diff --git a/README.md b/README.md index e8c5d42b5cc..ca026661a4d 100644 --- a/README.md +++ b/README.md @@ -400,6 +400,7 @@ Current number of checkpoints: ![](https://img.shields.io/endpoint?url=https://h 1. **[SegFormer](https://huggingface.co/docs/transformers/model_doc/segformer)** (from NVIDIA) released with the paper [SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers](https://arxiv.org/abs/2105.15203) by Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, Ping Luo. 1. **[SEW](https://huggingface.co/docs/transformers/model_doc/sew)** (from ASAPP) released with the paper [Performance-Efficiency Trade-offs in Unsupervised Pre-training for Speech Recognition](https://arxiv.org/abs/2109.06870) by Felix Wu, Kwangyoun Kim, Jing Pan, Kyu Han, Kilian Q. Weinberger, Yoav Artzi. 1. **[SEW-D](https://huggingface.co/docs/transformers/model_doc/sew_d)** (from ASAPP) released with the paper [Performance-Efficiency Trade-offs in Unsupervised Pre-training for Speech Recognition](https://arxiv.org/abs/2109.06870) by Felix Wu, Kwangyoun Kim, Jing Pan, Kyu Han, Kilian Q. Weinberger, Yoav Artzi. +1. **[SpeechT5](https://huggingface.co/docs/transformers/main/model_doc/speecht5)** (from Microsoft Research) released with the paper [SpeechT5: Unified-Modal Encoder-Decoder Pre-Training for Spoken Language Processing](https://arxiv.org/abs/2110.07205) by Junyi Ao, Rui Wang, Long Zhou, Chengyi Wang, Shuo Ren, Yu Wu, Shujie Liu, Tom Ko, Qing Li, Yu Zhang, Zhihua Wei, Yao Qian, Jinyu Li, Furu Wei. 1. **[SpeechToTextTransformer](https://huggingface.co/docs/transformers/model_doc/speech_to_text)** (from Facebook), released together with the paper [fairseq S2T: Fast Speech-to-Text Modeling with fairseq](https://arxiv.org/abs/2010.05171) by Changhan Wang, Yun Tang, Xutai Ma, Anne Wu, Dmytro Okhonko, Juan Pino. 1. **[SpeechToTextTransformer2](https://huggingface.co/docs/transformers/model_doc/speech_to_text_2)** (from Facebook), released together with the paper [Large-Scale Self- and Semi-Supervised Learning for Speech Translation](https://arxiv.org/abs/2104.06678) by Changhan Wang, Anne Wu, Juan Pino, Alexei Baevski, Michael Auli, Alexis Conneau. 1. **[Splinter](https://huggingface.co/docs/transformers/model_doc/splinter)** (from Tel Aviv University), released together with the paper [Few-Shot Question Answering by Pretraining Span Selection](https://arxiv.org/abs/2101.00438) by Ori Ram, Yuval Kirstain, Jonathan Berant, Amir Globerson, Omer Levy. diff --git a/README_es.md b/README_es.md index daadf631acc..baa63d9600e 100644 --- a/README_es.md +++ b/README_es.md @@ -328,8 +328,8 @@ Número actual de puntos de control: ![](https://img.shields.io/endpoint?url=htt 1. **[GPT NeoX](https://huggingface.co/docs/transformers/model_doc/gpt_neox)** (from EleutherAI) released with the paper [GPT-NeoX-20B: An Open-Source Autoregressive Language Model](https://arxiv.org/abs/2204.06745) by Sid Black, Stella Biderman, Eric Hallahan, Quentin Anthony, Leo Gao, Laurence Golding, Horace He, Connor Leahy, Kyle McDonell, Jason Phang, Michael Pieler, USVSN Sai Prashanth, Shivanshu Purohit, Laria Reynolds, Jonathan Tow, Ben Wang, Samuel Weinbach 1. **[GPT NeoX Japanese](https://huggingface.co/docs/transformers/model_doc/gpt_neox_japanese)** (from ABEJA) released by Shinya Otani, Takayoshi Makabe, Anuj Arora, and Kyo Hattori. 1. **[GPT-2](https://huggingface.co/docs/transformers/model_doc/gpt2)** (from OpenAI) released with the paper [Language Models are Unsupervised Multitask Learners](https://blog.openai.com/better-language-models/) by Alec Radford*, Jeffrey Wu*, Rewon Child, David Luan, Dario Amodei** and Ilya Sutskever**. -1. **[GPT-J](https://huggingface.co/docs/transformers/model_doc/gptj)** (from EleutherAI) released in the repository [kingoflolz/mesh-transformer-jax](https://github.com/kingoflolz/mesh-transformer-jax/) by Ben Wang and Aran Komatsuzaki. -1. **[GPT-Sw3](https://huggingface.co/docs/transformers/model_doc/gpt-sw3)** (from AI-Sweden) released with the paper [Lessons Learned from GPT-SW3: Building the First Large-Scale Generative Language Model for Swedish](http://www.lrec-conf.org/proceedings/lrec2022/pdf/2022.lrec-1.376.pdf) by Ariel Ekgren, Amaru Cuba Gyllensten, Evangelia Gogoulou, Alice Heiman, Severine Verlinden, Joey Öhman, Fredrik Carlsson, Magnus Sahlgren. +1. **[GPT-J](https://huggingface.co/docs/transformers/model_doc/gptj)** (from EleutherAI) released in the repository [kingoflolz/mesh-transformer-jax](https://github.com/kingoflolz/mesh-transformer-jax/) by Ben Wang and Aran Komatsuzaki. +1. **[GPT-Sw3](https://huggingface.co/docs/transformers/model_doc/gpt-sw3)** (from AI-Sweden) released with the paper [Lessons Learned from GPT-SW3: Building the First Large-Scale Generative Language Model for Swedish](http://www.lrec-conf.org/proceedings/lrec2022/pdf/2022.lrec-1.376.pdf) by Ariel Ekgren, Amaru Cuba Gyllensten, Evangelia Gogoulou, Alice Heiman, Severine Verlinden, Joey Öhman, Fredrik Carlsson, Magnus Sahlgren. 1. **[Graphormer](https://huggingface.co/docs/transformers/model_doc/graphormer)** (from Microsoft) released with the paper [Do Transformers Really Perform Bad for Graph Representation?](https://arxiv.org/abs/2106.05234) by Chengxuan Ying, Tianle Cai, Shengjie Luo, Shuxin Zheng, Guolin Ke, Di He, Yanming Shen, Tie-Yan Liu. 1. **[GroupViT](https://huggingface.co/docs/transformers/model_doc/groupvit)** (from UCSD, NVIDIA) released with the paper [GroupViT: Semantic Segmentation Emerges from Text Supervision](https://arxiv.org/abs/2202.11094) by Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, Xiaolong Wang. 1. **[Hubert](https://huggingface.co/docs/transformers/model_doc/hubert)** (from Facebook) released with the paper [HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units](https://arxiv.org/abs/2106.07447) by Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, Abdelrahman Mohamed. @@ -393,6 +393,7 @@ Número actual de puntos de control: ![](https://img.shields.io/endpoint?url=htt 1. **[SegFormer](https://huggingface.co/docs/transformers/model_doc/segformer)** (from NVIDIA) released with the paper [SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers](https://arxiv.org/abs/2105.15203) by Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, Ping Luo. 1. **[SEW](https://huggingface.co/docs/transformers/model_doc/sew)** (from ASAPP) released with the paper [Performance-Efficiency Trade-offs in Unsupervised Pre-training for Speech Recognition](https://arxiv.org/abs/2109.06870) by Felix Wu, Kwangyoun Kim, Jing Pan, Kyu Han, Kilian Q. Weinberger, Yoav Artzi. 1. **[SEW-D](https://huggingface.co/docs/transformers/model_doc/sew_d)** (from ASAPP) released with the paper [Performance-Efficiency Trade-offs in Unsupervised Pre-training for Speech Recognition](https://arxiv.org/abs/2109.06870) by Felix Wu, Kwangyoun Kim, Jing Pan, Kyu Han, Kilian Q. Weinberger, Yoav Artzi. +1. **[SpeechT5](https://huggingface.co/docs/transformers/main/model_doc/speecht5)** (from Microsoft Research) released with the paper [SpeechT5: Unified-Modal Encoder-Decoder Pre-Training for Spoken Language Processing](https://arxiv.org/abs/2110.07205) by Junyi Ao, Rui Wang, Long Zhou, Chengyi Wang, Shuo Ren, Yu Wu, Shujie Liu, Tom Ko, Qing Li, Yu Zhang, Zhihua Wei, Yao Qian, Jinyu Li, Furu Wei. 1. **[SpeechToTextTransformer](https://huggingface.co/docs/transformers/model_doc/speech_to_text)** (from Facebook), released together with the paper [fairseq S2T: Fast Speech-to-Text Modeling with fairseq](https://arxiv.org/abs/2010.05171) by Changhan Wang, Yun Tang, Xutai Ma, Anne Wu, Dmytro Okhonko, Juan Pino. 1. **[SpeechToTextTransformer2](https://huggingface.co/docs/transformers/model_doc/speech_to_text_2)** (from Facebook), released together with the paper [Large-Scale Self- and Semi-Supervised Learning for Speech Translation](https://arxiv.org/abs/2104.06678) by Changhan Wang, Anne Wu, Juan Pino, Alexei Baevski, Michael Auli, Alexis Conneau. 1. **[Splinter](https://huggingface.co/docs/transformers/model_doc/splinter)** (from Tel Aviv University), released together with the paper [Few-Shot Question Answering by Pretraining Span Selection](https://arxiv.org/abs/2101.00438) by Ori Ram, Yuval Kirstain, Jonathan Berant, Amir Globerson, Omer Levy. diff --git a/README_hd.md b/README_hd.md index 3adb4f3f611..b552761cba2 100644 --- a/README_hd.md +++ b/README_hd.md @@ -288,7 +288,7 @@ conda install -c huggingface transformers 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. **[ESM](https://huggingface.co/docs/transformers/model_doc/esm)** (मेटा AI से) ट्रांसफॉर्मर प्रोटीन भाषा मॉडल हैं। **ESM-1b** पेपर के साथ जारी किया गया था [ अलेक्जेंडर राइव्स, जोशुआ मेयर, टॉम सर्कु, सिद्धार्थ गोयल, ज़ेमिंग लिन द्वारा जैविक संरचना और कार्य असुरक्षित सीखने को 250 मिलियन प्रोटीन अनुक्रमों तक स्केल करने से उभरता है] (https://www.pnas.org/content/118/15/e2016239118) जेसन लियू, डेमी गुओ, मायल ओट, सी. लॉरेंस ज़िटनिक, जेरी मा और रॉब फर्गस। **ESM-1v** को पेपर के साथ जारी किया गया था [भाषा मॉडल प्रोटीन फ़ंक्शन पर उत्परिवर्तन के प्रभावों की शून्य-शॉट भविष्यवाणी को सक्षम करते हैं] (https://doi.org/10.1101/2021.07.09.450648) जोशुआ मेयर, रोशन राव, रॉबर्ट वेरकुइल, जेसन लियू, टॉम सर्कु और अलेक्जेंडर राइव्स द्वारा। **ESM-2** को पेपर के साथ जारी किया गया था [भाषा मॉडल विकास के पैमाने पर प्रोटीन अनुक्रम सटीक संरचना भविष्यवाणी को सक्षम करते हैं](https://doi.org/10.1101/2022.07.20.500902) ज़ेमिंग लिन, हलील अकिन, रोशन राव, ब्रायन ही, झोंगकाई झू, वेंटिंग लू, ए द्वारा लान डॉस सैंटोस कोस्टा, मरियम फ़ज़ल-ज़रंडी, टॉम सर्कू, साल कैंडिडो, अलेक्जेंडर राइव्स। -1. **[FLAN-T5](https://huggingface.co/docs/transformers/model_doc/flan-t5)** (from Google AI) released in the repository [google-research/t5x](https://github.com/google-research/t5x/blob/main/docs/models.md#flan-t5-checkpoints) by Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei +1. **[FLAN-T5](https://huggingface.co/docs/transformers/model_doc/flan-t5)** (from Google AI) released in the repository [google-research/t5x](https://github.com/google-research/t5x/blob/main/docs/models.md#flan-t5-checkpoints) by Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei 1. **[FlauBERT](https://huggingface.co/docs/transformers/model_doc/flaubert)** (CNRS से) साथ वाला पेपर [FlauBERT: Unsupervised Language Model Pre-training for फ़्रेंच](https://arxiv .org/abs/1912.05372) Hang Le, Loïc Vial, Jibril Frej, Vincent Segonne, Maximin Coavoux, बेंजामिन लेकोउटेक्स, अलेक्जेंड्रे अल्लाउज़ेन, बेनोइट क्रैबे, लॉरेंट बेसेसियर, डिडिएर श्वाब द्वारा। 1. **[FLAVA](https://huggingface.co/docs/transformers/model_doc/flava)** (FLAVA: A फाउंडेशनल लैंग्वेज एंड विजन अलाइनमेंट मॉडल) (https://arxiv) साथ वाला पेपर .org/abs/2112.04482) अमनप्रीत सिंह, रोंगहांग हू, वेदानुज गोस्वामी, गुइल्यूम कुएरॉन, वोज्शिएक गालुबा, मार्कस रोहरबैक, और डौवे कीला द्वारा। 1. **[FNet](https://huggingface.co/docs/transformers/model_doc/fnet)** (गूगल रिसर्च से) साथ वाला पेपर [FNet: मिक्सिंग टोकन विद फूरियर ट्रांसफॉर्म्स](https://arxiv.org /abs/2105.03824) जेम्स ली-थॉर्प, जोशुआ आइंस्ली, इल्या एकस्टीन, सैंटियागो ओंटानन द्वारा। @@ -365,6 +365,7 @@ conda install -c huggingface transformers 1. **[SegFormer](https://huggingface.co/docs/transformers/model_doc/segformer)** (from NVIDIA) released with the paper [SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers](https://arxiv.org/abs/2105.15203) by Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, Ping Luo. 1. **[SEW](https://huggingface.co/docs/transformers/model_doc/sew)** (ASAPP से) साथ देने वाला पेपर [भाषण पहचान के लिए अनसुपरवाइज्ड प्री-ट्रेनिंग में परफॉर्मेंस-एफिशिएंसी ट्रेड-ऑफ्स](https ://arxiv.org/abs/2109.06870) फेलिक्स वू, क्वांगयुन किम, जिंग पैन, क्यू हान, किलियन क्यू. वेनबर्गर, योव आर्टज़ी द्वारा। 1. **[SEW-D](https://huggingface.co/docs/transformers/model_doc/sew_d)** (ASAPP से) साथ में पेपर [भाषण पहचान के लिए अनसुपरवाइज्ड प्री-ट्रेनिंग में परफॉर्मेंस-एफिशिएंसी ट्रेड-ऑफ्स] (https://arxiv.org/abs/2109.06870) फेलिक्स वू, क्वांगयुन किम, जिंग पैन, क्यू हान, किलियन क्यू. वेनबर्गर, योआव आर्टज़ी द्वारा पोस्ट किया गया। +1. **[SpeechT5](https://huggingface.co/docs/transformers/main/model_doc/speecht5)** (from Microsoft Research) released with the paper [SpeechT5: Unified-Modal Encoder-Decoder Pre-Training for Spoken Language Processing](https://arxiv.org/abs/2110.07205) by Junyi Ao, Rui Wang, Long Zhou, Chengyi Wang, Shuo Ren, Yu Wu, Shujie Liu, Tom Ko, Qing Li, Yu Zhang, Zhihua Wei, Yao Qian, Jinyu Li, Furu Wei. 1. **[SpeechToTextTransformer](https://huggingface.co/docs/transformers/model_doc/speech_to_text)** (फेसबुक से), साथ में पेपर [फेयरसेक S2T: फास्ट स्पीच-टू-टेक्स्ट मॉडलिंग विद फेयरसेक](https: //arxiv.org/abs/2010.05171) चांगहान वांग, यूं तांग, जुताई मा, ऐनी वू, दिमित्रो ओखोनको, जुआन पिनो द्वारा पोस्ट किया गया。 1. **[SpeechToTextTransformer2](https://huggingface.co/docs/transformers/model_doc/speech_to_text_2)** (फेसबुक से) साथ में पेपर [लार्ज-स्केल सेल्फ- एंड सेमी-सुपरवाइज्ड लर्निंग फॉर स्पीच ट्रांसलेशन](https://arxiv.org/abs/2104.06678) चांगहान वांग, ऐनी वू, जुआन पिनो, एलेक्सी बेवस्की, माइकल औली, एलेक्सिस द्वारा Conneau द्वारा पोस्ट किया गया। 1. **[Splinter](https://huggingface.co/docs/transformers/model_doc/splinter)** (तेल अवीव यूनिवर्सिटी से) साथ में पेपर [स्पैन सिलेक्शन को प्री-ट्रेनिंग करके कुछ-शॉट क्वेश्चन आंसरिंग](https:// arxiv.org/abs/2101.00438) ओरि राम, युवल कर्स्टन, जोनाथन बेरेंट, अमीर ग्लोबर्सन, ओमर लेवी द्वारा। @@ -383,7 +384,7 @@ conda install -c huggingface transformers 1. **[Trajectory Transformer](https://huggingface.co/docs/transformers/model_doc/trajectory_transformers)** (from the University of California at Berkeley) released with the paper [Offline Reinforcement Learning as One Big Sequence Modeling Problem](https://arxiv.org/abs/2106.02039) by Michael Janner, Qiyang Li, Sergey Levine 1. **[Transformer-XL](https://huggingface.co/docs/transformers/model_doc/transfo-xl)** (Google/CMU की ओर से) कागज के साथ [संस्करण-एक्स: एक ब्लॉग मॉडल चौकस चौक मॉडल मॉडल] (https://arxivorg/abs/1901.02860) क्वोकोक वी. ले, रुस्लैन सलाखुतदी 1. **[TrOCR](https://huggingface.co/docs/transformers/model_doc/trocr)** (from Microsoft) released with the paper [TrOCR: Transformer-based Optical Character Recognition with Pre-trained Models](https://arxiv.org/abs/2109.10282) by Minghao Li, Tengchao Lv, Lei Cui, Yijuan Lu, Dinei Florencio, Cha Zhang, Zhoujun Li, Furu Wei. -1. **[UL2](https://huggingface.co/docs/transformers/model_doc/ul2)** (from Google Research) released with the paper [Unifying Language Learning Paradigms](https://arxiv.org/abs/2205.05131v1) by Yi Tay, Mostafa Dehghani, Vinh Q. Tran, Xavier Garcia, Dara Bahri, Tal Schuster, Huaixiu Steven Zheng, Neil Houlsby, Donald Metzler +1. **[UL2](https://huggingface.co/docs/transformers/model_doc/ul2)** (from Google Research) released with the paper [Unifying Language Learning Paradigms](https://arxiv.org/abs/2205.05131v1) by Yi Tay, Mostafa Dehghani, Vinh Q. Tran, Xavier Garcia, Dara Bahri, Tal Schuster, Huaixiu Steven Zheng, Neil Houlsby, Donald Metzler 1. **[UniSpeech](https://huggingface.co/docs/transformers/model_doc/unispeech)** (माइक्रोसॉफ्ट रिसर्च से) साथ में दिया गया पेपर [UniSpeech: यूनिफाइड स्पीच रिप्रेजेंटेशन लर्निंग विद लेबलेड एंड अनलेबल्ड डेटा](https:/ /arxiv.org/abs/2101.07597) चेंगई वांग, यू वू, याओ कियान, केनिची कुमातानी, शुजी लियू, फुरु वेई, माइकल ज़ेंग, ज़ुएदोंग हुआंग द्वारा। 1. **[UniSpeechSat](https://huggingface.co/docs/transformers/model_doc/unispeech-sat)** (माइक्रोसॉफ्ट रिसर्च से) कागज के साथ [UNISPEECH-SAT: यूनिवर्सल स्पीच रिप्रेजेंटेशन लर्निंग विद स्पीकर अवेयर प्री-ट्रेनिंग ](https://arxiv.org/abs/2110.05752) सानयुआन चेन, यू वू, चेंग्यी वांग, झेंगयांग चेन, झूओ चेन, शुजी लियू, जियान वू, याओ कियान, फुरु वेई, जिन्यु ली, जियांगज़ान यू द्वारा पोस्ट किया गया। 1. **[UPerNet](https://huggingface.co/docs/transformers/model_doc/upernet)** (from Peking University) released with the paper [Unified Perceptual Parsing for Scene Understanding](https://arxiv.org/abs/1807.10221) by Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, Jian Sun. diff --git a/README_ja.md b/README_ja.md index d24c61ad5ed..6157d65ff91 100644 --- a/README_ja.md +++ b/README_ja.md @@ -349,7 +349,7 @@ Flax、PyTorch、TensorFlowをcondaでインストールする方法は、それ 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. **[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. **[ESM](https://huggingface.co/docs/transformers/model_doc/esm)** (Meta AI から) はトランスフォーマープロテイン言語モデルです. **ESM-1b** は Alexander Rives, Joshua Meier, Tom Sercu, Siddharth Goyal, Zeming Lin, Jason Liu, Demi Guo, Myle Ott, C. Lawrence Zitnick, Jerry Ma, and Rob Fergus から公開された研究論文: [Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences](https://www.pnas.org/content/118/15/e2016239118). **ESM-1v** は Joshua Meier, Roshan Rao, Robert Verkuil, Jason Liu, Tom Sercu and Alexander Rives　から公開された研究論文: [Language models enable zero-shot prediction of the effects of mutations on protein function](https://doi.org/10.1101/2021.07.09.450648). **ESM-2** と　**ESMFold** は Zeming Lin, Halil Akin, Roshan Rao, Brian Hie, Zhongkai Zhu, Wenting Lu, Allan dos Santos Costa, Maryam Fazel-Zarandi, Tom Sercu, Sal Candido, Alexander Rives から公開された研究論文: [Language models of protein sequences at the scale of evolution enable accurate structure prediction](https://doi.org/10.1101/2022.07.20.500902) +1. **[ESM](https://huggingface.co/docs/transformers/model_doc/esm)** (Meta AI から) はトランスフォーマープロテイン言語モデルです. **ESM-1b** は Alexander Rives, Joshua Meier, Tom Sercu, Siddharth Goyal, Zeming Lin, Jason Liu, Demi Guo, Myle Ott, C. Lawrence Zitnick, Jerry Ma, and Rob Fergus から公開された研究論文: [Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences](https://www.pnas.org/content/118/15/e2016239118). **ESM-1v** は Joshua Meier, Roshan Rao, Robert Verkuil, Jason Liu, Tom Sercu and Alexander Rives　から公開された研究論文: [Language models enable zero-shot prediction of the effects of mutations on protein function](https://doi.org/10.1101/2021.07.09.450648). **ESM-2** と　**ESMFold** は Zeming Lin, Halil Akin, Roshan Rao, Brian Hie, Zhongkai Zhu, Wenting Lu, Allan dos Santos Costa, Maryam Fazel-Zarandi, Tom Sercu, Sal Candido, Alexander Rives から公開された研究論文: [Language models of protein sequences at the scale of evolution enable accurate structure prediction](https://doi.org/10.1101/2022.07.20.500902) 1. **[FLAN-T5](https://huggingface.co/docs/transformers/model_doc/flan-t5)** (Google AI から) Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V から公開されたレポジトリー [google-research/t5x](https://github.com/google-research/t5x/blob/main/docs/models.md#flan-t5-checkpoints) Le, and Jason Wei 1. **[FlauBERT](https://huggingface.co/docs/transformers/model_doc/flaubert)** (CNRS から) Hang Le, Loïc Vial, Jibril Frej, Vincent Segonne, Maximin Coavoux, Benjamin Lecouteux, Alexandre Allauzen, Benoît Crabbé, Laurent Besacier, Didier Schwab から公開された研究論文: [FlauBERT: Unsupervised Language Model Pre-training for French](https://arxiv.org/abs/1912.05372) 1. **[FLAVA](https://huggingface.co/docs/transformers/model_doc/flava)** (Facebook AI から) Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela から公開された研究論文: [FLAVA: A Foundational Language And Vision Alignment Model](https://arxiv.org/abs/2112.04482) @@ -362,8 +362,8 @@ Flax、PyTorch、TensorFlowをcondaでインストールする方法は、それ 1. **[GPT NeoX](https://huggingface.co/docs/transformers/model_doc/gpt_neox)** (EleutherAI から) Sid Black, Stella Biderman, Eric Hallahan, Quentin Anthony, Leo Gao, Laurence Golding, Horace He, Connor Leahy, Kyle McDonell, Jason Phang, Michael Pieler, USVSN Sai Prashanth, Shivanshu Purohit, Laria Reynolds, Jonathan Tow, Ben Wang, Samuel Weinbach から公開された研究論文: [GPT-NeoX-20B: An Open-Source Autoregressive Language Model](https://arxiv.org/abs/2204.06745) 1. **[GPT NeoX Japanese](https://huggingface.co/docs/transformers/model_doc/gpt_neox_japanese)** (ABEJA から) Shinya Otani, Takayoshi Makabe, Anuj Arora, and Kyo Hattori からリリース. 1. **[GPT-2](https://huggingface.co/docs/transformers/model_doc/gpt2)** (OpenAI から) Alec Radford*, Jeffrey Wu*, Rewon Child, David Luan, Dario Amodei** and Ilya Sutskever** から公開された研究論文: [Language Models are Unsupervised Multitask Learners](https://blog.openai.com/better-language-models/) -1. **[GPT-J](https://huggingface.co/docs/transformers/model_doc/gptj)** (EleutherAI から) Ben Wang and Aran Komatsuzaki から公開されたレポジトリー [kingoflolz/mesh-transformer-jax](https://github.com/kingoflolz/mesh-transformer-jax/) -1. **[GPT-Sw3](https://huggingface.co/docs/transformers/model_doc/gpt-sw3)** (AI-Sweden から) Ariel Ekgren, Amaru Cuba Gyllensten, Evangelia Gogoulou, Alice Heiman, Severine Verlinden, Joey Öhman, Fredrik Carlsson, Magnus Sahlgren から公開された研究論文: [Lessons Learned from GPT-SW3: Building the First Large-Scale Generative Language Model for Swedish](http://www.lrec-conf.org/proceedings/lrec2022/pdf/2022.lrec-1.376.pdf) +1. **[GPT-J](https://huggingface.co/docs/transformers/model_doc/gptj)** (EleutherAI から) Ben Wang and Aran Komatsuzaki から公開されたレポジトリー [kingoflolz/mesh-transformer-jax](https://github.com/kingoflolz/mesh-transformer-jax/) +1. **[GPT-Sw3](https://huggingface.co/docs/transformers/model_doc/gpt-sw3)** (AI-Sweden から) Ariel Ekgren, Amaru Cuba Gyllensten, Evangelia Gogoulou, Alice Heiman, Severine Verlinden, Joey Öhman, Fredrik Carlsson, Magnus Sahlgren から公開された研究論文: [Lessons Learned from GPT-SW3: Building the First Large-Scale Generative Language Model for Swedish](http://www.lrec-conf.org/proceedings/lrec2022/pdf/2022.lrec-1.376.pdf) 1. **[Graphormer](https://huggingface.co/docs/transformers/model_doc/graphormer)** (Microsoft から) Chengxuan Ying, Tianle Cai, Shengjie Luo, Shuxin Zheng, Guolin Ke, Di He, Yanming Shen, Tie-Yan Liu から公開された研究論文: [Do Transformers Really Perform Bad for Graph Representation?](https://arxiv.org/abs/2106.05234). 1. **[GroupViT](https://huggingface.co/docs/transformers/model_doc/groupvit)** (UCSD, NVIDIA から) Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, Xiaolong Wang から公開された研究論文: [GroupViT: Semantic Segmentation Emerges from Text Supervision](https://arxiv.org/abs/2202.11094) 1. **[Hubert](https://huggingface.co/docs/transformers/model_doc/hubert)** (Facebook から) Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, Abdelrahman Mohamed から公開された研究論文: [HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units](https://arxiv.org/abs/2106.07447) @@ -427,6 +427,7 @@ Flax、PyTorch、TensorFlowをcondaでインストールする方法は、それ 1. **[SegFormer](https://huggingface.co/docs/transformers/model_doc/segformer)** (NVIDIA から) Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, Ping Luo から公開された研究論文: [SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers](https://arxiv.org/abs/2105.15203) 1. **[SEW](https://huggingface.co/docs/transformers/model_doc/sew)** (ASAPP から) Felix Wu, Kwangyoun Kim, Jing Pan, Kyu Han, Kilian Q. Weinberger, Yoav Artzi から公開された研究論文: [Performance-Efficiency Trade-offs in Unsupervised Pre-training for Speech Recognition](https://arxiv.org/abs/2109.06870) 1. **[SEW-D](https://huggingface.co/docs/transformers/model_doc/sew_d)** (ASAPP から) Felix Wu, Kwangyoun Kim, Jing Pan, Kyu Han, Kilian Q. Weinberger, Yoav Artzi から公開された研究論文: [Performance-Efficiency Trade-offs in Unsupervised Pre-training for Speech Recognition](https://arxiv.org/abs/2109.06870) +1. **[SpeechT5](https://huggingface.co/docs/transformers/main/model_doc/speecht5)** (Microsoft Research から) Junyi Ao, Rui Wang, Long Zhou, Chengyi Wang, Shuo Ren, Yu Wu, Shujie Liu, Tom Ko, Qing Li, Yu Zhang, Zhihua Wei, Yao Qian, Jinyu Li, Furu Wei. から公開された研究論文 [SpeechT5: Unified-Modal Encoder-Decoder Pre-Training for Spoken Language Processing](https://arxiv.org/abs/2110.07205) 1. **[SpeechToTextTransformer](https://huggingface.co/docs/transformers/model_doc/speech_to_text)** (Facebook から), Changhan Wang, Yun Tang, Xutai Ma, Anne Wu, Dmytro Okhonko, Juan Pino から公開された研究論文: [fairseq S2T: Fast Speech-to-Text Modeling with fairseq](https://arxiv.org/abs/2010.05171) 1. **[SpeechToTextTransformer2](https://huggingface.co/docs/transformers/model_doc/speech_to_text_2)** (Facebook から), Changhan Wang, Anne Wu, Juan Pino, Alexei Baevski, Michael Auli, Alexis Conneau から公開された研究論文: [Large-Scale Self- and Semi-Supervised Learning for Speech Translation](https://arxiv.org/abs/2104.06678) 1. **[Splinter](https://huggingface.co/docs/transformers/model_doc/splinter)** (Tel Aviv University から), Ori Ram, Yuval Kirstain, Jonathan Berant, Amir Globerson, Omer Levy から公開された研究論文: [Few-Shot Question Answering by Pretraining Span Selection](https://arxiv.org/abs/2101.00438) diff --git a/README_ko.md b/README_ko.md index 66a76743967..02f25329aed 100644 --- a/README_ko.md +++ b/README_ko.md @@ -277,7 +277,7 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는 1. **[GPT NeoX](https://huggingface.co/docs/transformers/model_doc/gpt_neox)** (EleutherAI 에서) Sid Black, Stella Biderman, Eric Hallahan, Quentin Anthony, Leo Gao, Laurence Golding, Horace He, Connor Leahy, Kyle McDonell, Jason Phang, Michael Pieler, USVSN Sai Prashanth, Shivanshu Purohit, Laria Reynolds, Jonathan Tow, Ben Wang, Samuel Weinbac 의 [GPT-NeoX-20B: An Open-Source Autoregressive Language Model](https://arxiv.org/abs/2204.06745) 논문과 함께 발표했습니다. 1. **[GPT NeoX Japanese](https://huggingface.co/docs/transformers/model_doc/gpt_neox_japanese)** (from ABEJA) released by Shinya Otani, Takayoshi Makabe, Anuj Arora, and Kyo Hattori. 1. **[GPT-2](https://huggingface.co/docs/transformers/model_doc/gpt2)** (OpenAI 에서) Alec Radford*, Jeffrey Wu*, Rewon Child, David Luan, Dario Amodei** and Ilya Sutskever** 의 [Language Models are Unsupervised Multitask Learners](https://blog.openai.com/better-language-models/) 논문과 함께 발표했습니다. -1. **[GPT-J](https://huggingface.co/docs/transformers/model_doc/gptj)** (from EleutherAI) released in the repository [kingoflolz/mesh-transformer-jax](https://github.com/kingoflolz/mesh-transformer-jax/) by Ben Wang and Aran Komatsuzaki. +1. **[GPT-J](https://huggingface.co/docs/transformers/model_doc/gptj)** (from EleutherAI) released in the repository [kingoflolz/mesh-transformer-jax](https://github.com/kingoflolz/mesh-transformer-jax/) by Ben Wang and Aran Komatsuzaki. 1. **[GPT-Sw3](https://huggingface.co/docs/transformers/model_doc/gpt-sw3)** (AI-Sweden 에서) Ariel Ekgren, Amaru Cuba Gyllensten, Evangelia Gogoulou, Alice Heiman, Severine Verlinden, Joey Öhman, Fredrik Carlsson, Magnus Sahlgren. 의 [Lessons Learned from GPT-SW3: Building the First Large-Scale Generative Language Model for Swedish](http://www.lrec-conf.org/proceedings/lrec2022/pdf/2022.lrec-1.376.pdf) 논문과 함께 발표했습니다. 1. **[Graphormer](https://huggingface.co/docs/transformers/model_doc/graphormer)** (from Microsoft) Chengxuan Ying, Tianle Cai, Shengjie Luo, Shuxin Zheng, Guolin Ke, Di He, Yanming Shen, Tie-Yan Liu 의 [Do Transformers Really Perform Bad for Graph Representation?](https://arxiv.org/abs/2106.05234) 논문과 함께 발표했습니다. 1. **[GroupViT](https://huggingface.co/docs/transformers/model_doc/groupvit)** (UCSD, NVIDIA 에서) Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, Xiaolong Wang 의 [GroupViT: Semantic Segmentation Emerges from Text Supervision](https://arxiv.org/abs/2202.11094) 논문과 함께 발표했습니다. @@ -342,6 +342,7 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는 1. **[SegFormer](https://huggingface.co/docs/transformers/model_doc/segformer)** (NVIDIA 에서) Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, Ping Luo 의 [SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers](https://arxiv.org/abs/2105.15203) 논문과 함께 발표했습니다. 1. **[SEW](https://huggingface.co/docs/transformers/model_doc/sew)** (ASAPP 에서) Felix Wu, Kwangyoun Kim, Jing Pan, Kyu Han, Kilian Q. Weinberger, Yoav Artzi 의 [Performance-Efficiency Trade-offs in Unsupervised Pre-training for Speech Recognition](https://arxiv.org/abs/2109.06870) 논문과 함께 발표했습니다. 1. **[SEW-D](https://huggingface.co/docs/transformers/model_doc/sew_d)** (ASAPP 에서) Felix Wu, Kwangyoun Kim, Jing Pan, Kyu Han, Kilian Q. Weinberger, Yoav Artzi 의 [Performance-Efficiency Trade-offs in Unsupervised Pre-training for Speech Recognition](https://arxiv.org/abs/2109.06870) 논문과 함께 발표했습니다. +1. **[SpeechT5](https://huggingface.co/docs/transformers/main/model_doc/speecht5)** (Microsoft Research 에서 제공)은 Junyi Ao, Rui Wang, Long Zhou, Chengyi Wang, Shuo Ren, Yu Wu, Shujie Liu, Tom Ko, Qing Li, Yu Zhang, Zhihua Wei, Yao Qian, Jinyu Li, Furu Wei.의 [SpeechT5: Unified-Modal Encoder-Decoder Pre-Training for Spoken Language Processing](https://arxiv.org/abs/2110.07205)논문과 함께 발표했습니다. 1. **[SpeechToTextTransformer](https://huggingface.co/docs/transformers/model_doc/speech_to_text)** (Facebook 에서) Changhan Wang, Yun Tang, Xutai Ma, Anne Wu, Dmytro Okhonko, Juan Pino 의 [fairseq S2T: Fast Speech-to-Text Modeling with fairseq](https://arxiv.org/abs/2010.05171) 논문과 함께 발표했습니다. 1. **[SpeechToTextTransformer2](https://huggingface.co/docs/transformers/model_doc/speech_to_text_2)** (Facebook 에서) Changhan Wang, Anne Wu, Juan Pino, Alexei Baevski, Michael Auli, Alexis Conneau 의 [Large-Scale Self- and Semi-Supervised Learning for Speech Translation](https://arxiv.org/abs/2104.06678) 논문과 함께 발표했습니다. 1. **[Splinter](https://huggingface.co/docs/transformers/model_doc/splinter)** (Tel Aviv University 에서) Ori Ram, Yuval Kirstain, Jonathan Berant, Amir Globerson, Omer Levy 의 [Few-Shot Question Answering by Pretraining Span Selection](https://arxiv.org/abs/2101.00438) 논문과 함께 발표했습니다. diff --git a/README_zh-hans.md b/README_zh-hans.md index 0d069c36cb5..46f1057cd2b 100644 --- a/README_zh-hans.md +++ b/README_zh-hans.md @@ -301,8 +301,8 @@ conda install -c huggingface transformers 1. **[GPT NeoX](https://huggingface.co/docs/transformers/model_doc/gpt_neox)** (from EleutherAI) released with the paper [GPT-NeoX-20B: An Open-Source Autoregressive Language Model](https://arxiv.org/abs/2204.06745) by Sid Black, Stella Biderman, Eric Hallahan, Quentin Anthony, Leo Gao, Laurence Golding, Horace He, Connor Leahy, Kyle McDonell, Jason Phang, Michael Pieler, USVSN Sai Prashanth, Shivanshu Purohit, Laria Reynolds, Jonathan Tow, Ben Wang, Samuel Weinbach 1. **[GPT NeoX Japanese](https://huggingface.co/docs/transformers/model_doc/gpt_neox_japanese)** (来自 ABEJA) 由 Shinya Otani, Takayoshi Makabe, Anuj Arora, Kyo Hattori。 1. **[GPT-2](https://huggingface.co/docs/transformers/model_doc/gpt2)** (来自 OpenAI) 伴随论文 [Language Models are Unsupervised Multitask Learners](https://blog.openai.com/better-language-models/) 由 Alec Radford*, Jeffrey Wu*, Rewon Child, David Luan, Dario Amodei** and Ilya Sutskever** 发布。 -1. **[GPT-J](https://huggingface.co/docs/transformers/model_doc/gptj)** (来自 EleutherAI) 伴随论文 [kingoflolz/mesh-transformer-jax](https://github.com/kingoflolz/mesh-transformer-jax/) 由 Ben Wang and Aran Komatsuzaki 发布。 -1. **[GPT-Sw3](https://huggingface.co/docs/transformers/model_doc/gpt-sw3)** (from AI-Sweden) released with the paper [Lessons Learned from GPT-SW3: Building the First Large-Scale Generative Language Model for Swedish](http://www.lrec-conf.org/proceedings/lrec2022/pdf/2022.lrec-1.376.pdf) by Ariel Ekgren, Amaru Cuba Gyllensten, Evangelia Gogoulou, Alice Heiman, Severine Verlinden, Joey Öhman, Fredrik Carlsson, Magnus Sahlgren. +1. **[GPT-J](https://huggingface.co/docs/transformers/model_doc/gptj)** (来自 EleutherAI) 伴随论文 [kingoflolz/mesh-transformer-jax](https://github.com/kingoflolz/mesh-transformer-jax/) 由 Ben Wang and Aran Komatsuzaki 发布。 +1. **[GPT-Sw3](https://huggingface.co/docs/transformers/model_doc/gpt-sw3)** (from AI-Sweden) released with the paper [Lessons Learned from GPT-SW3: Building the First Large-Scale Generative Language Model for Swedish](http://www.lrec-conf.org/proceedings/lrec2022/pdf/2022.lrec-1.376.pdf) by Ariel Ekgren, Amaru Cuba Gyllensten, Evangelia Gogoulou, Alice Heiman, Severine Verlinden, Joey Öhman, Fredrik Carlsson, Magnus Sahlgren. 1. **[Graphormer](https://huggingface.co/docs/transformers/model_doc/graphormer)** (from Microsoft) released with the paper [Do Transformers Really Perform Bad for Graph Representation?](https://arxiv.org/abs/2106.05234) by Chengxuan Ying, Tianle Cai, Shengjie Luo, Shuxin Zheng, Guolin Ke, Di He, Yanming Shen, Tie-Yan Liu. 1. **[GroupViT](https://huggingface.co/docs/transformers/model_doc/groupvit)** (来自 UCSD, NVIDIA) 伴随论文 [GroupViT: Semantic Segmentation Emerges from Text Supervision](https://arxiv.org/abs/2202.11094) 由 Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, Xiaolong Wang 发布。 1. **[Hubert](https://huggingface.co/docs/transformers/model_doc/hubert)** (来自 Facebook) 伴随论文 [HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units](https://arxiv.org/abs/2106.07447) 由 Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, Abdelrahman Mohamed 发布。 @@ -366,6 +366,7 @@ conda install -c huggingface transformers 1. **[SegFormer](https://huggingface.co/docs/transformers/model_doc/segformer)** (来自 NVIDIA) 伴随论文 [SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers](https://arxiv.org/abs/2105.15203) 由 Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, Ping Luo 发布。 1. **[SEW](https://huggingface.co/docs/transformers/model_doc/sew)** (来自 ASAPP) 伴随论文 [Performance-Efficiency Trade-offs in Unsupervised Pre-training for Speech Recognition](https://arxiv.org/abs/2109.06870) 由 Felix Wu, Kwangyoun Kim, Jing Pan, Kyu Han, Kilian Q. Weinberger, Yoav Artzi 发布。 1. **[SEW-D](https://huggingface.co/docs/transformers/model_doc/sew_d)** (来自 ASAPP) 伴随论文 [Performance-Efficiency Trade-offs in Unsupervised Pre-training for Speech Recognition](https://arxiv.org/abs/2109.06870) 由 Felix Wu, Kwangyoun Kim, Jing Pan, Kyu Han, Kilian Q. Weinberger, Yoav Artzi 发布。 +1. **[SpeechT5](https://huggingface.co/docs/transformers/main/model_doc/speecht5)** (来自 Microsoft Research) 伴随论文 [SpeechT5: Unified-Modal Encoder-Decoder Pre-Training for Spoken Language Processing](https://arxiv.org/abs/2110.07205) 由 Junyi Ao, Rui Wang, Long Zhou, Chengyi Wang, Shuo Ren, Yu Wu, Shujie Liu, Tom Ko, Qing Li, Yu Zhang, Zhihua Wei, Yao Qian, Jinyu Li, Furu Wei 发布。 1. **[SpeechToTextTransformer](https://huggingface.co/docs/transformers/model_doc/speech_to_text)** (来自 Facebook), 伴随论文 [fairseq S2T: Fast Speech-to-Text Modeling with fairseq](https://arxiv.org/abs/2010.05171) 由 Changhan Wang, Yun Tang, Xutai Ma, Anne Wu, Dmytro Okhonko, Juan Pino 发布。 1. **[SpeechToTextTransformer2](https://huggingface.co/docs/transformers/model_doc/speech_to_text_2)** (来自 Facebook) 伴随论文 [Large-Scale Self- and Semi-Supervised Learning for Speech Translation](https://arxiv.org/abs/2104.06678) 由 Changhan Wang, Anne Wu, Juan Pino, Alexei Baevski, Michael Auli, Alexis Conneau 发布。 1. **[Splinter](https://huggingface.co/docs/transformers/model_doc/splinter)** (来自 Tel Aviv University) 伴随论文 [Few-Shot Question Answering by Pretraining Span Selection](https://arxiv.org/abs/2101.00438) 由 Ori Ram, Yuval Kirstain, Jonathan Berant, Amir Globerson, Omer Levy 发布。 @@ -373,7 +374,7 @@ conda install -c huggingface transformers 1. **[Swin Transformer](https://huggingface.co/docs/transformers/model_doc/swin)** (来自 Microsoft) 伴随论文 [Swin Transformer: Hierarchical Vision Transformer using Shifted Windows](https://arxiv.org/abs/2103.14030) 由 Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, Baining Guo 发布。 1. **[Swin Transformer V2](https://huggingface.co/docs/transformers/model_doc/swinv2)** (来自 Microsoft) 伴随论文 [Swin Transformer V2: Scaling Up Capacity and Resolution](https://arxiv.org/abs/2111.09883) 由 Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, Furu Wei, Baining Guo 发布。 1. **[Swin2SR](https://huggingface.co/docs/transformers/model_doc/swin2sr)** (来自 University of Würzburg) 伴随论文 [Swin2SR: SwinV2 Transformer for Compressed Image Super-Resolution and Restoration](https://arxiv.org/abs/2209.11345) 由 Marcos V. Conde, Ui-Jin Choi, Maxime Burchi, Radu Timofte 发布。 -1. **[SwitchTransformers](https://huggingface.co/docs/transformers/model_doc/switch_transformers)** (from Google) released with the paper [Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity](https://arxiv.org/abs/2101.03961) by William Fedus, Barret Zoph, Noam Shazeer. +1. **[SwitchTransformers](https://huggingface.co/docs/transformers/model_doc/switch_transformers)** (from Google) released with the paper [Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity](https://arxiv.org/abs/2101.03961) by William Fedus, Barret Zoph, Noam Shazeer. 1. **[T5](https://huggingface.co/docs/transformers/model_doc/t5)** (来自 Google AI) 伴随论文 [Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer](https://arxiv.org/abs/1910.10683) 由 Colin Raffel and Noam Shazeer and Adam Roberts and Katherine Lee and Sharan Narang and Michael Matena and Yanqi Zhou and Wei Li and Peter J. Liu 发布。 1. **[T5v1.1](https://huggingface.co/docs/transformers/model_doc/t5v1.1)** (来自 Google AI) 伴随论文 [google-research/text-to-text-transfer-transformer](https://github.com/google-research/text-to-text-transfer-transformer/blob/main/released_checkpoints.md#t511) 由 Colin Raffel and Noam Shazeer and Adam Roberts and Katherine Lee and Sharan Narang and Michael Matena and Yanqi Zhou and Wei Li and Peter J. Liu 发布。 1. **[Table Transformer](https://huggingface.co/docs/transformers/model_doc/table-transformer)** (来自 Microsoft Research) 伴随论文 [PubTables-1M: Towards Comprehensive Table Extraction From Unstructured Documents](https://arxiv.org/abs/2110.00061) 由 Brandon Smock, Rohith Pesala, Robin Abraham 发布。 diff --git a/README_zh-hant.md b/README_zh-hant.md index 56c5a5c42da..8bd3aea8a6c 100644 --- a/README_zh-hant.md +++ b/README_zh-hant.md @@ -313,8 +313,8 @@ conda install -c huggingface transformers 1. **[GPT NeoX](https://huggingface.co/docs/transformers/model_doc/gpt_neox)** (from EleutherAI) released with the paper [GPT-NeoX-20B: An Open-Source Autoregressive Language Model](https://arxiv.org/abs/2204.06745) by Sid Black, Stella Biderman, Eric Hallahan, Quentin Anthony, Leo Gao, Laurence Golding, Horace He, Connor Leahy, Kyle McDonell, Jason Phang, Michael Pieler, USVSN Sai Prashanth, Shivanshu Purohit, Laria Reynolds, Jonathan Tow, Ben Wang, Samuel Weinbach 1. **[GPT NeoX Japanese](https://huggingface.co/docs/transformers/model_doc/gpt_neox_japanese)** (from ABEJA) released by Shinya Otani, Takayoshi Makabe, Anuj Arora, and Kyo Hattori. 1. **[GPT-2](https://huggingface.co/docs/transformers/model_doc/gpt2)** (from OpenAI) released with the paper [Language Models are Unsupervised Multitask Learners](https://blog.openai.com/better-language-models/) by Alec Radford*, Jeffrey Wu*, Rewon Child, David Luan, Dario Amodei** and Ilya Sutskever**. -1. **[GPT-J](https://huggingface.co/docs/transformers/model_doc/gptj)** (from EleutherAI) released with the paper [kingoflolz/mesh-transformer-jax](https://github.com/kingoflolz/mesh-transformer-jax/) by Ben Wang and Aran Komatsuzaki. -1. **[GPT-Sw3](https://huggingface.co/docs/transformers/model_doc/gpt-sw3)** (from AI-Sweden) released with the paper [Lessons Learned from GPT-SW3: Building the First Large-Scale Generative Language Model for Swedish](http://www.lrec-conf.org/proceedings/lrec2022/pdf/2022.lrec-1.376.pdf) by Ariel Ekgren, Amaru Cuba Gyllensten, Evangelia Gogoulou, Alice Heiman, Severine Verlinden, Joey Öhman, Fredrik Carlsson, Magnus Sahlgren. +1. **[GPT-J](https://huggingface.co/docs/transformers/model_doc/gptj)** (from EleutherAI) released with the paper [kingoflolz/mesh-transformer-jax](https://github.com/kingoflolz/mesh-transformer-jax/) by Ben Wang and Aran Komatsuzaki. +1. **[GPT-Sw3](https://huggingface.co/docs/transformers/model_doc/gpt-sw3)** (from AI-Sweden) released with the paper [Lessons Learned from GPT-SW3: Building the First Large-Scale Generative Language Model for Swedish](http://www.lrec-conf.org/proceedings/lrec2022/pdf/2022.lrec-1.376.pdf) by Ariel Ekgren, Amaru Cuba Gyllensten, Evangelia Gogoulou, Alice Heiman, Severine Verlinden, Joey Öhman, Fredrik Carlsson, Magnus Sahlgren. 1. **[Graphormer](https://huggingface.co/docs/transformers/model_doc/graphormer)** (from Microsoft) released with the paper [Do Transformers Really Perform Bad for Graph Representation?](https://arxiv.org/abs/2106.05234) by Chengxuan Ying, Tianle Cai, Shengjie Luo, Shuxin Zheng, Guolin Ke, Di He, Yanming Shen, Tie-Yan Liu. 1. **[GroupViT](https://huggingface.co/docs/transformers/model_doc/groupvit)** (from UCSD, NVIDIA) released with the paper [GroupViT: Semantic Segmentation Emerges from Text Supervision](https://arxiv.org/abs/2202.11094) by Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, Xiaolong Wang. 1. **[Hubert](https://huggingface.co/docs/transformers/model_doc/hubert)** (from Facebook) released with the paper [HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units](https://arxiv.org/abs/2106.07447) by Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, Abdelrahman Mohamed. @@ -378,6 +378,7 @@ conda install -c huggingface transformers 1. **[SegFormer](https://huggingface.co/docs/transformers/model_doc/segformer)** (from NVIDIA) released with the paper [SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers](https://arxiv.org/abs/2105.15203) by Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, Ping Luo. 1. **[SEW](https://huggingface.co/docs/transformers/model_doc/sew)** (from ASAPP) released with the paper [Performance-Efficiency Trade-offs in Unsupervised Pre-training for Speech Recognition](https://arxiv.org/abs/2109.06870) by Felix Wu, Kwangyoun Kim, Jing Pan, Kyu Han, Kilian Q. Weinberger, Yoav Artzi. 