PartialFed: Cross-Domain Personalized Federated Learning via Partial Initialization

Authors: Benyuan Sun, Hongxing Huo, YI YANG, Bo Bai

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate the performance of Partial Fed, on real-world non i.i.d. tasks including cross-domain classification and detection.
Researcher Affiliation Industry Benyuan Sun Hongxing Huo Yi Yang Bo Bai Media Technology Lab, Huawei {sunbenyuan,huohongxing1,yangyi16,baibo3}@huawei.com
Pseudocode Yes Algorithm 1 Partial Fed: Partially-Loaded Federated Learning and Algorithm 2 Partial Fed-Adaptive
Open Source Code No The paper mentions using Py Torch [21] and Detectron2 [34] with their respective licenses, but does not provide a link or statement for the open-sourcing of the authors' own implementation code for Partial Fed.
Open Datasets Yes For federated classification, we use the popular Office-Home [30] dataset... For the detection task, we adopt challenging multi-domain detection dataset UODB [33]
Dataset Splits No Table 5 (UODB description) lists 'Train' and 'Test' data sizes for each dataset, but no explicit validation split or percentages are provided for any dataset, nor for Office-Home.
Hardware Specification No No specific hardware details (e.g., GPU/CPU models, memory) are mentioned for the experimental setup.
Software Dependencies No The paper mentions 'Py Torch [21]' and 'Detectron2 [34]' but does not provide specific version numbers for these software dependencies.
Experiment Setup Yes The temperature parameter τt in Equ. 5 is initialized as 5.0 and annealed to 0 as in [27]. The overall learning scheme is summarized in Alg. 2. For details, we list the hyper parameters in the Appendix.