def test_search_in_repository_by_model_id_correct(mock_init_tree): repo = ModelTypesRepository() model_names, _ = repo.search_models(desired_ids=[ModelGroupsIdsEnum.all]) assert ModelTypesIdsEnum.xgboost in model_names assert len(model_names) == 3
def _eval_strategy_for_task(model_type: ModelTypesIdsEnum, task_type_for_data: TaskTypesEnum): strategies_for_tasks = { MachineLearningTasksEnum.classification: [SkLearnClassificationStrategy, AutoMLEvaluationStrategy], MachineLearningTasksEnum.regression: [SkLearnRegressionStrategy], MachineLearningTasksEnum.auto_regression: [StatsModelsAutoRegressionStrategy], MachineLearningTasksEnum.clustering: [SkLearnClusteringStrategy] } models_for_strategies = { SkLearnClassificationStrategy: [ ModelTypesIdsEnum.xgboost, ModelTypesIdsEnum.knn, ModelTypesIdsEnum.logit, ModelTypesIdsEnum.dt, ModelTypesIdsEnum.rf, ModelTypesIdsEnum.mlp, ModelTypesIdsEnum.lda, ModelTypesIdsEnum.qda ], AutoMLEvaluationStrategy: [ModelTypesIdsEnum.tpot, ModelTypesIdsEnum.h2o], SkLearnClusteringStrategy: [ModelTypesIdsEnum.kmeans], SkLearnRegressionStrategy: [ ModelTypesIdsEnum.linear, ModelTypesIdsEnum.ridge, ModelTypesIdsEnum.lasso ], StatsModelsAutoRegressionStrategy: [ModelTypesIdsEnum.ar, ModelTypesIdsEnum.arima] } models_repo = ModelTypesRepository() _, model_info = models_repo.search_models(desired_ids=[model_type]) task_type_for_model = task_type_for_data task_types_acceptable_for_model = model_info[0].task_type # if the model can't be used directly for the task type from data if task_type_for_model not in task_types_acceptable_for_model: # search the supplementary task types, that can be included in chain which solves original task globally_compatible_task_types = compatible_task_types( task_type_for_model) compatible_task_types_acceptable_for_model = list( set(task_types_acceptable_for_model).intersection( set(globally_compatible_task_types))) if len(compatible_task_types_acceptable_for_model) == 0: raise ValueError( f'Model {model_type} can not be used as a part of {task_type_for_model}.' ) task_type_for_model = compatible_task_types_acceptable_for_model[0] eval_strategies = strategies_for_tasks[task_type_for_model] for strategy in eval_strategies: if model_type in models_for_strategies[strategy]: eval_strategy = strategy(model_type) return eval_strategy return None
def test_search_in_repository_by_metainfo_correct(mock_init_tree): repo = ModelTypesRepository() model_names, _ = repo.search_models(desired_metainfo=ModelMetaInfoTemplate( input_type=NumericalDataTypesEnum.table, output_type=CategoricalDataTypesEnum.vector, task_type=MachineLearningTasksEnum.classification)) assert ModelTypesIdsEnum.knn in model_names assert len(model_names) == 3
def test_search_in_repository_by_id_and_metainfo_correct(mock_init_tree): repo = ModelTypesRepository() model_names, _ = repo.search_models( desired_ids=[ModelGroupsIdsEnum.ml], desired_metainfo=ModelMetaInfoTemplate( task_type=MachineLearningTasksEnum.regression)) assert ModelTypesIdsEnum.xgboost in model_names assert len(model_names) == 1
def test_gp_composer_quality(data_fixture, request): random.seed(1) data = request.getfixturevalue(data_fixture) dataset_to_compose = data dataset_to_validate = data models_repo = ModelTypesRepository() available_model_types, _ = models_repo.search_models( desired_metainfo=ModelMetaInfoTemplate( input_type=NumericalDataTypesEnum.table, output_type=CategoricalDataTypesEnum.vector, task_type=MachineLearningTasksEnum.classification, can_be_initial=True, can_be_secondary=True)) metric_function = MetricsRepository().metric_by_id( ClassificationMetricsEnum.ROCAUC) baseline = baseline_chain() baseline.fit_from_scratch(input_data=dataset_to_compose) predict_baseline = baseline.predict(dataset_to_validate).predict dataset_to_compose.target = np.array( [int(round(i)) for i in predict_baseline]) composer_requirements = GPComposerRequirements( primary=available_model_types, secondary=available_model_types, max_arity=2, max_depth=3, pop_size=5, num_of_generations=5, crossover_prob=0.8, mutation_prob=0.8) # Create GP-based composer composer = GPComposer() composed_chain = composer.compose_chain( data=dataset_to_compose, initial_chain=None, composer_requirements=composer_requirements, metrics=metric_function) composed_chain.fit_from_scratch(input_data=dataset_to_compose) predict_composed = composed_chain.predict(dataset_to_validate).predict roc_auc_chain_created_by_hand = roc_auc(y_true=dataset_to_validate.target, y_score=predict_baseline) roc_auc_chain_evo_alg = roc_auc(y_true=dataset_to_validate.target, y_score=predict_composed) print("model created by hand prediction:", roc_auc_chain_created_by_hand) print("gp composed model prediction:", roc_auc_chain_evo_alg) assert composed_chain == baseline or composed_chain != baseline and abs( roc_auc_chain_created_by_hand - roc_auc_chain_evo_alg) < 0.01
