Skip to content

Latest commit

 

History

History
118 lines (85 loc) · 4.78 KB

README.md

File metadata and controls

118 lines (85 loc) · 4.78 KB
Raw

LightHuBERT

LightHuBERT: Lightweight and Configurable Speech Representation Learning with Once-for-All Hidden-Unit BERT

| Github | Huggingface | SUPERB Leaderboard |

The authors' PyTorch implementation and pretrained models of LightHuBERT.

  • September 2022: Published LightHuBERT paper has been available in ISCA Archive
  • June 2022: LightHuBERT has been accepted to Interspeech 2022
  • May 2022: LightHuBERT Small is available in s3prl.
  • March 2022: release preprint in arXiv and checkpoints in huggingface.

Pre-Trained Models

Model Pre-Training Dataset Download Link
LightHuBERT Base 960 hrs LibriSpeech huggingface: lighthubert/lighthubert_base.pt
LightHuBERT Small 960 hrs LibriSpeech huggingface: lighthubert/lighthubert_small.pt
LightHuBERT Stage 1 960 hrs LibriSpeech huggingface: lighthubert/lighthubert_stage1.pt

Actually, the pre-trained is trained in common.fp16: true so that we can perform model inference with fp16 weights.

Requirements and Installation

  • PyTorch version >= 1.8.1
  • Python version >= 3.6
  • numpy version >= 1.19.3
  • To install lighthubert:
git clone [email protected]:mechanicalsea/lighthubert.git
cd lighthubert
pip install --editable .

Load Pre-Trained Models for Inference

import torch
from lighthubert import LightHuBERT, LightHuBERTConfig

wav_input_16khz = torch.randn(1,10000).cuda()

# load the pre-trained checkpoints
checkpoint = torch.load('/path/to/lighthubert.pt')
cfg = LightHuBERTConfig(checkpoint['cfg']['model'])
cfg.supernet_type = 'base'
model = LightHuBERT(cfg)
model = model.cuda()
model = model.eval()
print(model.load_state_dict(checkpoint['model'], strict=False))

# (optional) set a subnet
subnet = model.supernet.sample_subnet()
model.set_sample_config(subnet)
params = model.calc_sampled_param_num()
print(f"subnet (Params {params / 1e6:.0f}M) | {subnet}")

# extract the the representation of last layer
rep = model.extract_features(wav_input_16khz)[0]

# extract the the representation of each layer
hs = model.extract_features(wav_input_16khz, ret_hs=True)[0]

print(f"Representation at bottom hidden states: {torch.allclose(rep, hs[-1])}")

More examples can be found in our tutorials.

Universal Representation Evaluation on SUPERB

SUPERB Leaderboard

Profiling LightHuBERT

As mentioned in Profiling Tool for SLT2022 SUPERB Challenge, we profiling the lighthubert in s3prl.

cd DeepSpeed
# lighthubert_small
python testing/s3prl_profiling_test.py -u lighthubert_small --libri_root "libri_root"
# lighthubert_base
python testing/s3prl_profiling_test.py -u lighthubert_base --libri_root "libri_root"
# lighthubert_stage1
python testing/s3prl_profiling_test.py -u lighthubert_stage1 --libri_root "libri_root"

License

This project is licensed under the license found in the LICENSE file in the root directory of this source tree. Portions of the source code are based on the FAIRSEQ project.

Reference

If you find our work is useful in your research, please cite the following paper:

@inproceedings{wang2022lighthubert,
  title={{LightHuBERT}: Lightweight and Configurable Speech Representation Learning with Once-for-All Hidden-Unit {BERT}},
  author={Rui Wang and Qibing Bai and Junyi Ao and Long Zhou and Zhixiang Xiong and Zhihua Wei and Yu Zhang and Tom Ko and Haizhou Li},
  booktitle={Proc. Interspeech},
  month = {sep},
  number = {September},
  pages = {1686--1690},
  year={2022},
  doi = {10.21437/Interspeech.2022-10269},
  url = {https://www.isca-speech.org/archive/interspeech_2022/wang22t_interspeech.html},
}

Contact Information

For help or issues using LightHuBERT models, please submit a GitHub issue.

For other communications related to LightHuBERT, please contact Rui Wang ([email protected]).