Adversarial Reweighting for Partial Domain Adaptation

Authors: Xiang Gu, Xi Yu, yan yang, Jian Sun, Zongben Xu

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments show that our method achieves state-of-the-art results on the benchmarks of Image Net-Caltech, Office-Home, Vis DA-2017, and Domain Net. Ablation studies also confirm the effectiveness of our approach.
Researcher Affiliation Academia Xiang Gu, Xi Yu, Yan Yang, Jian Sun , and Zongben Xu School of Mathematics and Statistics, Xi an Jiaotong University, P.R. China {xianggu,ericayu,yangyan92}@stu.xjtu.edu.cn {jiansun,zbxu}@xjtu.edu.cn
Pseudocode Yes We give the pseudo-code of the training algorithm in Supp. D.
Open Source Code Yes Our code is available at https://github.com/XJTU-XGU/ Adversarial-Reweighting-for-Partial-Domain-Adaptation.
Open Datasets Yes Office-31 dataset [36] contains 4,652 images of 31 categories, collected from three domains: Amazon (A), DSLR (D), and Webcam (W). ... Image Net-Caltech is built with Image Net (I) [35] and Caltech-256 (C) [12]... Office-Home [43] consists of four domains... Vis DA-2017 [32] is a large-scale challenging dataset... Domain Net [31] is another large-scale challenging dataset...
Dataset Splits No The paper describes how target domains are built (e.g., 'We use the first 6 classes in alphabetical order as the target domain' for Vis DA-2017) but does not provide explicit train/validation/test dataset split percentages or sample counts for the overall experimental setup.
Hardware Specification Yes We implement our method using Pytorch [30] on a Nvidia Tesla v100 GPU.
Software Dependencies No The paper mentions 'Pytorch [30]' and 'CVXPY [7] package' but does not specify their version numbers.
Experiment Setup Yes We use the SGD algorithm with momentum 0.9 to update θF and θC. The learning rate of θC is ten times that of θF . θD is updated by the Adam [18] algorithm with learning rate 0.001. Following [8], we adjust the learning rate η of θC by η = 0.01 (1+10p) 0.75 , where p is the training progress linearly changing from 0 to 1. We set the batchsize to 36, M = 500, and N = 36M.