Uncertainty Modeling for Out-of-Distribution Generalization
Authors: Xiaotong Li, Yongxing Dai, Yixiao Ge, Jun Liu, Ying Shan, LINGYU DUAN
ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate that our proposed method consistently improves the network generalization ability on multiple vision tasks, including image classification, semantic segmentation, and instance retrieval. |
| Researcher Affiliation | Collaboration | 1Peking University, Beijing, China 2ARC Lab, Tencent PCG 3Singapore University of Technology and Design, Singapore 4Peng Cheng Laboratory, Shenzhen, China |
| Pseudocode | Yes | The algorithm is described in the Appendix. A.1 ALGORITHM The algorithm of the proposed method is illustrated in Algorithm 1. |
| Open Source Code | Yes | The code can be available at https://github.com/lixiaotong97/DSU. |
| Open Datasets | Yes | We evaluate the proposed method on PACS (Li et al. (2017))... In addition to PACS, we also employ Office-Home (Venkateswara et al., 2017) for multi-domain generalization experiments in the Appendix. To evaluate the cross-scenario generalization ability of segmentation models, we adopt synthetic GTA5 for training while using real City Scapes for testing. Experiments are conducted on the widely used Duke MTMC (Ristani et al. (2016)) and Market1501 (Zheng et al. (2015)) datasets. We validate the proposed method for robustness towards corruptions on Image Net-C (Hendrycks & Dietterich (2019)), which contains 15 different pixel-level corruptions. Res Net50 is trained with 100 epochs for convergence on large-scale Image Net-1K (Deng et al. (2009)). |
| Dataset Splits | Yes | The implementation follows the official setup of Mix Style (Zhou et al. (2021b)) with a leave-one-domain-out protocol and Res Net18 (He et al., 2016) is used as the backbone. Res Net50 is trained with 100 epochs for convergence on large-scale Image Net-1K (Deng et al. (2009)). |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU models, CPU types, or memory) used for running its experiments. It only mentions training models on certain datasets with specific network backbones. |
| Software Dependencies | No | The paper mentions using Deep Lab-v2, ResNet backbones (ResNet18, ResNet50, ResNet101), and implementing based on released codes from FADA and MMT. However, it does not specify version numbers for these software components or any other ancillary software dependencies (e.g., Python, PyTorch versions). |
| Experiment Setup | Yes | Res Net50 is trained with 100 epochs for convergence on large-scale Image Net-1K (Deng et al. (2009)) and the hyperparameter p is set as 0.1 for training in Image Net. The hyper-parameter of the probability p is to trade off the strength of feature statistics augmentation... the accuracy reaches the best results when setting p as 0.5, which is also adopted as the default setting in all experiments if not specified. we use the batch size of 64 in our paper, following the original setting in PACS (Li et al., 2017) for fair comparison. |