A Sampling-Based Method for Tensor Ring Decomposition
Authors: Osman Asif Malik, Stephen Becker
ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We compare our proposal to existing methods in experiments on both synthetic and real data. Our method achieves substantial speedup sometimes two or three orders of magnitude over competing methods, while maintaining good accuracy. |
| Researcher Affiliation | Academia | 1Department of Applied Mathematics, University of Colorado Boulder, Boulder, CO, USA. Correspondence to: Osman Asif Malik <osman.malik@colorado.edu>. |
| Pseudocode | Yes | Algorithm 1: TR-ALS (Zhao et al., 2016) Algorithm 2: TR-ALS-Sampled (proposal) Algorithm 3: Sampling from q =n (proposal) Algorithm 4: Computation of G =n [2] (proposal) |
| Open Source Code | Yes | All our code used in the experiments is available online4. 4Available at https://github.com/OsmanMalik/tr-als-sampled. |
| Open Datasets | Yes | We consider five real image and video datasets5 (sizes in parentheses): Pavia University (610 340 103) and DC Mall (1280 307 191) are 3-way tensors containing hyperspectral images. ... Red Truck (128 128 3 72) is a 4-way tensor consisting of 72 color images of size 128 by 128 pixels depicting a red truck from different angles. It is a subset of the COIL-100 dataset (Nene et al., 1996). 5Links to the datasets are provided in Section S3 of the supplement. |
| Dataset Splits | Yes | The classification is done using 10-fold cross validation. |
| Hardware Specification | Yes | All experiments are run in Matlab R2017a with Tensor Toolbox 2.6 (Bader et al., 2015) on a computer with an Intel Xeon E5-2630 v3 @ 32x 3.2GHz CPU and 32 GB of RAM. |
| Software Dependencies | Yes | All experiments are run in Matlab R2017a with Tensor Toolbox 2.6 (Bader et al., 2015) |
| Experiment Setup | Yes | Initialize cores G(2), . . . , G(N), drawing each core entry independently from a standard normal distribution, until some termination criterion is met, target ranks (R1, . . . , RN), oversampling parameter of 10 in TR-SVD-Rand, The target rank is also set to R = 10., when all target ranks are R = 10 and R = 20, each rank Rn = 5, k-nearest neighbor algorithm with k = 1 |