def run_credit_scoring_problem(train_file_path, test_file_path, max_lead_time: datetime.timedelta = datetime.timedelta(minutes=5), gp_optimiser_params: Optional[GPChainOptimiserParameters] = None, pop_size=None, generations=None): dataset_to_compose = InputData.from_csv(train_file_path) dataset_to_validate = InputData.from_csv(test_file_path) models_repo = ModelTypesRepository() available_model_types, _ = models_repo.search_models( desired_metainfo=ModelMetaInfoTemplate(input_type=NumericalDataTypesEnum.table, output_type=CategoricalDataTypesEnum.vector, task_type=[MachineLearningTasksEnum.classification, MachineLearningTasksEnum.clustering], can_be_initial=True, can_be_secondary=True)) # the choice of the metric for the chain quality assessment during composition metric_function = MetricsRepository().metric_by_id(ClassificationMetricsEnum.ROCAUC) if gp_optimiser_params: optimiser_parameters = gp_optimiser_params else: selection_types = [SelectionTypesEnum.tournament] crossover_types = [CrossoverTypesEnum.subtree] mutation_types = [MutationTypesEnum.simple, MutationTypesEnum.growth, MutationTypesEnum.reduce] regularization_type = RegularizationTypesEnum.decremental optimiser_parameters = GPChainOptimiserParameters(selection_types=selection_types, crossover_types=crossover_types, mutation_types=mutation_types, regularization_type=regularization_type) composer_requirements = GPComposerRequirements( primary=available_model_types, secondary=available_model_types, max_arity=4, max_depth=3, pop_size=pop_size, num_of_generations=generations, crossover_prob=0.8, mutation_prob=0.8, max_lead_time=max_lead_time) # Create GP-based composer composer = GPComposer() chain_evo_composed = composer.compose_chain(data=dataset_to_compose, initial_chain=None, composer_requirements=composer_requirements, metrics=metric_function, optimiser_parameters=optimiser_parameters, is_visualise=False) chain_evo_composed.fit(input_data=dataset_to_compose, verbose=True) roc_on_valid_evo_composed = calculate_validation_metric(chain_evo_composed, dataset_to_validate) print(f'Composed ROC AUC is {round(roc_on_valid_evo_composed, 3)}') return roc_on_valid_evo_composed, chain_evo_composed, composer
def test_gp_composer_build_chain_correct(data_fixture, request): random.seed(1) np.random.seed(1) data = request.getfixturevalue(data_fixture) dataset_to_compose = data dataset_to_validate = data models_repo = ModelTypesRepository() available_model_types, _ = models_repo.search_models( desired_metainfo=ModelMetaInfoTemplate( input_type=NumericalDataTypesEnum.table, output_type=CategoricalDataTypesEnum.vector, task_type=MachineLearningTasksEnum.classification, can_be_initial=True, can_be_secondary=True)) metric_function = MetricsRepository().metric_by_id( ClassificationMetricsEnum.ROCAUC) gp_composer = GPComposer() req = GPComposerRequirements(primary=available_model_types, secondary=available_model_types, max_arity=2, max_depth=2, pop_size=2, num_of_generations=1, crossover_prob=0.4, mutation_prob=0.5) chain_gp_composed = gp_composer.compose_chain(data=dataset_to_compose, initial_chain=None, composer_requirements=req, metrics=metric_function) chain_gp_composed.fit_from_scratch(input_data=dataset_to_compose) predicted_gp_composed = chain_gp_composed.predict(dataset_to_validate) roc_on_valid_gp_composed = roc_auc(y_true=dataset_to_validate.target, y_score=predicted_gp_composed.predict) assert roc_on_valid_gp_composed > 0.6
def test_random_composer(data_fixture, request): random.seed(1) np.random.seed(1) data = request.getfixturevalue(data_fixture) dataset_to_compose = data dataset_to_validate = data models_repo = ModelTypesRepository() available_model_types, _ = models_repo.search_models( desired_metainfo=ModelMetaInfoTemplate( input_type=NumericalDataTypesEnum.table, output_type=CategoricalDataTypesEnum.vector, task_type=MachineLearningTasksEnum.classification, can_be_initial=True, can_be_secondary=True)) metric_function = MetricsRepository().metric_by_id( ClassificationMetricsEnum.ROCAUC) random_composer = RandomSearchComposer(iter_num=1) req = ComposerRequirements(primary=available_model_types, secondary=available_model_types) chain_random_composed = random_composer.compose_chain( data=dataset_to_compose, initial_chain=None, composer_requirements=req, metrics=metric_function) chain_random_composed.fit_from_scratch(input_data=dataset_to_compose) predicted_random_composed = chain_random_composed.predict( dataset_to_validate) roc_on_valid_random_composed = roc_auc( y_true=dataset_to_validate.target, y_score=predicted_random_composed.predict) assert roc_on_valid_random_composed > 0.6
