trim_right_ratio (`float`, *optional*, defaults to 1.0):
+            Ratio for trimming at the right of the transposed convolution under the `use_causal_conv = True` setup. If
+            equal to 1.0, it means that all the trimming is done at the right.
+        codebook_size (`int`, *optional*, defaults to 1024):
+            Number of discret codes that make up VQVAE.
+        codebook_dim (`int`, *optional*):
+            Dimension of the codebook vectors. If not defined, uses `hidden_size`.
+
+    Example:
+
+    ```python
+    >>> from transformers import EncodecModel, EncodecConfig
+
+    >>> # Initializing a "facebook/encodec_24khz" style configuration
+    >>> configuration = EncodecConfig()
+
+    >>> # Initializing a model (with random weights) from the "facebook/encodec_24khz" style configuration
+    >>> model = EncodecModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+    model_type = "encodec"
+
+    def __init__(
+        self,
+        target_bandwidths=[1.5, 3.0, 6.0, 12.0, 24.0],
+        sampling_rate=24_000,
+        audio_channels=1,
+        normalize=False,
+        chunk_length_s=None,
+        overlap=None,
+        hidden_size=128,
+        num_filters=32,
+        num_residual_layers=1,
+        upsampling_ratios=[8, 5, 4, 2],
+        norm_type="weight_norm",
+        kernel_size=7,
+        last_kernel_size=7,
+        residual_kernel_size=3,
+        dilation_growth_rate=2,
+        use_causal_conv=True,
+        pad_mode="reflect",
+        compress=2,
+        num_lstm_layers=2,
+        trim_right_ratio=1.0,
+        codebook_size=1024,
+        codebook_dim=None,
+        **kwargs,
+    ):
+        self.target_bandwidths = target_bandwidths
+        self.sampling_rate = sampling_rate
+        self.audio_channels = audio_channels
+        self.normalize = normalize
+        self.chunk_length_s = chunk_length_s
+        self.overlap = overlap
+        self.hidden_size = hidden_size
+        self.num_filters = num_filters
+        self.num_residual_layers = num_residual_layers
+        self.upsampling_ratios = upsampling_ratios
+        self.norm_type = norm_type
+        self.kernel_size = kernel_size
+        self.last_kernel_size = last_kernel_size
+        self.residual_kernel_size = residual_kernel_size
+        self.dilation_growth_rate = dilation_growth_rate
+        self.use_causal_conv = use_causal_conv
+        self.pad_mode = pad_mode
+        self.compress = compress
+        self.num_lstm_layers = num_lstm_layers
+        self.trim_right_ratio = trim_right_ratio
+        self.codebook_size = codebook_size
+        self.codebook_dim = codebook_dim if codebook_dim is not None else hidden_size
+
+        if self.norm_type not in ["weight_norm", "time_group_norm"]:
+            raise ValueError(
+                f'self.norm_type must be one of `"weight_norm"`, `"time_group_norm"`), got {self.norm_type}'
+            )
+
+        super().__init__(**kwargs)
+
+    # This is a property because you might want to change the chunk_length_s on the fly
+    @property
+    def chunk_length(self) -> Optional[int]:
+        if self.chunk_length_s is None:
+            return None
+        else:
+            return int(self.chunk_length_s * self.sampling_rate)
+
+    # This is a property because you might want to change the chunk_length_s on the fly
+    @property
+    def chunk_stride(self) -> Optional[int]:
+        if self.chunk_length_s is None or self.overlap is None:
+            return None
+        else:
+            return max(1, int((1.0 - self.overlap) * self.chunk_length))
+
+    @property
+    def frame_rate(self) -> int:
+        hop_length = np.prod(self.upsampling_ratios)
+        return math.ceil(self.sampling_rate / hop_length)
+
+    @property
+    def num_quantizers(self) -> int:
+        return int(1000 * self.target_bandwidths[-1] // (self.frame_rate * 10))

src/transformers/models/encodec/convert_encodec_checkpoint_to_pytorch.py

@@ -0,0 +1,352 @@
+# 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.
+"""Convert EnCodec checkpoints."""
+
+import argparse
+
+import torch
+
+from transformers import (
+    EncodecConfig,
+    EncodecFeatureExtractor,
+    EncodecModel,
+    logging,
+)
+
+
+# checkpoints downloaded from:
+# https://dl.fbaipublicfiles.com/encodec/v0/encodec_24khz-d7cc33bc.th
+# https://dl.fbaipublicfiles.com/encodec/v0/encodec_48khz-7e698e3e.th
+
+
+logging.set_verbosity_info()
+logger = logging.get_logger("transformers.models.encodec")
+
+MAPPING_QUANTIZER = {
+    "quantizer.vq.layers.*._codebook.inited": "quantizer.layers.*.codebook.inited",
+    "quantizer.vq.layers.*._codebook.cluster_size": "quantizer.layers.*.codebook.cluster_size",
+    "quantizer.vq.layers.*._codebook.embed": "quantizer.layers.*.codebook.embed",
+    "quantizer.vq.layers.*._codebook.embed_avg": "quantizer.layers.*.codebook.embed_avg",
+}
+MAPPING_ENCODER = {
+    "encoder.model.0.conv.conv": "encoder.layers.0.conv",
+    "encoder.model.1.block.1.conv.conv": "encoder.layers.1.block.1.conv",
+    "encoder.model.1.block.3.conv.conv": "encoder.layers.1.block.3.conv",
+    "encoder.model.1.shortcut.conv.conv": "encoder.layers.1.shortcut.conv",
+    "encoder.model.3.conv.conv": "encoder.layers.3.conv",
+    "encoder.model.4.block.1.conv.conv": "encoder.layers.4.block.1.conv",
+    "encoder.model.4.block.3.conv.conv": "encoder.layers.4.block.3.conv",
+    "encoder.model.4.shortcut.conv.conv": "encoder.layers.4.shortcut.conv",
+    "encoder.model.6.conv.conv": "encoder.layers.6.conv",
+    "encoder.model.7.block.1.conv.conv": "encoder.layers.7.block.1.conv",
+    "encoder.model.7.block.3.conv.conv": "encoder.layers.7.block.3.conv",
+    "encoder.model.7.shortcut.conv.conv": "encoder.layers.7.shortcut.conv",
+    "encoder.model.9.conv.conv": "encoder.layers.9.conv",
+    "encoder.model.10.block.1.conv.conv": "encoder.layers.10.block.1.conv",
+    "encoder.model.10.block.3.conv.conv": "encoder.layers.10.block.3.conv",
+    "encoder.model.10.shortcut.conv.conv": "encoder.layers.10.shortcut.conv",
+    "encoder.model.12.conv.conv": "encoder.layers.12.conv",
+    "encoder.model.13.lstm": "encoder.layers.13.lstm",
+    "encoder.model.15.conv.conv": "encoder.layers.15.conv",
+}
+MAPPING_ENCODER_48K = {
+    "encoder.model.0.conv.norm": "encoder.layers.0.norm",
+    "encoder.model.1.block.1.conv.norm": "encoder.layers.1.block.1.norm",
+    "encoder.model.1.block.3.conv.norm": "encoder.layers.1.block.3.norm",
+    "encoder.model.1.shortcut.conv.norm": "encoder.layers.1.shortcut.norm",
+    "encoder.model.3.conv.norm": "encoder.layers.3.norm",
+    "encoder.model.4.block.1.conv.norm": "encoder.layers.4.block.1.norm",
+    "encoder.model.4.block.3.conv.norm": "encoder.layers.4.block.3.norm",
+    "encoder.model.4.shortcut.conv.norm": "encoder.layers.4.shortcut.norm",
+    "encoder.model.6.conv.norm": "encoder.layers.6.norm",
+    "encoder.model.7.block.1.conv.norm": "encoder.layers.7.block.1.norm",
+    "encoder.model.7.block.3.conv.norm": "encoder.layers.7.block.3.norm",
+    "encoder.model.7.shortcut.conv.norm": "encoder.layers.7.shortcut.norm",
+    "encoder.model.9.conv.norm": "encoder.layers.9.norm",
+    "encoder.model.10.block.1.conv.norm": "encoder.layers.10.block.1.norm",
+    "encoder.model.10.block.3.conv.norm": "encoder.layers.10.block.3.norm",
+    "encoder.model.10.shortcut.conv.norm": "encoder.layers.10.shortcut.norm",
+    "encoder.model.12.conv.norm": "encoder.layers.12.norm",
+    "encoder.model.15.conv.norm": "encoder.layers.15.norm",
+}
+MAPPING_DECODER = {
+    "decoder.model.0.conv.conv": "decoder.layers.0.conv",
+    "decoder.model.1.lstm": "decoder.layers.1.lstm",
+    "decoder.model.3.convtr.convtr": "decoder.layers.3.conv",
+    "decoder.model.4.block.1.conv.conv": "decoder.layers.4.block.1.conv",
+    "decoder.model.4.block.3.conv.conv": "decoder.layers.4.block.3.conv",
+    "decoder.model.4.shortcut.conv.conv": "decoder.layers.4.shortcut.conv",
+    "decoder.model.6.convtr.convtr": "decoder.layers.6.conv",
+    "decoder.model.7.block.1.conv.conv": "decoder.layers.7.block.1.conv",
+    "decoder.model.7.block.3.conv.conv": "decoder.layers.7.block.3.conv",
+    "decoder.model.7.shortcut.conv.conv": "decoder.layers.7.shortcut.conv",
+    "decoder.model.9.convtr.convtr": "decoder.layers.9.conv",
+    "decoder.model.10.block.1.conv.conv": "decoder.layers.10.block.1.conv",
+    "decoder.model.10.block.3.conv.conv": "decoder.layers.10.block.3.conv",
+    "decoder.model.10.shortcut.conv.conv": "decoder.layers.10.shortcut.conv",
+    "decoder.model.12.convtr.convtr": "decoder.layers.12.conv",
+    "decoder.model.13.block.1.conv.conv": "decoder.layers.13.block.1.conv",
+    "decoder.model.13.block.3.conv.conv": "decoder.layers.13.block.3.conv",
+    "decoder.model.13.shortcut.conv.conv": "decoder.layers.13.shortcut.conv",
+    "decoder.model.15.conv.conv": "decoder.layers.15.conv",
+}
+MAPPING_DECODER_48K = {
+    "decoder.model.0.conv.norm": "decoder.layers.0.norm",
+    "decoder.model.3.convtr.norm": "decoder.layers.3.norm",
+    "decoder.model.4.block.1.conv.norm": "decoder.layers.4.block.1.norm",
+    "decoder.model.4.block.3.conv.norm": "decoder.layers.4.block.3.norm",
+    "decoder.model.4.shortcut.conv.norm": "decoder.layers.4.shortcut.norm",
+    "decoder.model.6.convtr.norm": "decoder.layers.6.norm",
+    "decoder.model.7.block.1.conv.norm": "decoder.layers.7.block.1.norm",
+    "decoder.model.7.block.3.conv.norm": "decoder.layers.7.block.3.norm",
+    "decoder.model.7.shortcut.conv.norm": "decoder.layers.7.shortcut.norm",
+    "decoder.model.9.convtr.norm": "decoder.layers.9.norm",
+    "decoder.model.10.block.1.conv.norm": "decoder.layers.10.block.1.norm",
+    "decoder.model.10.block.3.conv.norm": "decoder.layers.10.block.3.norm",
+    "decoder.model.10.shortcut.conv.norm": "decoder.layers.10.shortcut.norm",
+    "decoder.model.12.convtr.norm": "decoder.layers.12.norm",
+    "decoder.model.13.block.1.conv.norm": "decoder.layers.13.block.1.norm",
+    "decoder.model.13.block.3.conv.norm": "decoder.layers.13.block.3.norm",
+    "decoder.model.13.shortcut.conv.norm": "decoder.layers.13.shortcut.norm",
+    "decoder.model.15.conv.norm": "decoder.layers.15.norm",
+}
+MAPPING_24K = {
+    **MAPPING_QUANTIZER,
+    **MAPPING_ENCODER,
+    **MAPPING_DECODER,
+}
+MAPPING_48K = {
+    **MAPPING_QUANTIZER,
+    **MAPPING_ENCODER,
+    **MAPPING_ENCODER_48K,
+    **MAPPING_DECODER,
+    **MAPPING_DECODER_48K,
+}
+TOP_LEVEL_KEYS = []
+IGNORE_KEYS = []
+
+
+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"
+            f" {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
+    elif weight_type == "running_mean":
+        hf_pointer.running_mean.data = value
+    elif weight_type == "running_var":
+        hf_pointer.running_var.data = value
+    elif weight_type == "num_batches_tracked":
+        hf_pointer.num_batches_tracked.data = value
+    elif weight_type == "weight_ih_l0":
+        hf_pointer.weight_ih_l0.data = value
+    elif weight_type == "weight_hh_l0":
+        hf_pointer.weight_hh_l0.data = value
+    elif weight_type == "bias_ih_l0":
+        hf_pointer.bias_ih_l0.data = value
+    elif weight_type == "bias_hh_l0":
+        hf_pointer.bias_hh_l0.data = value
+    elif weight_type == "weight_ih_l1":
+        hf_pointer.weight_ih_l1.data = value
+    elif weight_type == "weight_hh_l1":
+        hf_pointer.weight_hh_l1.data = value
+    elif weight_type == "bias_ih_l1":
+        hf_pointer.bias_ih_l1.data = value
+    elif weight_type == "bias_hh_l1":
+        hf_pointer.bias_hh_l1.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 should_ignore(name, ignore_keys):
+    for key in ignore_keys:
