mirror of
https://github.com/huggingface/transformers.git
synced 2025-08-02 03:01:07 +06:00
Add Unispeech & Unispeech-SAT (#13963)
* unispeech * add copy from * remove hubert copy from * finish for today * add unispeech-sat * adapt more * up * up * up * up * add modeling * add tests * up * up * finish * up * Apply suggestions from code review * up * up * Apply suggestions from code review Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com> * up * up Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
This commit is contained in:
Patrick von Platen
GitHub
parent
9799f4e150
commit
9f3aa46f45
@ -282,6 +282,9 @@ Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih.
|
||||
1. **[TAPAS](https://huggingface.co/transformers/model_doc/tapas.html)** (from Google AI) released with the paper [TAPAS: Weakly Supervised Table Parsing via Pre-training](https://arxiv.org/abs/2004.02349) by Jonathan Herzig, Paweł Krzysztof Nowak, Thomas Müller, Francesco Piccinno and Julian Martin Eisenschlos.
|
||||
1. **[Transformer-XL](https://huggingface.co/transformers/model_doc/transformerxl.html)** (from Google/CMU) released with the paper [Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context](https://arxiv.org/abs/1901.02860) by Zihang Dai*, Zhilin Yang*, Yiming Yang, Jaime Carbonell, Quoc V. Le, Ruslan Salakhutdinov.
|
||||
1. **[TrOCR](https://huggingface.co/transformers/master/model_doc/trocr.html)** (from Microsoft), released together 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. **[UniSpeech](https://huggingface.co/transformers/master/model_doc/unispeech.html)** (from Microsoft Research) released with the paper [UniSpeech: Unified Speech Representation Learning with Labeled and Unlabeled Data](https://arxiv.org/abs/2101.07597) by Chengyi Wang, Yu Wu, Yao Qian, Kenichi Kumatani, Shujie Liu, Furu Wei, Michael Zeng, Xuedong Huang.
|
||||
1. **[UniSpeechSat](https://huggingface.co/transformers/master/model_doc/unispeech_sat.html)** (from Microsoft Research) released with the paper [UNISPEECH-SAT: UNIVERSAL SPEECH REPRESENTATION LEARNING WITH SPEAKER
|
||||
AWARE PRE-TRAINING](https://arxiv.org/abs/2110.05752) by Sanyuan Chen, Yu Wu, Chengyi Wang, Zhengyang Chen, Zhuo Chen, Shujie Liu, Jian Wu, Yao Qian, Furu Wei, Jinyu Li, Xiangzhan Yu.
|
||||
1. **[Vision Transformer (ViT)](https://huggingface.co/transformers/model_doc/vit.html)** (from Google AI) released with the paper [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) by Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby.
|
||||
1. **[VisualBERT](https://huggingface.co/transformers/model_doc/visual_bert.html)** (from UCLA NLP) released with the paper [VisualBERT: A Simple and Performant Baseline for Vision and Language](https://arxiv.org/pdf/1908.03557) by Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, Kai-Wei Chang.
|
||||
1. **[Wav2Vec2](https://huggingface.co/transformers/model_doc/wav2vec2.html)** (from Facebook AI) released with the paper [wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations](https://arxiv.org/abs/2006.11477) by Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, Michael Auli.
|
||||
|
||||
@ -280,6 +280,8 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
|
||||
1. **[TAPAS](https://huggingface.co/transformers/model_doc/tapas.html)** (from Google AI) released with the paper [TAPAS: Weakly Supervised Table Parsing via Pre-training](https://arxiv.org/abs/2004.02349) by Jonathan Herzig, Paweł Krzysztof Nowak, Thomas Müller, Francesco Piccinno and Julian Martin Eisenschlos.
|
||||
1. **[Transformer-XL](https://huggingface.co/transformers/model_doc/transformerxl.html)** (from Google/CMU) released with the paper [Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context](https://arxiv.org/abs/1901.02860) by Zihang Dai*, Zhilin Yang*, Yiming Yang, Jaime Carbonell, Quoc V. Le, Ruslan Salakhutdinov.
|
||||
1. **[TrOCR](https://huggingface.co/transformers/master/model_doc/trocr.html)** (from Microsoft), released together 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. **[UniSpeech](https://huggingface.co/transformers/master/model_doc/unispeech.html)** (from Microsoft Research) released with the paper [UniSpeech: Unified Speech Representation Learning with Labeled and Unlabeled Data](https://arxiv.org/abs/2101.07597) by Chengyi Wang, Yu Wu, Yao Qian, Kenichi Kumatani, Shujie Liu, Furu Wei, Michael Zeng, Xuedong Huang.
|
||||
1. **[UniSpeechSat](https://huggingface.co/transformers/master/model_doc/unispeech_sat.html)** (from Microsoft Research) released with the paper [UNISPEECH-SAT: UNIVERSAL SPEECH REPRESENTATION LEARNING WITH SPEAKER AWARE PRE-TRAINING](https://arxiv.org/abs/2110.05752) by Sanyuan Chen, Yu Wu, Chengyi Wang, Zhengyang Chen, Zhuo Chen, Shujie Liu, Jian Wu, Yao Qian, Furu Wei, Jinyu Li, Xiangzhan Yu.
|
||||
1. **[Vision Transformer (ViT)](https://huggingface.co/transformers/model_doc/vit.html)** (from Google AI) released with the paper [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) by Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby.
|
||||
1. **[VisualBERT](https://huggingface.co/transformers/model_doc/visual_bert.html)** (from UCLA NLP) released with the paper [VisualBERT: A Simple and Performant Baseline for Vision and Language](https://arxiv.org/pdf/1908.03557) by Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, Kai-Wei Chang.
|
||||
1. **[Wav2Vec2](https://huggingface.co/transformers/model_doc/wav2vec2.html)** (from Facebook AI) released with the paper [wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations](https://arxiv.org/abs/2006.11477) by Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, Michael Auli.
|
||||
|
||||
@ -305,6 +305,8 @@ conda install -c huggingface transformers
|
||||
1. **[TAPAS](https://huggingface.co/transformers/model_doc/tapas.html)** (来自 Google AI) 伴随论文 [TAPAS: Weakly Supervised Table Parsing via Pre-training](https://arxiv.org/abs/2004.02349) 由 Jonathan Herzig, Paweł Krzysztof Nowak, Thomas Müller, Francesco Piccinno and Julian Martin Eisenschlos 发布。
|
||||
1. **[Transformer-XL](https://huggingface.co/transformers/model_doc/transformerxl.html)** (来自 Google/CMU) 伴随论文 [Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context](https://arxiv.org/abs/1901.02860) 由 Zihang Dai*, Zhilin Yang*, Yiming Yang, Jaime Carbonell, Quoc V. Le, Ruslan Salakhutdinov 发布。
|
||||
1. **[TrOCR](https://huggingface.co/transformers/master/model_doc/trocr.html)** (来自 Microsoft) 伴随论文 [TrOCR: Transformer-based Optical Character Recognition with Pre-trained Models](https://arxiv.org/abs/2109.10282) 由 Minghao Li, Tengchao Lv, Lei Cui, Yijuan Lu, Dinei Florencio, Cha Zhang, Zhoujun Li, Furu Wei 发布。
|
||||
1. **[UniSpeech](https://huggingface.co/transformers/master/model_doc/unispeech.html)** (来自 Microsoft Research) 伴随论文 [UniSpeech: Unified Speech Representation Learning with Labeled and Unlabeled Data](https://arxiv.org/abs/2101.07597) 由 Chengyi Wang, Yu Wu, Yao Qian, Kenichi Kumatani, Shujie Liu, Furu Wei, Michael Zeng, Xuedong Huang 发布。
|
||||
1. **[UniSpeechSat](https://huggingface.co/transformers/master/model_doc/unispeech_sat.html)** (来自 Microsoft Research) 伴随论文 [UNISPEECH-SAT: UNIVERSAL SPEECH REPRESENTATION LEARNING WITH SPEAKER AWARE PRE-TRAINING](https://arxiv.org/abs/2110.05752) 由 Sanyuan Chen, Yu Wu, Chengyi Wang, Zhengyang Chen, Zhuo Chen, Shujie Liu, Jian Wu, Yao Qian, Furu Wei, Jinyu Li, Xiangzhan Yu 发布。
|
||||
1. **[Vision Transformer (ViT)](https://huggingface.co/transformers/model_doc/vit.html)** (来自 Google AI) 伴随论文 [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) 由 Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby 发布。
|
||||
1. **[VisionEncoderDecoder](https://huggingface.co/transformers/model_doc/visionencoderdecoder.html)**
|
||||
1. **[VisualBERT](https://huggingface.co/transformers/model_doc/visual_bert.html)** (来自 UCLA NLP) 伴随论文 [VisualBERT: A Simple and Performant Baseline for Vision and Language](https://arxiv.org/pdf/1908.03557) 由 Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, Kai-Wei Chang 发布。
|
||||
|
||||
@ -317,6 +317,8 @@ conda install -c huggingface transformers
|
||||
1. **[TAPAS](https://huggingface.co/transformers/model_doc/tapas.html)** (from Google AI) released with the paper [TAPAS: Weakly Supervised Table Parsing via Pre-training](https://arxiv.org/abs/2004.02349) by Jonathan Herzig, Paweł Krzysztof Nowak, Thomas Müller, Francesco Piccinno and Julian Martin Eisenschlos.
|
||||
1. **[Transformer-XL](https://huggingface.co/transformers/model_doc/transformerxl.html)** (from Google/CMU) released with the paper [Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context](https://arxiv.org/abs/1901.02860) by Zihang Dai*, Zhilin Yang*, Yiming Yang, Jaime Carbonell, Quoc V. Le, Ruslan Salakhutdinov.
|
||||
1. **[TrOCR](https://huggingface.co/transformers/master/model_doc/trocr.html)** (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. **[UniSpeech](https://huggingface.co/transformers/master/model_doc/unispeech.html)** (from Microsoft Research) released with the paper [UniSpeech: Unified Speech Representation Learning with Labeled and Unlabeled Data](https://arxiv.org/abs/2101.07597) by Chengyi Wang, Yu Wu, Yao Qian, Kenichi Kumatani, Shujie Liu, Furu Wei, Michael Zeng, Xuedong Huang.
|
||||
1. **[UniSpeechSat](https://huggingface.co/transformers/master/model_doc/unispeech_sat.html)** (from Microsoft Research) released with the paper [UNISPEECH-SAT: UNIVERSAL SPEECH REPRESENTATION LEARNING WITH SPEAKER AWARE PRE-TRAINING](https://arxiv.org/abs/2110.05752) by Sanyuan Chen, Yu Wu, Chengyi Wang, Zhengyang Chen, Zhuo Chen, Shujie Liu, Jian Wu, Yao Qian, Furu Wei, Jinyu Li, Xiangzhan Yu.
|
||||
1. **[Vision Transformer (ViT)](https://huggingface.co/transformers/model_doc/vit.html)** (from Google AI) released with the paper [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) by Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby.
|
||||
1. **[VisionEncoderDecoder](https://huggingface.co/transformers/model_doc/visionencoderdecoder.html)**
|
||||
1. **[VisualBERT](https://huggingface.co/transformers/model_doc/visual_bert.html)** (from UCLA NLP) released with the paper [VisualBERT: A Simple and Performant Baseline for Vision and Language](https://arxiv.org/pdf/1908.03557) by Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, Kai-Wei Chang.
|
||||
|
||||
@ -314,29 +314,37 @@ Supported models
|
||||
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.
|
||||
72. :doc:`Vision Transformer (ViT) <model_doc/vit>` (from Google AI) released with the paper `An Image is Worth 16x16
|
||||
72. `UniSpeech <https://huggingface.co/transformers/master/model_doc/unispeech.html>`__ (from Microsoft Research)
|
||||
released with the paper `UniSpeech: Unified Speech Representation Learning with Labeled and Unlabeled Data
|
||||
<https://arxiv.org/abs/2101.07597>`__ by Chengyi Wang, Yu Wu, Yao Qian, Kenichi Kumatani, Shujie Liu, Furu Wei,
|
||||
Michael Zeng, Xuedong Huang.
|
||||
73. `UniSpeechSat <https://huggingface.co/transformers/master/model_doc/unispeech_sat.html>`__ (from Microsoft
|
||||
Research) released with the paper `UNISPEECH-SAT: UNIVERSAL SPEECH REPRESENTATION LEARNING WITH SPEAKER AWARE
|
||||
PRE-TRAINING <https://arxiv.org/abs/2110.05752>`__ by Sanyuan Chen, Yu Wu, Chengyi Wang, Zhengyang Chen, Zhuo Chen,
|
||||
Shujie Liu, Jian Wu, Yao Qian, Furu Wei, Jinyu Li, Xiangzhan Yu.
|
||||
74. :doc:`Vision Transformer (ViT) <model_doc/vit>` (from Google AI) released with the paper `An Image is Worth 16x16
|
||||
Words: Transformers for Image Recognition at Scale <https://arxiv.org/abs/2010.11929>`__ by Alexey Dosovitskiy,
|
||||
Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias
|
||||
Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby.
|
||||
73. :doc:`VisualBERT <model_doc/visual_bert>` (from UCLA NLP) released with the paper `VisualBERT: A Simple and
|
||||
75. :doc:`VisualBERT <model_doc/visual_bert>` (from UCLA NLP) released with the paper `VisualBERT: A Simple and
|
||||
Performant Baseline for Vision and Language <https://arxiv.org/pdf/1908.03557>`__ by Liunian Harold Li, Mark
|
||||
Yatskar, Da Yin, Cho-Jui Hsieh, Kai-Wei Chang.
|
||||
74. :doc:`Wav2Vec2 <model_doc/wav2vec2>` (from Facebook AI) released with the paper `wav2vec 2.0: A Framework for
|
||||
76. :doc:`Wav2Vec2 <model_doc/wav2vec2>` (from Facebook AI) released with the paper `wav2vec 2.0: A Framework for
|
||||
Self-Supervised Learning of Speech Representations <https://arxiv.org/abs/2006.11477>`__ by Alexei Baevski, Henry
|
||||
Zhou, Abdelrahman Mohamed, Michael Auli.
|
||||
75. :doc:`XLM <model_doc/xlm>` (from Facebook) released together with the paper `Cross-lingual Language Model
|
||||
77. :doc:`XLM <model_doc/xlm>` (from Facebook) released together with the paper `Cross-lingual Language Model
|
||||
Pretraining <https://arxiv.org/abs/1901.07291>`__ by Guillaume Lample and Alexis Conneau.
|
||||
76. :doc:`XLM-ProphetNet <model_doc/xlmprophetnet>` (from Microsoft Research) released with the paper `ProphetNet:
|
||||
78. :doc:`XLM-ProphetNet <model_doc/xlmprophetnet>` (from Microsoft Research) released with the paper `ProphetNet:
|
||||
Predicting Future N-gram for Sequence-to-Sequence Pre-training <https://arxiv.org/abs/2001.04063>`__ by Yu Yan,
|
||||
Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou.
|
||||
77. :doc:`XLM-RoBERTa <model_doc/xlmroberta>` (from Facebook AI), released together with the paper `Unsupervised
|
||||
79. :doc:`XLM-RoBERTa <model_doc/xlmroberta>` (from Facebook AI), released together with the paper `Unsupervised
|
||||
Cross-lingual Representation Learning at Scale <https://arxiv.org/abs/1911.02116>`__ by Alexis Conneau*, Kartikay
|
||||
Khandelwal*, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke
|
||||
Zettlemoyer and Veselin Stoyanov.
|
||||
78. :doc:`XLNet <model_doc/xlnet>` (from Google/CMU) released with the paper `XLNet: Generalized Autoregressive
|
||||
80. :doc:`XLNet <model_doc/xlnet>` (from Google/CMU) released with the paper `XLNet: Generalized Autoregressive
|
||||
Pretraining for Language Understanding <https://arxiv.org/abs/1906.08237>`__ by Zhilin Yang*, Zihang Dai*, Yiming
|
||||
Yang, Jaime Carbonell, Ruslan Salakhutdinov, Quoc V. Le.
|
||||
79. :doc:`XLSR-Wav2Vec2 <model_doc/xlsr_wav2vec2>` (from Facebook AI) released with the paper `Unsupervised
|
||||
81. :doc:`XLSR-Wav2Vec2 <model_doc/xlsr_wav2vec2>` (from Facebook AI) released with the paper `Unsupervised
|
||||
Cross-Lingual Representation Learning For Speech Recognition <https://arxiv.org/abs/2006.13979>`__ by Alexis
|
||||
Conneau, Alexei Baevski, Ronan Collobert, Abdelrahman Mohamed, Michael Auli.
|
||||
|
||||
@ -484,6 +492,10 @@ Flax), PyTorch, and/or TensorFlow.
|
||||
+-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
|
||||
| TrOCR | ❌ | ❌ | ✅ | ❌ | ❌ |
|
||||
+-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
|
||||
| UniSpeech | ❌ | ❌ | ✅ | ❌ | ❌ |
|
||||
+-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
|
||||
| UniSpeechSat | ❌ | ❌ | ✅ | ❌ | ❌ |
|
||||
+-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
|
||||
| Vision Encoder decoder | ❌ | ❌ | ✅ | ❌ | ❌ |
|
||||
+-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
|
||||
| VisualBert | ❌ | ❌ | ✅ | ❌ | ❌ |
|
||||
@ -654,6 +666,8 @@ Flax), PyTorch, and/or TensorFlow.
|
||||
model_doc/tapas
|
||||
model_doc/transformerxl
|
||||
model_doc/trocr
|
||||
model_doc/unispeech
|
||||
model_doc/unispeech_sat
|
||||
model_doc/visionencoderdecoder
|
||||
model_doc/vit
|
||||
model_doc/visual_bert
|
||||
|
||||
88
docs/source/model_doc/unispeech.rst
Normal file
88
docs/source/model_doc/unispeech.rst
Normal file
@ -0,0 +1,88 @@
|
||||
..
|
||||
Copyright 2021 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.
|
||||
|
||||
UniSpeech
|
||||
-----------------------------------------------------------------------------------------------------------------------
|
||||
|
||||
Overview
|
||||
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
The UniSpeech model was proposed in `UniSpeech: Unified Speech Representation Learning with Labeled and Unlabeled Data
|
||||
<https://arxiv.org/abs/2101.07597>`__ by Chengyi Wang, Yu Wu, Yao Qian, Kenichi Kumatani, Shujie Liu, Furu Wei, Michael
|
||||
Zeng, Xuedong Huang .
