def apply_train_test(temporal_set, blocks, **kwargs):
# Connect to db
try:
db_engine = setup_environment.get_database()
except:
log.warning('Could not connect to the database')
raise
run_model = RunModels(labels=kwargs['labels'],
features=kwargs['features'],
schema_name=kwargs['schema_name'],
blocks=blocks,
feature_lookback_duration=kwargs['feature_lookback_duration'],
labels_config=kwargs['labels_config'],
labels_table_name=kwargs['labels_table_name'],
temporal_split=temporal_set,
grid_config=kwargs['grid_config'],
project_path=kwargs['project_path'],
misc_db_parameters=kwargs['misc_db_parameters'],
experiment_hash=kwargs['experiment_hash'],
db_engine=db_engine)
log.info('Run models for temporal set: {}'.format(temporal_set))
log.info('Run models for feature blocks: {}'.format(blocks))
model_storage = InMemoryModelStorageEngine('empty')
train_matrix_uuid, model_ids_generator = run_model.setup_train_models(model_storage)
if train_matrix_uuid is None:
return None
log.info('Run tests')
run_model.train_test_models(train_matrix_uuid, model_ids_generator, model_storage)
db_engine.dispose()
return None
def test_model_scoring_inspections():
with testing.postgresql.Postgresql() as postgresql:
db_engine = create_engine(postgresql.url())
ensure_db(db_engine)
metric_groups = [{
'metrics': ['precision@', 'recall@', 'fpr@'],
'thresholds': {
'percentiles': [50.0],
'top_n': [3]
}
}]
model_scorer = ModelScorer(metric_groups, db_engine)
_, model_id = fake_trained_model(
'myproject', InMemoryModelStorageEngine('myproject'), db_engine)
labels = numpy.array([True, False, numpy.nan, True, False])
prediction_probas = numpy.array([0.56, 0.4, 0.55, 0.5, 0.3])
evaluation_start = datetime.datetime(2016, 4, 1)
evaluation_end = datetime.datetime(2016, 7, 1)
example_frequency = '1d'
model_scorer.score(prediction_probas, labels, model_id,
evaluation_start, evaluation_end, example_frequency)
for record in db_engine.execute(
'''select * from results.evaluations
where model_id = %s and evaluation_start_time = %s order by 1''',
(model_id, evaluation_start)):
assert record['num_labeled_examples'] == 4
assert record['num_positive_labels'] == 2
if 'pct' in record['parameter']:
assert record['num_labeled_above_threshold'] == 1
else:
assert record['num_labeled_above_threshold'] == 2
def test_retry_recovery(self):
grid_config = {
'sklearn.ensemble.AdaBoostClassifier': {
'n_estimators': [10]
},
}
engine = None
trainer = None
port = None
with testing.postgresql.Postgresql() as postgresql:
port = postgresql.settings['port']
engine = create_engine(postgresql.url())
ensure_db(engine)
trainer = ModelTrainer(
project_path='econ-dev/inspections',
experiment_hash=None,
model_storage_engine=InMemoryModelStorageEngine(
project_path=''),
db_engine=engine,
model_group_keys=['label_name', 'label_window'])
matrix = pandas.DataFrame.from_dict({
'entity_id': [1, 2],
'feature_one': [3, 4],
'feature_two': [5, 6],
'label': ['good', 'bad']
})
matrix_store = InMemoryMatrixStore(
matrix, {
'label_window': '1d',
'end_time': datetime.datetime.now(),
'beginning_of_time': datetime.date(2012, 12, 20),
'label_name': 'label',
'metta-uuid': '1234',
'feature_names': ['ft1', 'ft2']
})
# start without a database server
# then bring it back up after the first sleep
# use self so it doesn't go out of scope too early and shut down
self.new_server = None
def replace_db(arg):
self.new_server = testing.postgresql.Postgresql(port=port)
engine = create_engine(self.new_server.url())
ensure_db(engine)
with patch('time.sleep') as time_mock:
time_mock.side_effect = replace_db
try:
trainer.train_models(grid_config, dict(), matrix_store)
finally:
if self.new_server is not None:
self.new_server.stop()
assert len(time_mock.mock_calls) == 1
def test_retry_max(self):
grid_config = {
'sklearn.ensemble.AdaBoostClassifier': {
'n_estimators': [10]
},
}
engine = None
trainer = None
# set up a basic model training run
# TODO abstract the setup of a basic model training run where
# we don't worry about the specific values used? it would make
# tests like this require a bit less noise to read past
with testing.postgresql.Postgresql() as postgresql:
engine = create_engine(postgresql.url())
ensure_db(engine)
trainer = ModelTrainer(
project_path='econ-dev/inspections',
experiment_hash=None,
model_storage_engine=InMemoryModelStorageEngine(
project_path=''),
db_engine=engine,
model_group_keys=['label_name', 'label_window'])
matrix = pandas.DataFrame.from_dict({
'entity_id': [1, 2],
'feature_one': [3, 4],
