AIQViT: Architecture-Informed Post-Training Quantization for Vision Transformers

Authors: Runqing Jiang, Ye Zhang, Longguang Wang, Pengpeng Yu, Yulan Guo

AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on five vision tasks, including image classification, object detection, instance segmentation, point cloud classification, and point cloud part segmentation, demonstrate the superiority of AIQVi T over state-of-the-art PTQ methods.
Researcher Affiliation Academia 1Shenzhen Campus, Sun Yat-sen University 2Aviation University of Air Force EMAIL, EMAIL, EMAIL, EMAIL
Pseudocode No The paper describes methods using mathematical equations and textual explanations, but it does not include any explicitly labeled pseudocode or algorithm blocks.
Open Source Code No The paper does not contain an explicit statement about releasing source code or provide any links to code repositories. The conclusion states 'In future works, we plan to develop novel PTQ methods for large models...'
Open Datasets Yes We adopt Image Net (Krizhevsky, Sutskever, and Hinton 2012), Model Net40 (Wu et al. 2015), and Shape Net Part (Yi et al. 2016) for image classification, point cloud classification, and point cloud part segmentation respectively. The COCO (Lin et al. 2014) dataset is used to evaluate object detection and instance segmentation tasks.
Dataset Splits Yes For image classification, we adopt the experimental settings from (Zhong et al. 2023) and randomly select a calibration set of 1,024 samples from the Image Net dataset. ... For object detection and instance segmentation, ... using a calibration set of 256 samples and a batch size of 1. ... For point cloud classification and part segmentation, ... a calibration set of 512 samples.
Hardware Specification No The paper does not provide specific details about the hardware (e.g., GPU models, CPU types) used for running the experiments.
Software Dependencies No The paper does not mention specific software dependencies with version numbers (e.g., Python 3.x, PyTorch 1.x) that were used for the experiments.
Experiment Setup Yes For image classification, ... with batch size set to 24. For object detection and instance segmentation, ... batch size of 1. For point cloud classification and part segmentation, ... a batch size of 32. The iteration numbers are set to 2,000 and 6,000 for network architecture search and calibration, respectively. We empirically set S = {10, 20, 50, 100, 150} for all model variants across all vision tasks. ...λ0 is the initial proportion of training samples which is set to 0.5 in our method