def test_n_jobs_not_new_model():
grid_config = {
'sklearn.ensemble.AdaBoostClassifier': {
'n_estimators': [10, 100, 1000]
},
'sklearn.ensemble.RandomForestClassifier': {
'n_estimators': [10, 100],
'max_features': ['sqrt', 'log2'],
'max_depth': [5, 10, 15, 20],
'criterion': ['gini', 'entropy'],
'n_jobs': [12, 24],
}
}
with testing.postgresql.Postgresql() as postgresql:
engine = create_engine(postgresql.url())
ensure_db(engine)
with mock_s3():
s3_conn = boto3.resource('s3')
s3_conn.create_bucket(Bucket='econ-dev')
trainer = ModelTrainer(
project_path='econ-dev/inspections',
experiment_hash=None,
model_storage_engine=S3ModelStorageEngine(s3_conn, 'econ-dev/inspections'),
db_engine=engine,
model_group_keys=['label_name', 'label_timespan']
)
matrix = pandas.DataFrame.from_dict({
'entity_id': [1, 2],
'feature_one': [3, 4],
'feature_two': [5, 6],
'label': ['good', 'bad']
})
train_tasks = trainer.generate_train_tasks(
grid_config,
dict(),
InMemoryMatrixStore(matrix, {
'label_timespan': '1d',
'end_time': datetime.datetime.now(),
'feature_start_time': datetime.date(2012, 12, 20),
'label_name': 'label',
'metta-uuid': '1234',
'feature_names': ['ft1', 'ft2'],
'indices': ['entity_id'],
})
)
assert len(train_tasks) == 35 # 32+3, would be (32*2)+3 if we didn't remove
assert len([
task for task in train_tasks
if 'n_jobs' in task['parameters']
]) == 32
for train_task in train_tasks:
trainer.process_train_task(**train_task)
for row in engine.execute(
'select model_parameters from results.model_groups'
):
assert 'n_jobs' not in row[0]
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 matrix_store(self):
data_dict = OrderedDict([('entity_id', [1, 2]),
('k_feature', [0.5, 0.4]),
('m_feature', [0.4, 0.5]), ('label', [0, 1])])
df = pandas.DataFrame.from_dict(data_dict)
metadata = {
'label_name': 'label',
'indices': ['entity_id'],
}
inmemory = InMemoryMatrixStore(matrix=df, metadata=metadata)
with tempfile.TemporaryDirectory() as tmpdir:
tmpcsv = os.path.join(tmpdir, 'df.csv')
tmpyaml = os.path.join(tmpdir, 'metadata.yaml')
tmphdf = os.path.join(tmpdir, 'df.h5')
with open(tmpyaml, 'w') as outfile:
yaml.dump(metadata, outfile, default_flow_style=False)
df.to_csv(tmpcsv, index=False)
df.to_hdf(tmphdf, 'matrix')
csv = CSVMatrixStore(matrix_path=tmpcsv, metadata_path=tmpyaml)
hdf = HDFMatrixStore(matrix_path=tmphdf, metadata_path=tmpyaml)
assert csv.matrix.to_dict() == inmemory.matrix.to_dict()
assert hdf.matrix.to_dict() == inmemory.matrix.to_dict()
assert csv.metadata == inmemory.metadata
assert hdf.metadata == inmemory.metadata
assert csv.head_of_matrix.to_dict(
) == inmemory.head_of_matrix.to_dict()
assert hdf.head_of_matrix.to_dict(
) == inmemory.head_of_matrix.to_dict()
assert csv.empty == inmemory.empty
assert hdf.empty == inmemory.empty
assert csv.labels().to_dict() == inmemory.labels().to_dict()
assert hdf.labels().to_dict() == inmemory.labels().to_dict()
matrix_store = [inmemory, hdf, csv]
return matrix_store
def test_as_of_dates_entity_date_index(self):
data = {
'entity_id': [1, 2, 1, 2],
'feature_one': [0.5, 0.6, 0.5, 0.6],
'feature_two': [0.5, 0.6, 0.5, 0.6],
'as_of_date':
['2016-01-01', '2016-01-01', '2017-01-01', '2017-01-01']
}
matrix = InMemoryMatrixStore(
matrix=pandas.DataFrame.from_dict(data),
metadata={'indices': ['entity_id', 'as_of_date']})
self.assertEqual(matrix.as_of_dates, ['2016-01-01', '2017-01-01'])
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_timespan']
)
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_timespan': '1d',
'end_time': datetime.datetime.now(),
'feature_start_time': datetime.date(2012, 12, 20),
'label_name': 'label',
'metta-uuid': '1234',
'feature_names': ['ft1', 'ft2'],
'indices': ['entity_id'],
})
# 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_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_timespan': '3month',
'metta-uuid': '1234',
'indices': ['entity_id'],
}
matrix_store = InMemoryMatrixStore(matrix, metadata)
