This repos is the implementation of Improving Federated Learning Personalization via Model Agnostic Meta Learning.
Thanks to the heavy dependency of FedLab and FedLab-benchmarks, my code has already been pulled to https://github.com/SMILELab-FL/FedLab-benchmarks/tree/master/fedlab_benchmarks/perfedavg. this repo is only for displaying README and as the interface for the people who interested in this algorithm implementation 😄.
- Reptile: On First-Order Meta-Learning Algorithms
- MAML: Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks
- LEAF: LEAF: A Benchmark for Federated Settings
Evaluation result after fine-tuned is shown below.
Communication round: 500
Fine-tune: outer loop: 100; inner loop: 10
Personalization round: 5
| FedAvg local training epochs (5 clients) | Initial loss | Initial Acc | Personalized loss | Personalized Acc |
|---|---|---|---|---|
| 20 | 2.3022 | 79.35% | 1.5766 | 84.86% |
| 10 | 1.8387 | 80.53% | 1.1231 | 87.22% |
| 5 | 1.4899 | 83.19% | 0.9809 | 88.97% |
| 2 | 1.4613 | 81.70% | 0.9359 | 88.49% |
| FedAvg local training epochs (20 clients) | Initial loss | Initial Acc | Personalized loss | Personalized Acc |
|---|---|---|---|---|
| 20 | 2.2398 | 82.40% | 0.9756 | 90.29% |
| 10 | 1.6560 | 83.23% | 0.8488 | 90.72% |
| 5 | 1.5485 | 81.48% | 0.7452 | 90.77% |
| 2 | 1.2707 | 82.48% | 0.7139 | 90.48% |
Experiment result from paper is shown below
