LightHuBERT

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

Authors: Rui Wang, Qibing Bai, Junyi Ao, Long Zhou, Zhixiang Xiong, Zhihua Wei, Yu Zhang, Tom Ko and Haizhou Li

The authors' PyTorch implementation and pre-trained models of LightHuBERT.

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

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])}")

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"

Reference

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

@article{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},
  journal={arXiv preprint arXiv:2203.15610},
  year={2022}
}

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]).

mechanicalsea
/

lighthubert