Proximal Alternating Direction Network: A Globally Converged Deep Unrolling Framework

Authors: Risheng Liu, Xin Fan, Shichao Cheng, Xiangyu Wang, Zhongxuan Luo

AAAI 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on various real-world applications verify the theoretical convergence and demonstrate the effectiveness of designed deep models.
Researcher Affiliation Academia 1DUT-RU International School of Information Science & Engineering, Dalian University of Technology, Dalian, China 2Key Laboratory for Ubiquitous Network and Service Software of Liaoning Province, Dalian, China 3School of Mathematical Science, Dalian University of Technology, Dalian, China {rsliu, xin.fan, zxluo}@dlut.edu.cn, shichao.cheng@outlook.com, wxy9326@gmail.com
Pseudocode Yes Algorithm 1 Proximal Alternating Direction Network
Open Source Code No The paper does not provide any explicit statement or link indicating that the source code for the methodology described in this paper is publicly available.
Open Datasets Yes We always use the set of 400 images of size 180x180 built in (Chen and Pock 2017) as our training data. Two commonly used image deblurring benchmarks respectively collected by Levin et. al. (Levin et al. 2009) ... and Sun et. al. (Sun et al. 2013) ... are used for testing. We build basic architecture G with 17 convolution layers ... and train it on synthetic hazy images (Ren et al. 2016) for our deep model. We evaluate the performance of our deep model ... on the commonly used Fattal s benchmark (Fattal 2008).
Dataset Splits No The paper mentions 'training data' and 'testing data' but does not provide specific percentages, sample counts, or detailed methodology for dataset splits, nor does it explicitly mention a validation set.
Hardware Specification Yes All experiments are conduced on a PC with Intel Core i7 CPU at 3.4 GHz, 32 GB RAM and a NVIDIA Ge Force GTX 1050 Ti GPU.
Software Dependencies No The paper does not provide specific software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions, or specific solver versions).
Experiment Setup Yes Here the basic architecture G (used in our deep models) consists of 7 convolution layers. The Re LU nonlinearities are added between each two linear layers accordingly and batch normalizations (BN) (Ioffe and Szegedy 2015) are also introduced for convolution operations from 2-nd to 6-th linear layers. Require: u0, v0, x0, λ0, G, tmax 1, kmax 1, ϵ > 0, CE > 0, {μk|2CE < μk < }, {ρk|ρk = (γ)kρ0, ρ0 > 0, γ > 1} ((γ)k denotes k-th power of γ) and {αk|αk = 1/ρk}.