Advancing Open-Set Domain Generalization Using Evidential Bi-Level Hardest Domain Scheduler
Authors: Kunyu Peng, Di Wen, Kailun Yang, Ao Luo, Yufan Chen, Jia Fu, M. Saquib Sarfraz, Alina Roitberg, Rainer Stiefelhagen
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We verify our approach on three OSDG benchmarks, i.e., PACS, Digits DG, and Office Home. The results show that our method substantially improves OSDG performance and achieves more discriminative embeddings for both the seen and unseen categories, underscoring the advantage of a judicious domain scheduler for the generalizability to unseen domains and unseen categories. The source code is publicly available at https://github.com/KPeng9510/EBi L-Ha DS. |
| Researcher Affiliation | Collaboration | Kunyu Peng1, Di Wen1, Kailun Yang2 , Ao Luo3, Yufan Chen1, Jia Fu4,5, M. Saquib Sarfraz1,6, Alina Roitberg7, Rainer Stiefelhagen1 1Karlsruhe Institute of Technology 2Hunan University 3Waseda University 4KTH Royal Institute of Technology 5RISE Research Institutes of Sweden 6Mercedes-Benz Tech Innovation 7University of Stuttgart |
| Pseudocode | Yes | Algorithm 1 Training with Evidential Bi-Level Hardest Domain Scheduler. |
| Open Source Code | Yes | The source code is publicly available at https://github.com/KPeng9510/EBi L-Ha DS. |
| Open Datasets | Yes | Our experiments demonstrate the effectiveness of domain scheduling via EBi L-Ha DS on three established datasets: PACS [31], Digits DG [63], and Office Home [51], which span a variety of image classification tasks. |
| Dataset Splits | Yes | The leave-one-domain-out DG setting is adopted. The open-set ratio is 6:1. The open-set ratio is 6:4, where digits 0, 1, 2, 3, 4, 5 are used as seen categories while digits 6, 7, 8, 9 are selected as unseen categories, in Table 5. Since the H-score relies on a predefined threshold derived from the source domain validation set to separate the seen categories and unseen categories, it is regarded as a secondary metric in our evaluation. |
| Hardware Specification | Yes | All the experiments use Py Toch 2.0 and one NVIDIA A100 GPU. This work was performed on the Hore Ka supercomputer funded by the Ministry of Science, Research and the Arts Baden-Württemberg and by the Federal Ministry of Education and Research. |
| Software Dependencies | Yes | All the experiments use Py Toch 2.0 and one NVIDIA A100 GPU. |
| Experiment Setup | Yes | We set the upper limit of the training step as 1e4 and use SGD optimizer, where the learning rate (lr) is set as 1e 3 and batch size is chosen as 16. The weights of LCLS, LREG, and LRBE are chosen as 1.0, 1e 4, and 5e 4. Lr decay is 1e 1 and conducted at 8e3 meta-training step. |