1. **[SEW-D](https://huggingface.co/docs/transformers/model_doc/sew_d)** (from ASAPP) released with the paper [Performance-Efficiency Trade-offs in Unsupervised Pre-training for Speech Recognition](https://arxiv.org/abs/2109.06870) by Felix Wu, Kwangyoun Kim, Jing Pan, Kyu Han, Kilian Q. Weinberger, Yoav Artzi. +1. **[SpeechT5](https://huggingface.co/docs/transformers/main/model_doc/speecht5)** (from Microsoft Research) released with the paper [SpeechT5: Unified-Modal Encoder-Decoder Pre-Training for Spoken Language Processing](https://arxiv.org/abs/2110.07205) by Junyi Ao, Rui Wang, Long Zhou, Chengyi Wang, Shuo Ren, Yu Wu, Shujie Liu, Tom Ko, Qing Li, Yu Zhang, Zhihua Wei, Yao Qian, Jinyu Li, Furu Wei. 1. **[SpeechToTextTransformer](https://huggingface.co/docs/transformers/model_doc/speech_to_text)** (from Facebook), released together with the paper [fairseq S2T: Fast Speech-to-Text Modeling with fairseq](https://arxiv.org/abs/2010.05171) by Changhan Wang, Yun Tang, Xutai Ma, Anne Wu, Dmytro Okhonko, Juan Pino. 1. **[SpeechToTextTransformer2](https://huggingface.co/docs/transformers/model_doc/speech_to_text_2)** (from Facebook) released with the paper [Large-Scale Self- and Semi-Supervised Learning for Speech Translation](https://arxiv.org/abs/2104.06678) by Changhan Wang, Anne Wu, Juan Pino, Alexei Baevski, Michael Auli, Alexis Conneau. 1. **[Splinter](https://huggingface.co/docs/transformers/model_doc/splinter)** (from Tel Aviv University) released with the paper [Few-Shot Question Answering by Pretraining Span Selection](https://arxiv.org/abs/2101.00438) by Ori Ram, Yuval Kirstain, Jonathan Berant, Amir Globerson, Omer Levy. @@ -385,7 +386,7 @@ conda install -c huggingface transformers 1. **[Swin Transformer](https://huggingface.co/docs/transformers/model_doc/swin)** (from Microsoft) released with the paper [Swin Transformer: Hierarchical Vision Transformer using Shifted Windows](https://arxiv.org/abs/2103.14030) by Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, Baining Guo. 1. **[Swin Transformer V2](https://huggingface.co/docs/transformers/model_doc/swinv2)** (from Microsoft) released with the paper [Swin Transformer V2: Scaling Up Capacity and Resolution](https://arxiv.org/abs/2111.09883) by Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, Furu Wei, Baining Guo. 1. **[Swin2SR](https://huggingface.co/docs/transformers/model_doc/swin2sr)** (from University of Würzburg) released with the paper [Swin2SR: SwinV2 Transformer for Compressed Image Super-Resolution and Restoration](https://arxiv.org/abs/2209.11345) by Marcos V. Conde, Ui-Jin Choi, Maxime Burchi, Radu Timofte. -1. **[SwitchTransformers](https://huggingface.co/docs/transformers/model_doc/switch_transformers)** (from Google) released with the paper [Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity](https://arxiv.org/abs/2101.03961) by William Fedus, Barret Zoph, Noam Shazeer. +1. **[SwitchTransformers](https://huggingface.co/docs/transformers/model_doc/switch_transformers)** (from Google) released with the paper [Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity](https://arxiv.org/abs/2101.03961) by William Fedus, Barret Zoph, Noam Shazeer. 1. **[T5](https://huggingface.co/docs/transformers/model_doc/t5)** (from Google AI) released with the paper [Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer](https://arxiv.org/abs/1910.10683) by Colin Raffel and Noam Shazeer and Adam Roberts and Katherine Lee and Sharan Narang and Michael Matena and Yanqi Zhou and Wei Li and Peter J. Liu. 1. **[T5v1.1](https://huggingface.co/docs/transformers/model_doc/t5v1.1)** (from Google AI) released with the paper [google-research/text-to-text-transfer-transformer](https://github.com/google-research/text-to-text-transfer-transformer/blob/main/released_checkpoints.md#t511) by Colin Raffel and Noam Shazeer and Adam Roberts and Katherine Lee and Sharan Narang and Michael Matena and Yanqi Zhou and Wei Li and Peter J. Liu. 1. **[Table Transformer](https://huggingface.co/docs/transformers/model_doc/table-transformer)** (from Microsoft Research) released with the paper [PubTables-1M: Towards Comprehensive Table Extraction From Unstructured Documents](https://arxiv.org/abs/2110.00061) by Brandon Smock, Rohith Pesala, Robin Abraham. diff --git a/docs/source/en/_toctree.yml b/docs/source/en/_toctree.yml index 7db5b6bd61f..db67a3e3c0c 100755 --- a/docs/source/en/_toctree.yml +++ b/docs/source/en/_toctree.yml @@ -491,6 +491,8 @@ title: Speech2Text - local: model_doc/speech_to_text_2 title: Speech2Text2 + - local: model_doc/speecht5 + title: SpeechT5 - local: model_doc/unispeech title: UniSpeech - local: model_doc/unispeech-sat diff --git a/docs/source/en/index.mdx b/docs/source/en/index.mdx index 87a2f09403e..a5ade5c6990 100644 --- a/docs/source/en/index.mdx +++ b/docs/source/en/index.mdx @@ -179,6 +179,7 @@ The documentation is organized into five sections: 1. **[SegFormer](model_doc/segformer)** (from NVIDIA) released with the paper [SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers](https://arxiv.org/abs/2105.15203) by Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, Ping Luo. 1. **[SEW](model_doc/sew)** (from ASAPP) released with the paper [Performance-Efficiency Trade-offs in Unsupervised Pre-training for Speech Recognition](https://arxiv.org/abs/2109.06870) by Felix Wu, Kwangyoun Kim, Jing Pan, Kyu Han, Kilian Q. Weinberger, Yoav Artzi. 1. **[SEW-D](model_doc/sew_d)** (from ASAPP) released with the paper [Performance-Efficiency Trade-offs in Unsupervised Pre-training for Speech Recognition](https://arxiv.org/abs/2109.06870) by Felix Wu, Kwangyoun Kim, Jing Pan, Kyu Han, Kilian Q. Weinberger, Yoav Artzi. +1. **[SpeechT5](model_doc/speecht5)** (from Microsoft Research) released with the paper [SpeechT5: Unified-Modal Encoder-Decoder Pre-Training for Spoken Language Processing](https://arxiv.org/abs/2110.07205) by Junyi Ao, Rui Wang, Long Zhou, Chengyi Wang, Shuo Ren, Yu Wu, Shujie Liu, Tom Ko, Qing Li, Yu Zhang, Zhihua Wei, Yao Qian, Jinyu Li, Furu Wei. 1. **[SpeechToTextTransformer](model_doc/speech_to_text)** (from Facebook), released together with the paper [fairseq S2T: Fast Speech-to-Text Modeling with fairseq](https://arxiv.org/abs/2010.05171) by Changhan Wang, Yun Tang, Xutai Ma, Anne Wu, Dmytro Okhonko, Juan Pino. 1. **[SpeechToTextTransformer2](model_doc/speech_to_text_2)** (from Facebook), released together with the paper [Large-Scale Self- and Semi-Supervised Learning for Speech Translation](https://arxiv.org/abs/2104.06678) by Changhan Wang, Anne Wu, Juan Pino, Alexei Baevski, Michael Auli, Alexis Conneau. 1. **[Splinter](model_doc/splinter)** (from Tel Aviv University), released together with the paper [Few-Shot Question Answering by Pretraining Span Selection](https://arxiv.org/abs/2101.00438) by Ori Ram, Yuval Kirstain, Jonathan Berant, Amir Globerson, Omer Levy. @@ -360,6 +361,7 @@ Flax), PyTorch, and/or TensorFlow. | Speech Encoder decoder | ❌ | ❌ | ✅ | ❌ | ✅ | | Speech2Text | ✅ | ❌ | ✅ | ✅ | ❌ | | Speech2Text2 | ✅ | ❌ | ❌ | ❌ | ❌ | +| SpeechT5 | ✅ | ❌ | ✅ | ❌ | ❌ | | Splinter | ✅ | ✅ | ✅ | ❌ | ❌ | | SqueezeBERT | ✅ | ✅ | ✅ | ❌ | ❌ | | Swin Transformer | ❌ | ❌ | ✅ | ✅ | ❌ | diff --git a/docs/source/en/main_classes/output.mdx b/docs/source/en/main_classes/output.mdx index 391955ce291..ced38976e84 100644 --- a/docs/source/en/main_classes/output.mdx +++ b/docs/source/en/main_classes/output.mdx @@ -136,6 +136,10 @@ documented on their corresponding model page. [[autodoc]] modeling_outputs.Seq2SeqQuestionAnsweringModelOutput +## Seq2SeqSpectrogramOutput + +[[autodoc]] modeling_outputs.Seq2SeqSpectrogramOutput + ## SemanticSegmenterOutput [[autodoc]] modeling_outputs.SemanticSegmenterOutput diff --git a/docs/source/en/model_doc/speecht5.mdx b/docs/source/en/model_doc/speecht5.mdx new file mode 100644 index 00000000000..848744b24d5 --- /dev/null +++ b/docs/source/en/model_doc/speecht5.mdx @@ -0,0 +1,81 @@ + + +# SpeechT5 + +## Overview + +The SpeechT5 model was proposed in [SpeechT5: Unified-Modal Encoder-Decoder Pre-Training for Spoken Language Processing](https://arxiv.org/abs/2110.07205) by Junyi Ao, Rui Wang, Long Zhou, Chengyi Wang, Shuo Ren, Yu Wu, Shujie Liu, Tom Ko, Qing Li, Yu Zhang, Zhihua Wei, Yao Qian, Jinyu Li, Furu Wei. + +The abstract from the paper is the following: + +*Motivated by the success of T5 (Text-To-Text Transfer Transformer) in pre-trained natural language processing models, we propose a unified-modal SpeechT5 framework that explores the encoder-decoder pre-training for self-supervised speech/text representation learning. The SpeechT5 framework consists of a shared encoder-decoder network and six modal-specific (speech/text) pre/post-nets. After preprocessing the input speech/text through the pre-nets, the shared encoder-decoder network models the sequence-to-sequence transformation, and then the post-nets generate the output in the speech/text modality based on the output of the decoder. Leveraging large-scale unlabeled speech and text data, we pre-train SpeechT5 to learn a unified-modal representation, hoping to improve the modeling capability for both speech and text. To align the textual and speech information into this unified semantic space, we propose a cross-modal vector quantization approach that randomly mixes up speech/text states with latent units as the interface between encoder and decoder. Extensive evaluations show the superiority of the proposed SpeechT5 framework on a wide variety of spoken language processing tasks, including automatic speech recognition, speech synthesis, speech translation, voice conversion, speech enhancement, and speaker identification.* + +This model was contributed by [Matthijs](https://huggingface.co/Matthijs). The original code can be found [here](https://github.com/microsoft/SpeechT5). + +## SpeechT5Config + +[[autodoc]] SpeechT5Config + +## SpeechT5HifiGanConfig + +[[autodoc]] SpeechT5HifiGanConfig + +## SpeechT5Tokenizer + +[[autodoc]] SpeechT5Tokenizer + - __call__ + - save_vocabulary + - decode + - batch_decode + +## SpeechT5FeatureExtractor + +[[autodoc]] SpeechT5FeatureExtractor + - __call__ + +## SpeechT5Processor + +[[autodoc]] SpeechT5Processor + - __call__ + - pad + - from_pretrained + - save_pretrained + - batch_decode + - decode + +## SpeechT5Model + +[[autodoc]] SpeechT5Model + - forward + +## SpeechT5ForSpeechToText + +[[autodoc]] SpeechT5ForSpeechToText + - forward + +## SpeechT5ForTextToSpeech + +[[autodoc]] SpeechT5ForTextToSpeech + - forward + - generate_speech + +## SpeechT5ForSpeechToSpeech + +[[autodoc]] SpeechT5ForSpeechToSpeech + - forward + - generate_speech + +## SpeechT5HifiGan + +[[autodoc]] SpeechT5HifiGan + - forward diff --git a/src/transformers/__init__.py b/src/transformers/__init__.py index e8cef7dfe4e..51989de777d 100755 --- a/src/transformers/__init__.py +++ b/src/transformers/__init__.py @@ -402,6 +402,12 @@ _import_structure = { "Speech2Text2Processor", "Speech2Text2Tokenizer", ], + "models.speecht5": [ + "SPEECHT5_PRETRAINED_CONFIG_ARCHIVE_MAP", + "SPEECHT5_PRETRAINED_HIFIGAN_CONFIG_ARCHIVE_MAP", + "SpeechT5Config", + "SpeechT5HifiGanConfig", + ], "models.splinter": ["SPLINTER_PRETRAINED_CONFIG_ARCHIVE_MAP", "SplinterConfig", "SplinterTokenizer"], "models.squeezebert": ["SQUEEZEBERT_PRETRAINED_CONFIG_ARCHIVE_MAP", "SqueezeBertConfig", "SqueezeBertTokenizer"], "models.swin": ["SWIN_PRETRAINED_CONFIG_ARCHIVE_MAP", "SwinConfig"], @@ -632,6 +638,7 @@ else: _import_structure["models.reformer"].append("ReformerTokenizer") _import_structure["models.rembert"].append("RemBertTokenizer") _import_structure["models.speech_to_text"].append("Speech2TextTokenizer") + _import_structure["models.speecht5"].append("SpeechT5Tokenizer") _import_structure["models.t5"].append("T5Tokenizer") _import_structure["models.xglm"].append("XGLMTokenizer") _import_structure["models.xlm_prophetnet"].append("XLMProphetNetTokenizer") @@ -734,6 +741,7 @@ else: _import_structure["models.audio_spectrogram_transformer"].append("ASTFeatureExtractor") _import_structure["models.mctct"].append("MCTCTFeatureExtractor") _import_structure["models.speech_to_text"].append("Speech2TextFeatureExtractor") + _import_structure["models.speecht5"].append("SpeechT5FeatureExtractor") # Tensorflow-text-specific objects try: @@ -772,6 +780,7 @@ except OptionalDependencyNotAvailable: ] else: _import_structure["models.speech_to_text"].append("Speech2TextProcessor") + _import_structure["models.speecht5"].append("SpeechT5Processor") # Vision-specific objects try: @@ -2193,6 +2202,17 @@ else: ] ) _import_structure["models.speech_to_text_2"].extend(["Speech2Text2ForCausalLM", "Speech2Text2PreTrainedModel"]) + _import_structure["models.speecht5"].extend( + [ + "SPEECHT5_PRETRAINED_MODEL_ARCHIVE_LIST", + "SpeechT5ForSpeechToSpeech", + "SpeechT5ForSpeechToText", + "SpeechT5ForTextToSpeech", + "SpeechT5HifiGan", + "SpeechT5Model", + "SpeechT5PreTrainedModel", + ] + ) _import_structure["models.splinter"].extend( [ "SPLINTER_PRETRAINED_MODEL_ARCHIVE_LIST", @@ -3842,6 +3862,12 @@ if TYPE_CHECKING: Speech2Text2Processor, Speech2Text2Tokenizer, ) + from .models.speecht5 import ( + SPEECHT5_PRETRAINED_CONFIG_ARCHIVE_MAP, + SPEECHT5_PRETRAINED_HIFIGAN_CONFIG_ARCHIVE_MAP, + SpeechT5Config, + SpeechT5HifiGanConfig, + ) from .models.splinter import SPLINTER_PRETRAINED_CONFIG_ARCHIVE_MAP, SplinterConfig, SplinterTokenizer from .models.squeezebert import SQUEEZEBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, SqueezeBertConfig, SqueezeBertTokenizer from .models.swin import SWIN_PRETRAINED_CONFIG_ARCHIVE_MAP, SwinConfig @@ -4054,6 +4080,7 @@ if TYPE_CHECKING: from .models.reformer import ReformerTokenizer from .models.rembert import RemBertTokenizer from .models.speech_to_text import Speech2TextTokenizer + from .models.speecht5 import SpeechT5Tokenizer from .models.t5 import T5Tokenizer from .models.xglm import XGLMTokenizer from .models.xlm_prophetnet import XLMProphetNetTokenizer @@ -4139,6 +4166,7 @@ if TYPE_CHECKING: from .models.audio_spectrogram_transformer import ASTFeatureExtractor from .models.mctct import MCTCTFeatureExtractor from .models.speech_to_text import Speech2TextFeatureExtractor + from .models.speecht5 import SpeechT5FeatureExtractor try: if not is_tensorflow_text_available(): @@ -4163,6 +4191,7 @@ if TYPE_CHECKING: from .utils.dummy_sentencepiece_and_speech_objects import * else: from .models.speech_to_text import Speech2TextProcessor + from .models.speecht5 import SpeechT5Processor try: if not is_vision_available(): @@ -5326,6 +5355,15 @@ if TYPE_CHECKING: Speech2TextPreTrainedModel, ) from .models.speech_to_text_2 import Speech2Text2ForCausalLM, Speech2Text2PreTrainedModel + from .models.speecht5 import ( + SPEECHT5_PRETRAINED_MODEL_ARCHIVE_LIST, + SpeechT5ForSpeechToSpeech, + SpeechT5ForSpeechToText, + SpeechT5ForTextToSpeech, + SpeechT5HifiGan, + SpeechT5Model, + SpeechT5PreTrainedModel, + ) from .models.splinter import ( SPLINTER_PRETRAINED_MODEL_ARCHIVE_LIST, SplinterForPreTraining, diff --git a/src/transformers/modeling_outputs.py b/src/transformers/modeling_outputs.py index d41e5804fa0..5c80a699056 100755 --- a/src/transformers/modeling_outputs.py +++ b/src/transformers/modeling_outputs.py @@ -1355,3 +1355,63 @@ class BaseModelOutputWithPoolingAndProjection(ModelOutput): hidden_states: Optional[Tuple[torch.FloatTensor]] = None attentions: Optional[Tuple[torch.FloatTensor]] = None projection_state: Optional[Tuple[torch.FloatTensor]] = None + + +@dataclass +class Seq2SeqSpectrogramOutput(ModelOutput): + """ + Base class for sequence-to-sequence spectrogram outputs. + + Args: + loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided): + Spectrogram generation loss. + spectrogram (`torch.FloatTensor` of shape `(batch_size, sequence_length, num_bins)`): + The predicted spectrogram. + past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`): + Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape + `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape + `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`. + + Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention + blocks) that can be used (see `past_key_values` input) to speed up sequential decoding. + decoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`): + Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, + + one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`. + + Hidden-states of the decoder at the output of each layer plus the initial embedding outputs. + decoder_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`): + Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length, + sequence_length)`. + + Attentions weights of the decoder, after the attention softmax, used to compute the weighted average in the + self-attention heads. + cross_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`): + Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length, + sequence_length)`. + + Attentions weights of the decoder's cross-attention layer, after the attention softmax, used to compute the + weighted average in the cross-attention heads. + encoder_last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*): + Sequence of hidden-states at the output of the last layer of the encoder of the model. + encoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`): + Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, + + one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`. + + Hidden-states of the encoder at the output of each layer plus the initial embedding outputs. + encoder_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`): + Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length, + sequence_length)`. + + Attentions weights of the encoder, after the attention softmax, used to compute the weighted average in the + self-attention heads. + """ + + loss: Optional[torch.FloatTensor] = None + spectrogram: torch.FloatTensor = None + past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None + decoder_hidden_states: Optional[Tuple[torch.FloatTensor]] = None + decoder_attentions: Optional[Tuple[torch.FloatTensor]] = None + cross_attentions: Optional[Tuple[torch.FloatTensor]] = None + encoder_last_hidden_state: Optional[torch.FloatTensor] = None + encoder_hidden_states: Optional[Tuple[torch.FloatTensor]] = None + encoder_attentions: Optional[Tuple[torch.FloatTensor]] = None diff --git a/src/transformers/models/__init__.py b/src/transformers/models/__init__.py index cf8e01baf5e..3beee2156d0 100644 --- a/src/transformers/models/__init__.py +++ b/src/transformers/models/__init__.py @@ -155,6 +155,7 @@ from . import ( speech_encoder_decoder, speech_to_text, speech_to_text_2, + speecht5, splinter, squeezebert, swin, diff --git a/src/transformers/models/auto/configuration_auto.py b/src/transformers/models/auto/configuration_auto.py index d153c10e829..ae1da07f15b 100755 --- a/src/transformers/models/auto/configuration_auto.py +++ b/src/transformers/models/auto/configuration_auto.py @@ -152,6 +152,7 @@ CONFIG_MAPPING_NAMES = OrderedDict( ("speech-encoder-decoder", "SpeechEncoderDecoderConfig"), ("speech_to_text", "Speech2TextConfig"), ("speech_to_text_2", "Speech2Text2Config"), + ("speecht5", "SpeechT5Config"), ("splinter", "SplinterConfig"), ("squeezebert", "SqueezeBertConfig"), ("swin", "SwinConfig"), @@ -310,6 +311,7 @@ CONFIG_ARCHIVE_MAP_MAPPING_NAMES = OrderedDict( ("sew-d", "SEW_D_PRETRAINED_CONFIG_ARCHIVE_MAP"), ("speech_to_text", "SPEECH_TO_TEXT_PRETRAINED_CONFIG_ARCHIVE_MAP"), ("speech_to_text_2", "SPEECH_TO_TEXT_2_PRETRAINED_CONFIG_ARCHIVE_MAP"), + ("speecht5", "SPEECHT5_PRETRAINED_CONFIG_ARCHIVE_MAP"), ("splinter", "SPLINTER_PRETRAINED_CONFIG_ARCHIVE_MAP"), ("squeezebert", "SQUEEZEBERT_PRETRAINED_CONFIG_ARCHIVE_MAP"), ("swin", "SWIN_PRETRAINED_CONFIG_ARCHIVE_MAP"), @@ -489,6 +491,7 @@ MODEL_NAMES_MAPPING = OrderedDict( ("speech-encoder-decoder", "Speech Encoder decoder"), ("speech_to_text", "Speech2Text"), ("speech_to_text_2", "Speech2Text2"), + ("speecht5", "SpeechT5"), ("splinter", "Splinter"), ("squeezebert", "SqueezeBERT"), ("swin", "Swin Transformer"), diff --git a/src/transformers/models/auto/feature_extraction_auto.py b/src/transformers/models/auto/feature_extraction_auto.py index 3726f9f238c..1f8c01d397d 100644 --- a/src/transformers/models/auto/feature_extraction_auto.py +++ b/src/transformers/models/auto/feature_extraction_auto.py @@ -76,6 +76,7 @@ FEATURE_EXTRACTOR_MAPPING_NAMES = OrderedDict( ("sew", "Wav2Vec2FeatureExtractor"), ("sew-d", "Wav2Vec2FeatureExtractor"), ("speech_to_text", "Speech2TextFeatureExtractor"), + ("speecht5", "SpeechT5FeatureExtractor"), ("swin", "ViTFeatureExtractor"), ("swinv2", "ViTFeatureExtractor"), ("table-transformer", "DetrFeatureExtractor"), diff --git a/src/transformers/models/auto/modeling_auto.py b/src/transformers/models/auto/modeling_auto.py index e23868fb6f2..1f8468e9a05 100755 --- a/src/transformers/models/auto/modeling_auto.py +++ b/src/transformers/models/auto/modeling_auto.py @@ -148,6 +148,7 @@ MODEL_MAPPING_NAMES = OrderedDict( ("sew", "SEWModel"), ("sew-d", "SEWDModel"), ("speech_to_text", "Speech2TextModel"), + ("speecht5", "SpeechT5Model"), ("splinter", "SplinterModel"), ("squeezebert", "SqueezeBertModel"), ("swin", "SwinModel"), @@ -591,6 +592,7 @@ MODEL_FOR_SPEECH_SEQ_2_SEQ_MAPPING_NAMES = OrderedDict( [ ("speech-encoder-decoder", "SpeechEncoderDecoderModel"), ("speech_to_text", "Speech2TextForConditionalGeneration"), + ("speecht5", "SpeechT5ForSpeechToText"), ("whisper", "WhisperForConditionalGeneration"), ] ) diff --git a/src/transformers/models/auto/processing_auto.py b/src/transformers/models/auto/processing_auto.py index e55e76aff6e..0b92dfaf791 100644 --- a/src/transformers/models/auto/processing_auto.py +++ b/src/transformers/models/auto/processing_auto.py @@ -61,6 +61,7 @@ PROCESSOR_MAPPING_NAMES = OrderedDict( ("sew-d", "Wav2Vec2Processor"), ("speech_to_text", "Speech2TextProcessor"), ("speech_to_text_2", "Speech2Text2Processor"), + ("speecht5", "SpeechT5Processor"), ("trocr", "TrOCRProcessor"), ("unispeech", "Wav2Vec2Processor"), ("unispeech-sat", "Wav2Vec2Processor"), diff --git a/src/transformers/models/auto/tokenization_auto.py b/src/transformers/models/auto/tokenization_auto.py index 7073221d744..cc91c11617f 100644 --- a/src/transformers/models/auto/tokenization_auto.py +++ b/src/transformers/models/auto/tokenization_auto.py @@ -264,6 +264,7 @@ else: ("roformer", ("RoFormerTokenizer", "RoFormerTokenizerFast" if is_tokenizers_available() else None)), ("speech_to_text", ("Speech2TextTokenizer" if is_sentencepiece_available() else None, None)), ("speech_to_text_2", ("Speech2Text2Tokenizer", None)), + ("speecht5", ("SpeechT5Tokenizer" if is_sentencepiece_available() else None, None)), ("splinter", ("SplinterTokenizer", "SplinterTokenizerFast")), ( "squeezebert", diff --git a/src/transformers/models/speecht5/__init__.py b/src/transformers/models/speecht5/__init__.py new file mode 100644 index 00000000000..1b173f33dfb --- /dev/null +++ b/src/transformers/models/speecht5/__init__.py @@ -0,0 +1,129 @@ +# flake8: noqa +# There's no way to ignore "F401 '...' imported but unused" warnings in this +# module, but to preserve other warnings. So, don't check this module at all. + +# Copyright 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_sentencepiece_available, + is_speech_available, + is_torch_available, +) + + +_import_structure = { + "configuration_speecht5": [ + "SPEECHT5_PRETRAINED_CONFIG_ARCHIVE_MAP", + "SPEECHT5_PRETRAINED_HIFIGAN_CONFIG_ARCHIVE_MAP", + "SpeechT5Config", + "SpeechT5HifiGanConfig", + ], +} + +try: + if not is_sentencepiece_available(): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + pass +else: + _import_structure["tokenization_speecht5"] = ["SpeechT5Tokenizer"] + +try: + if not is_speech_available(): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + pass +else: + _import_structure["feature_extraction_speecht5"] = ["SpeechT5FeatureExtractor"] + +try: + if not (is_speech_available() and is_sentencepiece_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + pass +else: + _import_structure["processing_speecht5"] = ["SpeechT5Processor"] + +try: + if not is_torch_available(): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + pass +else: + _import_structure["modeling_speecht5"] = [ + "SPEECHT5_PRETRAINED_MODEL_ARCHIVE_LIST", + "SpeechT5ForSpeechToText", + "SpeechT5ForSpeechToSpeech", + "SpeechT5ForTextToSpeech", + "SpeechT5Model", + "SpeechT5PreTrainedModel", + "SpeechT5HifiGan", + ] + +if TYPE_CHECKING: + from .configuration_speecht5 import ( + SPEECHT5_PRETRAINED_CONFIG_ARCHIVE_MAP, + SPEECHT5_PRETRAINED_HIFIGAN_CONFIG_ARCHIVE_MAP, + SpeechT5Config, + SpeechT5HifiGanConfig, + ) + + try: + if not is_sentencepiece_available(): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + pass + else: + from .tokenization_speecht5 import SpeechT5Tokenizer + + try: + if not is_speech_available(): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + pass + else: + from .feature_extraction_speecht5 import SpeechT5FeatureExtractor + + try: + if not (is_speech_available() and is_sentencepiece_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + pass + else: + from .processing_speecht5 import SpeechT5Processor + + try: + if not is_torch_available(): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + pass + else: + from .modeling_speecht5 import ( + SPEECHT5_PRETRAINED_MODEL_ARCHIVE_LIST, + SpeechT5ForSpeechToSpeech, + SpeechT5ForSpeechToText, + SpeechT5ForTextToSpeech, + SpeechT5HifiGan, + SpeechT5Model, + SpeechT5PreTrainedModel, + ) + +else: + import sys + + sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__) diff --git a/src/transformers/models/speecht5/configuration_speecht5.py b/src/transformers/models/speecht5/configuration_speecht5.py new file mode 100644 index 00000000000..c36d6e22faf --- /dev/null +++ b/src/transformers/models/speecht5/configuration_speecht5.py @@ -0,0 +1,412 @@ +# coding=utf-8 +# Copyright 2023 The Fairseq Authors, Microsoft Research, 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. +""" SpeechT5 model configuration""" + +import functools +import operator + +from ...configuration_utils import PretrainedConfig +from ...utils import logging + + +logger = logging.get_logger(__name__) + +SPEECHT5_PRETRAINED_CONFIG_ARCHIVE_MAP = { + "microsoft/speecht5_asr": "https://huggingface.co/microsoft/speecht5_asr/resolve/main/config.json", + "microsoft/speecht5_tts": "https://huggingface.co/microsoft/speecht5_tts/resolve/main/config.json", + "microsoft/speecht5_vc": "https://huggingface.co/microsoft/speecht5_vc/resolve/main/config.json", +} + +SPEECHT5_PRETRAINED_HIFIGAN_CONFIG_ARCHIVE_MAP = { + "microsoft/speecht5_hifigan": "https://huggingface.co/microsoft/speecht5_hifigan/resolve/main/config.json", +} + + +class SpeechT5Config(PretrainedConfig): + r""" + This is the configuration class to store the configuration of a [`SpeechT5Model`]. It is used to instantiate a + SpeechT5 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 SpeechT5 + [microsoft/speecht5_asr](https://huggingface.co/microsoft/speecht5_asr) architecture. + + Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the + documentation from [`PretrainedConfig`] for more information. + + Args: + vocab_size (`int`, *optional*, defaults to 81): + Vocabulary size of the SpeechT5 model. Defines the number of different tokens that can be represented by + the `inputs_ids` passed to the forward method of [`SpeechT5Model`]. + hidden_size (`int`, *optional*, defaults to 768): + Dimensionality of the encoder layers and the pooler layer. + encoder_layers (`int`, *optional*, defaults to 12): + Number of hidden layers in the Transformer encoder. + encoder_attention_heads (`int`, *optional*, defaults to 12): + Number of attention heads for each attention layer in the Transformer encoder. + encoder_ffn_dim (`int`, *optional*, defaults to 3072): + Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder. + encoder_layerdrop (`float`, *optional*, defaults to 0.1): + The LayerDrop probability for the encoder. See the [LayerDrop paper](see https://arxiv.org/abs/1909.11556) + for more details. + decoder_layers (`int`, *optional*, defaults to 6): + Number of hidden layers in the Transformer decoder. + decoder_attention_heads (`int`, *optional*, defaults to 12): + Number of attention heads for each attention layer in the Transformer decoder. + decoder_ffn_dim (`int`, *optional*, defaults to 3072): + Dimensionality of the "intermediate" (often named feed-forward) layer in the Transformer decoder. + decoder_layerdrop (`float`, *optional*, defaults to 0.1): + The LayerDrop probability for the decoder. See the [LayerDrop paper](see https://arxiv.org/abs/1909.11556) + for more details. + hidden_act (`str` or `function`, *optional*, defaults to `"gelu"`): + The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`, + `"relu"`, `"selu"` and `"gelu_new"` are supported. + positional_dropout (`float`, *optional*, defaults to 0.1): + The dropout probability for the text position encoding layers. + hidden_dropout (`float`, *optional*, defaults to 0.1): + The dropout probability for all fully connected layers in the embeddings, encoder, and pooler. + attention_dropout (`float`, *optional*, defaults to 0.1): + The dropout ratio for the attention probabilities. + activation_dropout (`float`, *optional*, defaults to 0.1): + The dropout ratio for activations inside the fully connected layer. + initializer_range (`float`, *optional*, defaults to 0.02): + The standard deviation of the truncated_normal_initializer for initializing all weight matrices. + layer_norm_eps (`float`, *optional*, defaults to 1e-5): + The epsilon used by the layer normalization layers. + scale_embedding (`bool`, *optional*, defaults to `False`): + Scale embeddings by diving by sqrt(d_model). + feat_extract_norm (`str`, *optional*, defaults to `"group"`): + The norm to be applied to 1D convolutional layers in the speech encoder pre-net. One of `"group"` for group + normalization of only the first 1D convolutional layer or `"layer"` for layer normalization of all 1D + convolutional layers. + feat_proj_dropout (`float`, *optional*, defaults to 0.0): + The dropout probability for output of the speech encoder pre-net. + feat_extract_activation (`str, `optional`, defaults to `"gelu"`): + The non-linear activation function (function or string) in the 1D convolutional layers of the feature + extractor. If string, `"gelu"`, `"relu"`, `"selu"` and `"gelu_new"` are supported. + conv_dim (`Tuple[int]` or `List[int]`, *optional*, defaults to `(512, 512, 512, 512, 512, 512, 512)`): + A tuple of integers defining the number of input and output channels of each 1D convolutional layer in the + speech encoder pre-net. The length of *conv_dim* defines the number of 1D convolutional layers. + conv_stride (`Tuple[int]` or `List[int]`, *optional*, defaults to `(5, 2, 2, 2, 2, 2, 2)`): + A tuple of integers defining the stride of each 1D convolutional layer in the speech encoder pre-net. The + length of *conv_stride* defines the number of convolutional layers and has to match the length of + *conv_dim*. + conv_kernel (`Tuple[int]` or `List[int]`, *optional*, defaults to `(10, 3, 3, 3, 3, 3, 3)`): + A tuple of integers defining the kernel size of each 1D convolutional layer in the speech encoder pre-net. + The length of *conv_kernel* defines the number of convolutional layers and has to match the length of + *conv_dim*. + conv_bias (`bool`, *optional*, defaults to `False`): + Whether the 1D convolutional layers have a bias. + num_conv_pos_embeddings (`int`, *optional*, defaults to 128): + Number of convolutional positional embeddings. Defines the kernel size of 1D convolutional positional + embeddings layer. + num_conv_pos_embedding_groups (`int`, *optional*, defaults to 16): + Number of groups of 1D convolutional positional embeddings layer. + apply_spec_augment (`bool`, *optional*, defaults to `True`): + Whether to apply *SpecAugment* data augmentation to the outputs of the speech encoder pre-net. For + reference see [SpecAugment: A Simple Data Augmentation Method for Automatic Speech + Recognition](https://arxiv.org/abs/1904.08779). + mask_time_prob (`float`, *optional*, defaults to 0.05): + Percentage (between 0 and 1) of all feature vectors along the time axis which will be masked. The masking + procecure generates ''mask_time_prob*len(time_axis)/mask_time_length'' independent masks over the axis. If + reasoning from the propability of each feature vector to be chosen as the start of the vector span to be + masked, *mask_time_prob* should be `prob_vector_start*mask_time_length`. Note that overlap may decrease the + actual percentage of masked vectors. This is only relevant if `apply_spec_augment is True`. + mask_time_length (`int`, *optional*, defaults to 10): + Length of vector span along the time axis. + mask_time_min_masks (`int`, *optional*, defaults to 2),: + The minimum number of masks of length `mask_feature_length` generated along the time axis, each time step, + irrespectively of `mask_feature_prob`. Only relevant if ''mask_time_prob*len(time_axis)/mask_time_length < + mask_time_min_masks'' + mask_feature_prob (`float`, *optional*, defaults to 0.0): + Percentage (between 0 and 1) of all feature vectors along the feature axis which will be masked. The + masking procecure generates ''mask_feature_prob*len(feature_axis)/mask_time_length'' independent masks over + the axis. If reasoning from the propability of each feature vector to be chosen as the start of the vector + span to be masked, *mask_feature_prob* should be `prob_vector_start*mask_feature_length`. Note that overlap + may decrease the actual percentage of masked vectors. This is only relevant if `apply_spec_augment is + True`. + mask_feature_length (`int`, *optional*, defaults to 10): + Length of vector span along the feature axis. + mask_feature_min_masks (`int`, *optional*, defaults to 0),: + The minimum number of masks of length `mask_feature_length` generated along the feature axis, each time + step, irrespectively of `mask_feature_prob`. Only relevant if + ''mask_feature_prob*len(feature_axis)/mask_feature_length < mask_feature_min_masks'' + num_mel_bins (`int`, *optional*, defaults to 80): + Number of mel features used per input features. Used by the speech decoder pre-net. Should correspond to + the value used in the [`SpeechT5Processor`] class. + speech_decoder_prenet_layers (`int`, *optional*, defaults to 2): + Number of layers in the speech decoder pre-net. + speech_decoder_prenet_units (`int`, *optional*, defaults to 256): + Dimensionality of the layers in the speech decoder pre-net. + speech_decoder_prenet_dropout (`float`, *optional*, defaults to 0.5): + The dropout probability for the speech decoder pre-net layers. + speaker_embedding_dim (`int`, *optional*, defaults to 512): + Dimensionality of the *XVector* embedding vectors. + speech_decoder_postnet_layers (`int`, *optional*, defaults to 5): + Number of layers in the speech decoder post-net. + speech_decoder_postnet_units (`int`, *optional*, defaults to 256): + Dimensionality of the layers in the speech decoder post-net. + speech_decoder_postnet_kernel (`int`, *optional*, defaults to 5): + Number of convolutional filter channels in the speech decoder post-net. + speech_decoder_postnet_dropout (`float`, *optional*, defaults to 0.5): + The dropout probability for the speech decoder post-net layers. + reduction_factor (`int`, *optional*, defaults to 2): + Spectrogram length reduction factor for the speech decoder post-net. + max_speech_positions (`int`, *optional*, defaults to 4000): + The maximum sequence length of speech features that this model might ever be used with. + max_text_positions (`int`, *optional*, defaults to 450): + The maximum sequence length of text features that this model might ever be used with. + encoder_max_relative_position (`int`, *optional*, defaults to 160): + Maximum distance for relative position embedding in the encoder. + decoder_max_relative_position (`int`, *optional*, defaults to 160): + Maximum distance for relative position embedding in the dencoder. + use_cache (`bool`, *optional*, defaults to `True`): + Whether or not the model should return the last key/values attentions (not used by all models). + + Example: + + ```python + >>> from transformers import SpeechT5Model, SpeechT5Config + + >>> # Initializing a "microsoft/speecht5_asr" style configuration + >>> configuration = SpeechT5Config() + + >>> # Initializing a model (with random weights) from the "microsoft/speecht5_asr" style configuration + >>> model = SpeechT5Model(configuration) + + >>> # Accessing the model configuration + >>> configuration = model.config + ```""" + model_type = "speecht5" + attribute_map = {"num_attention_heads": "encoder_attention_heads", "num_hidden_layers": "encoder_layers"} + + def __init__( + self, + vocab_size=81, + hidden_size=768, + encoder_layers=12, + encoder_attention_heads=12, + encoder_ffn_dim=3072, + encoder_layerdrop=0.1, + decoder_layers=6, + decoder_ffn_dim=3072, + decoder_attention_heads=12, + decoder_layerdrop=0.1, + hidden_act="gelu", + positional_dropout=0.1, + hidden_dropout=0.1, + attention_dropout=0.1, + activation_dropout=0.1, + initializer_range=0.02, + layer_norm_eps=1e-5, + scale_embedding=False, + feat_extract_norm="group", + feat_proj_dropout=0.0, + feat_extract_activation="gelu", + conv_dim=(512, 512, 512, 512, 512, 512, 512), + conv_stride=(5, 2, 2, 2, 2, 2, 2), + conv_kernel=(10, 3, 3, 3, 3, 2, 2), + conv_bias=False, + num_conv_pos_embeddings=128, + num_conv_pos_embedding_groups=16, + apply_spec_augment=True, + mask_time_prob=0.05, + mask_time_length=10, + mask_time_min_masks=2, + mask_feature_prob=0.0, + mask_feature_length=10, + mask_feature_min_masks=0, + pad_token_id=1, + bos_token_id=0, + eos_token_id=2, + decoder_start_token_id=2, + num_mel_bins=80, + speech_decoder_prenet_layers=2, + speech_decoder_prenet_units=256, + speech_decoder_prenet_dropout=0.5, + speaker_embedding_dim=512, + speech_decoder_postnet_layers=5, + speech_decoder_postnet_units=256, + speech_decoder_postnet_kernel=5, + speech_decoder_postnet_dropout=0.5, + reduction_factor=2, + max_speech_positions=4000, + max_text_positions=450, + encoder_max_relative_position=160, + decoder_max_relative_position=160, + use_cache=True, + is_encoder_decoder=True, + **kwargs + ): + self.vocab_size = vocab_size + self.hidden_size = hidden_size + self.encoder_layers = encoder_layers + self.encoder_ffn_dim = encoder_ffn_dim + self.encoder_attention_heads = encoder_attention_heads + self.encoder_layerdrop = encoder_layerdrop + self.decoder_layers = decoder_layers + self.decoder_ffn_dim = decoder_ffn_dim + self.decoder_attention_heads = decoder_attention_heads + self.decoder_layerdrop = decoder_layerdrop + self.hidden_act = hidden_act + self.positional_dropout = positional_dropout + self.hidden_dropout = hidden_dropout + self.attention_dropout = attention_dropout + self.activation_dropout = activation_dropout + self.initializer_range = initializer_range + self.layer_norm_eps = layer_norm_eps + self.scale_embedding = scale_embedding + + self.feat_extract_norm = feat_extract_norm + self.feat_proj_dropout = feat_proj_dropout + self.feat_extract_activation = feat_extract_activation + self.conv_dim = list(conv_dim) + self.conv_stride = list(conv_stride) + self.conv_kernel = list(conv_kernel) + self.conv_bias = conv_bias + self.num_conv_pos_embeddings = num_conv_pos_embeddings + self.num_conv_pos_embedding_groups = num_conv_pos_embedding_groups + self.num_feat_extract_layers = len(self.conv_dim) + + if ( + (len(self.conv_stride) != self.num_feat_extract_layers) + or (len(self.conv_kernel) != self.num_feat_extract_layers) + or (len(self.conv_dim) != self.num_feat_extract_layers) + ): + raise ValueError( + "Configuration for convolutional layers is incorrect. It is required that `len(config.conv_dim)` ==" + " `len(config.conv_stride)` == `len(config.conv_kernel)`, but is `len(config.conv_dim) =" + f" {len(self.conv_dim)}`, `len(config.conv_stride) = {len(self.conv_stride)}`," + f" `len(config.conv_kernel) = {len(self.conv_kernel)}`." + ) + + # fine-tuning config parameters for SpecAugment: https://arxiv.org/abs/1904.08779 + self.apply_spec_augment = apply_spec_augment + self.mask_time_prob = mask_time_prob + self.mask_time_length = mask_time_length + self.mask_time_min_masks = mask_time_min_masks + self.mask_feature_prob = mask_feature_prob + self.mask_feature_length = mask_feature_length + self.mask_feature_min_masks = mask_feature_min_masks + + self.num_mel_bins = num_mel_bins + self.speech_decoder_prenet_layers = speech_decoder_prenet_layers + self.speech_decoder_prenet_units = speech_decoder_prenet_units + self.speech_decoder_prenet_dropout = speech_decoder_prenet_dropout + self.speaker_embedding_dim = speaker_embedding_dim + + self.speech_decoder_postnet_layers = speech_decoder_postnet_layers + self.speech_decoder_postnet_units = speech_decoder_postnet_units + self.speech_decoder_postnet_kernel = speech_decoder_postnet_kernel + self.speech_decoder_postnet_dropout = speech_decoder_postnet_dropout + self.reduction_factor = reduction_factor + + self.max_speech_positions = max_speech_positions + self.max_text_positions = max_text_positions + self.encoder_max_relative_position = encoder_max_relative_position + self.decoder_max_relative_position = decoder_max_relative_position + self.use_cache = use_cache + self.is_encoder_decoder = is_encoder_decoder + + super().