Variational Model Perturbation for Source-Free Domain Adaptation

Authors: Mengmeng Jing, Xiantong Zhen, Jingjing Li, Cees Snoek

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on several source-free benchmarks under three different evaluation settings verify the effectiveness of the proposed variational model perturbation for source-free domain adaptation.
Researcher Affiliation Academia Mengmeng Jing1,2 , Xiantong Zhen2 , Jingjing Li1, Cees G. M. Snoek2; 1University of Electronic Science and Technology of China 2University of Amsterdam
Pseudocode No The paper describes its methods using prose and mathematical equations but does not include any structured pseudocode or algorithm blocks.
Open Source Code Yes Code is available at https://github.com/mmjing/Variational_Model_Perturbation.
Open Datasets Yes The Office [47] dataset includes 3 domains. ... Office-Home [48] consists of 4 domains. ... CIFAR10-C, CIFAR100-C [49] and Image Net-C [49]
Dataset Splits Yes Generalized SFDA [13]: we split the source data into 80% and 20% parts. In the source pre-training phase, we use the labeled 80% part to pre-train the source model. In the target adaptation phase, we use all the unlabeled target data to adapt the model. In the testing phase, we predict the remaining 20% source data and all target data.
Hardware Specification No The paper specifies network architectures used (e.g., Res Net-50, Wide Res Net-28, Res Ne Xt-29) but does not provide details on the specific hardware (e.g., GPU model, CPU type) used for experiments.
Software Dependencies No The paper mentions optimizers (Stochastic Gradient Descent, Adam) and network architectures but does not provide specific version numbers for software libraries or dependencies (e.g., PyTorch 1.x, TensorFlow 2.x).
Experiment Setup Yes As for the optimizer, following [7], we employ Stochastic Gradient Descent with weight decay 1e-3 and momentum 0.9. As for the learning rate, we set 2e-3 for the backbone model and 2e-2 for the bottleneck layer newly added in SHOT. ... β is set to 0.3 in both Office and Office-Home. The batch size is 64.