FedP3: Federated Personalized and Privacy-friendly Network Pruning under Model Heterogeneity

Authors: Kai Yi, Nidham Gazagnadou, Peter Richtárik, Lingjuan Lyu

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 4 EXPERIMENTS and subsequent subsections clearly outline empirical studies with data analysis and performance metrics. For example, We utilize benchmark datasets CIFAR10/100 Krizhevsky et al. (2009), a subset of EMNIST labeled EMNIST-L Cohen et al. (2017), and Fashion MNIST Xiao et al. (2017), maintaining standard train/test splits as in Mc Mahan et al. (2017) and Li et al. (2020b). and Figure 2 presents a comparison of different layer overlapping strategies.
Researcher Affiliation Collaboration 1KAUST, 2Sony AI
Pseudocode Yes Algorithm 1 Fed P3
Open Source Code No The paper mentions Given that the only partially open-source code available is from FjORD, referring to a third-party project, but does not provide any statement or link indicating that the authors' own source code for Fed P3 is publicly available.
Open Datasets Yes We utilize benchmark datasets CIFAR10/100 Krizhevsky et al. (2009), a subset of EMNIST labeled EMNIST-L Cohen et al. (2017), and Fashion MNIST Xiao et al. (2017), maintaining standard train/test splits as in Mc Mahan et al. (2017) and Li et al. (2020b).
Dataset Splits No While CIFAR100 has 100 labels, the others have 10, with a consistent data split of 70% for training and 30% for testing. The paper specifies train/test splits but does not explicitly mention a separate validation split.
Hardware Specification Yes Our experiments were conducted on NVIDIA A100 or V100 GPUs, depending on their availability in our cluster.
Software Dependencies Yes The framework was implemented in Py Torch 1.4.0 and torchvision 0.5.0 within a Python 3.8 environment.
Experiment Setup Yes We standardized the experiments to 500 epochs with a local training batch size of 48. The number of local updates was set at 10 to assess final performance. For the learning rate, we conducted a grid search, exploring a range from 10 5 to 0.1, with a fivefold increase at each step.