+        if key.endswith(".*"):
+            if name.startswith(key[:-1]):
+                return True
+        elif ".*." in key:
+            prefix, suffix = key.split(".*.")
+            if prefix in name and suffix in name:
+                return True
+        elif key in name:
+            return True
+    return False
+
+
+def recursively_load_weights(orig_dict, hf_model, model_name):
+    unused_weights = []
+
+    if model_name == "encodec_24khz":
+        MAPPING = MAPPING_24K
+    elif model_name == "encodec_48khz":
+        MAPPING = MAPPING_48K
+    else:
+        raise ValueError(f"Unsupported model: {model_name}")
+
+    for name, value in orig_dict.items():
+        if should_ignore(name, IGNORE_KEYS):
+            logger.info(f"{name} was ignored")
+            continue
+
+        is_used = False
+        for key, mapped_key in MAPPING.items():
+            if "*" in key:
+                prefix, suffix = key.split(".*.")
+                if prefix in name and suffix in name:
+                    key = suffix
+
+            if key in name:
+                # HACK otherwise .embed gets initialized with .embed_avg too
+                if key.endswith("embed") and name.endswith("embed_avg"):
+                    continue
+
+                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 "weight_ih_l0" in name:
+                    weight_type = "weight_ih_l0"
+                elif "weight_hh_l0" in name:
+                    weight_type = "weight_hh_l0"
+                elif "bias_ih_l0" in name:
+                    weight_type = "bias_ih_l0"
+                elif "bias_hh_l0" in name:
+                    weight_type = "bias_hh_l0"
+                elif "weight_ih_l1" in name:
+                    weight_type = "weight_ih_l1"
+                elif "weight_hh_l1" in name:
+                    weight_type = "weight_hh_l1"
+                elif "bias_ih_l1" in name:
+                    weight_type = "bias_ih_l1"
+                elif "bias_hh_l1" in name:
+                    weight_type = "bias_hh_l1"
+                elif "bias" in name:
+                    weight_type = "bias"
+                elif "weight" in name:
+                    weight_type = "weight"
+                elif "running_mean" in name:
+                    weight_type = "running_mean"
+                elif "running_var" in name:
+                    weight_type = "running_var"
+                elif "num_batches_tracked" in name:
+                    weight_type = "num_batches_tracked"
+                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}")
+
+
+@torch.no_grad()
+def convert_checkpoint(
+    model_name,
+    checkpoint_path,
+    pytorch_dump_folder_path,
+    config_path=None,
+    repo_id=None,
+):
+    """
+    Copy/paste/tweak model's weights to transformers design.
+    """
+    if config_path is not None:
+        config = EncodecConfig.from_pretrained(config_path)
+    else:
+        config = EncodecConfig()
+
+    if model_name == "encodec_24khz":
+        pass  # config is already correct
+    elif model_name == "encodec_48khz":
+        config.upsampling_ratios = [8, 5, 4, 2]
+        config.target_bandwidths = [3.0, 6.0, 12.0, 24.0]
+        config.sampling_rate = 48_000
+        config.audio_channels = 2
+        config.use_causal_conv = False
+        config.norm_type = "time_group_norm"
+        config.normalize = True
+        config.chunk_length_s = 1.0
+        config.overlap = 0.01
+    else:
+        raise ValueError(f"Unknown model name: {model_name}")
+
+    model = EncodecModel(config)
+
+    feature_extractor = EncodecFeatureExtractor(
+        feature_size=config.audio_channels,
+        sampling_rate=config.sampling_rate,
+        chunk_length_s=config.chunk_length_s,
+        overlap=config.overlap,
+    )
+    feature_extractor.save_pretrained(pytorch_dump_folder_path)
+
+    original_checkpoint = torch.load(checkpoint_path)
+    recursively_load_weights(original_checkpoint, model, model_name)
+    model.save_pretrained(pytorch_dump_folder_path)
+
+    if repo_id:
+        print("Pushing to the hub...")
+        feature_extractor.push_to_hub(repo_id)
+        model.push_to_hub(repo_id)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--model",
+        default="encodec_24khz",
+        type=str,
+        help="The model to convert. Should be one of 'encodec_24khz', 'encodec_48khz'.",
+    )
+    parser.add_argument("--checkpoint_path", required=True, default=None, type=str, help="Path to original checkpoint")
+    parser.add_argument("--config_path", default=None, type=str, help="Path to hf config.json of model to convert")
+    parser.add_argument(
+        "--pytorch_dump_folder_path", required=True, default=None, type=str, help="Path to the output PyTorch model."
+    )
+    parser.add_argument(
+        "--push_to_hub", default=None, type=str, help="Where to upload the converted model on the 🤗 hub."
+    )
+
+    args = parser.parse_args()
+    convert_checkpoint(
+        args.model,
+        args.checkpoint_path,
+        args.pytorch_dump_folder_path,
+        args.config_path,
+        args.push_to_hub,
+    )

src/transformers/models/encodec/feature_extraction_encodec.py

@@ -0,0 +1,206 @@
+# 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.
+"""Feature extractor class for EnCodec."""
+
+from typing import List, Optional, Union
+
+import numpy as np
+
+from ...feature_extraction_sequence_utils import SequenceFeatureExtractor
+from ...feature_extraction_utils import BatchFeature
+from ...utils import PaddingStrategy, TensorType, logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class EncodecFeatureExtractor(SequenceFeatureExtractor):
+    r"""
+    Constructs an EnCodec feature extractor.
+
+    This feature extractor inherits from [`~feature_extraction_sequence_utils.SequenceFeatureExtractor`] which contains
+    most of the main methods. Users should refer to this superclass for more information regarding those methods.
+
+    Instantiating a feature extractor with the defaults will yield a similar configuration to that of the
+    [facebook/encodec_24khz](https://huggingface.co/facebook/encodec_24khz) architecture.
+
+    Args:
+        feature_size (`int`, *optional*, defaults to 1):
+            The feature dimension of the extracted features. Use 1 for mono, 2 for stereo.
+        sampling_rate (`int`, *optional*, defaults to 24000):
+            The sampling rate at which the audio waveform should be digitalized expressed in hertz (Hz).
+        padding_value (`float`, *optional*, defaults to 0.0):
+            The value that is used to fill the padding values.
+        chunk_length_s (`float`, *optional*):
+            If defined the audio is pre-processed into chunks of lengths `chunk_length_s` and then encoded.
+        overlap (`float`, *optional*):
+            Defines the overlap between each chunk. It is used to compute the `chunk_stride` using the following
+            formulae : `int((1.0 - self.overlap) * self.chunk_length)`.
+    """
+
+    model_input_names = ["input_values", "padding_mask"]
+
+    def __init__(
+        self,
+        feature_size: int = 1,
+        sampling_rate: int = 24000,
+        padding_value: float = 0.0,
+        chunk_length_s: float = None,
+        overlap: float = None,
+        **kwargs,
+    ):
+        super().__init__(feature_size=feature_size, sampling_rate=sampling_rate, padding_value=padding_value, **kwargs)
+        self.chunk_length_s = chunk_length_s
+        self.overlap = overlap
+
+    # This is a property because you might want to change the chunk_length_s on the fly
+    @property
+    def chunk_length(self) -> Optional[int]:
+        if self.chunk_length_s is None:
+            return None
+        else:
+            return int(self.chunk_length_s * self.sampling_rate)
+
+    # This is a property because you might want to change the chunk_length_s on the fly
+    @property
+    def chunk_stride(self) -> Optional[int]:
+        if self.chunk_length_s is None or self.overlap is None:
+            return None
+        else:
+            return max(1, int((1.0 - self.overlap) * self.chunk_length))
+
+    def __call__(
+        self,
+        raw_audio: Union[np.ndarray, List[float], List[np.ndarray], List[List[float]]],
+        padding: Optional[Union[bool, str, PaddingStrategy]] = None,
+        truncation: Optional[bool] = False,
+        max_length: Optional[int] = None,
+        return_tensors: Optional[Union[str, TensorType]] = None,
+        sampling_rate: Optional[int] = None,
+    ) -> BatchFeature:
+        """
+        Main method to featurize and prepare for the model one or several sequence(s).
+
+        Args:
+            raw_audio (`np.ndarray`, `List[float]`, `List[np.ndarray]`, `List[List[float]]`):
+                The sequence or batch of sequences to be processed. Each sequence can be a numpy array, a list of float
+                values, a list of numpy arrays or a list of list of float values. The numpy array must be of shape
+                `(num_samples,)` for mono audio (`feature_size = 1`), or `(2, num_samples)` for stereo audio
+                (`feature_size = 2`).
+            padding (`bool`, `str` or [`~utils.PaddingStrategy`], *optional*, defaults to `True`):
+                Select a strategy to pad the returned sequences (according to the model's padding side and padding
+                index) among:
+
+                - `True` or `'longest'`: Pad to the longest sequence in the batch (or no padding if only a single
+                  sequence if provided).
+                - `'max_length'`: Pad to a maximum length specified with the argument `max_length` or to the maximum
+                  acceptable input length for the model if that argument is not provided.
+                - `False` or `'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of different
+                  lengths).
+            truncation (`bool`, *optional*, defaults to `False`):
+                Activates truncation to cut input sequences longer than `max_length` to `max_length`.
+            max_length (`int`, *optional*):
+                Maximum length of the returned list and optionally padding length (see above).
+            return_tensors (`str` or [`~utils.TensorType`], *optional*):
+                If set, will return tensors instead of list of python integers. Acceptable values are:
+
+                - `'tf'`: Return TensorFlow `tf.constant` objects.
+                - `'pt'`: Return PyTorch `torch.Tensor` objects.
+                - `'np'`: Return Numpy `np.ndarray` objects.
+            sampling_rate (`int`, *optional*):
+                The sampling rate at which the `audio` input was sampled. It is strongly recommended to pass
+                `sampling_rate` at the forward call to prevent silent errors.
+        """
+        if sampling_rate is not None:
+            if sampling_rate != self.sampling_rate:
+                raise ValueError(
+                    f"The model corresponding to this feature extractor: {self} was trained using a sampling rate of"
+                    f" {self.sampling_rate}. Please make sure that the provided audio input was sampled with"
+                    f" {self.sampling_rate} and not {sampling_rate}."
+                )
+        else:
+            logger.warning(
+                "It is strongly recommended to pass the `sampling_rate` argument to this function. "
+                "Failing to do so can result in silent errors that might be hard to debug."
+            )
+
+        if padding and truncation:
+            raise ValueError("Both padding and truncation were set. Make sure you only set one.")