__init__( + pad_token_id=pad_token_id, + bos_token_id=bos_token_id, + eos_token_id=eos_token_id, + is_encoder_decoder=is_encoder_decoder, + decoder_start_token_id=decoder_start_token_id, + **kwargs, + ) + + def inputs_to_logits_ratio(self): + return functools.reduce(operator.mul, self.conv_stride, 1) + + +class SpeechT5HifiGanConfig(PretrainedConfig): + r""" + This is the configuration class to store the configuration of a [`SpeechT5HifiGanModel`]. It is used to instantiate + a SpeechT5 HiFi-GAN vocoder 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 SpeechT5 + [microsoft/speecht5_hifigan](https://huggingface.co/microsoft/speecht5_hifigan) architecture. + + Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the + documentation from [`PretrainedConfig`] for more information. + + Args: + model_in_dim (`int`, *optional*, defaults to 80): + The number of frequency bins in the input log-mel spectrogram. + sampling_rate (`int`, *optional*, defaults to 16000): + The sampling rate at which the output audio will be generated, expressed in hertz (Hz). + upsample_initial_channel (`int`, *optional*, defaults to 512): + The number of input channels into the upsampling network. + upsample_rates (`Tuple[int]` or `List[int]`, *optional*, defaults to `[4, 4, 4, 4]`): + A tuple of integers defining the stride of each 1D convolutional layer in the upsampling network. The + length of *upsample_rates* defines the number of convolutional layers and has to match the length of + *upsample_kernel_sizes*. + upsample_kernel_sizes (`Tuple[int]` or `List[int]`, *optional*, defaults to `[8, 8, 8, 8]`): + A tuple of integers defining the kernel size of each 1D convolutional layer in the upsampling network. The + length of *upsample_kernel_sizes* defines the number of convolutional layers and has to match the length of + *upsample_rates*. + resblock_kernel_sizes (`Tuple[int]` or `List[int]`, *optional*, defaults to `[3, 7, 11]`): + A tuple of integers defining the kernel sizes of the 1D convolutional layers in the multi-receptive field + fusion (MRF) module. + resblock_dilation_sizes (`Tuple[Tuple[int]]` or `List[List[int]]`, *optional*, defaults to `[[1, 3, 5], [1, 3, 5], [1, 3, 5]]`): + A nested tuple of integers defining the dilation rates of the dilated 1D convolutional layers in the + multi-receptive field fusion (MRF) module. + initializer_range (`float`, *optional*, defaults to 0.01): + The standard deviation of the truncated_normal_initializer for initializing all weight matrices. + leaky_relu_slope (`float`, *optional*, defaults to 0.1): + The angle of the negative slope used by the leaky ReLU activation. + normalize_before (`bool`, *optional*, defaults to `True`): + Whether or not to normalize the spectrogram before vocoding using the vocoder's learned mean and variance. + + Example: + + ```python + >>> from transformers import SpeechT5HifiGan, SpeechT5HifiGanConfig + + >>> # Initializing a "microsoft/speecht5_hifigan" style configuration + >>> configuration = SpeechT5HifiGanConfig() + + >>> # Initializing a model (with random weights) from the "microsoft/speecht5_hifigan" style configuration + >>> model = SpeechT5HifiGan(configuration) + + >>> # Accessing the model configuration + >>> configuration = model.config + ```""" + model_type = "hifigan" + + def __init__( + self, + model_in_dim=80, + sampling_rate=16000, + upsample_initial_channel=512, + upsample_rates=[4, 4, 4, 4], + upsample_kernel_sizes=[8, 8, 8, 8], + resblock_kernel_sizes=[3, 7, 11], + resblock_dilation_sizes=[[1, 3, 5], [1, 3, 5], [1, 3, 5]], + initializer_range=0.01, + leaky_relu_slope=0.1, + normalize_before=True, + **kwargs, + ): + self.model_in_dim = model_in_dim + self.sampling_rate = sampling_rate + self.upsample_initial_channel = upsample_initial_channel + self.upsample_rates = upsample_rates + self.upsample_kernel_sizes = upsample_kernel_sizes + self.resblock_kernel_sizes = resblock_kernel_sizes + self.resblock_dilation_sizes = resblock_dilation_sizes + self.initializer_range = initializer_range + self.leaky_relu_slope = leaky_relu_slope + self.normalize_before = normalize_before + super().__init__(**kwargs) diff --git a/src/transformers/models/speecht5/convert_hifigan.py b/src/transformers/models/speecht5/convert_hifigan.py new file mode 100644 index 00000000000..4d78bb73af3 --- /dev/null +++ b/src/transformers/models/speecht5/convert_hifigan.py @@ -0,0 +1,108 @@ +# 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 SpeechT5 HiFi-GAN checkpoint.""" + +import argparse + +import numpy as np +import torch + +from transformers import SpeechT5HifiGan, SpeechT5HifiGanConfig, logging + + +logging.set_verbosity_info() +logger = logging.get_logger("transformers.models.speecht5") + + +def load_weights(checkpoint, hf_model, config): + hf_model.apply_weight_norm() + + hf_model.conv_pre.weight_g.data = checkpoint["input_conv.weight_g"] + hf_model.conv_pre.weight_v.data = checkpoint["input_conv.weight_v"] + hf_model.conv_pre.bias.data = checkpoint["input_conv.bias"] + + for i in range(len(config.upsample_rates)): + hf_model.upsampler[i].weight_g.data = checkpoint[f"upsamples.{i}.1.weight_g"] + hf_model.upsampler[i].weight_v.data = checkpoint[f"upsamples.{i}.1.weight_v"] + hf_model.upsampler[i].bias.data = checkpoint[f"upsamples.{i}.1.bias"] + + for i in range(len(config.upsample_rates) * len(config.resblock_kernel_sizes)): + for j in range(len(config.resblock_dilation_sizes)): + hf_model.resblocks[i].convs1[j].weight_g.data = checkpoint[f"blocks.{i}.convs1.{j}.1.weight_g"] + hf_model.resblocks[i].convs1[j].weight_v.data = checkpoint[f"blocks.{i}.convs1.{j}.1.weight_v"] + hf_model.resblocks[i].convs1[j].bias.data = checkpoint[f"blocks.{i}.convs1.{j}.1.bias"] + + hf_model.resblocks[i].convs2[j].weight_g.data = checkpoint[f"blocks.{i}.convs2.{j}.1.weight_g"] + hf_model.resblocks[i].convs2[j].weight_v.data = checkpoint[f"blocks.{i}.convs2.{j}.1.weight_v"] + hf_model.resblocks[i].convs2[j].bias.data = checkpoint[f"blocks.{i}.convs2.{j}.1.bias"] + + hf_model.conv_post.weight_g.data = checkpoint["output_conv.1.weight_g"] + hf_model.conv_post.weight_v.data = checkpoint["output_conv.1.weight_v"] + hf_model.conv_post.bias.data = checkpoint["output_conv.1.bias"] + + hf_model.remove_weight_norm() + + +@torch.no_grad() +def convert_hifigan_checkpoint( + checkpoint_path, + stats_path, + pytorch_dump_folder_path, + config_path=None, + repo_id=None, +): + if config_path is not None: + config = SpeechT5HifiGanConfig.from_pretrained(config_path) + else: + config = SpeechT5HifiGanConfig() + + model = SpeechT5HifiGan(config) + + orig_checkpoint = torch.load(checkpoint_path) + load_weights(orig_checkpoint["model"]["generator"], model, config) + + stats = np.load(stats_path) + mean = stats[0].reshape(-1) + scale = stats[1].reshape(-1) + model.mean = torch.from_numpy(mean).float() + model.scale = torch.from_numpy(scale).float() + + model.save_pretrained(pytorch_dump_folder_path) + + if repo_id: + print("Pushing to the hub...") + model.push_to_hub(repo_id) + + +if __name__ == "__main__": + parser = argparse.ArgumentParser() + parser.add_argument("--checkpoint_path", required=True, default=None, type=str, help="Path to original checkpoint") + parser.add_argument("--stats_path", required=True, default=None, type=str, help="Path to stats.npy file") + 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_hifigan_checkpoint( + args.checkpoint_path, + args.stats_path, + args.pytorch_dump_folder_path, + args.config_path, + args.push_to_hub, + ) diff --git a/src/transformers/models/speecht5/convert_speecht5_original_pytorch_checkpoint_to_pytorch.py b/src/transformers/models/speecht5/convert_speecht5_original_pytorch_checkpoint_to_pytorch.py new file mode 100644 index 00000000000..de1e81fcbd7 --- /dev/null +++ b/src/transformers/models/speecht5/convert_speecht5_original_pytorch_checkpoint_to_pytorch.py @@ -0,0 +1,402 @@ +# 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 SpeechT5 checkpoint.""" + +import argparse + +import torch + +from transformers import ( + SpeechT5Config, + SpeechT5FeatureExtractor, + SpeechT5ForSpeechToSpeech, + SpeechT5ForSpeechToText, + SpeechT5ForTextToSpeech, + SpeechT5Processor, + SpeechT5Tokenizer, + logging, +) +from transformers.tokenization_utils import AddedToken + + +logging.set_verbosity_info() +logger = logging.get_logger("transformers.models.speecht5") + +MAPPING_SPEECH_ENCODER_PRENET = { + "speech_encoder_prenet.layer_norm": "speecht5.encoder.prenet.feature_projection.layer_norm", + "speech_encoder_prenet.post_extract_proj": "speecht5.encoder.prenet.feature_projection.projection", + "speech_encoder_prenet.pos_conv.0": "speecht5.encoder.prenet.pos_conv_embed.conv", + "speech_encoder_prenet.mask_emb": "speecht5.encoder.prenet.masked_spec_embed", +} +MAPPING_TEXT_ENCODER_PRENET = { + "text_encoder_prenet.encoder_prenet.0": "speecht5.encoder.prenet.embed_tokens", + "text_encoder_prenet.encoder_prenet.1.alpha": "speecht5.encoder.prenet.encode_positions.alpha", +} +MAPPING_SPEECH_DECODER_PRENET = { + "speech_decoder_prenet.decoder_prenet.0.0.prenet.0.0": "speecht5.decoder.prenet.layers.0", + "speech_decoder_prenet.decoder_prenet.0.0.prenet.1.0": "speecht5.decoder.prenet.layers.1", + "speech_decoder_prenet.decoder_prenet.0.1": "speecht5.decoder.prenet.final_layer", + "speech_decoder_prenet.decoder_prenet.1.alpha": "speecht5.decoder.prenet.encode_positions.alpha", + "speech_decoder_prenet.spkembs_layer.0": "speecht5.decoder.prenet.speaker_embeds_layer", +} +MAPPING_SPEECH_DECODER_POSTNET = { + "speech_decoder_postnet.feat_out": "speech_decoder_postnet.feat_out", + "speech_decoder_postnet.prob_out": "speech_decoder_postnet.prob_out", + "speech_decoder_postnet.postnet.postnet.0.0": "speech_decoder_postnet.layers.0.conv", + "speech_decoder_postnet.postnet.postnet.0.1": "speech_decoder_postnet.layers.0.batch_norm", + "speech_decoder_postnet.postnet.postnet.1.0": "speech_decoder_postnet.layers.1.conv", + "speech_decoder_postnet.postnet.postnet.1.1": "speech_decoder_postnet.layers.1.batch_norm", + "speech_decoder_postnet.postnet.postnet.2.0": "speech_decoder_postnet.layers.2.conv", + "speech_decoder_postnet.postnet.postnet.2.1": "speech_decoder_postnet.layers.2.batch_norm", + "speech_decoder_postnet.postnet.postnet.3.0": "speech_decoder_postnet.layers.3.conv", + "speech_decoder_postnet.postnet.postnet.3.1": "speech_decoder_postnet.layers.3.batch_norm", + "speech_decoder_postnet.postnet.postnet.4.0": "speech_decoder_postnet.layers.4.conv", + "speech_decoder_postnet.postnet.postnet.4.1": "speech_decoder_postnet.layers.4.batch_norm", +} +MAPPING_TEXT_DECODER_PRENET = { + "text_decoder_prenet.embed_tokens": "speecht5.decoder.prenet.embed_tokens", +} +MAPPING_TEXT_DECODER_POSTNET = { + "text_decoder_postnet.output_projection": "text_decoder_postnet.lm_head", +} +MAPPING_ENCODER = { + "encoder.layers.*.self_attn.k_proj": "speecht5.encoder.wrapped_encoder.layers.*.attention.k_proj", + "encoder.layers.*.self_attn.v_proj": "speecht5.encoder.wrapped_encoder.layers.*.attention.v_proj", + "encoder.layers.*.self_attn.q_proj": "speecht5.encoder.wrapped_encoder.layers.*.attention.q_proj", + "encoder.layers.*.self_attn.out_proj": "speecht5.encoder.wrapped_encoder.layers.*.attention.out_proj", + "encoder.layers.*.self_attn_layer_norm": "speecht5.encoder.wrapped_encoder.layers.*.layer_norm", + "encoder.layers.*.fc1": "speecht5.encoder.wrapped_encoder.layers.*.feed_forward.intermediate_dense", + "encoder.layers.*.fc2": "speecht5.encoder.wrapped_encoder.layers.*.feed_forward.output_dense", + "encoder.layers.*.final_layer_norm": "speecht5.encoder.wrapped_encoder.layers.*.final_layer_norm", + "encoder.layer_norm": "speecht5.encoder.wrapped_encoder.layer_norm", + "encoder.pos_emb.pe_k": "speecht5.encoder.wrapped_encoder.embed_positions.pe_k", +} +MAPPING_DECODER = { + "decoder.layers.*.self_attn.k_proj": "speecht5.decoder.wrapped_decoder.layers.*.self_attn.k_proj", + "decoder.layers.*.self_attn.v_proj": "speecht5.decoder.wrapped_decoder.layers.*.self_attn.v_proj", + "decoder.layers.*.self_attn.q_proj": "speecht5.decoder.wrapped_decoder.layers.*.self_attn.q_proj", + "decoder.layers.*.self_attn.out_proj": "speecht5.decoder.wrapped_decoder.layers.*.self_attn.out_proj", + "decoder.layers.*.self_attn_layer_norm": "speecht5.decoder.wrapped_decoder.layers.*.self_attn_layer_norm", + "decoder.layers.*.encoder_attn.k_proj": "speecht5.decoder.wrapped_decoder.layers.*.encoder_attn.k_proj", + "decoder.layers.*.encoder_attn.v_proj": "speecht5.decoder.wrapped_decoder.layers.*.encoder_attn.v_proj", + "decoder.layers.*.encoder_attn.q_proj": "speecht5.decoder.wrapped_decoder.layers.*.encoder_attn.q_proj", + "decoder.layers.*.encoder_attn.out_proj": "speecht5.decoder.wrapped_decoder.layers.*.encoder_attn.out_proj", + "decoder.layers.*.encoder_attn_layer_norm": "speecht5.decoder.wrapped_decoder.layers.*.encoder_attn_layer_norm", + "decoder.layers.*.fc1": "speecht5.decoder.wrapped_decoder.layers.*.feed_forward.intermediate_dense", + "decoder.layers.*.fc2": "speecht5.decoder.wrapped_decoder.layers.*.feed_forward.output_dense", + "decoder.layers.*.final_layer_norm": "speecht5.decoder.wrapped_decoder.layers.*.final_layer_norm", +} +MAPPING_S2T = { + **MAPPING_SPEECH_ENCODER_PRENET, + **MAPPING_ENCODER, + **MAPPING_DECODER, + **MAPPING_TEXT_DECODER_PRENET, + **MAPPING_TEXT_DECODER_POSTNET, +} +MAPPING_T2S = { + **MAPPING_TEXT_ENCODER_PRENET, + **MAPPING_ENCODER, + **MAPPING_DECODER, + **MAPPING_SPEECH_DECODER_PRENET, + **MAPPING_SPEECH_DECODER_POSTNET, +} +MAPPING_S2S = { + **MAPPING_SPEECH_ENCODER_PRENET, + **MAPPING_ENCODER, + **MAPPING_DECODER, + **MAPPING_SPEECH_DECODER_PRENET, + **MAPPING_SPEECH_DECODER_POSTNET, +} +TOP_LEVEL_KEYS = [] +IGNORE_KEYS = [ + "encoder.version", + "encoder.layers.*.norm_k.weight", + "encoder.layers.*.norm_k.bias", + "decoder.version", + "decoder.layers.*.norm_k.weight", + "decoder.layers.*.norm_k.bias", + "decoder.pos_emb.pe_k", + "speech_encoder_prenet.embed_positions._float_tensor", + "text_decoder_prenet.embed_positions._float_tensor", +] +IGNORE_KEYS_S2T = IGNORE_KEYS + [ + "encoder.proj", + "text_encoder_prenet.*", + "speech_decoder_prenet.*", + "speech_decoder_postnet.*", +] +IGNORE_KEYS_T2S = IGNORE_KEYS + [ + "encoder.proj", + "speech_encoder_prenet.*", + "text_decoder_prenet.*", + "text_decoder_postnet.*", +] +IGNORE_KEYS_S2S = IGNORE_KEYS + [ + "encoder.proj", + "text_encoder_prenet.*", + "text_decoder_prenet.*", + "text_decoder_postnet.*", +] + + +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 + 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(fairseq_dict, hf_model, task): + unused_weights = [] + + if task == "s2t": + feature_encoder = hf_model.speecht5.encoder.prenet.feature_encoder + MAPPING = MAPPING_S2T + IGNORE_KEYS = IGNORE_KEYS_S2T + elif task == "t2s": + feature_encoder = None + MAPPING = MAPPING_T2S + IGNORE_KEYS = IGNORE_KEYS_T2S + elif task == "s2s": + feature_encoder = hf_model.speecht5.encoder.prenet.feature_encoder + MAPPING = MAPPING_S2S + IGNORE_KEYS = IGNORE_KEYS_S2S + else: + raise ValueError(f"Unsupported task: {task}") + + for name, value in fairseq_dict.items(): + if should_ignore(name, IGNORE_KEYS): + logger.info(f"{name} was ignored") + continue + + is_used = False + if "conv_layers" in name: + load_conv_layer( + name, + value, + feature_encoder, + unused_weights, + hf_model.config.feat_extract_norm == "group", + ) + is_used = True + else: + for key, mapped_key in MAPPING.items(): + # mapped_key = "speecht5." + mapped_key if mapped_key not in TOP_LEVEL_KEYS else mapped_key + + if "*" in key: + prefix, suffix = key.split(".*.") + if prefix in name and suffix in name: + key = suffix + + # if key in name or key.split("w2v_model.")[-1] == name.split(".")[0]: + if key in name: + is_used = True + if "*" in mapped_key: + layer_index = name.split(key)[0].split(".")[-2] + mapped_key = mapped_key.replace("*", layer_index) + if "weight_g" in name: + weight_type = "weight_g" + elif "weight_v" in name: + weight_type = "weight_v" + elif "bias" in name: + weight_type = "bias" + elif "weight" in name: + 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}") + + +def load_conv_layer(full_name, value, feature_extractor, unused_weights, use_group_norm): + name = full_name.split("conv_layers.")[-1] + items = name.split(".") + layer_id = int(items[0]) + type_id = int(items[1]) + + if type_id == 0: + if "bias" in name: + if value.shape != feature_extractor.conv_layers[layer_id].conv.bias.data.shape: + raise ValueError( + f"{full_name} has size {value.shape}, but" + f" {feature_extractor.conv_layers[layer_id].conv.bias.data.shape} was found." + ) + feature_extractor.conv_layers[layer_id].conv.bias.data = value + logger.info(f"Feat extract conv layer {layer_id} was initialized from {full_name}.") + elif "weight" in name: + if value.shape != feature_extractor.conv_layers[layer_id].conv.weight.data.shape: + raise ValueError( + f"{full_name} has size {value.shape}, but" + f" {feature_extractor.conv_layers[layer_id].conv.weight.data.shape} was found." + ) + feature_extractor.conv_layers[layer_id].conv.weight.data = value + logger.info(f"Feat extract conv layer {layer_id} was initialized from {full_name}.") + elif (type_id == 2 and not use_group_norm) or (type_id == 2 and layer_id == 0 and use_group_norm): + if "bias" in name: + if value.shape != feature_extractor.conv_layers[layer_id].layer_norm.bias.data.shape: + raise ValueError( + f"{full_name} has size {value.shape}, but" + f" {feature_extractor.conv_layers[layer_id].layer_norm.bias.data.shape} was found." + ) + feature_extractor.conv_layers[layer_id].layer_norm.bias.data = value + logger.info(f"Feat extract layer norm weight of layer {layer_id} was initialized from {full_name}.") + elif "weight" in name: + if value.shape != feature_extractor.conv_layers[layer_id].layer_norm.weight.data.shape: + raise ValueError( + f"{full_name} has size {value.shape}, but" + f" {feature_extractor.conv_layers[layer_id].layer_norm.weight.data.shape} was found." + ) + feature_extractor.conv_layers[layer_id].layer_norm.weight.data = value + logger.info(f"Feat extract layer norm weight of layer {layer_id} was initialized from {full_name}.") + else: + unused_weights.append(full_name) + + +@torch.no_grad() +def convert_speecht5_checkpoint( + task, + checkpoint_path, + pytorch_dump_folder_path, + config_path=None, + vocab_path=None, + repo_id=None, +): + """ + Copy/paste/tweak model's weights to transformers design. + """ + if config_path is not None: + config = SpeechT5Config.from_pretrained(config_path) + else: + config = SpeechT5Config() + + if task == "s2t": + config.max_length = config.max_text_positions + model = SpeechT5ForSpeechToText(config) + elif task == "t2s": + config.max_speech_positions = 1876 + config.max_text_positions = 600 + config.max_length = config.max_speech_positions + model = SpeechT5ForTextToSpeech(config) + elif task == "s2s": + config.max_speech_positions = 1876 + config.max_length = config.max_speech_positions + model = SpeechT5ForSpeechToSpeech(config) + else: + raise ValueError(f"Unknown task name: {task}") + + if vocab_path: + tokenizer = SpeechT5Tokenizer(vocab_path, model_max_length=config.max_text_positions) + + if task == "pretrain": + # Mask token behaves like a normal word, i.e. include the space before it + mask_token = AddedToken("", lstrip=True, rstrip=False) + tokenizer.mask_token = mask_token + tokenizer.add_special_tokens({"mask_token": mask_token}) + tokenizer.add_tokens([""]) + + feature_extractor = SpeechT5FeatureExtractor() + processor = SpeechT5Processor(tokenizer=tokenizer, feature_extractor=feature_extractor) + processor.save_pretrained(pytorch_dump_folder_path) + + fairseq_checkpoint = torch.load(checkpoint_path) + recursively_load_weights(fairseq_checkpoint["model"], model, task) + + model.save_pretrained(pytorch_dump_folder_path) + + if repo_id: + print("Pushing to the hub...") + processor.push_to_hub(repo_id) + model.push_to_hub(repo_id) + + +if __name__ == "__main__": + parser = argparse.ArgumentParser() + parser.add_argument( + "--task", + default="s2t", + type=str, + help="Type of the SpeechT5 model you'd like to convert. Should be one of 's2t', 't2s', 's2s'.", + ) + parser.add_argument("--checkpoint_path", required=True, default=None, type=str, help="Path to fairseq checkpoint") + parser.add_argument("--vocab_path", default=None, type=str, help="Path to SentencePiece model") + 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_speecht5_checkpoint( + args.task, + args.checkpoint_path, + args.pytorch_dump_folder_path, + args.config_path, + args.vocab_path, + args.push_to_hub, + ) diff --git a/src/transformers/models/speecht5/feature_extraction_speecht5.py b/src/transformers/models/speecht5/feature_extraction_speecht5.py new file mode 100644 index 00000000000..4f08b1cf577 --- /dev/null +++ b/src/transformers/models/speecht5/feature_extraction_speecht5.py @@ -0,0 +1,450 @@ +# 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 SpeechT5.""" + +from typing import List, Optional, Union + +import numpy as np +import torch +import torchaudio + +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 SpeechT5FeatureExtractor(SequenceFeatureExtractor): + r""" + Constructs a SpeechT5 feature extractor. + + This class can pre-process a raw speech signal by (optionally) normalizing to zero-mean unit-variance, for use by + the SpeechT5 speech encoder prenet. + + This class can also extract log-mel filter bank features from raw speech, for use by the SpeechT5 speech decoder + prenet. + + 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. + + Args: + feature_size (`int`, *optional*, defaults to 1): + The feature dimension of the extracted features. + sampling_rate (`int`, *optional*, defaults to 16000): + The sampling rate at which the audio files 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. + do_normalize (`bool`, *optional*, defaults to `False`): + Whether or not to zero-mean unit-variance normalize the input. Normalizing can help to significantly + improve the performance for some models. + num_mel_bins (`int`, *optional*, defaults to 80): + The number of mel-frequency bins in the extracted spectrogram features. + hop_length (`int`, *optional*, defaults to 16): + Number of ms between windows. Otherwise referred to as "shift" in many papers. + win_length (`int`, *optional*, defaults to 64): + Number of ms per window. + win_function (`str`, *optional*, defaults to `"hann_window"`): + Name for the window function used for windowing, must be accessible via `torch.{win_function}` + frame_signal_scale (`float`, *optional*, defaults to 1.0): + Constant multiplied in creating the frames before applying DFT. + fmin (`float`, *optional*, defaults to 80): + Minimum mel frequency in Hz. + fmax (`float`, *optional*, defaults to 7600): + Maximum mel frequency in Hz. + mel_floor (`float`, *optional*, defaults to 1e-10): + Minimum value of mel frequency banks. + reduction_factor (`int`, *optional*, defaults to 2): + Spectrogram length reduction factor. + return_attention_mask (`bool`, *optional*, defaults to `True`): + Whether or not [`~SpeechT5FeatureExtractor.__call__`] should return `attention_mask`. + """ + + model_input_names = ["input_values", "attention_mask"] + + def __init__( + self, + feature_size: int = 1, + sampling_rate: int = 16000, + padding_value: float = 0.0, + do_normalize: bool = False, + num_mel_bins: int = 80, + hop_length: int = 16, + win_length: int = 64, + win_function: str = "hann_window", + frame_signal_scale: float = 1.0, + fmin: float = 80, + fmax: float = 7600, + mel_floor: float = 1e-10, + reduction_factor: int = 2, + return_attention_mask: bool = True, + **kwargs + ): + super().__init__(feature_size=feature_size, sampling_rate=sampling_rate, padding_value=padding_value, **kwargs) + self.do_normalize = do_normalize + self.return_attention_mask = return_attention_mask + + self.num_mel_bins = num_mel_bins + self.hop_length = hop_length + self.win_length = win_length + self.win_function = win_function + self.frame_signal_scale = frame_signal_scale + self.fmin = fmin + self.fmax = fmax + self.mel_floor = mel_floor + self.reduction_factor = reduction_factor + + self.sample_size = win_length * sampling_rate // 1000 + self.sample_stride = hop_length * sampling_rate // 1000 + + self.n_fft = 2 ** int(np.ceil(np.log2(self.sample_size))) + self.n_freqs = (self.n_fft // 2) + 1 + + @staticmethod + # Copied from transformers.models.wav2vec2.feature_extraction_wav2vec2.Wav2Vec2FeatureExtractor.zero_mean_unit_var_norm + def zero_mean_unit_var_norm( + input_values: List[np.ndarray], attention_mask: List[np.ndarray], padding_value: float = 0.0 + ) -> List[np.ndarray]: + """ + Every array in the list is normalized to have zero mean and unit variance + """ + if attention_mask is not None: + attention_mask = np.array(attention_mask, np.int32) + normed_input_values = [] + + for vector, length in zip(input_values, attention_mask.sum(-1)): + normed_slice = (vector - vector[:length].mean()) / np.sqrt(vector[:length].var() + 1e-7) + if length < normed_slice.shape[0]: + normed_slice[length:] = padding_value + + normed_input_values.append(normed_slice) + else: + normed_input_values = [(x - x.mean()) / np.sqrt(x.var() + 1e-7) for x in input_values] + + return normed_input_values + + @staticmethod + def _center_pad(one_waveform, n_fft, pad_mode): + padding = [(int(n_fft // 2), int(n_fft // 2))] + return np.pad(one_waveform, padding, mode=pad_mode) + + @staticmethod + # Copied from transformers.models.mctct.feature_extraction_mctct.MCTCTFeatureExtractor._num_frames_calc + def _num_frames_calc(in_size, frame_size, frame_stride): + return int(1 + np.floor((in_size - frame_size) * 1 / frame_stride)) + + @staticmethod + # Copied from transformers.models.mctct.feature_extraction_mctct.MCTCTFeatureExtractor._frame_signal + def _frame_signal(one_waveform, n_frames, frame_signal_scale, window_length, sample_stride): + scale = frame_signal_scale + frames = np.zeros(n_frames * window_length) + for frame_idx in range(n_frames): + start = frame_idx * window_length + end = (frame_idx + 1) * window_length + wave_start = frame_idx * sample_stride + wave_end = frame_idx * sample_stride + window_length + frames[start:end] = scale * one_waveform[wave_start:wave_end] + + return frames + + @staticmethod + # Copied from transformers.models.mctct.feature_extraction_mctct.MCTCTFeatureExtractor._windowing + def _windowing(frames, window_length, window): + if frames.size % window_length != 0: + raise ValueError( + f"`frames` is supposed to have length divisble by `window_length`, but is {frames.size} with" + f" window_length={window_length}." + ) + + shaped = frames.reshape(-1, window_length) + shaped = window * shaped + return shaped + + @staticmethod + # Copied from transformers.models.mctct.feature_extraction_mctct.MCTCTFeatureExtractor._dft + def _dft(frames, K, n_frames, n_samples, n_fft): + dft = np.zeros([n_frames, K]) + + for frame in range(n_frames): + begin = frame * n_samples + + inwards_buffer = frames[begin : begin + n_samples] + inwards_buffer = np.pad(inwards_buffer, (0, n_fft - n_samples), "constant") + out = np.fft.rfft(inwards_buffer) + + dft[frame] = np.abs(out[:K]) + + return dft + + def _extract_fbank_features( + self, + one_waveform: np.ndarray, + ) -> np.ndarray: + """ + Extracts log-mel filterbank features for one waveform vector (unbatched). Adapted from Flashlight's C++ MFSC + code and librosa. + """ + one_waveform = self._center_pad(one_waveform, self.n_fft, "reflect") + + n_frames = self._num_frames_calc(one_waveform.size, self.sample_size, self.sample_stride) + + frames = self._frame_signal( + one_waveform, n_frames, self.frame_signal_scale, self.sample_size, self.sample_stride + ) + + window = getattr(torch, self.win_function)(window_length=self.sample_size, periodic=True) + window = window.numpy() + + frames = self._windowing(frames, self.sample_size, window) + + dft_out = self._dft(frames.flatten(), self.n_freqs, n_frames, self.sample_size, self.n_fft) + + fbanks = torchaudio.functional.melscale_fbanks( + n_freqs=self.n_freqs, + f_min=self.fmin, + f_max=self.fmax, + n_mels=self.num_mel_bins, + sample_rate=self.sampling_rate, + norm="slaney", + mel_scale="slaney", + ) + fbanks = fbanks.numpy() + + return np.log10(np.maximum(self.mel_floor, np.dot(dft_out, fbanks))) + + def _reduce(self, inputs): + reduced = [] + for i in range(len(inputs)): + reduced.append(inputs[i][self.reduction_factor - 1 :: self.reduction_factor]) + return reduced + + def __call__( + self, + audio: Optional[Union[np.ndarray, List[float], List[np.ndarray], List[List[float]]]] = None, + audio_target: Optional[Union[np.ndarray, List[float], List[np.ndarray], List[List[float]]]] = None, + padding: Union[bool, str, PaddingStrategy] = False, + max_length: Optional[int] = None, + truncation: bool = False, + pad_to_multiple_of: Optional[int] = None, + return_attention_mask: Optional[bool] = None, + return_tensors: Optional[Union[str, TensorType]] = None, + sampling_rate: Optional[int] = None, + **kwargs + ) -> BatchFeature: + """ + Main method to featurize and prepare for the model one or several sequence(s). + + Pass in a value for `audio` to extract waveform features. Pass in a value for `audio_target` to extract log-mel + spectrogram features. + + Args: + audio (`np.ndarray`, `List[float]`, `List[np.ndarray]`, `List[List[float]]`, *optional*): + 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. This outputs waveform features. + audio_target (`np.ndarray`, `List[float]`, `List[np.ndarray]`, `List[List[float]]`, *optional*): + The sequence or batch of sequences to be processed as targets. 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. This outputs log-mel + spectrogram features. + padding (`bool`, `str` or [`~utils.PaddingStrategy`], *optional*, defaults to `False`): + 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). + max_length (`int`, *optional*): + Maximum length of the returned list and optionally padding length (see above). + truncation (`bool`): + Activates truncation to cut input sequences longer than *max_length* to *max_length*. + pad_to_multiple_of (`int`, *optional*): + If set will pad the sequence to a multiple of the provided value. + + This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability + `>= 7.5` (Volta), or on TPUs which benefit from having sequence lengths be a multiple of 128. + return_attention_mask (`bool`, *optional*): + Whether to return the attention mask. If left to the default, will return the attention mask according + to the specific feature_extractor's default. + + [What are attention masks?](../glossary#attention-mask) + + 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` or `audio_target` input was sampled. It is strongly recommended + to pass `sampling_rate` at the forward call to prevent silent errors. + """ + if audio is None and audio_target is None: + raise ValueError("You must provide either `audio` or `audio_target` values.") + + 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 audio is not None: + inputs = self._process_audio( + audio, + False, + padding, + max_length, + truncation, + pad_to_multiple_of, + return_attention_mask, + return_tensors, + **kwargs, + ) + else: + inputs = None + + if audio_target is not None: + inputs_target = self._process_audio( + audio_target, + True, + padding, + max_length, + truncation, + pad_to_multiple_of, + return_attention_mask, + return_tensors, + **kwargs, + ) + + if inputs is None: + return inputs_target + else: + inputs["labels"] = inputs_target["input_values"] + inputs["stop_labels"] = inputs_target["stop_labels"] + decoder_attention_mask = inputs_target.get("attention_mask") + if decoder_attention_mask is not None: + inputs["decoder_attention_mask"] = decoder_attention_mask + + return inputs + + def _process_audio( + self, + speech: Union[np.ndarray, List[float], List[np.ndarray], List[List[float]]], + is_target: bool = False, + padding: Union[bool, str, PaddingStrategy] = False, + max_length: Optional[int] = None, + truncation: bool = False, + pad_to_multiple_of: Optional[int] = None, + return_attention_mask: Optional[bool] = None, + return_tensors: Optional[Union[str, TensorType]] = None, + **kwargs + ) -> BatchFeature: + is_batched = bool( + isinstance(speech, (list, tuple)) + and (isinstance(speech[0], np.ndarray) or isinstance(speech[0], (tuple, list))) + ) + + if is_batched: + speech = [np.asarray(speech, dtype=np.float32) for speech in speech] + elif not is_batched and not isinstance(speech, np.ndarray): + speech = np.asarray(speech, dtype=np.float32) + elif isinstance(speech, np.ndarray) and speech.dtype is np.dtype(np.float64): + speech = speech.astype(np.float32) + + # always return batch + if not is_batched: + speech = [speech] + + # needed to make pad() work on spectrogram inputs + feature_size_hack = self.feature_size + + # convert into correct format for padding + if is_target: + features = [self._extract_fbank_features(waveform) for waveform in speech] + fbank_sizes = [len(x) for x in features] + encoded_inputs = BatchFeature({"input_values": features}) + self.feature_size = self.num_mel_bins + else: + encoded_inputs = BatchFeature({"input_values": speech}) + + padded_inputs = self.pad( + encoded_inputs, + padding=padding, + max_length=max_length, + truncation=truncation, + pad_to_multiple_of=pad_to_multiple_of, + return_attention_mask=return_attention_mask, + **kwargs, + ) + + self.feature_size = feature_size_hack + + # convert input values to correct format + input_values = padded_inputs["input_values"] + if not isinstance(input_values[0], np.ndarray): + padded_inputs["input_values"] = [np.asarray(array, dtype=np.float32) for array in input_values] + elif ( + not isinstance(input_values, np.ndarray) + and isinstance(input_values[0], np.ndarray) + and input_values[0].dtype is np.dtype(np.float64) + ): + padded_inputs["input_values"] = [array.astype(np.float32) for array in input_values] + elif isinstance(input_values, np.ndarray) and input_values.dtype is np.dtype(np.float64): + padded_inputs["input_values"] = input_values.astype(np.float32) + + # convert attention_mask to correct format + attention_mask = padded_inputs.get("attention_mask") + if attention_mask is not None: + padded_inputs["attention_mask"] = [np.asarray(array, dtype=np.int32) for array in attention_mask] + + # zero-mean and unit-variance normalization + if not is_target and self.do_normalize: + attention_mask = ( + attention_mask + if self._get_padding_strategies(padding, max_length=max_length) is not PaddingStrategy.DO_NOT_PAD + else None + ) + padded_inputs["input_values"] = self.zero_mean_unit_var_norm( + padded_inputs["input_values"], attention_mask=attention_mask, padding_value=self.padding_value + ) + + if is_target: + # make labels for stop prediction + stop_labels = [] + for i, l in enumerate(fbank_sizes): + labels = np.zeros(len(padded_inputs["input_values"][i])) + labels[l - 1 :] = 1.0 + stop_labels.append(labels) + padded_inputs["stop_labels"] = stop_labels + + # thin out frames for reduction factor + if self.reduction_factor > 1: + padded_inputs["input_values"] = self._reduce(padded_inputs["input_values"]) + if attention_mask is not None: + padded_inputs["attention_mask"] = self._reduce(padded_inputs["attention_mask"]) + + if return_tensors is not None: + padded_inputs = padded_inputs.convert_to_tensors(return_tensors) + + return padded_inputs diff --git a/src/transformers/models/speecht5/modeling_speecht5.py b/src/transformers/models/speecht5/modeling_speecht5.py new file mode 100644 index 00000000000..e76470dee3a --- /dev/null +++ b/src/transformers/models/speecht5/modeling_speecht5.py @@ -0,0 +1,3060 @@ +# coding=utf-8 +# Copyright 2023 The Fairseq Authors, Microsoft Research, 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 SpeechT5 model.""" + +import math +import random +from typing import List, Optional, Tuple, Union + +import numpy as np +import torch +import torch.utils.checkpoint +from torch import nn +from torch.nn import CrossEntropyLoss + +from ...activations import ACT2FN +from ...deepspeed import is_deepspeed_zero3_enabled +from ...modeling_outputs import ( + BaseModelOutput, + BaseModelOutputWithPastAndCrossAttentions, + Seq2SeqLMOutput, + Seq2SeqModelOutput, + Seq2SeqSpectrogramOutput, +) +from ...modeling_utils import PreTrainedModel +from ...pytorch_utils import torch_int_div +from ...utils import add_start_docstrings, add_start_docstrings_to_model_forward, logging, replace_return_docstrings +from .configuration_speecht5 import SpeechT5Config, SpeechT5HifiGanConfig + + +logger = logging.get_logger(__name__) + + +_HIDDEN_STATES_START_POSITION = 1 + +# General docstring +_CONFIG_FOR_DOC = "SpeechT5Config" + + +SPEECHT5_PRETRAINED_MODEL_ARCHIVE_LIST = [ + "microsoft/speecht5_asr", + "microsoft/speecht5_tts", + "microsoft/speecht5_vc", + # See all SpeechT5 models at https://huggingface.co/models?filter=speecht5 +] + + +# Copied from transformers.models.bart.modeling_bart.shift_tokens_right +def shift_tokens_right(input_ids: torch.Tensor, pad_token_id: int, decoder_start_token_id: int): + """ + Shift input ids one token to the right. + """ + shifted_input_ids = input_ids.new_zeros(input_ids.shape) + shifted_input_ids[:, 1:] = input_ids[:, :-1].clone() + shifted_input_ids[:, 0] = decoder_start_token_id + + if pad_token_id is None: + raise ValueError("self.model.config.pad_token_id has to be defined.") + # replace possible -100 values in labels by `pad_token_id` + shifted_input_ids.masked_fill_(shifted_input_ids == -100, pad_token_id) + + return shifted_input_ids + + +def shift_spectrograms_right(input_values: torch.Tensor): + """ + Shift input spectrograms one timestep to the right. + """ + shifted_input_values = input_values.new_zeros(input_values.shape) + shifted_input_values[:, 1:] = input_values[:, :-1].clone() + return shifted_input_values + + +# Copied from transformers.models.bart.modeling_bart._make_causal_mask +def _make_causal_mask(input_ids_shape: torch.Size, dtype: torch.dtype, past_key_values_length: int = 0): + """ + Make causal mask used for bi-directional self-attention. + """ + bsz, tgt_len = input_ids_shape + mask = torch.full((tgt_len, tgt_len), torch.tensor(torch.finfo(dtype).min)) + mask_cond = torch.arange(mask.size(-1)) + mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0) + mask = mask.to(dtype) + + if past_key_values_length > 0: + mask = torch.cat([torch.zeros(tgt_len, past_key_values_length, dtype=dtype), mask], dim=-1) + return mask[None, None, :, :].expand(bsz, 1, tgt_len, tgt_len + past_key_values_length) + + +# Copied from transformers.models.bart.modeling_bart._expand_mask +def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None): + """ + Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`. + """ + bsz, src_len = mask.size() + tgt_len = tgt_len if tgt_len is not None else src_len + + expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype) + + inverted_mask = 1.0 - expanded_mask + + return inverted_mask.masked_fill(inverted_mask.to(torch.bool), torch.finfo(dtype).min) + + +# Copied from transformers.models.wav2vec2.modeling_wav2vec2._compute_mask_indices +def _compute_mask_indices( + shape: Tuple[int, int], + mask_prob: float, + mask_length: int, + attention_mask: Optional[torch.LongTensor] = None, + min_masks: int = 0, +) -> np.ndarray: + """ + Computes random mask spans for a given shape. Used to implement [SpecAugment: A Simple Data Augmentation Method for + ASR](https://arxiv.org/abs/1904.08779). Note that this method is not optimized to run on TPU and should be run on + CPU as part of the preprocessing during training. + + Args: + shape: The shape for which to compute masks. This should be of a tuple of size 2 where + the first element is the batch size and the second element is the length of the axis to span. + mask_prob: The percentage of the whole axis (between 0 and 1) which will be masked. The number of + independently generated mask spans of length `mask_length` is computed by + `mask_prob*shape[1]/mask_length`. Note that due to overlaps, `mask_prob` is an upper bound and the + actual percentage will be smaller. + mask_length: size of the mask + min_masks: minimum number of masked spans + attention_mask: A (right-padded) attention mask which independently shortens the feature axis of + each batch dimension. + """ + batch_size, sequence_length = shape + + if mask_length < 1: + raise ValueError("`mask_length` has to be bigger than 0.") + + if mask_length > sequence_length: + raise ValueError( + f"`mask_length` has to be smaller than `sequence_length`, but got `mask_length`: {mask_length}" + f" and `sequence_length`: {sequence_length}`" + ) + + # epsilon is used for probabilistic rounding + epsilon = np.random.rand(1).item() + + def compute_num_masked_span(input_length): + """Given input length, compute how many spans should be masked""" + num_masked_span = int(mask_prob * input_length / mask_length + epsilon) + num_masked_span = max(num_masked_span, min_masks) + + # make sure num masked span <= sequence_length + if num_masked_span * mask_length > sequence_length: + num_masked_span = sequence_length // mask_length + + # make sure num_masked span is also <= input_length - (mask_length - 1) + if input_length - (mask_length - 1) < num_masked_span: + num_masked_span = max(input_length - (mask_length - 1), 0) + + return num_masked_span + + # compute number of masked spans in batch + input_lengths = ( + attention_mask.sum(-1).detach().tolist() + if attention_mask is not None + else [sequence_length for _ in range(batch_size)] + ) + + # SpecAugment mask to fill + spec_aug_mask = np.zeros((batch_size, sequence_length), dtype=bool) + spec_aug_mask_idxs = [] + + max_num_masked_span = compute_num_masked_span(sequence_length) + + if max_num_masked_span == 0: + return spec_aug_mask + + for input_length in input_lengths: + # compute num of masked spans for this input + num_masked_span = compute_num_masked_span(input_length) + + # get random indices to mask + spec_aug_mask_idx = np.random.choice( + np.arange(input_length - (mask_length - 1)), num_masked_span, replace=False + ) + + # pick first sampled index that will serve as a dummy index to pad vector + # to ensure same dimension for all batches due to probabilistic rounding + # Picking first sample just pads those vectors twice. + if len(spec_aug_mask_idx) == 0: + # this case can only happen if `input_length` is strictly smaller then + # `sequence_length` in which case the last token has to be a padding + # token which we can use as a dummy mask id + dummy_mask_idx = sequence_length - 1 + else: + dummy_mask_idx = spec_aug_mask_idx[0] + + spec_aug_mask_idx = np.concatenate( + [spec_aug_mask_idx, np.ones(max_num_masked_span - num_masked_span, dtype=np.int32) * dummy_mask_idx] + ) + spec_aug_mask_idxs.append(spec_aug_mask_idx) + + spec_aug_mask_idxs = np.array(spec_aug_mask_idxs) + + # expand masked indices to masked spans + spec_aug_mask_idxs = np.broadcast_to( + spec_aug_mask_idxs[:, :, None], (batch_size, max_num_masked_span, mask_length) + ) + spec_aug_mask_idxs = spec_aug_mask_idxs.reshape(batch_size, max_num_masked_span * mask_length) + + # add offset to the starting indexes so that indexes now create a span + offsets = np.arange(mask_length)[None, None, :] + offsets = np.broadcast_to(offsets, (batch_size, max_num_masked_span, mask_length)).reshape( + batch_size, max_num_masked_span * mask_length + ) + spec_aug_mask_idxs = spec_aug_mask_idxs + offsets + + # ensure that we cannot have indices larger than sequence_length + if spec_aug_mask_idxs.max() > sequence_length - 1: + spec_aug_mask_idxs[spec_aug_mask_idxs > sequence_length - 1] = sequence_length - 1 + + # scatter indices to mask + np.put_along_axis(spec_aug_mask, spec_aug_mask_idxs, 1, -1) + + return spec_aug_mask + + +# Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2NoLayerNormConvLayer with Wav2Vec2->SpeechT5 +class SpeechT5NoLayerNormConvLayer(nn.Module): + def __init__(self, config, layer_id=0): + super().