def run_by_type(training_df: DataFrame, test_df: DataFrame,
job: Job) -> (dict, dict):
"""runs the specified training/evaluation run
:param training_df: training DataFrame
:param test_df: testing DataFrame
:param job: job configuration
:return: results and predictive_model split
"""
model_split = None
start_time = time.time()
if job.type == JobTypes.PREDICTION.value:
clusterer = _init_clusterer(job.clustering, training_df)
if job.predictive_model.predictive_model == PredictiveModels.CLASSIFICATION.value:
results, model_split = classification(training_df, test_df,
clusterer, job)
elif job.predictive_model.predictive_model == PredictiveModels.REGRESSION.value:
results, model_split = regression(training_df, test_df, clusterer,
job)
elif job.predictive_model.predictive_model == PredictiveModels.TIME_SERIES_PREDICTION.value:
results, model_split = time_series_prediction(
training_df, test_df, clusterer, job)
elif job.type == JobTypes.LABELLING.value:
results = _label_task(training_df)
elif job.type == JobTypes.UPDATE.value:
results, model_split = update_and_test(training_df, test_df, job)
else:
raise ValueError("Type {} not supported".format(job.type))
# TODO: integrateme
if job.type != JobTypes.LABELLING.value:
results['elapsed_time'] = timedelta(
seconds=time.time() - start_time) #todo find better place for this
if job.predictive_model.predictive_model == PredictiveModels.REGRESSION.value:
job.evaluation = Evaluation.init(
job.predictive_model.predictive_model, results)
elif job.predictive_model.predictive_model == PredictiveModels.CLASSIFICATION.value:
job.evaluation = Evaluation.init(
job.predictive_model.predictive_model, results,
len(model_split[ModelType.CLASSIFIER.value][0].classes_) <= 2)
job.evaluation.save()
elif job.type == JobTypes.LABELLING.value:
job.labelling.results = results
job.labelling.save()
if job.type == PredictiveModels.CLASSIFICATION.value: #todo this is an old workaround I should remove this
save_result(results, job, start_time)
print("End job {}, {} .".format(job.type, get_run(job)))
print("\tResults {} .".format(results))
return results, model_split
def run_by_type(training_df: DataFrame, test_df: DataFrame,
job: Job) -> (dict, dict):
"""runs the specified training/evaluation run
:param training_df: training DataFrame
:param test_df: testing DataFrame
:param job: job configuration
:return: results and predictive_model split
"""
model_split = None
# TODO fixme this needs to be fixed in the interface
# if job['incremental_train']['base_model'] is not None:
# job['type'] = JobTypes.UPDATE.value
start_time = time.time()
if job.type == JobTypes.PREDICTION.value:
clusterer = _init_clusterer(job.clustering, training_df)
if job.predictive_model.predictive_model == PredictiveModels.CLASSIFICATION.value:
results, model_split = classification(training_df, test_df,
clusterer, job)
elif job.predictive_model.predictive_model == PredictiveModels.REGRESSION.value:
results, model_split = regression(training_df, test_df, clusterer,
job)
elif job.predictive_model.predictive_model == PredictiveModels.TIME_SERIES_PREDICTION.value:
results, model_split = time_series_prediction(
training_df, test_df, clusterer, job)
elif job.type == JobTypes.LABELLING.value:
results = _label_task(training_df)
elif job.type == JobTypes.UPDATE.value:
results, model_split = update_and_test(training_df, test_df, job)
else:
raise ValueError("Type {} not supported".format(job.type))
# TODO: integrateme
if job.type != JobTypes.LABELLING.value:
if job.predictive_model.predictive_model == PredictiveModels.REGRESSION.value:
job.evaluation = Evaluation.init(
job.predictive_model.predictive_model, results)
elif job.predictive_model.predictive_model == PredictiveModels.CLASSIFICATION.value:
job.evaluation = Evaluation.init(
job.predictive_model.predictive_model, results,
len(model_split[job.predictive_model.predictive_model]
[0].classes_) <= 2)
job.save()
if job.type == PredictiveModels.CLASSIFICATION.value:
save_result(results, job, start_time)
print("End job {}, {} .".format(job.type, get_run(job)))
print("\tResults {} .".format(results))
return results, model_split
def calculate_hyperopt(job: Job) -> (dict, dict, dict):
"""main entry method for hyperopt calculations
returns the predictive_model for the best trial
:param job: job configuration
:return: tuple containing the results, config and predictive_model split from the search
"""
logger.info("Start hyperopt job {} with {}, performance_metric {}".format(
job.type, get_run(job),
job.hyperparameter_optimizer.__getattribute__(
job.hyperparameter_optimizer.optimization_method.lower()).
