TexQ: Zero-shot Network Quantization with Texture Feature Distribution Calibration

Authors: Xinrui Chen, Yizhi Wang, Renao YAN, Yiqing Liu, Tian Guan, Yonghong He

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on CIFAR10/100 and Image Net show that Tex Q is observed to perform state-of-the-art in low bit width quantization.
Researcher Affiliation Academia Xinrui Chen, Yizhi Wang, Renao Yan , Yiqin Liu, Tian Guan , Yonghong He Tsinghua Shenzhen International Graduate School, Tsinghua University {cxr22, yz-wang22, yra21, liuyiqin20}@mails.tsinghua.edu.cn {guantian, heyh}@sz.tsinghua.edu.cn
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide concrete access to source code for the methodology described in this paper.
Open Datasets Yes We report top-1 accuracy on validation sets of CIFAR-10/100 [52] and Image Net [53].
Dataset Splits Yes We report top-1 accuracy on validation sets of CIFAR-10/100 [52] and Image Net [53]. In fine-tuning, batchsize is 128 for CIFAR-10/100 and 16 for Image Net, adjusting by cosine annealing [57].
Hardware Specification Yes All experiments are implemented with Pytorch [56] via the code of FDDA [19] and Intra Q [29], and run on an NVIDIA Ge Force RTX 3090 GPU.
Software Dependencies No The paper mentions 'Pytorch' but does not specify its version number or any other software dependencies with their versions.
Experiment Setup Yes Calibration images are iterated with a constant learning rate of 0.05. Generator is imported from GDFQ [24] with a initial learning rate of 1e-3 multiplied by 0.1 every 100 epochs. In fine-tuning, batchsize is 128 for CIFAR-10/100 and 16 for Image Net, adjusting by cosine annealing [57]. We warm up G for 50 epochs, then update G and Q for 450 epochs. The optimal configurations on trade-off parameters from α1 to α6 obtained by grid search are 2, 10, 0.4, 0.02, 1.8, and 20. The optimal configurations are k=9, θU=0.3, θL=0.5, and ε=0.015.