__init__() + self.in_conv_dim = config.conv_dim[layer_id - 1] if layer_id > 0 else 1 + self.out_conv_dim = config.conv_dim[layer_id] + + self.conv = nn.Conv1d( + self.in_conv_dim, + self.out_conv_dim, + kernel_size=config.conv_kernel[layer_id], + stride=config.conv_stride[layer_id], + bias=config.conv_bias, + ) + self.activation = ACT2FN[config.feat_extract_activation] + + def forward(self, hidden_states): + hidden_states = self.conv(hidden_states) + hidden_states = self.activation(hidden_states) + return hidden_states + + +# Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2LayerNormConvLayer with Wav2Vec2->SpeechT5 +class SpeechT5LayerNormConvLayer(nn.Module): + def __init__(self, config, layer_id=0): + super().__init__() + self.in_conv_dim = config.conv_dim[layer_id - 1] if layer_id > 0 else 1 + self.out_conv_dim = config.conv_dim[layer_id] + + self.conv = nn.Conv1d( + self.in_conv_dim, + self.out_conv_dim, + kernel_size=config.conv_kernel[layer_id], + stride=config.conv_stride[layer_id], + bias=config.conv_bias, + ) + self.layer_norm = nn.LayerNorm(self.out_conv_dim, elementwise_affine=True) + self.activation = ACT2FN[config.feat_extract_activation] + + def forward(self, hidden_states): + hidden_states = self.conv(hidden_states) + + hidden_states = hidden_states.transpose(-2, -1) + hidden_states = self.layer_norm(hidden_states) + hidden_states = hidden_states.transpose(-2, -1) + + hidden_states = self.activation(hidden_states) + return hidden_states + + +# Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2GroupNormConvLayer with Wav2Vec2->SpeechT5 +class SpeechT5GroupNormConvLayer(nn.Module): + def __init__(self, config, layer_id=0): + super().__init__() + self.in_conv_dim = config.conv_dim[layer_id - 1] if layer_id > 0 else 1 + self.out_conv_dim = config.conv_dim[layer_id] + + self.conv = nn.Conv1d( + self.in_conv_dim, + self.out_conv_dim, + kernel_size=config.conv_kernel[layer_id], + stride=config.conv_stride[layer_id], + bias=config.conv_bias, + ) + self.activation = ACT2FN[config.feat_extract_activation] + + self.layer_norm = nn.GroupNorm(num_groups=self.out_conv_dim, num_channels=self.out_conv_dim, affine=True) + + def forward(self, hidden_states): + hidden_states = self.conv(hidden_states) + hidden_states = self.layer_norm(hidden_states) + hidden_states = self.activation(hidden_states) + return hidden_states + + +# Copied from transformers.models.speech_to_text.modeling_speech_to_text.Speech2TextSinusoidalPositionalEmbedding with Speech2Text->SpeechT5 +class SpeechT5SinusoidalPositionalEmbedding(nn.Module): + """This module produces sinusoidal positional embeddings of any length.""" + + def __init__(self, num_positions: int, embedding_dim: int, padding_idx: Optional[int] = None): + super().__init__() + self.offset = 2 + self.embedding_dim = embedding_dim + self.padding_idx = padding_idx + self.make_weights(num_positions + self.offset, embedding_dim, padding_idx) + + def make_weights(self, num_embeddings: int, embedding_dim: int, padding_idx: Optional[int] = None): + emb_weights = self.get_embedding(num_embeddings, embedding_dim, padding_idx) + if hasattr(self, "weights"): + # in forward put the weights on the correct dtype and device of the param + emb_weights = emb_weights.to(dtype=self.weights.dtype, device=self.weights.device) + + self.weights = nn.Parameter(emb_weights) + self.weights.requires_grad = False + self.weights.detach_() + + @staticmethod + def get_embedding(num_embeddings: int, embedding_dim: int, padding_idx: Optional[int] = None): + """ + Build sinusoidal embeddings. This matches the implementation in tensor2tensor, but differs slightly from the + description in Section 3.5 of "Attention Is All You Need". + """ + half_dim = embedding_dim // 2 + emb = math.log(10000) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=torch.float) * -emb) + emb = torch.arange(num_embeddings, dtype=torch.float).unsqueeze(1) * emb.unsqueeze(0) + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1).view(num_embeddings, -1) + if embedding_dim % 2 == 1: + # zero pad + emb = torch.cat([emb, torch.zeros(num_embeddings, 1)], dim=1) + if padding_idx is not None: + emb[padding_idx, :] = 0 + return emb.to(torch.get_default_dtype()) + + @torch.no_grad() + def forward(self, input_ids: torch.Tensor, past_key_values_length: int = 0): + bsz, seq_len = input_ids.size() + # Create the position ids from the input token ids. Any padded tokens remain padded. + position_ids = self.create_position_ids_from_input_ids(input_ids, self.padding_idx, past_key_values_length).to( + input_ids.device + ) + + # expand embeddings if needed + max_pos = self.padding_idx + 1 + seq_len + if max_pos > self.weights.size(0): + self.make_weights(max_pos + self.offset, self.embedding_dim, self.padding_idx) + + return self.weights.index_select(0, position_ids.view(-1)).view(bsz, seq_len, -1).detach() + + def create_position_ids_from_input_ids( + self, input_ids: torch.Tensor, padding_idx: int, past_key_values_length: Optional[int] = 0 + ): + """ + Replace non-padding symbols with their position numbers. Position numbers begin at padding_idx+1. Padding + symbols are ignored. This is modified from fairseq's `utils.make_positions`. + + Args: + x: torch.Tensor x: + Returns: torch.Tensor + """ + # The series of casts and type-conversions here are carefully balanced to both work with ONNX export and XLA. + mask = input_ids.ne(padding_idx).int() + incremental_indices = (torch.cumsum(mask, dim=1).type_as(mask) + past_key_values_length) * mask + return incremental_indices.long() + padding_idx + + +# Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2PositionalConvEmbedding with Wav2Vec2->SpeechT5 +class SpeechT5PositionalConvEmbedding(nn.Module): + def __init__(self, config): + super().__init__() + self.conv = nn.Conv1d( + config.hidden_size, + config.hidden_size, + kernel_size=config.num_conv_pos_embeddings, + padding=config.num_conv_pos_embeddings // 2, + groups=config.num_conv_pos_embedding_groups, + ) + + if is_deepspeed_zero3_enabled(): + import deepspeed + + with deepspeed.zero.GatheredParameters(self.conv.weight, modifier_rank=0): + self.conv = nn.utils.weight_norm(self.conv, name="weight", dim=2) + deepspeed.zero.register_external_parameter(self, self.conv.weight_v) + deepspeed.zero.register_external_parameter(self, self.conv.weight_g) + else: + self.conv = nn.utils.weight_norm(self.conv, name="weight", dim=2) + + self.padding = SpeechT5SamePadLayer(config.num_conv_pos_embeddings) + self.activation = ACT2FN[config.feat_extract_activation] + + def forward(self, hidden_states): + hidden_states = hidden_states.transpose(1, 2) + + hidden_states = self.conv(hidden_states) + hidden_states = self.padding(hidden_states) + hidden_states = self.activation(hidden_states) + + hidden_states = hidden_states.transpose(1, 2) + return hidden_states + + +class SpeechT5ScaledPositionalEncoding(nn.Module): + """ + Scaled positional encoding, see §3.2 in https://arxiv.org/abs/1809.08895 + """ + + def __init__(self, dropout, dim, max_len=5000): + pe = torch.zeros(max_len, dim) + position = torch.arange(0, max_len).unsqueeze(1) + div_term = torch.exp((torch.arange(0, dim, 2, dtype=torch.float) * -(math.log(10000.0) / dim))) + pe[:, 0::2] = torch.sin(position.float() * div_term) + pe[:, 1::2] = torch.cos(position.float() * div_term) + pe = pe.unsqueeze(0) + super().__init__() + self.register_buffer("pe", pe) + self.dropout = nn.Dropout(p=dropout) + self.dim = dim + self.alpha = torch.nn.Parameter(torch.tensor(1.0)) + + def forward(self, emb): + emb = emb + self.alpha * self.pe[:, : emb.size(1)] + emb = self.dropout(emb) + return emb + + +class SpeechT5RelativePositionalEncoding(torch.nn.Module): + def __init__(self, dim, max_length=1000): + super().__init__() + self.dim = dim + self.max_length = max_length + self.pe_k = torch.nn.Embedding(2 * max_length, dim) + + def forward(self, hidden_states): + seq_len = hidden_states.shape[1] + pos_seq = torch.arange(0, seq_len).long().to(hidden_states.device) + pos_seq = pos_seq[:, None] - pos_seq[None, :] + + pos_seq[pos_seq < -self.max_length] = -self.max_length + pos_seq[pos_seq >= self.max_length] = self.max_length - 1 + pos_seq = pos_seq + self.max_length + + return self.pe_k(pos_seq) + + +# Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2SamePadLayer with Wav2Vec2->SpeechT5 +class SpeechT5SamePadLayer(nn.Module): + def __init__(self, num_conv_pos_embeddings): + super().__init__() + self.num_pad_remove = 1 if num_conv_pos_embeddings % 2 == 0 else 0 + + def forward(self, hidden_states): + if self.num_pad_remove > 0: + hidden_states = hidden_states[:, :, : -self.num_pad_remove] + return hidden_states + + +# Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2FeatureEncoder with Wav2Vec2->SpeechT5 +class SpeechT5FeatureEncoder(nn.Module): + """Construct the features from raw audio waveform""" + + def __init__(self, config): + super().__init__() + + if config.feat_extract_norm == "group": + conv_layers = [SpeechT5GroupNormConvLayer(config, layer_id=0)] + [ + SpeechT5NoLayerNormConvLayer(config, layer_id=i + 1) for i in range(config.num_feat_extract_layers - 1) + ] + elif config.feat_extract_norm == "layer": + conv_layers = [ + SpeechT5LayerNormConvLayer(config, layer_id=i) for i in range(config.num_feat_extract_layers) + ] + else: + raise ValueError( + f"`config.feat_extract_norm` is {config.feat_extract_norm}, but has to be one of ['group', 'layer']" + ) + self.conv_layers = nn.ModuleList(conv_layers) + self.gradient_checkpointing = False + self._requires_grad = True + + def _freeze_parameters(self): + for param in self.parameters(): + param.requires_grad = False + self._requires_grad = False + + def forward(self, input_values): + hidden_states = input_values[:, None] + + # make sure hidden_states require grad for gradient_checkpointing + if self._requires_grad and self.training: + hidden_states.requires_grad = True + + for conv_layer in self.conv_layers: + if self._requires_grad and self.gradient_checkpointing and self.training: + + def create_custom_forward(module): + def custom_forward(*inputs): + return module(*inputs) + + return custom_forward + + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(conv_layer), + hidden_states, + ) + else: + hidden_states = conv_layer(hidden_states) + + return hidden_states + + +# Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2FeatureProjection with Wav2Vec2->SpeechT5 +class SpeechT5FeatureProjection(nn.Module): + def __init__(self, config): + super().__init__() + self.layer_norm = nn.LayerNorm(config.conv_dim[-1], eps=config.layer_norm_eps) + self.projection = nn.Linear(config.conv_dim[-1], config.hidden_size) + self.dropout = nn.Dropout(config.feat_proj_dropout) + + def forward(self, hidden_states): + # non-projected hidden states are needed for quantization + norm_hidden_states = self.layer_norm(hidden_states) + hidden_states = self.projection(norm_hidden_states) + hidden_states = self.dropout(hidden_states) + return hidden_states, norm_hidden_states + + +class SpeechT5SpeechEncoderPrenet(nn.Module): + def __init__(self, config): + super().__init__() + self.config = config + self.feature_encoder = SpeechT5FeatureEncoder(config) + self.feature_projection = SpeechT5FeatureProjection(config) + + # model only needs masking vector if mask prob is > 0.0 + if config.mask_time_prob > 0.0 or config.mask_feature_prob > 0.0: + self.masked_spec_embed = nn.Parameter(torch.FloatTensor(config.hidden_size).uniform_()) + + self.pos_conv_embed = SpeechT5PositionalConvEmbedding(config) + self.pos_sinusoidal_embed = SpeechT5SinusoidalPositionalEmbedding( + config.max_speech_positions + config.pad_token_id + 1, + config.hidden_size, + config.pad_token_id, + ) + + def freeze_feature_encoder(self): + self.feature_encoder._freeze_parameters() + + def forward( + self, + input_values: torch.Tensor, + attention_mask: Optional[torch.Tensor] = None, + mask_time_indices: Optional[torch.FloatTensor] = None, + ): + extract_features = self.feature_encoder(input_values) + extract_features = extract_features.transpose(1, 2) + + if attention_mask is not None: + # compute reduced attention_mask corresponding to feature vectors + attention_mask = self._get_feature_vector_attention_mask( + extract_features.shape[1], + attention_mask, + ) + + hidden_states, extract_features = self.feature_projection(extract_features) + hidden_states = self._mask_hidden_states( + hidden_states, mask_time_indices=mask_time_indices, attention_mask=attention_mask + ) + + positional_conv_embedding = self.pos_conv_embed(hidden_states) + hidden_states = hidden_states + positional_conv_embedding + + if attention_mask is not None: + padding_mask = attention_mask.ne(1).long() + else: + padding_mask = torch.zeros(hidden_states.shape[:2], dtype=torch.long, device=hidden_states.device) + + positional_sinusoidal_embeddings = self.pos_sinusoidal_embed(padding_mask) + hidden_states = hidden_states + positional_sinusoidal_embeddings + + return hidden_states, attention_mask + + # Copied from transformers.models.unispeech.modeling_unispeech.UniSpeechPreTrainedModel._get_feature_vector_attention_mask + def _get_feature_vector_attention_mask(self, feature_vector_length: int, attention_mask: torch.LongTensor): + # Effectively attention_mask.sum(-1), but not inplace to be able to run + # on inference mode. + non_padded_lengths = attention_mask.cumsum(dim=-1)[:, -1] + output_lengths = self._get_feat_extract_output_lengths(non_padded_lengths).to(torch.long) + batch_size = attention_mask.shape[0] + + attention_mask = torch.zeros( + (batch_size, feature_vector_length), dtype=attention_mask.dtype, device=attention_mask.device + ) + # these two operations makes sure that all values before the output lengths idxs are attended to + attention_mask[(torch.arange(attention_mask.shape[0], device=attention_mask.device), output_lengths - 1)] = 1 + attention_mask = attention_mask.flip([-1]).cumsum(-1).flip([-1]).bool() + return attention_mask + + # Copied from transformers.models.unispeech.modeling_unispeech.UniSpeechPreTrainedModel._get_feat_extract_output_lengths + def _get_feat_extract_output_lengths(self, input_lengths: Union[torch.LongTensor, int]): + """ + Computes the output length of the convolutional layers + """ + + def _conv_out_length(input_length, kernel_size, stride): + # 1D convolutional layer output length formula taken + # from https://pytorch.org/docs/stable/generated/torch.nn.Conv1d.html + return torch_int_div(input_length - kernel_size, stride) + 1 + + for kernel_size, stride in zip(self.config.conv_kernel, self.config.conv_stride): + input_lengths = _conv_out_length(input_lengths, kernel_size, stride) + + return input_lengths + + # Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2Model._mask_hidden_states + def _mask_hidden_states( + self, + hidden_states: torch.FloatTensor, + mask_time_indices: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.LongTensor] = None, + ): + """ + Masks extracted features along time axis and/or along feature axis according to + [SpecAugment](https://arxiv.org/abs/1904.08779). + """ + + # `config.apply_spec_augment` can set masking to False + if not getattr(self.config, "apply_spec_augment", True): + return hidden_states + + # generate indices & apply SpecAugment along time axis + batch_size, sequence_length, hidden_size = hidden_states.size() + + if mask_time_indices is not None: + # apply SpecAugment along time axis with given mask_time_indices + hidden_states[mask_time_indices] = self.masked_spec_embed.to(hidden_states.dtype) + elif self.config.mask_time_prob > 0 and self.training: + mask_time_indices = _compute_mask_indices( + (batch_size, sequence_length), + mask_prob=self.config.mask_time_prob, + mask_length=self.config.mask_time_length, + attention_mask=attention_mask, + min_masks=self.config.mask_time_min_masks, + ) + mask_time_indices = torch.tensor(mask_time_indices, device=hidden_states.device, dtype=torch.bool) + hidden_states[mask_time_indices] = self.masked_spec_embed.to(hidden_states.dtype) + + if self.config.mask_feature_prob > 0 and self.training: + # generate indices & apply SpecAugment along feature axis + mask_feature_indices = _compute_mask_indices( + (batch_size, hidden_size), + mask_prob=self.config.mask_feature_prob, + mask_length=self.config.mask_feature_length, + min_masks=self.config.mask_feature_min_masks, + ) + mask_feature_indices = torch.tensor(mask_feature_indices, device=hidden_states.device, dtype=torch.bool) + mask_feature_indices = mask_feature_indices[:, None].expand(-1, sequence_length, -1) + hidden_states[mask_feature_indices] = 0 + + return hidden_states + + +class SpeechT5SpeechDecoderPrenet(nn.Module): + def __init__(self, config): + super().__init__() + self.config = config + + self.layers = nn.ModuleList( + [ + nn.Linear( + config.num_mel_bins if i == 0 else config.speech_decoder_prenet_units, + config.speech_decoder_prenet_units, + ) + for i in range(config.speech_decoder_prenet_layers) + ] + ) + + self.final_layer = nn.Linear(config.speech_decoder_prenet_units, config.hidden_size) + + self.encode_positions = SpeechT5ScaledPositionalEncoding( + config.positional_dropout, + config.hidden_size, + config.max_speech_positions, + ) + + self.speaker_embeds_layer = nn.Linear(config.speaker_embedding_dim + config.hidden_size, config.hidden_size) + + def forward( + self, + input_values: torch.Tensor, + speaker_embeddings: Optional[torch.Tensor] = None, + ): + # Dropout is always applied, even when evaluating. See §2.2 in https://arxiv.org/abs/1712.05884. + + inputs_embeds = input_values + for layer in self.layers: + inputs_embeds = nn.functional.relu(layer(inputs_embeds)) + inputs_embeds = nn.functional.dropout( + inputs_embeds, self.config.speech_decoder_prenet_dropout, training=True + ) + + inputs_embeds = self.final_layer(inputs_embeds) + inputs_embeds = self.encode_positions(inputs_embeds) + + if speaker_embeddings is not None: + speaker_embeddings = nn.functional.normalize(speaker_embeddings) + speaker_embeddings = speaker_embeddings.unsqueeze(1) + speaker_embeddings = speaker_embeddings.expand(-1, inputs_embeds.size(1), -1) + inputs_embeds = torch.cat([inputs_embeds, speaker_embeddings], dim=-1) + inputs_embeds = nn.functional.relu(self.speaker_embeds_layer(inputs_embeds)) + + return inputs_embeds + + +class SpeechT5BatchNormConvLayer(nn.Module): + def __init__(self, config, layer_id=0): + super().__init__() + + if layer_id == 0: + in_conv_dim = config.num_mel_bins + else: + in_conv_dim = config.speech_decoder_postnet_units + + if layer_id == config.speech_decoder_postnet_layers - 1: + out_conv_dim = config.num_mel_bins + else: + out_conv_dim = config.speech_decoder_postnet_units + + self.conv = nn.Conv1d( + in_conv_dim, + out_conv_dim, + kernel_size=config.speech_decoder_postnet_kernel, + stride=1, + padding=(config.speech_decoder_postnet_kernel - 1) // 2, + bias=False, + ) + self.batch_norm = nn.BatchNorm1d(out_conv_dim) + + if layer_id < config.speech_decoder_postnet_layers - 1: + self.activation = nn.Tanh() + else: + self.activation = None + + self.dropout = nn.Dropout(config.speech_decoder_postnet_dropout) + + def forward(self, hidden_states): + hidden_states = self.conv(hidden_states) + hidden_states = self.batch_norm(hidden_states) + if self.activation is not None: + hidden_states = self.activation(hidden_states) + hidden_states = self.dropout(hidden_states) + return hidden_states + + +class SpeechT5SpeechDecoderPostnet(nn.Module): + def __init__(self, config): + super().__init__() + self.config = config + + self.feat_out = nn.Linear(config.hidden_size, config.num_mel_bins * config.reduction_factor) + self.prob_out = nn.Linear(config.hidden_size, config.reduction_factor) + + self.layers = nn.ModuleList( + [SpeechT5BatchNormConvLayer(config, i) for i in range(config.speech_decoder_postnet_layers)] + ) + + def forward(self, hidden_states: torch.Tensor): + outputs_before_postnet = self.feat_out(hidden_states).view(hidden_states.size(0), -1, self.config.num_mel_bins) + outputs_after_postnet = self.postnet(outputs_before_postnet) + logits = self.prob_out(hidden_states).view(hidden_states.size(0), -1) + return outputs_before_postnet, outputs_after_postnet, logits + + def postnet(self, hidden_states: torch.Tensor): + layer_output = hidden_states.transpose(1, 2) + for layer in self.layers: + layer_output = layer(layer_output) + return hidden_states + layer_output.transpose(1, 2) + + +class SpeechT5TextEncoderPrenet(nn.Module): + def __init__(self, config): + super().__init__() + self.config = config + self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, config.pad_token_id) + self.encode_positions = SpeechT5ScaledPositionalEncoding( + config.positional_dropout, + config.hidden_size, + config.max_text_positions, + ) + + def get_input_embeddings(self): + return self.embed_tokens + + def set_input_embeddings(self, value): + self.embed_tokens = value + + def forward(self, input_ids: torch.Tensor): + inputs_embeds = self.embed_tokens(input_ids) + inputs_embeds = self.encode_positions(inputs_embeds) + return inputs_embeds + + +class SpeechT5TextDecoderPrenet(nn.Module): + def __init__(self, config): + super().__init__() + self.config = config + self.dropout = nn.Dropout(config.positional_dropout) + self.embed_scale = math.sqrt(config.hidden_size) if config.scale_embedding else 1.0 + + self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, config.pad_token_id) + + self.embed_positions = SpeechT5SinusoidalPositionalEmbedding( + config.max_text_positions + config.pad_token_id + 1, + config.hidden_size, + config.pad_token_id, + ) + + def get_input_embeddings(self): + return self.embed_tokens + + def set_input_embeddings(self, value): + self.embed_tokens = value + + def forward( + self, + input_ids: torch.Tensor, + attention_mask: Optional[torch.Tensor] = None, + past_key_values: Optional[List[torch.FloatTensor]] = None, + ): + if input_ids is not None: + input_shape = input_ids.size() + input_ids = input_ids.view(-1, input_shape[-1]) + else: + raise ValueError("You have to specify `decoder_input_ids`") + + past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0 + positions = self.embed_positions(input_ids, past_key_values_length) + + inputs_embeds = self.embed_tokens(input_ids) * self.embed_scale + inputs_embeds += positions + inputs_embeds = self.dropout(inputs_embeds) + + return inputs_embeds, attention_mask + + +class SpeechT5TextDecoderPostnet(nn.Module): + def __init__(self, config): + super().__init__() + self.config = config + self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False) + + def forward(self, hidden_states: torch.Tensor): + return self.lm_head(hidden_states) + + def get_output_embeddings(self): + return self.lm_head + + def set_output_embeddings(self, new_embeddings): + self.lm_head = new_embeddings + + +class SpeechT5Attention(nn.Module): + """ + Multi-headed attention from 'Attention Is All You Need' paper with relative position bias (see + https://aclanthology.org/N18-2074.pdf) + """ + + def __init__( + self, + embed_dim: int, + num_heads: int, + dropout: float = 0.0, + is_decoder: bool = False, + bias: bool = True, + ): + super().__init__() + self.embed_dim = embed_dim + self.num_heads = num_heads + self.dropout = dropout + self.head_dim = embed_dim // num_heads + + if (self.head_dim * num_heads) != self.embed_dim: + raise ValueError( + f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim}" + f" and `num_heads`: {num_heads})." + ) + self.scaling = self.head_dim**-0.5 + self.is_decoder = is_decoder + + self.k_proj = nn.Linear(embed_dim, embed_dim, bias=bias) + self.v_proj = nn.Linear(embed_dim, embed_dim, bias=bias) + self.q_proj = nn.Linear(embed_dim, embed_dim, bias=bias) + self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias) + + def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int): + return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous() + + def forward( + self, + hidden_states: torch.Tensor, + key_value_states: Optional[torch.Tensor] = None, + past_key_value: Optional[Tuple[torch.Tensor]] = None, + attention_mask: Optional[torch.Tensor] = None, + layer_head_mask: Optional[torch.Tensor] = None, + position_bias: Optional[torch.Tensor] = None, + output_attentions: bool = False, + ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: + """Input shape: Batch x Time x Channel""" + + # if key_value_states are provided this layer is used as a cross-attention layer + # for the decoder + is_cross_attention = key_value_states is not None + + bsz, tgt_len, _ = hidden_states.size() + + # get query proj + query_states = self.q_proj(hidden_states) * self.scaling + # get key, value proj + if is_cross_attention and past_key_value is not None: + # reuse k,v, cross_attentions + key_states = past_key_value[0] + value_states = past_key_value[1] + elif is_cross_attention: + # cross_attentions + key_states = self._shape(self.k_proj(key_value_states), -1, bsz) + value_states = self._shape(self.v_proj(key_value_states), -1, bsz) + elif past_key_value is not None: + # reuse k, v, self_attention + key_states = self._shape(self.k_proj(hidden_states), -1, bsz) + value_states = self._shape(self.v_proj(hidden_states), -1, bsz) + key_states = torch.cat([past_key_value[0], key_states], dim=2) + value_states = torch.cat([past_key_value[1], value_states], dim=2) + else: + # self_attention + key_states = self._shape(self.k_proj(hidden_states), -1, bsz) + value_states = self._shape(self.v_proj(hidden_states), -1, bsz) + + if self.is_decoder: + # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states. + # Further calls to cross_attention layer can then reuse all cross-attention + # key/value_states (first "if" case) + # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of + # all previous decoder key/value_states. Further calls to uni-directional self-attention + # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case) + # if encoder bi-directional self-attention `past_key_value` is always `None` + past_key_value = (key_states, value_states) + + proj_shape = (bsz * self.num_heads, -1, self.head_dim) + query_states = self._shape(query_states, tgt_len, bsz).view(*proj_shape) + key_states = key_states.view(*proj_shape) + value_states = value_states.view(*proj_shape) + + src_len = key_states.size(1) + attn_weights = torch.bmm(query_states, key_states.transpose(1, 2)) + + if attn_weights.size() != (bsz * self.num_heads, tgt_len, src_len): + raise ValueError( + f"Attention weights should be of size {(bsz * self.num_heads, tgt_len, src_len)}, but is" + f" {attn_weights.size()}" + ) + + # relative attention bias + if position_bias is not None: + reshape_q = query_states.contiguous().view(bsz * self.num_heads, -1, self.head_dim).transpose(0, 1) + rel_pos_bias = torch.matmul(reshape_q, position_bias.transpose(-2, -1)) + rel_pos_bias = rel_pos_bias.transpose(0, 1).view( + bsz * self.num_heads, position_bias.size(0), position_bias.size(1) + ) + attn_weights += rel_pos_bias + + if attention_mask is not None: + if attention_mask.size() != (bsz, 1, tgt_len, src_len): + raise ValueError( + f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is {attention_mask.size()}" + ) + attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + attention_mask + attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len) + + attn_weights = nn.functional.softmax(attn_weights, dim=-1) + + if layer_head_mask is not None: + if layer_head_mask.size() != (self.num_heads,): + raise ValueError( + f"Head mask for a single layer should be of size {(self.num_heads,)}, but is" + f" {layer_head_mask.size()}" + ) + attn_weights = layer_head_mask.view(1, -1, 1, 1) * attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len) + + if output_attentions: + # this operation is a bit awkward, but it's required to + # make sure that attn_weights keeps its gradient. + # In order to do so, attn_weights have to be reshaped + # twice and have to be reused in the following + attn_weights_reshaped = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + attn_weights = attn_weights_reshaped.view(bsz * self.num_heads, tgt_len, src_len) + else: + attn_weights_reshaped = None + + attn_probs = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training) + + attn_output = torch.bmm(attn_probs, value_states) + + if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim): + raise ValueError( + f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is" + f" {attn_output.size()}" + ) + + attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim) + attn_output = attn_output.transpose(1, 2) + + # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be + # partitioned aross GPUs when using tensor-parallelism. + attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim) + + attn_output = self.out_proj(attn_output) + + return attn_output, attn_weights_reshaped, past_key_value + + +class SpeechT5FeedForward(nn.Module): + def __init__(self, config, intermediate_size): + super().__init__() + self.intermediate_dropout = nn.Dropout(config.activation_dropout) + + self.intermediate_dense = nn.Linear(config.hidden_size, intermediate_size) + if isinstance(config.hidden_act, str): + self.intermediate_act_fn = ACT2FN[config.hidden_act] + else: + self.intermediate_act_fn = config.hidden_act + + self.output_dense = nn.Linear(intermediate_size, config.hidden_size) + self.output_dropout = nn.Dropout(config.hidden_dropout) + + def forward(self, hidden_states): + hidden_states = self.intermediate_dense(hidden_states) + hidden_states = self.intermediate_act_fn(hidden_states) + hidden_states = self.intermediate_dropout(hidden_states) + + hidden_states = self.output_dense(hidden_states) + hidden_states = self.output_dropout(hidden_states) + return hidden_states + + +class SpeechT5EncoderLayer(nn.Module): + def __init__(self, config: SpeechT5Config): + super().__init__() + self.attention = SpeechT5Attention( + embed_dim=config.hidden_size, + num_heads=config.encoder_attention_heads, + dropout=config.attention_dropout, + is_decoder=False, + ) + self.dropout = nn.Dropout(config.hidden_dropout) + self.layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps) + self.feed_forward = SpeechT5FeedForward(config, config.encoder_ffn_dim) + self.final_layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps) + + def forward( + self, + hidden_states: torch.Tensor, + attention_mask: Optional[torch.Tensor] = None, + layer_head_mask: Optional[torch.Tensor] = None, + position_bias: Optional[torch.Tensor] = None, + output_attentions: bool = False, + ): + """ + Args: + hidden_states (`torch.FloatTensor`): + input to the layer of shape `(batch, seq_len, hidden_size)` + attention_mask (`torch.FloatTensor`): + attention mask of size `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very + large negative values. + layer_head_mask (`torch.FloatTensor`): mask for attention heads in a given layer of size + `(config.encoder_attention_heads,)`. + position_bias (`torch.FloatTensor`): + relative position embeddings of size `(seq_len, seq_len, hidden_size // encoder_attention_heads)` + output_attentions (`bool`, *optional*): + Whether or not to return the attentions tensors of all attention layers. See `attentions` under + returned tensors for more detail. + """ + residual = hidden_states + hidden_states, attn_weights, _ = self.attention( + hidden_states=hidden_states, + attention_mask=attention_mask, + layer_head_mask=layer_head_mask, + position_bias=position_bias, + output_attentions=output_attentions, + ) + + hidden_states = self.dropout(hidden_states) + hidden_states = residual + hidden_states + + hidden_states = self.layer_norm(hidden_states) + hidden_states = hidden_states + self.feed_forward(hidden_states) + hidden_states = self.final_layer_norm(hidden_states) + + outputs = (hidden_states,) + + if output_attentions: + outputs += (attn_weights,) + + return outputs + + +class SpeechT5DecoderLayer(nn.Module): + def __init__(self, config: SpeechT5Config): + super().__init__() + self.self_attn = SpeechT5Attention( + embed_dim=config.hidden_size, + num_heads=config.decoder_attention_heads, + dropout=config.attention_dropout, + is_decoder=True, + ) + self.dropout = nn.Dropout(config.hidden_dropout) + self.self_attn_layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps) + + self.encoder_attn = SpeechT5Attention( + config.hidden_size, + config.decoder_attention_heads, + dropout=config.attention_dropout, + is_decoder=True, + ) + self.encoder_attn_layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps) + + self.feed_forward = SpeechT5FeedForward(config, config.decoder_ffn_dim) + self.final_layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps) + + def forward( + self, + hidden_states: torch.Tensor, + attention_mask: Optional[torch.Tensor] = None, + encoder_hidden_states: Optional[torch.Tensor] = None, + encoder_attention_mask: Optional[torch.Tensor] = None, + layer_head_mask: Optional[torch.Tensor] = None, + cross_attn_layer_head_mask: Optional[torch.Tensor] = None, + past_key_value: Optional[Tuple[torch.Tensor]] = None, + output_attentions: Optional[bool] = False, + use_cache: Optional[bool] = True, + ): + """ + Args: + hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, hidden_size)` + attention_mask (`torch.FloatTensor`): attention mask of size + `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values. + encoder_hidden_states (`torch.FloatTensor`): + cross attention input to the layer of shape `(batch, seq_len, hidden_size)` + encoder_attention_mask (`torch.FloatTensor`): encoder attention mask of size + `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values. + layer_head_mask (`torch.FloatTensor`): mask for attention heads in a given layer of size + `(encoder_attention_heads,)`. + cross_attn_layer_head_mask (`torch.FloatTensor`): mask for cross-attention heads in a given layer of + size `(decoder_attention_heads,)`. + past_key_value (`Tuple(torch.FloatTensor)`): cached past key and value projection states + output_attentions (`bool`, *optional*): + Whether or not to return the attentions tensors of all attention layers. See `attentions` under + returned tensors for more detail. + """ + residual = hidden_states + + # Self Attention + # decoder uni-directional self-attention cached key/values tuple is at positions 1,2 + self_attn_past_key_value = past_key_value[:2] if past_key_value is not None else None + # add present self-attn cache to positions 1,2 of present_key_value tuple + hidden_states, self_attn_weights, present_key_value = self.self_attn( + hidden_states=hidden_states, + past_key_value=self_attn_past_key_value, + attention_mask=attention_mask, + layer_head_mask=layer_head_mask, + output_attentions=output_attentions, + ) + hidden_states = self.dropout(hidden_states) + hidden_states = residual + hidden_states + hidden_states = self.self_attn_layer_norm(hidden_states) + + # Cross-Attention Block + cross_attn_present_key_value = None + cross_attn_weights = None + if encoder_hidden_states is not None: + residual = hidden_states + + # cross_attn cached key/values tuple is at positions 3,4 of present_key_value tuple + cross_attn_past_key_value = past_key_value[-2:] if past_key_value is not None else None + hidden_states, cross_attn_weights, cross_attn_present_key_value = self.encoder_attn( + hidden_states=hidden_states, + key_value_states=encoder_hidden_states, + attention_mask=encoder_attention_mask, + layer_head_mask=cross_attn_layer_head_mask, + past_key_value=cross_attn_past_key_value, + output_attentions=output_attentions, + ) + hidden_states = self.dropout(hidden_states) + hidden_states = residual + hidden_states + hidden_states = self.encoder_attn_layer_norm(hidden_states) + + # add cross-attn to positions 3,4 of present_key_value tuple + present_key_value = present_key_value + cross_attn_present_key_value + + # Fully Connected + hidden_states = hidden_states + self.feed_forward(hidden_states) + hidden_states = self.final_layer_norm(hidden_states) + + outputs = (hidden_states,) + + if output_attentions: + outputs += (self_attn_weights, cross_attn_weights) + + if use_cache: + outputs += (present_key_value,) + + return outputs + + +class SpeechT5PreTrainedModel(PreTrainedModel): + """ + An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained + models. + """ + + config_class = SpeechT5Config + base_model_prefix = "speecht5" + main_input_name = "input_values" + supports_gradient_checkpointing = True + + _keys_to_ignore_on_load_missing = [r"position_ids"] + + def _init_weights(self, module): + """Initialize the weights""" + if isinstance(module, SpeechT5PositionalConvEmbedding): + nn.init.normal_( + module.conv.weight, + mean=0, + std=2 * math.sqrt(1 / (module.conv.kernel_size[0] * module.conv.in_channels)), + ) + nn.init.constant_(module.conv.bias, 0) + elif isinstance(module, SpeechT5FeatureProjection): + k = math.sqrt(1 / module.projection.in_features) + nn.init.uniform_(module.projection.weight, a=-k, b=k) + nn.init.uniform_(module.projection.bias, a=-k, b=k) + elif 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_() + + def _set_gradient_checkpointing(self, module, value=False): + if isinstance(module, (SpeechT5Encoder, SpeechT5Decoder, SpeechT5FeatureEncoder)): + module.gradient_checkpointing = value + + +class SpeechT5Encoder(SpeechT5PreTrainedModel): + """ + Transformer encoder consisting of *config.encoder_layers* layers. Each layer is a [`SpeechT5EncoderLayer`]. + """ + + def __init__(self, config: SpeechT5Config): + super().