CHASe

CHASe: Client Heterogeneity-Aware Data Selection for Effective Federated Active Learning

Requirements

• PyTorch
• NumPy

Files

• options.py: Hyperparameter setting for CHASe.
• sampling.py: Sample the MNIST, EMNIST, CIFAR10, CIFAR-100 and Shakespeare in a IID/NonIID manner.
• utils.py:
  • Construction of labeled, unlabeled and global test sets for sampled dataset；
  • Server's aggregation & Definition of log detail.
• model.py: Models for MNIST, EMNIST, CIFAR10, CIFAR-100 and Shakespeare datasets.
• localtraining.py:
  • Clients' local training ;
    • Quantify Epistemic Variation;
    • Calibrate Decision Boundary;
  • Inference of local & global model.
• slected_strategy.py: Definition of the sampling with EV.
• main: Core code for CHASe, Logic and interaction throughout the pipeline.

Usage

1. Download MNIST, EMNIST, CIFAR10 , CIFAR-100 and Shakespeare datasets or Execute the program default download;
2. Set parameters in options.py;
3. Execute main.py to run the CHASe.