|
||||
|
||||
The abstract from the paper is the following:
|
||||
|
||||
*In this paper, we propose a unified pre-training approach called UniSpeech to learn speech representations with both
|
||||
unlabeled and labeled data, in which supervised phonetic CTC learning and phonetically-aware contrastive
|
||||
self-supervised learning are conducted in a multi-task learning manner. The resultant representations can capture
|
||||
information more correlated with phonetic structures and improve the generalization across languages and domains. We
|
||||
evaluate the effectiveness of UniSpeech for cross-lingual representation learning on public CommonVoice corpus. The
|
||||
results show that UniSpeech outperforms self-supervised pretraining and supervised transfer learning for speech
|
||||
recognition by a maximum of 13.4% and 17.8% relative phone error rate reductions respectively (averaged over all
|
||||
testing languages). The transferability of UniSpeech is also demonstrated on a domain-shift speech recognition task,
|
||||
i.e., a relative word error rate reduction of 6% against the previous approach.*
|
||||
|
||||
Tips:
|
||||
|
||||
- UniSpeech is a speech model that accepts a float array corresponding to the raw waveform of the speech signal. Please
|
||||
use :class:`~transformers.Wav2Vec2Processor` for the feature extraction.
|
||||
- UniSpeech model can be fine-tuned using connectionist temporal classification (CTC) so the model output has to be
|
||||
decoded using :class:`~transformers.Wav2Vec2CTCTokenizer`.
|
||||
|
||||
This model was contributed by `patrickvonplaten <https://huggingface.co/patrickvonplaten>`__. The Authors' code can be
|
||||
found `here <https://github.com/microsoft/UniSpeech/tree/main/UniSpeech>`__.
|
||||
|
||||
|
||||
UniSpeechConfig
|
||||
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
.. autoclass:: transformers.UniSpeechConfig
|
||||
:members:
|
||||
|
||||
|
||||
UniSpeech specific outputs
|
||||
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
.. autoclass:: transformers.models.unispeech.modeling_unispeech.UniSpeechBaseModelOutput
|
||||
:members:
|
||||
|
||||
.. autoclass:: transformers.models.unispeech.modeling_unispeech.UniSpeechForPreTrainingOutput
|
||||
:members:
|
||||
|
||||
|
||||
UniSpeechModel
|
||||
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
.. autoclass:: transformers.UniSpeechModel
|
||||
:members: forward
|
||||
|
||||
|
||||
UniSpeechForCTC
|
||||
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
.. autoclass:: transformers.UniSpeechForCTC
|
||||
:members: forward
|
||||
|
||||
|
||||
UniSpeechForSequenceClassification
|
||||
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
.. autoclass:: transformers.UniSpeechForSequenceClassification
|
||||
:members: forward
|
||||
|
||||
|
||||
UniSpeechForPreTraining
|
||||
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
.. autoclass:: transformers.UniSpeechForPreTraining
|
||||
:members: forward
|
||||
92
docs/source/model_doc/unispeech_sat.rst
Normal file
92
docs/source/model_doc/unispeech_sat.rst
Normal file
@ -0,0 +1,92 @@
|
||||
..
|
||||
Copyright 2021 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.
|
||||
|
||||
UniSpeech-SAT
|
||||
-----------------------------------------------------------------------------------------------------------------------
|
||||
|
||||
Overview
|
||||
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
The UniSpeech-SAT model was proposed in `UniSpeech-SAT: Universal Speech Representation Learning with Speaker Aware
|
||||
Pre-Training <https://arxiv.org/abs/2110.05752>`__ by Sanyuan Chen, Yu Wu, Chengyi Wang, Zhengyang Chen, Zhuo Chen,
|
||||
Shujie Liu, Jian Wu, Yao Qian, Furu Wei, Jinyu Li, Xiangzhan Yu .
|
||||
|
||||
The abstract from the paper is the following:
|
||||
|
||||
*Self-supervised learning (SSL) is a long-standing goal for speech processing, since it utilizes large-scale unlabeled
|
||||
data and avoids extensive human labeling. Recent years witness great successes in applying self-supervised learning in
|
||||
speech recognition, while limited exploration was attempted in applying SSL for modeling speaker characteristics. In
|
||||
this paper, we aim to improve the existing SSL framework for speaker representation learning. Two methods are
|
||||
introduced for enhancing the unsupervised speaker information extraction. First, we apply the multi-task learning to
|
||||
the current SSL framework, where we integrate the utterance-wise contrastive loss with the SSL objective function.
|
||||
Second, for better speaker discrimination, we propose an utterance mixing strategy for data augmentation, where
|
||||
additional overlapped utterances are created unsupervisely and incorporate during training. We integrate the proposed
|
||||
methods into the HuBERT framework. Experiment results on SUPERB benchmark show that the proposed system achieves
|
||||
state-of-the-art performance in universal representation learning, especially for speaker identification oriented
|
||||
tasks. An ablation study is performed verifying the efficacy of each proposed method. Finally, we scale up training
|
||||
dataset to 94 thousand hours public audio data and achieve further performance improvement in all SUPERB tasks.*
|
||||
|
||||
Tips:
|
||||
|
||||
- UniSpeechSat is a speech model that accepts a float array corresponding to the raw waveform of the speech signal.
|
||||
Please use :class:`~transformers.Wav2Vec2Processor` for the feature extraction.
|
||||
- UniSpeechSat model can be fine-tuned using connectionist temporal classification (CTC) so the model output has to be
|
||||
decoded using :class:`~transformers.Wav2Vec2CTCTokenizer`.
|
||||
- UniSpeechSat performs especially well on speaker verification, speaker identification, and speaker diarization tasks.
|
||||
|
||||
This model was contributed by `patrickvonplaten <https://huggingface.co/patrickvonplaten>`__. The Authors' code can be
|
||||
found `here <https://github.com/microsoft/UniSpeech/tree/main/UniSpeech-SAT>`__.
|
||||
|
||||
|
||||
UniSpeechSatConfig
|
||||
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
.. autoclass:: transformers.UniSpeechSatConfig
|
||||
:members:
|
||||
|
||||
|
||||
UniSpeechSat specific outputs
|
||||
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
.. autoclass:: transformers.models.unispeech_sat.modeling_unispeech_sat.UniSpeechSatBaseModelOutput
|
||||
:members:
|
||||
|
||||
.. autoclass:: transformers.models.unispeech_sat.modeling_unispeech_sat.UniSpeechSatForPreTrainingOutput
|
||||
:members:
|
||||
|
||||
|
||||
UniSpeechSatModel
|
||||
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
.. autoclass:: transformers.UniSpeechSatModel
|
||||
:members: forward
|
||||
|
||||
|
||||
UniSpeechSatForCTC
|
||||
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
.. autoclass:: transformers.UniSpeechSatForCTC
|
||||
:members: forward
|
||||
|
||||
|
||||
UniSpeechSatForSequenceClassification
|
||||
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
.. autoclass:: transformers.UniSpeechSatForSequenceClassification
|
||||
:members: forward
|
||||
|
||||
|
||||
UniSpeechSatForPreTraining
|
||||
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
.. autoclass:: transformers.UniSpeechSatForPreTraining
|
||||
:members: forward
|
||||
@ -280,6 +280,14 @@ _import_structure = {
|
||||
"TrOCRConfig",
|
||||
"TrOCRProcessor",
|
||||
],
|
||||
"models.unispeech": [
|
||||
"UNISPEECH_PRETRAINED_CONFIG_ARCHIVE_MAP",
|
||||
"UniSpeechConfig",
|
||||
],
|
||||
"models.unispeech_sat": [
|
||||
"UNISPEECH_SAT_PRETRAINED_CONFIG_ARCHIVE_MAP",
|
||||
"UniSpeechSatConfig",
|
||||
],
|
||||
"models.vision_encoder_decoder": ["VisionEncoderDecoderConfig"],
|
||||
"models.visual_bert": ["VISUAL_BERT_PRETRAINED_CONFIG_ARCHIVE_MAP", "VisualBertConfig"],
|
||||
"models.vit": ["VIT_PRETRAINED_CONFIG_ARCHIVE_MAP", "ViTConfig"],
|
||||
@ -1213,6 +1221,26 @@ if is_torch_available():
|
||||
_import_structure["models.trocr"].extend(
|
||||
["TROCR_PRETRAINED_MODEL_ARCHIVE_LIST", "TrOCRForCausalLM", "TrOCRPreTrainedModel"]
|
||||
)
|
||||
_import_structure["models.unispeech"].extend(
|
||||
[
|
||||
"UNISPEECH_PRETRAINED_MODEL_ARCHIVE_LIST",
|
||||
"UniSpeechForCTC",
|
||||
"UniSpeechForPreTraining",
|
||||
"UniSpeechForSequenceClassification",
|
||||
"UniSpeechModel",
|
||||
"UniSpeechPreTrainedModel",
|
||||
]
|
||||
)
|
||||
_import_structure["models.unispeech_sat"].extend(
|
||||
[
|
||||
"UNISPEECH_SAT_PRETRAINED_MODEL_ARCHIVE_LIST",
|
||||
"UniSpeechSatForCTC",
|
||||
"UniSpeechSatForPreTraining",
|
||||
"UniSpeechSatForSequenceClassification",
|
||||
"UniSpeechSatModel",
|
||||
"UniSpeechSatPreTrainedModel",
|
||||
]
|
||||
)
|
||||
_import_structure["models.vision_encoder_decoder"].extend(["VisionEncoderDecoderModel"])
|
||||
_import_structure["models.visual_bert"].extend(
|
||||
[
|
||||
@ -2138,6 +2166,8 @@ if TYPE_CHECKING:
|
||||
TransfoXLTokenizer,
|
||||
)
|
||||
from .models.trocr import TROCR_PRETRAINED_CONFIG_ARCHIVE_MAP, TrOCRConfig, TrOCRProcessor
|
||||
from .models.unispeech import UNISPEECH_PRETRAINED_CONFIG_ARCHIVE_MAP, UniSpeechConfig
|
||||
from .models.unispeech_sat import UNISPEECH_SAT_PRETRAINED_CONFIG_ARCHIVE_MAP, UniSpeechSatConfig
|
||||
from .models.vision_encoder_decoder import VisionEncoderDecoderConfig
|
||||
from .models.visual_bert import VISUAL_BERT_PRETRAINED_CONFIG_ARCHIVE_MAP, VisualBertConfig
|
||||
from .models.vit import VIT_PRETRAINED_CONFIG_ARCHIVE_MAP, ViTConfig
|
||||
@ -2918,6 +2948,22 @@ if TYPE_CHECKING:
|
||||
load_tf_weights_in_transfo_xl,
|
||||
)
|
||||
from .models.trocr import TROCR_PRETRAINED_MODEL_ARCHIVE_LIST, TrOCRForCausalLM, TrOCRPreTrainedModel
|
||||
from .models.unispeech import (
|
||||
UNISPEECH_PRETRAINED_MODEL_ARCHIVE_LIST,
|
||||
UniSpeechForCTC,
|
||||
UniSpeechForPreTraining,
|
||||
UniSpeechForSequenceClassification,
|
||||
UniSpeechModel,
|
||||
UniSpeechPreTrainedModel,
|
||||
)
|
||||
from .models.unispeech_sat import (
|
||||
UNISPEECH_SAT_PRETRAINED_MODEL_ARCHIVE_LIST,
|
||||
UniSpeechSatForCTC,
|
||||
UniSpeechSatForPreTraining,
|
||||
UniSpeechSatForSequenceClassification,
|
||||
UniSpeechSatModel,
|
||||
UniSpeechSatPreTrainedModel,
|
||||
)
|
||||
from .models.vision_encoder_decoder import VisionEncoderDecoderModel
|
||||
from .models.visual_bert import (
|
||||
VISUAL_BERT_PRETRAINED_MODEL_ARCHIVE_LIST,
|
||||
|
||||
@ -98,6 +98,8 @@ CONFIG_MAPPING_NAMES = OrderedDict(
|
||||
("splinter", "SplinterConfig"),
|
||||
("sew-d", "SEWDConfig"),
|
||||
("sew", "SEWConfig"),
|
||||
("unispeech-sat", "UniSpeechSatConfig"),
|
||||
("unispeech", "UniSpeechConfig"),
|
||||
]
|
||||
)
|
||||
|
||||
@ -166,6 +168,8 @@ CONFIG_ARCHIVE_MAP_MAPPING_NAMES = OrderedDict(
|
||||
("splinter", "SPLINTER_PRETRAINED_CONFIG_ARCHIVE_MAP"),
|
||||
("sew-d", "SEW_D_PRETRAINED_CONFIG_ARCHIVE_MAP"),
|
||||
("sew", "SEW_PRETRAINED_CONFIG_ARCHIVE_MAP"),
|
||||
("unispeech-sat", "UNISPEECH_SAT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
|
||||
("unispeech", "UNISPEECH_PRETRAINED_CONFIG_ARCHIVE_MAP"),
|
||||
]
|
||||
)
|
||||
|
||||
@ -252,6 +256,8 @@ MODEL_NAMES_MAPPING = OrderedDict(
|
||||
("splinter", "Splinter"),
|
||||
("sew-d", "SEW-D"),
|
||||
("sew", "SEW"),
|
||||
("unispeech-sat", "UniSpeechSat"),
|
||||
("unispeech", "UniSpeech"),
|
||||
]
|
||||
)
|
||||
|
||||
|
||||
@ -46,6 +46,8 @@ MODEL_MAPPING_NAMES = OrderedDict(
|
||||
("speech_to_text", "Speech2TextModel"),
|
||||
("vit", "ViTModel"),
|
||||
("wav2vec2", "Wav2Vec2Model"),
|
||||
("unispeech-sat", "UniSpeechSatModel"),
|
||||
("unispeech", "UniSpeechModel"),
|
||||
("hubert", "HubertModel"),
|
||||
("m2m_100", "M2M100Model"),
|
||||
("convbert", "ConvBertModel"),
|
||||
@ -134,6 +136,8 @@ MODEL_FOR_PRETRAINING_MAPPING_NAMES = OrderedDict(
|
||||
("deberta", "DebertaForMaskedLM"),
|
||||
("deberta-v2", "DebertaV2ForMaskedLM"),
|
||||
("wav2vec2", "Wav2Vec2ForPreTraining"),
|
||||
("unispeech-sat", "UniSpeechSatForPreTraining"),
|
||||
("unispeech", "UniSpeechForPreTraining"),
|
||||
]
|
||||
)
|
||||
|
||||
@ -475,6 +479,8 @@ MODEL_FOR_AUDIO_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
|
||||
[
|
||||
# Model for Audio Classification mapping
|
||||
("wav2vec2", "Wav2Vec2ForSequenceClassification"),
|
||||
("unispeech-sat", "UniSpeechSatForSequenceClassification"),
|
||||
("unispeech", "UniSpeechForSequenceClassification"),
|
||||
("hubert", "HubertForSequenceClassification"),
|
||||
("sew", "SEWForSequenceClassification"),
|
||||
("sew-d", "SEWDForSequenceClassification"),
|
||||
@ -485,6 +491,8 @@ MODEL_FOR_CTC_MAPPING_NAMES = OrderedDict(
|
||||
[
|
||||
# Model for Connectionist temporal classification (CTC) mapping
|
||||
("wav2vec2", "Wav2Vec2ForCTC"),
|
||||
("unispeech-sat", "UniSpeechSatForCTC"),
|
||||
("unispeech", "UniSpeechForCTC"),
|
||||
("hubert", "HubertForCTC"),
|
||||
("sew", "SEWForCTC"),
|
||||
("sew-d", "SEWDForCTC"),
|
||||
|
||||
@ -139,9 +139,12 @@ class BartAttention(nn.Module):
|
||||
self.num_heads = num_heads
|
||||
self.dropout = dropout
|
||||
self.head_dim = embed_dim // num_heads
|
||||
assert (
|
||||
self.head_dim * num_heads == self.embed_dim
|
||||
), f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`: {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
|
||||
|
||||
|
||||
@ -1213,9 +1213,12 @@ class BigBirdPegasusDecoderAttention(nn.Module):
|
||||
self.num_heads = num_heads
|
||||
self.dropout = dropout
|
||||
self.head_dim = embed_dim // num_heads
|
||||
assert (
|
||||
self.head_dim * num_heads == self.embed_dim
|
||||
), f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`: {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
|
||||
|
||||
|
||||
@ -141,9 +141,12 @@ class BlenderbotAttention(nn.Module):
|
||||
self.num_heads = num_heads
|
||||
self.dropout = dropout
|
||||
self.head_dim = embed_dim // num_heads
|
||||
assert (
|
||||
self.head_dim * num_heads == self.embed_dim
|
||||
), f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`: {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
|
||||
|
||||
|
||||
@ -139,9 +139,12 @@ class BlenderbotSmallAttention(nn.Module):
|
||||
self.num_heads = num_heads
|
||||
self.dropout = dropout
|
||||
self.head_dim = embed_dim // num_heads
|
||||
assert (
|
||||
self.head_dim * num_heads == self.embed_dim
|
||||
), f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`: {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
|
||||
|
||||
|
||||
@ -43,7 +43,7 @@ _CONFIG_FOR_DOC = "HubertConfig"
|
||||
_CHECKPOINT_FOR_DOC = "facebook/hubert-base-ls960"
|
||||
_PROCESSOR_FOR_DOC = "Wav2Vec2Processor"
|
||||
|
||||
_SEQ_CLASS_CHECKPOINT = ("superb/hubert-base-superb-ks",)
|
||||
_SEQ_CLASS_CHECKPOINT = "superb/hubert-base-superb-ks"
|
||||
_SEQ_CLASS_PROCESSOR_FOR_DOC = "Wav2Vec2FeatureExtractor"
|
||||
|
||||
_HIDDEN_STATES_START_POSITION = 1
|
||||
@ -354,9 +354,12 @@ class HubertAttention(nn.Module):
|
||||
self.num_heads = num_heads
|
||||
self.dropout = dropout
|
||||
self.head_dim = embed_dim // num_heads
|
||||
assert (
|
||||
self.head_dim * num_heads == self.embed_dim
|
||||
), f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`: {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
|
||||
|
||||
@ -929,9 +932,8 @@ class HubertModel(HubertPreTrainedModel):
|
||||
mask_prob=self.config.mask_feature_prob,
|
||||
mask_length=self.config.mask_feature_length,
|
||||
)
|
||||
mask_feature_indices = torch.tensor(mask_feature_indices, device=hidden_states.device, dtype=torch.bool)[
|
||||
:, None
|
||||
].expand(-1, sequence_length, -1)
|
||||
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
|
||||
|
||||
@ -210,9 +210,12 @@ class M2M100Attention(nn.Module):
|
||||
self.num_heads = num_heads
|
||||
self.dropout = dropout
|
||||
self.head_dim = embed_dim // num_heads
|
||||
assert (
|
||||
self.head_dim * num_heads == self.embed_dim
|
||||
), f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`: {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
|
||||
|
||||
|
||||
@ -156,9 +156,12 @@ class MarianAttention(nn.Module):
|
||||
self.num_heads = num_heads
|
||||
self.dropout = dropout
|
||||
self.head_dim = embed_dim // num_heads
|
||||
assert (
|
||||
self.head_dim * num_heads == self.embed_dim
|
||||
), f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`: {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
|
||||
|
||||
|
||||
@ -146,9 +146,12 @@ class MBartAttention(nn.Module):
|
||||
self.num_heads = num_heads
|
||||
self.dropout = dropout
|
||||
self.head_dim = embed_dim // num_heads
|
||||
assert (
|
||||
self.head_dim * num_heads == self.embed_dim
|
||||
), f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`: {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
|
||||
|
||||
|
||||
@ -156,9 +156,12 @@ class PegasusAttention(nn.Module):
|
||||
self.num_heads = num_heads
|
||||
self.dropout = dropout
|
||||
self.head_dim = embed_dim // num_heads
|
||||
assert (
|
||||
self.head_dim * num_heads == self.embed_dim
|
||||
), f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`: {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
|
||||
|
||||
|
||||
@ -361,9 +361,12 @@ class SEWAttention(nn.Module):
|
||||
self.num_heads = num_heads
|
||||
self.dropout = dropout
|
||||
self.head_dim = embed_dim // num_heads
|
||||
assert (
|
||||
self.head_dim * num_heads == self.embed_dim
|
||||
), f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`: {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
|
||||
|
||||
@ -831,9 +834,8 @@ class SEWModel(SEWPreTrainedModel):
|
||||
mask_prob=self.config.mask_feature_prob,
|
||||
mask_length=self.config.mask_feature_length,
|
||||
)
|
||||
mask_feature_indices = torch.tensor(mask_feature_indices, device=hidden_states.device, dtype=torch.bool)[
|
||||
:, None
|
||||
].expand(-1, sequence_length, -1)
|
||||
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
|
||||
|
||||
@ -1331,9 +1331,8 @@ class SEWDModel(SEWDPreTrainedModel):
|
||||
mask_prob=self.config.mask_feature_prob,
|
||||
mask_length=self.config.mask_feature_length,
|
||||
)
|
||||
mask_feature_indices = torch.tensor(mask_feature_indices, device=hidden_states.device, dtype=torch.bool)[
|
||||
:, None
|
||||
].expand(-1, sequence_length, -1)
|
||||
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
|
||||
|
||||
@ -223,9 +223,12 @@ class Speech2TextAttention(nn.Module):
|
||||
self.num_heads = num_heads
|
||||
self.dropout = dropout
|
||||
self.head_dim = embed_dim // num_heads
|
||||
assert (
|
||||
self.head_dim * num_heads == self.embed_dim
|
||||
), f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`: {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
|
||||
|
||||
|
||||
@ -164,9 +164,12 @@ class Speech2Text2Attention(nn.Module):
|
||||
self.num_heads = num_heads
|
||||
self.dropout = dropout
|
||||
self.head_dim = embed_dim // num_heads
|
||||
assert (
|
||||
self.head_dim * num_heads == self.embed_dim
|
||||
), f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`: {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
|
||||
|
||||
|
||||
53
src/transformers/models/unispeech/__init__.py
Normal file
53
src/transformers/models/unispeech/__init__.py
Normal file
@ -0,0 +1,53 @@
|
||||
# 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 2021 The HuggingFace Team. All rights reserved.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
from typing import TYPE_CHECKING
|
||||
|
||||
from ...file_utils import _LazyModule, is_flax_available, is_tf_available, is_torch_available
|
||||
|
||||
|
||||
_import_structure = {
|
||||
"configuration_unispeech": ["UNISPEECH_PRETRAINED_CONFIG_ARCHIVE_MAP", "UniSpeechConfig"],
|
||||
}
|
||||
|
||||
if is_torch_available():
|
||||
_import_structure["modeling_unispeech"] = [
|
||||
"UNISPEECH_PRETRAINED_MODEL_ARCHIVE_LIST",
|
||||
"UniSpeechForCTC",
|
||||
"UniSpeechForPreTraining",
|
||||
"UniSpeechForSequenceClassification",
|
||||
"UniSpeechModel",
|
||||
"UniSpeechPreTrainedModel",
|
||||
]
|
||||
|
||||
if TYPE_CHECKING:
|
||||
from .configuration_unispeech import UNISPEECH_PRETRAINED_CONFIG_ARCHIVE_MAP, UniSpeechConfig
|
||||
|
||||
if is_torch_available():
|
||||
from .modeling_unispeech import (
|
||||
UNISPEECH_PRETRAINED_MODEL_ARCHIVE_LIST,
|
||||
UniSpeechForCTC,
|
||||
UniSpeechForPreTraining,
|
||||
UniSpeechForSequenceClassification,
|
||||
UniSpeechModel,
|
||||
UniSpeechPreTrainedModel,
|
||||
)
|
||||
|
||||
else:
|
||||
import sys
|
||||
|
||||
sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure)