predict_proba = predictor.predict(
model_id,
matrix_store,
misc_db_parameters=dict(),
train_matrix_columns=['feature_one', 'feature_two'])
# 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_uniform_distribution_entity_date_index():
with testing.postgresql.Postgresql() as postgresql:
db_engine = create_engine(postgresql.url())
ensure_db(db_engine)
init_engine(db_engine)
model = ModelFactory()
feature_importances = [
FeatureImportanceFactory(model_rel=model,
feature='feature_{}'.format(i))
for i in range(0, 10)
]
data_dict = {
'entity_id': [1, 1],
'as_of_date': ['2016-01-01', '2017-01-01']
}
for imp in feature_importances:
data_dict[imp.feature] = [0.5, 0.5]
test_store = InMemoryMatrixStore(
matrix=pandas.DataFrame.from_dict(data_dict).set_index(
['entity_id', 'as_of_date']),
metadata=sample_metadata())
session.commit()
results = uniform_distribution(db_engine,
model_id=model.model_id,
as_of_date='2016-01-01',
test_matrix_store=test_store,
n_ranks=5)
assert len(results) == 5 # 5 features x 1 entity for this as_of_date
for result in results:
assert 'entity_id' in result
assert 'feature_name' in result
assert 'score' in result
assert 'feature_value' in result
assert result['feature_value'] == 0.5
assert result['score'] >= 0
assert result['score'] <= 1
assert isinstance(result['feature_name'], str)
assert result['entity_id'] in [1, 2]
def test_integration():
with testing.postgresql.Postgresql() as postgresql:
db_engine = create_engine(postgresql.url())
ensure_db(db_engine)
with mock_s3():
s3_conn = boto3.resource('s3')
s3_conn.create_bucket(Bucket='econ-dev')
project_path = 'econ-dev/inspections'
# create train and test matrices
train_matrix = pandas.DataFrame.from_dict({
'entity_id': [1, 2],
'feature_one': [3, 4],
'feature_two': [5, 6],
'label': [7, 8]
}).set_index('entity_id')
train_metadata = {
'beginning_of_time': datetime.date(2012, 12, 20),
'end_time': datetime.date(2016, 12, 20),
'label_name': 'label',
'label_window': '1y',
'feature_names': ['ft1', 'ft2'],
'metta-uuid': '1234',
}
train_store = InMemoryMatrixStore(train_matrix, train_metadata)
as_of_dates = [
datetime.date(2016, 12, 21),
datetime.date(2017, 1, 21)
]
test_stores = [
InMemoryMatrixStore(
pandas.DataFrame.from_dict({
'entity_id': [3],
'feature_one': [8],
'feature_two': [5],
'label': [5]
}).set_index('entity_id'), {
'label_name': 'label',
'label_window': '1y',
'end_time': as_of_date,
'metta-uuid': '1234',
}) for as_of_date in as_of_dates
]
model_storage_engine = S3ModelStorageEngine(s3_conn, project_path)
experiment_hash = save_experiment_and_get_hash({}, db_engine)
# instantiate pipeline objects
trainer = ModelTrainer(
project_path=project_path,
experiment_hash=experiment_hash,
model_storage_engine=model_storage_engine,
db_engine=db_engine,
model_group_keys=['label_name', 'label_window'])
predictor = Predictor(project_path, model_storage_engine,
db_engine)
model_evaluator = ModelEvaluator([{
'metrics': ['precision@'],
'thresholds': {
'top_n': [5]
}
}], db_engine)
# run the pipeline
grid_config = {
'sklearn.linear_model.LogisticRegression': {
'C': [0.00001, 0.0001],
'penalty': ['l1', 'l2'],
'random_state': [2193]
}
}
model_ids = trainer.train_models(grid_config=grid_config,
misc_db_parameters=dict(),
matrix_store=train_store)
for model_id in model_ids:
for as_of_date, test_store in zip(as_of_dates, test_stores):
predictions_proba = predictor.predict(
model_id,
test_store,
misc_db_parameters=dict(),
train_matrix_columns=['feature_one', 'feature_two'])
model_evaluator.evaluate(predictions_proba,
test_store.labels(), model_id,
as_of_date, as_of_date, '6month')
# assert
# 1. 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, model_id, as_of_date
from results.predictions
join results.models using (model_id)
order by 3, 2''')
]
assert records == [
(3, 1, datetime.datetime(2016, 12, 21)),
(3, 2, datetime.datetime(2016, 12, 21)),
(3, 3, datetime.datetime(2016, 12, 21)),
(3, 4, datetime.datetime(2016, 12, 21)),
