Ability to override the default threshold at model import / to use optimal at eval

[paulhenricalas-dku]

To be completed by the according PR on DIP

import_mlflow_version_from_managed_folder and import_mlflow_version_from_path take an optional binary_classification_threshold parameter. In bidimensional classification models, it allows the user to override its 0.5 value.

evaluatetakes an optional use_optimal_threshold parameter, set by default to False. It allows the user to evaluate according to the previously computed optimal threshold at training, depending on the metric set in the saved model.

for example:

settings = saved_model.get_settings()
settings.prediction_metrics_settings['thresholdOptimizationMetric'] = 'ACCURACY'
settings.save()

# ...
model_version = saved_model.import_mlflow_version_from_path("v0", model_dir, binary_classification_threshold=0.2)
# ...

# will evaluate using the optimal threshold according to accuracy
model_version.evaluate("DATASET", use_optimal_threshold=True) 

# will evaluate using a threshold set to 0.2
model_version.evaluate("DATASET") 

[shortcut-integration]

This pull request has been linked to Shortcut Story #95506: Experiment Tracking: Allow to override the default threshold when importing a MLflow binary classification model.