|
||||
273
src/transformers/models/unispeech/configuration_unispeech.py
Normal file
273
src/transformers/models/unispeech/configuration_unispeech.py
Normal file
@ -0,0 +1,273 @@
|
||||
# coding=utf-8
|
||||
# Copyright 2021 The Fairseq Authors 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.
|
||||
""" UniSpeech model configuration """
|
||||
|
||||
from ...configuration_utils import PretrainedConfig
|
||||
from ...utils import logging
|
||||
|
||||
|
||||
logger = logging.get_logger(__name__)
|
||||
|
||||
UNISPEECH_PRETRAINED_CONFIG_ARCHIVE_MAP = {
|
||||
"facebook/unispeech-base-960h": "https://huggingface.co/facebook/unispeech-base-960h/resolve/main/config.json",
|
||||
# See all UniSpeech models at https://huggingface.co/models?filter=unispeech
|
||||
}
|
||||
|
||||
|
||||
class UniSpeechConfig(PretrainedConfig):
|
||||
r"""
|
||||
This is the configuration class to store the configuration of a :class:`~transformers.UniSpeechModel`. It is used
|
||||
to instantiate an UniSpeech 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 UniSpeech
|
||||
`facebook/unispeech-base-960h <https://huggingface.co/facebook/unispeech-base-960h>`__ architecture.
|
||||
|
||||
Configuration objects inherit from :class:`~transformers.PretrainedConfig` and can be used to control the model
|
||||
outputs. Read the documentation from :class:`~transformers.PretrainedConfig` for more information.
|
||||
|
||||
|
||||
Args:
|
||||
vocab_size (:obj:`int`, `optional`, defaults to 32):
|
||||
Vocabulary size of the UniSpeech model. Defines the number of different tokens that can be represented by
|
||||
the :obj:`inputs_ids` passed when calling :class:`~transformers.UniSpeechModel`. Vocabulary size of the
|
||||
model. Defines the different tokens that can be represented by the `inputs_ids` passed to the forward
|
||||
method of :class:`~transformers.UniSpeechModel`.
|
||||
hidden_size (:obj:`int`, `optional`, defaults to 768):
|
||||
Dimensionality of the encoder layers and the pooler layer.
|
||||
num_hidden_layers (:obj:`int`, `optional`, defaults to 12):
|
||||
Number of hidden layers in the Transformer encoder.
|
||||
num_attention_heads (:obj:`int`, `optional`, defaults to 12):
|
||||
Number of attention heads for each attention layer in the Transformer encoder.
|
||||
intermediate_size (:obj:`int`, `optional`, defaults to 3072):
|
||||
Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
|
||||
hidden_act (:obj:`str` or :obj:`function`, `optional`, defaults to :obj:`"gelu"`):
|
||||
The non-linear activation function (function or string) in the encoder and pooler. If string,
|
||||
:obj:`"gelu"`, :obj:`"relu"`, :obj:`"selu"` and :obj:`"gelu_new"` are supported.
|
||||
hidden_dropout (:obj:`float`, `optional`, defaults to 0.1):
|
||||
The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
|
||||
attention_dropout (:obj:`float`, `optional`, defaults to 0.1):
|
||||
The dropout ratio for the attention probabilities.
|
||||
final_dropout (:obj:`float`, `optional`, defaults to 0.1):
|
||||
The dropout probability for the final projection layer of :class:`UniSpeechForCTC`.
|
||||
initializer_range (:obj:`float`, `optional`, defaults to 0.02):
|
||||
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
|
||||
layer_norm_eps (:obj:`float`, `optional`, defaults to 1e-12):
|
||||
The epsilon used by the layer normalization layers.
|
||||
feat_extract_norm (:obj:`str`, `optional`, defaults to :obj:`"group"`):
|
||||
The norm to be applied to 1D convolutional layers in feature extractor. One of :obj:`"group"` for group
|
||||
normalization of only the first 1D convolutional layer or :obj:`"layer"` for layer normalization of all 1D
|
||||
convolutional layers.
|
||||
feat_proj_dropout (:obj:`float`, `optional`, defaults to 0.0):
|
||||
The dropout probability for output of the feature extractor.
|
||||
feat_extract_activation (:obj:`str, `optional`, defaults to :obj:`"gelu"`):
|
||||
The non-linear activation function (function or string) in the 1D convolutional layers of the feature
|
||||
extractor. If string, :obj:`"gelu"`, :obj:`"relu"`, :obj:`"selu"` and :obj:`"gelu_new"` are supported.
|
||||
feat_quantizer_dropout (obj:`float`, `optional`, defaults to 0.0):
|
||||
The dropout probabilitiy for quantized feature extractor states.
|
||||
conv_dim (:obj:`Tuple[int]`, `optional`, defaults to :obj:`(512, 512, 512, 512, 512, 512, 512)`):
|
||||
A tuple of integers defining the number of input and output channels of each 1D convolutional layer in the
|
||||
feature extractor. The length of `conv_dim` defines the number of 1D convolutional layers.
|
||||
conv_stride (:obj:`Tuple[int]`, `optional`, defaults to :obj:`(5, 2, 2, 2, 2, 2, 2)`):
|
||||
A tuple of integers defining the stride of each 1D convolutional layer in the feature extractor. The length
|
||||
of `conv_stride` defines the number of convolutional layers and has to match the the length of `conv_dim`.
|
||||
conv_kernel (:obj:`Tuple[int]`, `optional`, defaults to :obj:`(10, 3, 3, 3, 3, 3, 3)`):
|
||||
A tuple of integers defining the kernel size of each 1D convolutional layer in the feature extractor. The
|
||||
length of `conv_kernel` defines the number of convolutional layers and has to match the the length of
|
||||
`conv_dim`.
|
||||
conv_bias (:obj:`bool`, `optional`, defaults to :obj:`False`):
|
||||
Whether the 1D convolutional layers have a bias.
|
||||
num_conv_pos_embeddings (:obj:`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 (:obj:`int`, `optional`, defaults to 16):
|
||||
Number of groups of 1D convolutional positional embeddings layer.
|
||||
do_stable_layer_norm (:obj:`bool`, `optional`, defaults to :obj:`False`):
|
||||
Whether to apply `stable` layer norm architecture of the Transformer encoder. ``do_stable_layer_norm is
|
||||
True`` corresponds to applying layer norm before the attention layer, whereas ``do_stable_layer_norm is
|
||||
False`` corresponds to applying layer norm after the attention layer.
|
||||
apply_spec_augment (:obj:`bool`, `optional`, defaults to :obj:`True`):
|
||||
Whether to apply *SpecAugment* data augmentation to the outputs of the feature extractor. For reference see
|
||||
`SpecAugment: A Simple Data Augmentation Method for Automatic Speech Recognition
|
||||
<https://arxiv.org/abs/1904.08779>`__.
|
||||
mask_time_prob (:obj:`float`, `optional`, defaults to 0.05):
|
||||
Propability of each feature vector along the time axis to be chosen as the start of the vector span to be
|
||||
masked. Approximately ``mask_time_prob * sequence_length // mask_time_length`` feature vectors will be
|
||||
masked along the time axis. This is only relevant if ``apply_spec_augment is True``.
|
||||
mask_time_length (:obj:`int`, `optional`, defaults to 10):
|
||||
Length of vector span along the time axis.
|
||||
mask_feature_prob (:obj:`float`, `optional`, defaults to 0.0):
|
||||
Propability of each feature vector along the feature axis to be chosen as the start of the vector span to
|
||||
be masked. Approximately ``mask_time_prob * hidden_size // mask_time_length`` feature vectors will be
|
||||
masked along the time axis. This is only relevant if ``apply_spec_augment is True``.
|
||||
mask_feature_length (:obj:`int`, `optional`, defaults to 10):
|
||||
Length of vector span along the feature axis.
|
||||
num_codevectors_per_group (:obj:`int`, `optional`, defaults to 320):
|
||||
Number of entries in each quantization codebook (group).
|
||||
num_codevector_groups (:obj:`int`, `optional`, defaults to 2):
|
||||
Number of codevector groups for product codevector quantization.
|
||||
contrastive_logits_temperature (:obj:`float`, `optional`, defaults to 0.1):
|
||||
The temperature `kappa` in the contrastive loss.
|
||||
feat_quantizer_dropout (:obj:`float`, `optional`, defaults to 0.0):
|
||||
The dropout probabilitiy for the output of the feature extractor that's used by the quantizer.
|
||||
num_negatives (:obj:`int`, `optional`, defaults to 100):
|
||||
Number of negative samples for the contrastive loss.
|
||||
codevector_dim (:obj:`int`, `optional`, defaults to 256):
|
||||
Dimensionality of the quantized feature vectors.
|
||||
proj_codevector_dim (:obj:`int`, `optional`, defaults to 256):
|
||||
Dimensionality of the final projection of both the quantized and the transformer features.
|
||||
diversity_loss_weight (:obj:`int`, `optional`, defaults to 0.1):
|
||||
The weight of the codebook diversity loss component.
|
||||
ctc_loss_reduction (:obj:`str`, `optional`, defaults to :obj:`"mean"`):
|
||||
Specifies the reduction to apply to the output of ``torch.nn.CTCLoss``. Only relevant when training an
|
||||
instance of :class:`~transformers.UniSpeechForCTC`.
|
||||
ctc_zero_infinity (:obj:`bool`, `optional`, defaults to :obj:`False`):
|
||||
Whether to zero infinite losses and the associated gradients of ``torch.nn.CTCLoss``. Infinite losses
|
||||
mainly occur when the inputs are too short to be aligned to the targets. Only relevant when training an
|
||||
instance of :class:`~transformers.UniSpeechForCTC`.
|
||||
use_weighted_layer_sum (:obj:`bool`, `optional`, defaults to :obj:`False`):
|
||||
Whether to use a weighted average of layer outputs with learned weights. Only relevant when using an
|
||||
instance of :class:`~transformers.UniSpeechForSequenceClassification`.
|
||||
classifier_proj_size (:obj:`int`, `optional`, defaults to 256):
|
||||
Dimensionality of the projection before token mean-pooling for classification.
|
||||
replace_prob (:obj:`float`, `optional`, defaults to 0.5):
|
||||
Propability that transformer feature is replaced by quantized feature for pretraining.
|
||||
|
||||
Example::
|
||||
|
||||
>>> from transformers import UniSpeechModel, UniSpeechConfig
|
||||
|
||||
>>> # Initializing a UniSpeech facebook/unispeech-base-960h style configuration
|
||||
>>> configuration = UniSpeechConfig()
|
||||
|
||||
>>> # Initializing a model from the facebook/unispeech-base-960h style configuration
|
||||
>>> model = UniSpeechModel(configuration)
|
||||
|
||||
>>> # Accessing the model configuration
|
||||
>>> configuration = model.config
|
||||
"""
|
||||
model_type = "unispeech"
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
vocab_size=32,
|
||||
hidden_size=768,
|
||||
num_hidden_layers=12,
|
||||
num_attention_heads=12,
|
||||
intermediate_size=3072,
|
||||
hidden_act="gelu",
|
||||
hidden_dropout=0.1,
|
||||
activation_dropout=0.1,
|
||||
attention_dropout=0.1,
|
||||
feat_proj_dropout=0.0,
|
||||
feat_quantizer_dropout=0.0,
|
||||
final_dropout=0.1,
|
||||
layerdrop=0.1,
|
||||
initializer_range=0.02,
|
||||
layer_norm_eps=1e-5,
|
||||
feat_extract_norm="group",
|
||||
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,
|
||||
do_stable_layer_norm=False,
|
||||
apply_spec_augment=True,
|
||||
mask_time_prob=0.05,
|
||||
mask_time_length=10,
|
||||
mask_feature_prob=0.0,
|
||||
mask_feature_length=10,
|
||||
num_codevectors_per_group=320,
|
||||
num_codevector_groups=2,
|
||||
contrastive_logits_temperature=0.1,
|
||||
num_negatives=100,
|
||||
codevector_dim=256,
|
||||
proj_codevector_dim=256,
|
||||
diversity_loss_weight=0.1,
|
||||
ctc_loss_reduction="mean",
|
||||
ctc_zero_infinity=False,
|
||||
use_weighted_layer_sum=False,
|
||||
classifier_proj_size=256,
|
||||
num_ctc_classes=80,
|
||||
pad_token_id=0,
|
||||
bos_token_id=1,
|
||||
eos_token_id=2,
|
||||
replace_prob=0.5,
|
||||
**kwargs
|
||||
):
|
||||
super().__init__(**kwargs, pad_token_id=pad_token_id, bos_token_id=bos_token_id, eos_token_id=eos_token_id)
|
||||
self.hidden_size = hidden_size
|
||||
self.feat_extract_norm = feat_extract_norm
|
||||
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)
|
||||
self.num_hidden_layers = num_hidden_layers
|
||||
self.intermediate_size = intermediate_size
|
||||
self.hidden_act = hidden_act
|
||||
self.num_attention_heads = num_attention_heads
|
||||
self.hidden_dropout = hidden_dropout
|
||||
self.attention_dropout = attention_dropout
|
||||
self.activation_dropout = activation_dropout
|
||||
self.feat_proj_dropout = feat_proj_dropout
|
||||
self.final_dropout = final_dropout
|
||||
self.layerdrop = layerdrop
|
||||
self.layer_norm_eps = layer_norm_eps
|
||||
self.initializer_range = initializer_range
|
||||
self.num_ctc_classes = num_ctc_classes
|
||||
self.vocab_size = vocab_size
|
||||
self.do_stable_layer_norm = do_stable_layer_norm
|
||||
self.use_weighted_layer_sum = use_weighted_layer_sum
|
||||
self.classifier_proj_size = classifier_proj_size
|
||||
|
||||
if (
|
||||
(len(self.conv_stride) != self.num_feat_extract_layers)
|
||||
or (len(self.conv_kernel) != self.num_feat_extract_layers)
|
||||
or (len(self.conv_dim) != self.num_feat_extract_layers)
|
||||
):
|
||||
raise ValueError(
|
||||
"Configuration for convolutional layers is incorrect. "
|
||||
"It is required that `len(config.conv_dim)` == `len(config.conv_stride)` == `len(config.conv_kernel)`, "
|
||||
f"but is `len(config.conv_dim) = {len(self.conv_dim)}`, `len(config.conv_stride) "
|
||||
f"= {len(self.conv_stride)}`, `len(config.conv_kernel) = {len(self.conv_kernel)}`."
