Working with the med3pa Subpackage

This tutorial guides you through the process of setting up and running comprehensive experiments using the med3pa subpackage. It includes steps to execute MED3pa experiment with Med3paExperiment and the combination of MED3pa and Detectron using Med3paDetectronExperiment.

Running the MED3pa Experiment

Step 1: Setting up the Datasets

First, configure the DatasetsManager. In the case of MED3pa only experiment you only need to set the DatasetManager with either testing and reference dataset:

from MED3pa.datasets import DatasetsManager

# Initialize the DatasetsManager
datasets = DatasetsManager()

# Load datasets for reference, and testing
datasets.set_from_file(dataset_type="reference", file='./path_to_reference_data.csv', target_column_name='Outcome')
datasets.set_from_file(dataset_type="testing", file='./path_to_test_data.6.csv', target_column_name='Outcome')

Step 2: Configuring the Model

Next, utilize the ModelFactory to load a pre-trained model, and set it as the base model for the experiment. Alternatively, you can train your own model and use it.

from MED3pa.models import BaseModelManager, ModelFactory

# Initialize the model factory and load the pre-trained model
factory = ModelFactory()
model = factory.create_model_from_pickled("./path_to_model.pkl")

# Set the base model using BaseModelManager
base_model_manager = BaseModelManager()
base_model_manager.set_base_model(model=model)

Step 3: Running the med3pa Experiment

Execute the MED3PA experiment with the specified datasets and base model. You can also specify other parameters as needed. See the documentation of the subpackage for more information about the parameters.

The experiment outputs two structure one for the reference set and the other for the testing set, both containing files indicating the extracted profiles at different declaration rates, the performance of the model on these profiles..etc.

from MED3pa.med3pa import Med3paExperiment
from MED3pa.med3pa.uncertainty import AbsoluteError
from MED3pa.models.concrete_regressors import RandomForestRegressorModel

# Define parameters for the experiment
ipc_params = {'n_estimators': 100}
apc_params = {'max_depth': 3}
med3pa_metrics = ['Auc', 'Accuracy', 'BalancedAccuracy']

# Execute the MED3PA experiment
ipc_params = {'n_estimators': 100}
apc_params = {'max_depth': 3}
med3pa_metrics = ['Auc', 'Accuracy', 'BalancedAccuracy']

# Execute the MED3PA experiment
reference_results, test_results = Med3paExperiment.run(
                                    datasets_manager=datasets,
                                    base_model_manager=base_model_manager,
                                    uncertainty_metric=AbsoluteError,
                                    ipc_type=RandomForestRegressorModel,
                                    ipc_params=ipc_params,
                                    apc_params=apc_params,
                                    samples_ratio_min=0,
                                    samples_ratio_max=10,
                                    samples_ratio_step=5,
                                    med3pa_metrics=med3pa_metrics,
                                    evaluate_models=True,
                                    models_metrics=['MSE', 'RMSE']
                                )

Step 4: Analyzing and Saving the Results

After running the experiment, you can analyze and save the results using the returned Med3paResults instance.

# Save the results to a specified directory
reference_results.save(file_path='./med3pa_experiment_results/reference')
test_results.save(file_path='./med3pa_experiment_results/test')

Running the MED3pa and Detectron Experiment

You can also run an experiment that combines the forces of Detectron in covariate shift detection with MED3pa problematic profiles extraction using the Med3paDetectronExperiment class. To be able to run this experiment, all datasets of the DatasetsManager should be set, alongside the BaseModelManager. This experiment will run MED3pa experiment on the testing and reference sets and then run the detectron experiment on the testing set as a whole, and then on the extracted profiles from MED3pa:

from MED3pa.med3pa import Med3paDetectronExperiment
from MED3pa.detectron.strategies import EnhancedDisagreementStrategy

# Execute the integrated MED3PA and Detectron experiment
reference_results, test_results, detectron_results = Med3paDetectronExperiment.run(
    datasets=datasets,
    base_model_manager=base_model_manager,
    uncertainty_metric=AbsoluteError,
    samples_size=20,
    ensemble_size=10,
    num_calibration_runs=100,
    patience=3,
    test_strategies=EnhancedDisagreementStrategy,
    allow_margin=False,
    margin=0.05,
    ipc_params=ipc_params,
    apc_params=apc_params,
    samples_ratio_min=0,
    samples_ratio_max=50,
    samples_ratio_step=5,
    med3pa_metrics=med3pa_metrics,
    evaluate_models=True,
    models_metrics=['MSE', 'RMSE']
)

# Save the results to a specified directory
reference_results.save(file_path='./med3pa_detectron_experiment_results/reference')
test_results.save(file_path='./med3pa_detectron_experiment_results/test')
detectron_results.save(file_path='./med3pa_detectron_experiment_results/detectron')