performance_metric) #Todo: WHY DO I NEED TO GET HYPEROPT?
)
global train_df, validation_df, test_df, global_job
global_job = job
train_df, test_df = get_encoded_logs(job)
train_df, validation_df, test_df = _retrieve_train_validate_test(
train_df, test_df)
train_start_time = time.time()
space = _get_space(job)
max_evaluations = job.hyperparameter_optimizer.__getattribute__(
job.hyperparameter_optimizer.optimization_method.lower(
)).max_evaluations #Todo: WHY DO I NEED TO GET HYPEROPT?
trials = Trials()
algorithm = algorithm = OPTIMISATION_ALGORITHM[
job.hyperparameter_optimizer.__getattribute__(
job.hyperparameter_optimizer.optimization_method.lower(
)).algorithm_type]
_run_hyperoptimisation(space, algorithm.suggest, max_evaluations, trials)
best_candidate = trials.best_trial['result']
job.predictive_model = PredictiveModel.objects.filter(
pk=best_candidate['predictive_model_id'])[0]
job.predictive_model.save()
job.save()
best_candidate['results']['elapsed_time'] = timedelta(
seconds=time.time() -
train_start_time) # todo find better place for this
job.evaluation.elapsed_time = best_candidate['results']['elapsed_time']
job.evaluation.save()
results_df, auc = _test_best_candidate(
best_candidate, job.labelling.type,
job.predictive_model.predictive_model)
if job.predictive_model.predictive_model == PredictiveModels.CLASSIFICATION.value:
results = classification_prepare_results(results_df, auc)
else:
results = regression_prepare_results(results_df, job.labelling)
results['elapsed_time'] = job.evaluation.elapsed_time
job.evaluation = Evaluation.init(job.predictive_model.predictive_model,
results,
len(set(validation_df['label'])) <= 2)
job.evaluation.save()
job.save()
logger.info(
"End hyperopt job {}, {}. \n\tResults on validation {}. \n\tResults on test {}."
.format(job.type, get_run(job), best_candidate['results'], results)) #
return results, best_candidate['config'], best_candidate['model_split']
def run_by_type(training_df: DataFrame, test_df: DataFrame, job: Job) -> (dict, dict):
"""runs the specified training/evaluation run
:param training_df: training DataFrame
:param test_df: testing DataFrame
:param job: job configuration
:return: results and predictive_model split
"""
model_split = None
start_time = time.time()
if job.type == JobTypes.PREDICTION.value:
clusterer = init_clusterer(job.clustering, training_df)
results, model_split = MODEL[job.predictive_model.predictive_model][ModelActions.BUILD_MODEL_AND_TEST.value](training_df, test_df, clusterer, job)
elif job.type == JobTypes.LABELLING.value:
results = _label_task(training_df)
elif job.type == JobTypes.UPDATE.value:
results, model_split = MODEL[job.predictive_model.predictive_model][ModelActions.UPDATE_AND_TEST.value](training_df, test_df, job)
else:
raise ValueError("Type {} not supported".format(job.type))
# TODO: integrateme
if job.type != JobTypes.LABELLING.value:
results['elapsed_time'] = timedelta(seconds=time.time() - start_time) #todo find better place for this
if job.predictive_model.predictive_model == PredictiveModels.REGRESSION.value:
job.evaluation = Evaluation.init(
job.predictive_model.predictive_model,
results
)
elif job.predictive_model.predictive_model == PredictiveModels.CLASSIFICATION.value:
job.evaluation = Evaluation.init(
job.predictive_model.predictive_model,
results,
len(set(test_df['label'])) <= 2
)
elif job.predictive_model.predictive_model == PredictiveModels.TIME_SERIES_PREDICTION.value:
job.evaluation = Evaluation.init(
job.predictive_model.predictive_model,
results
)
job.evaluation.save()