|
||||
)
|
||||
|
||||
# fine-tuning config parameters for SpecAugment: https://arxiv.org/abs/1904.08779
|
||||
self.apply_spec_augment = apply_spec_augment
|
||||
self.mask_time_prob = mask_time_prob
|
||||
self.mask_time_length = mask_time_length
|
||||
self.mask_feature_prob = mask_feature_prob
|
||||
self.mask_feature_length = mask_feature_length
|
||||
|
||||
# parameters for pretraining with codevector quantized representations
|
||||
self.num_codevectors_per_group = num_codevectors_per_group
|
||||
self.num_codevector_groups = num_codevector_groups
|
||||
self.contrastive_logits_temperature = contrastive_logits_temperature
|
||||
self.feat_quantizer_dropout = feat_quantizer_dropout
|
||||
self.num_negatives = num_negatives
|
||||
self.codevector_dim = codevector_dim
|
||||
self.proj_codevector_dim = proj_codevector_dim
|
||||
self.diversity_loss_weight = diversity_loss_weight
|
||||
|
||||
# ctc loss
|
||||
self.ctc_loss_reduction = ctc_loss_reduction
|
||||
self.ctc_zero_infinity = ctc_zero_infinity
|
||||
|
||||
# pretraining loss
|
||||
self.replace_prob = replace_prob
|
||||
@ -0,0 +1,191 @@
|
||||
# coding=utf-8
|
||||
# Copyright 2021 The HuggingFace Inc. team.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
"""Convert UniSpeech checkpoint."""
|
||||
|
||||
|
||||
import argparse
|
||||
|
||||
import fairseq
|
||||
import torch
|
||||
|
||||
from transformers import UniSpeechConfig, UniSpeechForPreTraining, logging
|
||||
|
||||
|
||||
logging.set_verbosity_info()
|
||||
logger = logging.get_logger(__name__)
|
||||
|
||||
MAPPING = {
|
||||
"post_extract_proj": "feature_projection.projection",
|
||||
"encoder.pos_conv.0": "encoder.pos_conv_embed.conv",
|
||||
"self_attn.k_proj": "encoder.layers.*.attention.k_proj",
|
||||
"self_attn.v_proj": "encoder.layers.*.attention.v_proj",
|
||||
"self_attn.q_proj": "encoder.layers.*.attention.q_proj",
|
||||
"self_attn.out_proj": "encoder.layers.*.attention.out_proj",
|
||||
"self_attn_layer_norm": "encoder.layers.*.layer_norm",
|
||||
"fc1": "encoder.layers.*.feed_forward.intermediate_dense",
|
||||
"fc2": "encoder.layers.*.feed_forward.output_dense",
|
||||
"final_layer_norm": "encoder.layers.*.final_layer_norm",
|
||||
"encoder.layer_norm": "encoder.layer_norm",
|
||||
"w2v_model.layer_norm": "feature_projection.layer_norm",
|
||||
"quantizer.weight_proj": "quantizer.weight_proj",
|
||||
"quantizer.vars": "quantizer.codevectors",
|
||||
"project_q": "project_q",
|
||||
"final_proj": "project_hid",
|
||||
"w2v_encoder.proj": "ctc_proj",
|
||||
"mask_emb": "masked_spec_embed",
|
||||
}
|
||||
TOP_LEVEL_KEYS = [
|
||||
"ctc_proj",
|
||||
"quantizer.weight_proj",
|
||||
"quantizer.codevectors",
|
||||
"project_q",
|
||||
"project_hid",
|
||||
]
|
||||
|
||||
|
||||
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
|
||||
|
||||
assert (
|
||||
hf_shape == value.shape
|
||||
), f"Shape of hf {key + '.' + weight_type if weight_type is not None else ''} is {hf_shape}, but should be {value.shape} for {full_name}"
|
||||
|
||||
if weight_type == "weight":
|
||||
hf_pointer.weight.data = value
|
||||
elif weight_type == "weight_g":
|
||||
hf_pointer.weight_g.data = value
|
||||
elif weight_type == "weight_v":
|
||||
hf_pointer.weight_v.data = value
|
||||
elif weight_type == "bias":
|
||||
hf_pointer.bias.data = value
|
||||
else:
|
||||
hf_pointer.data = value
|
||||
|
||||
logger.info(f"{key + '.' + weight_type if weight_type is not None else ''} was initialized from {full_name}.")
|
||||
|
||||
|
||||
def recursively_load_weights(fairseq_model, hf_model):
|
||||
unused_weights = []
|
||||
fairseq_dict = fairseq_model.state_dict()
|
||||
|
||||
feature_extractor = hf_model.unispeech.feature_extractor
|
||||
|
||||
for name, value in fairseq_dict.items():
|
||||
is_used = False
|
||||
if "conv_layers" in name:
|
||||
load_conv_layer(
|
||||
name,
|
||||
value,
|
||||
feature_extractor,
|
||||
unused_weights,
|
||||
hf_model.config.feat_extract_norm == "group",
|
||||
)
|
||||
is_used = True
|
||||
else:
|
||||
for key, mapped_key in MAPPING.items():
|
||||
mapped_key = "unispeech." + mapped_key if mapped_key not in TOP_LEVEL_KEYS else mapped_key
|
||||
if key in name or key.split("w2v_model.")[-1] == name.split(".")[0]:
|
||||
is_used = True
|
||||
if "*" in mapped_key:
|
||||
layer_index = name.split(key)[0].split(".")[-2]
|
||||
mapped_key = mapped_key.replace("*", layer_index)
|
||||
if "weight_g" in name:
|
||||
weight_type = "weight_g"
|
||||
elif "weight_v" in name:
|
||||
weight_type = "weight_v"
|
||||
elif "bias" in name:
|
||||
weight_type = "bias"
|
||||
elif "weight" in name:
|
||||
# TODO: don't match quantizer.weight_proj
|
||||
weight_type = "weight"
|
||||
else:
|
||||
weight_type = None
|
||||
set_recursively(hf_model, mapped_key, value, name, weight_type)
|
||||
continue
|
||||
if not is_used:
|
||||
unused_weights.append(name)
|
||||
|
||||
logger.warning(f"Unused weights: {unused_weights}")
|
||||
|
||||
|
||||
def 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:
|
||||
assert (
|
||||
value.shape == feature_extractor.conv_layers[layer_id].conv.bias.data.shape
|
||||
), f"{full_name} has size {value.shape}, but {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:
|
||||
assert (
|
||||
value.shape == feature_extractor.conv_layers[layer_id].conv.weight.data.shape
|
||||
), f"{full_name} has size {value.shape}, but {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:
|
||||
assert (
|
||||
value.shape == feature_extractor.conv_layers[layer_id].layer_norm.bias.data.shape
|
||||
), f"{full_name} has size {value.shape}, but {feature_extractor[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:
|
||||
assert (
|
||||
value.shape == feature_extractor.conv_layers[layer_id].layer_norm.weight.data.shape
|
||||
), f"{full_name} has size {value.shape}, but {feature_extractor[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_unispeech_checkpoint(checkpoint_path, pytorch_dump_folder_path, config_path=None):
|
||||
"""
|
||||
Copy/paste/tweak model's weights to transformers design.
|
||||
"""
|
||||
if config_path is not None:
|
||||
config = UniSpeechConfig.from_pretrained(config_path)
|
||||
else:
|
||||
config = UniSpeechConfig()
|
||||
|
||||
hf_unispeech = UniSpeechForPreTraining(config)
|
||||
|
||||
model, _, _ = fairseq.checkpoint_utils.load_model_ensemble_and_task([checkpoint_path])
|
||||
model = model[0].eval()
|
||||
|
||||
recursively_load_weights(model, hf_unispeech)
|
||||
|
||||
hf_unispeech.save_pretrained(pytorch_dump_folder_path)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
parser = argparse.ArgumentParser()
|
||||
parser.add_argument("--pytorch_dump_folder_path", default=None, type=str, help="Path to the output PyTorch model.")
|
||||
parser.add_argument("--checkpoint_path", default=None, type=str, help="Path to fairseq checkpoint")
|
||||
parser.add_argument("--config_path", default=None, type=str, help="Path to hf config.json of model to convert")
|
||||
args = parser.parse_args()
|
||||
convert_unispeech_checkpoint(args.checkpoint_path, args.pytorch_dump_folder_path, args.config_path)
|
||||
1524
src/transformers/models/unispeech/modeling_unispeech.py
Executable file
1524
src/transformers/models/unispeech/modeling_unispeech.py
Executable file
File diff suppressed because it is too large
Load Diff
53
src/transformers/models/unispeech_sat/__init__.py
Normal file
53
src/transformers/models/unispeech_sat/__init__.py
Normal file
@ -0,0 +1,53 @@
|
||||
# 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 2021 The HuggingFace Team. All rights reserved.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
from typing import TYPE_CHECKING
|
||||
|
||||
from ...file_utils import _LazyModule, is_flax_available, is_tf_available, is_torch_available
|
||||
|
||||
|
||||
_import_structure = {
|
||||
"configuration_unispeech_sat": ["UNISPEECH_SAT_PRETRAINED_CONFIG_ARCHIVE_MAP", "UniSpeechSatConfig"],
|
||||
}
|
||||
|
||||
if is_torch_available():
|
||||
_import_structure["modeling_unispeech_sat"] = [
|
||||
"UNISPEECH_SAT_PRETRAINED_MODEL_ARCHIVE_LIST",
|
||||
"UniSpeechSatForCTC",
|
||||
"UniSpeechSatForPreTraining",
|
||||
"UniSpeechSatForSequenceClassification",
|
||||
"UniSpeechSatModel",
|
||||
"UniSpeechSatPreTrainedModel",
|
||||
]
|
||||
|
||||
if TYPE_CHECKING:
|
||||
from .configuration_unispeech_sat import UNISPEECH_SAT_PRETRAINED_CONFIG_ARCHIVE_MAP, UniSpeechSatConfig
|
||||
|
||||
if is_torch_available():
|
||||
from .modeling_unispeech_sat import (
|
||||
UNISPEECH_SAT_PRETRAINED_MODEL_ARCHIVE_LIST,
|
||||
UniSpeechSatForCTC,
|
||||
UniSpeechSatForPreTraining,
|
||||
UniSpeechSatForSequenceClassification,
|
||||
UniSpeechSatModel,
|
||||
UniSpeechSatPreTrainedModel,
|
||||
)
|
||||
|
||||
else:
|
||||
import sys
|
||||
|
||||
sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure)
|
||||
@ -0,0 +1,267 @@
|
||||
# coding=utf-8
|
||||
# Copyright 2021 The Fairseq Authors 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.
|
||||
""" UniSpeechSat model configuration """
|
||||
|
||||
from ...configuration_utils import PretrainedConfig
|
||||
from ...utils import logging
|
||||
|
||||
|
||||
logger = logging.get_logger(__name__)
|
||||
|
||||
UNISPEECH_SAT_PRETRAINED_CONFIG_ARCHIVE_MAP = {
|
||||
"facebook/unispeech_sat-base-960h": "https://huggingface.co/facebook/unispeech_sat-base-960h/resolve/main/config.json",
|
||||
# See all UniSpeechSat models at https://huggingface.co/models?filter=unispeech_sat
|
||||
}
|
||||
|
||||
|
||||
class UniSpeechSatConfig(PretrainedConfig):
|
||||
r"""
|
||||
This is the configuration class to store the configuration of a :class:`~transformers.UniSpeechSatModel`. It is
|
||||
used to instantiate an UniSpeechSat 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 UniSpeechSat
|
||||
`facebook/unispeech_sat-base-960h <https://huggingface.co/facebook/unispeech_sat-base-960h>`__ architecture.
|
||||
|
||||
Configuration objects inherit from :class:`~transformers.PretrainedConfig` and can be used to control the model
|
||||
outputs. Read the documentation from :class:`~transformers.PretrainedConfig` for more information.
|
||||
|
||||
|
||||
Args:
|
||||
vocab_size (:obj:`int`, `optional`, defaults to 32):
|
||||
Vocabulary size of the UniSpeechSat model. Defines the number of different tokens that can be represented
|
||||
by the :obj:`inputs_ids` passed when calling :class:`~transformers.UniSpeechSatModel`. Vocabulary size of
|
||||
the model. Defines the different tokens that can be represented by the `inputs_ids` passed to the forward
|
||||
method of :class:`~transformers.UniSpeechSatModel`.
|
||||
hidden_size (:obj:`int`, `optional`, defaults to 768):
|
||||
Dimensionality of the encoder layers and the pooler layer.
|
||||
num_hidden_layers (:obj:`int`, `optional`, defaults to 12):
|
||||
Number of hidden layers in the Transformer encoder.
|
||||
num_attention_heads (:obj:`int`, `optional`, defaults to 12):
|
||||
Number of attention heads for each attention layer in the Transformer encoder.
|
||||
intermediate_size (:obj:`int`, `optional`, defaults to 3072):
|
||||
Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
|
||||
hidden_act (:obj:`str` or :obj:`function`, `optional`, defaults to :obj:`"gelu"`):
|
||||
The non-linear activation function (function or string) in the encoder and pooler. If string,
|
||||
:obj:`"gelu"`, :obj:`"relu"`, :obj:`"selu"` and :obj:`"gelu_new"` are supported.
|
||||
hidden_dropout (:obj:`float`, `optional`, defaults to 0.1):
|
||||
The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
|
||||
attention_dropout (:obj:`float`, `optional`, defaults to 0.1):
|
||||
The dropout ratio for the attention probabilities.
|
||||
final_dropout (:obj:`float`, `optional`, defaults to 0.1):
|
||||
The dropout probability for the final projection layer of :class:`UniSpeechSatForCTC`.
|
||||
initializer_range (:obj:`float`, `optional`, defaults to 0.02):
|
||||
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
|
||||
layer_norm_eps (:obj:`float`, `optional`, defaults to 1e-12):
|
||||
The epsilon used by the layer normalization layers.
|
||||
feat_extract_norm (:obj:`str`, `optional`, defaults to :obj:`"group"`):
|
||||
The norm to be applied to 1D convolutional layers in feature extractor. One of :obj:`"group"` for group
|
||||
normalization of only the first 1D convolutional layer or :obj:`"layer"` for layer normalization of all 1D
|
||||
convolutional layers.
|
||||
feat_proj_dropout (:obj:`float`, `optional`, defaults to 0.0):
|
||||
The dropout probability for output of the feature extractor.
|
||||
feat_extract_activation (:obj:`str, `optional`, defaults to :obj:`"gelu"`):
|
||||
The non-linear activation function (function or string) in the 1D convolutional layers of the feature
|
||||
extractor. If string, :obj:`"gelu"`, :obj:`"relu"`, :obj:`"selu"` and :obj:`"gelu_new"` are supported.
|
||||
feat_quantizer_dropout (obj:`float`, `optional`, defaults to 0.0):
|
||||
The dropout probabilitiy for quantized feature extractor states.
|
||||
conv_dim (:obj:`Tuple[int]`, `optional`, defaults to :obj:`(512, 512, 512, 512, 512, 512, 512)`):
|
||||
A tuple of integers defining the number of input and output channels of each 1D convolutional layer in the
|
||||
feature extractor. The length of `conv_dim` defines the number of 1D convolutional layers.
|
||||
conv_stride (:obj:`Tuple[int]`, `optional`, defaults to :obj:`(5, 2, 2, 2, 2, 2, 2)`):
|
||||
A tuple of integers defining the stride of each 1D convolutional layer in the feature extractor. The length
|
||||
of `conv_stride` defines the number of convolutional layers and has to match the the length of `conv_dim`.
|
||||
conv_kernel (:obj:`Tuple[int]`, `optional`, defaults to :obj:`(10, 3, 3, 3, 3, 3, 3)`):
|
||||
A tuple of integers defining the kernel size of each 1D convolutional layer in the feature extractor. The
|
||||
length of `conv_kernel` defines the number of convolutional layers and has to match the the length of
|
||||
`conv_dim`.
|
||||
conv_bias (:obj:`bool`, `optional`, defaults to :obj:`False`):
|
||||
Whether the 1D convolutional layers have a bias.
|
||||
num_conv_pos_embeddings (:obj:`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 (:obj:`int`, `optional`, defaults to 16):
|
||||
Number of groups of 1D convolutional positional embeddings layer.
|
||||
do_stable_layer_norm (:obj:`bool`, `optional`, defaults to :obj:`False`):
|
||||
Whether to apply `stable` layer norm architecture of the Transformer encoder. ``do_stable_layer_norm is
|
||||
True`` corresponds to applying layer norm before the attention layer, whereas ``do_stable_layer_norm is
|
||||
False`` corresponds to applying layer norm after the attention layer.
|
||||
apply_spec_augment (:obj:`bool`, `optional`, defaults to :obj:`True`):
|
||||
Whether to apply *SpecAugment* data augmentation to the outputs of the feature extractor. For reference see
|
||||
`SpecAugment: A Simple Data Augmentation Method for Automatic Speech Recognition
|
||||
<https://arxiv.org/abs/1904.08779>`__.
|
||||
mask_time_prob (:obj:`float`, `optional`, defaults to 0.05):
|
||||
Propability of each feature vector along the time axis to be chosen as the start of the vector span to be
|
||||
masked. Approximately ``mask_time_prob * sequence_length // mask_time_length`` feature vectors will be
|
||||
masked along the time axis. This is only relevant if ``apply_spec_augment is True``.
|
||||
mask_time_length (:obj:`int`, `optional`, defaults to 10):
|
||||
Length of vector span along the time axis.
|
||||
mask_feature_prob (:obj:`float`, `optional`, defaults to 0.0):
|
||||
Propability of each feature vector along the feature axis to be chosen as the start of the vector span to
|
||||
be masked. Approximately ``mask_time_prob * hidden_size // mask_time_length`` feature vectors will be
|
||||
masked along the time axis. This is only relevant if ``apply_spec_augment is True``.
|
||||
mask_feature_length (:obj:`int`, `optional`, defaults to 10):
|
||||
Length of vector span along the feature axis.
|
||||
num_codevectors_per_group (:obj:`int`, `optional`, defaults to 320):
|
||||
Number of entries in each quantization codebook (group).
|
||||
num_codevector_groups (:obj:`int`, `optional`, defaults to 2):
|
||||
Number of codevector groups for product codevector quantization.
|
||||
contrastive_logits_temperature (:obj:`float`, `optional`, defaults to 0.1):
|
||||
The temperature `kappa` in the contrastive loss.
|
||||
feat_quantizer_dropout (:obj:`float`, `optional`, defaults to 0.0):
|
||||
The dropout probabilitiy for the output of the feature extractor that's used by the quantizer.
|
||||
num_negatives (:obj:`int`, `optional`, defaults to 100):
|
||||
Number of negative samples for the contrastive loss.
|
||||
codevector_dim (:obj:`int`, `optional`, defaults to 256):
|
||||
Dimensionality of the quantized feature vectors.
|
||||
proj_codevector_dim (:obj:`int`, `optional`, defaults to 256):
|
||||
Dimensionality of the final projection of both the quantized and the transformer features.
|
||||
diversity_loss_weight (:obj:`int`, `optional`, defaults to 0.1):
|
||||
The weight of the codebook diversity loss component.
|
||||
ctc_loss_reduction (:obj:`str`, `optional`, defaults to :obj:`"mean"`):
|
||||
Specifies the reduction to apply to the output of ``torch.nn.CTCLoss``. Only relevant when training an
|
||||
instance of :class:`~transformers.UniSpeechSatForCTC`.
