Unified Optimal Transport Framework for Universal Domain Adaptation

Authors: Wanxing Chang, Ye Shi, Hoang Tuan, Jingya Wang

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

Reproducibility Variable Result LLM Response
Research Type Experimental 4 Experiments
Researcher Affiliation Academia Wanxing Chang1 Ye Shi1,3 Hoang Duong Tuan2 Jingya Wang1,3 1Shanghai Tech University 2University of Technology Sydney 3Shanghai Engineering Research Center of Intelligent Vision and Imaging
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code Yes Our code is available at: https://github.com/changwxx/Uni OT-for-Uni DA.
Open Datasets Yes Office [31] contains 31 categories and about 4K images in 3 domains: Amazon(A), DSLR(D) and Webcam(W). Office-Home [37] contains 65 categories and about 15K images in 4 domains: Artistic images(Ar), Clip-Art images(Cl), Product images(Pr) and Real-World images(Rw). Vis DA [30] is a large dataset which conains 12 categories in source domain with 15K synthetic images and target domain with 5K real-world images. Domain Net [29] is the largest domain adaptation dataset which contains 345 categories and about 0.6 million in 6 domains but we only use 3 domain: Painting (P), Real (R), and Sketch (S) like [17].
Dataset Splits Yes We follow the dataset split in [17] to conduct experiments.
Hardware Specification Yes All experiments are implemented on a GPU of NVIDIA TITAN V with 12GB.
Software Dependencies No The paper mentions 'POT [15]' but does not provide specific version numbers for software components like POT, PyTorch, or CUDA.
Experiment Setup Yes The batch size is set to 36 for both source and target domains. For all experiments, the initial learning rate is set to 1 10 2 for all new layers and 1 10 3 for pretrained backbone. The total training steps are set to be 10K for all datasets. For the pre-defined number of target prototypes, a larger size of target domain indicates a larger K. Therefore, we empirically set K = 50 for Office, K = 150 for Office-Home, K = 500 for Vis DA, K = 1000 for Domain Net. We default γ = 0.7, µ = 0.7, τ = 0.1, ε = 0.01, κ = 0.5 and λ = 0.1 for all datasets. We set the size of memory queue 2K for Office and Office-Home, 10K for Vis DA and Domain Net.