Lightweight Image Super-Resolution via Flexible Meta Pruning

Authors: Yulun Zhang, Kai Zhang, Luc Van Gool, Martin Danelljan, Fisher Yu

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct extensive experiments to investigate critical factors in the flexible channel and weight pruning for image SR, demonstrating the superiority of our FMP when applied to baseline image SR architectures.
Researcher Affiliation Academia 1Mo E Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University, China 2School of Intelligence Science and Technology, Nanjing University, China 3ETH Z urich, Switzerland 4KU Leuven, Belgium 5INSAIT, Bulgaria.
Pseudocode No The paper describes the proposed method using text and mathematical equations, but does not include any explicitly labeled pseudocode or algorithm blocks.
Open Source Code No The paper does not contain any explicit statement about releasing source code or a link to a code repository.
Open Datasets Yes Following most recent works (Timofte et al., 2017; Lim et al., 2017; Haris et al., 2018), we use DIV2K dataset (Timofte et al., 2017) and Flickr2K (Lim et al., 2017) as training data.
Dataset Splits Yes Following most recent works (Timofte et al., 2017; Lim et al., 2017; Haris et al., 2018), we use DIV2K dataset (Timofte et al., 2017) and Flickr2K (Lim et al., 2017) as training data. We use five standard benchmark datasets: Set5 (Bevilacqua et al., 2012), Set14 (Zeyde et al., 2010), B100 (Martin et al., 2001), Urban100 (Huang et al., 2015), and Manga109 (Matsui et al., 2017). ... We report inference time on Urban100 and DIV2K validation and test data.
Hardware Specification Yes We use Py Torch (Paszke et al., 2017) to implement our models with RTX 3090 GPUs.
Software Dependencies No We use Py Torch (Paszke et al., 2017) to implement our models with RTX 3090 GPUs.
Experiment Setup Yes We perform data augmentation on the training images, which are randomly rotated by 90 , 180 , 270 and flipped horizontally. Each training batch consists of 16 LR color patches, whose size is 64 64. Our FMP model is trained by ADAM optimizer (Kingma & Ba, 2015) with β1=0.9, β2=0.999, and ϵ=10 8. We set the initial learning rate as 10 4 and then decrease it to half every 2 105 iterations.