|
||||
ctc_zero_infinity (:obj:`bool`, `optional`, defaults to :obj:`False`):
|
||||
Whether to zero infinite losses and the associated gradients of ``torch.nn.CTCLoss``. Infinite losses
|
||||
mainly occur when the inputs are too short to be aligned to the targets. Only relevant when training an
|
||||
instance of :class:`~transformers.UniSpeechSatForCTC`.
|
||||
use_weighted_layer_sum (:obj:`bool`, `optional`, defaults to :obj:`False`):
|
||||
Whether to use a weighted average of layer outputs with learned weights. Only relevant when using an
|
||||
instance of :class:`~transformers.UniSpeechSatForSequenceClassification`.
|
||||
classifier_proj_size (:obj:`int`, `optional`, defaults to 256):
|
||||
Dimensionality of the projection before token mean-pooling for classification.
|
||||
|
||||
Example::
|
||||
|
||||
>>> from transformers import UniSpeechSatModel, UniSpeechSatConfig
|
||||
|
||||
>>> # Initializing a UniSpeechSat facebook/unispeech_sat-base-960h style configuration
|
||||
>>> configuration = UniSpeechSatConfig()
|
||||
|
||||
>>> # Initializing a model from the facebook/unispeech_sat-base-960h style configuration
|
||||
>>> model = UniSpeechSatModel(configuration)
|
||||
|
||||
>>> # Accessing the model configuration
|
||||
>>> configuration = model.config
|
||||
"""
|
||||
model_type = "unispeech-sat"
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
vocab_size=32,
|
||||
hidden_size=768,
|
||||
num_hidden_layers=12,
|
||||
num_attention_heads=12,
|
||||
intermediate_size=3072,
|
||||
hidden_act="gelu",
|
||||
hidden_dropout=0.1,
|
||||
activation_dropout=0.1,
|
||||
attention_dropout=0.1,
|
||||
feat_proj_dropout=0.0,
|
||||
feat_quantizer_dropout=0.0,
|
||||
final_dropout=0.1,
|
||||
layerdrop=0.1,
|
||||
initializer_range=0.02,
|
||||
layer_norm_eps=1e-5,
|
||||
feat_extract_norm="group",
|
||||
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,
|
||||
do_stable_layer_norm=False,
|
||||
apply_spec_augment=True,
|
||||
mask_time_prob=0.05,
|
||||
mask_time_length=10,
|
||||
mask_feature_prob=0.0,
|
||||
mask_feature_length=10,
|
||||
num_codevectors_per_group=320,
|
||||
num_codevector_groups=2,
|
||||
contrastive_logits_temperature=0.1,
|
||||
num_negatives=100,
|
||||
codevector_dim=256,
|
||||
proj_codevector_dim=256,
|
||||
diversity_loss_weight=0.1,
|
||||
ctc_loss_reduction="mean",
|
||||
ctc_zero_infinity=False,
|
||||
use_weighted_layer_sum=False,
|
||||
classifier_proj_size=256,
|
||||
pad_token_id=0,
|
||||
bos_token_id=1,
|
||||
eos_token_id=2,
|
||||
num_clusters=504,
|
||||
**kwargs
|
||||
):
|
||||
super().__init__(**kwargs, pad_token_id=pad_token_id, bos_token_id=bos_token_id, eos_token_id=eos_token_id)
|
||||
self.hidden_size = hidden_size
|
||||
self.feat_extract_norm = feat_extract_norm
|
||||
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)
|
||||
self.num_hidden_layers = num_hidden_layers
|
||||
self.intermediate_size = intermediate_size
|
||||
self.hidden_act = hidden_act
|
||||
self.num_attention_heads = num_attention_heads
|
||||
self.hidden_dropout = hidden_dropout
|
||||
self.attention_dropout = attention_dropout
|
||||
self.activation_dropout = activation_dropout
|
||||
self.feat_proj_dropout = feat_proj_dropout
|
||||
self.final_dropout = final_dropout
|
||||
self.layerdrop = layerdrop
|
||||
self.layer_norm_eps = layer_norm_eps
|
||||
self.initializer_range = initializer_range
|
||||
self.vocab_size = vocab_size
|
||||
self.num_clusters = num_clusters
|
||||
self.do_stable_layer_norm = do_stable_layer_norm
|
||||
self.use_weighted_layer_sum = use_weighted_layer_sum
|
||||
self.classifier_proj_size = classifier_proj_size
|
||||
|
||||
if (
|
||||
(len(self.conv_stride) != self.num_feat_extract_layers)
|
||||
or (len(self.conv_kernel) != self.num_feat_extract_layers)
|
||||
or (len(self.conv_dim) != self.num_feat_extract_layers)
|
||||
):
|
||||
raise ValueError(
|
||||
"Configuration for convolutional layers is incorrect. "
|
||||
"It is required that `len(config.conv_dim)` == `len(config.conv_stride)` == `len(config.conv_kernel)`, "
|
||||
f"but is `len(config.conv_dim) = {len(self.conv_dim)}`, `len(config.conv_stride) "
|
||||
f"= {len(self.conv_stride)}`, `len(config.conv_kernel) = {len(self.conv_kernel)}`."
|
||||
)
|
||||
|
||||
# fine-tuning config parameters for SpecAugment: https://arxiv.org/abs/1904.08779
|
||||
self.apply_spec_augment = apply_spec_augment
|
||||
self.mask_time_prob = mask_time_prob
|
||||
self.mask_time_length = mask_time_length
|
||||
self.mask_feature_prob = mask_feature_prob
|
||||
self.mask_feature_length = mask_feature_length
|
||||
|
||||
# parameters for pretraining with codevector quantized representations
|
||||
self.num_codevectors_per_group = num_codevectors_per_group
|
||||
self.num_codevector_groups = num_codevector_groups
|
||||
self.contrastive_logits_temperature = contrastive_logits_temperature
|
||||
self.feat_quantizer_dropout = feat_quantizer_dropout
|
||||
self.num_negatives = num_negatives
|
||||
self.codevector_dim = codevector_dim
|
||||
self.proj_codevector_dim = proj_codevector_dim
|
||||
self.diversity_loss_weight = diversity_loss_weight
|
||||
|
||||
# ctc loss
|
||||
self.ctc_loss_reduction = ctc_loss_reduction
|
||||
self.ctc_zero_infinity = ctc_zero_infinity
|
||||
@ -0,0 +1,220 @@
|
||||
# coding=utf-8
|
||||
# Copyright 2021 The HuggingFace Inc. team.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
"""Convert UniSpeechSat checkpoint."""
|
||||
|
||||
|
||||
import argparse
|
||||
|
||||
import fairseq
|
||||
import torch
|
||||
|
||||
from transformers import ( # UniSpeechSatCTCTokenizer,; UniSpeechSatFeatureExtractor,; UniSpeechSatProcessor,
|
||||
UniSpeechSatConfig,
|
||||
UniSpeechSatForCTC,
|
||||
UniSpeechSatForPreTraining,
|
||||
logging,
|
||||
)
|
||||
|
||||
|
||||
logging.set_verbosity_info()
|
||||
logger = logging.get_logger(__name__)
|
||||
|
||||
MAPPING = {
|
||||
"post_extract_proj": "feature_projection.projection",
|
||||
"encoder.pos_conv.0": "encoder.pos_conv_embed.conv",
|
||||
"self_attn.k_proj": "encoder.layers.*.attention.k_proj",
|
||||
"self_attn.v_proj": "encoder.layers.*.attention.v_proj",
|
||||
"self_attn.q_proj": "encoder.layers.*.attention.q_proj",
|
||||
"self_attn.out_proj": "encoder.layers.*.attention.out_proj",
|
||||
"self_attn_layer_norm": "encoder.layers.*.layer_norm",
|
||||
"fc1": "encoder.layers.*.feed_forward.intermediate_dense",
|
||||
"fc2": "encoder.layers.*.feed_forward.output_dense",
|
||||
"final_layer_norm": "encoder.layers.*.final_layer_norm",
|
||||
"encoder.layer_norm": "encoder.layer_norm",
|
||||
"encoder.layer_norm_for_extract": "layer_norm_for_extract",
|
||||
"w2v_model.layer_norm": "feature_projection.layer_norm",
|
||||
"quantizer.weight_proj": "quantizer.weight_proj",
|
||||
"quantizer.vars": "quantizer.codevectors",
|
||||
"project_q": "project_q",
|
||||
"final_proj": "project_hid",
|
||||
"w2v_encoder.proj": "lm_head",
|
||||
"label_embs_concat": "label_embeddings_concat",
|
||||
"mask_emb": "masked_spec_embed",
|
||||
"spk_proj": "speaker_proj",
|
||||
}
|
||||
TOP_LEVEL_KEYS = [
|
||||
"lm_head",
|
||||
"quantizer.weight_proj",
|
||||
"quantizer.codevectors",
|
||||
"project_q",
|
||||
"project_hid",
|
||||
"label_embeddings_concat",
|
||||
"speaker_proj",
|
||||
"layer_norm_for_extract",
|
||||
]
|
||||
|
||||
|
||||
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
|
||||
|
||||
assert (
|
||||
hf_shape == value.shape
|
||||
), f"Shape of hf {key + '.' + weight_type if weight_type is not None else ''} is {hf_shape}, but should be {value.shape} for {full_name}"
|
||||
|
||||
if weight_type == "weight":
|
||||
hf_pointer.weight.data = value
|
||||
elif weight_type == "weight_g":
|
||||
hf_pointer.weight_g.data = value
|
||||
elif weight_type == "weight_v":
|
||||
hf_pointer.weight_v.data = value
|
||||
elif weight_type == "bias":
|
||||
hf_pointer.bias.data = value
|
||||
else:
|
||||
hf_pointer.data = value
|
||||
|
||||
logger.info(f"{key + '.' + weight_type if weight_type is not None else ''} was initialized from {full_name}.")
|
||||
|
||||
|
||||
def recursively_load_weights(fairseq_model, hf_model):
|
||||
unused_weights = []
|
||||
fairseq_dict = fairseq_model.state_dict()
|
||||
|
||||
feature_extractor = hf_model.unispeech_sat.feature_extractor
|
||||
|
||||
for name, value in fairseq_dict.items():
|
||||
is_used = False
|
||||
if "conv_layers" in name:
|
||||
load_conv_layer(
|
||||
name,
|
||||
value,
|
||||
feature_extractor,
|
||||
unused_weights,
|
||||
hf_model.config.feat_extract_norm == "group",
|
||||
)
|
||||
is_used = True
|
||||
else:
|
||||
for key, mapped_key in MAPPING.items():
|
||||
mapped_key = "unispeech_sat." + mapped_key if mapped_key not in TOP_LEVEL_KEYS else mapped_key
|
||||
if key in name or key.split("w2v_model.")[-1] == name.split(".")[0]:
|
||||
if "layer_norm_for_extract" in name and (".".join(name.split(".")[:-1]) != key):
|
||||
# special case since naming is very similar
|
||||
continue
|
||||
is_used = True
|
||||
if "*" in mapped_key:
|
||||
layer_index = name.split(key)[0].split(".")[-2]
|
||||
mapped_key = mapped_key.replace("*", layer_index)
|
||||
if "weight_g" in name:
|
||||
weight_type = "weight_g"
|
||||
elif "weight_v" in name:
|
||||
weight_type = "weight_v"
|
||||
elif "bias" in name:
|
||||
weight_type = "bias"
|
||||
elif "weight" in name:
|
||||
# TODO: don't match quantizer.weight_proj
|
||||
weight_type = "weight"
|
||||
else:
|
||||
weight_type = None
|
||||
set_recursively(hf_model, mapped_key, value, name, weight_type)
|
||||
continue
|
||||
if not is_used:
|
||||
unused_weights.append(name)
|
||||
|
||||
logger.warning(f"Unused weights: {unused_weights}")
|
||||
|
||||
|
||||
def 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:
|
||||
assert (
|
||||
value.shape == feature_extractor.conv_layers[layer_id].conv.bias.data.shape
|
||||
), f"{full_name} has size {value.shape}, but {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:
|
||||
assert (
|
||||
value.shape == feature_extractor.conv_layers[layer_id].conv.weight.data.shape
|
||||
), f"{full_name} has size {value.shape}, but {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:
|
||||
assert (
|
||||
value.shape == feature_extractor.conv_layers[layer_id].layer_norm.bias.data.shape
|
||||
), f"{full_name} has size {value.shape}, but {feature_extractor[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:
|
||||
assert (
|
||||
value.shape == feature_extractor.conv_layers[layer_id].layer_norm.weight.data.shape
|
||||
), f"{full_name} has size {value.shape}, but {feature_extractor[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_unispeech_sat_checkpoint(
|
||||
checkpoint_path, pytorch_dump_folder_path, config_path=None, dict_path=None, is_finetuned=True
|
||||
):
|
||||
"""
|
||||
Copy/paste/tweak model's weights to transformers design.
|
||||
"""
|
||||
if config_path is not None:
|
||||
config = UniSpeechSatConfig.from_pretrained(config_path)
|
||||
else:
|
||||
config = UniSpeechSatConfig()
|
||||
|
||||
dict_path = ""
|
||||
|
||||
if is_finetuned:
|
||||
hf_wav2vec = UniSpeechSatForCTC(config)
|
||||
else:
|
||||
hf_wav2vec = UniSpeechSatForPreTraining(config)
|
||||
|
||||
model, _, _ = fairseq.checkpoint_utils.load_model_ensemble_and_task(
|
||||
[checkpoint_path], arg_overrides={"data": "/".join(dict_path.split("/")[:-1])}
|
||||
)
|
||||
model = model[0].eval()
|
||||
|
||||
recursively_load_weights(model, hf_wav2vec)
|
||||
|
||||
hf_wav2vec.save_pretrained(pytorch_dump_folder_path)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
parser = argparse.ArgumentParser()
|
||||
parser.add_argument("--pytorch_dump_folder_path", default=None, type=str, help="Path to the output PyTorch model.")