'feature_two': [5, 6],
'label': ['good', 'bad']
})
matrix_store = InMemoryMatrixStore(
matrix, {
'label_window': '1d',
'end_time': datetime.datetime.now(),
'beginning_of_time': datetime.date(2012, 12, 20),
'label_name': 'label',
'metta-uuid': '1234',
'feature_names': ['ft1', 'ft2']
})
# the postgres server goes out of scope here and thus no longer exists
with patch('time.sleep') as time_mock:
with self.assertRaises(sqlalchemy.exc.OperationalError):
trainer.train_models(grid_config, dict(), matrix_store)
# we want to make sure that we are using the retrying module sanely
# as opposed to matching the exact # of calls specified by the code
assert len(time_mock.mock_calls) > 5
def test_model_scoring_inspections():
with testing.postgresql.Postgresql() as postgresql:
db_engine = create_engine(postgresql.url())
ensure_db(db_engine)
metric_groups = [{
'metrics': ['precision@', 'recall@'],
'thresholds': {
'percentiles': [5.0, 10.0],
'top_n': [5, 10]
}
}]
model_scorer = ModelScorer(metric_groups, db_engine)
trained_model, model_id = fake_trained_model(
'myproject', InMemoryModelStorageEngine('myproject'), db_engine)
labels = fake_labels(5)
as_of_date = datetime.date(2016, 5, 5)
evaluation_start = datetime.datetime(2016, 4, 1)
evaluation_end = datetime.datetime(2016, 7, 1)
prediction_frequency = '1d'
model_scorer.score(
trained_model.predict_proba(labels)[:, 1],
trained_model.predict(labels), labels, model_id, evaluation_start,
evaluation_end, prediction_frequency)
# assert
# that all of the records are there
results = db_engine.execute(
'''select distinct(metric || parameter) from results.evaluations
where model_id = %s and evaluation_start_time = %s order by 1''',
(model_id, evaluation_start))
records = [row[0] for row in results]
assert records == [
'[email protected]_pct',
'precision@10_abs',
'[email protected]_pct',
'precision@5_abs',
'[email protected]_pct',
'recall@10_abs',
'[email protected]_pct',
'recall@5_abs',
]
def apply_score_day(temporal_set, blocks, chosen_model_group_id, date, **kwargs):
# Connect to db
try:
db_engine = setup_environment.get_database(production=True)
except:
log.warning('Could not connect to the database')
raise
run_model = RunModels(labels=kwargs['labels'],
features=kwargs['features'],
schema_name=kwargs['schema_name'],
blocks=blocks,
feature_lookback_duration=kwargs['feature_lookback_duration'],
labels_config=kwargs['labels_config'],
labels_table_name=kwargs['labels_table_name'],
temporal_split=temporal_set,
grid_config=kwargs['grid_config'],
project_path=kwargs['project_path'],
misc_db_parameters=kwargs['misc_db_parameters'],
db_engine=db_engine)
log.info('Run models for temporal set: {}'.format(temporal_set))
log.info('Run models for feature blocks: {}'.format(blocks))
model_storage = InMemoryModelStorageEngine('empty')
train_matrix_uuid, model_ids_generator = run_model.setup_train_models(model_storage)
if train_matrix_uuid is None:
return None
log.info('Score model')
run_model.train_score_models(model_ids_generator, model_storage)
### update statistic tables
# time delta
query = "select production.populate_time_delta();"
conn = db_engine.raw_connection()
conn.cursor().execute(query)
conn.commit()
db_engine.dispose()
return None
def test_predictor_composite_index():
with testing.postgresql.Postgresql() as postgresql:
db_engine = create_engine(postgresql.url())
ensure_db(db_engine)
project_path = 'econ-dev/inspections'
model_storage_engine = InMemoryModelStorageEngine(project_path)
_, model_id = \
fake_trained_model(project_path, model_storage_engine, db_engine)
predictor = Predictor(project_path, model_storage_engine, db_engine)
dayone = datetime.datetime(2011, 1, 1)
daytwo = datetime.datetime(2011, 1, 2)
# create prediction set
matrix = pandas.DataFrame.from_dict({
'entity_id': [1, 2, 1, 2],
'as_of_date': [dayone, dayone, daytwo, daytwo],
'feature_one': [3, 4, 5, 6],
'feature_two': [5, 6, 7, 8],
'label': [7, 8, 8, 7]
}).set_index(['entity_id', 'as_of_date'])
metadata = {
'label_name': 'label',
'end_time': AS_OF_DATE,
'label_window': '3month',
'metta-uuid': '1234',
}
matrix_store = InMemoryMatrixStore(matrix, metadata)
predict_proba = predictor.predict(model_id, matrix_store, misc_db_parameters=dict())
# assert
# 1. that the returned predictions are of the desired length
assert len(predict_proba) == 4
# 2. that the predictions table entries are present and
# can be linked to the original models
records = [
row for row in
db_engine.execute('''select entity_id, as_of_date
