add VITS model (#24085)

* add VITS model * let's vits * finish TextEncoder (mostly) * rename VITS to Vits * add StochasticDurationPredictor * ads flow model * add generator * correctly set vocab size * add tokenizer * remove processor & feature extractor * add PosteriorEncoder * add missing weights to SDP * also convert LJSpeech and VCTK checkpoints * add training stuff in forward * add placeholder tests for tokenizer * add placeholder tests for model * starting cleanup * let the great renaming begin! * use config * global_conditioning * more cleaning * renaming variables * more renaming * more renaming * it never ends * reticulating the splines * more renaming * HiFi-GAN * doc strings for main model * fixup * fix-copies * don't make it a PreTrainedModel * fixup * rename config options * remove training logic from forward pass * simplify relative position * use actual checkpoint * style * PR review fixes * more review changes * fixup * more unit tests * fixup * fix doc test * add integration test * improve tokenizer tests * add tokenizer integration test * fix tests on GPU (gave OOM) * conversion script can handle repos from hub * add conversion script for all MMS-TTS checkpoints * automatically create a README for the converted checkpoint * small changes to config * push README to hub * only show uroman note for checkpoints that need it * remove conversion script because code formatting breaks the readme * make WaveNet layers configurable * rename variables * simplifying the math * output attentions and hidden states * remove VitsFlip in flow model * also got rid of the other flip * fix tests * rename more variables * rename tokenizer, add phonemization * raise error when phonemizer missing * re-order config docstrings to match method * change config naming * remove redundant str -> list * fix copyright: vits authors -> kakao enterprise * (mean, log_variances) -> (prior_mean, prior_log_variances) * if return dict -> if not return dict * speed -> speaking rate * Apply suggestions from code review Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com> * update fused tanh sigmoid * reduce dims in tester * audio -> output_values * audio -> output_values in tuple out * fix return type * fix return type * make _unconstrained_rational_quadratic_spline a function * all nn's to accept a config * add spectro to output * move {speaking rate, noise scale, noise scale duration} to config * path -> attn_path * idxs -> valid idxs -> padded idxs * output values -> waveform * use config for attention * make generation work * harden integration test * add spectrogram to dict output * tokenizer refactor * make style * remove 'fake' padding token * harden tokenizer tests * ron norm test * fprop / save tests deterministic * move uroman to tokenizer as much as possible * better logger message * fix vivit imports * add uroman integration test * make style * up * matthijs -> sanchit-gandhi * fix tokenizer test * make fix-copies * fix dict comprehension * fix config tests * fix model tests * make outputs consistent with reverse/not reverse * fix key concat * more model details * add author * return dict * speaker error * labels error * Apply suggestions from code review Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com> * Update src/transformers/models/vits/convert_original_checkpoint.py Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com> * remove uromanize * add docstrings * add docstrings for tokenizer * upper-case skip messages * fix return dict * style * finish tests * update checkpoints * make style * remove doctest file * revert * fix docstring * fix tokenizer * remove uroman integration test * add sampling rate * fix docs / docstrings * style * add sr to model output * fix outputs * style / copies * fix docstring * fix copies * remove sr from model outputs * Update utils/documentation_tests.txt Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com> * add sr as allowed attr --------- Co-authored-by: sanchit-gandhi <sanchit@huggingface.co> Co-authored-by: Sanchit Gandhi <93869735+sanchit-gandhi@users.noreply.github.com> Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>
2023-09-01 11:50:06 +02:00
parent ef10dbce5c
commit 4ece3b9433
26 changed files with 3694 additions and 0 deletions
--- a/README_zh-hans.md
+++ b/README_zh-hans.md
@@ -439,6 +439,7 @@ conda install -c huggingface transformers
 1. **[VitDet](https://huggingface.co/docs/transformers/main/model_doc/vitdet)** (来自 Meta AI) 伴随论文 [Exploring Plain Vision Transformer Backbones for Object Detection](https://arxiv.org/abs/2203.16527) 由 Yanghao Li, Hanzi Mao, Ross Girshick, Kaiming He 发布。
 1. **[ViTMAE](https://huggingface.co/docs/transformers/model_doc/vit_mae)** (来自 Meta AI) 伴随论文 [Masked Autoencoders Are Scalable Vision Learners](https://arxiv.org/abs/2111.06377) 由 Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, Ross Girshick 发布。
 1. **[ViTMSN](https://huggingface.co/docs/transformers/model_doc/vit_msn)** (来自 Meta AI) 伴随论文 [Masked Siamese Networks for Label-Efficient Learning](https://arxiv.org/abs/2204.07141) by Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Michael Rabbat, Nicolas Ballas 发布.
+1. **[VITS](https://huggingface.co/docs/transformers/main/model_doc/vits)** (来自 Kakao Enterprise) 伴随论文 [Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech](https://arxiv.org/abs/2106.06103) 由 Jaehyeon Kim, Jungil Kong, Juhee Son 发布。
 1. **[ViViT](https://huggingface.co/docs/transformers/model_doc/vivit)** (来自 Google Research) released with the paper [ViViT: A Video Vision Transformer](https://arxiv.org/abs/2103.15691) 由 Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, Cordelia Schmid.
 1. **[Wav2Vec2](https://huggingface.co/docs/transformers/model_doc/wav2vec2)** (来自 Facebook AI) 伴随论文 [wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations](https://arxiv.org/abs/2006.11477) 由 Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, Michael Auli 发布。
 1. **[Wav2Vec2-Conformer](https://huggingface.co/docs/transformers/model_doc/wav2vec2-conformer)** (来自 Facebook AI) 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, Sravya Popuri, Dmytro Okhonko, Juan Pino 发布。