Code for ICCV2025 paper 'Semantic Alignment and Reinforcement for Data-Free Quantization of Vision Transformers'
Environment Setup
conda create -n zsqvit python=3.8
pip install torch==1.10.0+cu111 torchvision==0.11.0+cu111 torchaudio==0.10.0 -f https://download.pytorch.org/whl/torch_stable.html
pip install timm==0.4.12 IPython tqdm scipy matplotlib
pip install torch==1.10.0+cu111 torchvision==0.11.0+cu111 torchaudio==0.10.0 -f https://download.pytorch.org/whl/torch_stable.html
pip install timm==0.4.12 IPython tqdm scipy matplotlib
Sample Synthesis
CUDA_VISIBLE_DEVICES=0 python -u generate.py [--model]
[--calib_batchsize] [--save_fake] [--softlabel] [--coe_sf] [--coe_attn]
optional arguments:
--model: Model architecture
--calib_batchsize: Total number of synthesized samples, e.g., 32 or 1024.
--save_fake: save fake data?
--coe_sf: weight of SL loss
--coe_attn: weight of the APA loss
[--calib_batchsize] [--save_fake] [--softlabel] [--coe_sf] [--coe_attn]
optional arguments:
--model: Model architecture
--calib_batchsize: Total number of synthesized samples, e.g., 32 or 1024.
--save_fake: save fake data?
--coe_sf: weight of SL loss
--coe_attn: weight of the APA loss
Quantization
CUDA_VISIBLE_DEVICES=0 python -u test_quant.py [--model]
[--dataset] [--w_bit] [--a_bit] [--calib_batchsize]
[--iter] [--optim_batchsize] [--fake_path] [--rep] [--fake_path] [--box_path]
[--softtagets_path]
optional arguments:
--model: Model architecture
--dataset: Path to the ImageNet dataset
--w_bit: Bit-precision of weights
--a_bit: Bit-precision of activations
--optim_batchsize: Batch size per optimization iteration; varies with bit-width. Refer to PTQViT settings, see line 384 of test_quant.py
--rep: use reparameters or not (Please see RepQ-ViT and I\&s-vit)
--fake_path: Path to the synthesized samples
--box_path: Path to the boxes generated by MSR
--softtagets_path: Path to the softtarget generated by SL
[--dataset] [--w_bit] [--a_bit] [--calib_batchsize]
[--iter] [--optim_batchsize] [--fake_path] [--rep] [--fake_path] [--box_path]
[--softtagets_path]
optional arguments:
--model: Model architecture
--dataset: Path to the ImageNet dataset
--w_bit: Bit-precision of weights
--a_bit: Bit-precision of activations
--optim_batchsize: Batch size per optimization iteration; varies with bit-width. Refer to PTQViT settings, see line 384 of test_quant.py
--rep: use reparameters or not (Please see RepQ-ViT and I\&s-vit)
--fake_path: Path to the synthesized samples
--box_path: Path to the boxes generated by MSR
--softtagets_path: Path to the softtarget generated by SL
Citation
We appreciate it if you would please cite our paper if you found the code/idea useful for your work:
@article{zhong2024semantics,
title={Semantics Prompting Data-Free Quantization for Low-Bit Vision Transformers},
author={Zhong, Yunshan and Zhou, Yuyao and Zhang, Yuxin and Li, Shen and Li, Yong and Chao, Fei and Zeng, Zhanpeng and Ji, Rongrong},
booktitle={International Conference on Computer Vision},
year={2025}
}
title={Semantics Prompting Data-Free Quantization for Low-Bit Vision Transformers},
author={Zhong, Yunshan and Zhou, Yuyao and Zhang, Yuxin and Li, Shen and Li, Yong and Chao, Fei and Zeng, Zhanpeng and Ji, Rongrong},
booktitle={International Conference on Computer Vision},
year={2025}
}
Acknowledgments
Our code is based on the opensource code listed in the following. We highly appreciate their contribution!
@inproceedings{li2022psaqvit,
title={Patch Similarity Aware Data-Free Quantization for Vision Transformers},
author={Li, Zhikai and Ma, Liping and Chen, Mengjuan and Xiao, Junrui and Gu, Qingyi},
booktitle={European Conference on Computer Vision},
pages={154--170},
year={2022}
}
@inproceedings{li2023repq,
title={Repq-vit: Scale reparameterization for post-training quantization of vision transformers},
author={Li, Zhikai and Xiao, Junrui and Yang, Lianwei and Gu, Qingyi},
booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
pages={17227--17236},
year={2023}
}
@article{zhong2023s,
title={I\&s-vit: An inclusive \& stable method for pushing the limit of post-training vits quantization},
author={Zhong, Yunshan and Hu, Jiawei and Lin, Mingbao and Chen, Mengzhao and Ji, Rongrong},
journal={arXiv preprint arXiv:2311.10126},
year={2023}
}
title={Patch Similarity Aware Data-Free Quantization for Vision Transformers},
author={Li, Zhikai and Ma, Liping and Chen, Mengjuan and Xiao, Junrui and Gu, Qingyi},
booktitle={European Conference on Computer Vision},
pages={154--170},
year={2022}
}
@inproceedings{li2023repq,
title={Repq-vit: Scale reparameterization for post-training quantization of vision transformers},
author={Li, Zhikai and Xiao, Junrui and Yang, Lianwei and Gu, Qingyi},
booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
pages={17227--17236},
year={2023}
}
@article{zhong2023s,
title={I\&s-vit: An inclusive \& stable method for pushing the limit of post-training vits quantization},
author={Zhong, Yunshan and Hu, Jiawei and Lin, Mingbao and Chen, Mengzhao and Ji, Rongrong},
journal={arXiv preprint arXiv:2311.10126},
year={2023}
}