(3, 1, datetime.datetime(2017, 1, 21)),
(3, 2, datetime.datetime(2017, 1, 21)),
(3, 3, datetime.datetime(2017, 1, 21)),
(3, 4, datetime.datetime(2017, 1, 21)),
]
# that evaluations are there
records = [
row for row in db_engine.execute('''
select model_id, evaluation_start_time, metric, parameter
from results.evaluations order by 2, 1''')
]
assert records == [
(1, datetime.datetime(2016, 12, 21), 'precision@', '5_abs'),
(2, datetime.datetime(2016, 12, 21), 'precision@', '5_abs'),
(3, datetime.datetime(2016, 12, 21), 'precision@', '5_abs'),
(4, datetime.datetime(2016, 12, 21), 'precision@', '5_abs'),
(1, datetime.datetime(2017, 1, 21), 'precision@', '5_abs'),
(2, datetime.datetime(2017, 1, 21), 'precision@', '5_abs'),
(3, datetime.datetime(2017, 1, 21), 'precision@', '5_abs'),
(4, datetime.datetime(2017, 1, 21), 'precision@', '5_abs'),
]
def test_predictor():
with testing.postgresql.Postgresql() as postgresql:
db_engine = create_engine(postgresql.url())
ensure_db(db_engine)
with mock_s3():
s3_conn = boto3.resource('s3')
s3_conn.create_bucket(Bucket='econ-dev')
project_path = 'econ-dev/inspections'
model_storage_engine = S3ModelStorageEngine(s3_conn, project_path)
_, model_id = \
fake_trained_model(project_path, model_storage_engine, db_engine)
predictor = Predictor(project_path, model_storage_engine,
db_engine)
# create prediction set
matrix = pandas.DataFrame.from_dict({
'entity_id': [1, 2],
'feature_one': [3, 4],
'feature_two': [5, 6],
'label': [7, 8]
}).set_index('entity_id')
metadata = {
'label_name': 'label',
'end_time': AS_OF_DATE,
'label_window': '3month',
'metta-uuid': '1234',
}
matrix_store = InMemoryMatrixStore(matrix, metadata)
train_matrix_columns = ['feature_one', 'feature_two']
predict_proba = predictor.predict(
model_id,
matrix_store,
misc_db_parameters=dict(),
train_matrix_columns=train_matrix_columns)
# assert
# 1. that the returned predictions are of the desired length
assert len(predict_proba) == 2
# 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) == 2
# 3. that the contained as_of_dates match what we sent in
for record in records:
assert record[1].date() == AS_OF_DATE
# 4. that the entity ids match the given dataset
assert sorted([record[0] for record in records]) == [1, 2]
# 5. running with same model_id, different as of date
# then with same as of date only replaces the records
# with the same date
new_matrix = pandas.DataFrame.from_dict({
'entity_id': [1, 2],
'feature_one': [3, 4],
'feature_two': [5, 6],
'label': [7, 8]
}).set_index('entity_id')
new_metadata = {
'label_name': 'label',
'end_time': AS_OF_DATE + datetime.timedelta(days=1),
'label_window': '3month',
'metta-uuid': '1234',
}
new_matrix_store = InMemoryMatrixStore(new_matrix, new_metadata)
predictor.predict(model_id,
new_matrix_store,
misc_db_parameters=dict(),
train_matrix_columns=train_matrix_columns)
predictor.predict(model_id,
matrix_store,
misc_db_parameters=dict(),
train_matrix_columns=train_matrix_columns)
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
# 6. That we can delete the model when done prediction on it
predictor.delete_model(model_id)
assert predictor.load_model(model_id) == None
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).strftime(predictor.expected_matrix_ts_format)
daytwo = datetime.date(2011, 1,
2).strftime(predictor.expected_matrix_ts_format)
# 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(),
train_matrix_columns=['feature_one', 'feature_two'])
# 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(),
train_matrix_columns=['feature_one', 'feature_two'])
assert_array_equal(new_predict_proba, predict_proba)
assert not predictor.load_model.called
def test_model_trainer():
with testing.postgresql.Postgresql() as postgresql:
engine = create_engine(postgresql.url())
ensure_db(engine)
grid_config = {
'sklearn.linear_model.LogisticRegression': {
'C': [0.00001, 0.0001],
'penalty': ['l1', 'l2'],
'random_state': [2193]
}
}
with mock_s3():
s3_conn = boto3.resource('s3')
