Experiments ON Federated Learning ---------Some Questions

• IN FedEM’s code, what is learner, cluster?
• How to add new algorithm to it?

I plan to run 3 algorithms (i.e. FedAvg, FedProx, FedMine) ON 3 datasets (i.e. CIFAR10, CIFAR100, Shakespeare)

In my paper, there should be some figures about the experiment.

• 1 from the distribution about the dataset.
• 4 from the experiment’s result (i.e. Train/Acc, Train/Loss; Test/Acc, Test/Loss), and which one has 3 lines (FedAvg, FedProx, FedMine)

Also, it must contains some tables.

• 1 about the three dataset’s model’s architecture
• 1 about used GPU and coresponding run time on each dataset

Yes, there was some introductions at each bottom of the figures and tables.So , maybe that’s all.

Tips on running:(latter convenients)

• CIFAR10:
  Split:

  python generate_data.py  --n_tasks 80  --n_components 3 --alpha 0.4 --s_frac 1.0 --tr_frac 0.8  --seed 12345
  

  Run:

  python run_experiment.py cifar10 FedAvg --n_learners 1 --n_rounds 200 --bz 64 --lr 0.01  --lr_scheduler multi_step --log_freq 10 --device cuda --optimizer sgd --seed 1234 --verbose 1
  

• CIFAR100:
  Split:

  python generate_data.py --n_tasks 100 --pachinko_allocation_split --alpha 0.4 --beta 10 --s_frac 1.0 --tr_frac 0.8 --seed 12345  
  

  Run:

  python run_experiment.py cifar100 FedAvg --n_learners 1 --n_rounds 200 --bz 128 --lr 0.01 --lr_scheduler multi_step --log_freq 5 --device cuda --optimizer sgd --seed 1234 --verbose 1
  

• ShakeSpeare:
  Split: First, Run ./get_data.sh, then run generate_data.py

  python generate_data.py --s_frac 1.0 --tr_frac 0.8 --seed 12345
  

  Run:

  python run_experiment.py shakespeare FedAvg --n_learners 1 --n_rounds 200 --bz 128 --lr 0.01 --lr_scheduler multi_step --log_freq 5 --device cuda --optimizer sgd --seed 1234 --verbose 1