Semi-Supervised Single Domain Generalization with Label-Free Adversarial Data Augmentation
Authors: Ronghang Zhu, Xiang Yu, Sheng Li
TMLR 2023 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on Office Home, PACS and Domain Net20 datasets show that our method surpasses the latest SDG and semi-supervised methods. Moreover, on PACS and Domain Net20, our method approaches the fully supervised ERM upper bound within 5% gap, but only uses less than 8% of the labels. |
| Researcher Affiliation | Collaboration | Ronghang Zhu EMAIL University of Georgia Xiang Yu EMAIL Amazon Sheng Li EMAIL University of Virginia |
| Pseudocode | Yes | Algorithm 1 Our proposed Algorithm. Require: Ds, Dg, Dw g , Ds g, initialized G, F, C. Ensure: Learned G and classifier C 1: for t = 1 to TP do TP #iterations 2: Apply Eqn. 2 for Pre-train (Sec. 3.2) 3: if Mod(t,Q)=0 then Q iteration interval 4: Apply Eqn. 8 for Data Augmentation (Sec. 3.3) 5: end if 6: end for 7: for t = 1 to TJ do TJ #iterations 8: Group xs g, xw g , x, xl 9: Apply S-Aug on xs g, xw g , x and W-Aug on x, xl 10: Apply overall loss Eqn. 11 to train G, C 11: end for |
| Open Source Code | No | The paper does not provide an explicit statement or a link to their source code for the methodology described. |
| Open Datasets | Yes | (1) PACS (Li et al., 2017) is a recent challenging domain adaptation/generalization benchmark... (2) Office Home (Venkateswara et al., 2017) contains four domains... (3) Domain Net20 is a subset of Domain Net. |
| Dataset Splits | Yes | For this dataset, we evaluate on two SS-SDG settings: 15 labeled samples per class (total 105 labels) and 25 labeled samples per class (total 175 labels). ... We design two SS-SDG settings on this dataset: 10 labeled samples per class (total 650 labels) and 15 labeled samples per class (total 975 labels). ... We adopt two SS-SDG settings: 15 labeled samples per class (total 300 labels) and 25 labeled samples per class (total 500 labels). ... For each of the benchmarks, amongst all the domains defined, we iteratively take one domain as the source domain and test on all the rest domains. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for experiments, such as GPU or CPU models. |
| Software Dependencies | No | The paper mentions "We adopt an Image Net-pretrained Res Net18 as the feature extractor G" but does not specify software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions). |
| Experiment Setup | Yes | For our framework: at stage1 in Sec. 3.2 and 3.3, We adopt an Image Net-pretrained Res Net18 as the feature extractor G and a 2-layer MLP head (hidden layer 512-d, with Re LU) as projection head F. We set memory bank size to 1600 and batch size to 32 in whole training process. At the model pre-train stage in Sec. 3.2, we use SGD optimizer with learning rate 0.0005, weight decay 0.0005 and momentum 0.9, and train for 1500 iterations. At data augmentation stage in Sec. 3.3, we adopt SGD optimizer with learning rate 50.0 and 15 iterations to maximize Equation 7. The data augmentation is involved only once at the 200-th iteration during model pre-train. At Stage 2 in Sec. 3.4, we adopt SGD with learning rate 0.001, weight decay 0.0005, momentum 0.9, batch size 32 and train for 8500 iterations. ... where λ,γ are balancing hyper-parameters and empirically set as 1 and 0.5, respectively. |