federated-learning-models

In this repo, we aim to predict the remaining useful life (RUL) of turbofan engines from the NASA turbofan engine degradation simulation data set FD001, using federated learning models.

This repo is divided into four main segments:

1. data - contains the train and test data for all workers in the federated learning process
2. script - contains the federated learning pipeline for the FATE model
3. extracted - contains the train and test RUL prediction output of all workers from the FATE model
4. dc_turbofan_models - contains the federated learning pipeline for the dc_federated model
5. dc_extracted - contains the train and test RUL prediction output of all workers from the dc_federated model

FATE

The model for federated Gradient Boosted Decision Tree (GBDT) can be found in pipeline_homo_sbt_regression.py, which runs the customised federated learning pipeline for both training and testing purposes.

pipeline-upload.py supports the modelling process by uploading the necessary data for each worker onto FATE.

The number of workers can be adjusted by updating the variable num_parties in both of these files.

Remember to update the config file config_5.yaml and config_3.yaml to reflect the correct number of guests and hosts, in line with the number of workers.

dc_federated

The model for federated Neural Network (NN) using the federated averaging (FedAvg) algorithm can be found in turbofan_fed_model.py.

Both turbofan_fed_avg_server.py and turbofan_fed_avg_worker.py make calls to the model.

An example call to the server is

python FATE-Ubuntu/dc_turbofan_models/turbofan_fed_avg_server.py --batch-size 24 --learn-rate 0.03 --iter-rounds 12 --layer-one 64 --layer-two 128 --layer-three 256 --drop-out 0 --acti-func sigmoid

An example call to each worker is

python FATE-Ubuntu/dc_turbofan_models/turbofan_fed_avg_worker.py --server-host-ip 127.0.1.1 --server-port 8080 --batch-size 24 --learn-rate 0.03 --iter-rounds 12 --layer-one 64 --layer-two 128 --layer-three 256 --drop-out 0 --acti-func sigmoid --bal-imbal _balanced --party-code 5

The arguments supplied to each of the above command line call determines the hyperparameters that the FedAvg algorithm will be trained on.

This is the same set of commands that is being used in run_optimisation.py which executes a randomised grid search to tune the best performing federated NN model.

References

• FATE - https://github.com/FederatedAI/FATE
• dc_federated - https://github.com/digicatapult/dc-federated