FedBABU: Toward Enhanced Representation for Federated Image Classification

Authors: Jaehoon Oh, SangMook Kim, Se-Young Yun

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments show consistent performance improvements and an efficient personalization of Fed BABU. Table 1 describes the accuracy of Fed Avg on CIFAR100 according to different FL settings (f, τ, and s) with 100 clients.
Researcher Affiliation Academia Jaehoon Oh Graduate School of KSE, KAIST jhoon.oh@kaist.ac.kr Sangmook Kim , Se-Young Yun Graduate School of AI, KAIST {sangmook.kim, yunseyoung}@kaist.ac.kr
Pseudocode Yes Algorithm 1 Training procedure of Fed BABU.
Open Source Code Yes The code is available at https://github.com/jhoon-oh/Fed BABU.
Open Datasets Yes We used two public datasets, the CIFAR (Krizhevsky et al., 2009) and EMNIST (Cohen et al., 2017) , for performance evaluation.
Dataset Splits Yes We set the number of clients to 100 and then each client has 500 training data and 100 test data; the classes in the training and test data sets are the same.
Hardware Specification Yes For computation costs, we used a single TITAN RTX and the entire training time of Fed BABU on CIFAR100 using Mobile Net takes 2 hours when f=1.0 and τ=1.
Software Dependencies No The paper mentions 'Py Torch (Paszke et al., 2019)' but does not specify a version number for PyTorch or other software dependencies.
Experiment Setup Yes We set the number of clients to 100 and then each client has 500 training data and 100 test data; the classes in the training and test data sets are the same. We control FL environments with three hyperparameters: client fraction ratio f, local epochs τ, and shards per user s. ... The learning rate starts with 0.1 and is decayed by a factor of 0.1 at half and three-quarters of total updates. ... fine-tuning with the fine-tuning epochs of τf; Here, one epoch is equal to 10 updates in our case because each client has 500 training samples and the batch size is 50.