def __init__(
self,
config,
db_engine,
project_path=None,
matrix_storage_class=CSVMatrixStore,
replace=True,
cleanup=False,
cleanup_timeout=None,
materialize_subquery_fromobjs=True,
profile=False,
):
self._check_config_version(config)
self.config = config
self.project_storage = ProjectStorage(project_path)
self.model_storage_engine = ModelStorageEngine(self.project_storage)
self.matrix_storage_engine = MatrixStorageEngine(
self.project_storage, matrix_storage_class
)
self.project_path = project_path
self.replace = replace
self.db_engine = db_engine
upgrade_db(db_engine=self.db_engine)
self.features_schema_name = "features"
self.materialize_subquery_fromobjs = materialize_subquery_fromobjs
self.experiment_hash = save_experiment_and_get_hash(self.config, self.db_engine)
self.labels_table_name = "labels_{}".format(self.experiment_hash)
self.cohort_table_name = "cohort_{}".format(self.experiment_hash)
self.initialize_components()
self.cleanup = cleanup
if self.cleanup:
logging.info(
"cleanup is set to True, so intermediate tables (labels and states) "
"will be removed after matrix creation"
)
else:
logging.info(
"cleanup is set to False, so intermediate tables (labels and states) "
"will not be removed after matrix creation"
)
self.cleanup_timeout = (
self.cleanup_timeout if cleanup_timeout is None else cleanup_timeout
)
self.profile = profile
logging.info("Generate profiling stats? (profile option): %s", self.profile)
def __init__(
self,
config,
db_engine,
project_path=None,
matrix_storage_class=CSVMatrixStore,
replace=True,
cleanup=False,
cleanup_timeout=None,
):
self._check_config_version(config)
self.config = config
if isinstance(db_engine, Engine):
logging.warning(
"Raw, unserializable SQLAlchemy engine passed. "
"URL will be used, other options may be lost in multi-process environments"
)
self.db_engine = create_engine(db_engine.url)
else:
self.db_engine = db_engine
self.project_storage = ProjectStorage(project_path)
self.model_storage_engine = ModelStorageEngine(self.project_storage)
self.matrix_storage_engine = MatrixStorageEngine(
self.project_storage, matrix_storage_class)
self.project_path = project_path
self.replace = replace
upgrade_db(db_engine=self.db_engine)
self.features_schema_name = "features"
self.experiment_hash = save_experiment_and_get_hash(
self.config, self.db_engine)
self.labels_table_name = "labels_{}".format(self.experiment_hash)
self.initialize_components()
self.cleanup = cleanup
if self.cleanup:
logging.info(
"cleanup is set to True, so intermediate tables (labels and states) "
"will be removed after matrix creation")
else:
logging.info(
"cleanup is set to False, so intermediate tables (labels and states) "
"will not be removed after matrix creation")
self.cleanup_timeout = (self.cleanup_timeout if cleanup_timeout is None
else cleanup_timeout)
def test_ModelTrainTester_generate_tasks(db_engine_with_results_schema,
project_storage,
sample_timechop_splits,
sample_grid_config):
db_engine = db_engine_with_results_schema
model_storage_engine = ModelStorageEngine(project_storage)
matrix_storage_engine = MatrixStorageEngine(project_storage)
sample_matrix_store = get_matrix_store(project_storage)
experiment_hash = save_experiment_and_get_hash({}, db_engine)
run_id = initialize_tracking_and_get_run_id(
experiment_hash,
experiment_class_path="",
random_seed=5,
experiment_kwargs={},
db_engine=db_engine_with_results_schema)
# instantiate pipeline objects
trainer = ModelTrainer(
experiment_hash=experiment_hash,
model_storage_engine=model_storage_engine,
db_engine=db_engine,
run_id=run_id,
)
train_tester = ModelTrainTester(
matrix_storage_engine=matrix_storage_engine,
model_trainer=trainer,
model_evaluator=None,
individual_importance_calculator=None,
predictor=None,
subsets=None,
protected_groups_generator=None,
)
with patch.object(matrix_storage_engine,
'get_store',
return_value=sample_matrix_store):
batches = train_tester.generate_task_batches(
splits=sample_timechop_splits, grid_config=sample_grid_config)
assert len(batches) == 3
# we expect to have a task for each combination of split and classifier
flattened_tasks = list(task for batch in batches
for task in batch.tasks)
assert len(flattened_tasks) == \
len(sample_timechop_splits) * len(list(flatten_grid_config(sample_grid_config)))
# we also expect each task to match the call signature of process_task
with patch.object(train_tester, 'process_task', autospec=True):
for task in flattened_tasks:
train_tester.process_task(**task)
def setup_model_train_tester(project_storage,
replace,
additional_bigtrain_classnames=None):
matrix_storage_engine = MatrixStorageEngine(project_storage)
train_matrix_store = get_matrix_store(
project_storage,
metadata=matrix_metadata_creator(matrix_type="train"),
write_to_db=False)
test_matrix_store = get_matrix_store(
project_storage,
metadata=matrix_metadata_creator(matrix_type="test"),
write_to_db=False)
sample_train_kwargs = {
'matrix_store': train_matrix_store,
'class_path': None,
'parameters': {},
'model_hash': None,
'misc_db_parameters': {}
}
train_test_task = {
'train_kwargs': sample_train_kwargs,
'train_store': train_matrix_store,
'test_store': test_matrix_store
