Energy-Based Test Sample Adaptation for Domain Generalization

Authors: Zehao Xiao, Xiantong Zhen, Shengcai Liao, Cees G. M. Snoek

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on six benchmarks for classification of images and microblog threads demonstrate the effectiveness of our proposal.
Researcher Affiliation Collaboration 1AIM Lab, University of Amsterdam 2Inception Institute of Artificial Intelligence
Pseudocode Yes We provide the detailed training and test algorithm of our energy-based sample adaptation in Algorithm 1.
Open Source Code Yes 1 Code available: https://github.com/zzzx1224/EBTSA-ICLR2023.
Open Datasets Yes We conduct our experiments on five widely used datasets for domain generalization, PACS (Li et al., 2017), Office-Home (Venkateswara et al., 2017), Domain Net (Peng et al., 2019), and Rotated MNIST and Fashion-MNIST. ... PHEME (Zubiaga et al., 2016)
Dataset Splits Yes We use the same training and validation split as (Li et al., 2017) and follow their leaveone-out protocol.
Hardware Specification Yes We train all models on an NVIDIA Tesla V100 GPU for 10,000 iterations.
Software Dependencies No The paper mentions software components like 'Adam optimization' and 'Distil BERT' but does not provide specific version numbers for any libraries or frameworks, which is required for reproducibility.
Experiment Setup Yes We use Adam optimization and train for 10,000 iterations with a batch size of 128. We set the learning rate to 0.00005 for Res Net-18, 0.00001 for Res Net-50, and 0.0001 for the energy-based model and classification model. We use 20 steps of Langevin dynamics sampling to adapt the target samples to source distributions, with a step size of 50.