+        elif padding is None:
+            # by default let's pad the inputs
+            padding = True
+
+        is_batched = bool(
+            isinstance(raw_audio, (list, tuple)) and (isinstance(raw_audio[0], (np.ndarray, tuple, list)))
+        )
+
+        if is_batched:
+            raw_audio = [np.asarray(audio, dtype=np.float32).T for audio in raw_audio]
+        elif not is_batched and not isinstance(raw_audio, np.ndarray):
+            raw_audio = np.asarray(raw_audio, dtype=np.float32)
+        elif isinstance(raw_audio, np.ndarray) and raw_audio.dtype is np.dtype(np.float64):
+            raw_audio = raw_audio.astype(np.float32)
+
+        # always return batch
+        if not is_batched:
+            raw_audio = [np.asarray(raw_audio).T]
+
+        # verify inputs are valid
+        for idx, example in enumerate(raw_audio):
+            if example.ndim > 2:
+                raise ValueError(f"Expected input shape (channels, length) but got shape {example.shape}")
+            if self.feature_size == 1 and example.ndim != 1:
+                raise ValueError(f"Expected mono audio but example has {example.shape[-1]} channels")
+            if self.feature_size == 2 and example.shape[-1] != 2:
+                raise ValueError(f"Expected stereo audio but example has {example.shape[-1]} channels")
+
+        padded_inputs = None
+        input_values = BatchFeature({"input_values": raw_audio})
+        if self.chunk_stride is not None and self.chunk_length is not None and max_length is None:
+            if truncation:
+                max_length = min(array.shape[0] for array in raw_audio)
+                nb_step = int(np.floor(max_length / self.chunk_stride))
+                max_length = (nb_step - 1) * self.chunk_stride + self.chunk_length
+            elif padding:
+                max_length = max(array.shape[0] for array in raw_audio)
+                nb_step = int(np.ceil(max_length / self.chunk_stride))
+                max_length = (nb_step - 1) * self.chunk_stride + self.chunk_length
+                padding = "max_length"
+            else:
+                padded_inputs = input_values
+
+        # normal padding on batch
+        if padded_inputs is None:
+            padded_inputs = self.pad(
+                input_values,
+                max_length=max_length,
+                truncation=truncation,
+                padding=padding,
+                return_attention_mask=padding,
+            )
+            if padding:
+                padded_inputs["padding_mask"] = padded_inputs.pop("attention_mask")
+
+        input_values = []
+        for example in padded_inputs.pop("input_values"):
+            if self.feature_size == 1:
+                example = example[..., None]
+            input_values.append(example.T)
+
+        padded_inputs["input_values"] = input_values
+        if return_tensors is not None:
+            padded_inputs = padded_inputs.convert_to_tensors(return_tensors)
+
+        return padded_inputs

src/transformers/models/encodec/modeling_encodec.py

@@ -0,0 +1,808 @@
+# coding=utf-8
+# Copyright 2023 Meta Platforms, Inc. and affiliates, and 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.
+""" PyTorch EnCodec model."""
+
+import math
+from dataclasses import dataclass
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+
+from ...modeling_utils import PreTrainedModel
+from ...utils import (
+    ModelOutput,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    logging,
+    replace_return_docstrings,
+)
+from .configuration_encodec import EncodecConfig
+
+
+logger = logging.get_logger(__name__)
+
+
+# General docstring
+_CONFIG_FOR_DOC = "EncodecConfig"
+
+
+ENCODEC_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "facebook/encodec_24khz",
+    "facebook/encodec_48khz",
+    # See all EnCodec models at https://huggingface.co/models?filter=encodec
+]
+
+
+@dataclass
+class EncodecOutput(ModelOutput):
+    """
+    Args:
+        audio_codes (`torch.FloatTensor`  of shape `(batch_size, nb_chunks, chunk_length)`, *optional*):
+            Discret code embeddings computed using `model.encode`.
+        audio_values (`torch.FlaotTensor` of shape `(batch_size, sequence_length)`, *optional*)
+            Decoded audio values, obtained using the decoder part of Encodec.
+    """
+
+    audio_codes: torch.FloatTensor = None
+    audio_values: torch.FloatTensor = None
+
+
+@dataclass
+class EncodecEncoderOutput(ModelOutput):
+    """
+    Args:
+        audio_codes (`torch.FloatTensor`  of shape `(batch_size, nb_chunks, chunk_length)`, *optional*):
+            Discret code embeddings computed using `model.encode`.
+        audio_scales (`torch.Tensor` of shape `(batch_size, nb_chunks)`, *optional*):
+            Scaling factor for each `audio_codes` input. This is used to unscale each chunk of audio when decoding.
+    """
+
+    audio_codes: torch.FloatTensor = None
+    audio_scales: torch.FloatTensor = None
+
+
+@dataclass
+class EncodecDecoderOutput(ModelOutput):
+    """
+    Args:
+        audio_values (`torch.FloatTensor`  of shape `(batch_size, segment_length)`, *optional*):
+            Decoded audio values, obtained using the decoder part of Encodec.
+    """
+
+    audio_values: torch.FloatTensor = None
+
+
+class EncodecConv1d(nn.Module):
+    """Conv1d with asymmetric or causal padding and normalization."""
+
+    def __init__(
+        self, config, in_channels: int, out_channels: int, kernel_size: int, stride: int = 1, dilation: int = 1
+    ):
+        super().__init__()
+        self.causal = config.use_causal_conv
+        self.pad_mode = config.pad_mode
+        self.norm_type = config.norm_type
+
+        if self.norm_type not in ["weight_norm", "time_group_norm"]:
+            raise ValueError(
+                f'self.norm_type must be one of `"weight_norm"`, `"time_group_norm"`), got {self.norm_type}'
+            )
+
+        # warn user on unusual setup between dilation and stride
+        if stride > 1 and dilation > 1:
+            logger.warning(
+                "EncodecConv1d has been initialized with stride > 1 and dilation > 1"
+                f" (kernel_size={kernel_size} stride={stride}, dilation={dilation})."
+            )
+
+        self.conv = nn.Conv1d(in_channels, out_channels, kernel_size, stride, dilation=dilation)
+        if self.norm_type == "weight_norm":
+            self.conv = nn.utils.weight_norm(self.conv)
+        elif self.norm_type == "time_group_norm":
+            self.norm = nn.GroupNorm(1, out_channels)
+
+    @staticmethod
+    def _get_extra_padding_for_conv1d(
+        hidden_states: torch.Tensor, kernel_size: int, stride: int, padding_total: int = 0
+    ) -> int:
+        """See `pad_for_conv1d`."""
+        length = hidden_states.shape[-1]
+        n_frames = (length - kernel_size + padding_total) / stride + 1
+        ideal_length = (math.ceil(n_frames) - 1) * stride + (kernel_size - padding_total)
+        return ideal_length - length
+
+    @staticmethod
+    def _pad1d(hidden_states: torch.Tensor, paddings: Tuple[int, int], mode: str = "zero", value: float = 0.0):
+        """Tiny wrapper around torch.nn.functional.pad, just to allow for reflect padding on small input.
+        If this is the case, we insert extra 0 padding to the right before the reflection happens.
+        """
+        length = hidden_states.shape[-1]
+        padding_left, padding_right = paddings
+        if not mode == "reflect":
+            return nn.functional.pad(hidden_states, paddings, mode, value)
+
+        max_pad = max(padding_left, padding_right)
+        extra_pad = 0
+        if length <= max_pad:
+            extra_pad = max_pad - length + 1
+            hidden_states = nn.functional.pad(hidden_states, (0, extra_pad))
+        padded = nn.functional.pad(hidden_states, paddings, mode, value)
+        end = padded.shape[-1] - extra_pad
+        return padded[..., :end]
+
+    def forward(self, hidden_states):
+        kernel_size = self.conv.kernel_size[0]
+        stride = self.conv.stride[0]
+        dilation = self.conv.dilation[0]
+        kernel_size = (kernel_size - 1) * dilation + 1  # effective kernel size with dilations
+        padding_total = kernel_size - stride
+        extra_padding = self._get_extra_padding_for_conv1d(hidden_states, kernel_size, stride, padding_total)
+
+        if self.causal:
+            # Left padding for causal
+            hidden_states = self._pad1d(hidden_states, (padding_total, extra_padding), mode=self.pad_mode)
+        else:
+            # Asymmetric padding required for odd strides
+            padding_right = padding_total // 2
+            padding_left = padding_total - padding_right
+            hidden_states = self._pad1d(
+                hidden_states, (padding_left, padding_right + extra_padding), mode=self.pad_mode
+            )
+
+        hidden_states = self.conv(hidden_states)
+
+        if self.norm_type == "time_group_norm":
+            hidden_states = self.norm(hidden_states)
+
+        return hidden_states
+
+
+class EncodecConvTranspose1d(nn.Module):
+    """ConvTranspose1d with asymmetric or causal padding and normalization."""
+
+    def __init__(self, config, in_channels: int, out_channels: int, kernel_size: int, stride: int = 1):
+        super().__init__()
+        self.causal = config.use_causal_conv
+        self.trim_right_ratio = config.trim_right_ratio
+        self.norm_type = config.norm_type
+        if self.norm_type not in ["weight_norm", "time_group_norm"]:
+            raise ValueError(
+                f'self.norm_type must be one of `"weight_norm"`, `"time_group_norm"`), got {self.norm_type}'
+            )
+
+        self.conv = nn.ConvTranspose1d(in_channels, out_channels, kernel_size, stride)
+        if config.norm_type == "weight_norm":
+            self.conv = nn.utils.weight_norm(self.conv)
+        elif config.norm_type == "time_group_norm":
+            self.norm = nn.GroupNorm(1, out_channels)
+
+        if not (self.causal or self.trim_right_ratio == 1.0):
+            raise ValueError("`trim_right_ratio` != 1.0 only makes sense for causal convolutions")
+
+    def forward(self, hidden_states):
+        kernel_size = self.conv.kernel_size[0]
+        stride = self.conv.stride[0]
+        padding_total = kernel_size - stride
+
+        hidden_states = self.conv(hidden_states)
+
+        if self.norm_type == "time_group_norm":
+            hidden_states = self.norm(hidden_states)
+
+        # We will only trim fixed padding. Extra padding from `pad_for_conv1d` would be
+        # removed at the very end, when keeping only the right length for the output,
+        # as removing it here would require also passing the length at the matching layer
+        # in the encoder.
+        if self.causal:
+            # Trim the padding on the right according to the specified ratio
+            # if trim_right_ratio = 1.0, trim everything from right
+            padding_right = math.ceil(padding_total * self.trim_right_ratio)
+        else:
+            # Asymmetric padding required for odd strides
+            padding_right = padding_total // 2
+
+        padding_left = padding_total - padding_right
+
+        # unpad
+        end = hidden_states.shape[-1] - padding_right
+        hidden_states = hidden_states[..., padding_left:end]
+        return hidden_states
+
+
+class EncodecLSTM(nn.Module):
+    """
+    LSTM without worrying about the hidden state, nor the layout of the data. Expects input as convolutional layout.
+    """
+
+    def __init__(self, config, dimension):
+        super().__init__()
+        self.lstm = nn.LSTM(dimension, dimension, config.num_lstm_layers)