__init__(config) + self.layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps) + self.dropout = nn.Dropout(config.hidden_dropout) + self.layerdrop = config.encoder_layerdrop + + self.layers = nn.ModuleList([SpeechT5EncoderLayer(config) for _ in range(config.encoder_layers)]) + + self.embed_positions = SpeechT5RelativePositionalEncoding( + config.hidden_size // config.encoder_attention_heads, config.encoder_max_relative_position + ) + + self.gradient_checkpointing = False + + # Initialize weights and apply final processing + self.post_init() + + def forward( + self, + hidden_states: torch.FloatTensor, + attention_mask: Optional[torch.Tensor] = None, + head_mask: Optional[torch.Tensor] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + ) -> Union[Tuple, BaseModelOutput]: + """ + Args: + hidden_states (`torch.FloatTensor` of shape `(batch_size, sequence_length, feature_size)`): + Features extracted from the speech or text input by the encoder prenet. + attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*): + Mask to avoid performing convolution and attention on padding token indices. Mask values selected in + `[0, 1]`: + + - 1 for tokens that are **not masked**, + - 0 for tokens that are **masked**. + + [What are attention masks?](../glossary#attention-mask) + output_attentions (`bool`, *optional*): + Whether or not to return the attentions tensors of all attention layers. See `attentions` under + returned tensors for more detail. + head_mask (`torch.Tensor` of shape `(encoder_layers, encoder_attention_heads)`, *optional*): + Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`: + + - 1 indicates the head is **not masked**, + - 0 indicates the head is **masked**. + + output_hidden_states (`bool`, *optional*): + Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors + for more detail. + return_dict (`bool`, *optional*): + Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. + """ + output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions + output_hidden_states = ( + output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states + ) + return_dict = return_dict if return_dict is not None else self.config.use_return_dict + + # expand attention_mask + if attention_mask is not None: + # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len] + attention_mask = _expand_mask(attention_mask, hidden_states.dtype) + + hidden_states = self.layer_norm(hidden_states) + hidden_states = self.dropout(hidden_states) + + position_bias = self.embed_positions(hidden_states) + + deepspeed_zero3_is_enabled = is_deepspeed_zero3_enabled() + + all_hidden_states = () if output_hidden_states else None + all_self_attentions = () if output_attentions else None + + # check if head_mask has a correct number of layers specified if desired + if head_mask is not None: + if head_mask.size()[0] != len(self.layers): + raise ValueError( + f"The head_mask should be specified for {len(self.layers)} layers, but it is for" + f" {head_mask.size()[0]}." + ) + + for idx, encoder_layer in enumerate(self.layers): + if output_hidden_states: + all_hidden_states = all_hidden_states + (hidden_states,) + + # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description) + dropout_probability = np.random.uniform(0, 1) + + skip_the_layer = self.training and (dropout_probability < self.layerdrop) + if not skip_the_layer or deepspeed_zero3_is_enabled: + # under deepspeed zero3 all gpus must run in sync + if self.gradient_checkpointing and self.training: + # create gradient checkpointing function + def create_custom_forward(module): + def custom_forward(*inputs): + return module(*inputs, output_attentions) + + return custom_forward + + layer_outputs = torch.utils.checkpoint.checkpoint( + create_custom_forward(encoder_layer), + hidden_states, + attention_mask, + (head_mask[idx] if head_mask is not None else None), + position_bias, + ) + else: + layer_outputs = encoder_layer( + hidden_states, + attention_mask=attention_mask, + position_bias=position_bias, + layer_head_mask=(head_mask[idx] if head_mask is not None else None), + output_attentions=output_attentions, + ) + hidden_states = layer_outputs[0] + + if skip_the_layer: + layer_outputs = (None, None) + + if output_attentions: + all_self_attentions = all_self_attentions + (layer_outputs[1],) + + if output_hidden_states: + all_hidden_states = all_hidden_states + (hidden_states,) + + if not return_dict: + return tuple(v for v in [hidden_states, all_hidden_states, all_self_attentions] if v is not None) + + return BaseModelOutput( + last_hidden_state=hidden_states, + hidden_states=all_hidden_states, + attentions=all_self_attentions, + ) + + +class SpeechT5EncoderWithSpeechPrenet(SpeechT5PreTrainedModel): + """ + Wrapper around SpeechT5Encoder that applies SpeechT5SpeechEncoderPrenet to convert the audio waveform data to + hidden features. + """ + + def __init__(self, config: SpeechT5Config): + super().__init__(config) + self.prenet = SpeechT5SpeechEncoderPrenet(config) + self.wrapped_encoder = SpeechT5Encoder(config) + self.gradient_checkpointing = False + + # Initialize weights and apply final processing + self.post_init() + + def forward( + self, + input_values: torch.FloatTensor, + attention_mask: Optional[torch.Tensor] = None, + head_mask: Optional[torch.Tensor] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + ) -> Union[Tuple, BaseModelOutput]: + + hidden_states, attention_mask = self.prenet(input_values, attention_mask) + + outputs = self.wrapped_encoder( + hidden_states=hidden_states, + attention_mask=attention_mask, + head_mask=head_mask, + output_attentions=output_attentions, + output_hidden_states=output_hidden_states, + return_dict=return_dict, + ) + + return outputs + + +class SpeechT5EncoderWithTextPrenet(SpeechT5PreTrainedModel): + """ + Wrapper around SpeechT5Encoder that applies SpeechT5TextEncoderPrenet to convert the input_ids to hidden features. + """ + + def __init__(self, config: SpeechT5Config): + super().__init__(config) + self.prenet = SpeechT5TextEncoderPrenet(config) + self.wrapped_encoder = SpeechT5Encoder(config) + self.gradient_checkpointing = False + + # Initialize weights and apply final processing + self.post_init() + + def get_input_embeddings(self): + return self.prenet.get_input_embeddings() + + def set_input_embeddings(self, value): + self.prenet.set_input_embeddings(value) + + def forward( + self, + input_values: torch.FloatTensor, + attention_mask: Optional[torch.Tensor] = None, + head_mask: Optional[torch.Tensor] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + ) -> Union[Tuple, BaseModelOutput]: + + hidden_states = self.prenet(input_values) + + outputs = self.wrapped_encoder( + hidden_states=hidden_states, + attention_mask=attention_mask, + head_mask=head_mask, + output_attentions=output_attentions, + output_hidden_states=output_hidden_states, + return_dict=return_dict, + ) + + return outputs + + +class SpeechT5EncoderWithoutPrenet(SpeechT5PreTrainedModel): + """ + This wrapper class is a helper class to correctly load pretrained checkpoints when used in combination with + [`SpeechT5Model`]. + """ + + def __init__(self, config: SpeechT5Config): + super().__init__(config) + self.wrapped_encoder = SpeechT5Encoder(config) + self.gradient_checkpointing = False + + # Initialize weights and apply final processing + self.post_init() + + def forward( + self, + input_values: torch.FloatTensor, + attention_mask: Optional[torch.Tensor] = None, + head_mask: Optional[torch.Tensor] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + ) -> Union[Tuple, BaseModelOutput]: + + return self.wrapped_encoder( + hidden_states=input_values, + attention_mask=attention_mask, + head_mask=head_mask, + output_attentions=output_attentions, + output_hidden_states=output_hidden_states, + return_dict=return_dict, + ) + + +class SpeechT5Decoder(SpeechT5PreTrainedModel): + """ + Transformer decoder consisting of *config.decoder_layers* layers. Each layer is a [`SpeechT5DecoderLayer`] + """ + + def __init__(self, config: SpeechT5Config): + super().__init__(config) + self.layerdrop = config.decoder_layerdrop + + self.layers = nn.ModuleList([SpeechT5DecoderLayer(config) for _ in range(config.decoder_layers)]) + + self.gradient_checkpointing = False + + # Initialize weights and apply final processing + self.post_init() + + # Copied from transformers.models.bart.modeling_bart.BartDecoder._prepare_decoder_attention_mask + def _prepare_decoder_attention_mask(self, attention_mask, input_shape, inputs_embeds, past_key_values_length): + # create causal mask + # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len] + combined_attention_mask = None + if input_shape[-1] > 1: + combined_attention_mask = _make_causal_mask( + input_shape, inputs_embeds.dtype, past_key_values_length=past_key_values_length + ).to(inputs_embeds.device) + + if attention_mask is not None: + # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len] + expanded_attn_mask = _expand_mask(attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]).to( + inputs_embeds.device + ) + combined_attention_mask = ( + expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask + combined_attention_mask + ) + + return combined_attention_mask + + def forward( + self, + hidden_states: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.Tensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + encoder_attention_mask: Optional[torch.LongTensor] = None, + head_mask: Optional[torch.Tensor] = None, + cross_attn_head_mask: Optional[torch.Tensor] = None, + past_key_values: Optional[List[torch.FloatTensor]] = None, + use_cache: Optional[bool] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + ) -> Union[Tuple, BaseModelOutputWithPastAndCrossAttentions]: + r""" + Args: + hidden_states (`torch.FloatTensor` of shape `(batch_size, sequence_length, feature_size)`): + Features extracted from the speech or text input by the decoder prenet. + attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*): + Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`: + + - 1 for tokens that are **not masked**, + - 0 for tokens that are **masked**. + + [What are attention masks?](../glossary#attention-mask) + encoder_hidden_states (`torch.FloatTensor` of shape `(batch_size, encoder_sequence_length, hidden_size)`, *optional*): + Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention + of the decoder. + encoder_attention_mask (`torch.LongTensor` of shape `(batch_size, encoder_sequence_length)`, *optional*): + Mask to avoid performing cross-attention on padding tokens indices of encoder input_ids. Mask values + selected in `[0, 1]`: + + - 1 for tokens that are **not masked**, + - 0 for tokens that are **masked**. + + [What are attention masks?](../glossary#attention-mask) + head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*): + Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`: + + - 1 indicates the head is **not masked**, + - 0 indicates the head is **masked**. + + cross_attn_head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*): + Mask to nullify selected heads of the cross-attention modules in the decoder to avoid performing + cross-attention on hidden heads. Mask values selected in `[0, 1]`: + + - 1 indicates the head is **not masked**, + - 0 indicates the head is **masked**. + + past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`): + Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of + shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of + shape `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`. + + Contains pre-computed hidden-states (key and values in the self-attention blocks and in the + cross-attention blocks) that can be used (see `past_key_values` input) to speed up sequential decoding. + + If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those + that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of + all `decoder_input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`torch.FloatTensor` of + shape `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing + `input_ids` you can choose to directly pass an embedded representation. This is useful if you want more + control over how to convert `input_ids` indices into associated vectors than the model's internal + embedding lookup matrix. + output_attentions (`bool`, *optional*): + Whether or not to return the attentions tensors of all attention layers. See `attentions` under + returned tensors for more detail. + output_hidden_states (`bool`, *optional*): + Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors + for more detail. + return_dict (`bool`, *optional*): + Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. + """ + output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions + output_hidden_states = ( + output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states + ) + use_cache = use_cache if use_cache is not None else self.config.use_cache + return_dict = return_dict if return_dict is not None else self.config.use_return_dict + + input_shape = hidden_states.size()[:-1] + + past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0 + + attention_mask = self._prepare_decoder_attention_mask( + attention_mask, input_shape, hidden_states, past_key_values_length + ) + + # expand encoder attention mask + if encoder_hidden_states is not None and encoder_attention_mask is not None: + # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len] + encoder_attention_mask = _expand_mask(encoder_attention_mask, hidden_states.dtype, tgt_len=input_shape[-1]) + + deepspeed_zero3_is_enabled = is_deepspeed_zero3_enabled() + + # decoder layers + all_hidden_states = () if output_hidden_states else None + all_self_attentions = () if output_attentions else None + all_cross_attentions = () if (output_attentions and encoder_hidden_states is not None) else None + next_decoder_cache = () if use_cache else None + + # check if head_mask/cross_attn_head_mask has a correct number of layers specified if desired + for attn_mask, mask_name in zip([head_mask, cross_attn_head_mask], ["head_mask", "cross_attn_head_mask"]): + if attn_mask is not None: + if attn_mask.size()[0] != (len(self.layers)): + raise ValueError( + f"The `{mask_name}` should be specified for {len(self.layers)} layers, but it is for" + f" {head_mask.size()[0]}." + ) + + for idx, decoder_layer in enumerate(self.layers): + if output_hidden_states: + all_hidden_states = all_hidden_states + (hidden_states,) + + # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description) + dropout_probability = random.uniform(0, 1) + + skip_the_layer = self.training and (dropout_probability < self.layerdrop) + if skip_the_layer and not deepspeed_zero3_is_enabled: + continue + + past_key_value = past_key_values[idx] if past_key_values is not None else None + + if self.gradient_checkpointing and self.training: + if use_cache: + logger.warning( + "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..." + ) + use_cache = False + + def create_custom_forward(module): + def custom_forward(*inputs): + # None for past_key_value + return module(*inputs, output_attentions, use_cache) + + return custom_forward + + layer_outputs = torch.utils.checkpoint.checkpoint( + create_custom_forward(decoder_layer), + hidden_states, + attention_mask, + encoder_hidden_states, + encoder_attention_mask, + head_mask[idx] if head_mask is not None else None, + cross_attn_head_mask[idx] if cross_attn_head_mask is not None else None, + None, + ) + else: + layer_outputs = decoder_layer( + hidden_states, + attention_mask=attention_mask, + encoder_hidden_states=encoder_hidden_states, + encoder_attention_mask=encoder_attention_mask, + layer_head_mask=(head_mask[idx] if head_mask is not None else None), + cross_attn_layer_head_mask=( + cross_attn_head_mask[idx] if cross_attn_head_mask is not None else None + ), + past_key_value=past_key_value, + output_attentions=output_attentions, + use_cache=use_cache, + ) + hidden_states = layer_outputs[0] + + if use_cache: + next_decoder_cache += (layer_outputs[3 if output_attentions else 1],) + + if output_attentions: + all_self_attentions = all_self_attentions + (layer_outputs[1],) + + if encoder_hidden_states is not None: + all_cross_attentions = all_cross_attentions + (layer_outputs[2],) + + if output_hidden_states: + all_hidden_states = all_hidden_states + (hidden_states,) + + next_cache = next_decoder_cache if use_cache else None + if not return_dict: + return tuple( + v + for v in [hidden_states, next_cache, all_hidden_states, all_self_attentions, all_cross_attentions] + if v is not None + ) + + return BaseModelOutputWithPastAndCrossAttentions( + last_hidden_state=hidden_states, + past_key_values=next_cache, + hidden_states=all_hidden_states, + attentions=all_self_attentions, + cross_attentions=all_cross_attentions, + ) + + +class SpeechT5DecoderWithSpeechPrenet(SpeechT5PreTrainedModel): + """ + Wrapper around SpeechT5Decoder that applies SpeechT5SpeechDecoderPrenet to convert log-mel filterbanks to hidden + features. + """ + + def __init__(self, config: SpeechT5Config): + super().__init__(config) + self.prenet = SpeechT5SpeechDecoderPrenet(config) + self.wrapped_decoder = SpeechT5Decoder(config) + self.gradient_checkpointing = False + + # Initialize weights and apply final processing + self.post_init() + + def forward( + self, + input_values: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.Tensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + encoder_attention_mask: Optional[torch.LongTensor] = None, + speaker_embeddings: Optional[torch.Tensor] = None, + head_mask: Optional[torch.Tensor] = None, + cross_attn_head_mask: Optional[torch.Tensor] = None, + past_key_values: Optional[List[torch.FloatTensor]] = None, + use_cache: Optional[bool] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + ) -> Union[Tuple, BaseModelOutputWithPastAndCrossAttentions]: + + decoder_hidden_states = self.prenet(input_values, speaker_embeddings) + + outputs = self.wrapped_decoder( + hidden_states=decoder_hidden_states, + attention_mask=attention_mask, + encoder_hidden_states=encoder_hidden_states, + encoder_attention_mask=encoder_attention_mask, + head_mask=head_mask, + cross_attn_head_mask=cross_attn_head_mask, + past_key_values=past_key_values, + use_cache=use_cache, + output_attentions=output_attentions, + output_hidden_states=output_hidden_states, + return_dict=return_dict, + ) + + return outputs + + +class SpeechT5DecoderWithTextPrenet(SpeechT5PreTrainedModel): + """ + Wrapper around SpeechT5Decoder that applies SpeechT5TextDecoderPrenet to convert input tokens to hidden features. + """ + + def __init__(self, config: SpeechT5Config): + super().__init__(config) + self.prenet = SpeechT5TextDecoderPrenet(config) + self.wrapped_decoder = SpeechT5Decoder(config) + self.gradient_checkpointing = False + + # Initialize weights and apply final processing + self.post_init() + + def get_input_embeddings(self): + return self.prenet.get_input_embeddings() + + def set_input_embeddings(self, value): + self.prenet.set_input_embeddings(value) + + def forward( + self, + input_values: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.Tensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + encoder_attention_mask: Optional[torch.LongTensor] = None, + head_mask: Optional[torch.Tensor] = None, + cross_attn_head_mask: Optional[torch.Tensor] = None, + past_key_values: Optional[List[torch.FloatTensor]] = None, + use_cache: Optional[bool] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + ) -> Union[Tuple, BaseModelOutputWithPastAndCrossAttentions]: + + decoder_hidden_states, attention_mask = self.prenet(input_values, attention_mask, past_key_values) + + outputs = self.wrapped_decoder( + hidden_states=decoder_hidden_states, + attention_mask=attention_mask, + encoder_hidden_states=encoder_hidden_states, + encoder_attention_mask=encoder_attention_mask, + head_mask=head_mask, + cross_attn_head_mask=cross_attn_head_mask, + past_key_values=past_key_values, + use_cache=use_cache, + output_attentions=output_attentions, + output_hidden_states=output_hidden_states, + return_dict=return_dict, + ) + + return outputs + + +class SpeechT5DecoderWithoutPrenet(SpeechT5PreTrainedModel): + """ + This wrapper class is a helper class to correctly load pretrained checkpoints when used in combination with + [`SpeechT5Model`]. + """ + + def __init__(self, config: SpeechT5Config): + super().__init__(config) + self.wrapped_decoder = SpeechT5Decoder(config) + self.gradient_checkpointing = False + + # Initialize weights and apply final processing + self.post_init() + + def forward( + self, + input_values: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.Tensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + encoder_attention_mask: Optional[torch.LongTensor] = None, + head_mask: Optional[torch.Tensor] = None, + cross_attn_head_mask: Optional[torch.Tensor] = None, + past_key_values: Optional[List[torch.FloatTensor]] = None, + use_cache: Optional[bool] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + ) -> Union[Tuple, BaseModelOutputWithPastAndCrossAttentions]: + + outputs = self.wrapped_decoder( + hidden_states=input_values, + attention_mask=attention_mask, + encoder_hidden_states=encoder_hidden_states, + encoder_attention_mask=encoder_attention_mask, + head_mask=head_mask, + cross_attn_head_mask=cross_attn_head_mask, + past_key_values=past_key_values, + use_cache=use_cache, + output_attentions=output_attentions, + output_hidden_states=output_hidden_states, + return_dict=return_dict, + ) + return outputs + + +SPEECHT5_BASE_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 ([`SpeechT5Config`]): + 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. + encoder ([`SpeechT5EncoderWithSpeechPrenet`] or [`SpeechT5EncoderWithTextPrenet`] or `None`): + The Transformer encoder module that applies the appropiate speech or text encoder prenet. If `None`, + [`SpeechT5EncoderWithoutPrenet`] will be used and the `input_values` are assumed to be hidden states. + decoder ([`SpeechT5DecoderWithSpeechPrenet`] or [`SpeechT5DecoderWithTextPrenet`] or `None`): + The Transformer decoder module that applies the appropiate speech or text decoder prenet. If `None`, + [`SpeechT5DecoderWithoutPrenet`] will be used and the `decoder_input_values` are assumed to be hidden + states. +""" + + +SPEECHT5_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 ([`SpeechT5Config`]): + 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. +""" + + +SPEECHT5_INPUTS_DOCSTRING = r""" + Args: + attention_mask (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): + Mask to avoid performing convolution and attention on padding token indices. Mask values selected in `[0, + 1]`: + + - 1 for tokens that are **not masked**, + - 0 for tokens that are **masked**. + + [What are attention masks?](../glossary#attention-mask) + + + + `attention_mask` should only be passed if the corresponding processor has `config.return_attention_mask == + True`. For all models whose processor has `config.return_attention_mask == False`, `attention_mask` should + **not** be passed to avoid degraded performance when doing batched inference. For such models + `input_values` should simply be padded with 0 and passed without `attention_mask`. Be aware that these + models also yield slightly different results depending on whether `input_values` is padded or not. + + + + decoder_attention_mask (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*): + Default behavior: generate a tensor that ignores pad tokens in `decoder_input_values`. Causal mask will + also be used by default. + + If you want to change padding behavior, you should read [`SpeechT5Decoder._prepare_decoder_attention_mask`] + and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more + information on the default strategy. + + head_mask (`torch.Tensor` of shape `(encoder_layers, encoder_attention_heads)`, *optional*): + Mask to nullify selected heads of the attention modules in the encoder. Mask values selected in `[0, 1]`: + + - 1 indicates the head is **not masked**, + - 0 indicates the head is **masked**. + + decoder_head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*): + Mask to nullify selected heads of the attention modules in the decoder. Mask values selected in `[0, 1]`: + + - 1 indicates the head is **not masked**, + - 0 indicates the head is **masked**. + + cross_attn_head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*): + Mask to nullify selected heads of the cross-attention modules. Mask values selected in `[0, 1]`: + + - 1 indicates the head is **not masked**, + - 0 indicates the head is **masked**. + + encoder_outputs (`tuple(tuple(torch.FloatTensor)`, *optional*): + Tuple consists of (`last_hidden_state`, *optional*: `hidden_states`, *optional*: `attentions`) + `last_hidden_state` of shape `(batch_size, sequence_length, hidden_size)`, *optional*) is a sequence of + hidden-states at the output of the last layer of the encoder. Used in the cross-attention of the decoder. + + past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`): + Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape + `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape + `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`. + + Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention + blocks) that can be used (see `past_key_values` input) to speed up sequential decoding. + + If `past_key_values` are used, the user can optionally input only the last `decoder_input_values` (those + that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all + `decoder_input_values` of shape `(batch_size, sequence_length)`. decoder_inputs_embeds (`torch.FloatTensor` + of shape `(batch_size, target_sequence_length, hidden_size)`, *optional*): Optionally, instead of passing + `decoder_input_values` you can choose to directly pass an embedded representation. If `past_key_values` is + used, optionally only the last `decoder_inputs_embeds` have to be input (see `past_key_values`). This is + useful if you want more control over how to convert `decoder_input_values` indices into associated vectors + than the model's internal embedding lookup matrix. + + use_cache (`bool`, *optional*): + If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see + `past_key_values`). + + output_attentions (`bool`, *optional*): + Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned + tensors for more detail. + + output_hidden_states (`bool`, *optional*): + Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for + more detail. + + return_dict (`bool`, *optional*): + Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. +""" + + +@add_start_docstrings( + "The bare SpeechT5 Encoder-Decoder Model outputting raw hidden-states without any specific pre- or post-nets.", + SPEECHT5_BASE_START_DOCSTRING, +) +class SpeechT5Model(SpeechT5PreTrainedModel): + def __init__( + self, + config: SpeechT5Config, + encoder: Optional[nn.Module] = None, + decoder: Optional[nn.Module] = None, + ): + super().__init__(config) + self.config = config + self.encoder = SpeechT5EncoderWithoutPrenet(config) if encoder is None else encoder + self.decoder = SpeechT5DecoderWithoutPrenet(config) if decoder is None else decoder + + # Initialize weights and apply final processing + self.post_init() + + def get_input_embeddings(self): + if isinstance(self.encoder, SpeechT5EncoderWithTextPrenet): + return self.encoder.get_input_embeddings() + if isinstance(self.decoder, SpeechT5DecoderWithTextPrenet): + return self.decoder.get_input_embeddings() + return None + + def set_input_embeddings(self, value): + if isinstance(self.encoder, SpeechT5EncoderWithTextPrenet): + self.encoder.set_input_embeddings(value) + if isinstance(self.decoder, SpeechT5DecoderWithTextPrenet): + self.decoder.set_input_embeddings(value) + + def get_encoder(self): + return self.encoder + + def get_decoder(self): + return self.decoder + + def freeze_feature_encoder(self): + """ + Calling this function will disable the gradient computation for the feature encoder so that its parameter will + not be updated during training. + """ + if isinstance(self.encoder, SpeechT5EncoderWithSpeechPrenet): + self.encoder.prenet.freeze_feature_encoder() + + @add_start_docstrings_to_model_forward(SPEECHT5_INPUTS_DOCSTRING) + @replace_return_docstrings(output_type=Seq2SeqModelOutput, config_class=_CONFIG_FOR_DOC) + def forward( + self, + input_values: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.Tensor] = None, + decoder_input_values: Optional[torch.Tensor] = None, + decoder_attention_mask: Optional[torch.BoolTensor] = None, + head_mask: Optional[torch.FloatTensor] = None, + decoder_head_mask: Optional[torch.FloatTensor] = None, + cross_attn_head_mask: Optional[torch.Tensor] = None, + encoder_outputs: Optional[Tuple[Tuple[torch.FloatTensor]]] = None, + past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None, + use_cache: Optional[bool] = None, + speaker_embeddings: Optional[torch.Tensor] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + ) -> Union[Tuple[torch.FloatTensor], Seq2SeqModelOutput]: + r""" + input_values (`torch.FloatTensor` of shape `(batch_size, sequence_length)`): + Depending on which encoder is being used, the `input_values` are either: float values of the input raw + speech waveform, or indices of input sequence tokens in the vocabulary, or hidden states. + + decoder_input_values (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*): + Depending on which decoder is being used, the `decoder_input_values` are either: float values of log-mel + filterbank features extracted from the raw speech waveform, or indices of decoder input sequence tokens in + the vocabulary, or hidden states. + + speaker_embeddings (`torch.FloatTensor` of shape `(batch_size, config.speaker_embedding_dim)`, *optional*): + Tensor containing the speaker embeddings. + + Returns: + """ + output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions + output_hidden_states = ( + output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states + ) + use_cache = use_cache if use_cache is not None else self.config.use_cache + return_dict = return_dict if return_dict is not None else self.config.use_return_dict + + # Encode if needed (training, first prediction pass) + if encoder_outputs is None: + encoder_outputs = self.encoder( + input_values=input_values, + attention_mask=attention_mask, + head_mask=head_mask, + output_attentions=output_attentions, + output_hidden_states=output_hidden_states, + return_dict=return_dict, + ) + # If the user passed a tuple for encoder_outputs, we wrap it in a BaseModelOutput when return_dict=True + elif return_dict and not isinstance(encoder_outputs, BaseModelOutput): + encoder_outputs = BaseModelOutput( + last_hidden_state=encoder_outputs[0], + hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None, + attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None, + ) + + # downsample encoder attention mask (only for encoders with speech input) + if attention_mask is not None and isinstance(self.encoder, SpeechT5EncoderWithSpeechPrenet): + encoder_attention_mask = self.encoder.prenet._get_feature_vector_attention_mask( + encoder_outputs[0].shape[1], attention_mask + ) + else: + encoder_attention_mask = attention_mask + + if isinstance(self.decoder, SpeechT5DecoderWithSpeechPrenet): + decoder_args = {"speaker_embeddings": speaker_embeddings} + else: + decoder_args = {} + + decoder_outputs = self.decoder( + input_values=decoder_input_values, + attention_mask=decoder_attention_mask, + encoder_hidden_states=encoder_outputs[0], + encoder_attention_mask=encoder_attention_mask, + head_mask=decoder_head_mask, + cross_attn_head_mask=cross_attn_head_mask, + past_key_values=past_key_values, + use_cache=use_cache, + output_attentions=output_attentions, + output_hidden_states=output_hidden_states, + return_dict=return_dict, + **decoder_args, + ) + + if not return_dict: + return decoder_outputs + encoder_outputs + + return Seq2SeqModelOutput( + last_hidden_state=decoder_outputs.last_hidden_state, + past_key_values=decoder_outputs.past_key_values, + decoder_hidden_states=decoder_outputs.hidden_states, + decoder_attentions=decoder_outputs.attentions, + cross_attentions=decoder_outputs.cross_attentions, + encoder_last_hidden_state=encoder_outputs.last_hidden_state, + encoder_hidden_states=encoder_outputs.hidden_states, + encoder_attentions=encoder_outputs.attentions, + ) + + +@add_start_docstrings( + """SpeechT5 Model with a speech encoder and a text decoder.""", + SPEECHT5_START_DOCSTRING, +) +class SpeechT5ForSpeechToText(SpeechT5PreTrainedModel): + _keys_to_ignore_on_load_missing = [ + r"speecht5.encoder.prenet.pos_sinusoidal_embed.weights", + r"text_decoder_postnet.lm_head.weight", + ] + _keys_to_ignore_on_save = [ + r"speecht5.encoder.prenet.pos_sinusoidal_embed.weights", + ] + + def __init__(self, config: SpeechT5Config): + super().__init__(config) + + if config.vocab_size is None: + raise ValueError( + f"You are trying to instantiate {self.__class__} with a configuration that does not define the" + " vocabulary size of the language model head. Please instantiate the model as follows:" + " `SpeechT5ForSpeechToText.from_pretrained(..., vocab_size=vocab_size)`. or define `vocab_size` of" + " your model's configuration." + ) + + speech_encoder = SpeechT5EncoderWithSpeechPrenet(config) + text_decoder = SpeechT5DecoderWithTextPrenet(config) + self.speecht5 = SpeechT5Model(config, speech_encoder, text_decoder) + + self.text_decoder_postnet = SpeechT5TextDecoderPostnet(config) + + # Initialize weights and apply final processing + self.post_init() + + def get_encoder(self): + return self.speecht5.get_encoder() + + def get_decoder(self): + return self.speecht5.get_decoder() + + def freeze_feature_encoder(self): + """ + Calling this function will disable the gradient computation for the feature encoder so that its parameter will + not be updated during training. + """ + self.get_encoder().prenet.freeze_feature_encoder() + + def resize_token_embeddings(self, new_num_tokens: int) -> nn.Embedding: + new_embeddings = super().resize_token_embeddings(new_num_tokens) + return new_embeddings + + def get_output_embeddings(self): + return self.text_decoder_postnet.get_output_embeddings() + + def set_output_embeddings(self, new_embeddings): + self.text_decoder_postnet.set_output_embeddings(new_embeddings) + + @add_start_docstrings_to_model_forward(SPEECHT5_INPUTS_DOCSTRING) + @replace_return_docstrings(output_type=Seq2SeqLMOutput, config_class=_CONFIG_FOR_DOC) + def forward( + self, + input_values: Optional[torch.Tensor] = None, + attention_mask: Optional[torch.Tensor] = None, + decoder_input_ids: Optional[torch.Tensor] = None, + decoder_attention_mask: Optional[torch.BoolTensor] = None, + head_mask: Optional[torch.FloatTensor] = None, + decoder_head_mask: Optional[torch.FloatTensor] = None, + cross_attn_head_mask: Optional[torch.Tensor] = None, + encoder_outputs: Optional[Tuple[Tuple[torch.FloatTensor]]] = None, + past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None, + use_cache: Optional[bool] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + labels: Optional[torch.Tensor] = None, + ) -> Union[Tuple, Seq2SeqLMOutput]: + r""" + input_values (`torch.FloatTensor` of shape `(batch_size, sequence_length)`): + Float values of input raw speech waveform. Values can be obtained by loading a *.flac* or *.wav* audio file + into an array of type `List[float]` or a `numpy.ndarray`, *e.g.* via the soundfile library (*pip install + soundfile*). To prepare the array into `input_values`, the [`SpeechT5Processor`] should be used for padding + and conversion into a tensor of type `torch.FloatTensor`. See [`SpeechT5Processor.__call__`] for details. + + decoder_input_ids (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*): + Indices of decoder input sequence tokens in the vocabulary. + + Indices can be obtained using [`SpeechT5Tokenizer`]. See [`PreTrainedTokenizer.encode`] and + [`PreTrainedTokenizer.__call__`] for details. + + [What are decoder input IDs?](../glossary#decoder-input-ids) + + SpeechT5 uses the `eos_token_id` as the starting token for `decoder_input_ids` generation. If + `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see + `past_key_values`). + + labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): + Labels for computing the language modeling loss. Indices should either be in `[0, ..., config.vocab_size]` + or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored (masked), the loss is + only computed for the tokens with labels in `[0, ..., config.vocab_size]`. + + Label indices can be obtained using [`SpeechT5Tokenizer`]. See [`PreTrainedTokenizer.encode`] and + [`PreTrainedTokenizer.