Adversarial Style Mining for One-Shot Unsupervised Domain Adaptation

Authors: Yawei Luo, Ping Liu, Tao Guan, Junqing Yu, Yi Yang

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on both cross-domain classification and segmentation benchmarks verify that ASM achieves state-of-the-art adaptation performance under the challenging one-shot setting.
Researcher Affiliation Collaboration 1School of Computer Science & Technology, Huazhong University of Science & Technology 2CCAI, Zhejiang University 3Baidu Research 4Re LER, University of Technology Sydney 5Institute of High Performance Computing, A*STAR, Singapore
Pseudocode Yes Algorithm 1: Adversarial Style Mining
Open Source Code Yes The code is publicly available at https://github. com/Royal Vane/ASM.
Open Datasets Yes We use MNIST [21]-USPS [16]-SVHN [35] benchmarks to evaluate ASM on one-shot cross domain classification task. For one-shot cross-domain segmentation task, we evaluate ASM on two benchmarks, i.e., SYNTHIA [37] Cityscapes [5] and GTA5 [36] Cityscapes.
Dataset Splits No We run each OSUDA experiment for 5 times to get the average result, where each time we randomly select one-shot sample from the target domain. The paper describes how the one-shot sample is selected but does not specify general train/validation/test splits for the full datasets used.
Hardware Specification No The paper does not provide specific details on the hardware used, such as GPU or CPU models.
Software Dependencies No We evaluate ASM together with several state-of-the-art UDA algorithms on both classification and segmentation tasks using Paddle Paddle and Pytorch. The paper mentions frameworks but does not provide specific version numbers for any software dependencies.
Experiment Setup No More details on experimental settings are given in Appendix A and B. The main text does not include specific hyperparameters or system-level training settings.