+
+    def forward(self, hidden_states):
+        hidden_states = hidden_states.permute(2, 0, 1)
+        hidden_states = self.lstm(hidden_states)[0] + hidden_states
+        hidden_states = hidden_states.permute(1, 2, 0)
+        return hidden_states
+
+
+class EncodecResnetBlock(nn.Module):
+    """
+    Residual block from SEANet model as used by EnCodec.
+    """
+
+    def __init__(self, config: EncodecConfig, dim: int, dilations: List[int]):
+        super().__init__()
+        kernel_sizes = (config.residual_kernel_size, 1)
+        if len(kernel_sizes) != len(dilations):
+            raise ValueError("Number of kernel sizes should match number of dilations")
+
+        hidden = dim // config.compress
+        block = []
+        for i, (kernel_size, dilation) in enumerate(zip(kernel_sizes, dilations)):
+            in_chs = dim if i == 0 else hidden
+            out_chs = dim if i == len(kernel_sizes) - 1 else hidden
+            block += [nn.ELU()]
+            block += [EncodecConv1d(config, in_chs, out_chs, kernel_size, dilation=dilation)]
+        self.block = nn.ModuleList(block)
+
+        self.shortcut = EncodecConv1d(config, dim, dim, kernel_size=1)
+
+    def forward(self, hidden_states):
+        residual = hidden_states
+        for layer in self.block:
+            hidden_states = layer(hidden_states)
+
+        return self.shortcut(residual) + hidden_states
+
+
+class EncodecEncoder(nn.Module):
+    """SEANet encoder as used by EnCodec."""
+
+    def __init__(self, config: EncodecConfig):
+        super().__init__()
+        model = [EncodecConv1d(config, config.audio_channels, config.num_filters, config.kernel_size)]
+        scaling = 1
+
+        # Downsample to raw audio scale
+        for ratio in reversed(config.upsampling_ratios):
+            current_scale = scaling * config.num_filters
+            # Add residual layers
+            for j in range(config.num_residual_layers):
+                model += [EncodecResnetBlock(config, current_scale, [config.dilation_growth_rate**j, 1])]
+            # Add downsampling layers
+            model += [nn.ELU()]
+            model += [EncodecConv1d(config, current_scale, current_scale * 2, kernel_size=ratio * 2, stride=ratio)]
+            scaling *= 2
+
+        model += [EncodecLSTM(config, scaling * config.num_filters)]
+        model += [nn.ELU()]
+        model += [EncodecConv1d(config, scaling * config.num_filters, config.hidden_size, config.last_kernel_size)]
+
+        self.layers = nn.ModuleList(model)
+
+    def forward(self, hidden_states):
+        for layer in self.layers:
+            hidden_states = layer(hidden_states)
+        return hidden_states
+
+
+class EncodecDecoder(nn.Module):
+    """SEANet decoder as used by EnCodec."""
+
+    def __init__(self, config: EncodecConfig):
+        super().__init__()
+        scaling = int(2 ** len(config.upsampling_ratios))
+        model = [EncodecConv1d(config, config.hidden_size, scaling * config.num_filters, config.kernel_size)]
+
+        model += [EncodecLSTM(config, scaling * config.num_filters)]
+
+        # Upsample to raw audio scale
+        for ratio in config.upsampling_ratios:
+            current_scale = scaling * config.num_filters
+            # Add upsampling layers
+            model += [nn.ELU()]
+            model += [
+                EncodecConvTranspose1d(config, current_scale, current_scale // 2, kernel_size=ratio * 2, stride=ratio)
+            ]
+            # Add residual layers
+            for j in range(config.num_residual_layers):
+                model += [EncodecResnetBlock(config, current_scale // 2, (config.dilation_growth_rate**j, 1))]
+            scaling //= 2
+
+        # Add final layers
+        model += [nn.ELU()]
+        model += [EncodecConv1d(config, config.num_filters, config.audio_channels, config.last_kernel_size)]
+        self.layers = nn.ModuleList(model)
+
+    def forward(self, hidden_states):
+        for layer in self.layers:
+            hidden_states = layer(hidden_states)
+        return hidden_states
+
+
+class EncodecEuclideanCodebook(nn.Module):
+    """Codebook with Euclidean distance."""
+
+    def __init__(self, config: EncodecConfig):
+        super().__init__()
+        embed = torch.zeros(config.codebook_size, config.codebook_dim)
+
+        self.codebook_size = config.codebook_size
+
+        self.register_buffer("inited", torch.Tensor([True]))
+        self.register_buffer("cluster_size", torch.zeros(config.codebook_size))
+        self.register_buffer("embed", embed)
+        self.register_buffer("embed_avg", embed.clone())
+
+    def quantize(self, hidden_states):
+        embed = self.embed.t()
+        scaled_states = hidden_states.pow(2).sum(1, keepdim=True)
+        dist = -(scaled_states - 2 * hidden_states @ embed + embed.pow(2).sum(0, keepdim=True))
+        embed_ind = dist.max(dim=-1).indices
+        return embed_ind
+
+    def encode(self, hidden_states):
+        shape = hidden_states.shape
+        # pre-process
+        hidden_states = hidden_states.reshape((-1, shape[-1]))
+        # quantize
+        embed_ind = self.quantize(hidden_states)
+        # post-process
+        embed_ind = embed_ind.view(*shape[:-1])
+        return embed_ind
+
+    def decode(self, embed_ind):
+        quantize = nn.functional.embedding(embed_ind, self.embed)
+        return quantize
+
+
+class EncodecVectorQuantization(nn.Module):
+    """
+    Vector quantization implementation. Currently supports only euclidean distance.
+    """
+
+    def __init__(self, config: EncodecConfig):
+        super().__init__()
+        self.codebook = EncodecEuclideanCodebook(config)
+
+    def encode(self, hidden_states):
+        hidden_states = hidden_states.permute(0, 2, 1)
+        embed_in = self.codebook.encode(hidden_states)
+        return embed_in
+
+    def decode(self, embed_ind):
+        quantize = self.codebook.decode(embed_ind)
+        quantize = quantize.permute(0, 2, 1)
+        return quantize
+
+
+class EncodecResidualVectorQuantizer(nn.Module):
+    """Residual Vector Quantizer."""
+
+    def __init__(self, config: EncodecConfig):
+        super().__init__()
+        self.codebook_size = config.codebook_size
+        self.frame_rate = config.frame_rate
+        self.num_quantizers = config.num_quantizers
+        self.layers = nn.ModuleList([EncodecVectorQuantization(config) for _ in range(config.num_quantizers)])
+
+    def get_num_quantizers_for_bandwidth(self, bandwidth: Optional[float] = None) -> int:
+        """Return num_quantizers based on specified target bandwidth."""
+        bw_per_q = math.log2(self.codebook_size) * self.frame_rate
+        num_quantizers = self.num_quantizers
+        if bandwidth is not None and bandwidth > 0.0:
+            num_quantizers = int(max(1, math.floor(bandwidth * 1000 / bw_per_q)))
+        return num_quantizers
+
+    def encode(self, embeddings: torch.Tensor, bandwidth: Optional[float] = None) -> torch.Tensor:
+        """
+        Encode a given input tensor with the specified frame rate at the given bandwidth. The RVQ encode method sets
+        the appropriate number of quantizers to use and returns indices for each quantizer.
+        """
+        num_quantizers = self.get_num_quantizers_for_bandwidth(bandwidth)
+        residual = embeddings
+        all_indices = []
+        for layer in self.layers[:num_quantizers]:
+            indices = layer.encode(residual)
+            quantized = layer.decode(indices)
+            residual = residual - quantized
+            all_indices.append(indices)
+        out_indices = torch.stack(all_indices)
+        return out_indices
+
+    def decode(self, codes: torch.Tensor) -> torch.Tensor:
+        """Decode the given codes to the quantized representation."""
+        quantized_out = torch.tensor(0.0, device=codes.device)
+        for i, indices in enumerate(codes):
+            layer = self.layers[i]
+            quantized = layer.decode(indices)
+            quantized_out = quantized_out + quantized
+        return quantized_out
+
+
+class EncodecPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = EncodecConfig
+    base_model_prefix = "encodec"
+    main_input_name = "input_values"
+    supports_gradient_checkpointing = True
+
+    def _init_weights(self, module):
+        """Initialize the weights"""
+        if isinstance(module, nn.Linear):
+            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, (nn.LayerNorm, nn.GroupNorm)):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+        elif isinstance(module, nn.Conv1d):
+            nn.init.kaiming_normal_(module.weight)
+            if module.bias is not None:
+                k = math.sqrt(module.groups / (module.in_channels * module.kernel_size[0]))
+                nn.init.uniform_(module.bias, a=-k, b=k)
+        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_()
+        elif isinstance(module, nn.LSTM):
+            for name, param in module.named_parameters():
+                if "weight" in name:
+                    nn.init.xavier_uniform_(param)
+                elif "bias" in name:
+                    nn.init.constant_(param, 0.0)
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, (EncodecEncoder, EncodecDecoder)):
+            module.gradient_checkpointing = value
+
+
+ENCODEC_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 ([`EncodecConfig`]):
+            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.
+"""
+
+
+ENCODEC_INPUTS_DOCSTRING = r"""
+    Args:
+        input_values (`torch.FloatTensor` of shape `(batch_size, channels, sequence_length)`, *optional*):
+            Raw audio input converted to Float and padded to the approriate length in order to be encoded using chunks
+            of length self.chunk_length and a stride of `config.chunk_stride`.
+        padding_mask (`torch.BoolTensor` of shape `(batch_size, channels, sequence_length)`, *optional*):
+            Mask to avoid computing scaling factors on padding token indices (can we avoid computing conv on these+).
+            Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            <Tip warning={true}>
+
+             `padding_mask` should always be passed, unless the input was truncated or not padded. This is because in
+             order to process tensors effectively, the input audio should be padded so that `input_length % stride =
+             step` with `step = chunk_length-stride`. This ensures that all chunks are of the same shape
+
+            </Tip>
+
+        bandwidth (`float`, *optional*):
+            The target bandwidth. Must be one of `config.target_bandwidths`. If `None`, uses the smallest possible
+            bandwidth. bandwidth is represented as a thousandth of what it is, e.g. 6kbps bandwidth is represented as
+            `bandwidth == 6.0`
+        audio_codes (`torch.FloatTensor`  of shape `(batch_size, nb_chunks, chunk_length)`, *optional*):
+            Discret code embeddings computed using `model.encode`.
+        audio_scales (`torch.Tensor` of shape `(batch_size, nb_chunks)`, *optional*):
+            Scaling factor for each `audio_codes` input.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+@add_start_docstrings(
+    "The EnCodec neural audio codec model.",
+    ENCODEC_START_DOCSTRING,
+)
+class EncodecModel(EncodecPreTrainedModel):
+    def __init__(self, config: EncodecConfig):
+        super().__init__(config)
+        self.config = config
+
+        self.encoder = EncodecEncoder(config)
+        self.decoder = EncodecDecoder(config)
+
+        self.quantizer = EncodecResidualVectorQuantizer(config)
+