__call__`] for details. + + Returns: + + Example: + + ```python + >>> from transformers import SpeechT5Processor, SpeechT5ForSpeechToText + >>> from datasets import load_dataset + + >>> dataset = load_dataset("hf-internal-testing/librispeech_asr_demo", "clean", split="validation") + >>> dataset = dataset.sort("id") + >>> sampling_rate = dataset.features["audio"].sampling_rate + + >>> processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_asr") + >>> model = SpeechT5ForSpeechToText.from_pretrained("microsoft/speecht5_asr") + + >>> # audio file is decoded on the fly + >>> inputs = processor(audio=dataset[0]["audio"]["array"], sampling_rate=sampling_rate, return_tensors="pt") + >>> predicted_ids = model.generate(**inputs, max_length=100) + + >>> # transcribe speech + >>> transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True) + >>> transcription[0] + 'mister quilter is the apostle of the middle classes and we are glad to welcome his gospel' + ``` + + ```python + >>> inputs["labels"] = processor(text_target=dataset[0]["text"], return_tensors="pt").input_ids + + >>> # compute loss + >>> loss = model(**inputs).loss + >>> round(loss.item(), 2) + 19.88 + ``` + """ + return_dict = return_dict if return_dict is not None else self.config.use_return_dict + + if labels is not None: + if decoder_input_ids is None: + decoder_input_ids = shift_tokens_right( + labels, self.config.pad_token_id, self.config.decoder_start_token_id + ) + + outputs = self.speecht5( + input_values=input_values, + attention_mask=attention_mask, + decoder_input_values=decoder_input_ids, + decoder_attention_mask=decoder_attention_mask, + head_mask=head_mask, + decoder_head_mask=decoder_head_mask, + cross_attn_head_mask=cross_attn_head_mask, + encoder_outputs=encoder_outputs, + past_key_values=past_key_values, + use_cache=use_cache, + output_attentions=output_attentions, + output_hidden_states=output_hidden_states, + return_dict=True, + ) + + logits = self.text_decoder_postnet(outputs[0]) + + loss = None + if labels is not None: + loss_fct = CrossEntropyLoss() + loss = loss_fct(logits.view(-1, self.config.vocab_size), labels.view(-1)) + + if not return_dict: + output = (logits,) + outputs[1:] + return ((loss,) + output) if loss is not None else output + + return Seq2SeqLMOutput( + loss=loss, + logits=logits, + past_key_values=outputs.past_key_values, + decoder_hidden_states=outputs.decoder_hidden_states, + decoder_attentions=outputs.decoder_attentions, + cross_attentions=outputs.cross_attentions, + encoder_last_hidden_state=outputs.encoder_last_hidden_state, + encoder_hidden_states=outputs.encoder_hidden_states, + encoder_attentions=outputs.encoder_attentions, + ) + + def prepare_inputs_for_generation( + self, + decoder_input_ids, + past=None, + attention_mask=None, + head_mask=None, + decoder_head_mask=None, + cross_attn_head_mask=None, + use_cache=None, + encoder_outputs=None, + **kwargs + ): + # cut decoder_input_ids if past is used + if past is not None: + decoder_input_ids = decoder_input_ids[:, -1:] + + return { + "encoder_outputs": encoder_outputs, + "past_key_values": past, + "decoder_input_ids": decoder_input_ids, + "attention_mask": attention_mask, + "head_mask": head_mask, + "decoder_head_mask": decoder_head_mask, + "cross_attn_head_mask": cross_attn_head_mask, + "use_cache": use_cache, # change this to avoid caching (presumably for debugging) + } + + @staticmethod + def _reorder_cache(past, beam_idx): + reordered_past = () + for layer_past in past: + reordered_past += (tuple(past_state.index_select(0, beam_idx) for past_state in layer_past),) + return reordered_past + + +def _generate_speech( + model: SpeechT5PreTrainedModel, + input_values: torch.FloatTensor, + speaker_embeddings: Optional[torch.FloatTensor] = None, + threshold: float = 0.5, + minlenratio: float = 0.0, + maxlenratio: float = 20.0, + vocoder: Optional[nn.Module] = None, +) -> torch.FloatTensor: + encoder_attention_mask = torch.ones_like(input_values) + + encoder_out = model.speecht5.encoder( + input_values=input_values, + attention_mask=encoder_attention_mask, + return_dict=True, + ) + + encoder_last_hidden_state = encoder_out.last_hidden_state + + # downsample encoder attention mask + if isinstance(model.speecht5.encoder, SpeechT5EncoderWithSpeechPrenet): + encoder_attention_mask = model.speecht5.encoder.prenet._get_feature_vector_attention_mask( + encoder_out[0].shape[1], encoder_attention_mask + ) + + maxlen = int(encoder_last_hidden_state.size(1) * maxlenratio / model.config.reduction_factor) + minlen = int(encoder_last_hidden_state.size(1) * minlenratio / model.config.reduction_factor) + + # Start the output sequence with a mel spectrum that is all zeros. + output_sequence = encoder_last_hidden_state.new_zeros(1, 1, model.config.num_mel_bins) + + spectrogram = [] + past_key_values = None + idx = 0 + + while True: + idx += 1 + + # Run the decoder prenet on the entire output sequence. + decoder_hidden_states = model.speecht5.decoder.prenet(output_sequence, speaker_embeddings) + + # Run the decoder layers on the last element of the prenet output. + decoder_out = model.speecht5.decoder.wrapped_decoder( + hidden_states=decoder_hidden_states[:, -1:], + attention_mask=None, + encoder_hidden_states=encoder_last_hidden_state, + encoder_attention_mask=encoder_attention_mask, + past_key_values=past_key_values, + use_cache=True, + return_dict=True, + ) + + last_decoder_output = decoder_out.last_hidden_state[0, -1] + past_key_values = decoder_out.past_key_values + + # Predict the new mel spectrum for this step in the sequence. + spectrum = model.speech_decoder_postnet.feat_out(last_decoder_output) + spectrum = spectrum.view(model.config.reduction_factor, model.config.num_mel_bins) + spectrogram.append(spectrum) + + # Extend the output sequence with the new mel spectrum. + output_sequence = torch.cat((output_sequence, spectrum[-1].view(1, 1, model.config.num_mel_bins)), dim=1) + + # Predict the probability that this is the stop token. + prob = torch.sigmoid(model.speech_decoder_postnet.prob_out(last_decoder_output)) + + # Finished when stop token or maximum length is reached. + if idx >= minlen and (int(sum(prob >= threshold)) > 0 or idx >= maxlen): + spectrogram = torch.cat(spectrogram, dim=0).unsqueeze(0) + spectrogram = model.speech_decoder_postnet.postnet(spectrogram) + spectrogram = spectrogram.squeeze(0) + break + + if vocoder is not None: + return vocoder(spectrogram) + else: + return spectrogram + + +@add_start_docstrings( + """SpeechT5 Model with a text encoder and a speech decoder.""", + SPEECHT5_START_DOCSTRING, +) +class SpeechT5ForTextToSpeech(SpeechT5PreTrainedModel): + _keys_to_ignore_on_load_missing = [] + _keys_to_ignore_on_save = [] + + main_input_name = "input_ids" + + def __init__(self, config: SpeechT5Config): + super().__init__(config) + + if config.vocab_size is None: + raise ValueError( + f"You are trying to instantiate {self.__class__} with a configuration that does not define the" + " vocabulary size of the language model head. Please instantiate the model as follows:" + " `SpeechT5ForTextToSpeech.from_pretrained(..., vocab_size=vocab_size)`. or define `vocab_size` of" + " your model's configuration." + ) + + text_encoder = SpeechT5EncoderWithTextPrenet(config) + speech_decoder = SpeechT5DecoderWithSpeechPrenet(config) + self.speecht5 = SpeechT5Model(config, text_encoder, speech_decoder) + + self.speech_decoder_postnet = SpeechT5SpeechDecoderPostnet(config) + + # Initialize weights and apply final processing + self.post_init() + + def get_encoder(self): + return self.speecht5.get_encoder() + + def get_decoder(self): + return self.speecht5.get_decoder() + + @add_start_docstrings_to_model_forward(SPEECHT5_INPUTS_DOCSTRING) + @replace_return_docstrings(output_type=Seq2SeqSpectrogramOutput, config_class=_CONFIG_FOR_DOC) + def forward( + self, + input_ids: Optional[torch.Tensor] = None, + attention_mask: Optional[torch.Tensor] = None, + decoder_input_values: Optional[torch.Tensor] = None, + decoder_attention_mask: Optional[torch.BoolTensor] = None, + head_mask: Optional[torch.FloatTensor] = None, + decoder_head_mask: Optional[torch.FloatTensor] = None, + cross_attn_head_mask: Optional[torch.Tensor] = None, + encoder_outputs: Optional[Tuple[Tuple[torch.FloatTensor]]] = None, + past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None, + use_cache: Optional[bool] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + speaker_embeddings: Optional[torch.Tensor] = None, + labels: Optional[torch.Tensor] = None, + stop_labels: Optional[torch.Tensor] = None, + ) -> Union[Tuple, Seq2SeqSpectrogramOutput]: + r""" + input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`): + Indices of input sequence tokens in the vocabulary. The `batch_size` should be 1 currently. + + Indices can be obtained using [`SpeechT5Tokenizer`]. See [`~PreTrainedTokenizer.encode`] and + [`~PreTrainedTokenizer.__call__`] for details. + + [What are input IDs?](../glossary#input-ids) + decoder_input_values (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.num_mel_bins)`): + Float values of input mel spectrogram. + + SpeechT5 uses an all-zero spectrum as the starting token for `decoder_input_values` generation. If + `past_key_values` is used, optionally only the last `decoder_input_values` have to be input (see + `past_key_values`). + speaker_embeddings (`torch.FloatTensor` of shape `(batch_size, config.speaker_embedding_dim)`, *optional*): + Tensor containing the speaker embeddings. + labels (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.num_mel_bins)`, *optional*): + Float values of target mel spectrogram. Spectrograms can be obtained using [`SpeechT5Processor`]. See + [`SpeechT5Processor.__call__`] for details. + stop_labels (`torch.FloatTensor` of shape `(batch_size, unreduced_sequence_length)`, *optional*): + Labels for computing the stop token loss. Values are 0.0 until the end of the sequence, after which they + become 1.0. The sequence length of this tensor is `config.reduction_factor` times larger than the length of + the target mel spectrogram. Labels can be obtained using [`SpeechT5Processor`]. See + [`SpeechT5Processor.__call__`] for details. + + Returns: + + Example: + + ```python + >>> from transformers import SpeechT5Processor, SpeechT5ForTextToSpeech, SpeechT5HifiGan + >>> import torch + + >>> processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts") + >>> model = SpeechT5ForTextToSpeech.from_pretrained("microsoft/speecht5_tts") + >>> vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan") + + >>> inputs = processor(text="Hello, my dog is cute", return_tensors="pt") + >>> speaker_embeddings = torch.zeros((1, 512)) # or load xvectors from a file + + >>> # generate speech + >>> speech = model.generate_speech(inputs["input_ids"], speaker_embeddings, vocoder=vocoder) + >>> speech.shape + torch.Size([15872]) + ``` + """ + return_dict = return_dict if return_dict is not None else self.config.use_return_dict + + if labels is not None: + if decoder_input_values is None: + decoder_input_values = shift_spectrograms_right(labels) + + outputs = self.speecht5( + input_values=input_ids, + attention_mask=attention_mask, + decoder_input_values=decoder_input_values, + decoder_attention_mask=decoder_attention_mask, + head_mask=head_mask, + decoder_head_mask=decoder_head_mask, + cross_attn_head_mask=cross_attn_head_mask, + encoder_outputs=encoder_outputs, + past_key_values=past_key_values, + use_cache=use_cache, + speaker_embeddings=speaker_embeddings, + output_attentions=output_attentions, + output_hidden_states=output_hidden_states, + return_dict=True, + ) + + _, spectrogram, logits = self.speech_decoder_postnet(outputs[0]) + + loss = None + + if not return_dict: + output = (spectrogram,) + outputs[1:] + return ((loss,) + output) if loss is not None else output + + return Seq2SeqSpectrogramOutput( + loss=loss, + spectrogram=spectrogram, + past_key_values=outputs.past_key_values, + decoder_hidden_states=outputs.decoder_hidden_states, + decoder_attentions=outputs.decoder_attentions, + cross_attentions=outputs.cross_attentions, + encoder_last_hidden_state=outputs.encoder_last_hidden_state, + encoder_hidden_states=outputs.encoder_hidden_states, + encoder_attentions=outputs.encoder_attentions, + ) + + @torch.no_grad() + def generate_speech( + self, + input_ids: torch.LongTensor, + speaker_embeddings: Optional[torch.FloatTensor] = None, + threshold: float = 0.5, + minlenratio: float = 0.0, + maxlenratio: float = 20.0, + vocoder: Optional[nn.Module] = None, + ) -> torch.FloatTensor: + r""" + Converts a sequence of input tokens into a sequence of mel spectrograms, which are subsequently turned into a + speech waveform using a vocoder. + + Args: + input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`): + Indices of input sequence tokens in the vocabulary. The `batch_size` should be 1 currently. + + Indices can be obtained using [`SpeechT5Tokenizer`]. See [`~PreTrainedTokenizer.encode`] and + [`~PreTrainedTokenizer.__call__`] for details. + + [What are input IDs?](../glossary#input-ids) + speaker_embeddings (`torch.FloatTensor` of shape `(batch_size, config.speaker_embedding_dim)`, *optional*): + Tensor containing the speaker embeddings. + threshold (`float`, *optional*, defaults to 0.5): + The generated sequence ends when the predicted stop token probability exceeds this value. + minlenratio (`float`, *optional*, defaults to 0.0): + Used to calculate the minimum required length for the output sequence. + maxlenratio (`float`, *optional*, defaults to 20.0): + Used to calculate the maximum allowed length for the output sequence. + vocoder (`nn.Module`, *optional*, defaults to `None`): + The vocoder that converts the mel spectrogram into a speech waveform. If `None`, the output is the mel + spectrogram. + + Returns: + `torch.FloatTensor`: Tensor of shape `(output_sequence_length, config.num_mel_bins)` containing the + predicted mel spectrogram, or a tensor with shape `(num_frames,)` containing the speech waveform. + """ + return _generate_speech( + self, + input_ids, + speaker_embeddings, + threshold, + minlenratio, + maxlenratio, + vocoder, + ) + + +@add_start_docstrings( + """SpeechT5 Model with a speech encoder and a speech decoder.""", + SPEECHT5_START_DOCSTRING, +) +class SpeechT5ForSpeechToSpeech(SpeechT5PreTrainedModel): + _keys_to_ignore_on_load_missing = [ + r"speecht5.encoder.prenet.pos_sinusoidal_embed.weights", + ] + _keys_to_ignore_on_save = [ + r"speecht5.encoder.prenet.pos_sinusoidal_embed.weights", + ] + + def __init__(self, config: SpeechT5Config): + super().__init__(config) + + speech_encoder = SpeechT5EncoderWithSpeechPrenet(config) + speech_decoder = SpeechT5DecoderWithSpeechPrenet(config) + self.speecht5 = SpeechT5Model(config, speech_encoder, speech_decoder) + + self.speech_decoder_postnet = SpeechT5SpeechDecoderPostnet(config) + + # Initialize weights and apply final processing + self.post_init() + + def get_encoder(self): + return self.speecht5.get_encoder() + + def get_decoder(self): + return self.speecht5.get_decoder() + + def freeze_feature_encoder(self): + """ + Calling this function will disable the gradient computation for the feature encoder so that its parameter will + not be updated during training. + """ + self.get_encoder().prenet.freeze_feature_encoder() + + @add_start_docstrings_to_model_forward(SPEECHT5_INPUTS_DOCSTRING) + @replace_return_docstrings(output_type=Seq2SeqSpectrogramOutput, config_class=_CONFIG_FOR_DOC) + def forward( + self, + input_values: Optional[torch.Tensor] = None, + attention_mask: Optional[torch.Tensor] = None, + decoder_input_values: Optional[torch.Tensor] = None, + decoder_attention_mask: Optional[torch.BoolTensor] = None, + head_mask: Optional[torch.FloatTensor] = None, + decoder_head_mask: Optional[torch.FloatTensor] = None, + cross_attn_head_mask: Optional[torch.Tensor] = None, + encoder_outputs: Optional[Tuple[Tuple[torch.FloatTensor]]] = None, + past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None, + use_cache: Optional[bool] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + speaker_embeddings: Optional[torch.Tensor] = None, + labels: Optional[torch.Tensor] = None, + stop_labels: Optional[torch.Tensor] = None, + ) -> Union[Tuple, Seq2SeqSpectrogramOutput]: + r""" + input_values (`torch.FloatTensor` of shape `(batch_size, sequence_length)`): + Float values of input raw speech waveform. Values can be obtained by loading a *.flac* or *.wav* audio file + into an array of type `List[float]` or a `numpy.ndarray`, *e.g.* via the soundfile library (*pip install + soundfile*). To prepare the array into `input_values`, the [`SpeechT5Processor`] should be used for padding + and conversion into a tensor of type `torch.FloatTensor`. See [`SpeechT5Processor.__call__`] for details. + decoder_input_values (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.num_mel_bins)`): + Float values of input mel spectrogram. + + SpeechT5 uses an all-zero spectrum as the starting token for `decoder_input_values` generation. If + `past_key_values` is used, optionally only the last `decoder_input_values` have to be input (see + `past_key_values`). + speaker_embeddings (`torch.FloatTensor` of shape `(batch_size, config.speaker_embedding_dim)`, *optional*): + Tensor containing the speaker embeddings. + labels (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.num_mel_bins)`, *optional*): + Float values of target mel spectrogram. Spectrograms can be obtained using [`SpeechT5Processor`]. See + [`SpeechT5Processor.__call__`] for details. + stop_labels (`torch.FloatTensor` of shape `(batch_size, unreduced_sequence_length)`, *optional*): + Labels for computing the stop token loss. Values are 0.0 until the end of the sequence, after which they + become 1.0. The sequence length of this tensor is `config.reduction_factor` times larger than the length of + the target mel spectrogram. Labels can be obtained using [`SpeechT5Processor`]. See + [`SpeechT5Processor.__call__`] for details. + + Returns: + + Example: + + ```python + >>> from transformers import SpeechT5Processor, SpeechT5ForSpeechToSpeech, SpeechT5HifiGan + >>> from datasets import load_dataset + >>> import torch + + >>> dataset = load_dataset("hf-internal-testing/librispeech_asr_demo", "clean", split="validation") + >>> dataset = dataset.sort("id") + >>> sampling_rate = dataset.features["audio"].sampling_rate + + >>> processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_vc") + >>> model = SpeechT5ForSpeechToSpeech.from_pretrained("microsoft/speecht5_vc") + >>> vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan") + + >>> # audio file is decoded on the fly + >>> inputs = processor(audio=dataset[0]["audio"]["array"], sampling_rate=sampling_rate, return_tensors="pt") + + >>> speaker_embeddings = torch.zeros((1, 512)) # or load xvectors from a file + + >>> # generate speech + >>> speech = model.generate_speech(inputs["input_values"], speaker_embeddings, vocoder=vocoder) + >>> speech.shape + torch.Size([77312]) + ``` + """ + return_dict = return_dict if return_dict is not None else self.config.use_return_dict + + if labels is not None: + if decoder_input_values is None: + decoder_input_values = shift_spectrograms_right(labels) + + outputs = self.speecht5( + input_values=input_values, + attention_mask=attention_mask, + decoder_input_values=decoder_input_values, + decoder_attention_mask=decoder_attention_mask, + head_mask=head_mask, + decoder_head_mask=decoder_head_mask, + cross_attn_head_mask=cross_attn_head_mask, + encoder_outputs=encoder_outputs, + past_key_values=past_key_values, + use_cache=use_cache, + speaker_embeddings=speaker_embeddings, + output_attentions=output_attentions, + output_hidden_states=output_hidden_states, + return_dict=True, + ) + + _, spectrogram, logits = self.speech_decoder_postnet(outputs[0]) + + loss = None + + if not return_dict: + output = (spectrogram,) + outputs[1:] + return ((loss,) + output) if loss is not None else output + + return Seq2SeqSpectrogramOutput( + loss=loss, + spectrogram=spectrogram, + past_key_values=outputs.past_key_values, + decoder_hidden_states=outputs.decoder_hidden_states, + decoder_attentions=outputs.decoder_attentions, + cross_attentions=outputs.cross_attentions, + encoder_last_hidden_state=outputs.encoder_last_hidden_state, + encoder_hidden_states=outputs.encoder_hidden_states, + encoder_attentions=outputs.encoder_attentions, + ) + + @torch.no_grad() + def generate_speech( + self, + input_values: torch.FloatTensor, + speaker_embeddings: Optional[torch.FloatTensor] = None, + threshold: float = 0.5, + minlenratio: float = 0.0, + maxlenratio: float = 20.0, + vocoder: Optional[nn.Module] = None, + ) -> torch.FloatTensor: + r""" + Converts a raw speech waveform into a sequence of mel spectrograms, which are subsequently turned back into a + speech waveform using a vocoder. + + Args: + input_values (`torch.FloatTensor` of shape `(batch_size, sequence_length)`): + Float values of input raw speech waveform. The `batch_size` should be 1 currently. + + Values can be obtained by loading a *.flac* or *.wav* audio file into an array of type `List[float]` or + a `numpy.ndarray`, *e.g.* via the soundfile library (*pip install soundfile*). To prepare the array + into `input_values`, the [`SpeechT5Processor`] should be used for padding and conversion into a tensor + of type `torch.FloatTensor`. See [`SpeechT5Processor.__call__`] for details. + speaker_embeddings (`torch.FloatTensor` of shape `(batch_size, config.speaker_embedding_dim)`, *optional*): + Tensor containing the speaker embeddings. + threshold (`float`, *optional*, defaults to 0.5): + The generated sequence ends when the predicted stop token probability exceeds this value. + minlenratio (`float`, *optional*, defaults to 0.0): + Used to calculate the minimum required length for the output sequence. + maxlenratio (`float`, *optional*, defaults to 20.0): + Used to calculate the maximum allowed length for the output sequence. + vocoder (`nn.Module`, *optional*, defaults to `None`): + The vocoder that converts the mel spectrogram into a speech waveform. If `None`, the output is the mel + spectrogram. + + Returns: + `torch.FloatTensor`: Tensor of shape `(output_sequence_length, config.num_mel_bins)` containing the + predicted mel spectrogram, or a tensor with shape `(num_frames,)` containing the speech waveform. + """ + if speaker_embeddings is None: + speaker_embeddings = torch.zeros((1, 512)) + + return _generate_speech( + self, + input_values, + speaker_embeddings, + threshold, + minlenratio, + maxlenratio, + vocoder, + ) + + +HIFIGAN_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 ([`SpeechT5HifiGanConfig`]): + 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. +""" + + +class HifiGanResidualBlock(nn.Module): + def __init__(self, channels, kernel_size=3, dilation=(1, 3, 5), leaky_relu_slope=0.1): + super().__init__() + self.leaky_relu_slope = leaky_relu_slope + + self.convs1 = nn.ModuleList( + [ + nn.Conv1d( + channels, + channels, + kernel_size, + stride=1, + dilation=dilation[i], + padding=self.get_padding(kernel_size, dilation[i]), + ) + for i in range(len(dilation)) + ] + ) + self.convs2 = nn.ModuleList( + [ + nn.Conv1d( + channels, + channels, + kernel_size, + stride=1, + dilation=1, + padding=self.get_padding(kernel_size, 1), + ) + for _ in range(len(dilation)) + ] + ) + + def get_padding(self, kernel_size, dilation=1): + return (kernel_size * dilation - dilation) // 2 + + def apply_weight_norm(self): + for layer in self.convs1: + nn.utils.weight_norm(layer) + for layer in self.convs2: + nn.utils.weight_norm(layer) + + def remove_weight_norm(self): + for layer in self.convs1: + nn.utils.remove_weight_norm(layer) + for layer in self.convs2: + nn.utils.remove_weight_norm(layer) + + def forward(self, hidden_states): + for conv1, conv2 in zip(self.convs1, self.convs2): + residual = hidden_states + hidden_states = nn.functional.leaky_relu(hidden_states, self.leaky_relu_slope) + hidden_states = conv1(hidden_states) + hidden_states = nn.functional.leaky_relu(hidden_states, self.leaky_relu_slope) + hidden_states = conv2(hidden_states) + hidden_states = hidden_states + residual + return hidden_states + + +@add_start_docstrings( + """HiFi-GAN vocoder.""", + HIFIGAN_START_DOCSTRING, +) +class SpeechT5HifiGan(PreTrainedModel): + config_class = SpeechT5HifiGanConfig + main_input_name = "spectrogram" + + def __init__(self, config: SpeechT5HifiGanConfig): + super().__init__(config) + self.num_kernels = len(config.resblock_kernel_sizes) + self.num_upsamples = len(config.upsample_rates) + self.conv_pre = nn.Conv1d( + config.model_in_dim, + config.upsample_initial_channel, + kernel_size=7, + stride=1, + padding=3, + ) + + self.upsampler = nn.ModuleList() + for i, (upsample_rate, kernel_size) in enumerate(zip(config.upsample_rates, config.upsample_kernel_sizes)): + self.upsampler.append( + nn.ConvTranspose1d( + config.upsample_initial_channel // (2**i), + config.upsample_initial_channel // (2 ** (i + 1)), + kernel_size=kernel_size, + stride=upsample_rate, + padding=(kernel_size - upsample_rate) // 2, + ) + ) + + self.resblocks = nn.ModuleList() + for i in range(len(self.upsampler)): + channels = config.upsample_initial_channel // (2 ** (i + 1)) + for kernel_size, dilation in zip(config.resblock_kernel_sizes, config.resblock_dilation_sizes): + self.resblocks.append(HifiGanResidualBlock(channels, kernel_size, dilation, config.leaky_relu_slope)) + + self.conv_post = nn.Conv1d(channels, 1, kernel_size=7, stride=1, padding=3) + + self.register_buffer("mean", torch.zeros(config.model_in_dim)) + self.register_buffer("scale", torch.ones(config.model_in_dim)) + + # Initialize weights and apply final processing + self.post_init() + + def _init_weights(self, module): + """Initialize the weights.""" + if isinstance(module, (nn.Linear, nn.Conv1d)): + module.weight.data.normal_(mean=0.0, std=self.config.initializer_range) + if module.bias is not None: + module.bias.data.zero_() + + def apply_weight_norm(self): + nn.utils.weight_norm(self.conv_pre) + for layer in self.upsampler: + nn.utils.weight_norm(layer) + for layer in self.resblocks: + layer.apply_weight_norm() + nn.utils.weight_norm(self.conv_post) + + def remove_weight_norm(self): + nn.utils.remove_weight_norm(self.conv_pre) + for layer in self.upsampler: + nn.utils.remove_weight_norm(layer) + for layer in self.resblocks: + layer.remove_weight_norm() + nn.utils.remove_weight_norm(self.conv_post) + + def forward(self, spectrogram): + r""" + Converts a single log-mel spectogram into a speech waveform. + + Args: + spectrogram (`torch.FloatTensor` of shape `(sequence_length, config.model_in_dim)`): + Tensor containing the log-mel spectrogram. + + Returns: + `torch.FloatTensor`: Tensor of shape `(num_frames,)` containing the speech waveform. + """ + if self.config.normalize_before: + spectrogram = (spectrogram - self.mean) / self.scale + + hidden_states = spectrogram.transpose(1, 0).unsqueeze(0) + + hidden_states = self.conv_pre(hidden_states) + for i in range(self.num_upsamples): + hidden_states = nn.functional.leaky_relu(hidden_states, self.config.leaky_relu_slope) + hidden_states = self.upsampler[i](hidden_states) + + res_state = self.resblocks[i * self.num_kernels](hidden_states) + for j in range(1, self.num_kernels): + res_state += self.resblocks[i * self.num_kernels + j](hidden_states) + hidden_states = res_state / self.num_kernels + + hidden_states = nn.functional.leaky_relu(hidden_states) + hidden_states = self.conv_post(hidden_states) + hidden_states = torch.tanh(hidden_states) + + waveform = hidden_states.squeeze(0).transpose(1, 0).view(-1) + return waveform diff --git a/src/transformers/models/speecht5/processing_speecht5.py b/src/transformers/models/speecht5/processing_speecht5.py new file mode 100644 index 00000000000..3cb66498337 --- /dev/null +++ b/src/transformers/models/speecht5/processing_speecht5.py @@ -0,0 +1,156 @@ +# 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. +"""Speech processor class for SpeechT5.""" + +from ...processing_utils import ProcessorMixin + + +class SpeechT5Processor(ProcessorMixin): + r""" + Constructs a SpeechT5 processor which wraps a feature extractor and a tokenizer into a single processor. + + [`SpeechT5Processor`] offers all the functionalities of [`SpeechT5FeatureExtractor`] and [`SpeechT5Tokenizer`]. See + the docstring of [`~SpeechT5Processor.__call__`] and [`~SpeechT5Processor.decode`] for more information. + + Args: + feature_extractor (`SpeechT5FeatureExtractor`): + An instance of [`SpeechT5FeatureExtractor`]. The feature extractor is a required input. + tokenizer (`SpeechT5Tokenizer`): + An instance of [`SpeechT5Tokenizer`]. The tokenizer is a required input. + """ + feature_extractor_class = "SpeechT5FeatureExtractor" + tokenizer_class = "SpeechT5Tokenizer" + + def __init__(self, feature_extractor, tokenizer): + super().__init__(feature_extractor, tokenizer) + + def __call__(self, *args, **kwargs): + """ + Processes audio and text input, as well as audio and text targets. + + You can process audio by using the argument `audio`, or process audio targets by using the argument + `audio_target`. This forwards the arguments to SpeechT5FeatureExtractor's + [`~SpeechT5FeatureExtractor.__call__`]. + + You can process text by using the argument `text`, or process text labels by using the argument `text_target`. + This forwards the arguments to SpeechT5Tokenizer's [`~SpeechT5Tokenizer.__call__`]. + + Valid input combinations are: + + - `text` only + - `audio` only + - `text_target` only + - `audio_target` only + - `text` and `audio_target` + - `audio` and `audio_target` + - `text` and `text_target` + - `audio` and `text_target` + + Please refer to the docstring of the above two methods for more information. + """ + audio = kwargs.pop("audio", None) + text = kwargs.pop("text", None) + text_target = kwargs.pop("text_target", None) + audio_target = kwargs.pop("audio_target", None) + sampling_rate = kwargs.pop("sampling_rate", None) + + if audio is not None and text is not None: + raise ValueError( + "Cannot process both `audio` and `text` inputs. Did you mean `audio_target` or `text_target`?" + ) + if audio_target is not None and text_target is not None: + raise ValueError( + "Cannot process both `audio_target` and `text_target` inputs. Did you mean `audio` or `text`?" + ) + if audio is None and audio_target is None and text is None and text_target is None: + raise ValueError( + "You need to specify either an `audio`, `audio_target`, `text`, or `text_target` input to process." + ) + + if audio is not None: + inputs = self.feature_extractor(audio, *args, sampling_rate=sampling_rate, **kwargs) + elif text is not None: + inputs = self.tokenizer(text, **kwargs) + else: + inputs = None + + if audio_target is not None: + audio_target_features = self.feature_extractor( + audio_target=audio_target, *args, sampling_rate=sampling_rate, **kwargs + ) + if inputs is None: + return audio_target_features + else: + inputs["labels"] = audio_target_features["input_values"] + inputs["stop_labels"] = audio_target_features["stop_labels"] + decoder_attention_mask = audio_target_features.get("attention_mask") + if decoder_attention_mask is not None: + inputs["decoder_attention_mask"] = decoder_attention_mask + + if text_target is not None: + encodings_target = self.tokenizer(text_target, **kwargs) + if inputs is None: + return encodings_target + else: + inputs["labels"] = encodings_target["input_ids"] + decoder_attention_mask = encodings_target.get("attention_mask") + if decoder_attention_mask is not None: + inputs["decoder_attention_mask"] = decoder_attention_mask + + return inputs + + def pad(self, *args, **kwargs): + """ + This method forwards all its arguments to SpeechT5FeatureExtractor's [`~SpeechT5FeatureExtractor.pad`] and + returns its output. + + You can process your labels by using the argument `text` (either in the same call as your audio inputs, or in a + separate call). This forwards its arguments to SpeechT5Tokenizer's [`~SpeechT5Tokenizer.pad`]. + + Please refer to the docstring of the above two methods for more information. + """ + input_features = kwargs.pop("input_features", None) + labels = kwargs.pop("labels", None) + + if len(args) > 0: + input_features = args[0] + args = args[1:] + + if input_features is not None: + input_features = self.feature_extractor.pad(input_features, *args, **kwargs) + if labels is not None: + labels = self.tokenizer.pad(labels, **kwargs) + + if labels is None: + return input_features + elif input_features is None: + return labels + else: + input_features["labels"] = labels["input_ids"] + return input_features + + def batch_decode(self, *args, **kwargs): + """ + This method forwards all its arguments to SpeechT5Tokenizer's [`~SpeechT5Tokenizer.batch_decode`]. Please refer + to the docstring of this method for more information. + """ + return self.tokenizer.batch_decode(*args, **kwargs) + + def decode(self, *args, **kwargs): + """ + This method forwards all its arguments to SpeechT5Tokenizer's [`~SpeechT5Tokenizer.decode`]. Please refer to + the docstring of this method for more information. + """ + return self.tokenizer.decode(*args, **kwargs) diff --git a/src/transformers/models/speecht5/tokenization_speecht5.py b/src/transformers/models/speecht5/tokenization_speecht5.py new file mode 100644 index 00000000000..4b641558b7c --- /dev/null +++ b/src/transformers/models/speecht5/tokenization_speecht5.py @@ -0,0 +1,185 @@ +# coding=utf-8 +# Copyright 2023 The Facebook Inc. 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. +"""Tokenization class for SpeechT5.""" + + +import os +from shutil import copyfile +from typing import Any, Dict, List, Optional, Tuple + +import sentencepiece as spm + +from ...tokenization_utils import PreTrainedTokenizer +from ...utils import logging + + +logger = logging.get_logger(__name__) + +VOCAB_FILES_NAMES = {"vocab_file": "spm_char.model"} + +PRETRAINED_VOCAB_FILES_MAP = { + "vocab_file": { + "microsoft/speecht5_asr": "https://huggingface.co/microsoft/speecht5_asr/resolve/main/spm_char.model", + "microsoft/speecht5_tts": "https://huggingface.co/microsoft/speecht5_tts/resolve/main/spm_char.model", + "microsoft/speecht5_vc": "https://huggingface.co/microsoft/speecht5_vc/resolve/main/spm_char.model", + } +} + +PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = { + "microsoft/speecht5_asr": 1024, + "microsoft/speecht5_tts": 1024, + "microsoft/speecht5_vc": 1024, +} + + +class SpeechT5Tokenizer(PreTrainedTokenizer): + """ + Construct a SpeechT5 tokenizer. Based on [SentencePiece](https://github.com/google/sentencepiece). + + This tokenizer inherits from [`PreTrainedTokenizer`] which contains most of the main methods. Users should refer to + this superclass for more information regarding those methods. + + Args: + vocab_file (`str`): + [SentencePiece](https://github.com/google/sentencepiece) file (generally has a *.spm* extension) that + contains the vocabulary necessary to instantiate a tokenizer. + eos_token (`str`, *optional*, defaults to `""`): + The end of sequence token. + bos_token (`str`, *optional*, defaults to `""`): + The begin of sequence token. + unk_token (`str`, *optional*, defaults to `""`): + The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this + token instead. + pad_token (`str`, *optional*, defaults to `""`): + The token used for padding, for example when batching sequences of different lengths. + sp_model_kwargs (`dict`, *optional*): + Will be passed to the `SentencePieceProcessor.__init__()` method. The [Python wrapper for + SentencePiece](https://github.com/google/sentencepiece/tree/master/python) can be used, among other things, + to set: + + - `enable_sampling`: Enable subword regularization. + - `nbest_size`: Sampling parameters for unigram. Invalid for BPE-Dropout. + + - `nbest_size = {0,1}`: No sampling is performed. + - `nbest_size > 1`: samples from the nbest_size results. + - `nbest_size < 0`: assuming that nbest_size is infinite and samples from the all hypothesis (lattice) + using forward-filtering-and-backward-sampling algorithm. + + - `alpha`: Smoothing parameter for unigram sampling, and dropout probability of merge operations for + BPE-dropout. + + Attributes: + sp_model (`SentencePieceProcessor`): + The *SentencePiece* processor that is used for every conversion (string, tokens and IDs). + """ + + vocab_files_names = VOCAB_FILES_NAMES + pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP + max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES + model_input_names = ["input_ids", "attention_mask"] + + def __init__( + self, + vocab_file, + bos_token="", + eos_token="", + unk_token="", + pad_token="", + sp_model_kwargs: Optional[Dict[str, Any]] = None, + **kwargs + ) -> None: + self.sp_model_kwargs = {} if sp_model_kwargs is None else sp_model_kwargs + + super().__init__( + bos_token=bos_token, + eos_token=eos_token, + unk_token=unk_token, + pad_token=pad_token, + sp_model_kwargs=self.sp_model_kwargs, + **kwargs, + ) + + self.vocab_file = vocab_file + + self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs) + self.sp_model.Load(vocab_file) + + @property + def vocab_size(self): + return self.sp_model.get_piece_size() + + def get_vocab(self): + vocab = {self.convert_ids_to_tokens(i): i for i in range(self.vocab_size)} + vocab.update(self.added_tokens_encoder) + return vocab + + def __getstate__(self): + state = self.__dict__.copy() + state["sp_model"] = None + return state + + def __setstate__(self, d): + self.