Semi-Supervised Domain Generalization with Known and Unknown Classes
Authors: Lei Zhang, Ji-Fu Li, Wei Wang
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The experiments conducted on real-world datasets verify the effectiveness and superiority of our method. |
| Researcher Affiliation | Academia | Lei Zhang, Ji-Fu Li, Wei Wang National Key Laboratory for Novel Software Technology, Nanjing University, China School of Artificial Intelligence, Nanjing University, China {zhangl,lijf,wangw}@lamda.nju.edu.cn |
| Pseudocode | Yes | The paper includes "Procedure 1 Calculating Thresholds" and "Algorithm 1 Class-Wise Adaptive Exploration and Exploitation (CWAEE)" which are clearly labeled algorithm blocks. |
| Open Source Code | No | The paper does not provide an explicit statement about releasing source code or a link to a code repository. |
| Open Datasets | Yes | We use PACS [45], Office-Home [46] and mini Domain Net [47] datasets in the experiments. |
| Dataset Splits | Yes | We adopt the common leave-one-domain-out protocol [12, 16]: three domains are used as the source domains and the remaining one as the target domain. Similar to [23, 38] we split the classes into known and unknown classes, specifically we split the original label set into 3:2:2, 25:20:20 and 42:42:42 (known classes, seen unknown classes and unseen unknown classes) in PACS [45], Office Home [46] and mini Domain Net [47] respectively in alphabetical order of the class name. On each source domain, 10 labeled samples of each known class are randomly sampled to construct the labeled data, and the remaining samples of known classes and seen unknown classes construct the unlabeled data. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for experiments, such as CPU or GPU models, or memory specifications. It states, "The Image Net-pretrained Res Net18 [2] is used as the CNN backbone," but this refers to the model, not the hardware it was run on. |
| Software Dependencies | No | The paper mentions using a ResNet18 backbone and SGD optimizer, but it does not specify any software libraries or dependencies with version numbers (e.g., PyTorch version, Python version). |
| Experiment Setup | Yes | The initial learning rate of SGD optimizer is set to 0.003 for the pretrained backbone and 0.01 for the randomly initialized stochastic classifier, both decaying following the cosine annealing rule. The running epochs are 40, 20 and 20 for PACS, Office Home and mini Domain Net respectively. For each mini-batch, we randomly sample 16 labeled samples and 16 unlabeled samples from each source domain. We set λ1 = 1.0 on PACS, λ1 = 0.4 on Office Home and λ1 = 0.1 on mini Domain Net. We set λ2 = 0.4 and λ3 = 1.0 for all three datasets. We evaluate the accuracy on known classes and AUROC on unknown classes of the methods with 3 different random seeds, and report the average results. |