Implementation of the algorithm presented in the paper titled "PHiFL-TL: Personalized Hierarchical Federated Learning using Transfer Learning" with Tensorflow.

• Here is one example to run this code (IID MNIST Scenario):

    dataset="mnist"
    flag1=1
    model="cnn1"  
    batch_size=32
    communication_round=6          
    epochs=20                         
    num_edge_aggregation=4           
    num_edges=3   
    num_clients=30 
    fraction_clients=0.5              
    lr=0.01      
    val_ratio=0.1     
    image_shape=(28,28,1)
    lr=0.00001      # for Transfer learning phase
  

• Here is one example to run this code (non-IID MNIST Scenario):

    dataset="mnist"
    flag1=3
    k1=4
    k2=2
    model="cnn1"  
    batch_size=32
    communication_round=6          
    epochs=20                         
    num_edge_aggregation=4           
    num_edges=3   
    num_clients=30 
    fraction_clients=0.5              
    lr=0.01
    val_ratio=0.1     
    image_shape=(28,28,1)
    lr=0.00001      # for Transfer learning phase
  

• Here is one example to run this code (non-IID FEMNIST Scenario):

    dataset="femnist"
    num_labels=20   # number classes of 62 classes  
    train_size=21000
    test_size=9000 
    label_reduce=12
    model="cnn1"  
    batch_size=32
    communication_round=6          
    epochs=20                         
    num_edge_aggregation=4           
    num_edges=3   
    num_clients=30 
    fraction_clients=0.5              
    lr=0.01
    val_ratio=0.25     
    image_shape=(28,28,1)
    lr=0.001      # for Transfer learning phase
  

Notice: You need to create the following folders where the program is located: results\global_models, results\edges_models\itr_i (i : 1 to communication_round) and results\fig.

Citation

If you find this repository useful, please cite our paper:

@article{afzali2024phifl,
  title={PHiFL-TL: Personalized Hierarchical Federated Learning using Transfer Learning},
  author={Afzali, Afsaneh and Shamsinejadbabaki, Pirooz},
  journal={Future Generation Computer Systems},
  pages={107672},
  year={2024},
  publisher={Elsevier},
  doi={10.1016/j.future.2024.107672},
}