Choosing your framework and fusion method
From your Federated Learning experiment, you must choose the framework and fusion method for the Federated Learning model.
Choosing the framework
These are the default frameworks provided by Watson Studio:
|Frameworks||Short description||Models types supported|
|Scikit-learn||The Scikit-learn Python Machine Learning package that is used for predictive data analysis and more.||Requires pre-existing untrained model provided by your team’s data scientist and configuration to save the model as a pickle file. - Classification
|TensorFlow 2||TensorFlow 2.x based platform used to build neural networks and more.||Requires a pre-existing untrained model provided by your team’s data scientist in HD5 format. Browse and upload the Model file and give it a name.|
Note: If you choose Scikit-learn as your framework, you must modify the model configuration to save the model in Federated Learning as it is not a feature by default.
- Scikit-learn model configuration for examples of Scikit-learn model files.
- Tensorflow 2 model configuration for examples of Tensorflow 2 model files.
Choosing the fusion method
Based on the framework that you chose, the corresponding fusion method that supports the framework is available for selection:
|Fusion Method||Short description||Models types supported|
|Iterative Average||Simplest aggregation used as a baseline where all parties’ model updates are equally weighted.||Neural Networks, Scikit-learn linear models|
|Weighted average fusion||Weights the average of updates based on the number of each party sample. Use with training datasets of widely differing sizes.||Neural Networks, Scikit-learn linear models|
|XGBoost classification fusion||Use to perform classification tasks using XGBoost.||XGboost classification tasks|
|XGBoost regression fusion||Use to perform regression tasks using XGBoost.||XGboost regression tasks|
|Kmeans fusion (SPAHM)||Use to train KMeans (unsupervised learning) models when parties have heterogeneous datasets.||Kmeans Scikit-learn models|