iter_avg

Running simple average with Keras CNN in IBM federated learning

This example explains how to run federated learning on CNNs implemented with Keras training on MNIST data. Data in this example is preprocessed by scaling down to range from [0, 255] to [0, 1]. No other preprocessing is performed.

Model Setup

This experiment can be run using models with different underlying framework. By default, configs with keras(tf 1.15) based model are generated, but other models like PYTORCH, Scikit Learn, keras(tf 2.1) can be created by changing -m param.

Model Type	Params
Keras (with tf 1.15)	keras
Pytorch	pytorch
Scikit Learn	sklearn
Tensorflow/keras( tf 2.1)	tf

Dataset Setup

Iterative Avg fusion can be run on different datasets by just changing -d param while generating config. Model definition changes as dataset changes, we currently only support below shown combinations.

Dataset	Params	Keras	Pytorch	TF	sklearn
MNIST	mnist	YES	YES	YES	YES
Adult Dataset	adult	NO	NO	NO	YES
Cifar-10	cifar10	YES	NO	NO	NO
FEMNIST	femnist	YES	NO	NO	NO

• Split data by running:

  python examples/generate_data.py -n <num_parties> -d <dataset> -pp <points_per_party>
  

• Generate config files by running:

  python examples/generate_configs.py -f iter_avg -m keras -n <num_parties> -d <dataset> -p <path>
  

• In a terminal running an activated IBM FL environment (refer to Quickstart in our website to learn more about how to set up the running environment), start the aggregator by running:

  python -m ibmfl.aggregator.aggregator <agg_config>
  

  Type START and press enter to start accepting connections

• In a terminal running an activated IBM FL environment, start each party by running:

  python -m ibmfl.party.party <party_config>
  

  Type START and press enter to start accepting connections.

  Type REGISTER and press enter to register the party with the aggregator.

• Finally, start training by entering TRAIN in the aggregator terminal.