Cross Site Model Evaluation / Federated Evaluation

The cross site model evaluation workflow uses the data from clients to run evaluation with the models of other clients. Data is not shared, rather the collection of models is distributed to each client site to run local validation. The results of local validation are collected by the server to construct an all-to-all matrix of model performance vs. client dataset.

The server’s global model is also distributed to each client for evaluation on the client’s local dataset for global model evaluation.

The hello-numpy-cross-val example is a simple example that implements the cross site model evaluation workflow.


Previously in NVFlare before version 2.0, cross-site validation was built into the framework itself, and there was an admin command to retrieve cross-site validation results. In NVFlare 2.0, with the ability to have customized workflows, cross-site validation is no longer in the NVFlare framework but is instead handled by the workflow. The the cifar10 example is configured to run cross-site model evaluation and config_fed_server.json is configured with ValidationJsonGenerator to write the results to a JSON file on the server.

Examples with Cross Site Model Evaluation / Federated Evaluation Workflow

See Hello Numpy Cross-Site Validation and Step-by-step Cross-site Evaluation for examples using server-controlled cross-site evaluation workflows.