Frameworks, fusion methods, and Python versions

Last updated: Oct 09, 2024

These are the available machine learning model frameworks and model fusion methods for the Federated Learning model. The software spec and frameworks are also compatible with specific Python versions.

Frameworks and fusion methods

This table lists supported software frameworks for building Federated Learning models. For each framework you can see the supported model types, fusion methods, and hyperparameter options.

Table 1. Frameworks and fusion methods
Frameworks	Model Type	Fusion Method	Description	Hyperparameters
TensorFlow Used to build neural networks. See Save the Tensorflow model.	Any	Simple Avg	Simplest aggregation that is used as a baseline where all parties' model updates are equally weighted.	- Rounds - Termination predicate (Optional) - Quorum (Optional) - Max Timeout (Optional)
		Weighted Avg	Weights the average of updates based on the number of each party sample. Use with training data sets of widely differing sizes.	- Rounds - Termination predicate (Optional) - Quorum (Optional) - Max Timeout (Optional)
Scikit-learn Used for predictive data analysis. See Save the Scikit-learn model.	Classification	Simple Avg	Simplest aggregation that is used as a baseline where all parties' model updates are equally weighted.	- Rounds - Termination predicate (Optional)
		Weighted Avg	Weights the average of updates based on the number of each party sample. Use with training data sets of widely differing sizes.	- Rounds - Termination predicate (Optional)
	Regression	Simple Avg	Simplest aggregation that is used as a baseline where all parties' model updates are equally weighted.	Rounds
		Weighted Avg	Weights the average of updates based on the number of each party sample. Use with training data sets of widely differing sizes.	Rounds
	XGBoost	XGBoost Classification	Use to build classification models that use XGBoost.	- Learning rate - Loss - Rounds - Number of classes
		XGBoost Regression	Use to build regression models that use XGBoost.	- Learning rate - Rounds - Loss
		K-Means/SPAHM	Used to train KMeans (unsupervised learning) models when parties have heterogeneous data sets.	- Max Iter - N cluster
Pytorch Used for training neural network models. See Save the Pytorch model.	Any	Simple Avg	Simplest aggregation that is used as a baseline where all parties' model updates are equally weighted.	- Rounds - Epochs - Quorum (Optional) - Max Timeout (Optional)
	Neural Networks	Probabilistic Federated Neural Matching (PFNM)	Communication-efficient method for fully connected neural networks when parties have heterogeneous data sets.	- Rounds - Termination accuracy (Optional) - Epochs - sigma - sigma0 - gamma - iters

Software specifications and Python version by framework

This table lists the software spec and Python versions available for each framework.

Software specifications and Python version by framework
Watson Studio frameworks	Python version	Software Spec	Python Client Extras	Framework package
scikit-learn	3.10	runtime-22.2-py3.10	fl-rt22.2-py3.10	scikit-learn 1.1.1
Tensorflow	3.10	runtime-22.2-py3.10	fl-rt22.2-py3.10	tensorflow 2.9.2
PyTorch	3.10	runtime-22.2-py3.10	fl-rt22.2-py3.10	torch 1.12.1

Learn more

Hyperparameter definitions

Parent topic: IBM Federated Learning