Adaptive Dataset Quantization

Authors: Muquan Li, Dongyang Zhang, Qiang Dong, Xiurui Xie, Ke Qin

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on CIFAR-10, CIFAR-100 (Krizhevsky, Hinton et al. 2009), Image Net-1K (Russakovsky et al. 2015) and Tiny-Image Net (Le and Yang 2015) substantiate a marked enhancement in performance over the baseline DQ by average 3%, establishing the new state-of-the-art results.
Researcher Affiliation Academia Institute of Intelligent Computing, University of Electronic Science and Technology of China, China EMAIL, EMAIL
Pseudocode Yes Algorithm 1: Adaptive Dataset Quantization
Open Source Code Yes Code https://github.com/SLGSP/ADQ
Open Datasets Yes Datasets Following the evaluation protocol of previous DQ (Zhou et al. 2023), we utilize image classification as a proxy task for evaluation and mainly assess our method on CIFAR-10 (Krizhevsky, Hinton et al. 2009) and Image Net-1K (Russakovsky et al. 2015).
Dataset Splits Yes CIFAR-10 contains 50,000 samples for training and 10,000 samples for validation, with a resolution of 32 × 32. Image Net-1K comprises 128,1126 samples from 1000 categories for training, with each category containing 50 images for validation.
Hardware Specification No The paper mentions 'GPU hours' in Table 2, but does not specify any particular GPU models or other hardware components used for the experiments.
Software Dependencies No The paper mentions models like Res Net-18 and Vision Transformer, but does not provide specific version numbers for any software libraries, frameworks, or programming languages used.
Experiment Setup Yes For comparison, we conduct training for 200 epochs on the CIFAR-10 with batch size 128, and we employ a cosineannealed learning rate that initializes at 0.1.