Awesome

Zero-shot Adversarial Quantization (ZAQ)

[paper] accepted as Oral by CVPR 2021.

Author: Yuang Liu, Wei Zhang, Jun Wang

East China Normal University (ECNU)

Intro

<div align="center"> <img src="./images/overview.png" width="50%" alt="overview" /> <br> <div style="color:orange; border-bottom: 1px solid #d9d9d9; display: inline-block; color: #999; padding: 2px;">Figure 1. Overview. (a) previous methods; (b) ours.</div> </div>

To address the quantization issue without data, we propose a zero-shot adversarial quantization (ZAQ) framework, facilitating effective discrepancy estimation and knowledge transfer from a full-precision model to its quantized model. This is achieved by a novel two-level discrepancy modeling to drive a generator to synthesize informative and diverse data examples to optimize the quantized model in an adversarial learning fashion.

<div align="center"> <img src="./images/ZAQ_arch.png" width="60%" alt="framework" /> <br> <div style="color:orange; border-bottom: 1px solid #d9d9d9; display: inline-block; color: #999; padding: 2px;">Figure 2. ZAQ framework</div> </div>

Requirements

python>=3.6
torch>=1.2
torchvision
visdom
numpy
pillow
scikit-learn

Usage

To obtain a full-precision model, please refer train.py.

QAT on original dataset:

python quantize.py --model resnet18 --ckpt 'path/' --data_root './data/' --weight_bit 6 --activation_bit 8

Zero-shot quantization without data:

python main.py --model resnet18 --ckpt 'path/' --data_root './data/' --weight_bit 6 --activation_bit 8

Todo

Segmentation networks
Object detection networks
Quantization supported by PyTorch >= 1.7
Mixed-/Arbitrary- precision quantization

Note: This code is temporarily for reference and we will upload a improved version in the future.

Citation

@InProceedings{yuang2021zaq,
    tilte = {Zero-shot Adversarial Quantization},
    author = {Liu, Yuang and Zhang, Wei and Wang, Jun},
    booktitle = {Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
    month = {June},
    year = {2021}
}