|
||||
parser.add_argument("--checkpoint_path", default=None, type=str, help="Path to fairseq checkpoint")
|
||||
parser.add_argument("--dict_path", default=None, type=str, help="Path to dict of fine-tuned model")
|
||||
parser.add_argument("--config_path", default=None, type=str, help="Path to hf config.json of model to convert")
|
||||
parser.add_argument(
|
||||
"--not_finetuned", action="store_true", help="Whether the model to convert is a fine-tuned model or not"
|
||||
)
|
||||
args = parser.parse_args()
|
||||
convert_unispeech_sat_checkpoint(
|
||||
args.checkpoint_path, args.pytorch_dump_folder_path, args.config_path, args.dict_path, not args.not_finetuned
|
||||
)
|
||||
1511
src/transformers/models/unispeech_sat/modeling_unispeech_sat.py
Executable file
1511
src/transformers/models/unispeech_sat/modeling_unispeech_sat.py
Executable file
File diff suppressed because it is too large
Load Diff
@ -46,7 +46,7 @@ _CONFIG_FOR_DOC = "Wav2Vec2Config"
|
||||
_CHECKPOINT_FOR_DOC = "facebook/wav2vec2-base-960h"
|
||||
_PROCESSOR_FOR_DOC = "Wav2Vec2Processor"
|
||||
|
||||
_SEQ_CLASS_CHECKPOINT = ("superb/wav2vec2-base-superb-ks",)
|
||||
_SEQ_CLASS_CHECKPOINT = "superb/wav2vec2-base-superb-ks"
|
||||
_SEQ_CLASS_PROCESSOR_FOR_DOC = "Wav2Vec2FeatureExtractor"
|
||||
|
||||
_HIDDEN_STATES_START_POSITION = 2
|
||||
@ -462,9 +462,12 @@ class Wav2Vec2Attention(nn.Module):
|
||||
self.num_heads = num_heads
|
||||
self.dropout = dropout
|
||||
self.head_dim = embed_dim // num_heads
|
||||
assert (
|
||||
self.head_dim * num_heads == self.embed_dim
|
||||
), f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`: {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
|
||||
|
||||
@ -858,9 +861,11 @@ class Wav2Vec2GumbelVectorQuantizer(nn.Module):
|
||||
self.num_groups = config.num_codevector_groups
|
||||
self.num_vars = config.num_codevectors_per_group
|
||||
|
||||
assert (
|
||||
config.codevector_dim % self.num_groups == 0
|
||||
), f"`config.codevector_dim {config.codevector_dim} must be divisible by `config.num_codevector_groups` {self.num_groups} for concatenation"
|
||||
if config.codevector_dim % self.num_groups != 0:
|
||||
raise ValueError(
|
||||
f"`config.codevector_dim {config.codevector_dim} must be divisible "
|
||||
f"by `config.num_codevector_groups` {self.num_groups} for concatenation"
|
||||
)
|
||||
|
||||
# storage for codebook variables (codewords)
|
||||
self.codevectors = nn.Parameter(
|
||||
@ -871,9 +876,6 @@ class Wav2Vec2GumbelVectorQuantizer(nn.Module):
|
||||
# can be decayed for training
|
||||
self.temperature = 2
|
||||
|
||||
def set_temperature(self, temperature: int):
|
||||
self.temperature = temperature
|
||||
|
||||
@staticmethod
|
||||
def _compute_perplexity(probs, mask=None):
|
||||
if mask is not None:
|
||||
@ -1118,9 +1120,8 @@ class Wav2Vec2Model(Wav2Vec2PreTrainedModel):
|
||||
mask_prob=self.config.mask_feature_prob,
|
||||
mask_length=self.config.mask_feature_length,
|
||||
)
|
||||
mask_feature_indices = torch.tensor(mask_feature_indices, device=hidden_states.device, dtype=torch.bool)[
|
||||
:, None
|
||||
].expand(-1, sequence_length, -1)
|
||||
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
|
||||
@ -1200,7 +1201,7 @@ class Wav2Vec2ForPreTraining(Wav2Vec2PreTrainedModel):
|
||||
"""
|
||||
Set the Gumbel softmax temperature to a given value. Only necessary for training
|
||||
"""
|
||||
return self.quantizer.set_temperature(temperature)
|
||||
self.quantizer.temperature = temperature
|
||||
|
||||
def freeze_feature_extractor(self):
|
||||
"""
|
||||
|
||||
@ -3626,6 +3626,86 @@ class TrOCRPreTrainedModel:
|
||||
requires_backends(cls, ["torch"])
|
||||
|
||||
|
||||
UNISPEECH_PRETRAINED_MODEL_ARCHIVE_LIST = None
|
||||
|
||||
|
||||
class UniSpeechForCTC:
|
||||
def __init__(self, *args, **kwargs):
|
||||
requires_backends(self, ["torch"])
|
||||
|
||||
|
||||
class UniSpeechForPreTraining:
|
||||
def __init__(self, *args, **kwargs):
|
||||
requires_backends(self, ["torch"])
|
||||
|
||||
|
||||
class UniSpeechForSequenceClassification:
|
||||
def __init__(self, *args, **kwargs):
|
||||
requires_backends(self, ["torch"])
|
||||
|
||||
@classmethod
|
||||
def from_pretrained(cls, *args, **kwargs):
|
||||
requires_backends(cls, ["torch"])
|
||||
|
||||
|
||||
class UniSpeechModel:
|
||||
def __init__(self, *args, **kwargs):
|
||||
requires_backends(self, ["torch"])
|
||||
|
||||
@classmethod
|
||||
def from_pretrained(cls, *args, **kwargs):
|
||||
requires_backends(cls, ["torch"])
|
||||
|
||||
|
||||
class UniSpeechPreTrainedModel:
|
||||
def __init__(self, *args, **kwargs):
|
||||
requires_backends(self, ["torch"])
|
||||
|
||||
@classmethod
|
||||
def from_pretrained(cls, *args, **kwargs):
|
||||
requires_backends(cls, ["torch"])
|
||||
|
||||
|
||||
UNISPEECH_SAT_PRETRAINED_MODEL_ARCHIVE_LIST = None
|
||||
|
||||
|
||||
class UniSpeechSatForCTC:
|
||||
def __init__(self, *args, **kwargs):
|
||||
requires_backends(self, ["torch"])
|
||||
|
||||
|
||||
class UniSpeechSatForPreTraining:
|
||||
def __init__(self, *args, **kwargs):
|
||||
requires_backends(self, ["torch"])
|
||||
|
||||
|
||||
class UniSpeechSatForSequenceClassification:
|
||||
def __init__(self, *args, **kwargs):
|
||||
requires_backends(self, ["torch"])
|
||||
|
||||
@classmethod
|
||||
def from_pretrained(cls, *args, **kwargs):
|
||||
requires_backends(cls, ["torch"])
|
||||
|
||||
|
||||
class UniSpeechSatModel:
|
||||
def __init__(self, *args, **kwargs):
|
||||
requires_backends(self, ["torch"])
|
||||
|
||||
@classmethod
|
||||
def from_pretrained(cls, *args, **kwargs):
|
||||
requires_backends(cls, ["torch"])
|
||||
|
||||
|
||||
class UniSpeechSatPreTrainedModel:
|
||||
def __init__(self, *args, **kwargs):
|
||||
requires_backends(self, ["torch"])
|
||||
|
||||
@classmethod
|
||||
def from_pretrained(cls, *args, **kwargs):
|
||||
requires_backends(cls, ["torch"])
|
||||
|
||||
|
||||
class VisionEncoderDecoderModel:
|
||||
def __init__(self, *args, **kwargs):
|
||||
requires_backends(self, ["torch"])
|
||||
|
||||
@ -1,3 +1,3 @@
|
||||
{
|
||||
"feature_extractor_type": "Wav2Vec2FeatureExtractor"
|
||||
}
|
||||
}
|
||||
|
||||
584
tests/test_modeling_unispeech.py
Normal file
584
tests/test_modeling_unispeech.py
Normal file
@ -0,0 +1,584 @@
|
||||
# coding=utf-8
|
||||
# Copyright 2021 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 UniSpeech model. """
|
||||
|
||||
import math
|
||||
import unittest
|
||||
|
||||
import numpy as np
|
||||
import pytest
|
||||
from datasets import load_dataset
|
||||
|
||||
from tests.test_modeling_common import floats_tensor, ids_tensor, random_attention_mask
|
||||
from transformers import UniSpeechConfig, is_torch_available
|
||||
from transformers.testing_utils import require_datasets, require_soundfile, require_torch, slow, torch_device
|
||||
|
||||
from .test_configuration_common import ConfigTester
|
||||
from .test_modeling_common import ModelTesterMixin, _config_zero_init
|
||||
|
||||
|
||||
if is_torch_available():
|
||||
import torch
|
||||
|
||||
from transformers import (
|
||||
UniSpeechForCTC,
|
||||
UniSpeechForPreTraining,
|
||||
UniSpeechForSequenceClassification,
|
||||
UniSpeechModel,
|
||||
Wav2Vec2FeatureExtractor,
|
||||
Wav2Vec2Processor,
|
||||
)
|
||||
|
||||
|
||||
class UniSpeechModelTester:
|
||||
def __init__(
|
||||
self,
|
||||
parent,
|
||||
batch_size=13,
|
||||
seq_length=1024, # speech is longer
|
||||
is_training=False,
|
||||
hidden_size=16,
|
||||
feat_extract_norm="group",
|
||||
feat_extract_dropout=0.0,
|
||||
feat_extract_activation="gelu",
|
||||
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,
|
||||
num_hidden_layers=4,
|
||||
num_attention_heads=2,
|
||||
hidden_dropout_prob=0.1, # this is most likely not correctly set yet
|
||||
intermediate_size=20,
|
||||
layer_norm_eps=1e-5,
|
||||
hidden_act="gelu",
|
||||
initializer_range=0.02,
|
||||
vocab_size=32,
|
||||
do_stable_layer_norm=False,
|
||||
scope=None,
|
||||
):
|
||||
self.parent = parent
|
||||
self.batch_size = batch_size
|
||||
self.seq_length = seq_length
|
||||
self.is_training = is_training
|
||||
self.hidden_size = hidden_size
|
||||
self.feat_extract_norm = feat_extract_norm
|
||||
self.feat_extract_dropout = feat_extract_dropout
|
||||
self.feat_extract_activation = feat_extract_activation
|
||||
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.num_hidden_layers = num_hidden_layers
|
||||
self.num_attention_heads = num_attention_heads
|
||||
self.hidden_dropout_prob = hidden_dropout_prob
|
||||
self.intermediate_size = intermediate_size
|
||||
self.layer_norm_eps = layer_norm_eps
|
||||
self.hidden_act = hidden_act
|
||||
self.initializer_range = initializer_range
|
||||
self.vocab_size = vocab_size
|
||||
self.do_stable_layer_norm = do_stable_layer_norm
|
||||
self.scope = scope
|
||||
|
||||
output_seq_length = self.seq_length
|
||||
for kernel, stride in zip(self.conv_kernel, self.conv_stride):
|
||||
output_seq_length = (output_seq_length - (kernel - 1)) / stride
|
||||
self.output_seq_length = int(math.ceil(output_seq_length))
|
||||
self.encoder_seq_length = self.output_seq_length
|
||||
|
||||
def prepare_config_and_inputs(self):
|
||||
input_values = floats_tensor([self.batch_size, self.seq_length], self.vocab_size)
|
||||
attention_mask = random_attention_mask([self.batch_size, self.seq_length])
|
||||
|
||||
config = self.get_config()
|
||||
|
||||
return config, input_values, attention_mask
|
||||
|
||||
def get_config(self):
|
||||
return UniSpeechConfig(
|
||||
hidden_size=self.hidden_size,
|
||||
feat_extract_norm=self.feat_extract_norm,
|
||||
feat_extract_dropout=self.feat_extract_dropout,
|
||||
feat_extract_activation=self.feat_extract_activation,
|
||||
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,
|
||||
num_hidden_layers=self.num_hidden_layers,
|
||||
num_attention_heads=self.num_attention_heads,
|
||||
hidden_dropout_prob=self.hidden_dropout_prob,
|
||||
intermediate_size=self.intermediate_size,
|
||||
layer_norm_eps=self.layer_norm_eps,
|
||||
hidden_act=self.hidden_act,
|
||||
initializer_range=self.initializer_range,
|
||||
vocab_size=self.vocab_size,
|
||||
)
|
||||
|
||||
def create_and_check_model(self, config, input_values, attention_mask):
|
||||
model = UniSpeechModel(config=config)
|
||||
model.to(torch_device)
|
||||
model.eval()
|
||||
result = model(input_values, attention_mask=attention_mask)
|
||||
self.parent.assertEqual(
|
||||
result.last_hidden_state.shape, (self.batch_size, self.output_seq_length, self.hidden_size)
|
||||
)
|
||||
|
||||
def create_and_check_batch_inference(self, config, input_values, *args):
|
||||
# test does not pass for models making use of `group_norm`
|
||||
# check: https://github.com/pytorch/fairseq/issues/3227
|
||||
model = UniSpeechModel(config=config)
|
||||
model.to(torch_device)
|
||||
model.eval()
|
||||
|
||||
input_values = input_values[:3]
|
||||
attention_mask = torch.ones(input_values.shape, device=torch_device, dtype=torch.bool)
|
||||
|
||||
input_lengths = [input_values.shape[-1] // i for i in [4, 2, 1]]
|
||||
|
||||
# pad input
|
||||
for i in range(len(input_lengths)):
|
||||
input_values[i, input_lengths[i] :] = 0.0
|
||||
attention_mask[i, input_lengths[i] :] = 0.0
|
||||
|
||||
batch_outputs = model(input_values, attention_mask=attention_mask).last_hidden_state
|
||||
|
||||
for i in range(input_values.shape[0]):
|
||||
input_slice = input_values[i : i + 1, : input_lengths[i]]
|
||||
output = model(input_slice).last_hidden_state
|
||||
|
||||
batch_output = batch_outputs[i : i + 1, : output.shape[1]]
|
||||
self.parent.assertTrue(torch.allclose(output, batch_output, atol=1e-3))
|
||||
|
||||
def check_ctc_loss(self, config, input_values, *args):
|
||||
model = UniSpeechForCTC(config=config)
|
||||
model.to(torch_device)
|
||||
|
||||
# make sure that dropout is disabled
|
||||
model.eval()
|
||||
|
||||
input_values = input_values[:3]
|
||||
attention_mask = torch.ones(input_values.shape, device=torch_device, dtype=torch.long)
|
||||
|
||||
input_lengths = [input_values.shape[-1] // i for i in [4, 2, 1]]
|
||||
max_length_labels = model._get_feat_extract_output_lengths(torch.tensor(input_lengths))
|
||||
labels = ids_tensor((input_values.shape[0], min(max_length_labels) - 1), model.config.vocab_size)
|
||||
|
||||
# pad input
|
||||
for i in range(len(input_lengths)):
|
||||
input_values[i, input_lengths[i] :] = 0.0
|
||||
attention_mask[i, input_lengths[i] :] = 0
|
||||
|
||||
model.config.ctc_loss_reduction = "sum"
|
||||
sum_loss = model(input_values, attention_mask=attention_mask, labels=labels).loss.item()
|
||||
|
||||
model.config.ctc_loss_reduction = "mean"
|
||||
mean_loss = model(input_values, attention_mask=attention_mask, labels=labels).loss.item()
|
||||
|
||||
self.parent.assertTrue(isinstance(sum_loss, float))
|
||||
self.parent.assertTrue(isinstance(mean_loss, float))
|
||||
|
||||
def check_seq_classifier_loss(self, config, input_values, *args):
|
||||
model = UniSpeechForSequenceClassification(config=config)
|
||||
model.to(torch_device)
|
||||
|
||||
# make sure that dropout is disabled
|
||||
model.eval()
|
||||
|
||||
input_values = input_values[:3]
|
||||
attention_mask = torch.ones(input_values.shape, device=torch_device, dtype=torch.long)
|
||||
|
||||
input_lengths = [input_values.shape[-1] // i for i in [4, 2, 1]]
|
||||
labels = ids_tensor((input_values.shape[0], 1), len(model.config.id2label))
|
||||
|
||||
# pad input
|
||||
for i in range(len(input_lengths)):
|
||||
input_values[i, input_lengths[i] :] = 0.0
|
||||
attention_mask[i, input_lengths[i] :] = 0
|
||||
|
||||
masked_loss = model(input_values, attention_mask=attention_mask, labels=labels).loss.item()
|
||||
unmasked_loss = model(input_values, labels=labels).loss.item()
|
||||
|
||||
self.parent.assertTrue(isinstance(masked_loss, float))
|
||||
self.parent.assertTrue(isinstance(unmasked_loss, float))
|
||||
self.parent.assertTrue(masked_loss != unmasked_loss)
|
||||
|
||||
def check_ctc_training(self, config, input_values, *args):
|
||||
config.ctc_zero_infinity = True
|
||||
model = UniSpeechForCTC(config=config)
|
||||
model.to(torch_device)
|
||||
model.train()
|
||||
|
||||
# freeze feature encoder
|
||||
model.freeze_feature_extractor()
|
||||
|
||||
input_values = input_values[:3]
|
||||
|
||||
input_lengths = [input_values.shape[-1] // i for i in [4, 2, 1]]
|
||||
max_length_labels = model._get_feat_extract_output_lengths(torch.tensor(input_lengths))
|
||||
labels = ids_tensor((input_values.shape[0], max(max_length_labels) - 2), model.config.vocab_size)
|
||||
|
||||
# pad input
|
||||
for i in range(len(input_lengths)):
|
||||
input_values[i, input_lengths[i] :] = 0.0
|
||||
|
||||
if max_length_labels[i] < labels.shape[-1]:
|
||||
# it's important that we make sure that target lenghts are at least
|
||||
# one shorter than logit lenghts to prevent -inf
|
||||
labels[i, max_length_labels[i] - 1 :] = -100
|
||||
|
||||
loss = model(input_values, labels=labels).loss
|
||||
self.parent.assertFalse(torch.isinf(loss).item())
|
||||
|
||||
loss.backward()
|
||||
|
||||
def check_seq_classifier_training(self, config, input_values, *args):
|
||||
config.ctc_zero_infinity = True
|
||||
model = UniSpeechForSequenceClassification(config=config)
|
||||
model.to(torch_device)
|
||||
model.train()
|
||||
|
||||
# freeze everything but the classification head
|
||||
model.freeze_base_model()
|
||||
|
||||
input_values = input_values[:3]
|
||||
|
||||
input_lengths = [input_values.shape[-1] // i for i in [4, 2, 1]]
|
||||
labels = ids_tensor((input_values.shape[0], 1), len(model.config.id2label))
|
||||
|
||||
# pad input
|
||||
for i in range(len(input_lengths)):
|
||||
input_values[i, input_lengths[i] :] = 0.0
|
||||
|
||||
loss = model(input_values, labels=labels).loss
|
||||
self.parent.assertFalse(torch.isinf(loss).item())
|
||||
|
||||
loss.backward()
|
||||
|
||||
def check_labels_out_of_vocab(self, config, input_values, *args):
|
||||
model = UniSpeechForCTC(config)
|
||||
model.to(torch_device)
|
||||
model.train()
|
||||
|
||||
input_values = input_values[:3]
|
||||
|
||||
input_lengths = [input_values.shape[-1] // i for i in [4, 2, 1]]
|
||||
max_length_labels = model._get_feat_extract_output_lengths(torch.tensor(input_lengths))
|
||||
labels = ids_tensor((input_values.shape[0], max(max_length_labels) - 2), model.config.vocab_size + 100)
|
||||
|
||||
with pytest.raises(ValueError):
|
||||
model(input_values, labels=labels)
|
||||
|
||||
def prepare_config_and_inputs_for_common(self):
|
||||
config, input_values, attention_mask = self.prepare_config_and_inputs()
|
||||
inputs_dict = {"input_values": input_values, "attention_mask": attention_mask}
|
||||
return config, inputs_dict
|
||||
|
||||
|
||||
@require_torch
|
||||
class UniSpeechRobustModelTest(ModelTesterMixin, unittest.TestCase):
|
||||
all_model_classes = (
|
||||
(UniSpeechForCTC, UniSpeechModel, UniSpeechForSequenceClassification, UniSpeechForPreTraining)
|
||||
if is_torch_available()
|
||||
else ()
|
||||
)
|
||||
test_pruning = False
|
||||
test_headmasking = False
|
||||
test_torchscript = False
|
||||
|
||||
def setUp(self):
|
||||
self.model_tester = UniSpeechModelTester(
|
||||
self, conv_stride=(3, 3, 3), feat_extract_norm="layer", do_stable_layer_norm=True
|
||||
)
|
||||
self.config_tester = ConfigTester(self, config_class=UniSpeechConfig, hidden_size=37)
|
||||
|
||||
def test_config(self):
|
||||
self.config_tester.run_common_tests()
|
||||
|
||||
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_batched_inference(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
||||
self.model_tester.create_and_check_batch_inference(*config_and_inputs)
|
||||
|
||||
