This is partly the reproduction of the paper of Communication-Efficient Learning of Deep Networks from Decentralized Data
Only experiments on MNIST and CIFAR10 (both IID and non-IID) is produced by far.
Note: The scripts will be slow without the implementation of parallel computing.
The MLP and CNN models are produced by:
python main_nn.py
The testing accuracy of MLP on MINST: 92.14% (10 epochs training) with the learning rate of 0.01. The testing accuracy of CNN on MINST: 98.37% (10 epochs training) with the learning rate of 0.01.
Federated learning with MLP and CNN is produced by:
python main_fed.py
See the arguments in options.py.
For example:
python main_fed.py --dataset mnist --num_channels 1 --model cnn --epochs 50 --gpu 0
Results are shown in Table 1 and Table 2, with the parameters C=0.1, B=10, E=5.
Table 1. results of 10 epochs training with the learning rate of 0.01
| Model | Acc. of IID | Acc. of Non-IID |
|---|---|---|
| FedAVG-MLP | 85.66% | 72.08% |
| FedAVG-CNN | 95.00% | 74.92% |
Table 2. results of 50 epochs training with the learning rate of 0.01
| Model | Acc. of IID | Acc. of Non-IID |
|---|---|---|
| FedAVG-MLP | 84.42% | 88.17% |
| FedAVG-CNN | 98.17% | 89.92% |
python 3.6
pytorch 0.4