[WIP] add SpeechT5 model (#18922)

* make SpeechT5 model by copying Wav2Vec2 * add paper to docs * whoops added docs in wrong file * remove SpeechT5Tokenizer + put CTC back in the name * remove deprecated class * remove unused docstring * delete SpeechT5FeatureExtractor, use Wav2Vec2FeatureExtractor instead * remove classes we don't need right now * initial stab at speech encoder prenet * add more speech encoder prenet stuff * improve SpeechEncoderPrenet * add encoder (not finished yet) * add relative position bias to self-attention * add encoder CTC layers * fix formatting * add decoder from BART, doesn't work yet * make it work with generate loop * wrap the encoder into a speech encoder class * wrap the decoder in a text decoder class * changed my mind * changed my mind again ;-) * load decoder weights, make it work * add weights for text decoder postnet * add SpeechT5ForCTC model that uses only the encoder * clean up EncoderLayer and DecoderLayer * implement _init_weights in SpeechT5PreTrainedModel * cleanup config + Encoder and Decoder * add head + cross attention masks * improve doc comments * fixup * more cleanup * more fixup * TextDecoderPrenet works now, thanks Kendall * add CTC loss * add placeholders for other pre/postnets * add type annotation * fix freeze_feature_encoder * set padding tokens to 0 in decoder attention mask * encoder attention mask downsampling * remove features_pen calculation * disable the padding tokens thing again * fixup * more fixup * code review fixes * rename encoder/decoder wrapper classes * allow checkpoints to be loaded into SpeechT5Model * put encoder into wrapper for CTC model * clean up conversion script * add encoder for TTS model * add speech decoder prenet * add speech decoder post-net * attempt to reconstruct the generation loop * add speech generation loop * clean up generate_speech * small tweaks * fix forward pass * enable always dropout on speech decoder prenet * sort declaration * rename models * fixup * fix copies * more fixup * make consistency checker happy * add Seq2SeqSpectrogramOutput class * doc comments * quick note about loss and labels * add HiFi-GAN implementation (from Speech2Speech PR) * rename file * add vocoder to TTS model * improve vocoder * working on tokenizer * more better tokenizer * add CTC tokenizer * fix decode and batch_code in CTC tokenizer * fix processor * two processors and feature extractors * use SpeechT5WaveformFeatureExtractor instead of Wav2Vec2 * cleanup * more cleanup * even more fixup * notebooks * fix log-mel spectrograms * support reduction factor * fixup * shift spectrograms to right to create decoder inputs * return correct labels * add labels for stop token prediction * fix doc comments * fixup * remove SpeechT5ForPreTraining * more fixup * update copyright headers * add usage examples * add SpeechT5ProcessorForCTC * fixup * push unofficial checkpoints to hub * initial version of tokenizer unit tests * add slow test * fix failing tests * tests for CTC tokenizer * finish CTC tokenizer tests * processor tests * initial test for feature extractors * tests for spectrogram feature extractor * fixup * more fixup * add decorators * require speech for tests * modeling tests * more tests for ASR model * fix imports * add fake tests for the other models * fixup * remove jupyter notebooks * add missing SpeechT5Model tests * add missing tests for SpeechT5ForCTC * add missing tests for SpeechT5ForTextToSpeech * sort tests by name * fix Hi-Fi GAN tests * fixup * add speech-to-speech model * refactor duplicate speech generation code * add processor for SpeechToSpeech model * add usage example * add tests for speech-to-speech model * fixup * enable gradient checkpointing for SpeechT5FeatureEncoder * code review * push_to_hub now takes repo_id * improve doc comments for HiFi-GAN config * add missing test * add integration tests * make number of layers in speech decoder prenet configurable * rename variable * rename variables * add auto classes for TTS and S2S * REMOVE CTC!!! * S2S processor does not support save/load_pretrained * fixup * these models are now in an auto mapping * fix doc links * rename HiFiGAN to HifiGan, remove separate config file * REMOVE auto classes * there can be only one * fixup * replace assert * reformat * feature extractor can process input and target at same time * update checkpoint names * fix commit hash
2023-02-03 18:43:46 +01:00
parent fb13a7df95
commit e4bacf6614
39 changed files with 7545 additions and 14 deletions
--- 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.