from results.predictions
join results.models using (model_id)''')
]
assert len(records) == 4
def test_predictor_retrieve():
with testing.postgresql.Postgresql() as postgresql:
db_engine = create_engine(postgresql.url())
ensure_db(db_engine)
project_path = 'econ-dev/inspections'
model_storage_engine = InMemoryModelStorageEngine(project_path)
_, model_id = \
fake_trained_model(project_path, model_storage_engine, db_engine)
predictor = Predictor(project_path, model_storage_engine, db_engine, replace=False)
dayone = datetime.date(2011, 1, 1).isoformat()
daytwo = datetime.date(2011, 1, 2).isoformat()
# create prediction set
matrix_data = {
'entity_id': [1, 2, 1, 2],
'as_of_date': [dayone, dayone, daytwo, daytwo],
'feature_one': [3, 4, 5, 6],
'feature_two': [5, 6, 7, 8],
'label': [7, 8, 8, 7]
}
matrix = pandas.DataFrame.from_dict(matrix_data)\
.set_index(['entity_id', 'as_of_date'])
metadata = {
'label_name': 'label',
'end_time': AS_OF_DATE,
'label_window': '3month',
'metta-uuid': '1234',
}
matrix_store = InMemoryMatrixStore(matrix, metadata)
predict_proba = predictor.predict(model_id, matrix_store, misc_db_parameters=dict())
# When run again, the predictions retrieved from the database
# should match.
#
# Some trickiness here. Let's explain:
#
# If we are not careful, retrieving predictions from the database and
# presenting them as a numpy array can result in a bad ordering,
# since the given matrix may not be 'ordered' by some criteria
# that can be easily represented by an ORDER BY clause.
#
# It will sometimes work, because without ORDER BY you will get
# it back in the table's physical order, which unless something has
# happened to the table will be the order you inserted it,
# which could very well be the order in the matrix.
# So it's not a bug that would necessarily immediately show itself,
# but when it does go wrong your scores will be garbage.
#
# So we simulate a table order mutation that can happen over time:
# Remove the first row and put it at the end.
# If the Predictor doesn't explicitly reorder the results, this will fail
session = sessionmaker(bind=db_engine)()
obj = session.query(Prediction).first()
session.delete(obj)
session.commit()
make_transient(obj)
session = sessionmaker(bind=db_engine)()
session.add(obj)
session.commit()
predictor.load_model = Mock()
new_predict_proba = predictor.predict(model_id, matrix_store, misc_db_parameters=dict())
assert_array_equal(new_predict_proba, predict_proba)
assert not predictor.load_model.called
def test_model_scoring_early_warning():
with testing.postgresql.Postgresql() as postgresql:
db_engine = create_engine(postgresql.url())
ensure_db(db_engine)
metric_groups = [{
'metrics': [
'precision@', 'recall@', 'true positives@', 'true negatives@',
'false positives@', 'false negatives@'
],
'thresholds': {
'percentiles': [5.0, 10.0],
'top_n': [5, 10]
}
}, {
'metrics': [
'f1', 'mediocre', 'accuracy', 'roc_auc',
'average precision score'
],
}, {
'metrics': ['fbeta@'],
'parameters': [{
'beta': 0.75
}, {
'beta': 1.25
}]
}]
custom_metrics = {'mediocre': always_half}
model_scorer = ModelScorer(metric_groups, db_engine, custom_metrics)
trained_model, model_id = fake_trained_model(
'myproject', InMemoryModelStorageEngine('myproject'), db_engine)
labels = fake_labels(5)
as_of_date = datetime.date(2016, 5, 5)
model_scorer.score(
trained_model.predict_proba(labels)[:, 1],
trained_model.predict(labels), labels, model_id, as_of_date,
as_of_date, '1y')
# assert
# that all of the records are there
records = [
row[0] for row in db_engine.execute(
'''select distinct(metric || parameter) from results.evaluations
where model_id = %s and evaluation_start_time = %s order by 1''',
(model_id, as_of_date))
]
assert records == [
'accuracy', 'average precision score', 'f1',
'false [email protected]_pct', 'false negatives@10_abs',
'false [email protected]_pct', 'false negatives@5_abs',
'false [email protected]_pct', 'false positives@10_abs',
'false [email protected]_pct', 'false positives@5_abs',
'[email protected]_beta', '[email protected]_beta', 'mediocre',
'[email protected]_pct', 'precision@10_abs', '[email protected]_pct',
'precision@5_abs', '[email protected]_pct', 'recall@10_abs',
'[email protected]_pct', 'recall@5_abs', 'roc_auc',
'true [email protected]_pct', 'true negatives@10_abs',
'true [email protected]_pct', 'true negatives@5_abs',
'true [email protected]_pct', 'true positives@10_abs',
'true [email protected]_pct', 'true positives@5_abs'
]