+        self.bits_per_codebook = int(math.log2(self.config.codebook_size))
+        if 2**self.bits_per_codebook != self.config.codebook_size:
+            raise ValueError("The codebook_size must be a power of 2.")
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_encoder(self):
+        return self.encoder
+
+    def get_decoder(self):
+        return self.decoder
+
+    def _encode_frame(
+        self, input_values: torch.Tensor, bandwidth: float, padding_mask: int
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
+        """
+        Encodes the given input using the underlying VQVAE. If `config.normalize` is set to `True` the input is first
+        normalized. The padding mask is required to compute the correct scale.
+        """
+        length = input_values.shape[-1]
+        duration = length / self.config.sampling_rate
+
+        if self.config.chunk_length_s is not None and duration > 1e-5 + self.config.chunk_length_s:
+            raise RuntimeError(f"Duration of frame ({duration}) is longer than chunk {self.config.chunk_length_s}")
+
+        scale = None
+        if self.config.normalize:
+            # if the padding is non zero
+            input_values = input_values * padding_mask
+            mono = torch.sum(input_values, 1, keepdim=True) / input_values.shape[1]
+            scale = mono.pow(2).mean(dim=-1, keepdim=True).sqrt() + 1e-8
+            input_values = input_values / scale
+
+        embeddings = self.encoder(input_values)
+        codes = self.quantizer.encode(embeddings, bandwidth)
+        codes = codes.transpose(0, 1)
+        return codes, scale
+
+    def encode(
+        self,
+        input_values: torch.Tensor,
+        padding_mask: torch.Tensor = None,
+        bandwidth: Optional[float] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple[torch.Tensor, Optional[torch.Tensor]], EncodecEncoderOutput]:
+        """
+        Encodes the input audio waveform into discrete codes.
+
+        Args:
+            input_values (`torch.Tensor` of shape `(batch_size, channels, sequence_length)`):
+                Float values of the input audio waveform.
+            padding_mask (`torch.Tensor` of shape `(batch_size, channels, sequence_length)`):
+                Padding mask used to pad the `input_values`.
+            bandwidth (`float`, *optional*):
+                The target bandwidth. Must be one of `config.target_bandwidths`. If `None`, uses the smallest possible
+                bandwidth. bandwidth is represented as a thousandth of what it is, e.g. 6kbps bandwidth is represented
+                as bandwidth == 6.0
+
+        Returns:
+            A list of frames containing the discrete encoded codes for the input audio waveform, along with rescaling
+            factors for each chunk when `normalize` is True. Each frames is a tuple `(codebook, scale)`, with
+            `codebook` of shape `[batch_size, num_codebooks, frames]`.
+        """
+        return_dict = return_dict if return_dict is not None else self.config.return_dict
+
+        if bandwidth is None:
+            bandwidth = self.config.target_bandwidths[0]
+        if bandwidth not in self.config.target_bandwidths:
+            raise ValueError(
+                f"This model doesn't support the bandwidth {bandwidth}. "
+                f"Select one of {self.config.target_bandwidths}."
+            )
+
+        _, channels, input_length = input_values.shape
+
+        if channels < 1 or channels > 2:
+            raise ValueError(f"Number of audio channels must be 1 or 2, but got {channels}")
+
+        chunk_length = self.config.chunk_length
+        if chunk_length is None:
+            chunk_length = input_length
+            stride = input_length
+        else:
+            stride = self.config.chunk_stride
+
+        if padding_mask is None:
+            padding_mask = torch.ones_like(input_values).bool()
+
+        encoded_frames = []
+        scales = []
+
+        step = chunk_length - stride
+        if (input_length % stride) - step != 0:
+            raise ValueError(
+                "The input length is not properly padded for batched chunked decoding. Make sure to pad the input correctly."
+            )
+
+        for offset in range(0, input_length - step, stride):
+            mask = padding_mask[..., offset : offset + chunk_length].bool()
+            frame = input_values[:, :, offset : offset + chunk_length]
+            encoded_frame, scale = self._encode_frame(frame, bandwidth, mask)
+            encoded_frames.append(encoded_frame)
+            scales.append(scale)
+
+        encoded_frames = torch.stack(encoded_frames)
+
+        if not return_dict:
+            return (encoded_frames, scales)
+
+        return EncodecEncoderOutput(encoded_frames, scales)
+
+    @staticmethod
+    def _linear_overlap_add(frames: List[torch.Tensor], stride: int):
+        # Generic overlap add, with linear fade-in/fade-out, supporting complex scenario
+        # e.g., more than 2 frames per position.
+        # The core idea is to use a weight function that is a triangle,
+        # with a maximum value at the middle of the chunk.
+        # We use this weighting when summing the frames, and divide by the sum of weights
+        # for each positions at the end. Thus:
+        #   - if a frame is the only one to cover a position, the weighting is a no-op.
+        #   - if 2 frames cover a position:
+        #          ...  ...
+        #         /   \/   \
+        #        /    /\    \
+        #            S  T       , i.e. S offset of second frame starts, T end of first frame.
+        # Then the weight function for each one is: (t - S), (T - t), with `t` a given offset.
+        # After the final normalization, the weight of the second frame at position `t` is
+        # (t - S) / (t - S + (T - t)) = (t - S) / (T - S), which is exactly what we want.
+        #
+        #   - if more than 2 frames overlap at a given point, we hope that by induction
+        #      something sensible happens.
+        if len(frames) == 0:
+            raise ValueError("`frames` cannot be an empty list.")
+
+        device = frames[0].device
+        dtype = frames[0].dtype
+        shape = frames[0].shape[:-1]
+        total_size = stride * (len(frames) - 1) + frames[-1].shape[-1]
+
+        frame_length = frames[0].shape[-1]
+        time_vec = torch.linspace(0, 1, frame_length + 2, device=device, dtype=dtype)[1:-1]
+        weight = 0.5 - (time_vec - 0.5).abs()
+
+        sum_weight = torch.zeros(total_size, device=device, dtype=dtype)
+        out = torch.zeros(*shape, total_size, device=device, dtype=dtype)
+        offset: int = 0
+
+        for frame in frames:
+            frame_length = frame.shape[-1]
+            out[..., offset : offset + frame_length] += weight[:frame_length] * frame
+            sum_weight[offset : offset + frame_length] += weight[:frame_length]
+            offset += stride
+
+        if sum_weight.min() == 0:
+            raise ValueError(f"`sum_weight` minimum element must be bigger than zero: {sum_weight}`")
+
+        return out / sum_weight
+
+    def _decode_frame(self, codes: torch.Tensor, scale: Optional[torch.Tensor] = None) -> torch.Tensor:
+        codes = codes.transpose(0, 1)
+        embeddings = self.quantizer.decode(codes)
+        outputs = self.decoder(embeddings)
+        if scale is not None:
+            outputs = outputs * scale.view(-1, 1, 1)
+        return outputs
+
+    def decode(
+        self,
+        audio_codes: torch.Tensor,
+        audio_scales: torch.Tensor,
+        padding_mask: Optional[torch.Tensor] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple[torch.Tensor, torch.Tensor], EncodecDecoderOutput]:
+        """
+        Decodes the given frames into an output audio waveform.
+
+        Note that the output might be a bit bigger than the input. In that case, any extra steps at the end can be
+        trimmed.
+
+        Args:
+            audio_codes (`torch.FloatTensor`  of shape `(batch_size, nb_chunks, chunk_length)`, *optional*):
+                Discret code embeddings computed using `model.encode`.
+            audio_scales (`torch.Tensor` of shape `(batch_size, nb_chunks)`, *optional*):
+                Scaling factor for each `audio_codes` input.
+            padding_mask (`torch.Tensor` of shape `(batch_size, channels, sequence_length)`):
+                Padding mask used to pad the `input_values`.
+            return_dict (`bool`, *optional*):
+                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+
+        """
+        return_dict = return_dict or self.config.return_dict
+
+        chunk_length = self.config.chunk_length
+        if chunk_length is None:
+            if len(audio_codes) != 1:
+                raise ValueError(f"Expected one frame, got {len(audio_codes)}")
+            audio_values = self._decode_frame(audio_codes[0], audio_scales[0])
+        else:
+            decoded_frames = []
+
+            for frame, scale in zip(audio_codes, audio_scales):
+                frames = self._decode_frame(frame, scale)
+                decoded_frames.append(frames)
+
+            audio_values = self._linear_overlap_add(decoded_frames, self.config.chunk_stride or 1)
+
+        # truncate based on padding mask
+        if padding_mask is not None and padding_mask.shape[-1] < audio_values.shape[-1]:
+            audio_values = audio_values[..., : padding_mask.shape[-1]]
+
+        if not return_dict:
+            return (audio_values,)
+        return EncodecDecoderOutput(audio_values)
+
+    @add_start_docstrings_to_model_forward(ENCODEC_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=EncodecOutput, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_values: torch.Tensor,
+        padding_mask: Optional[torch.Tensor] = None,
+        bandwidth: Optional[float] = None,
+        audio_codes: Optional[torch.Tensor] = None,
+        audio_scales: Optional[torch.Tensor] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple[torch.Tensor, torch.Tensor], EncodecOutput]:
+        r"""
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> from datasets import load_dataset
+        >>> from transformers import AutoProcessor, EncodecModel
+
+        >>> dataset = load_dataset("ashraq/esc50")
+        >>> audio_sample = dataset["train"]["audio"][0]["array"]
+
+        >>> model_id = "facebook/encodec_24khz"
+        >>> model = EncodecModel.from_pretrained(model_id)
+        >>> processor = AutoProcessor.from_pretrained(model_id)
+
+        >>> inputs = processor(raw_audio=audio_sample, return_tensors="pt")
+
+        >>> outputs = model(**inputs)
+        >>> audio_codes = outputs.audio_codes
+        >>> audio_values = outputs.audio_values
+        ```"""
+        return_dict = return_dict or self.config.return_dict
+
+        if padding_mask is None:
+            padding_mask = torch.ones_like(input_values).bool()
+
+        if audio_codes is not None and audio_scales is None:
+            raise ValueError("You specified `audio_codes` but did not specify the `audio_scales`")
+
+        if audio_scales is not None and audio_codes is None:
+            raise ValueError("You specified `audio_scales` but did not specify the `audio_codes`")
+
+        if audio_scales is None and audio_codes is None:
+            audio_codes, audio_scales = self.encode(input_values, padding_mask, bandwidth, False)
+
+        audio_values = self.decode(audio_codes, audio_scales, padding_mask, return_dict=return_dict)[0]
+        if not return_dict:
+            return (audio_codes, audio_values)
+
+        return EncodecOutput(audio_codes=audio_codes, audio_values=audio_values)

