A Separation and Alignment Framework for Black-Box Domain Adaptation
Authors: Mingxuan Xia, Junbo Zhao, Gengyu Lyu, Zenan Huang, Tianlei Hu, Gang Chen, Haobo Wang
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate that our proposed method outperforms best baselines on benchmark datasets, e.g. improving the averaged per-class accuracy by 4.1% on the Vis DA dataset. |
| Researcher Affiliation | Academia | 1School of Software Technology, Zhejiang University 2State Key Laboratory of Blockchain and Data Security, Zhejiang University 3Faculty of Information Technology, Beijing University of Technology |
| Pseudocode | Yes | The pseudo-code is summarized in Appendix. |
| Open Source Code | Yes | The source code is available at: https: //github.com/Mingxuan Xia/SEAL. |
| Open Datasets | Yes | Datasets. Office (Saenko et al. 2010) is the most popular image recognition benchmark for UDA. It contains 31 categories in three domains, namely Amazon, DSLR, and Webcam, with 4,652 images in total. Office-Home (Venkateswara et al. 2017) is a more challenging medium-sized dataset with 65 classes from four domains: Art, Clipart, Product, and Real-World, which contains 15,500 images. Vis DA (Peng et al. 2017) is a large-scaled synthetic-to-real dataset for computer vision. It contains 12 categories with 152k synthetic images in the source domain and 55k real object images from Microsoft COCO in the target domain. Domain Net (Peng et al. 2019) is the largest domain adaptation benchmark with over 596k images from 6 domains in 345 categories. |
| Dataset Splits | No | The paper describes the datasets used (Office, Office-Home, Vis DA, Domain Net) but does not provide specific train/validation/test dataset splits (e.g., percentages or sample counts) needed to reproduce the data partitioning. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., exact GPU/CPU models, memory amounts, or detailed computer specifications) used for running its experiments. |
| Software Dependencies | No | The paper mentions software components like ResNet backbones, SGD optimizer, SimAugment, and RandAugment, but does not specify version numbers for any key software libraries or dependencies (e.g., Python, PyTorch, CUDA versions). |
| Experiment Setup | Yes | The model is trained using a standard SGD optimizer with a learning rate of ϵ0 = 1e 3 for the backbone and ϵ0 = 1e 2 for other layers. The momentum, weight decay, and batch size are set to 0.9, 1e 3, and 64 respectively, and the learning rate schedule is set as ϵ = ϵ0 (1 + 10 e E ) 0.75. The total training epochs E is set to 100 for all datasets except 20 for Vis DA and 50 for Domain Net. For hyper-parameters, we fix the high threshold τ, temperature t, and contrastive loss weight η as 0.95, 0.07, and 0.5 respectively. The number of neighbors k is 5/3/100 for Office/Office-Home/Vis DA. |