Topology-aware Robust Optimization for Out-of-Distribution Generalization

Authors: Fengchun Qiao, Xi Peng

ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate TRO in a wide range of tasks including classification, regression, and semantic segmentation. We compare TRO with SOTA baselines on OOD generalization and conduct ablation study on the key components of TRO.
Researcher Affiliation Academia Fengchun Qiao University of Delaware fengchun@udel.edu Xi Peng University of Delaware xipeng@udel.edu
Pseudocode Yes Algorithm 1: TRO Algorithm
Open Source Code Yes The source code and pre-trained models are available at: https://github.com/joffery/TRO.
Open Datasets Yes DG-15 (Xu et al., 2022) is a synthetic binary classification dataset... TPT-48 (Vose et al., 2014) contains the monthly average temperature... Sen1Floods11 (Bonafilia et al., 2020) is a public dataset for flood mapping at the global scale... we have conducted experiments on PACS (Li et al., 2017), Terra (Beery et al., 2018), and VLCS (Fang et al., 2013).
Dataset Splits Yes Following Gulrajani & Lopez-Paz (2021), we perform model selection based on a validation set constructed from training groups only... data of other events are split into training and validation sets with a random 80-20 split.
Hardware Specification No The paper does not provide specific details about the hardware used for running the experiments, such as GPU or CPU models, memory, or cloud instance types.
Software Dependencies No The paper mentions using default values from other implementations, such as 'default values from the official implementation1 of Tong et al. (2021)' and 'default values from Xu et al. (2022) (DG-15/-60 and TPT-48) and Bonafilia et al. (2020) (Sen1Floods11)'. However, it does not explicitly list specific software or library versions like Python, PyTorch, or TensorFlow versions.
Experiment Setup Yes We select λ from {1e-3, 1e-2, 1e-1, 1, 10, 100} and select ηq from {1e-4, 1e-3, 1e-2, 1e-1, 1}.