def test_ctc_loss_inference(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
||||
self.model_tester.check_ctc_loss(*config_and_inputs)
|
||||
|
||||
def test_seq_classifier_loss_inference(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
||||
self.model_tester.check_seq_classifier_loss(*config_and_inputs)
|
||||
|
||||
def test_ctc_train(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
||||
self.model_tester.check_ctc_training(*config_and_inputs)
|
||||
|
||||
def test_seq_classifier_train(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
||||
self.model_tester.check_seq_classifier_training(*config_and_inputs)
|
||||
|
||||
def test_labels_out_of_vocab(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
||||
self.model_tester.check_labels_out_of_vocab(*config_and_inputs)
|
||||
|
||||
# UniSpeech has no inputs_embeds
|
||||
def test_inputs_embeds(self):
|
||||
pass
|
||||
|
||||
# `input_ids` is renamed to `input_values`
|
||||
def test_forward_signature(self):
|
||||
pass
|
||||
|
||||
# UniSpeech cannot resize token embeddings
|
||||
# since it has no tokens embeddings
|
||||
def test_resize_tokens_embeddings(self):
|
||||
pass
|
||||
|
||||
# UniSpeech has no inputs_embeds
|
||||
# and thus the `get_input_embeddings` fn
|
||||
# is not implemented
|
||||
def test_model_common_attributes(self):
|
||||
pass
|
||||
|
||||
def test_retain_grad_hidden_states_attentions(self):
|
||||
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
config.output_hidden_states = True
|
||||
config.output_attentions = True
|
||||
|
||||
# no need to test all models as different heads yield the same functionality
|
||||
model_class = self.all_model_classes[0]
|
||||
model = model_class(config)
|
||||
model.to(torch_device)
|
||||
|
||||
# set layer drop to 0
|
||||
model.config.layerdrop = 0.0
|
||||
|
||||
input_values = inputs_dict["input_values"]
|
||||
|
||||
input_lengths = torch.tensor(
|
||||
[input_values.shape[1] for _ in range(input_values.shape[0])], dtype=torch.long, device=torch_device
|
||||
)
|
||||
output_lengths = model._get_feat_extract_output_lengths(input_lengths)
|
||||
|
||||
labels = ids_tensor((input_values.shape[0], output_lengths[0] - 2), self.model_tester.vocab_size)
|
||||
inputs_dict["attention_mask"] = torch.ones_like(inputs_dict["attention_mask"])
|
||||
inputs_dict["labels"] = labels
|
||||
|
||||
outputs = model(**inputs_dict)
|
||||
|
||||
output = outputs[0]
|
||||
|
||||
# Encoder-/Decoder-only models
|
||||
hidden_states = outputs.hidden_states[0]
|
||||
attentions = outputs.attentions[0]
|
||||
|
||||
hidden_states.retain_grad()
|
||||
attentions.retain_grad()
|
||||
|
||||
output.flatten()[0].backward(retain_graph=True)
|
||||
|
||||
self.assertIsNotNone(hidden_states.grad)
|
||||
self.assertIsNotNone(attentions.grad)
|
||||
|
||||
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",
|
||||
"codevectors",
|
||||
"quantizer.weight_proj.weight",
|
||||
"project_hid.weight",
|
||||
"project_hid.bias",
|
||||
"project_q.weight",
|
||||
"project_q.bias",
|
||||
"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",
|
||||
)
|
||||
|
||||
# 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, "codevectors") and module.codevectors is not None:
|
||||
module.codevectors.data.fill_(3)
|
||||
if hasattr(module, "masked_spec_embed") and module.masked_spec_embed is not None:
|
||||
module.masked_spec_embed.data.fill_(3)
|
||||
|
||||
def test_mask_feature_prob_ctc(self):
|
||||
model = UniSpeechForCTC.from_pretrained(
|
||||
"hf-internal-testing/tiny-random-unispeech", mask_feature_prob=0.2, mask_feature_length=2
|
||||
)
|
||||
model.to(torch_device).train()
|
||||
processor = Wav2Vec2Processor.from_pretrained(
|
||||
"hf-internal-testing/tiny-random-unispeech", return_attention_mask=True
|
||||
)
|
||||
|
||||
batch_duration_in_seconds = [1, 3, 2, 6]
|
||||
input_features = [np.random.random(16_000 * s) for s in batch_duration_in_seconds]
|
||||
|
||||
batch = processor(
|
||||
input_features, padding=True, sampling_rate=processor.feature_extractor.sampling_rate, return_tensors="pt"
|
||||
)
|
||||
|
||||
logits = model(
|
||||
input_values=batch["input_values"].to(torch_device),
|
||||
attention_mask=batch["attention_mask"].to(torch_device),
|
||||
).logits
|
||||
|
||||
self.assertEqual(logits.shape, (4, 1498, 32))
|
||||
|
||||
def test_mask_time_prob_ctc(self):
|
||||
model = UniSpeechForCTC.from_pretrained(
|
||||
"hf-internal-testing/tiny-random-unispeech", mask_time_prob=0.2, mask_time_length=2
|
||||
)
|
||||
model.to(torch_device).train()
|
||||
processor = Wav2Vec2Processor.from_pretrained(
|
||||
"hf-internal-testing/tiny-random-unispeech", return_attention_mask=True
|
||||
)
|
||||
|
||||
batch_duration_in_seconds = [1, 3, 2, 6]
|
||||
input_features = [np.random.random(16_000 * s) for s in batch_duration_in_seconds]
|
||||
|
||||
batch = processor(
|
||||
input_features, padding=True, sampling_rate=processor.feature_extractor.sampling_rate, return_tensors="pt"
|
||||
)
|
||||
|
||||
logits = model(
|
||||
input_values=batch["input_values"].to(torch_device),
|
||||
attention_mask=batch["attention_mask"].to(torch_device),
|
||||
).logits
|
||||
|
||||
self.assertEqual(logits.shape, (4, 1498, 32))
|
||||
|
||||
def test_mask_time_feature_prob_ctc_single_batch(self):
|
||||
model = UniSpeechForCTC.from_pretrained(
|
||||
"hf-internal-testing/tiny-random-unispeech",
|
||||
mask_time_prob=0.2,
|
||||
mask_feature_prob=0.2,
|
||||
mask_time_length=2,
|
||||
mask_feature_length=2,
|
||||
)
|
||||
model.to(torch_device).train()
|
||||
processor = Wav2Vec2Processor.from_pretrained(
|
||||
"hf-internal-testing/tiny-random-unispeech", return_attention_mask=True
|
||||
)
|
||||
|
||||
batch_duration_in_seconds = [6]
|
||||
input_features = [np.random.random(16_000 * s) for s in batch_duration_in_seconds]
|
||||
|
||||
batch = processor(
|
||||
input_features, padding=True, sampling_rate=processor.feature_extractor.sampling_rate, return_tensors="pt"
|
||||
)
|
||||
|
||||
logits = model(
|
||||
input_values=batch["input_values"].to(torch_device),
|
||||
attention_mask=batch["attention_mask"].to(torch_device),
|
||||
).logits
|
||||
|
||||
self.assertEqual(logits.shape, (1, 1498, 32))
|
||||
|
||||
@slow
|
||||
def test_model_from_pretrained(self):
|
||||
model = UniSpeechModel.from_pretrained("microsoft/unispeech-large-1500h-cv")
|
||||
self.assertIsNotNone(model)
|
||||
|
||||
|
||||
@require_torch
|
||||
@require_datasets
|
||||
@require_soundfile
|
||||
@slow
|
||||
class UniSpeechModelIntegrationTest(unittest.TestCase):
|
||||
def _load_datasamples(self, num_samples):
|
||||
import soundfile as sf
|
||||
|
||||
ids = [f"1272-141231-000{i}" for i in range(num_samples)]
|
||||
|
||||
# map files to raw
|
||||
def map_to_array(batch):
|
||||
speech, _ = sf.read(batch["file"])
|
||||
batch["speech"] = speech
|
||||
return batch
|
||||
|
||||
ds = load_dataset("patrickvonplaten/librispeech_asr_dummy", "clean", split="validation")
|
||||
|
||||
ds = ds.filter(lambda x: x["id"] in ids).sort("id").map(map_to_array)
|
||||
|
||||
return ds["speech"][:num_samples]
|
||||
|
||||
def _load_superb(self, task, num_samples):
|
||||
|
||||
ds = load_dataset("anton-l/superb_dummy", task, split="test")
|
||||
|
||||
return ds[:num_samples]
|
||||
|
||||
def test_inference_pretraining(self):
|
||||
model = UniSpeechForPreTraining.from_pretrained("microsoft/unispeech-large-1500h-cv")
|
||||
model.to(torch_device)
|
||||
feature_extractor = Wav2Vec2FeatureExtractor.from_pretrained("facebook/wav2vec2-large-xlsr-53")
|
||||
input_speech = self._load_datasamples(2)
|
||||
|
||||
inputs_dict = feature_extractor(input_speech, return_tensors="pt", padding=True)
|
||||
|
||||
with torch.no_grad():
|
||||
torch.manual_seed(0)
|
||||
outputs = model(
|
||||
inputs_dict.input_values.to(torch_device),
|
||||
attention_mask=inputs_dict.attention_mask.to(torch_device),
|
||||
)
|
||||
|
||||
# compute cosine similarity
|
||||
cosine_sim = torch.cosine_similarity(outputs.projected_states, outputs.projected_quantized_states, dim=-1)
|
||||
|
||||
# pretrained model should have learned a high cosine similarity
|
||||
self.assertTrue(cosine_sim.mean() > 0.5)
|
||||
|
||||
# fmt: off
|
||||
expected_cosine_sim_slice = torch.tensor(
|
||||
[[0.8290, 0.8335, 0.8815, 0.8580, 0.8249],
|
||||
[0.8892, 0.9221, 0.8711, 0.8601, 0.8482]],
|
||||
device=torch_device,
|
||||
)
|
||||
# fmt: on
|
||||
|
||||
self.assertTrue(torch.allclose(cosine_sim[:, :5], expected_cosine_sim_slice, atol=1e-3))
|
||||
800
tests/test_modeling_unispeech_sat.py
Normal file
800
tests/test_modeling_unispeech_sat.py
Normal file
@ -0,0 +1,800 @@
|
||||
# coding=utf-8
|
||||
# Copyright 2021 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 UniSpeechSat model. """
|
||||
|
||||
import math
|
||||
import unittest
|
||||
|
||||
import numpy as np
|
||||
import pytest
|
||||
from datasets import load_dataset
|
||||
|
||||
from tests.test_modeling_common import floats_tensor, ids_tensor, random_attention_mask
|
||||
from transformers import UniSpeechSatConfig, is_torch_available
|
||||
from transformers.testing_utils import require_datasets, require_soundfile, require_torch, slow, torch_device
|
||||
|
||||
from .test_configuration_common import ConfigTester
|
||||
from .test_modeling_common import ModelTesterMixin, _config_zero_init
|
||||
|
||||
|
||||
if is_torch_available():
|
||||
import torch
|
||||
|
||||
from transformers import (
|
||||
UniSpeechSatForCTC,
|
||||
UniSpeechSatForSequenceClassification,
|
||||
UniSpeechSatModel,
|
||||
Wav2Vec2FeatureExtractor,
|
||||
Wav2Vec2Processor,
|
||||
)
|
||||
|
||||
|
||||
class UniSpeechSatModelTester:
|
||||
def __init__(
|
||||
self,
|
||||
parent,
|
||||
batch_size=13,
|
||||
seq_length=1024, # speech is longer
|
||||
is_training=False,
|
||||
hidden_size=16,
|
||||
feat_extract_norm="group",
|
||||
feat_extract_dropout=0.0,
|
||||
feat_extract_activation="gelu",
|
||||
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,
|
||||
num_hidden_layers=4,
|
||||
num_attention_heads=2,
|
||||
hidden_dropout_prob=0.1, # this is most likely not correctly set yet
|
||||
intermediate_size=20,
|
||||
layer_norm_eps=1e-5,
|
||||
hidden_act="gelu",
|
||||
initializer_range=0.02,
|
||||
vocab_size=32,
|
||||
do_stable_layer_norm=False,
|
||||
scope=None,
|
||||
):
|
||||
self.parent = parent
|
||||
self.batch_size = batch_size
|
||||
self.seq_length = seq_length
|
||||
self.is_training = is_training
|
||||
self.hidden_size = hidden_size
|
||||
self.feat_extract_norm = feat_extract_norm
|
||||
self.feat_extract_dropout = feat_extract_dropout
|
||||
self.feat_extract_activation = feat_extract_activation
|
||||
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.num_hidden_layers = num_hidden_layers
|
||||
self.num_attention_heads = num_attention_heads
|
||||
self.hidden_dropout_prob = hidden_dropout_prob
|
||||
self.intermediate_size = intermediate_size
|
||||
self.layer_norm_eps = layer_norm_eps
|
||||
self.hidden_act = hidden_act
|
||||
self.initializer_range = initializer_range
|
||||
self.vocab_size = vocab_size
|
||||
self.do_stable_layer_norm = do_stable_layer_norm
|
||||
self.scope = scope
|
||||
|
||||
output_seq_length = self.seq_length
|
||||
for kernel, stride in zip(self.conv_kernel, self.conv_stride):
|
||||
output_seq_length = (output_seq_length - (kernel - 1)) / stride
|
||||
self.output_seq_length = int(math.ceil(output_seq_length))
|
||||
self.encoder_seq_length = self.output_seq_length
|
||||
|
||||
def prepare_config_and_inputs(self):
|
||||
input_values = floats_tensor([self.batch_size, self.seq_length], self.vocab_size)
|
||||
attention_mask = random_attention_mask([self.batch_size, self.seq_length])
|
||||
|
||||
config = self.get_config()
|
||||
|
||||
return config, input_values, attention_mask
|
||||
|
||||
def get_config(self):
|
||||
return UniSpeechSatConfig(
|
||||
hidden_size=self.hidden_size,
|
||||
feat_extract_norm=self.feat_extract_norm,
|
||||
feat_extract_dropout=self.feat_extract_dropout,
|
||||
feat_extract_activation=self.feat_extract_activation,
|
||||
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,
|
||||
num_hidden_layers=self.num_hidden_layers,
|
||||
num_attention_heads=self.num_attention_heads,
|
||||
hidden_dropout_prob=self.hidden_dropout_prob,
|
||||
intermediate_size=self.intermediate_size,
|
||||
layer_norm_eps=self.layer_norm_eps,
|
||||
hidden_act=self.hidden_act,
|
||||
initializer_range=self.initializer_range,
|
||||
vocab_size=self.vocab_size,
|
||||
)
|
||||
|
||||
def create_and_check_model(self, config, input_values, attention_mask):
|
||||
model = UniSpeechSatModel(config=config)
|
||||
model.to(torch_device)
|
||||
model.eval()
|
||||
result = model(input_values, attention_mask=attention_mask)
|
||||
self.parent.assertEqual(
|
||||
result.last_hidden_state.shape, (self.batch_size, self.output_seq_length, self.hidden_size)
|
||||
)
|
||||
|
||||
def create_and_check_batch_inference(self, config, input_values, *args):
|
||||
# test does not pass for models making use of `group_norm`
|
||||
# check: https://github.com/pytorch/fairseq/issues/3227
|
||||
model = UniSpeechSatModel(config=config)
|
||||
model.to(torch_device)
|
||||
model.eval()
|
||||
|
||||
input_values = input_values[:3]
|
||||
attention_mask = torch.ones(input_values.shape, device=torch_device, dtype=torch.bool)
|
||||
|
||||
input_lengths = [input_values.shape[-1] // i for i in [4, 2, 1]]
|
||||
|
||||
# pad input
|
||||
for i in range(len(input_lengths)):
|
||||
input_values[i, input_lengths[i] :] = 0.0
|
||||
attention_mask[i, input_lengths[i] :] = 0.0
|
||||
|
||||
batch_outputs = model(input_values, attention_mask=attention_mask).last_hidden_state
|
||||
|
||||
for i in range(input_values.shape[0]):
|
||||
input_slice = input_values[i : i + 1, : input_lengths[i]]
|
||||
output = model(input_slice).last_hidden_state
|
||||
|
||||
batch_output = batch_outputs[i : i + 1, : output.shape[1]]
|
||||
self.parent.assertTrue(torch.allclose(output, batch_output, atol=1e-3))
|
||||
|
||||
def check_ctc_loss(self, config, input_values, *args):
|
||||
model = UniSpeechSatForCTC(config=config)
|
||||
model.to(torch_device)
|
||||
|
||||
# make sure that dropout is disabled
|
||||
model.eval()
|
||||
|
||||
input_values = input_values[:3]
|
||||
attention_mask = torch.ones(input_values.shape, device=torch_device, dtype=torch.long)
|
||||
|
||||
input_lengths = [input_values.shape[-1] // i for i in [4, 2, 1]]
|
||||
max_length_labels = model._get_feat_extract_output_lengths(torch.tensor(input_lengths))
|
||||
labels = ids_tensor((input_values.shape[0], min(max_length_labels) - 1), model.config.vocab_size)
|
||||
|
||||
# pad input
|
||||
for i in range(len(input_lengths)):
|
||||
input_values[i, input_lengths[i] :] = 0.0
|
||||
attention_mask[i, input_lengths[i] :] = 0
|
||||
|
||||
model.config.ctc_loss_reduction = "sum"
|
||||
sum_loss = model(input_values, attention_mask=attention_mask, labels=labels).loss.item()
|
||||
|
||||
model.config.ctc_loss_reduction = "mean"
|
||||
mean_loss = model(input_values, attention_mask=attention_mask, labels=labels).loss.item()
|
||||
|
||||
self.parent.assertTrue(isinstance(sum_loss, float))
|
||||
self.parent.assertTrue(isinstance(mean_loss, float))
|
||||
|
||||
def check_seq_classifier_loss(self, config, input_values, *args):
|
||||
model = UniSpeechSatForSequenceClassification(config=config)
|
||||
model.to(torch_device)
|
||||
|
||||
# make sure that dropout is disabled
|
||||
model.eval()
|
||||
|
||||
input_values = input_values[:3]
|
||||
attention_mask = torch.ones(input_values.shape, device=torch_device, dtype=torch.long)
|
||||
|
||||
input_lengths = [input_values.shape[-1] // i for i in [4, 2, 1]]
|
||||
labels = ids_tensor((input_values.shape[0], 1), len(model.config.id2label))
|
||||
|
||||
# pad input
|
||||
for i in range(len(input_lengths)):
|
||||
input_values[i, input_lengths[i] :] = 0.0
|
||||
attention_mask[i, input_lengths[i] :] = 0
|
||||
|
||||
masked_loss = model(input_values, attention_mask=attention_mask, labels=labels).loss.item()
|
||||
unmasked_loss = model(input_values, labels=labels).loss.item()
|
||||
|
||||
self.parent.assertTrue(isinstance(masked_loss, float))
|
||||
self.parent.assertTrue(isinstance(unmasked_loss, float))
|
||||
self.parent.assertTrue(masked_loss != unmasked_loss)
|
||||
|
||||
def check_ctc_training(self, config, input_values, *args):
|
||||
config.ctc_zero_infinity = True
|
||||
model = UniSpeechSatForCTC(config=config)
|
||||
model.to(torch_device)
|
||||
model.train()
|
||||
|
||||
# freeze feature encoder
|
||||
model.freeze_feature_extractor()
|
||||
|
||||
input_values = input_values[:3]
|
||||
|
||||
input_lengths = [input_values.shape[-1] // i for i in [4, 2, 1]]
|
||||
max_length_labels = model._get_feat_extract_output_lengths(torch.tensor(input_lengths))
|
||||
labels = ids_tensor((input_values.shape[0], max(max_length_labels) - 2), model.config.vocab_size)
|
||||
|
||||
# pad input
|
||||
for i in range(len(input_lengths)):
|
||||
input_values[i, input_lengths[i] :] = 0.0
|
||||
|
||||
if max_length_labels[i] < labels.shape[-1]:
|
||||
# it's important that we make sure that target lenghts are at least
|
||||
# one shorter than logit lenghts to prevent -inf
|
||||
labels[i, max_length_labels[i] - 1 :] = -100
|
||||
|
||||
loss = model(input_values, labels=labels).loss
|
||||
self.parent.assertFalse(torch.isinf(loss).item())
|
||||
|
||||
loss.backward()
|
||||
|
||||
def check_seq_classifier_training(self, config, input_values, *args):
|
||||
config.ctc_zero_infinity = True
|
||||
model = UniSpeechSatForSequenceClassification(config=config)
|
||||
model.to(torch_device)
|
||||
model.train()
|
||||
|
||||
# freeze everything but the classification head
|
||||