s3_conn.create_bucket(Bucket='econ-dev')
# create training set
matrix = pandas.DataFrame.from_dict({
'entity_id': [1, 2],
'feature_one': [3, 4],
'feature_two': [5, 6],
'label': ['good', 'bad']
})
metadata = {
'feature_start_time': datetime.date(2012, 12, 20),
'end_time': datetime.date(2016, 12, 20),
'label_name': 'label',
'label_timespan': '1y',
'metta-uuid': '1234',
'feature_names': ['ft1', 'ft2'],
'indices': ['entity_id'],
}
project_path = 'econ-dev/inspections'
model_storage_engine = S3ModelStorageEngine(s3_conn, project_path)
trainer = ModelTrainer(
project_path=project_path,
experiment_hash=None,
model_storage_engine=model_storage_engine,
db_engine=engine,
model_group_keys=['label_name', 'label_timespan']
)
matrix_store = InMemoryMatrixStore(matrix, metadata)
model_ids = trainer.train_models(
grid_config=grid_config,
misc_db_parameters=dict(),
matrix_store=matrix_store
)
# assert
# 1. that the models and feature importances table entries are present
records = [
row for row in
engine.execute('select * from results.feature_importances')
]
assert len(records) == 4 * 2 # maybe exclude entity_id? yes
records = [
row for row in
engine.execute('select model_hash from results.models')
]
assert len(records) == 4
cache_keys = [
model_cache_key(project_path, model_row[0], s3_conn)
for model_row in records
]
# 2. that the model groups are distinct
records = [
row for row in
engine.execute('select distinct model_group_id from results.models')
]
assert len(records) == 4
# 3. that all four models are cached
model_pickles = [
pickle.loads(cache_key.get()['Body'].read())
for cache_key in cache_keys
]
assert len(model_pickles) == 4
assert len([x for x in model_pickles if x is not None]) == 4
# 4. that their results can have predictions made on it
test_matrix = pandas.DataFrame.from_dict({
'entity_id': [3, 4],
'feature_one': [4, 4],
'feature_two': [6, 5],
})
test_matrix = InMemoryMatrixStore(matrix=test_matrix, metadata=metadata).matrix
for model_pickle in model_pickles:
predictions = model_pickle.predict(test_matrix)
assert len(predictions) == 2
# 5. when run again, same models are returned
new_model_ids = trainer.train_models(
grid_config=grid_config,
misc_db_parameters=dict(),
matrix_store=matrix_store
)
assert len([
row for row in
engine.execute('select model_hash from results.models')
]) == 4
assert model_ids == new_model_ids
# 6. if replace is set, update non-unique attributes and feature importances
max_batch_run_time = [
row[0] for row in
engine.execute('select max(batch_run_time) from results.models')
][0]
trainer = ModelTrainer(
project_path=project_path,
experiment_hash=None,
model_storage_engine=model_storage_engine,
db_engine=engine,
model_group_keys=['label_name', 'label_timespan'],
replace=True
)
new_model_ids = trainer.train_models(
grid_config=grid_config,
misc_db_parameters=dict(),
matrix_store=matrix_store,
)
assert model_ids == new_model_ids
assert [
row['model_id'] for row in
engine.execute('select model_id from results.models order by 1 asc')
] == model_ids
new_max_batch_run_time = [
row[0] for row in
engine.execute('select max(batch_run_time) from results.models')
][0]
assert new_max_batch_run_time > max_batch_run_time
records = [
row for row in
engine.execute('select * from results.feature_importances')
]
assert len(records) == 4 * 2 # maybe exclude entity_id? yes
# 7. if the cache is missing but the metadata is still there, reuse the metadata
for row in engine.execute('select model_hash from results.models'):
model_storage_engine.get_store(row[0]).delete()
new_model_ids = trainer.train_models(
grid_config=grid_config,
misc_db_parameters=dict(),
matrix_store=matrix_store
)
assert model_ids == sorted(new_model_ids)
# 8. that the generator interface works the same way
new_model_ids = trainer.generate_trained_models(
grid_config=grid_config,
misc_db_parameters=dict(),
matrix_store=matrix_store
)
assert model_ids == \
sorted([model_id for model_id in new_model_ids])