job.save()
elif job.type == JobTypes.LABELLING.value:
# job.labelling = duplicate_orm_row(job.labelling) #todo: replace with simple CREATE
job.labelling = Labelling.objects.create(
type=job.labelling.type,
attribute_name=job.labelling.attribute_name,
threshold_type=job.labelling.threshold_type,
threshold=job.labelling.threshold
) #todo: futurebug if object changes
job.labelling.results = results
job.labelling.save()
job.save()
# if job.type == PredictiveModels.CLASSIFICATION.value: #todo this is an old workaround I should remove this
# save_result(results, job, start_time)
logger.info("End job {}, {} .".format(job.type, get_run(job)))
logger.info("\tResults {} .".format(results))
return results, model_split
def calculate_hyperopt(job: Job) -> (dict, dict, dict):
"""main entry method for hyperopt calculations
returns the predictive_model for the best trial
:param job: job configuration
:return: tuple containing the results, config and predictive_model split from the search
"""
logger.info("Start hyperopt job {} with {}, performance_metric {}".format(
job.type, get_run(job),
job.hyperparameter_optimizer.__getattribute__(
job.hyperparameter_optimizer.optimization_method.lower()
).performance_metric) #Todo: WHY DO I NEED TO GET HYPEROPT?
)
global training_df, test_df, global_job
global_job = job
training_df, test_df = get_encoded_logs(job)
#TODO evaluate on validation set
if holdout:
validation_df = test_df
# test_df = training_df.sample(frac=.2)
test_df = training_df.tail(int(len(training_df) * 20 / 100))
training_df = training_df.drop(test_df.index)
train_start_time = time.time()
space = _get_space(job)
max_evaluations = job.hyperparameter_optimizer.__getattribute__(
job.hyperparameter_optimizer.optimization_method.lower()
).max_evaluations #Todo: WHY DO I NEED TO GET HYPEROPT?
trials = Trials()
algorithm = _choose_algorithm(job)
try:
fmin(_calculate_and_evaluate, space, algo=algorithm.suggest, max_evals=max_evaluations, trials=trials)
except ValueError:
raise ValueError("All jobs failed, cannot find best configuration")
current_best = {'loss': 100, 'results': {}, 'predictive_model_id': {}, 'model_split': {}, 'config': {}}
for trial in trials:
a = trial['result']
if current_best['loss'] > a['loss']:
current_best = a
job.predictive_model = PredictiveModel.objects.filter(pk=current_best['predictive_model_id'])[0]
job.predictive_model.save()
job.save()
current_best['results']['elapsed_time'] = timedelta(seconds=time.time() - train_start_time) # todo find better place for this
job.evaluation.elapsed_time = current_best['results']['elapsed_time']
job.evaluation.save()
#TODO evaluate on validation set
if holdout:
results_df, auc = _test(
current_best['model_split'],
validation_df.drop(['trace_id'], 1),
evaluation=True,
is_binary_classifier=_check_is_binary_classifier(job.labelling.type)
)
results = _prepare_results(results_df, auc)
results['elapsed_time'] = job.evaluation.elapsed_time
job.evaluation = Evaluation.init(
job.predictive_model.predictive_model,
results,
len(set(test_df['label'])) <= 2
)
job.evaluation.save()
job.save()
if holdout:
logger.info("End hyperopt job {}, {}. \n\tResults on test {}. \n\tResults on validation {}.".format(job.type, get_run(job), current_best['results'], results))
return results, current_best['config'], current_best['model_split']
else:
logger.info("End hyperopt job {}, {}. \n\tResults on test {}.".format(job.type, get_run(job), current_best['results']))
return current_best['results'], current_best['config'], current_best['model_split']