CVPR 2021 : Zero-shot Adversarial Quantization (ZAQ)
[paper] accepted as Oral by CVPR 2021.
Author: Yuang Liu, Wei Zhang, Jun Wang
East China Normal University (ECNU)
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.
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
PyTorch >= 1.7
Note: This code is temporarily for reference and we will upload a improved version in the future.
@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}
}