zzp1012 / federated-learning-environment

federated learning standalone modeling environment

Federated Averaging

Requirements

1. Download the repo on Github FedML
2. Change the root dir in main_fedavg.py to the absolute path of FedML. For example,

       sys.path.insert(0, os.path.abspath("/home/zzp1012/FedML")) # add the root dir of FedML

3. Follow the instruction or documentation of FedML to install required package in python environment.

Experimental Tracking Platform

1. report real-time result to wandb.com, please change ID to your own

   wandb login `Your ID`
   

Experiment Scripts

1. Before any experiments, remember to kill any process occupying port 8999. Simply, you can run the following script:

   lsof -i:8999 # this instruction could show the PID of the process occupying port 8999
   kill PID

2. To test whether program are correctly configured, you can run following commands to see whether training process starts correctly.

   sh begin.sh
   

Or, you can try other heterogeneous distribution (Non-IID) experiment:

## MNIST (non-i.i.d)
sh ./src/run_fedavg_standalone_pytorch.sh 0 10 mnist /home/zzp1012/FedML/data/mnist lr hetero 50 20 0.03 sgd 0 sch_random -v

## shakespeare (non-i.i.d LEAF)
sh ./src/run_fedavg_standalone_pytorch.sh 0 10 shakespeare /home/zzp1012/FedML/data/shakespeare rnn hetero 50 1 0.8 sgd 0 sch_random -v

# fed_shakespeare (non-i.i.d Google)
sh ./src/run_fedavg_standalone_pytorch.sh 0 10 fed_shakespeare /home/zzp1012/FedML/data/fed_shakespeare rnn hetero 50 1 0.8 sgd 0 sch_random -v

## Federated EMNIST (non-i.i.d)
sh ./src/run_fedavg_standalone_pytorch.sh 0 10 femnist /home/zzp1012/FedML/data/FederatedEMNIST cnn hetero 50 1 0.03 sgd 0 sch_random -v

## Fed_CIFAR100 (non-i.i.d)
sh ./src/run_fedavg_standalone_pytorch.sh 0 10 fed_cifar100 /home/zzp1012/FedML/data/fed_cifar100 resnet18_gn hetero 50 1 0.03 adam 0 sch_random -v

# Stackoverflow_LR (non-i.i.d)
sh ./src/run_fedavg_standalone_pytorch.sh 0 10 stackoverflow_lr /home/zzp1012/FedML/data/stackoverflow lr hetero 50 1 0.03 sgd 0 sch_random -v

# Stackoverflow_NWP (non-i.i.d)
sh ./src/run_fedavg_standalone_pytorch.sh 0 10 stackoverflow_nwp /home/zzp1012/FedML/data/stackoverflow rnn hetero 50 1 0.03 sgd 0 sch_random -v
 
# CIFAR10 (non-i.i.d) 
sh ./src/run_fedavg_standalone_pytorch.sh 0 10 cifar10 /home/zzp1012/FedML/data/cifar10 resnet56 hetero 50 1 0.03 sgd 0 sch_random -v

All above datasets are heterogeneous non-i.i.d distributed.

Benchmark Results

We publish benchmark experimental results at wanb.com:
https://app.wandb.ai/automl/fedml