Test-Time Style Shifting: Handling Arbitrary Styles in Domain Generalization
Authors: Jungwuk Park, Dong-Jun Han, Soyeong Kim, Jaekyun Moon
ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results on different datasets show the effectiveness of our methods. |
| Researcher Affiliation | Academia | 1School of Electrical Engineering, Korea Advanced Institute of Science and Technology (KAIST), South Korea. 2Purdue University, USA. |
| Pseudocode | Yes | M. Algorithm in Pseudo Code. Algorithm 1 shows the sample selection process in style balancing. The process for test-time style shifting is provided in Algorithm 2. |
| Open Source Code | No | The paper does not contain any explicit statement about open-sourcing its code or providing a link to a code repository. |
| Open Datasets | Yes | Targeting multi-domain classification, we perform experiments using PACS (Li et al., 2017) with 4 domains (Art, Cartoon, Photo, Sketch) and VLCS (Fang et al., 2013) with 4 domains (Caltech, Label Me, Pascal, Sun), which are the commonly adopted benchmarks for DG. We also considered Office-Home (Venkateswara et al., 2017) dataset in Appendix. We adopt Market1501 (Zheng et al., 2015) and GRID (Loy et al., 2009) datasets, and train the model in one dataset and test on the other one. We also compare our scheme with Bo DA (Yang et al., 2022) on Domain Net dataset. |
| Dataset Splits | Yes | We focus on the leave-one-domain-out setting where the model is trained on three domains and tested on the remaining one domain. (Section 5.1); Training-domain validation strategy is used for selecting the model in Domain Bed setup. (Appendix A) |
| Hardware Specification | No | The paper does not specify any particular hardware (e.g., GPU models, CPU types, memory) used for running the experiments. |
| Software Dependencies | No | The paper states its work is 'built upon the official setup of EFDMix' and that 'Other setups are exactly the same as in Mix Style (Zhou et al., 2021), DSU (Li et al., 2022) and EFDMix (Zhang et al., 2022)'. However, it does not explicitly list the specific software dependencies with version numbers within this paper. |
| Experiment Setup | Yes | Our work is built upon the official setup of EFDMix (Zhang et al., 2022). Different from the original setting of EFDMix, for image classification tasks, we trained the model for 150 epochs with a mini-batch size of 128. We also randomly sampled the data from all source domains in each mini-batch. Other setups are exactly the same as in Mix Style (Zhou et al., 2021), DSU (Li et al., 2022) and EFDMix (Zhang et al., 2022) when implementing each module; each module is activated with probability 0.5. Following the original setups, Mixstyle and EFDM are inserted after the 1st, 2nd and 3rd residual blocks for PACS. For other datasets, Mixstyle and EFDM are inserted after the 1st and 2nd residual blocks. DSU is inserted after 1st convolutional layer, max pooling, 1,2,3,4-th residual blocks. Here, our SB module is operated at the moment where Mix Style, DSU, EFDMix are first activated. The TS module is operated at first residual blocks during testing for VLCS, Office-Home and person re-ID task. We set α = 3 for all experiments on image classification tasks, while α = 5 is utilized for person re-ID task. (Appendix N) |