__dict__ = d + + # for backward compatibility + if not hasattr(self, "sp_model_kwargs"): + self.sp_model_kwargs = {} + + self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs) + self.sp_model.Load(self.vocab_file) + + def _tokenize(self, text: str) -> List[str]: + """Take as input a string and return a list of strings (tokens) for words/sub-words""" + return self.sp_model.encode(text, out_type=str) + + def _convert_token_to_id(self, token): + """Converts a token (str) in an id using the vocab.""" + return self.sp_model.piece_to_id(token) + + def _convert_id_to_token(self, index): + """Converts an index (integer) in a token (str) using the vocab.""" + token = self.sp_model.IdToPiece(index) + return token + + def convert_tokens_to_string(self, tokens): + """Converts a sequence of tokens (string) in a single string.""" + current_sub_tokens = [] + out_string = "" + for token in tokens: + # make sure that special tokens are not decoded using sentencepiece model + if token in self.all_special_tokens: + out_string += self.sp_model.decode(current_sub_tokens) + token + current_sub_tokens = [] + else: + current_sub_tokens.append(token) + out_string += self.sp_model.decode(current_sub_tokens) + return out_string.strip() + + def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]: + if not os.path.isdir(save_directory): + logger.error(f"Vocabulary path ({save_directory}) should be a directory") + return + out_vocab_file = os.path.join( + save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"] + ) + + if os.path.abspath(self.vocab_file) != os.path.abspath(out_vocab_file) and os.path.isfile(self.vocab_file): + copyfile(self.vocab_file, out_vocab_file) + elif not os.path.isfile(self.vocab_file): + with open(out_vocab_file, "wb") as fi: + content_spiece_model = self.sp_model.serialized_model_proto() + fi.write(content_spiece_model) + + return (out_vocab_file,) diff --git a/src/transformers/utils/dummy_pt_objects.py b/src/transformers/utils/dummy_pt_objects.py index e1e52b7dbad..da52ac42cc0 100644 --- a/src/transformers/utils/dummy_pt_objects.py +++ b/src/transformers/utils/dummy_pt_objects.py @@ -5505,6 +5505,51 @@ class Speech2Text2PreTrainedModel(metaclass=DummyObject): requires_backends(self, ["torch"]) +SPEECHT5_PRETRAINED_MODEL_ARCHIVE_LIST = None + + +class SpeechT5ForSpeechToSpeech(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + +class SpeechT5ForSpeechToText(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + +class SpeechT5ForTextToSpeech(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + +class SpeechT5HifiGan(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + +class SpeechT5Model(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + +class SpeechT5PreTrainedModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + SPLINTER_PRETRAINED_MODEL_ARCHIVE_LIST = None diff --git a/src/transformers/utils/dummy_sentencepiece_and_speech_objects.py b/src/transformers/utils/dummy_sentencepiece_and_speech_objects.py index b9b971f1f15..11074dd4635 100644 --- a/src/transformers/utils/dummy_sentencepiece_and_speech_objects.py +++ b/src/transformers/utils/dummy_sentencepiece_and_speech_objects.py @@ -8,3 +8,10 @@ class Speech2TextProcessor(metaclass=DummyObject): def __init__(self, *args, **kwargs): requires_backends(self, ["sentencepiece", "speech"]) + + +class SpeechT5Processor(metaclass=DummyObject): + _backends = ["sentencepiece", "speech"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["sentencepiece", "speech"]) diff --git a/src/transformers/utils/dummy_sentencepiece_objects.py b/src/transformers/utils/dummy_sentencepiece_objects.py index 9d9c7613cd0..330f4032f90 100644 --- a/src/transformers/utils/dummy_sentencepiece_objects.py +++ b/src/transformers/utils/dummy_sentencepiece_objects.py @@ -164,6 +164,13 @@ class Speech2TextTokenizer(metaclass=DummyObject): requires_backends(self, ["sentencepiece"]) +class SpeechT5Tokenizer(metaclass=DummyObject): + _backends = ["sentencepiece"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["sentencepiece"]) + + class T5Tokenizer(metaclass=DummyObject): _backends = ["sentencepiece"] diff --git a/src/transformers/utils/dummy_speech_objects.py b/src/transformers/utils/dummy_speech_objects.py index d1929dd2853..7aab43f17be 100644 --- a/src/transformers/utils/dummy_speech_objects.py +++ b/src/transformers/utils/dummy_speech_objects.py @@ -22,3 +22,10 @@ class Speech2TextFeatureExtractor(metaclass=DummyObject): def __init__(self, *args, **kwargs): requires_backends(self, ["speech"]) + + +class SpeechT5FeatureExtractor(metaclass=DummyObject): + _backends = ["speech"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["speech"]) diff --git a/tests/fixtures/test_sentencepiece_bpe_char.model b/tests/fixtures/test_sentencepiece_bpe_char.model new file mode 100644 index 00000000000..82ee359fb5e Binary files /dev/null and b/tests/fixtures/test_sentencepiece_bpe_char.model differ diff --git a/tests/models/speecht5/__init__.py b/tests/models/speecht5/__init__.py new file mode 100644 index 00000000000..e69de29bb2d diff --git a/tests/models/speecht5/test_feature_extraction_speecht5.py b/tests/models/speecht5/test_feature_extraction_speecht5.py new file mode 100644 index 00000000000..2720f531af9 --- /dev/null +++ b/tests/models/speecht5/test_feature_extraction_speecht5.py @@ -0,0 +1,421 @@ +# 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 SpeechT5 feature extractors.""" + +import itertools +import random +import unittest + +import numpy as np + +from transformers import BatchFeature, is_speech_available +from transformers.testing_utils import require_torch, require_torchaudio +from transformers.utils.import_utils import is_torch_available + +from ...test_sequence_feature_extraction_common import SequenceFeatureExtractionTestMixin + + +if is_speech_available(): + from transformers import SpeechT5FeatureExtractor + +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 SpeechT5FeatureExtractionTester(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=16000, + do_normalize=True, + num_mel_bins=80, + hop_length=16, + win_length=64, + win_function="hann_window", + frame_signal_scale=1.0, + fmin=80, + fmax=7600, + mel_floor=1e-10, + reduction_factor=2, + 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.do_normalize = do_normalize + self.num_mel_bins = num_mel_bins + self.hop_length = hop_length + self.win_length = win_length + self.win_function = win_function + self.frame_signal_scale = frame_signal_scale + self.fmin = fmin + self.fmax = fmax + self.mel_floor = mel_floor + self.reduction_factor = reduction_factor + 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, + "do_normalize": self.do_normalize, + "num_mel_bins": self.num_mel_bins, + "hop_length": self.hop_length, + "win_length": self.win_length, + "win_function": self.win_function, + "frame_signal_scale": self.frame_signal_scale, + "fmin": self.fmin, + "fmax": self.fmax, + "mel_floor": self.mel_floor, + "reduction_factor": self.reduction_factor, + "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: + speech_inputs = floats_list((self.batch_size, self.max_seq_length)) + else: + # make sure that inputs increase in size + speech_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: + speech_inputs = [np.asarray(x) for x in speech_inputs] + + return speech_inputs + + def prepare_inputs_for_target(self, equal_length=False, numpify=False): + if equal_length: + speech_inputs = [floats_list((self.max_seq_length, self.num_mel_bins)) for _ in range(self.batch_size)] + else: + # make sure that inputs increase in size + speech_inputs = [ + floats_list((x, self.num_mel_bins)) + for x in range(self.min_seq_length, self.max_seq_length, self.seq_length_diff) + ] + + if numpify: + speech_inputs = [np.asarray(x) for x in speech_inputs] + + return speech_inputs + + +@require_torch +@require_torchaudio +class SpeechT5FeatureExtractionTest(SequenceFeatureExtractionTestMixin, unittest.TestCase): + + feature_extraction_class = SpeechT5FeatureExtractor if is_speech_available() else None + + def setUp(self): + self.feat_extract_tester = SpeechT5FeatureExtractionTester(self) + + def _check_zero_mean_unit_variance(self, input_vector): + self.assertTrue(np.all(np.mean(input_vector, axis=0) < 1e-3)) + self.assertTrue(np.all(np.abs(np.var(input_vector, axis=0) - 1) < 1e-3)) + + 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 + speech_inputs = [floats_list((1, x))[0] for x in range(800, 1400, 200)] + np_speech_inputs = [np.asarray(speech_input) for speech_input in speech_inputs] + + # Test not batched input + encoded_sequences_1 = feat_extract(speech_inputs[0], return_tensors="np").input_values + encoded_sequences_2 = feat_extract(np_speech_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(speech_inputs, return_tensors="np").input_values + encoded_sequences_2 = feat_extract(np_speech_inputs, 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_zero_mean_unit_variance_normalization_np(self): + feat_extract = self.feature_extraction_class(**self.feat_extract_tester.prepare_feat_extract_dict()) + speech_inputs = [floats_list((1, x))[0] for x in range(800, 1400, 200)] + + paddings = ["longest", "max_length", "do_not_pad"] + max_lengths = [None, 1600, None] + for max_length, padding in zip(max_lengths, paddings): + processed = feat_extract(speech_inputs, padding=padding, max_length=max_length, return_tensors="np") + input_values = processed.input_values + + self._check_zero_mean_unit_variance(input_values[0][:800]) + self.assertTrue(input_values[0][800:].sum() < 1e-6) + self._check_zero_mean_unit_variance(input_values[1][:1000]) + self.assertTrue(input_values[0][1000:].sum() < 1e-6) + self._check_zero_mean_unit_variance(input_values[2][:1200]) + + def test_zero_mean_unit_variance_normalization(self): + feat_extract = self.feature_extraction_class(**self.feat_extract_tester.prepare_feat_extract_dict()) + lengths = range(800, 1400, 200) + speech_inputs = [floats_list((1, x))[0] for x in lengths] + + paddings = ["longest", "max_length", "do_not_pad"] + max_lengths = [None, 1600, None] + + for max_length, padding in zip(max_lengths, paddings): + processed = feat_extract(speech_inputs, max_length=max_length, padding=padding) + input_values = processed.input_values + + self._check_zero_mean_unit_variance(input_values[0][:800]) + self._check_zero_mean_unit_variance(input_values[1][:1000]) + self._check_zero_mean_unit_variance(input_values[2][:1200]) + + def test_zero_mean_unit_variance_normalization_trunc_np_max_length(self): + feat_extract = self.feature_extraction_class(**self.feat_extract_tester.prepare_feat_extract_dict()) + speech_inputs = [floats_list((1, x))[0] for x in range(800, 1400, 200)] + processed = feat_extract( + speech_inputs, truncation=True, max_length=1000, padding="max_length", return_tensors="np" + ) + input_values = processed.input_values + + self._check_zero_mean_unit_variance(input_values[0, :800]) + self._check_zero_mean_unit_variance(input_values[1]) + self._check_zero_mean_unit_variance(input_values[2]) + + def test_zero_mean_unit_variance_normalization_trunc_np_longest(self): + feat_extract = self.feature_extraction_class(**self.feat_extract_tester.prepare_feat_extract_dict()) + speech_inputs = [floats_list((1, x))[0] for x in range(800, 1400, 200)] + processed = feat_extract( + speech_inputs, truncation=True, max_length=1000, padding="longest", return_tensors="np" + ) + input_values = processed.input_values + + self._check_zero_mean_unit_variance(input_values[0, :800]) + self._check_zero_mean_unit_variance(input_values[1, :1000]) + self._check_zero_mean_unit_variance(input_values[2]) + + # make sure that if max_length < longest -> then pad to max_length + self.assertTrue(input_values.shape == (3, 1000)) + + speech_inputs = [floats_list((1, x))[0] for x in range(800, 1400, 200)] + processed = feat_extract( + speech_inputs, truncation=True, max_length=2000, padding="longest", return_tensors="np" + ) + input_values = processed.input_values + + self._check_zero_mean_unit_variance(input_values[0, :800]) + self._check_zero_mean_unit_variance(input_values[1, :1000]) + self._check_zero_mean_unit_variance(input_values[2]) + + # make sure that if max_length > longest -> then pad to longest + self.assertTrue(input_values.shape == (3, 1200)) + + def test_double_precision_pad(self): + feature_extractor = self.feature_extraction_class(**self.feat_extract_tester.prepare_feat_extract_dict()) + np_speech_inputs = np.random.rand(100).astype(np.float64) + py_speech_inputs = np_speech_inputs.tolist() + + for inputs in [py_speech_inputs, np_speech_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 test_call_target(self): + # Tests that all call wrap to encode_plus and batch_encode_plus + feature_extractor = self.feature_extraction_class(**self.feat_extract_tester.prepare_feat_extract_dict()) + # create three inputs of length 8000, 14000, and 2000 + speech_inputs = [floats_list((1, x))[0] for x in range(8000, 14000, 2000)] + np_speech_inputs = [np.asarray(speech_input) for speech_input in speech_inputs] + + # Test feature size + input_values = feature_extractor(audio_target=np_speech_inputs, padding=True, return_tensors="np").input_values + self.assertTrue(input_values.ndim == 3) + self.assertTrue(input_values.shape[-1] == feature_extractor.num_mel_bins) + + # Test not batched input + encoded_sequences_1 = feature_extractor(speech_inputs[0], return_tensors="np").input_values + encoded_sequences_2 = feature_extractor(np_speech_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 = feature_extractor(speech_inputs, return_tensors="np").input_values + encoded_sequences_2 = feature_extractor(np_speech_inputs, 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_batch_feature_target(self): + speech_inputs = self.feat_extract_tester.prepare_inputs_for_target() + feat_extract = self.feature_extraction_class(**self.feat_extract_dict) + input_name = feat_extract.model_input_names[0] + + processed_features = BatchFeature({input_name: speech_inputs}) + + self.assertTrue(all(len(x) == len(y) for x, y in zip(speech_inputs, processed_features[input_name]))) + + speech_inputs = self.feat_extract_tester.prepare_inputs_for_target(equal_length=True) + processed_features = BatchFeature({input_name: speech_inputs}, tensor_type="np") + + batch_features_input = processed_features[input_name] + + if len(batch_features_input.shape) < 3: + batch_features_input = batch_features_input[:, :, None] + + self.assertTrue( + batch_features_input.shape + == (self.feat_extract_tester.batch_size, len(speech_inputs[0]), self.feat_extract_tester.num_mel_bins) + ) + + @require_torch + def test_batch_feature_target_pt(self): + speech_inputs = self.feat_extract_tester.prepare_inputs_for_target(equal_length=True) + feat_extract = self.feature_extraction_class(**self.feat_extract_dict) + input_name = feat_extract.model_input_names[0] + + processed_features = BatchFeature({input_name: speech_inputs}, tensor_type="pt") + + batch_features_input = processed_features[input_name] + + if len(batch_features_input.shape) < 3: + batch_features_input = batch_features_input[:, :, None] + + self.assertTrue( + batch_features_input.shape + == (self.feat_extract_tester.batch_size, len(speech_inputs[0]), self.feat_extract_tester.num_mel_bins) + ) + + @require_torch + def test_padding_accepts_tensors_target_pt(self): + feat_extract = self.feature_extraction_class(**self.feat_extract_dict) + speech_inputs = self.feat_extract_tester.prepare_inputs_for_target() + input_name = feat_extract.model_input_names[0] + + processed_features = BatchFeature({input_name: speech_inputs}) + + feat_extract.feature_size = feat_extract.num_mel_bins # hack! + + input_np = feat_extract.pad(processed_features, padding="longest", return_tensors="np")[input_name] + input_pt = feat_extract.pad(processed_features, padding="longest", return_tensors="pt")[input_name] + + self.assertTrue(abs(input_np.astype(np.float32).sum() - input_pt.numpy().astype(np.float32).sum()) < 1e-2) + + def test_attention_mask_target(self): + feat_dict = self.feat_extract_dict + feat_dict["return_attention_mask"] = True + feat_extract = self.feature_extraction_class(**feat_dict) + speech_inputs = self.feat_extract_tester.prepare_inputs_for_target() + input_lenghts = [len(x) for x in speech_inputs] + input_name = feat_extract.model_input_names[0] + + processed = BatchFeature({input_name: speech_inputs}) + + feat_extract.feature_size = feat_extract.num_mel_bins # hack! + + processed = feat_extract.pad(processed, padding="longest", return_tensors="np") + self.assertIn("attention_mask", processed) + self.assertListEqual(list(processed.attention_mask.shape), list(processed[input_name].shape[:2])) + self.assertListEqual(processed.attention_mask.sum(-1).tolist(), input_lenghts) + + def test_attention_mask_with_truncation_target(self): + feat_dict = self.feat_extract_dict + feat_dict["return_attention_mask"] = True + feat_extract = self.feature_extraction_class(**feat_dict) + speech_inputs = self.feat_extract_tester.prepare_inputs_for_target() + input_lenghts = [len(x) for x in speech_inputs] + input_name = feat_extract.model_input_names[0] + + processed = BatchFeature({input_name: speech_inputs}) + max_length = min(input_lenghts) + + feat_extract.feature_size = feat_extract.num_mel_bins # hack! + + processed_pad = feat_extract.pad( + processed, padding="max_length", max_length=max_length, truncation=True, return_tensors="np" + ) + self.assertIn("attention_mask", processed_pad) + self.assertListEqual( + list(processed_pad.attention_mask.shape), list((processed_pad[input_name].shape[0], max_length)) + ) + self.assertListEqual( + processed_pad.attention_mask[:, :max_length].sum(-1).tolist(), [max_length for x in speech_inputs] + ) + + 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 + speech_samples = ds.sort("id").select(range(num_samples))[:num_samples]["audio"] + + return [x["array"] for x in speech_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_speech = self._load_datasamples(1) + feature_extractor = SpeechT5FeatureExtractor() + input_values = feature_extractor(input_speech, return_tensors="pt").input_values + self.assertTrue(torch.allclose(input_values[0, :30], EXPECTED_INPUT_VALUES, atol=1e-4)) + + def test_integration_target(self): + # fmt: off + EXPECTED_INPUT_VALUES = torch.tensor( + [-2.7713, -2.8896, -3.2619, -3.0843, -2.9919, -3.0084, -3.2796, -3.3169, + -3.2397, -3.2053, -2.9151, -2.7921, -2.9403, -2.7411, -3.0654, -2.8314, + -3.0026, -2.9797, -3.1314, -2.9939, -2.6748, -2.7725, -2.8563, -2.9462, + -3.2623, -3.3044, -3.1318, -3.2672, -3.4030, -3.1988] + ) + # fmt: on + + input_speech = self._load_datasamples(1) + feature_extractor = SpeechT5FeatureExtractor() + input_values = feature_extractor(audio_target=input_speech, return_tensors="pt").input_values + self.assertTrue(torch.allclose(input_values[0, 0, :30], EXPECTED_INPUT_VALUES, atol=1e-4)) diff --git a/tests/models/speecht5/test_modeling_speecht5.py b/tests/models/speecht5/test_modeling_speecht5.py new file mode 100644 index 00000000000..6c37135b417 --- /dev/null +++ b/tests/models/speecht5/test_modeling_speecht5.py @@ -0,0 +1,1547 @@ +# coding=utf-8 +# Copyright 2022 The HuggingFace Inc. team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" Testing suite for the PyTorch SpeechT5 model. """ + +import copy +import inspect +import tempfile +import unittest + +from transformers import SpeechT5Config, SpeechT5HifiGanConfig +from transformers.testing_utils import ( + is_torch_available, + require_sentencepiece, + require_tokenizers, + require_torch, + require_torchaudio, + slow, + torch_device, +) +from transformers.utils import cached_property + +from ...test_configuration_common import ConfigTester +from ...test_modeling_common import ( + ModelTesterMixin, + _config_zero_init, + floats_tensor, + ids_tensor, + random_attention_mask, +) + + +if is_torch_available(): + import torch + + from transformers import ( + SpeechT5ForSpeechToSpeech, + SpeechT5ForSpeechToText, + SpeechT5ForTextToSpeech, + SpeechT5HifiGan, + SpeechT5Model, + SpeechT5Processor, + ) + + +def prepare_inputs_dict( + config, + input_ids=None, + input_values=None, + decoder_input_ids=None, + decoder_input_values=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} + + if decoder_input_ids is not None: + decoder_dict = {"decoder_input_ids": decoder_input_ids} + else: + decoder_dict = {"decoder_input_values": decoder_input_values} + + if head_mask is None: + head_mask = torch.ones(config.encoder_layers, config.encoder_attention_heads, device=torch_device) + if decoder_head_mask is None: + decoder_head_mask = torch.ones(config.decoder_layers, config.decoder_attention_heads, device=torch_device) + if cross_attn_head_mask is None: + cross_attn_head_mask = torch.ones(config.decoder_layers, config.decoder_attention_heads, device=torch_device) + + return { + **encoder_dict, + **decoder_dict, + "attention_mask": attention_mask, + "decoder_attention_mask": decoder_attention_mask, + "head_mask": head_mask, + "decoder_head_mask": decoder_head_mask, + "cross_attn_head_mask": cross_attn_head_mask, + } + + +@require_torch +class SpeechT5ModelTester: + def __init__( + self, + parent, + batch_size=13, + seq_length=7, + is_training=False, + hidden_size=24, + num_hidden_layers=4, + num_attention_heads=2, + intermediate_size=4, + ): + self.parent = parent + self.batch_size = batch_size + self.seq_length = seq_length + self.is_training = is_training + self.hidden_size = hidden_size + self.num_hidden_layers = num_hidden_layers + self.num_attention_heads = num_attention_heads + self.intermediate_size = intermediate_size + + def prepare_config_and_inputs(self): + input_values = floats_tensor([self.batch_size, self.seq_length, self.hidden_size], scale=1.0) + attention_mask = random_attention_mask([self.batch_size, self.seq_length]) + + decoder_input_values = floats_tensor([self.batch_size, self.seq_length, self.hidden_size], scale=1.0) + decoder_attention_mask = random_attention_mask([self.batch_size, self.seq_length]) + + config = self.get_config() + inputs_dict = prepare_inputs_dict( + config, + input_values=input_values, + decoder_input_values=decoder_input_values, + attention_mask=attention_mask, + decoder_attention_mask=decoder_attention_mask, + ) + 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 SpeechT5Config( + hidden_size=self.hidden_size, + encoder_layers=self.num_hidden_layers, + decoder_layers=self.num_hidden_layers, + encoder_attention_heads=self.num_attention_heads, + decoder_attention_heads=self.num_attention_heads, + encoder_ffn_dim=self.intermediate_size, + decoder_ffn_dim=self.intermediate_size, + ) + + def create_and_check_model_forward(self, config, inputs_dict): + model = SpeechT5Model(config=config).to(torch_device).eval() + + input_values = inputs_dict["input_values"] + attention_mask = inputs_dict["attention_mask"] + decoder_input_values = inputs_dict["decoder_input_values"] + + result = model(input_values, attention_mask=attention_mask, decoder_input_values=decoder_input_values) + self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size)) + + +@require_torch +class SpeechT5ModelTest(ModelTesterMixin, unittest.TestCase): + all_model_classes = (SpeechT5Model,) if is_torch_available() else () + is_encoder_decoder = True + test_pruning = False + test_headmasking = False + test_resize_embeddings = False + + input_name = "input_values" + + def setUp(self): + self.model_tester = SpeechT5ModelTester(self) + self.config_tester = ConfigTester(self, config_class=SpeechT5Config, hidden_size=37) + + 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", + "attention_mask", + "decoder_input_values", + "decoder_attention_mask", + ] + expected_arg_names.extend( + ["head_mask", "decoder_head_mask", "cross_attn_head_mask", "encoder_outputs"] + if "head_mask" and "decoder_head_mask" and "cross_attn_head_mask" in arg_names + else ["encoder_outputs"] + ) + self.assertListEqual(arg_names[: len(expected_arg_names)], expected_arg_names) + + # this model has no inputs_embeds + def test_inputs_embeds(self): + pass + + # this model has no input embeddings + def test_model_common_attributes(self): + pass + + def test_retain_grad_hidden_states_attentions(self): + # decoder cannot keep gradients + pass + + @slow + def test_torchscript_output_attentions(self): + # disabled because this model doesn't have decoder_input_ids + pass + + @slow + def test_torchscript_output_hidden_state(self): + # disabled because this model doesn't have decoder_input_ids + pass + + @slow + def test_torchscript_simple(self): + # disabled because this model doesn't have decoder_input_ids + pass + + +@require_torch +class SpeechT5ForSpeechToTextTester: + def __init__( + self, + parent, + batch_size=13, + encoder_seq_length=1024, # speech is longer + decoder_seq_length=7, + is_training=False, + hidden_size=24, + num_hidden_layers=4, + num_attention_heads=2, + intermediate_size=4, + conv_dim=(32, 32, 32), + conv_stride=(4, 4, 4), + conv_kernel=(8, 8, 8), + conv_bias=False, + num_conv_pos_embeddings=16, + num_conv_pos_embedding_groups=2, + vocab_size=81, + ): + self.parent = parent + self.batch_size = batch_size + self.encoder_seq_length = encoder_seq_length + self.decoder_seq_length = decoder_seq_length + self.is_training = is_training + self.hidden_size = hidden_size + self.num_hidden_layers = num_hidden_layers + self.num_attention_heads = num_attention_heads + self.intermediate_size = intermediate_size + self.conv_dim = conv_dim + self.conv_stride = conv_stride + self.conv_kernel = conv_kernel + self.conv_bias = conv_bias + self.num_conv_pos_embeddings = num_conv_pos_embeddings + self.num_conv_pos_embedding_groups = num_conv_pos_embedding_groups + self.vocab_size = vocab_size + + def prepare_config_and_inputs(self): + input_values = floats_tensor([self.batch_size, self.encoder_seq_length], scale=1.0) + attention_mask = random_attention_mask([self.batch_size, self.encoder_seq_length]) + + decoder_input_ids = ids_tensor([self.batch_size, self.decoder_seq_length], self.vocab_size).clamp(2) + decoder_attention_mask = random_attention_mask([self.batch_size, self.decoder_seq_length]) + + config = self.get_config() + inputs_dict = prepare_inputs_dict( + config, + input_values=input_values, + decoder_input_ids=decoder_input_ids, + attention_mask=attention_mask, + decoder_attention_mask=decoder_attention_mask, + ) + 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 SpeechT5Config( + hidden_size=self.hidden_size, + encoder_layers=self.num_hidden_layers, + decoder_layers=self.num_hidden_layers, + encoder_attention_heads=self.num_attention_heads, + decoder_attention_heads=self.num_attention_heads, + encoder_ffn_dim=self.intermediate_size, + decoder_ffn_dim=self.intermediate_size, + conv_dim=self.conv_dim, + conv_stride=self.conv_stride, + conv_kernel=self.conv_kernel, + conv_bias=self.conv_bias, + num_conv_pos_embeddings=self.num_conv_pos_embeddings, + num_conv_pos_embedding_groups=self.num_conv_pos_embedding_groups, + vocab_size=self.vocab_size, + ) + + def create_and_check_model_forward(self, config, inputs_dict): + model = SpeechT5ForSpeechToText(config=config).to(torch_device).eval() + + input_values = inputs_dict["input_values"] + attention_mask = inputs_dict["attention_mask"] + decoder_input_ids = inputs_dict["decoder_input_ids"] + + result = model(input_values, attention_mask=attention_mask, decoder_input_ids=decoder_input_ids) + self.parent.assertEqual(result.logits.shape, (self.batch_size, self.decoder_seq_length, self.vocab_size)) + + def create_and_check_decoder_model_past_large_inputs(self, config, inputs_dict): + model = SpeechT5ForSpeechToText(config=config).get_decoder().to(torch_device).eval() + input_ids = inputs_dict["decoder_input_ids"] + attention_mask = inputs_dict["decoder_attention_mask"] + + # first forward pass + outputs = model(input_ids, attention_mask=attention_mask, use_cache=True) + + output, past_key_values = outputs.to_tuple() + + # create hypothetical multiple next token and extent to next_input_ids + next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size).clamp(2) + next_attn_mask = ids_tensor((self.batch_size, 3), 2) + + # append to next input_ids and + next_input_ids = torch.cat([input_ids, next_tokens], dim=-1) + next_attention_mask = torch.cat([attention_mask, next_attn_mask], dim=-1) + + output_from_no_past = model(next_input_ids, attention_mask=next_attention_mask)["last_hidden_state"] + output_from_past = model(next_tokens, attention_mask=next_attention_mask, past_key_values=past_key_values)[ + "last_hidden_state" + ] + + # select random slice + random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item() + output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach() + output_from_past_slice = output_from_past[:, :, random_slice_idx].detach() + + self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1]) + + # test that outputs are equal for slice + self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-2)) + + +@require_torch +class SpeechT5ForSpeechToTextTest(ModelTesterMixin, unittest.TestCase): + all_model_classes = (SpeechT5ForSpeechToText,) if is_torch_available() else () + all_generative_model_classes = (SpeechT5ForSpeechToText,) if is_torch_available() else () + is_encoder_decoder = True + test_pruning = False + test_headmasking = False + + input_name = "input_values" + + def setUp(self): + self.model_tester = SpeechT5ForSpeechToTextTester(self) + self.config_tester = ConfigTester(self, config_class=SpeechT5Config, hidden_size=37) + + def test_config(self): + self.config_tester.run_common_tests() + + def test_save_load_strict(self): + config, inputs_dict = self.model_tester.prepare_config_and_inputs() + for model_class in self.all_model_classes: + model = model_class(config) + + with tempfile.TemporaryDirectory() as tmpdirname: + model.save_pretrained(tmpdirname) + model2, info = model_class.from_pretrained(tmpdirname, output_loading_info=True) + self.assertEqual(info["missing_keys"], []) + + 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_decoder_model_past_with_large_inputs(self): + config_and_inputs = self.model_tester.prepare_config_and_inputs() + self.model_tester.create_and_check_decoder_model_past_large_inputs(*config_and_inputs) + + def test_attention_outputs(self): + config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common() + config.return_dict = True + + seq_len = getattr(self.model_tester, "seq_length", None) + decoder_seq_length = getattr(self.model_tester, "decoder_seq_length", seq_len) + encoder_seq_length = getattr(self.model_tester, "encoder_seq_length", seq_len) + decoder_key_length = getattr(self.model_tester, "decoder_key_length", decoder_seq_length) + encoder_key_length = getattr(self.model_tester, "key_length", encoder_seq_length) + + for model_class in self.all_model_classes: + inputs_dict["output_attentions"] = True + inputs_dict["output_hidden_states"] = False + config.return_dict = True + model = model_class(config) + model.to(torch_device) + model.eval() + + subsampled_encoder_seq_length = model.speecht5.encoder.prenet._get_feat_extract_output_lengths( + encoder_seq_length + ) + subsampled_encoder_key_length = model.speecht5.encoder.prenet._get_feat_extract_output_lengths( + encoder_key_length + ) + + with torch.no_grad(): + outputs = model(**self._prepare_for_class(inputs_dict, model_class)) + attentions = outputs.encoder_attentions if config.is_encoder_decoder else outputs.attentions + self.assertEqual(len(attentions), self.model_tester.num_hidden_layers) + + # check that output_attentions also work using config + del inputs_dict["output_attentions"] + config.output_attentions = True + model = model_class(config) + model.to(torch_device) + model.eval() + with torch.no_grad(): + outputs = model(**self._prepare_for_class(inputs_dict, model_class)) + attentions = outputs.encoder_attentions if config.is_encoder_decoder else outputs.attentions + self.assertEqual(len(attentions), self.model_tester.num_hidden_layers) + + self.assertListEqual( + list(attentions[0].shape[-3:]), + [self.model_tester.num_attention_heads, subsampled_encoder_seq_length, subsampled_encoder_key_length], + ) + out_len = len(outputs) + + correct_outlen = 5 + + # loss is at first position + if "labels" in inputs_dict: + correct_outlen += 1 # loss is added to beginning + if "past_key_values" in outputs: + correct_outlen += 1 # past_key_values have been returned + + self.assertEqual(out_len, correct_outlen) + + # decoder attentions + decoder_attentions = outputs.decoder_attentions + self.assertIsInstance(decoder_attentions, (list, tuple)) + self.assertEqual(len(decoder_attentions), self.model_tester.num_hidden_layers) + self.assertListEqual( + list(decoder_attentions[0].shape[-3:]), + [self.model_tester.num_attention_heads, decoder_seq_length, decoder_key_length], + ) + + # cross attentions + cross_attentions = outputs.cross_attentions + self.assertIsInstance(cross_attentions, (list, tuple)) + self.assertEqual(len(cross_attentions), self.model_tester.num_hidden_layers) + self.assertListEqual( + list(cross_attentions[0].shape[-3:]), + [ + self.model_tester.num_attention_heads, + decoder_seq_length, + subsampled_encoder_key_length, + ], + ) + + # Check attention is always last and order is fine + inputs_dict["output_attentions"] = True + inputs_dict["output_hidden_states"] = True + model = model_class(config) + model.to(torch_device) + model.eval() + with torch.no_grad(): + outputs = model(**self._prepare_for_class(inputs_dict, model_class)) + + added_hidden_states = 2 + self.assertEqual(out_len + added_hidden_states, len(outputs)) + + self_attentions = outputs.encoder_attentions if config.is_encoder_decoder else outputs.attentions + + self.assertEqual(len(self_attentions), self.model_tester.num_hidden_layers) + self.assertListEqual( + list(self_attentions[0].shape[-3:]), + [self.model_tester.num_attention_heads, subsampled_encoder_seq_length, subsampled_encoder_key_length], + ) + + 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", + "attention_mask", + "decoder_input_ids", + "decoder_attention_mask", + ] + expected_arg_names.extend( + ["head_mask", "decoder_head_mask", "cross_attn_head_mask", "encoder_outputs"] + if "head_mask" and "decoder_head_mask" and "cross_attn_head_mask" in arg_names + else ["encoder_outputs"] + ) + self.assertListEqual(arg_names[: len(expected_arg_names)], expected_arg_names) + + def test_hidden_states_output(self): + def check_hidden_states_output(inputs_dict, config, model_class): + model = model_class(config) + model.to(torch_device) + model.eval() + + with torch.no_grad(): + outputs = model(**self._prepare_for_class(inputs_dict, model_class)) + + hidden_states = outputs.encoder_hidden_states if config.is_encoder_decoder else outputs.hidden_states + + expected_num_layers = getattr( + self.model_tester, "expected_num_hidden_layers", self.model_tester.num_hidden_layers + 1 + ) + self.assertEqual(len(hidden_states), expected_num_layers) + + if hasattr(self.model_tester, "encoder_seq_length"): + seq_length = self.model_tester.encoder_seq_length + else: + seq_length = self.model_tester.seq_length + + subsampled_seq_length = model.speecht5.encoder.prenet._get_feat_extract_output_lengths(seq_length) + + self.assertListEqual( + list(hidden_states[0].shape[-2:]), + [subsampled_seq_length, self.model_tester.hidden_size], + ) + + if config.is_encoder_decoder: + hidden_states = outputs.decoder_hidden_states + + self.assertIsInstance(hidden_states, (list, tuple)) + self.assertEqual(len(hidden_states), expected_num_layers) + seq_len = getattr(self.model_tester, "seq_length", None) + decoder_seq_length = getattr(self.model_tester, "decoder_seq_length", seq_len) + + self.assertListEqual( + list(hidden_states[0].shape[-2:]), + [decoder_seq_length, self.model_tester.hidden_size], + ) + + config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common() + + for model_class in self.all_model_classes: + inputs_dict["output_hidden_states"] = True + check_hidden_states_output(inputs_dict, config, model_class) + + # check that output_hidden_states also work using config + del inputs_dict["output_hidden_states"] + config.output_hidden_states = True + + check_hidden_states_output(inputs_dict, config, model_class) + + 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.weight", + "masked_spec_embed", + "feature_projection.projection.weight", + "feature_projection.projection.bias", + ] + 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", + ) + else: + 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", + ) + + # this model has no inputs_embeds + def test_inputs_embeds(self): + pass + + def test_resize_embeddings_untied(self): + original_config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common() + if not self.test_resize_embeddings: + return + + original_config.tie_word_embeddings = False + + # if model cannot untied embeddings -> leave test + if original_config.tie_word_embeddings: + return + + for model_class in self.all_model_classes: + config = copy.deepcopy(original_config) + model = model_class(config).to(torch_device) + + # if no output embeddings -> leave test + if model.get_output_embeddings() is None: + continue + + # Check that resizing the token embeddings with a larger vocab size increases the model's vocab size + model_vocab_size = config.vocab_size + model.resize_token_embeddings(model_vocab_size + 10) + self.assertEqual(model.config.vocab_size, model_vocab_size + 10) + output_embeds = model.get_output_embeddings() + self.assertEqual(output_embeds.weight.shape[0], model_vocab_size + 10) + # Check bias if present + if output_embeds.bias is not None: + self.assertEqual(output_embeds.bias.shape[0], model_vocab_size + 10) + # Check that the model can still do a forward pass successfully (every parameter should be resized) + model(**self._prepare_for_class(inputs_dict, model_class)) + + # Check that resizing the token embeddings with a smaller vocab size decreases the model's vocab size + model.resize_token_embeddings(model_vocab_size - 15) + self.assertEqual(model.config.vocab_size, model_vocab_size - 15) + # Check that it actually resizes the embeddings matrix + output_embeds = model.get_output_embeddings() + self.assertEqual(output_embeds.weight.shape[0], model_vocab_size - 15) + # Check bias if present + if output_embeds.bias is not None: + self.assertEqual(output_embeds.bias.shape[0], model_vocab_size - 15) + # Check that the model can still do a forward pass successfully (every parameter should be resized) + if "decoder_input_ids" in inputs_dict: + inputs_dict["decoder_input_ids"].clamp_(max=model_vocab_size - 15 - 1) + # Check that the model can still do a forward pass successfully (every parameter should be resized) + model(**self._prepare_for_class(inputs_dict, model_class)) + + def test_resize_tokens_embeddings(self): + original_config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common() + if not self.test_resize_embeddings: + return + + for model_class in self.all_model_classes: + config = copy.deepcopy(original_config) + model = model_class(config) + model.to(torch_device) + + if self.model_tester.is_training is False: + model.eval() + + model_vocab_size = config.vocab_size + # Retrieve the embeddings and clone theme + model_embed = model.resize_token_embeddings(model_vocab_size) + cloned_embeddings = model_embed.weight.clone() + + # Check that resizing the token embeddings with a larger vocab size increases the model's vocab size + model_embed = model.resize_token_embeddings(model_vocab_size + 10) + self.assertEqual(model.config.vocab_size, model_vocab_size + 10) + # Check that it actually resizes the embeddings matrix + self.assertEqual(model_embed.weight.shape[0], cloned_embeddings.shape[0] + 10) + # Check that the model can still do a forward pass successfully (every parameter should be resized) + model(**self._prepare_for_class(inputs_dict, model_class)) + + # Check that resizing the token embeddings with a smaller vocab size decreases the model's vocab size + model_embed = model.resize_token_embeddings(model_vocab_size - 15) + self.assertEqual(model.config.vocab_size, model_vocab_size - 15) + # Check that it actually resizes the embeddings matrix + self.assertEqual(model_embed.weight.shape[0], cloned_embeddings.shape[0] - 15) + + # make sure that decoder_input_ids are resized + if "decoder_input_ids" in inputs_dict: + inputs_dict["decoder_input_ids"].clamp_(max=model_vocab_size - 15 - 1) + model(**self._prepare_for_class(inputs_dict, model_class)) + + # Check that adding and removing tokens has not modified the first part of the embedding matrix. + models_equal = True + for p1, p2 in zip(cloned_embeddings, model_embed.weight): + if p1.data.ne(p2.data).sum() > 0: + models_equal = False + + self.assertTrue(models_equal) + + def test_retain_grad_hidden_states_attentions(self): + # decoder cannot keep gradients + pass + + # training is not supported yet + def test_training(self): + pass + + def test_training_gradient_checkpointing(self): + pass + + # overwrite from test_modeling_common + def _mock_init_weights(self, module): + if hasattr(module, "weight") and module.weight is not None: + module.weight.data.fill_(3) + if hasattr(module, "weight_g") and module.weight_g is not None: + module.weight_g.data.fill_(3) + if hasattr(module, "weight_v") and module.weight_v is not None: + module.weight_v.data.fill_(3) + if hasattr(module, "bias") and module.bias is not None: + module.bias.data.fill_(3) + if hasattr(module, "masked_spec_embed") and module.masked_spec_embed is not None: + module.masked_spec_embed.data.fill_(3) + + +@require_torch +@require_torchaudio +@require_sentencepiece +@require_tokenizers +@slow +class SpeechT5ForSpeechToTextIntegrationTests(unittest.TestCase): + @cached_property + def default_processor(self): + return SpeechT5Processor.from_pretrained("microsoft/speecht5_asr") + + 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 + speech_samples = ds.sort("id").select(range(num_samples))[:num_samples]["audio"] + + return [x["array"] for x in speech_samples] + + def test_generation_librispeech(self): + model = SpeechT5ForSpeechToText.from_pretrained("microsoft/speecht5_asr") + model.to(torch_device) + processor = self.default_processor + + input_speech = self._load_datasamples(1) + + input_values = processor(audio=input_speech, return_tensors="pt").input_values.to(torch_device) + + generated_ids = model.generate(input_values) + generated_transcript = processor.batch_decode(generated_ids, skip_special_tokens=True) + + EXPECTED_TRANSCRIPTIONS = [ + "mister quilter is the apostle of the middle classes and we are glad to welcome his gospel" + ] + self.assertListEqual(generated_transcript, EXPECTED_TRANSCRIPTIONS) + + def test_generation_librispeech_batched(self): + model = SpeechT5ForSpeechToText.from_pretrained("microsoft/speecht5_asr") + model.to(torch_device) + processor = self.default_processor + + input_speech = self._load_datasamples(4) + + inputs = processor(audio=input_speech, return_tensors="pt", padding=True) + + input_values = inputs.input_values.to(torch_device) + attention_mask = inputs.attention_mask.to(torch_device) + + generated_ids = model.generate(input_values, attention_mask=attention_mask) + generated_transcripts = processor.batch_decode(generated_ids, skip_special_tokens=True) + + EXPECTED_TRANSCRIPTIONS = [ + "mister quilter is the apostle of the middle classes and we are glad to welcome his gospel", + "nor is mister quilter's manner less interesting than his matter", + "he tells us that at this festive season of the year with christmas and rosebeaf looming before us" + " similars drawn from eating and its results occur most readily to the mind", + "he has grave doubts whether sir frederick latin's work is really greek after all and can discover in it" + " but little of rocky ithica", + ] + self.assertListEqual(generated_transcripts, EXPECTED_TRANSCRIPTIONS) + + +@require_torch +class SpeechT5ForTextToSpeechTester: + def __init__( + self, + parent, + batch_size=13, + encoder_seq_length=7, + decoder_seq_length=1024, # speech is longer + is_training=False, + hidden_size=24, + num_hidden_layers=4, + num_attention_heads=2, + intermediate_size=4, + vocab_size=81, + num_mel_bins=20, + reduction_factor=2, + ): + self.parent = parent + self.batch_size = batch_size + self.encoder_seq_length = encoder_seq_length + self.decoder_seq_length = decoder_seq_length + self.is_training = is_training + self.hidden_size = hidden_size + self.num_hidden_layers = num_hidden_layers + self.num_attention_heads = num_attention_heads + self.intermediate_size = intermediate_size + self.vocab_size = vocab_size + self.num_mel_bins = num_mel_bins + self.reduction_factor = reduction_factor + + def prepare_config_and_inputs(self): + input_ids = ids_tensor([self.batch_size, self.encoder_seq_length], self.vocab_size).clamp(2) + attention_mask = random_attention_mask([self.batch_size, self.encoder_seq_length]) + + decoder_input_values = floats_tensor([self.batch_size, self.decoder_seq_length, self.num_mel_bins], scale=1.0) + decoder_attention_mask = random_attention_mask([self.batch_size, self.decoder_seq_length]) + + config = self.get_config() + inputs_dict = prepare_inputs_dict( + config, + input_ids=input_ids, + decoder_input_values=decoder_input_values, + attention_mask=attention_mask, + decoder_attention_mask=decoder_attention_mask, + ) + 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 SpeechT5Config( + hidden_size=self.hidden_size, + encoder_layers=self.num_hidden_layers, + decoder_layers=self.num_hidden_layers, + encoder_attention_heads=self.num_attention_heads, + decoder_attention_heads=self.num_attention_heads, + encoder_ffn_dim=self.intermediate_size, + decoder_ffn_dim=self.intermediate_size, + vocab_size=self.vocab_size, + num_mel_bins=self.num_mel_bins, + reduction_factor=self.reduction_factor, + ) + + def create_and_check_model_forward(self, config, inputs_dict): + model = SpeechT5ForTextToSpeech(config=config).to(torch_device).eval() + + input_ids = inputs_dict["input_ids"] + attention_mask = inputs_dict["attention_mask"] + decoder_input_values = inputs_dict["decoder_input_values"] + + result = model(input_ids, attention_mask=attention_mask, decoder_input_values=decoder_input_values) + self.parent.assertEqual( + result.spectrogram.shape, + (self.batch_size, self.decoder_seq_length * self.reduction_factor, self.num_mel_bins), + ) + + +@require_torch +class SpeechT5ForTextToSpeechTest(ModelTesterMixin, unittest.TestCase): + all_model_classes = (SpeechT5ForTextToSpeech,) if is_torch_available() else () + all_generative_model_classes = (SpeechT5ForTextToSpeech,) if is_torch_available() else () + is_encoder_decoder = True + test_pruning = False + test_headmasking = False + + input_name = "input_ids" + + def setUp(self): + self.model_tester = SpeechT5ForTextToSpeechTester(self) + self.config_tester = ConfigTester(self, config_class=SpeechT5Config, hidden_size=37) + + def test_config(self): + self.config_tester.run_common_tests() + + def test_save_load_strict(self): + config, inputs_dict = self.model_tester.prepare_config_and_inputs() + for model_class in self.all_model_classes: + model = model_class(config) + + with tempfile.TemporaryDirectory() as tmpdirname: + model.save_pretrained(tmpdirname) + model2, info = model_class.from_pretrained(tmpdirname, output_loading_info=True) + self.assertEqual(info["missing_keys"], []) + + 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) + + # skipped because there is always dropout in SpeechT5SpeechDecoderPrenet + def test_decoder_model_past_with_large_inputs(self): + pass + + # skipped because there is always dropout in SpeechT5SpeechDecoderPrenet + def test_determinism(self): + pass + + 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_ids", + "attention_mask", + "decoder_input_values", + "decoder_attention_mask", + ] + expected_arg_names.extend( + ["head_mask", "decoder_head_mask", "cross_attn_head_mask", "encoder_outputs"] + if "head_mask" and "decoder_head_mask" and "cross_attn_head_mask" in arg_names + else ["encoder_outputs"] + ) + self.assertListEqual(arg_names[: len(expected_arg_names)], expected_arg_names) + + 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.weight", + ] + 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", + ) + else: + 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", + ) + + # this model has no inputs_embeds + def test_inputs_embeds(self): + pass + + # skipped because there is always dropout in SpeechT5SpeechDecoderPrenet + def test_model_outputs_equivalence(self): + pass + + # skipped because there is always dropout in SpeechT5SpeechDecoderPrenet + def test_save_load(self): + pass + + def test_retain_grad_hidden_states_attentions(self): + # decoder cannot keep gradients + pass + + @slow + def test_torchscript_output_attentions(self): + # disabled because this model doesn't have decoder_input_ids + pass + + @slow + def test_torchscript_output_hidden_state(self): + # disabled because this model doesn't have decoder_input_ids + pass + + @slow + def test_torchscript_simple(self): + # disabled because this model doesn't have decoder_input_ids + pass + + # training is not supported yet + def test_training(self): + pass + + def test_training_gradient_checkpointing(self): + pass + + # overwrite from test_modeling_common + def _mock_init_weights(self, module): + if hasattr(module, "weight") and module.weight is not None: + module.weight.data.fill_(3) + if hasattr(module, "weight_g") and module.weight_g is not None: + module.weight_g.data.fill_(3) + if hasattr(module, "weight_v") and module.weight_v is not None: + module.weight_v.data.fill_(3) + if hasattr(module, "bias") and module.bias is not None: + module.bias.data.fill_(3) + + +@require_torch +@require_torchaudio +@require_sentencepiece +@require_tokenizers +@slow +class SpeechT5ForTextToSpeechIntegrationTests(unittest.TestCase): + @cached_property + def default_processor(self): + return SpeechT5Processor.from_pretrained("microsoft/speecht5_tts") + + def test_generation(self): + model = SpeechT5ForTextToSpeech.from_pretrained("microsoft/speecht5_tts") + model.to(torch_device) + processor = self.default_processor + + input_text = "mister quilter is the apostle of the middle classes and we are glad to welcome his gospel" + input_ids = processor(text=input_text, return_tensors="pt").input_ids.to(torch_device) + + generated_speech = model.generate_speech(input_ids) + self.assertEqual(generated_speech.shape, (1800, model.config.num_mel_bins)) + + +@require_torch +class SpeechT5ForSpeechToSpeechTester: + def __init__( + self, + parent, + batch_size=13, + encoder_seq_length=1024, # speech is longer + decoder_seq_length=1024, + is_training=False, + hidden_size=24, + num_hidden_layers=4, + num_attention_heads=2, + intermediate_size=4, + conv_dim=(32, 32, 32), + conv_stride=(4, 4, 4), + conv_kernel=(8, 8, 8), + conv_bias=False, + num_conv_pos_embeddings=16, + num_conv_pos_embedding_groups=2, + vocab_size=81, + num_mel_bins=20, + reduction_factor=2, + ): + self.parent = parent + self.batch_size = batch_size + self.encoder_seq_length = encoder_seq_length + self.decoder_seq_length = decoder_seq_length + self.is_training = is_training + self.hidden_size = hidden_size + self.num_hidden_layers = num_hidden_layers + self.num_attention_heads = num_attention_heads + self.intermediate_size = intermediate_size + self.conv_dim = conv_dim + self.conv_stride = conv_stride + self.conv_kernel = conv_kernel + self.conv_bias = conv_bias + self.num_conv_pos_embeddings = num_conv_pos_embeddings + self.num_conv_pos_embedding_groups = num_conv_pos_embedding_groups + self.vocab_size = vocab_size + self.num_mel_bins = num_mel_bins + self.reduction_factor = reduction_factor + + def prepare_config_and_inputs(self): + input_values = floats_tensor([self.batch_size, self.encoder_seq_length], scale=1.0) + attention_mask = random_attention_mask([self.batch_size, self.encoder_seq_length]) + + decoder_input_values = floats_tensor([self.batch_size, self.decoder_seq_length, self.num_mel_bins], scale=1.0) + decoder_attention_mask = random_attention_mask([self.batch_size, self.decoder_seq_length]) + + config = self.get_config() + inputs_dict = prepare_inputs_dict( + config, + input_values=input_values, + decoder_input_values=decoder_input_values, + attention_mask=attention_mask, + decoder_attention_mask=decoder_attention_mask, + ) + 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 SpeechT5Config( + hidden_size=self.hidden_size, + encoder_layers=self.num_hidden_layers, + decoder_layers=self.num_hidden_layers, + encoder_attention_heads=self.num_attention_heads, + decoder_attention_heads=self.num_attention_heads, + encoder_ffn_dim=self.intermediate_size, + decoder_ffn_dim=self.intermediate_size, + conv_dim=self.conv_dim, + conv_stride=self.conv_stride, + conv_kernel=self.conv_kernel, + conv_bias=self.conv_bias, + num_conv_pos_embeddings=self.num_conv_pos_embeddings, + num_conv_pos_embedding_groups=self.num_conv_pos_embedding_groups, + vocab_size=self.vocab_size, + num_mel_bins=self.num_mel_bins, + reduction_factor=self.reduction_factor, + ) + + def create_and_check_model_forward(self, config, inputs_dict): + model = SpeechT5ForSpeechToSpeech(config=config).to(torch_device).eval() + + input_values = inputs_dict["input_values"] + attention_mask = inputs_dict["attention_mask"] + decoder_input_values = inputs_dict["decoder_input_values"] + + result = model(input_values, attention_mask=attention_mask, decoder_input_values=decoder_input_values) + self.parent.assertEqual( + result.spectrogram.shape, + (self.batch_size, self.decoder_seq_length * self.reduction_factor, self.num_mel_bins), + ) + + +@require_torch +class SpeechT5ForSpeechToSpeechTest(ModelTesterMixin, unittest.TestCase): + all_model_classes = (SpeechT5ForSpeechToSpeech,) if is_torch_available() else () + all_generative_model_classes = (SpeechT5ForSpeechToSpeech,) if is_torch_available() else () + is_encoder_decoder = True + test_pruning = False + test_headmasking = False + test_resize_embeddings = False + + input_name = "input_values" + + def setUp(self): + self.model_tester = SpeechT5ForSpeechToSpeechTester(self) + self.config_tester = ConfigTester(self, config_class=SpeechT5Config, hidden_size=37) + + def test_config(self): + self.config_tester.run_common_tests() + + def test_save_load_strict(self): + config, inputs_dict = self.model_tester.prepare_config_and_inputs() + for model_class in self.all_model_classes: + model = model_class(config) + + with tempfile.TemporaryDirectory() as tmpdirname: + model.save_pretrained(tmpdirname) + model2, info = model_class.from_pretrained(tmpdirname, output_loading_info=True) + self.assertEqual(info["missing_keys"], []) + + 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) + + # skipped because there is always dropout in SpeechT5SpeechDecoderPrenet + def test_decoder_model_past_with_large_inputs(self): + pass + + # skipped because there is always dropout in SpeechT5SpeechDecoderPrenet + def test_determinism(self): + pass + + def test_attention_outputs(self): + config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common() + config.return_dict = True + + seq_len = getattr(self.model_tester, "seq_length", None) + decoder_seq_length = getattr(self.model_tester, "decoder_seq_length", seq_len) + encoder_seq_length = getattr(self.model_tester, "encoder_seq_length", seq_len) + decoder_key_length = getattr(self.model_tester, "decoder_key_length", decoder_seq_length) + encoder_key_length = getattr(self.model_tester, "key_length", encoder_seq_length) + + for model_class in self.all_model_classes: + inputs_dict["output_attentions"] = True + inputs_dict["output_hidden_states"] = False + config.return_dict = True + model = model_class(config) + model.to(torch_device) + model.eval() + + subsampled_encoder_seq_length = model.speecht5.encoder.prenet._get_feat_extract_output_lengths( + encoder_seq_length + ) + subsampled_encoder_key_length = model.speecht5.encoder.prenet._get_feat_extract_output_lengths( + encoder_key_length + ) + + with torch.no_grad(): + outputs = model(**self._prepare_for_class(inputs_dict, model_class)) + attentions = outputs.encoder_attentions if config.is_encoder_decoder else outputs.attentions + self.assertEqual(len(attentions), self.model_tester.num_hidden_layers) + + # check that output_attentions also work using config + del inputs_dict["output_attentions"] + config.output_attentions = True + model = model_class(config) + model.to(torch_device) + model.eval() + with torch.no_grad(): + outputs = model(**self._prepare_for_class(inputs_dict, model_class)) + attentions = outputs.encoder_attentions if config.is_encoder_decoder else outputs.attentions + self.assertEqual(len(attentions), self.model_tester.num_hidden_layers) + + self.assertListEqual( + list(attentions[0].shape[-3:]), + [self.model_tester.num_attention_heads, subsampled_encoder_seq_length, subsampled_encoder_key_length], + ) + out_len = len(outputs) + + correct_outlen = 5 + + # loss is at first position + if "labels" in inputs_dict: + correct_outlen += 1 # loss is added to beginning + if "past_key_values" in outputs: + correct_outlen += 1 # past_key_values have been returned + + self.assertEqual(out_len, correct_outlen) + + # decoder attentions + decoder_attentions = outputs.decoder_attentions + self.assertIsInstance(decoder_attentions, (list, tuple)) + self.assertEqual(len(decoder_attentions), self.model_tester.num_hidden_layers) + self.assertListEqual( + list(decoder_attentions[0].shape[-3:]), + [self.model_tester.num_attention_heads, decoder_seq_length, decoder_key_length], + ) + + # cross attentions + cross_attentions = outputs.cross_attentions + self.assertIsInstance(cross_attentions, (list, tuple)) + self.assertEqual(len(cross_attentions), self.model_tester.num_hidden_layers) + self.assertListEqual( + list(cross_attentions[0].shape[-3:]), + [ + self.model_tester.num_attention_heads, + decoder_seq_length, + subsampled_encoder_key_length, + ], + ) + + # Check attention is always last and order is fine + inputs_dict["output_attentions"] = True + inputs_dict["output_hidden_states"] = True + model = model_class(config) + model.to(torch_device) + model.eval() + with torch.no_grad(): + outputs = model(**self._prepare_for_class(inputs_dict, model_class)) + + added_hidden_states = 2 + self.assertEqual(out_len + added_hidden_states, len(outputs)) + + self_attentions = outputs.encoder_attentions if config.is_encoder_decoder else outputs.attentions + + self.assertEqual(len(self_attentions), self.model_tester.num_hidden_layers) + self.assertListEqual( + list(self_attentions[0].shape[-3:]), + [self.model_tester.num_attention_heads, subsampled_encoder_seq_length, subsampled_encoder_key_length], + ) + + 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", + "attention_mask", + "decoder_input_values", + "decoder_attention_mask", + ] + expected_arg_names.extend( + ["head_mask", "decoder_head_mask", "cross_attn_head_mask", "encoder_outputs"] + if "head_mask" and "decoder_head_mask" and "cross_attn_head_mask" in arg_names + else ["encoder_outputs"] + ) + self.assertListEqual(arg_names[: len(expected_arg_names)], expected_arg_names) + + def test_hidden_states_output(self): + def check_hidden_states_output(inputs_dict, config, model_class): + model = model_class(config) + model.to(torch_device) + model.eval() + + with torch.no_grad(): + outputs = model(**self._prepare_for_class(inputs_dict, model_class)) + + hidden_states = outputs.encoder_hidden_states if config.is_encoder_decoder else outputs.hidden_states + + expected_num_layers = getattr( + self.model_tester, "expected_num_hidden_layers", self.model_tester.num_hidden_layers + 1 + ) + self.assertEqual(len(hidden_states), expected_num_layers) + + if hasattr(self.model_tester, "encoder_seq_length"): + seq_length = self.model_tester.encoder_seq_length + else: + seq_length = self.model_tester.seq_length + + subsampled_seq_length = model.speecht5.encoder.prenet._get_feat_extract_output_lengths(seq_length) + + self.assertListEqual( + list(hidden_states[0].shape[-2:]), + [subsampled_seq_length, self.model_tester.hidden_size], + ) + + if config.is_encoder_decoder: + hidden_states = outputs.decoder_hidden_states + + self.assertIsInstance(hidden_states, (list, tuple)) + self.assertEqual(len(hidden_states), expected_num_layers) + seq_len = getattr(self.model_tester, "seq_length", None) + decoder_seq_length = getattr(self.model_tester, "decoder_seq_length", seq_len) + + self.assertListEqual( + list(hidden_states[0].shape[-2:]), + [decoder_seq_length, self.model_tester.hidden_size], + ) + + config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common() + + for model_class in self.all_model_classes: + inputs_dict["output_hidden_states"] = True + check_hidden_states_output(inputs_dict, config, model_class) + + # check that output_hidden_states also work using config + del inputs_dict["output_hidden_states"] + config.output_hidden_states = True + + check_hidden_states_output(inputs_dict, config, model_class) + + 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.weight", + "masked_spec_embed", + "feature_projection.projection.weight", + "feature_projection.projection.bias", + ] + 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", + ) + else: + 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", + ) + + # this model has no inputs_embeds + def test_inputs_embeds(self): + pass + + # this model has no input embeddings + def test_model_common_attributes(self): + pass + + # skipped because there is always dropout in SpeechT5SpeechDecoderPrenet + def test_model_outputs_equivalence(self): + pass + + def test_retain_grad_hidden_states_attentions(self): + # decoder cannot keep gradients + pass + + # skipped because there is always dropout in SpeechT5SpeechDecoderPrenet + def test_save_load(self): + pass + + @slow + def test_torchscript_output_attentions(self): + # disabled because this model doesn't have decoder_input_ids + pass + + @slow + def test_torchscript_output_hidden_state(self): + # disabled because this model doesn't have decoder_input_ids + pass + + @slow + def test_torchscript_simple(self): + # disabled because this model doesn't have decoder_input_ids + pass + + # training is not supported yet + def test_training(self): + pass + + def test_training_gradient_checkpointing(self): + pass + + # overwrite from test_modeling_common + def _mock_init_weights(self, module): + if hasattr(module, "weight") and module.weight is not None: + module.weight.data.fill_(3) + if hasattr(module, "weight_g") and module.weight_g is not None: + module.weight_g.data.fill_(3) + if hasattr(module, "weight_v") and module.weight_v is not None: + module.weight_v.data.fill_(3) + if hasattr(module, "bias") and module.bias is not None: + module.bias.data.fill_(3) + if hasattr(module, "masked_spec_embed") and module.masked_spec_embed is not None: + module.masked_spec_embed.data.fill_(3) + + +@require_torch +@require_torchaudio +@require_sentencepiece +@require_tokenizers +@slow +class SpeechT5ForSpeechToSpeechIntegrationTests(unittest.TestCase): + @cached_property + def default_processor(self): + return SpeechT5Processor.from_pretrained("microsoft/speecht5_vc") + + 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 + speech_samples = ds.sort("id").select(range(num_samples))[:num_samples]["audio"] + + return [x["array"] for x in speech_samples] + + def test_generation_librispeech(self): + model = SpeechT5ForSpeechToSpeech.from_pretrained("microsoft/speecht5_vc") + model.to(torch_device) + processor = self.default_processor + + input_speech = self._load_datasamples(1) + input_values = processor(audio=input_speech, return_tensors="pt").input_values.to(torch_device) + + speaker_embeddings = torch.zeros((1, 512)) + generated_speech = model.generate_speech(input_values, speaker_embeddings=speaker_embeddings) + + self.assertEqual(generated_speech.shape[1], model.config.num_mel_bins) + self.assertGreaterEqual(generated_speech.shape[0], 300) + self.assertLessEqual(generated_speech.shape[0], 310) + + +class SpeechT5HifiGanTester: + def __init__( + self, + parent, + batch_size=13, + seq_length=7, + is_training=False, + num_mel_bins=20, + ): + self.parent = parent + self.batch_size = batch_size + self.seq_length = seq_length + self.is_training = is_training + self.num_mel_bins = num_mel_bins + + def prepare_config_and_inputs(self): + input_values = floats_tensor([self.seq_length, self.num_mel_bins], scale=1.0) + config = self.get_config() + return config, input_values + + def get_config(self): + return SpeechT5HifiGanConfig( + model_in_dim=self.num_mel_bins, + ) + + def create_and_check_model(self, config, input_values): + model = SpeechT5HifiGan(config=config).to(torch_device).eval() + result = model(input_values) + self.parent.assertEqual(result.shape, (self.seq_length * 256,)) + + def prepare_config_and_inputs_for_common(self): + config, input_values = self.prepare_config_and_inputs() + inputs_dict = {"spectrogram": input_values} + return config, inputs_dict + + +@require_torch +class SpeechT5HifiGanTest(ModelTesterMixin, unittest.TestCase): + all_model_classes = (SpeechT5HifiGan,) if is_torch_available() else () + test_torchscript = False + test_pruning = False + test_resize_embeddings = False + test_resize_position_embeddings = False + test_head_masking = False + test_mismatched_shapes = False + test_missing_keys = False + test_model_parallel = False + is_encoder_decoder = False + has_attentions = False + + input_name = "spectrogram" + + def setUp(self): + self.model_tester = SpeechT5HifiGanTester(self) + self.config_tester = ConfigTester(self, config_class=SpeechT5HifiGanConfig) + + def test_config(self): + self.config_tester.create_and_test_config_to_json_string() + self.config_tester.create_and_test_config_to_json_file() + self.config_tester.create_and_test_config_from_and_save_pretrained() + self.config_tester.create_and_test_config_from_and_save_pretrained_subfolder() + self.config_tester.create_and_test_config_with_num_labels() + self.config_tester.check_config_can_be_init_without_params() + self.config_tester.check_config_arguments_init() + + def test_model(self): + config_and_inputs = self.model_tester.prepare_config_and_inputs() + self.model_tester.create_and_check_model(*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 = [ + "spectrogram", + ] + self.assertListEqual(arg_names[: len(expected_arg_names)], expected_arg_names) + + # this model does not output hidden states + def test_hidden_states_output(self): + pass + + # skip + def test_initialization(self): + pass + + # this model has no inputs_embeds + def test_inputs_embeds(self): + pass + + # this model has no input embeddings + def test_model_common_attributes(self): + pass + + # skip as this model doesn't support all arguments tested + def test_model_outputs_equivalence(self): + pass + + # this model does not output hidden states + def test_retain_grad_hidden_states_attentions(self): + pass + + # skip because it fails on automapping of SpeechT5HifiGanConfig + def test_save_load_fast_init_from_base(self): + pass + + # skip because it fails on automapping of SpeechT5HifiGanConfig + def test_save_load_fast_init_to_base(self): + pass diff --git a/tests/models/speecht5/test_processor_speecht5.py b/tests/models/speecht5/test_processor_speecht5.py new file mode 100644 index 00000000000..d3f28738cb7 --- /dev/null +++ b/tests/models/speecht5/test_processor_speecht5.py @@ -0,0 +1,187 @@ +# Copyright 2022 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +"""Tests for the SpeechT5 processors.""" + +import json +import os +import shutil +import tempfile +import unittest + +from transformers import is_speech_available, is_torch_available +from transformers.models.speecht5 import SpeechT5Tokenizer +from transformers.testing_utils import get_tests_dir, require_torch, require_torchaudio +from transformers.utils import FEATURE_EXTRACTOR_NAME + + +if is_speech_available() and is_torch_available(): + from transformers import SpeechT5FeatureExtractor, SpeechT5Processor + + from .test_feature_extraction_speecht5 import floats_list + + +SAMPLE_VOCAB = get_tests_dir("fixtures/test_sentencepiece_bpe_char.model") + + +@require_torch +@require_torchaudio +class SpeechT5ProcessorTest(unittest.TestCase): + def setUp(self): + self.tmpdirname = tempfile.mkdtemp() + + tokenizer = SpeechT5Tokenizer(SAMPLE_VOCAB) + tokenizer.save_pretrained(self.tmpdirname) + + feature_extractor_map = { + "feature_size": 1, + "padding_value": 0.0, + "sampling_rate": 16000, + "do_normalize": False, + "num_mel_bins": 80, + "hop_length": 16, + "win_length": 64, + "win_function": "hann_window", + "frame_signal_scale": 1.0, + "fmin": 80, + "fmax": 7600, + "mel_floor": 1e-10, + "reduction_factor": 2, + "return_attention_mask": True, + } + + self.feature_extraction_file = os.path.join(self.tmpdirname, FEATURE_EXTRACTOR_NAME) + with open(self.feature_extraction_file, "w", encoding="utf-8") as fp: + fp.write(json.dumps(feature_extractor_map) + "\n") + + def get_tokenizer(self, **kwargs): + return SpeechT5Tokenizer.from_pretrained(self.tmpdirname, **kwargs) + + def get_feature_extractor(self, **kwargs): + return SpeechT5FeatureExtractor.from_pretrained(self.tmpdirname, **kwargs) + + def tearDown(self): + shutil.rmtree(self.tmpdirname) + + def test_save_load_pretrained_default(self): + tokenizer = self.get_tokenizer() + feature_extractor = self.get_feature_extractor() + + processor = SpeechT5Processor(tokenizer=tokenizer, feature_extractor=feature_extractor) + + processor.save_pretrained(self.tmpdirname) + processor = SpeechT5Processor.from_pretrained(self.tmpdirname) + + self.assertEqual(processor.tokenizer.get_vocab(), tokenizer.get_vocab()) + self.assertIsInstance(processor.tokenizer, SpeechT5Tokenizer) + + self.assertEqual(processor.feature_extractor.to_json_string(), feature_extractor.to_json_string()) + self.assertIsInstance(processor.feature_extractor, SpeechT5FeatureExtractor) + + def test_save_load_pretrained_additional_features(self): + processor = SpeechT5Processor(tokenizer=self.get_tokenizer(), feature_extractor=self.get_feature_extractor()) + processor.save_pretrained(self.tmpdirname) + + tokenizer_add_kwargs = self.get_tokenizer(bos_token="(BOS)", eos_token="(EOS)") + feature_extractor_add_kwargs = self.get_feature_extractor(do_normalize=False, padding_value=1.0) + + processor = SpeechT5Processor.from_pretrained( + self.tmpdirname, bos_token="(BOS)", eos_token="(EOS)", do_normalize=False, padding_value=1.0 + ) + + self.assertEqual(processor.tokenizer.get_vocab(), tokenizer_add_kwargs.get_vocab()) + self.assertIsInstance(processor.tokenizer, SpeechT5Tokenizer) + + self.assertEqual(processor.feature_extractor.to_json_string(), feature_extractor_add_kwargs.to_json_string()) + self.assertIsInstance(processor.feature_extractor, SpeechT5FeatureExtractor) + + def test_feature_extractor(self): + feature_extractor = self.get_feature_extractor() + tokenizer = self.get_tokenizer() + + processor = SpeechT5Processor(tokenizer=tokenizer, feature_extractor=feature_extractor) + + raw_speech = floats_list((3, 1000)) + + input_feat_extract = feature_extractor(audio=raw_speech, return_tensors="np") + input_processor = processor(audio=raw_speech, return_tensors="np") + + for key in input_feat_extract.keys(): + self.assertAlmostEqual(input_feat_extract[key].sum(), input_processor[key].sum(), delta=1e-2) + + def test_feature_extractor_target(self): + feature_extractor = self.get_feature_extractor() + tokenizer = self.get_tokenizer() + + processor = SpeechT5Processor(tokenizer=tokenizer, feature_extractor=feature_extractor) + + raw_speech = floats_list((3, 1000)) + + input_feat_extract = feature_extractor(audio_target=raw_speech, return_tensors="np") + input_processor = processor(audio_target=raw_speech, return_tensors="np") + + for key in input_feat_extract.keys(): + self.assertAlmostEqual(input_feat_extract[key].sum(), input_processor[key].sum(), delta=1e-2) + + def test_tokenizer(self): + feature_extractor = self.get_feature_extractor() + tokenizer = self.get_tokenizer() + + processor = SpeechT5Processor(tokenizer=tokenizer, feature_extractor=feature_extractor) + + input_str = "This is a test string" + + encoded_processor = processor(text=input_str) + encoded_tok = tokenizer(input_str) + + for key in encoded_tok.keys(): + self.assertListEqual(encoded_tok[key], encoded_processor[key]) + + def test_tokenizer_target(self): + feature_extractor = self.get_feature_extractor() + tokenizer = self.get_tokenizer() + + processor = SpeechT5Processor(tokenizer=tokenizer, feature_extractor=feature_extractor) + + input_str = "This is a test string" + + encoded_processor = processor(text_target=input_str) + encoded_tok = tokenizer(input_str) + + for key in encoded_tok.keys(): + self.assertListEqual(encoded_tok[key], encoded_processor[key]) + + def test_tokenizer_decode(self): + feature_extractor = self.get_feature_extractor() + tokenizer = self.get_tokenizer() + + processor = SpeechT5Processor(tokenizer=tokenizer, feature_extractor=feature_extractor) + + predicted_ids = [[1, 4, 5, 8, 1, 0, 8], [3, 4, 3, 1, 1, 8, 9]] + + decoded_processor = processor.batch_decode(predicted_ids) + decoded_tok = tokenizer.batch_decode(predicted_ids) + + self.assertListEqual(decoded_tok, decoded_processor) + + def test_model_input_names(self): + feature_extractor = self.get_feature_extractor() + tokenizer = self.get_tokenizer() + + processor = SpeechT5Processor(tokenizer=tokenizer, feature_extractor=feature_extractor) + + self.assertListEqual( + processor.model_input_names, + feature_extractor.model_input_names, + msg="`processor` and `feature_extractor` model input names do not match", + ) diff --git a/tests/models/speecht5/test_tokenization_speecht5.py b/tests/models/speecht5/test_tokenization_speecht5.py new file mode 100644 index 00000000000..ec8dca122ff --- /dev/null +++ b/tests/models/speecht5/test_tokenization_speecht5.py @@ -0,0 +1,186 @@ +# coding=utf-8 +# Copyright 2022 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +"""Tests for the SpeechT5 tokenizers.""" + +import unittest + +from transformers import SPIECE_UNDERLINE +from transformers.models.speecht5 import SpeechT5Tokenizer +from transformers.testing_utils import get_tests_dir, require_sentencepiece, require_tokenizers, slow + +from ...test_tokenization_common import TokenizerTesterMixin + + +SAMPLE_VOCAB = get_tests_dir("fixtures/test_sentencepiece_bpe_char.model") + + +@require_sentencepiece +@require_tokenizers +class SpeechT5TokenizerTest(TokenizerTesterMixin, unittest.TestCase): + tokenizer_class = SpeechT5Tokenizer + test_rust_tokenizer = False + test_sentencepiece = True + + def setUp(self): + super().setUp() + + # We have a SentencePiece fixture for testing + tokenizer = SpeechT5Tokenizer(SAMPLE_VOCAB) + tokenizer.save_pretrained(self.tmpdirname) + + def get_input_output_texts(self, tokenizer): + input_text = "this is a test" + output_text = "this is a test" + return input_text, output_text + + def get_clean_sequence(self, tokenizer, with_prefix_space=False, max_length=20, min_length=5): + input_text, output_text = self.get_input_output_texts(tokenizer) + ids = tokenizer.encode(output_text, add_special_tokens=False) + text = tokenizer.decode(ids, clean_up_tokenization_spaces=False) + return text, ids + + def test_convert_token_and_id(self): + """Test ``_convert_token_to_id`` and ``_convert_id_to_token``.""" + token = "" + token_id = 1 + + self.assertEqual(self.get_tokenizer()._convert_token_to_id(token), token_id) + self.assertEqual(self.get_tokenizer()._convert_id_to_token(token_id), token) + + def test_get_vocab(self): + vocab_keys = list(self.get_tokenizer().get_vocab().keys()) + + self.assertEqual(vocab_keys[0], "") + self.assertEqual(vocab_keys[1], "") + self.assertEqual(vocab_keys[-2], "œ") + self.assertEqual(len(vocab_keys), 79) + + def test_vocab_size(self): + self.assertEqual(self.get_tokenizer().vocab_size, 79) + + def test_add_tokens_tokenizer(self): + tokenizers = self.get_tokenizers(do_lower_case=False) + for tokenizer in tokenizers: + with self.subTest(f"{tokenizer.__class__.__name__}"): + vocab_size = tokenizer.vocab_size + all_size = len(tokenizer) + + self.assertNotEqual(vocab_size, 0) + + # We usually have added tokens from the start in tests because our vocab fixtures are + # smaller than the original vocabs - let's not assert this + # self.assertEqual(vocab_size, all_size) + + new_toks = ["aaaaa bbbbbb", "cccccccccdddddddd"] + added_toks = tokenizer.add_tokens(new_toks) + vocab_size_2 = tokenizer.vocab_size + all_size_2 = len(tokenizer) + + self.assertNotEqual(vocab_size_2, 0) + self.assertEqual(vocab_size, vocab_size_2) + self.assertEqual(added_toks, len(new_toks)) + self.assertEqual(all_size_2, all_size + len(new_toks)) + + tokens = tokenizer.encode("aaaaa bbbbbb low cccccccccdddddddd l", add_special_tokens=False) + + self.assertGreaterEqual(len(tokens), 4) + self.assertGreater(tokens[0], tokenizer.vocab_size - 1) + self.assertGreater(tokens[-3], tokenizer.vocab_size - 1) + + new_toks_2 = {"eos_token": ">>>>|||<||<<|<<", "pad_token": "<<<<<|||>|>>>>|>"} + added_toks_2 = tokenizer.add_special_tokens(new_toks_2) + vocab_size_3 = tokenizer.vocab_size + all_size_3 = len(tokenizer) + + self.assertNotEqual(vocab_size_3, 0) + self.assertEqual(vocab_size, vocab_size_3) + self.assertEqual(added_toks_2, len(new_toks_2)) + self.assertEqual(all_size_3, all_size_2 + len(new_toks_2)) + + tokens = tokenizer.encode( + ">>>>|||<||<<|<< aaaaabbbbbb low cccccccccdddddddd <<<<<|||>|>>>>|> l", add_special_tokens=False + ) + + self.assertGreaterEqual(len(tokens), 6) + self.assertGreater(tokens[0], tokenizer.vocab_size - 1) + self.assertGreater(tokens[0], tokens[1]) + self.assertGreater(tokens[-3], tokenizer.vocab_size - 1) + self.assertGreater(tokens[-3], tokens[-4]) + self.assertEqual(tokens[0], tokenizer.eos_token_id) + self.assertEqual(tokens[-3], tokenizer.pad_token_id) + + def test_pickle_subword_regularization_tokenizer(self): + pass + + def test_subword_regularization_tokenizer(self): + pass + + def test_full_tokenizer(self): + tokenizer = self.get_tokenizer() + + tokens = tokenizer.tokenize("This is a test") + # fmt: off + self.assertListEqual(tokens, [SPIECE_UNDERLINE, 'T', 'h', 'i', 's', SPIECE_UNDERLINE, 'i', 's', SPIECE_UNDERLINE, 'a', SPIECE_UNDERLINE, 't', 'e', 's', 't']) + # fmt: on + + self.assertListEqual( + tokenizer.convert_tokens_to_ids(tokens), + [4, 32, 11, 10, 12, 4, 10, 12, 4, 7, 4, 6, 5, 12, 6], + ) + + tokens = tokenizer.tokenize("I was born in 92000, and this is falsé.") + self.assertListEqual( + tokens, + # fmt: off + [SPIECE_UNDERLINE, 'I', SPIECE_UNDERLINE, 'w', 'a', 's', SPIECE_UNDERLINE, 'b', 'o', 'r', 'n', SPIECE_UNDERLINE, 'i', 'n', SPIECE_UNDERLINE, '92000', ',', SPIECE_UNDERLINE, 'a', 'n', 'd', SPIECE_UNDERLINE, 't', 'h', 'i', 's', SPIECE_UNDERLINE, 'i', 's', SPIECE_UNDERLINE, 'f', 'a', 'l', 's', 'é', '.'] + # fmt: on + ) + + ids = tokenizer.convert_tokens_to_ids(tokens) + # fmt: off + self.assertListEqual(ids, [4, 30, 4, 20, 7, 12, 4, 25, 8, 13, 9, 4, 10, 9, 4, 3, 23, 4, 7, 9, 14, 4, 6, 11, 10, 12, 4, 10, 12, 4, 19, 7, 15, 12, 73, 26]) + # fmt: on + + back_tokens = tokenizer.convert_ids_to_tokens(ids) + self.assertListEqual( + back_tokens, + # fmt: off + [SPIECE_UNDERLINE, 'I', SPIECE_UNDERLINE, 'w', 'a', 's', SPIECE_UNDERLINE, 'b', 'o', 'r', 'n', SPIECE_UNDERLINE, 'i', 'n', SPIECE_UNDERLINE, '', ',', SPIECE_UNDERLINE, 'a', 'n', 'd', SPIECE_UNDERLINE, 't', 'h', 'i', 's', SPIECE_UNDERLINE, 'i', 's', SPIECE_UNDERLINE, 'f', 'a', 'l', 's', 'é', '.'] + # fmt: on + ) + + @slow + def test_tokenizer_integration(self): + # Use custom sequence because this tokenizer does not handle numbers. + sequences = [ + "Transformers (formerly known as pytorch-transformers and pytorch-pretrained-bert) provides " + "general-purpose architectures (BERT, GPT, RoBERTa, XLM, DistilBert, XLNet...) for Natural " + "Language Understanding (NLU) and Natural Language Generation (NLG) with over thirty-two pretrained " + "models in one hundred plus languages and deep interoperability between Jax, PyTorch and TensorFlow.", + "BERT is designed to pre-train deep bidirectional representations from unlabeled text by jointly " + "conditioning on both left and right context in all layers.", + "The quick brown fox jumps over the lazy dog.", + ] + + # fmt: off + expected_encoding = {'input_ids': [[4, 32, 13, 7, 9, 12, 19, 8, 13, 18, 5, 13, 12, 4, 64, 19, 8, 13, 18, 5, 13, 15, 22, 4, 28, 9, 8, 20, 9, 4, 7, 12, 4, 24, 22, 6, 8, 13, 17, 11, 39, 6, 13, 7, 9, 12, 19, 8, 13, 18, 5, 13, 12, 4, 7, 9, 14, 4, 24, 22, 6, 8, 13, 17, 11, 39, 24, 13, 5, 6, 13, 7, 10, 9, 5, 14, 39, 25, 5, 13, 6, 63, 4, 24, 13, 8, 27, 10, 14, 5, 12, 4, 21, 5, 9, 5, 13, 7, 15, 39, 24, 16, 13, 24, 8, 12, 5, 4, 7, 13, 17, 11, 10, 6, 5, 17, 6, 16, 13, 5, 12, 4, 64, 40, 47, 54, 32, 23, 4, 53, 49, 32, 23, 4, 54, 8, 40, 47, 54, 32, 7, 23, 4, 69, 52, 43, 23, 4, 51, 10, 12, 6, 10, 15, 40, 5, 13, 6, 23, 4, 69, 52, 48, 5, 6, 26, 26, 26, 63, 4, 19, 8, 13, 4, 48, 7, 6, 16, 13, 7, 15, 4, 52, 7, 9, 21, 16, 7, 21, 5, 4, 61, 9, 14, 5, 13, 12, 6, 7, 9, 14, 10, 9, 21, 4, 64, 48, 52, 61, 63, 4, 7, 9, 14, 4, 48, 7, 6, 16, 13, 7, 15, 4, 52, 7, 9, 21, 16, 7, 21, 5, 4, 53, 5, 9, 5, 13, 7, 6, 10, 8, 9, 4, 64, 48, 52, 53, 63, 4, 20, 10, 6, 11, 4, 8, 27, 5, 13, 4, 6, 11, 10, 13, 6, 22, 39, 6, 20, 8, 4, 24, 13, 5, 6, 13, 7, 10, 9, 5, 14, 4, 18, 8, 14, 5, 15, 12, 4, 10, 9, 4, 8, 9, 5, 4, 11, 16, 9, 14, 13, 5, 14, 4, 24, 15, 16, 12, 4, 15, 7, 9, 21, 16, 7, 21, 5, 12, 4, 7, 9, 14, 4, 14, 5, 5, 24, 4, 10, 9, 6, 5, 13, 8, 24, 5, 13, 7, 25, 10, 15, 10, 6, 22, 4, 25, 5, 6, 20, 5, 5, 9, 4, 58, 7, 37, 23, 4, 49, 22, 32, 8, 13, 17, 11, 4, 7, 9, 14, 4, 32, 5, 9, 12, 8, 13, 55, 15, 8, 20, 26], [4, 40, 47, 54, 32, 4, 10, 12, 4, 14, 5, 12, 10, 21, 9, 5, 14, 4, 6, 8, 4, 24, 13, 5, 39, 6, 13, 7, 10, 9, 4, 14, 5, 5, 24, 4, 25, 10, 14, 10, 13, 5, 17, 6, 10, 8, 9, 7, 15, 4, 13, 5, 24, 13, 5, 12, 5, 9, 6, 7, 6, 10, 8, 9, 12, 4, 19, 13, 8, 18, 4, 16, 9, 15, 7, 25, 5, 15, 5, 14, 4, 6, 5, 37, 6, 4, 25, 22, 4, 46, 8, 10, 9, 6, 15, 22, 4, 17, 8, 9, 14, 10, 6, 10, 8, 9, 10, 9, 21, 4, 8, 9, 4, 25, 8, 6, 11, 4, 15, 5, 19, 6, 4, 7, 9, 14, 4, 13, 10, 21, 11, 6, 4, 17, 8, 9, 6, 5, 37, 6, 4, 10, 9, 4, 7, 15, 15, 4, 15, 7, 22, 5, 13, 12, 26, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [4, 32, 11, 5, 4, 45, 16, 10, 17, 28, 4, 25, 13, 8, 20, 9, 4, 19, 8, 37, 4, 46, 16, 18, 24, 12, 4, 8, 27, 5, 13, 4, 6, 11, 5, 4, 15, 7, 57, 22, 4, 14, 8, 21, 26, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]], 'attention_mask': [[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]]} + # fmt: on + + self.tokenizer_integration_test_util( + expected_encoding=expected_encoding, + model_name="microsoft/speecht5_asr", + revision="c5ef64c71905caeccde0e4462ef3f9077224c524", + sequences=sequences, + ) diff --git a/utils/check_repo.py b/utils/check_repo.py index ad68a43c8a2..37fd498b64e 100755 --- a/utils/check_repo.py +++ b/utils/check_repo.py @@ -133,6 +133,18 @@ IGNORE_NON_TESTED = PRIVATE_MODELS.copy() + [ "BlipTextLMHeadModel", # No need to test it as it is tested by BlipTextVision models "BridgeTowerTextModel", # No need to test it as it is tested by BridgeTowerModel model. "BridgeTowerVisionModel", # No need to test it as it is tested by BridgeTowerModel model. + "SpeechT5Decoder", # Building part of bigger (tested) model. + "SpeechT5DecoderWithoutPrenet", # Building part of bigger (tested) model. + "SpeechT5DecoderWithSpeechPrenet", # Building part of bigger (tested) model. + "SpeechT5DecoderWithTextPrenet", # Building part of bigger (tested) model. + "SpeechT5Encoder", # Building part of bigger (tested) model. + "SpeechT5EncoderWithoutPrenet", # Building part of bigger (tested) model. + "SpeechT5EncoderWithSpeechPrenet", # Building part of bigger (tested) model. + "SpeechT5EncoderWithTextPrenet", # Building part of bigger (tested) model. + "SpeechT5SpeechDecoder", # Building part of bigger (tested) model. + "SpeechT5SpeechEncoder", # Building part of bigger (tested) model. + "SpeechT5TextDecoder", # Building part of bigger (tested) model. + "SpeechT5TextEncoder", # Building part of bigger (tested) model. ] # Update this list with test files that don't have a tester with a `all_model_classes` variable and which don't @@ -269,6 +281,9 @@ IGNORE_NON_AUTO_CONFIGURED = PRIVATE_MODELS.copy() + [ "AltCLIPTextModel", "AltCLIPVisionModel", "AltRobertaModel", + "SpeechT5ForSpeechToSpeech", + "SpeechT5ForTextToSpeech", + "SpeechT5HifiGan", ] # Update this list for models that have multiple model types for the same @@ -378,6 +393,8 @@ def is_a_private_model(model): return True if model.endswith("Decoder"): return True + if model.endswith("Prenet"): + return True return False diff --git a/utils/documentation_tests.txt b/utils/documentation_tests.txt index 2827f7a248b..40349c10aaf 100644 --- a/utils/documentation_tests.txt +++ b/utils/documentation_tests.txt @@ -169,6 +169,8 @@ src/transformers/models/speech_to_text/configuration_speech_to_text.py src/transformers/models/speech_to_text/modeling_speech_to_text.py src/transformers/models/speech_to_text_2/configuration_speech_to_text_2.py src/transformers/models/speech_to_text_2/modeling_speech_to_text_2.py +src/transformers/models/speecht5/modeling_speecht5.py +src/transformers/models/speecht5/tokenization_speecht5.py src/transformers/models/segformer/modeling_tf_segformer.py src/transformers/models/squeezebert/configuration_squeezebert.py src/transformers/models/swin/configuration_swin.py