Skip to content

Implementation of Communication-Efficient Learning of Deep Networks from Decentralized Data

License

Notifications You must be signed in to change notification settings

EnfangCui/Federated-Learning-PyTorch

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

63 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Federated-Learning (PyTorch)

Implementation of the vanilla federated learning paper : Communication-Efficient Learning of Deep Networks from Decentralized Data.

Experiments are produced on MNIST, Fashion MNIST and CIFAR10 (both IID and non-IID). In case of non-IID, the data amongst the users can be split equally or unequally.

Since the purpose of these experiments are to illustrate the effectiveness of the federated learning paradigm, only simple models such as MLP and CNN are used.

Requirments

Install all the packages from requirments.txt

  • Python3
  • Pytorch
  • Torchvision

Data

  • Download train and test datasets manually or they will be automatically downloaded from torchvision datasets.
  • Experiments are run on Mnist, Fashion Mnist and Cifar.
  • To use your own dataset: Move your dataset to data directory and write a wrapper on pytorch dataset class.

Running the experiments

The baseline experiment trains the model in the conventional way.

  • To run the baseline experiment with MNIST on MLP using CPU:
python src/baseline_main.py --model=mlp --dataset=mnist --epochs=10
  • Or to run it on GPU (eg: if gpu:0 is available):
python src/baseline_main.py --model=mlp --dataset=mnist --gpu=0 --epochs=10

Federated experiment involves training a global model using many local models.

  • To run the federated experiment with CIFAR on CNN (IID):
python src/federated_main.py --model=cnn --dataset=cifar --gpu=0 --iid=1 --epochs=10
  • To run the same experiment under non-IID condition:
python src/federated_main.py --model=cnn --dataset=cifar --gpu=0 --iid=0 --epochs=10

You can change the default values of other parameters to simulate different conditions. Refer to the options section.

Options

The default values for various paramters parsed to the experiment are given in options.py. Details are given some of those parameters:

  • --dataset: Default: 'mnist'. Options: 'mnist', 'fmnist', 'cifar'
  • --model: Default: 'mlp'. Options: 'mlp', 'cnn'
  • --gpu: Default: None (runs on CPU). Can also be set to the specific gpu id.
  • --epochs: Number of rounds of training.
  • --lr: Learning rate set to 0.01 by default.
  • --verbose: Detailed log outputs. Activated by default, set to 0 to deactivate.
  • --seed: Random Seed. Default set to 1.

Federated Parameters

  • --iid: Distribution of data amongst users. Default set to IID. Set to 0 for non-IID.
  • --num_users:Number of users. Default is 100.
  • --frac: Fraction of users to be used for federated updates. Default is 0.1.
  • --local_ep: Number of local training epochs in each user. Default is 10.
  • --local_bs: Batch size of local updates in each user. Default is 10.
  • --unequal: Used in non-iid setting. Option to split the data amongst users equally or unequally. Default set to 0 for equal splits. Set to 1 for unequal splits.

Results on MNIST

Baseline Experiment:

The experiment involves training a single model in the conventional way.

Parameters:

  • Optimizer: : SGD
  • Learning Rate: 0.01

Table 1: Test accuracy after training for 10 epochs:

Model Test Acc
MLP 92.71%
CNN 98.42%

Federated Experiment:

The experiment involves training a global model in the federated setting.

Federated parameters (default values):

  • Fraction of users (C): 0.1
  • Local Batch size (B): 10
  • Local Epochs (E): 10
  • Optimizer : SGD
  • Learning Rate : 0.01

Table 2: Test accuracy after training for 10 global epochs with:

Model IID Non-IID (equal)
MLP 88.38% 73.49%
CNN 97.28% 75.94%

Further Readings

Papers:

Blog Posts:

About

Implementation of Communication-Efficient Learning of Deep Networks from Decentralized Data

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 100.0%