Interpretable Lightweight Transformer via Unrolling of Learned Graph Smoothness Priors

Authors: VIET HO TAM THUC DO, Parham Eftekhar, Seyed Alireza Hosseini, Gene Cheung, Philip Chou

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments for two image interpolation applications verify the restoration performance, parameter efficiency and robustness to covariate shift of our graph-based unrolled networks compared to conventional transformers. (Abstract)
Researcher Affiliation Collaboration Tam Thuc Do York University Toronto, Canada dtamthuc@yorku.ca Parham Eftekhar York University Toronto, Canada eftekhar@yorku.ca Seyed Alireza Hosseini York University Toronto, Canada ahoseini@yorku.ca Gene Cheung York University Toronto, Canada genec@yorku.ca Philip A. Chou packet.media Seattle, USA pachou@ieee.org
Pseudocode No The paper describes iterative algorithms (CG, ADMM) through mathematical equations and textual explanations, but it does not provide a clearly labeled 'Pseudocode' or 'Algorithm' block with structured steps.
Open Source Code No The NeurIPS Paper Checklist for 'Open access to data and code' explicitly states 'No' and justifies it by saying, 'We used open source data: Mc M [43], Kodak [44], and Urban100 [45] datasets.', implying they did not release their own code.
Open Datasets Yes To train each learned model, we used the DIV2K dataset, which contains 800 and 100 high-resolution (HR) training and validation images, respectively. ... To test a model, we used the Mc M [43], Kodak [44], and Urban100 [45] datasets, running each model on the whole images. (Section 6.1)
Dataset Splits Yes Since the images are HR, we patchified the images into small images and used only about 1 to 4% of the patches for training and validation sets. We randomly sampled patches of 64 × 64 pixels to train the model. (Section 6.1)
Hardware Specification Yes All models were developed using Python 3.11. We leveraged Py Torch to implement all models and trained them using NVIDIA Ge Force RTX 2080 Ti. (Section 6.1)
Software Dependencies Yes All models were developed using Python 3.11. We leveraged Py Torch to implement all models and trained them using NVIDIA Ge Force RTX 2080 Ti. (Section 6.1)
Experiment Setup Yes A training dataset was created consisting of 5000, 10000, or 20000 image patches, each of size 64 × 64, to train the model. ... The parameters γ, α, β, and the metric matrix M were learned during training. ... In all ADMM blocks, we set the number of ADMM iterations to 5 and the number of CG iterations to 10. (Appendix A.8)