Diverse Weight Averaging for Out-of-Distribution Generalization

Authors: Alexandre Rame, Matthieu Kirchmeyer, Thibaud Rahier, Alain Rakotomamonjy, Patrick Gallinari, Matthieu Cord

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimentally, Di WA consistently improves the state of the art on Domain Bed without inference overhead. We now present our evaluation on Domain Bed [12]. By imposing the code, the training procedures and the Res Net50 [52] architecture, Domain Bed is arguably the fairest benchmark for OOD generalization. It includes 5 multi-domain real-world datasets: PACS [51], VLCS [53], Office Home [50], Terra Incognita [54] and Domain Net [55].
Researcher Affiliation Collaboration 1Sorbonne Université, CNRS, ISIR, F-75005 Paris, France 2Criteo AI Lab, Paris, France 3Valeo.ai, Paris, France 4Université de Rouen, LITIS, France
Pseudocode Yes Algorithm 1 Di WA Pseudo-code
Open Source Code Yes Our code is available at https://github.com/alexrame/diwa.
Open Datasets Yes We now present our evaluation on Domain Bed [12]. By imposing the code, the training procedures and the Res Net50 [52] architecture, Domain Bed is arguably the fairest benchmark for OOD generalization. It includes 5 multi-domain real-world datasets: PACS [51], VLCS [53], Office Home [50], Terra Incognita [54] and Domain Net [55].
Dataset Splits Yes The validation dataset is sampled from S, i.e., we follow Domain Bed s training-domain model selection.
Hardware Specification Yes Approximately 20000 hours of GPUs (Nvidia V100) on an internal cluster, mostly for the 2640 runs needed in Table 1.
Software Dependencies No The paper does not provide specific software version numbers for ancillary software dependencies.
Experiment Setup Yes The experimental setup is further described in Appendix G.1. In our experiments, we thus use the mild search space defined in Table 7, first introduced in SWAD [14].