AdaBits: Neural Network Quantization with Adaptive Bit-widths

SAT: Neural Network Quantization with Scale-Adjusted Training

SAT arXiv BMVC2020 | AdaBits arXiv CVPR2020 | Model Zoo | BibTex

Illustration of neural network quantization with scale-adjusted training and adaptive bit-widths. The same model can run at different bit-widths, permitting instant and adaptive accuracy-efficiency trade-offs.

Run

Requirements:
- python3, pytorch 1.0, torchvision 0.2.1, pyyaml 3.13.
- Prepare ImageNet-1k data following pytorch example.
Training and Testing:
- The codebase is a general ImageNet training framework using yaml config under apps dir, based on PyTorch.
- To test, download pretrained models to logs dir and directly run command.
- To train, comment test_only and pretrained in config file. You will need to manage visible gpus by yourself.
- Command: python train.py app:{apps/***.yml}. {apps/***.yml} is config file. Do not miss app: prefix.
Still have questions?
- If you still have questions, please search closed issues first. If the problem is not solved, please open a new.

Training Settings

All settings are

fp_pretrained=True

weight_only=False

rescale=True

clamp=True

rescale_conv=False

For all adaptive models, switchbn=True, switch_alpha=True

Centered Weight Quantization for Low Precision

For convergent training, it is beneficial to have vanishing mean for weight distribution, besides proper variance. To this end, we should use the following quantization method for weights:

$$ q = 2 \cdot \frac{1}{2^b}\Bigg(\mathrm{clip}\bigg(\Big\lfloor 2^b\cdot\frac{w+1}{2} \Big\rfloor,0,2^b-1\bigg)+\frac{1}{2}\Bigg) - 1 $$

(also check this)

The following table compares the different quantization schemes.

Original

Modified

Centered Symmetric

Centered Asymmetric

The code for the above is also available.

Technical Details

Implementing network quantization with adaptive bit-widths is straightforward:

Switchable batchnorm and adaptive bit-widths layers are implemented in models/quant_ops.
Quantization training with adaptive bit-widths is implemented in the [run_one_epoch] function in train.py.

Acknowledgement

This repo is based on slimmable_networks and benefits from the following projects

Neural Network Distiller

License

CC 4.0 Attribution-NonCommercial International

The software is for educaitonal and academic research purpose only.

Citing

@article{jin2019quantization,
  title={Towards efficient training for neural network quantization},
  author={Jin, Qing and Yang, Linjie and Liao, Zhenyu},
  journal={arXiv preprint arXiv:1912.10207},
  year={2019}
}
@article{jin2020sat,
  title={Neural Network Quantization with Scale-Adjusted Training},
  author={Jin, Qing and Yang, Linjie and Liao, Zhenyu and Qian, Xiaoning},
  booktitle={The British Machine Vision Conference},
  year={2020}
}
@article{jin2020adabits,
  title={AdaBits: Neural Network Quantization with Adaptive Bit-Widths},
  author={Jin, Qing and Yang, Linjie and Liao, Zhenyu},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  year={2020}
}

deJQK / AdaBits

readme