model.freeze_base_model()
|
||||
|
||||
input_values = input_values[:3]
|
||||
|
||||
input_lengths = [input_values.shape[-1] // i for i in [4, 2, 1]]
|
||||
labels = ids_tensor((input_values.shape[0], 1), len(model.config.id2label))
|
||||
|
||||
# pad input
|
||||
for i in range(len(input_lengths)):
|
||||
input_values[i, input_lengths[i] :] = 0.0
|
||||
|
||||
loss = model(input_values, labels=labels).loss
|
||||
self.parent.assertFalse(torch.isinf(loss).item())
|
||||
|
||||
loss.backward()
|
||||
|
||||
def check_labels_out_of_vocab(self, config, input_values, *args):
|
||||
model = UniSpeechSatForCTC(config)
|
||||
model.to(torch_device)
|
||||
model.train()
|
||||
|
||||
input_values = input_values[:3]
|
||||
|
||||
input_lengths = [input_values.shape[-1] // i for i in [4, 2, 1]]
|
||||
max_length_labels = model._get_feat_extract_output_lengths(torch.tensor(input_lengths))
|
||||
labels = ids_tensor((input_values.shape[0], max(max_length_labels) - 2), model.config.vocab_size + 100)
|
||||
|
||||
with pytest.raises(ValueError):
|
||||
model(input_values, labels=labels)
|
||||
|
||||
def prepare_config_and_inputs_for_common(self):
|
||||
config, input_values, attention_mask = self.prepare_config_and_inputs()
|
||||
inputs_dict = {"input_values": input_values, "attention_mask": attention_mask}
|
||||
return config, inputs_dict
|
||||
|
||||
|
||||
@require_torch
|
||||
class UniSpeechSatModelTest(ModelTesterMixin, unittest.TestCase):
|
||||
all_model_classes = (
|
||||
(UniSpeechSatForCTC, UniSpeechSatModel, UniSpeechSatForSequenceClassification) if is_torch_available() else ()
|
||||
)
|
||||
test_pruning = False
|
||||
test_headmasking = False
|
||||
test_torchscript = False
|
||||
|
||||
def setUp(self):
|
||||
self.model_tester = UniSpeechSatModelTester(self)
|
||||
self.config_tester = ConfigTester(self, config_class=UniSpeechSatConfig, hidden_size=37)
|
||||
|
||||
def test_config(self):
|
||||
self.config_tester.run_common_tests()
|
||||
|
||||
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_ctc_loss_inference(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
||||
self.model_tester.check_ctc_loss(*config_and_inputs)
|
||||
|
||||
def test_seq_classifier_loss_inference(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
||||
self.model_tester.check_seq_classifier_loss(*config_and_inputs)
|
||||
|
||||
def test_ctc_train(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
||||
self.model_tester.check_ctc_training(*config_and_inputs)
|
||||
|
||||
def test_seq_classifier_train(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
||||
self.model_tester.check_seq_classifier_training(*config_and_inputs)
|
||||
|
||||
def test_labels_out_of_vocab(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
||||
self.model_tester.check_labels_out_of_vocab(*config_and_inputs)
|
||||
|
||||
# UniSpeechSat has no inputs_embeds
|
||||
def test_inputs_embeds(self):
|
||||
pass
|
||||
|
||||
# `input_ids` is renamed to `input_values`
|
||||
def test_forward_signature(self):
|
||||
pass
|
||||
|
||||
# UniSpeechSat cannot resize token embeddings
|
||||
# since it has no tokens embeddings
|
||||
def test_resize_tokens_embeddings(self):
|
||||
pass
|
||||
|
||||
# UniSpeechSat has no inputs_embeds
|
||||
# and thus the `get_input_embeddings` fn
|
||||
# is not implemented
|
||||
def test_model_common_attributes(self):
|
||||
pass
|
||||
|
||||
def test_retain_grad_hidden_states_attentions(self):
|
||||
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
config.output_hidden_states = True
|
||||
config.output_attentions = True
|
||||
|
||||
# no need to test all models as different heads yield the same functionality
|
||||
model_class = self.all_model_classes[0]
|
||||
model = model_class(config)
|
||||
model.to(torch_device)
|
||||
|
||||
# set layer drop to 0
|
||||
model.config.layerdrop = 0.0
|
||||
|
||||
input_values = inputs_dict["input_values"]
|
||||
|
||||
input_lengths = torch.tensor(
|
||||
[input_values.shape[1] for _ in range(input_values.shape[0])], dtype=torch.long, device=torch_device
|
||||
)
|
||||
output_lengths = model._get_feat_extract_output_lengths(input_lengths)
|
||||
|
||||
labels = ids_tensor((input_values.shape[0], output_lengths[0] - 2), self.model_tester.vocab_size)
|
||||
inputs_dict["attention_mask"] = torch.ones_like(inputs_dict["attention_mask"])
|
||||
inputs_dict["labels"] = labels
|
||||
|
||||
outputs = model(**inputs_dict)
|
||||
|
||||
output = outputs[0]
|
||||
|
||||
# Encoder-/Decoder-only models
|
||||
hidden_states = outputs.hidden_states[0]
|
||||
attentions = outputs.attentions[0]
|
||||
|
||||
hidden_states.retain_grad()
|
||||
attentions.retain_grad()
|
||||
|
||||
output.flatten()[0].backward(retain_graph=True)
|
||||
|
||||
self.assertIsNotNone(hidden_states.grad)
|
||||
self.assertIsNotNone(attentions.grad)
|
||||
|
||||
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",
|
||||
"codevectors",
|
||||
"quantizer.weight_proj.weight",
|
||||
"project_hid.weight",
|
||||
"project_hid.bias",
|
||||
"project_q.weight",
|
||||
"project_q.bias",
|
||||
"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",
|
||||
)
|
||||
|
||||
# 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, "codevectors") and module.codevectors is not None:
|
||||
module.codevectors.data.fill_(3)
|
||||
if hasattr(module, "masked_spec_embed") and module.masked_spec_embed is not None:
|
||||
module.masked_spec_embed.data.fill_(3)
|
||||
|
||||
def test_mask_feature_prob_ctc(self):
|
||||
model = UniSpeechSatForCTC.from_pretrained(
|
||||
"hf-internal-testing/tiny-random-unispeech-sat", mask_feature_prob=0.2, mask_feature_length=2
|
||||
)
|
||||
model.to(torch_device).train()
|
||||
processor = Wav2Vec2Processor.from_pretrained(
|
||||
"hf-internal-testing/tiny-random-unispeech-sat", return_attention_mask=True
|
||||
)
|
||||
|
||||
batch_duration_in_seconds = [1, 3, 2, 6]
|
||||
input_features = [np.random.random(16_000 * s) for s in batch_duration_in_seconds]
|
||||
|
||||
batch = processor(
|
||||
input_features, padding=True, sampling_rate=processor.feature_extractor.sampling_rate, return_tensors="pt"
|
||||
)
|
||||
logits = model(
|
||||
input_values=batch["input_values"].to(torch_device),
|
||||
attention_mask=batch["attention_mask"].to(torch_device),
|
||||
).logits
|
||||
|
||||
self.assertEqual(logits.shape, (4, 1498, 32))
|
||||
|
||||
def test_mask_time_prob_ctc(self):
|
||||
model = UniSpeechSatForCTC.from_pretrained(
|
||||
"hf-internal-testing/tiny-random-unispeech-sat", mask_time_prob=0.2, mask_time_length=2
|
||||
)
|
||||
model.to(torch_device).train()
|
||||
processor = Wav2Vec2Processor.from_pretrained(
|
||||
"hf-internal-testing/tiny-random-unispeech-sat", return_attention_mask=True
|
||||
)
|
||||
|
||||
batch_duration_in_seconds = [1, 3, 2, 6]
|
||||
input_features = [np.random.random(16_000 * s) for s in batch_duration_in_seconds]
|
||||
|
||||
batch = processor(
|
||||
input_features, padding=True, sampling_rate=processor.feature_extractor.sampling_rate, return_tensors="pt"
|
||||
)
|
||||
|
||||
logits = model(
|
||||
input_values=batch["input_values"].to(torch_device),
|
||||
attention_mask=batch["attention_mask"].to(torch_device),
|
||||
).logits
|
||||
|
||||
self.assertEqual(logits.shape, (4, 1498, 32))
|
||||
|
||||
@slow
|
||||
def test_model_from_pretrained(self):
|
||||
model = UniSpeechSatModel.from_pretrained("microsoft/unispeech-sat-base-plus")
|
||||
self.assertIsNotNone(model)
|
||||
|
||||
|
||||
@require_torch
|
||||
class UniSpeechSatRobustModelTest(ModelTesterMixin, unittest.TestCase):
|
||||
all_model_classes = (
|
||||
(UniSpeechSatForCTC, UniSpeechSatModel, UniSpeechSatForSequenceClassification) if is_torch_available() else ()
|
||||
)
|
||||
test_pruning = False
|
||||
test_headmasking = False
|
||||
test_torchscript = False
|
||||
|
||||
def setUp(self):
|
||||
self.model_tester = UniSpeechSatModelTester(
|
||||
self, conv_stride=(3, 3, 3), feat_extract_norm="layer", do_stable_layer_norm=True
|
||||
)
|
||||
self.config_tester = ConfigTester(self, config_class=UniSpeechSatConfig, hidden_size=37)
|
||||
|
||||
def test_config(self):
|
||||
self.config_tester.run_common_tests()
|
||||
|
||||
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_batched_inference(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
||||
self.model_tester.create_and_check_batch_inference(*config_and_inputs)
|
||||
|
||||
def test_ctc_loss_inference(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
||||
self.model_tester.check_ctc_loss(*config_and_inputs)
|
||||
|
||||
def test_seq_classifier_loss_inference(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
||||
self.model_tester.check_seq_classifier_loss(*config_and_inputs)
|
||||
|
||||
def test_ctc_train(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
||||
self.model_tester.check_ctc_training(*config_and_inputs)
|
||||
|
||||
def test_seq_classifier_train(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
||||
self.model_tester.check_seq_classifier_training(*config_and_inputs)
|
||||
|
||||
def test_labels_out_of_vocab(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
||||
self.model_tester.check_labels_out_of_vocab(*config_and_inputs)
|
||||
|
||||
# UniSpeechSat has no inputs_embeds
|
||||
def test_inputs_embeds(self):
|
||||
pass
|
||||
|
||||
# `input_ids` is renamed to `input_values`
|
||||
def test_forward_signature(self):
|
||||
pass
|
||||
|
||||
# UniSpeechSat cannot resize token embeddings
|
||||
# since it has no tokens embeddings
|
||||
def test_resize_tokens_embeddings(self):
|
||||
pass
|
||||
|
||||
# UniSpeechSat has no inputs_embeds
|
||||
# and thus the `get_input_embeddings` fn
|
||||
# is not implemented
|
||||
def test_model_common_attributes(self):
|
||||
pass
|
||||
|
||||
def test_retain_grad_hidden_states_attentions(self):
|
||||
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
config.output_hidden_states = True
|
||||
config.output_attentions = True
|
||||
|
||||
# no need to test all models as different heads yield the same functionality
|
||||
model_class = self.all_model_classes[0]
|
||||
model = model_class(config)
|
||||
model.to(torch_device)
|
||||
|
||||
# set layer drop to 0
|
||||
model.config.layerdrop = 0.0
|
||||
|
||||
input_values = inputs_dict["input_values"]
|
||||
|
||||
input_lengths = torch.tensor(
|
||||
[input_values.shape[1] for _ in range(input_values.shape[0])], dtype=torch.long, device=torch_device
|
||||
)
|
||||
output_lengths = model._get_feat_extract_output_lengths(input_lengths)
|
||||
|
||||
labels = ids_tensor((input_values.shape[0], output_lengths[0] - 2), self.model_tester.vocab_size)
|
||||
inputs_dict["attention_mask"] = torch.ones_like(inputs_dict["attention_mask"])
|
||||
inputs_dict["labels"] = labels
|
||||
|
||||
outputs = model(**inputs_dict)
|
||||
|
||||
output = outputs[0]
|
||||
|
||||
# Encoder-/Decoder-only models
|
||||
hidden_states = outputs.hidden_states[0]
|
||||
attentions = outputs.attentions[0]
|
||||
|
||||
hidden_states.retain_grad()
|
||||
attentions.retain_grad()
|
||||
|
||||
output.flatten()[0].backward(retain_graph=True)
|
||||
|
||||
self.assertIsNotNone(hidden_states.grad)
|
||||
self.assertIsNotNone(attentions.grad)
|
||||
|
||||
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",
|
||||
"codevectors",
|
||||
"quantizer.weight_proj.weight",
|
||||
"project_hid.weight",
|
||||
"project_hid.bias",
|
||||
"project_q.weight",
|
||||
"project_q.bias",
|
||||
"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",
|
||||
)
|
||||
|
||||
# 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, "codevectors") and module.codevectors is not None:
|
||||
module.codevectors.data.fill_(3)
|
||||
if hasattr(module, "masked_spec_embed") and module.masked_spec_embed is not None:
|
||||
module.masked_spec_embed.data.fill_(3)
|
||||
|
||||
def test_mask_feature_prob_ctc(self):
|
||||
model = UniSpeechSatForCTC.from_pretrained(
|
||||
"hf-internal-testing/tiny-random-unispeech-sat", mask_feature_prob=0.2, mask_feature_length=2
|
||||
)
|
||||
model.to(torch_device).train()
|
||||
processor = Wav2Vec2Processor.from_pretrained(
|
||||
"hf-internal-testing/tiny-random-unispeech-sat", return_attention_mask=True
|
||||
)
|
||||
|
||||
batch_duration_in_seconds = [1, 3, 2, 6]
|
||||
input_features = [np.random.random(16_000 * s) for s in batch_duration_in_seconds]
|
||||
|
||||
batch = processor(
|
||||
input_features, padding=True, sampling_rate=processor.feature_extractor.sampling_rate, return_tensors="pt"
|
||||
)
|
||||
|
||||
logits = model(
|
||||
input_values=batch["input_values"].to(torch_device),
|
||||
attention_mask=batch["attention_mask"].to(torch_device),
|
||||
).logits
|
||||
|
||||
self.assertEqual(logits.shape, (4, 1498, 32))
|
||||
|
||||
def test_mask_time_prob_ctc(self):
|
||||
model = UniSpeechSatForCTC.from_pretrained(
|
||||
"hf-internal-testing/tiny-random-unispeech-sat", mask_time_prob=0.2, mask_time_length=2
|
||||
)
|
||||
model.to(torch_device).train()
|
||||
processor = Wav2Vec2Processor.from_pretrained(
|
||||
"hf-internal-testing/tiny-random-unispeech-sat", return_attention_mask=True
|
||||
)
|
||||
|
||||
batch_duration_in_seconds = [1, 3, 2, 6]
|
||||
input_features = [np.random.random(16_000 * s) for s in batch_duration_in_seconds]
|
||||
|
||||
batch = processor(
|
||||
input_features, padding=True, sampling_rate=processor.feature_extractor.sampling_rate, return_tensors="pt"
|
||||
)
|
||||
|
||||
logits = model(
|
||||
input_values=batch["input_values"].to(torch_device),
|
||||
attention_mask=batch["attention_mask"].to(torch_device),
|
||||
).logits
|
||||
|
||||
self.assertEqual(logits.shape, (4, 1498, 32))
|
||||
|
||||
def test_mask_time_feature_prob_ctc_single_batch(self):
|
||||
model = UniSpeechSatForCTC.from_pretrained(
|
||||
"hf-internal-testing/tiny-random-unispeech-sat",
|
||||
mask_time_prob=0.2,
|
||||
mask_feature_prob=0.2,
|
||||
mask_time_length=2,
|
||||
mask_feature_length=2,
|
||||
)
|
||||
model.to(torch_device).train()
|
||||
processor = Wav2Vec2Processor.from_pretrained(
|
||||
"hf-internal-testing/tiny-random-unispeech-sat", return_attention_mask=True
|
||||
)
|
||||
|
||||
batch_duration_in_seconds = [6]
|
||||
input_features = [np.random.random(16_000 * s) for s in batch_duration_in_seconds]
|
||||
|
||||
batch = processor(
|
||||
input_features, padding=True, sampling_rate=processor.feature_extractor.sampling_rate, return_tensors="pt"
|
||||
)
|
||||
|
||||
logits = model(
|
||||
input_values=batch["input_values"].to(torch_device),
|
||||
attention_mask=batch["attention_mask"].to(torch_device),
|
||||
).logits
|
||||
|
||||
self.assertEqual(logits.shape, (1, 1498, 32))
|
||||
|
||||
@slow
|
||||
def test_model_from_pretrained(self):
|
||||
model = UniSpeechSatModel.from_pretrained("microsoft/unispeech-sat-large")
|
||||
self.assertIsNotNone(model)
|
||||
|
||||
|
||||
@require_torch
|
||||
@require_datasets
|
||||
@require_soundfile
|
||||
@slow
|
||||
class UniSpeechSatModelIntegrationTest(unittest.TestCase):
|
||||
def _load_datasamples(self, num_samples):
|
||||
import soundfile as sf
|
||||
|
||||
ids = [f"1272-141231-000{i}" for i in range(num_samples)]
|
||||
|
||||
# map files to raw
|
||||
def map_to_array(batch):
|
||||
speech, _ = sf.read(batch["file"])
|
||||
batch["speech"] = speech
|
||||
return batch
|
||||
|
||||
ds = load_dataset("patrickvonplaten/librispeech_asr_dummy", "clean", split="validation")
|
||||
|
||||
ds = ds.filter(lambda x: x["id"] in ids).sort("id").map(map_to_array)
|
||||
|
||||
return ds["speech"][:num_samples]
|
||||
|
||||
def _load_superb(self, task, num_samples):
|
||||
ds = load_dataset("anton-l/superb_dummy", task, split="test")
|
||||
|
||||
return ds[:num_samples]
|
||||
|
||||
def test_inference_encoder_base(self):
|
||||
model = UniSpeechSatModel.from_pretrained("microsoft/unispeech-sat-base-plus")
|
||||
model.to(torch_device)
|
||||
feature_extractor = Wav2Vec2FeatureExtractor.from_pretrained(
|
||||
"facebook/wav2vec2-base", return_attention_mask=True
|
||||
)
|
||||
input_speech = self._load_datasamples(2)
|
||||
|
||||
inputs_dict = feature_extractor(input_speech, return_tensors="pt", padding=True)
|
||||
|
||||
with torch.no_grad():
|
||||
outputs = model(
|
||||
inputs_dict.input_values.to(torch_device),
|
||||
attention_mask=inputs_dict.attention_mask.to(torch_device),
|
||||
)
|
||||
|
||||
# fmt: off
|
||||
expected_hidden_states_slice = torch.tensor(
|
||||
[[[-0.0743, 0.1384],
|
||||
[-0.0845, 0.1704]],
|
||||
[[-0.0954, 0.1936],
|
||||
[-0.1123, 0.2095]]],
|
||||
device=torch_device,
|
||||
)
|
||||
# fmt: on
|
||||
|
||||
self.assertTrue(torch.allclose(outputs.last_hidden_state[:, :2, -2:], expected_hidden_states_slice, atol=1e-3))
|
||||
|
||||
def test_inference_encoder_large(self):
|
||||
model = UniSpeechSatModel.from_pretrained("microsoft/unispeech-sat-large")
|
||||
model.to(torch_device)
|
||||
feature_extractor = Wav2Vec2FeatureExtractor.from_pretrained("facebook/wav2vec2-large-xlsr-53")
|
||||
input_speech = self._load_datasamples(2)
|
||||
|
||||
inputs_dict = feature_extractor(input_speech, return_tensors="pt", padding=True)
|
||||
|
||||
with torch.no_grad():
|
||||
outputs = model(
|
||||
inputs_dict.input_values.to(torch_device),
|
||||
attention_mask=inputs_dict.attention_mask.to(torch_device),
|
||||
)
|
||||
|
||||
# fmt: off
|
||||
expected_hidden_states_slice = torch.tensor(
|
||||
[[[-0.1172, -0.0797],
|
||||
[-0.0012, 0.0213]],
|
||||
[[-0.1225, -0.1277],
|
||||
[-0.0668, -0.0585]]],
|
||||
device=torch_device,
|
||||
)
|
||||
# fmt: on
|
||||
|
||||
self.assertTrue(torch.allclose(outputs.last_hidden_state[:, :2, -2:], expected_hidden_states_slice, atol=1e-3))
|
||||
Loading…
Reference in New Issue
Block a user