0.44.0rc2

0.44.0rc2

Added

• Add documentation on how to change SubstraFL log level (#194)
• Add the simulate_experiment function, that will execute the Compute Plan in RAM only. It returns Python objects containing the computed Performances and the saved intermediate States. More information about this feature is available in docstrings (#184).

Example of usage:

from substrafl.experiment import simulate_experiment

scores, intermediate_state_train, intermediate_state_agg = simulate_experiment(
    client=my_substra_client,
    strategy=my_strategy,
    train_data_nodes=train_data_nodes,
    evaluation_strategy=my_eval_strategy,
    aggregation_node=aggregation_node,
    clean_models=False,
    num_rounds=NUM_ROUNDS,
)

Changed

• BREAKING: rename datasamplesto data_from_opener (#193)
• Bump documentation dependencies to Sphinx 7.2.6 (#195)
• The predict task does not exist anymore. The evaluation of a model is done in a single task #177
• Strategy implement an evaluate method, with the @remote_data decorator, to compute the evaluation of the model. The evaluate method is the same for all strategies #177
• BREAKING: the perform_predict method of Strategy changed in favor of perform_evaluation that calls the new evaluate method #177
• BREAKING: metric_functions are now passed to the Strategy instead of the TestDataNode #177
• BREAKING: the predict method of Algo has no @remote_data decorator anymore. It signatures does not take prediction_path anymore, and the predictions are return by the method #177
• Abstract base class Node is replaced by Protocols, defined in substrafl.nodes.protocol.py (#185)
• BREAKING: rename test_data_sample_keys, test_tasks and register_test_operations, tasks to data_sample_keys and register_operations in TestDataNodes (#185)
• BREAKING: InputIdentifiers and OutputIdentifiers move from substrafl.nodes.node to substrafl.nodes.schemas (#185)
• Switch to python-slim as base image, instead of substra-tools (#197)
• Dropped support for Python 3.8

Fixed

• Numerical stability of the NewtonRaphson strategy is improved by symmetrizing the Hessian