src/transformers/utils/dummy_pt_objects.py

@@ -2655,6 +2655,23 @@ def load_tf_weights_in_electra(*args, **kwargs):
     requires_backends(load_tf_weights_in_electra, ["torch"])
 
 
+ENCODEC_PRETRAINED_MODEL_ARCHIVE_LIST = None
+
+
+class EncodecModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class EncodecPreTrainedModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
 class EncoderDecoderModel(metaclass=DummyObject):
     _backends = ["torch"]
 

src/transformers/utils/dummy_speech_objects.py

@@ -9,6 +9,13 @@ class ASTFeatureExtractor(metaclass=DummyObject):
         requires_backends(self, ["speech"])
 
 
+class EncodecFeatureExtractor(metaclass=DummyObject):
+    _backends = ["speech"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["speech"])
+
+
 class MCTCTFeatureExtractor(metaclass=DummyObject):
     _backends = ["speech"]
 

tests/models/encodec/test_feature_extraction_encodec.py

@@ -0,0 +1,255 @@
+# coding=utf-8
+# Copyright 2021-2023 HuggingFace Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Tests for the EnCodec feature extractor."""
+
+import itertools
+import random
+import unittest
+
+import numpy as np
+
+from transformers import is_speech_available
+from transformers.testing_utils import require_torch
+from transformers.utils.import_utils import is_torch_available
+
+from ...test_sequence_feature_extraction_common import SequenceFeatureExtractionTestMixin
+
+
+if is_speech_available():
+    from transformers import EncodecFeatureExtractor
+
+if is_torch_available():
+    import torch
+
+
+global_rng = random.Random()
+
+
+def floats_list(shape, scale=1.0, rng=None, name=None):
+    """Creates a random float32 tensor"""
+    if rng is None:
+        rng = global_rng
+
+    values = []
+    for batch_idx in range(shape[0]):
+        values.append([])
+        for _ in range(shape[1]):
+            values[-1].append(rng.random() * scale)
+
+    return values
+
+
+@require_torch
+class EnCodecFeatureExtractionTester(unittest.TestCase):
+    def __init__(
+        self,
+        parent,
+        batch_size=7,
+        min_seq_length=400,
+        max_seq_length=2000,
+        feature_size=1,
+        padding_value=0.0,
+        sampling_rate=24000,
+        return_attention_mask=True,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.min_seq_length = min_seq_length
+        self.max_seq_length = max_seq_length
+        self.seq_length_diff = (self.max_seq_length - self.min_seq_length) // (self.batch_size - 1)
+        self.feature_size = feature_size
+        self.padding_value = padding_value
+        self.sampling_rate = sampling_rate
+        self.return_attention_mask = return_attention_mask
+
+    def prepare_feat_extract_dict(self):
+        return {
+            "feature_size": self.feature_size,
+            "padding_value": self.padding_value,
+            "sampling_rate": self.sampling_rate,
+            "return_attention_mask": self.return_attention_mask,
+        }
+
+    def prepare_inputs_for_common(self, equal_length=False, numpify=False):
+        def _flatten(list_of_lists):
+            return list(itertools.chain(*list_of_lists))
+
+        if equal_length:
+            audio_inputs = floats_list((self.batch_size, self.max_seq_length))
+        else:
+            # make sure that inputs increase in size
+            audio_inputs = [
+                _flatten(floats_list((x, self.feature_size)))
+                for x in range(self.min_seq_length, self.max_seq_length, self.seq_length_diff)
+            ]
+
+        if numpify:
+            audio_inputs = [np.asarray(x) for x in audio_inputs]
+
+        return audio_inputs
+
+
+@require_torch
+class EnCodecFeatureExtractionTest(SequenceFeatureExtractionTestMixin, unittest.TestCase):
+    feature_extraction_class = EncodecFeatureExtractor if is_speech_available() else None
+
+    def setUp(self):
+        self.feat_extract_tester = EnCodecFeatureExtractionTester(self)
+
+    def test_call(self):
+        # Tests that all call wrap to encode_plus and batch_encode_plus
+        feat_extract = self.feature_extraction_class(**self.feat_extract_tester.prepare_feat_extract_dict())
+        # create three inputs of length 800, 1000, and 1200
+        audio_inputs = [floats_list((1, x))[0] for x in range(800, 1400, 200)]
+        np_audio_inputs = [np.asarray(audio_input) for audio_input in audio_inputs]
+
+        # Test not batched input
+        encoded_sequences_1 = feat_extract(audio_inputs[0], return_tensors="np").input_values
+        encoded_sequences_2 = feat_extract(np_audio_inputs[0], return_tensors="np").input_values
+        self.assertTrue(np.allclose(encoded_sequences_1, encoded_sequences_2, atol=1e-3))
+
+        # Test batched
+        encoded_sequences_1 = feat_extract(audio_inputs, padding=True, return_tensors="np").input_values
+        encoded_sequences_2 = feat_extract(np_audio_inputs, padding=True, return_tensors="np").input_values
+        for enc_seq_1, enc_seq_2 in zip(encoded_sequences_1, encoded_sequences_2):
+            self.assertTrue(np.allclose(enc_seq_1, enc_seq_2, atol=1e-3))
+
+    def test_double_precision_pad(self):
+        feature_extractor = self.feature_extraction_class(**self.feat_extract_tester.prepare_feat_extract_dict())
+        np_audio_inputs = np.random.rand(100).astype(np.float64)
+        py_audio_inputs = np_audio_inputs.tolist()
+
+        for inputs in [py_audio_inputs, np_audio_inputs]:
+            np_processed = feature_extractor.pad([{"input_values": inputs}], return_tensors="np")
+            self.assertTrue(np_processed.input_values.dtype == np.float32)
+            pt_processed = feature_extractor.pad([{"input_values": inputs}], return_tensors="pt")
+            self.assertTrue(pt_processed.input_values.dtype == torch.float32)
+
+    def _load_datasamples(self, num_samples):
+        from datasets import load_dataset
+
+        ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
+        # automatic decoding with librispeech
+        audio_samples = ds.sort("id").select(range(num_samples))[:num_samples]["audio"]
+
+        return [x["array"] for x in audio_samples]
+
+    def test_integration(self):
+        # fmt: off
+        EXPECTED_INPUT_VALUES = torch.tensor(
+            [2.3804e-03, 2.0752e-03, 1.9836e-03, 2.1057e-03, 1.6174e-03,
+             3.0518e-04, 9.1553e-05, 3.3569e-04, 9.7656e-04, 1.8311e-03,
+             2.0142e-03, 2.1057e-03, 1.7395e-03, 4.5776e-04, -3.9673e-04,
+             4.5776e-04, 1.0071e-03, 9.1553e-05, 4.8828e-04, 1.1597e-03,
+             7.3242e-04, 9.4604e-04, 1.8005e-03, 1.8311e-03, 8.8501e-04,
+             4.2725e-04, 4.8828e-04, 7.3242e-04, 1.0986e-03, 2.1057e-03]
+        )
+        # fmt: on
+        input_audio = self._load_datasamples(1)
+        feature_extractor = EncodecFeatureExtractor()
+        input_values = feature_extractor(input_audio, return_tensors="pt").input_values
+        self.assertEquals(input_values.shape, (1, 1, 93680))
+        self.assertTrue(torch.allclose(input_values[0, 0, :30], EXPECTED_INPUT_VALUES, atol=1e-6))
+
+    def test_integration_stereo(self):
+        # fmt: off
+        EXPECTED_INPUT_VALUES = torch.tensor(
+            [2.3804e-03, 2.0752e-03, 1.9836e-03, 2.1057e-03, 1.6174e-03,
+             3.0518e-04, 9.1553e-05, 3.3569e-04, 9.7656e-04, 1.8311e-03,
+             2.0142e-03, 2.1057e-03, 1.7395e-03, 4.5776e-04, -3.9673e-04,
+             4.5776e-04, 1.0071e-03, 9.1553e-05, 4.8828e-04, 1.1597e-03,
+             7.3242e-04, 9.4604e-04, 1.8005e-03, 1.8311e-03, 8.8501e-04,
+             4.2725e-04, 4.8828e-04, 7.3242e-04, 1.0986e-03, 2.1057e-03]
+        )
+        # fmt: on
+        input_audio = self._load_datasamples(1)
+        input_audio = [np.tile(input_audio[0][None], reps=(2, 1))]
+        input_audio[0][1] *= 0.5
+        feature_extractor = EncodecFeatureExtractor(feature_size=2)
+        input_values = feature_extractor(input_audio, return_tensors="pt").input_values
+        self.assertEquals(input_values.shape, (1, 2, 93680))
+        self.assertTrue(torch.allclose(input_values[0, 0, :30], EXPECTED_INPUT_VALUES, atol=1e-6))
+        self.assertTrue(torch.allclose(input_values[0, 1, :30], EXPECTED_INPUT_VALUES * 0.5, atol=1e-6))
+
+    def test_truncation_and_padding(self):
+        input_audio = self._load_datasamples(2)
+        # would be easier if the stride was like
+        feature_extractor = EncodecFeatureExtractor(feature_size=1, chunk_length_s=1, overlap=0.01)
+
+        # pad and trunc raise an error ?
+        with self.assertRaisesRegex(
+            ValueError,
+            "^Both padding and truncation were set. Make sure you only set one.$",
+        ):
+            truncated_outputs = feature_extractor(
+                input_audio, padding="max_length", truncation=True, return_tensors="pt"
+            ).input_values
+
+        # truncate to chunk
+        truncated_outputs = feature_extractor(input_audio, truncation=True, return_tensors="pt").input_values
+        self.assertEquals(truncated_outputs.shape, (2, 1, 71520))  # 2 chunks
+
+        # force truncate to max_length
+        truncated_outputs = feature_extractor(
+            input_audio, truncation=True, max_length=48000, return_tensors="pt"
+        ).input_values
+        self.assertEquals(truncated_outputs.shape, (2, 1, 48000))
+
+        # pad to chunk
+        padded_outputs = feature_extractor(input_audio, padding=True, return_tensors="pt").input_values
+        self.assertEquals(padded_outputs.shape, (2, 1, 95280))
+
+        # pad to chunk
+        truncated_outputs = feature_extractor(input_audio, return_tensors="pt").input_values
+        self.assertEquals(truncated_outputs.shape, (2, 1, 95280))
+
+        # force pad to max length
+        truncated_outputs = feature_extractor(
+            input_audio, padding="max_length", max_length=100000, return_tensors="pt"
+        ).input_values
+        self.assertEquals(truncated_outputs.shape, (2, 1, 100000))
+
+        # force no pad
+        with self.assertRaisesRegex(
+            ValueError,
+            "^Unable to create tensor, you should probably activate padding with 'padding=True' to have batched tensors with the same length.$",
+        ):
+            truncated_outputs = feature_extractor(input_audio, padding=False, return_tensors="pt").input_values
+
+        truncated_outputs = feature_extractor(input_audio[0], padding=False, return_tensors="pt").input_values
+        self.assertEquals(truncated_outputs.shape, (1, 1, 93680))
+
+        # no pad if no chunk_length_s
+        feature_extractor.chunk_length_s = None
+        with self.assertRaisesRegex(
+            ValueError,
+            "^Unable to create tensor, you should probably activate padding with 'padding=True' to have batched tensors with the same length.$",
+        ):
+            truncated_outputs = feature_extractor(input_audio, padding=False, return_tensors="pt").input_values
+
+        truncated_outputs = feature_extractor(input_audio[0], padding=False, return_tensors="pt").input_values
+        self.assertEquals(truncated_outputs.shape, (1, 1, 93680))
+
+        # no pad if no overlap
+        feature_extractor.chunk_length_s = 2
+        feature_extractor.overlap = None
+        with self.assertRaisesRegex(
+            ValueError,
+            "^Unable to create tensor, you should probably activate padding with 'padding=True' to have batched tensors with the same length.$",
+        ):
+            truncated_outputs = feature_extractor(input_audio, padding=False, return_tensors="pt").input_values
+
+        truncated_outputs = feature_extractor(input_audio[0], padding=False, return_tensors="pt").input_values
+        self.assertEquals(truncated_outputs.shape, (1, 1, 93680))

