Multi-level Consistency Learning for Semi-supervised Domain Adaptation

Authors: Zizheng Yan, Yushuang Wu, Guanbin Li, Yipeng Qin, Xiaoguang Han, Shuguang Cui

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

Reproducibility Variable Result LLM Response
Research Type Experimental Empirically, we verified the effectiveness of our MCL framework on three popular SSDA benchmarks, i.e., Vis DA2017, Domain Net, and Office-Home datasets, and the experimental results demonstrate that our MCL framework achieves the state-of-the-art performance.
Researcher Affiliation Academia 1Shenzhen Research Institute of Big Data, The Chinese University of Hong Kong, Shenzhen 2Sun Yat-sen University 3Cardiff University {zizhengyan@link, yushuangwu@link, hanxiaoguang@, shuguangcui@}.cuhk.edu.cn, liguanbin@mail.sysu.edu.cn, qiny16@cardiff.ac.uk
Pseudocode No No pseudocode or algorithm block was found.
Open Source Code Yes Code is available at https://github.com/chester256/MCL.
Open Datasets Yes We evaluate our proposed MCL on several popular benchmark datasets, including Vis DA2017 [Peng et al., 2017], Domain Net [Peng et al., 2019], and Office Home [Venkateswara et al., 2017].
Dataset Splits No No explicit train/validation/test split percentages or counts were provided; it mentions '1-shot and 3-shot experiments' referring to the number of labeled target samples per class.
Hardware Specification No No specific hardware details (e.g., GPU/CPU models, memory amounts) used for running experiments were found. Only a general mention of 'High Performance Computing Services' was present.
Software Dependencies No The paper mentions 'Pytorch' and 'POT' but does not provide specific version numbers for these software dependencies.
Experiment Setup Yes Similar to [Li et al., 2021a], the threshold τ (Eq. 12) is set as 0.95, and we use the softmax temperature T to control the sharpness of the prediction for the thresholding operation (1 for Domainnet, 1.25 for Office-Home and Vis DA). The loss weight balancing hyperparameters λ1 is set as 1, and λ2 is set to 1 for Domain Net, 0.2 for Office-Home , and 0.1 for Vis DA. We use Random Flip and Random Crop as the augmentation methods for view A and Rand Augment [Cubuk et al., 2020] for view B. Moreover, the momentum m used to update source prototypes is set to 0.9.