full_path_train = os.path.join(str(project_root()), file_path_train) dataset_to_compose = InputData.from_csv(full_path_train, task_type=problem_class) # a dataset for a final validation of the composed model file_path_test = 'cases/data/ts/metocean_data_test.csv' full_path_test = os.path.join(str(project_root()), file_path_test) dataset_to_validate = InputData.from_csv(full_path_test, task_type=problem_class) # the search of the models provided by the framework that can be used as nodes in a chain for the selected task models_repo = ModelTypesRepository() available_model_types, _ = models_repo.search_models( desired_metainfo=ModelMetaInfoTemplate( input_type=NumericalDataTypesEnum.table, output_type=CategoricalDataTypesEnum.vector, task_type=problem_class, can_be_initial=True, can_be_secondary=True)) # the choice of the metric for the chain quality assessment during composition metric_function = MetricsRepository().metric_by_id(RegressionMetricsEnum.RMSE) # the choice and initialisation single_composer_requirements = ComposerRequirements( primary=[ModelTypesIdsEnum.ar], secondary=[]) chain_single = DummyComposer(DummyChainTypeEnum.flat).compose_chain( data=dataset_to_compose, initial_chain=None, composer_requirements=single_composer_requirements,
def run_credit_scoring_problem(train_file_path, test_file_path, max_lead_time: datetime.timedelta = datetime.timedelta(minutes=20), gp_optimiser_params: Optional[GPChainOptimiserParameters] = None): dataset_to_compose = InputData.from_csv(train_file_path) dataset_to_validate = InputData.from_csv(test_file_path) # the search of the models provided by the framework that can be used as nodes in a chain for the selected task models_repo = ModelTypesRepository() available_model_types, _ = models_repo.search_models( desired_metainfo=ModelMetaInfoTemplate(input_type=NumericalDataTypesEnum.table, output_type=CategoricalDataTypesEnum.vector, task_type=[MachineLearningTasksEnum.classification, MachineLearningTasksEnum.clustering], can_be_initial=True, can_be_secondary=True)) # the choice of the metric for the chain quality assessment during composition metric_function = MetricsRepository().metric_by_id(ClassificationMetricsEnum.ROCAUC) if gp_optimiser_params: optimiser_parameters = gp_optimiser_params else: optimiser_parameters = GPChainOptimiserParameters(selection_types=[SelectionTypesEnum.tournament], crossover_types=[CrossoverTypesEnum.subtree], mutation_types=[MutationTypesEnum.growth], regularization_type=RegularizationTypesEnum.decremental, chain_generation_function=random_ml_chain, crossover_types_dict=crossover_by_type, mutation_types_dict=mutation_by_type) composer_requirements = GPComposerRequirements( primary=available_model_types, secondary=available_model_types, max_arity=4, max_depth=3, pop_size=5, num_of_generations=5, crossover_prob=0.8, mutation_prob=0.8, max_lead_time=max_lead_time) # Create GP-based composer composer = GPComposer() # the optimal chain generation by composition - the most time-consuming task chain_evo_composed = composer.compose_chain(data=dataset_to_compose, initial_chain=None, composer_requirements=composer_requirements, metrics=metric_function, optimiser_parameters=optimiser_parameters, is_visualise=False) chain_evo_composed.fit(input_data=dataset_to_compose, verbose=True) # the choice and initialisation of the dummy_composer dummy_composer = DummyComposer(DummyChainTypeEnum.hierarchical) chain_static = dummy_composer.compose_chain(data=dataset_to_compose, initial_chain=None, composer_requirements=composer_requirements, metrics=metric_function, is_visualise=True) chain_static.fit(input_data=dataset_to_compose, verbose=True) # the single-model variant of optimal chain single_composer_requirements = ComposerRequirements(primary=[ModelTypesIdsEnum.xgboost], secondary=[]) chain_single = DummyComposer(DummyChainTypeEnum.flat).compose_chain(data=dataset_to_compose, initial_chain=None, composer_requirements=single_composer_requirements, metrics=metric_function) chain_single.fit(input_data=dataset_to_compose, verbose=True) print("Composition finished") ComposerVisualiser.visualise(chain_static) ComposerVisualiser.visualise(chain_evo_composed) # the quality assessment for the obtained composite models roc_on_valid_static = calculate_validation_metric(chain_static, dataset_to_validate) roc_on_valid_single = calculate_validation_metric(chain_single, dataset_to_validate) roc_on_valid_evo_composed = calculate_validation_metric(chain_evo_composed, dataset_to_validate) print(f'Composed ROC AUC is {round(roc_on_valid_evo_composed, 3)}') print(f'Static ROC AUC is {round(roc_on_valid_static, 3)}') print(f'Single-model ROC AUC is {round(roc_on_valid_single, 3)}') return (roc_on_valid_evo_composed, chain_evo_composed), (chain_static, roc_on_valid_static), ( chain_single, roc_on_valid_single)