tests/models/encodec/test_modeling_encodec.py

@@ -0,0 +1,571 @@
+# 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.
+""" Testing suite for the PyTorch Encodec model. """
+
+import copy
+import inspect
+import os
+import tempfile
+import unittest
+from typing import Dict, List, Tuple
+
+import numpy as np
+from datasets import Audio, load_dataset
+
+from transformers import AutoProcessor, EncodecConfig
+from transformers.testing_utils import (
+    is_torch_available,
+    require_torch,
+    slow,
+    torch_device,
+)
+
+from ...test_configuration_common import ConfigTester
+from ...test_modeling_common import (
+    ModelTesterMixin,
+    _config_zero_init,
+    floats_tensor,
+)
+from ...test_pipeline_mixin import PipelineTesterMixin
+
+
+if is_torch_available():
+    import torch
+
+    from transformers import EncodecModel
+
+
+def prepare_inputs_dict(
+    config,
+    input_ids=None,
+    input_values=None,
+    decoder_input_ids=None,
+    attention_mask=None,
+    decoder_attention_mask=None,
+    head_mask=None,
+    decoder_head_mask=None,
+    cross_attn_head_mask=None,
+):
+    if input_ids is not None:
+        encoder_dict = {"input_ids": input_ids}
+    else:
+        encoder_dict = {"input_values": input_values}
+
+    decoder_dict = {"decoder_input_ids": decoder_input_ids} if decoder_input_ids is not None else {}
+
+    return {**encoder_dict, **decoder_dict}
+
+
+@require_torch
+class EncodecModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=13,
+        num_channels=2,
+        is_training=False,
+        num_hidden_layers=4,
+        intermediate_size=40,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.num_channels = num_channels
+        self.is_training = is_training
+
+        self.num_hidden_layers = num_hidden_layers
+        self.intermediate_size = intermediate_size
+
+    def prepare_config_and_inputs(self):
+        input_values = floats_tensor([self.batch_size, self.num_channels, self.intermediate_size], scale=1.0)
+        config = self.get_config()
+        inputs_dict = {"input_values": input_values}
+        return config, inputs_dict
+
+    def prepare_config_and_inputs_for_common(self):
+        config, inputs_dict = self.prepare_config_and_inputs()
+        return config, inputs_dict
+
+    def get_config(self):
+        return EncodecConfig(audio_channels=self.num_channels, chunk_in_sec=None)
+
+    def create_and_check_model_forward(self, config, inputs_dict):
+        model = EncodecModel(config=config).to(torch_device).eval()
+
+        input_values = inputs_dict["input_values"]
+        result = model(input_values)
+        self.parent.assertEqual(
+            result.audio_values.shape, (self.batch_size, self.num_channels, self.intermediate_size)
+        )
+
+
+@require_torch
+class EncodecModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
+    all_model_classes = (EncodecModel,) if is_torch_available() else ()
+    is_encoder_decoder = True
+    test_pruning = False
+    test_headmasking = False
+    test_resize_embeddings = False
+    pipeline_model_mapping = {}
+    input_name = "input_values"
+
+    def _prepare_for_class(self, inputs_dict, model_class, return_labels=False):
+        # model does not have attention and does not support returning hidden states
+        inputs_dict = super()._prepare_for_class(inputs_dict, model_class, return_labels=return_labels)
+        if "output_attentions" in inputs_dict:
+            inputs_dict.pop("output_attentions")
+        if "output_hidden_states" in inputs_dict:
+            inputs_dict.pop("output_hidden_states")
+        return inputs_dict
+
+    def setUp(self):
+        self.model_tester = EncodecModelTester(self)
+        self.config_tester = ConfigTester(
+            self, config_class=EncodecConfig, hidden_size=37, common_properties=[], has_text_modality=False
+        )
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_model_forward(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_model_forward(*config_and_inputs)
+
+    def test_forward_signature(self):
+        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            model = model_class(config)
+            signature = inspect.signature(model.forward)
+            # signature.parameters is an OrderedDict => so arg_names order is deterministic
+            arg_names = [*signature.parameters.keys()]
+
+            expected_arg_names = ["input_values", "padding_mask", "bandwidth"]
+            self.assertListEqual(arg_names[: len(expected_arg_names)], expected_arg_names)
+
+    @unittest.skip("The EncodecModel is not transformers based, thus it does not have `inputs_embeds` logics")
+    def test_inputs_embeds(self):
+        pass
+
+    @unittest.skip("The EncodecModel is not transformers based, thus it does not have `inputs_embeds` logics")
+    def test_model_common_attributes(self):
+        pass
+
+    @unittest.skip("The EncodecModel is not transformers based, thus it does not have the usual `attention` logic")
+    def test_retain_grad_hidden_states_attentions(self):
+        pass
+
+    @unittest.skip("The EncodecModel is not transformers based, thus it does not have the usual `attention` logic")
+    def test_torchscript_output_attentions(self):
+        pass
+
+    @unittest.skip("The EncodecModel is not transformers based, thus it does not have the usual `hidden_states` logic")
+    def test_torchscript_output_hidden_state(self):
+        pass
+
+    def _create_and_check_torchscript(self, config, inputs_dict):
+        if not self.test_torchscript:
+            return
+
+        configs_no_init = _config_zero_init(config)  # To be sure we have no Nan
+        configs_no_init.torchscript = True
+        configs_no_init.return_dict = False
+        for model_class in self.all_model_classes:
+            model = model_class(config=configs_no_init)
+            model.to(torch_device)
+            model.eval()
+            inputs = self._prepare_for_class(inputs_dict, model_class)
+
+            main_input_name = model_class.main_input_name
+
+            try:
+                main_input = inputs[main_input_name]
+                model(main_input)
+                traced_model = torch.jit.trace(model, main_input)
+            except RuntimeError:
+                self.fail("Couldn't trace module.")
+
+            with tempfile.TemporaryDirectory() as tmp_dir_name:
+                pt_file_name = os.path.join(tmp_dir_name, "traced_model.pt")
+
+                try:
+                    torch.jit.save(traced_model, pt_file_name)
+                except Exception:
+                    self.fail("Couldn't save module.")
+
+                try:
+                    loaded_model = torch.jit.load(pt_file_name)
+                except Exception:
+                    self.fail("Couldn't load module.")
+
+            model.to(torch_device)
+            model.eval()
+
+            loaded_model.to(torch_device)
+            loaded_model.eval()
+
+            model_state_dict = model.state_dict()
+            loaded_model_state_dict = loaded_model.state_dict()
+
+            non_persistent_buffers = {}
+            for key in loaded_model_state_dict.keys():
+                if key not in model_state_dict.keys():
+                    non_persistent_buffers[key] = loaded_model_state_dict[key]
+
+            loaded_model_state_dict = {
+                key: value for key, value in loaded_model_state_dict.items() if key not in non_persistent_buffers
+            }
+
+            self.assertEqual(set(model_state_dict.keys()), set(loaded_model_state_dict.keys()))
+
+            model_buffers = list(model.buffers())
+            for non_persistent_buffer in non_persistent_buffers.values():
+                found_buffer = False
+                for i, model_buffer in enumerate(model_buffers):
+                    if torch.equal(non_persistent_buffer, model_buffer):
+                        found_buffer = True
+                        break
+
+                self.assertTrue(found_buffer)
+                model_buffers.pop(i)
+
+            models_equal = True
+            for layer_name, p1 in model_state_dict.items():
+                if layer_name in loaded_model_state_dict:
+                    p2 = loaded_model_state_dict[layer_name]
+                    if p1.data.ne(p2.data).sum() > 0:
+                        models_equal = False
+
+            self.assertTrue(models_equal)
+
+            # Avoid memory leak. Without this, each call increase RAM usage by ~20MB.
+            # (Even with this call, there are still memory leak by ~0.04MB)
+            self.clear_torch_jit_class_registry()
+
+    @unittest.skip("The EncodecModel is not transformers based, thus it does not have the usual `attention` logic")
+    def test_attention_outputs(self):
+        pass
+
+    def test_feed_forward_chunking(self):
+        (original_config, inputs_dict) = self.model_tester.prepare_config_and_inputs_for_common()
+        for model_class in self.all_model_classes:
+            torch.manual_seed(0)
+            config = copy.deepcopy(original_config)
+            config.chunk_length_s = None
+            config.overlap = None
+            config.sampling_rate = 10
+
+            model = model_class(config)
+            model.to(torch_device)
+            model.eval()
+            inputs = self._prepare_for_class(inputs_dict, model_class)
+            inputs["input_values"] = inputs["input_values"].repeat(1, 1, 10)
+
+            hidden_states_no_chunk = model(**inputs)[0]
+
+            torch.manual_seed(0)
+            config.chunk_length_s = 1
+            config.overlap = 0
+            config.sampling_rate = 10
+
+            model = model_class(config)
+            model.to(torch_device)
+            model.eval()
+
+            hidden_states_with_chunk = model(**inputs)[0]
+            self.assertTrue(torch.allclose(hidden_states_no_chunk, hidden_states_with_chunk, atol=1e-3))
+
+    @unittest.skip("The EncodecModel is not transformers based, thus it does not have the usual `hidden_states` logic")
+    def test_hidden_states_output(self):
+        pass
+
+    def test_determinism(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        def check_determinism(first, second):
+            # outputs are not tensors but list (since each sequence don't have the same frame_length)
+            out_1 = first.cpu().numpy()
+            out_2 = second.cpu().numpy()
+            out_1 = out_1[~np.isnan(out_1)]
+            out_2 = out_2[~np.isnan(out_2)]
+            max_diff = np.amax(np.abs(out_1 - out_2))
+            self.assertLessEqual(max_diff, 1e-5)
+
+        for model_class in self.all_model_classes:
+            model = model_class(config)
+            model.to(torch_device)
+            model.eval()
+            with torch.no_grad():
+                first = model(**self._prepare_for_class(inputs_dict, model_class))[0]
+                second = model(**self._prepare_for_class(inputs_dict, model_class))[0]
+
+            if isinstance(first, tuple) and isinstance(second, tuple):
+                for tensor1, tensor2 in zip(first, second):
+                    check_determinism(tensor1, tensor2)
+            else:
+                check_determinism(first, second)
+
+    def test_model_outputs_equivalence(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        def set_nan_tensor_to_zero(t):
+            t[t != t] = 0
+            return t
+
+        def check_equivalence(model, tuple_inputs, dict_inputs, additional_kwargs={}):
+            with torch.no_grad():
+                tuple_output = model(**tuple_inputs, return_dict=False, **additional_kwargs)
+                dict_output = model(**dict_inputs, return_dict=True, **additional_kwargs)
+
+                def recursive_check(tuple_object, dict_object):
+                    if isinstance(tuple_object, (List, Tuple)):
+                        for tuple_iterable_value, dict_iterable_value in zip(tuple_object, dict_object):
+                            recursive_check(tuple_iterable_value, dict_iterable_value)
+                    elif isinstance(tuple_object, Dict):
+                        for tuple_iterable_value, dict_iterable_value in zip(
+                            tuple_object.values(), dict_object.values()
+                        ):
+                            recursive_check(tuple_iterable_value, dict_iterable_value)
+                    elif tuple_object is None:
+                        return
+                    else:
+                        self.assertTrue(
+                            torch.allclose(
+                                set_nan_tensor_to_zero(tuple_object), set_nan_tensor_to_zero(dict_object), atol=1e-5
+                            ),
+                            msg=(
+                                "Tuple and dict output are not equal. Difference:"
+                                f" {torch.max(torch.abs(tuple_object - dict_object))}. Tuple has `nan`:"
+                                f" {torch.isnan(tuple_object).any()} and `inf`: {torch.isinf(tuple_object)}. Dict has"
+                                f" `nan`: {torch.isnan(dict_object).any()} and `inf`: {torch.isinf(dict_object)}."
+                            ),
+                        )
+
+                recursive_check(tuple_output, dict_output)
+
+        for model_class in self.all_model_classes:
+            model = model_class(config)
+            model.to(torch_device)
+            model.eval()
+
+            tuple_inputs = self._prepare_for_class(inputs_dict, model_class)
+            dict_inputs = self._prepare_for_class(inputs_dict, model_class)
+            check_equivalence(model, tuple_inputs, dict_inputs)
+
+    def test_initialization(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        configs_no_init = _config_zero_init(config)
+        for model_class in self.all_model_classes:
+            model = model_class(config=configs_no_init)
+            for name, param in model.named_parameters():
+                uniform_init_parms = ["conv"]
+                ignore_init = ["lstm"]
+                if param.requires_grad:
+                    if any([x in name for x in uniform_init_parms]):
+                        self.assertTrue(
+                            -1.0 <= ((param.data.mean() * 1e9).round() / 1e9).item() <= 1.0,
+                            msg=f"Parameter {name} of model {model_class} seems not properly initialized",
+                        )
+                    elif not any([x in name for x in ignore_init]):
+                        self.assertIn(
+                            ((param.data.mean() * 1e9).round() / 1e9).item(),
+                            [0.0, 1.0],
+                            msg=f"Parameter {name} of model {model_class} seems not properly initialized",
+                        )
+
+
+def normalize(arr):
+    norm = np.linalg.norm(arr)
+    normalized_arr = arr / norm
+    return normalized_arr
+
+
+def compute_rmse(arr1, arr2):
+    arr1_normalized = normalize(arr1)
+    arr2_normalized = normalize(arr2)
+    return np.sqrt(((arr1_normalized - arr2_normalized) ** 2).mean())
+
+
+@slow
+@require_torch
+class EncodecIntegrationTest(unittest.TestCase):
+    def test_integration_24kHz(self):
+        expected_rmse = {
+            "1.5": 0.0025,
+            "24.0": 0.0015,
+        }
+        expected_codesums = {
+            "1.5": [367184],
+            "24.0": [6648961],
+        }
+        librispeech_dummy = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
+        model_id = "facebook/encodec_24khz"
+
+        model = EncodecModel.from_pretrained(model_id).to(torch_device)
+        processor = AutoProcessor.from_pretrained(model_id)
+
+        librispeech_dummy = librispeech_dummy.cast_column("audio", Audio(sampling_rate=processor.sampling_rate))
+        audio_sample = librispeech_dummy[-1]["audio"]["array"]
+
+        inputs = processor(
+            raw_audio=audio_sample,
+            sampling_rate=processor.sampling_rate,
+            return_tensors="pt",
+        ).to(torch_device)
+
+        for bandwidth, expected_rmse in expected_rmse.items():
+            with torch.no_grad():
+                # use max bandwith for best possible reconstruction
+                encoder_outputs = model.encode(inputs["input_values"], bandwidth=float(bandwidth))
+
+                audio_code_sums = [a[0].sum().cpu().item() for a in encoder_outputs[0]]
+
+                # make sure audio encoded codes are correct
+                self.assertListEqual(audio_code_sums, expected_codesums[bandwidth])
+
+                audio_codes, scales = encoder_outputs.to_tuple()
+                input_values_dec = model.decode(audio_codes, scales, inputs["padding_mask"])[0]
+                input_values_enc_dec = model(
+                    inputs["input_values"], inputs["padding_mask"], bandwidth=float(bandwidth)
+                )[-1]
+
+            # make sure forward and decode gives same result
+            self.assertTrue(torch.allclose(input_values_dec, input_values_enc_dec, atol=1e-3))
+
+            # make sure shape matches
+            self.assertTrue(inputs["input_values"].shape == input_values_enc_dec.shape)
+
+            arr = inputs["input_values"][0].cpu().numpy()
+            arr_enc_dec = input_values_enc_dec[0].cpu().numpy()
+
+            # make sure audios are more or less equal
+            # the RMSE of two random gaussian noise vectors with ~N(0, 1) is around 1.0
+            rmse = compute_rmse(arr, arr_enc_dec)
+            self.assertTrue(rmse < expected_rmse)
+
+    def test_integration_48kHz(self):
+        expected_rmse = {
+            "3.0": 0.001,
+            "24.0": 0.0005,
+        }
+        expected_codesums = {
+            "3.0": [142174, 147901, 154090, 178965, 161879],
+            "24.0": [1561048, 1284593, 1278330, 1487220, 1659404],
+        }
+        librispeech_dummy = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
+        model_id = "facebook/encodec_48khz"
+
+        model = EncodecModel.from_pretrained(model_id).to(torch_device)
+        model = model.eval()
+        processor = AutoProcessor.from_pretrained(model_id)
+
+        librispeech_dummy = librispeech_dummy.cast_column("audio", Audio(sampling_rate=processor.sampling_rate))
+        audio_sample = librispeech_dummy[-1]["audio"]["array"]
+
+        # transform mono to stereo
+        audio_sample = np.array([audio_sample, audio_sample])
+
+        inputs = processor(raw_audio=audio_sample, sampling_rate=processor.sampling_rate, return_tensors="pt").to(
+            torch_device
+        )
+
+        for bandwidth, expected_rmse in expected_rmse.items():
+            with torch.no_grad():
+                # use max bandwith for best possible reconstruction
+                encoder_outputs = model.encode(
+                    inputs["input_values"], inputs["padding_mask"], bandwidth=float(bandwidth), return_dict=False
+                )
+                audio_code_sums = [a[0].sum().cpu().item() for a in encoder_outputs[0]]
+
+                # make sure audio encoded codes are correct
+                self.assertListEqual(audio_code_sums, expected_codesums[bandwidth])
+                audio_codes, scales = encoder_outputs
+                input_values_dec = model.decode(audio_codes, scales, inputs["padding_mask"])[0]
+                input_values_enc_dec = model(
+                    inputs["input_values"], inputs["padding_mask"], bandwidth=float(bandwidth)
+                )[-1]
+
+            # make sure forward and decode gives same result
+            self.assertTrue(torch.allclose(input_values_dec, input_values_enc_dec, atol=1e-3))
+
+            # make sure shape matches
+            self.assertTrue(inputs["input_values"].shape == input_values_enc_dec.shape)
+
+            arr = inputs["input_values"][0].cpu().numpy()
+            arr_enc_dec = input_values_enc_dec[0].cpu().numpy()
+
+            # make sure audios are more or less equal
+            # the RMSE of two random gaussian noise vectors with ~N(0, 1) is around 1.0
+            rmse = compute_rmse(arr, arr_enc_dec)
+            self.assertTrue(rmse < expected_rmse)
+
+    def test_batch_48kHz(self):
+        expected_rmse = {
+            "3.0": 0.001,
+            "24.0": 0.0005,
+        }
+        expected_codesums = {
+            "3.0": [
+                [71689, 78549, 75644, 88889, 73100, 82509, 71449, 82835],
+                [84427, 82356, 75809, 52509, 80137, 87672, 87436, 70456],
+            ],
+            "24.0": [
+                [71689, 78549, 75644, 88889, 73100, 82509, 71449, 82835],
+                [84427, 82356, 75809, 52509, 80137, 87672, 87436, 70456],
+            ],
+        }
+        librispeech_dummy = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
+        model_id = "facebook/encodec_48khz"
+
+        model = EncodecModel.from_pretrained(model_id).to(torch_device)
+        processor = AutoProcessor.from_pretrained(model_id, chunk_length_s=1, overlap=0.01)
+
+        librispeech_dummy = librispeech_dummy.cast_column("audio", Audio(sampling_rate=processor.sampling_rate))
+
+        audio_samples = [
+            np.array([audio_sample["array"], audio_sample["array"]])
+            for audio_sample in librispeech_dummy[-2:]["audio"]
+        ]
+
+        inputs = processor(raw_audio=audio_samples, sampling_rate=processor.sampling_rate, return_tensors="pt")
+        input_values = inputs["input_values"].to(torch_device)
+        for bandwidth, expected_rmse in expected_rmse.items():
+            with torch.no_grad():
+                # use max bandwith for best possible reconstruction
+                encoder_outputs = model.encode(input_values, bandwidth=float(bandwidth), return_dict=False)
+                audio_code_sums_0 = [a[0][0].sum().cpu().item() for a in encoder_outputs[0]]
+                audio_code_sums_1 = [a[0][1].sum().cpu().item() for a in encoder_outputs[0]]
+
+                # make sure audio encoded codes are correct
+                self.assertListEqual(audio_code_sums_0, expected_codesums[bandwidth][0])
+                self.assertListEqual(audio_code_sums_1, expected_codesums[bandwidth][1])
+
+                audio_codes, scales = encoder_outputs
+                input_values_dec = model.decode(audio_codes, scales)[0]
+                input_values_enc_dec = model(input_values, bandwidth=float(bandwidth))[-1]
+
+            # make sure forward and decode gives same result
+            self.assertTrue(torch.allclose(input_values_dec, input_values_enc_dec, atol=1e-3))
+
+            # make sure shape matches
+            self.assertTrue(input_values.shape == input_values_enc_dec.shape)
+
+            arr = input_values[0].cpu().numpy()
+            arr_enc_dec = input_values_enc_dec[0].cpu().numpy()
+
+            # make sure audios are more or less equal
+            # the RMSE of two random gaussian noise vectors with ~N(0, 1) is around 1.0
+            rmse = compute_rmse(arr, arr_enc_dec)
+            self.assertTrue(rmse < expected_rmse)

