Multi-Source Collaborative Gradient Discrepancy Minimization for Federated Domain Generalization
Authors: Yikang Wei, Yahong Han
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The extensive experiments on federated domain generalization and adaptation indicate that our method outperforms the state-of-the-art methods significantly. |
| Researcher Affiliation | Academia | Yikang Wei1,2, Yahong Han1,2 1College of Intelligence and Computing, Tianjin University, Tianjin, China 2Tianjin Key Lab of Machine Learning, Tianjin University, Tianjin, China {yikang,yahong}@tju.edu.cn |
| Pseudocode | No | No pseudocode or algorithm blocks are present in the paper. |
| Open Source Code | No | The paper does not provide an explicit statement or a link for the open-source code of their methodology. It only mentions reproducing results using the Mind Spore framework. |
| Open Datasets | Yes | We conduct Fed DG experiments on three image classification datasets including PACS(Li et al. 2017), VLCS(Ghifary et al. 2015), and Office-Home(Venkateswara et al. 2017). ... We conduct Fed DA experiments on two image classification datasets including Digit-5 and Office-Caltech10. Digit-5 contains five domains MNIST(mt), MNIST-M(mm), SVHN(sv), Syn(syn), and USPS(up) of 10 categories. Office-Caltech10 contains four domains Amazon(A), Caltech(C), Webcam(W), and Dslr(D) of 10 categories. |
| Dataset Splits | Yes | PACS contains 9,991 images of 7 categories from four domains: Artpainting (A), Cartoon (C), Photo (P), and Sketch (S), the data of each domain are split into 80% for training and 20% for testing. Office-Home contains 15,500 images of 65 categories from four domains: Artist (A), Clipart (C), Product (P), and Real-world (R), each domain is split into 90% as the training set and 10% as the test set. VLCS contains 10,729 images of 5 categories from four domains: Pascal (P), Label Me (L), Caltech (C), and Sun (S), each domain is split into 80% for training and 20% for testing. The leave-one-domain-out protocol (Zhou et al. 2021b) is used to evaluate the generalization performance on one domain and train the model on the rest source domains. |
| Hardware Specification | Yes | We implement our method with Py Torch and use a single NVIDIA RTX3090. |
| Software Dependencies | No | The paper mentions 'Py Torch' and 'Mind Spore framework (Huawei 2020)' but does not provide specific version numbers for these software components or any other libraries. |
| Experiment Setup | Yes | Following the previous works (Yuan et al. 2023; Wu and Gong 2021), we use the Res Net-18 (He et al. 2016) pretrained on Image Net as the backbone for PACS, Office Home, and VLCS datasets. The SGD optimizer with momentum 0.9 and weight decay 5e-4 is adopted for PACS, Office-Home, and VLCS. For PACS and VLCS, the batch size is 16 and the initial learning rate is 0.001, decayed by the cosine schedule from 0.001 to 0.0001 during training. For Office-Home, the batch size is 30 and the initial learning rate is 0.002, decayed by cosine scheduled from 0.002 to 0.0001. The hyper-parameter λ on Equation 8 is 0.3 for PACS, 0.8 for Office-Home and VLCS. For each local client, we train the local model 1 epoch and then upload the local model to the server side for conducting model aggregation. The local training on the client side and global model aggregation on the server side are conducted iteratively. The total training rounds E is 40 on PACS, VLCS, and Office-Home datasets. All experiments are repeated three times with different random seeds and the mean accuracy (%) is reported. For the federated domain adaptation datasets, we use the three-layer CNN as the backbone for Digit-5 and the Res Net-101 pre-trained on Images Net as the backbone for Office-Caltech10 following the setting of previous work KD3A (Feng et al. 2021). The total training rounds are 50 for Digit-5 and 40 for Office-Caltech10. Following previous work (Feng et al. 2021), we generate the pseudo-labels on the unlabeled target domain by knowledge vote. |