OT4P: Unlocking Effective Orthogonal Group Path for Permutation Relaxation
Authors: Yaming Guo, chen zhu, Hengshu Zhu, Tieru Wu
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments involving the optimization over permutation matrices validate the effectiveness of the proposed method. ... This section conducts experiments to evaluate the performance of OT4P in optimization problems and probabilistic tasks. |
| Researcher Affiliation | Academia | Yaming Guo1,4, Chen Zhu2 , Hengshu Zhu3,4 , Tieru Wu1 1School of Artificial Intelligence, Jilin University 2School of Management, University of Science and Technology of China 3Computer Network Information Center, Chinese Academy of Sciences 4The Hong Kong University of Science and Technology (Guangzhou) |
| Pseudocode | Yes | We summarize the pseudo-code of OT4P in Algorithm 1. |
| Open Source Code | Yes | The core code for OT4P is available at https://github.com/Yaming Guo98/OT4P. |
| Open Datasets | Yes | We explore a variety of network architectures, including MLP5 (5-layer MLP) [54], VGG11 [60], and Res Net18 [22]. The weights for these networks are derived from official pre-trained models in Py Torch [52]... We use the WILLOW-Object Class dataset [9] to generate problem instances... We use the CMU House [5] image sequence to generate problem instances. |
| Dataset Splits | No | The paper mentions tuning learning rates ('The initial learning rates are tuned within the set {0.1, 0.01, 0.001, 0.0001}'), but it does not explicitly describe a separate validation dataset split or a validation procedure for hyperparameter tuning. It mainly focuses on training and testing. |
| Hardware Specification | Yes | We conducted the experiments (Appendices F.2 and F.3) using a single NVIDIA A800... on the NVIDIA Ge Force RTX 3090. |
| Software Dependencies | Yes | the runtime environment was Python=3.10, CUDA=11.7, and Py Torch=2.01. |
| Experiment Setup | Yes | The Adam W [45] with an initial learning rate of 0.1 is employed to minimize the loss w.r.t. Equation (17), with a maximum of 500 iterations. ... We conduct 500 iterations using the Adam optimizer [31] with an initial learning rate of 0.01. ... All algorithms employ the Adam optimizer for 100 iterations, with Rieman Birk utilizing Riemannian Adam [4, 34]. The initial learning rates are tuned within the set {0.1, 0.01, 0.001, 0.0001}. |