utils/check_config_attributes.py

@@ -31,6 +31,8 @@ transformers = direct_transformers_import(PATH_TO_TRANSFORMERS)
 CONFIG_MAPPING = transformers.models.auto.configuration_auto.CONFIG_MAPPING
 
 SPECIAL_CASES_TO_ALLOW = {
+    # used to compute the property `self.chunk_length`
+    "EncodecConfig": ["overlap"],
     # used as `self.bert_model = BertModel(config, ...)`
     "DPRConfig": True,
     # not used in modeling files, but it's an important information

utils/documentation_tests.txt

@@ -83,6 +83,7 @@ src/transformers/models/electra/configuration_electra.py
 src/transformers/models/electra/modeling_electra.py
 src/transformers/models/electra/modeling_tf_electra.py
 src/transformers/models/efficientformer/modeling_tf_efficientformer.py
+src/transformers/models/encodec/modeling_encodec.py
 src/transformers/models/ernie/configuration_ernie.py
 src/transformers/models/ernie_m/configuration_ernie_m.py
 src/transformers/models/ernie_m/modeling_ernie_m.py
@@ -395,6 +396,7 @@ src/transformers/models/deit/feature_extraction_deit.py
 src/transformers/models/detr/feature_extraction_detr.py
 src/transformers/models/donut/feature_extraction_donut.py
 src/transformers/models/dpt/feature_extraction_dpt.py
+src/transformers/models/encodec/feature_extraction_encodec.py
 src/transformers/models/flava/feature_extraction_flava.py
 src/transformers/models/glpn/feature_extraction_glpn.py
 src/transformers/models/imagegpt/feature_extraction_imagegpt.py