}
predictor = MagicMock(spec_set=Predictor)
trainer = MagicMock(spec_set=ModelTrainer)
evaluator = MagicMock(spec_set=ModelEvaluator)
individual_importance_calculator = MagicMock(
spec_set=IndividualImportanceCalculator)
protected_groups_generator = MagicMock(spec_set=ProtectedGroupsGenerator)
train_tester = ModelTrainTester(
matrix_storage_engine=matrix_storage_engine,
model_trainer=trainer,
model_evaluator=evaluator,
individual_importance_calculator=individual_importance_calculator,
predictor=predictor,
subsets=[],
replace=replace,
protected_groups_generator=protected_groups_generator,
additional_bigtrain_classnames=additional_bigtrain_classnames)
return train_tester, train_test_task
def __init__(
self,
config,
db_engine,
project_path=None,
matrix_storage_class=CSVMatrixStore,
replace=True,
cleanup=False,
cleanup_timeout=None,
materialize_subquery_fromobjs=True,
features_ignore_cohort=False,
profile=False,
save_predictions=True,
skip_validation=False,
partial_run=False,
):
# For a partial run, skip validation and avoid cleaning up
# we'll also skip filling default config values below
if partial_run:
cleanup = False
skip_validation = True
experiment_kwargs = bind_kwargs(
self.__class__, **{
key: value
for (key, value) in locals().items()
if key not in {'db_engine', 'config', 'self'}
})
self._check_config_version(config)
self.config = config
self.config['random_seed'] = self.config.get('random_seed',
random.randint(1, 1e7))
random.seed(self.config['random_seed'])
self.project_storage = ProjectStorage(project_path)
self.model_storage_engine = ModelStorageEngine(self.project_storage)
self.matrix_storage_engine = MatrixStorageEngine(
self.project_storage, matrix_storage_class)
self.project_path = project_path
self.replace = replace
self.save_predictions = save_predictions
self.skip_validation = skip_validation
self.db_engine = db_engine
results_schema.upgrade_if_clean(dburl=self.db_engine.url)
self.features_schema_name = "features"
self.materialize_subquery_fromobjs = materialize_subquery_fromobjs
self.features_ignore_cohort = features_ignore_cohort
# only fill default values for full runs
if not partial_run:
## Defaults to sane values
self.config['temporal_config'] = fill_timechop_config_missing(
self.config, self.db_engine)
## Defaults to all the entities found in the features_aggregation's from_obj
self.config['cohort_config'] = fill_cohort_config_missing(
self.config)
## Defaults to all the feature_aggregation's prefixes
self.config[
'feature_group_definition'] = fill_feature_group_definition(
self.config)
grid_config = fill_model_grid_presets(self.config)
self.config.pop('model_grid_preset', None)
if grid_config is not None:
self.config['grid_config'] = grid_config
###################### RUBICON ######################
self.experiment_hash = save_experiment_and_get_hash(
self.config, self.db_engine)
self.run_id = initialize_tracking_and_get_run_id(
self.experiment_hash,
experiment_class_path=classpath(self.__class__),
experiment_kwargs=experiment_kwargs,
db_engine=self.db_engine)
self.initialize_components()
self.cleanup = cleanup
if self.cleanup:
logging.info(
"cleanup is set to True, so intermediate tables (labels and cohort) "
"will be removed after matrix creation and subset tables will be "
"removed after model training and testing")
else:
logging.info(
"cleanup is set to False, so intermediate tables (labels, cohort, and subsets) "
"will not be removed")
self.cleanup_timeout = (self.cleanup_timeout if cleanup_timeout is None
else cleanup_timeout)
self.profile = profile
logging.info("Generate profiling stats? (profile option): %s",
self.profile)
class ExperimentBase(ABC):
"""The base class for all Experiments.
Subclasses must implement the following four methods:
process_query_tasks
process_matrix_build_tasks
process_train_tasks
process_model_test_tasks
Look at singlethreaded.py for reference implementation of each.
Args:
config (dict)
db_engine (triage.util.db.SerializableDbEngine or sqlalchemy.engine.Engine)
project_path (string)
replace (bool)
cleanup_timeout (int)
"""
cleanup_timeout = 60 # seconds
def __init__(
self,
config,
db_engine,
project_path=None,
matrix_storage_class=CSVMatrixStore,
replace=True,
cleanup=False,
cleanup_timeout=None,
):
self._check_config_version(config)
self.config = config
if isinstance(db_engine, Engine):
logging.warning('Raw, unserializable SQLAlchemy engine passed. URL will be used, other options may be lost in multi-process environments')
self.db_engine = create_engine(db_engine.url)
else:
self.db_engine = db_engine
self.project_storage = ProjectStorage(project_path)
self.model_storage_engine = ModelStorageEngine(self.project_storage)
self.matrix_storage_engine = MatrixStorageEngine(self.project_storage, matrix_storage_class)
self.project_path = project_path
self.replace = replace
upgrade_db(db_engine=self.db_engine)
self.features_schema_name = 'features'
self.experiment_hash = save_experiment_and_get_hash(self.config,
self.db_engine)
self.labels_table_name = 'labels_{}'.format(self.experiment_hash)
self.initialize_components()
self.cleanup = cleanup
if self.cleanup:
logging.info('cleanup is set to True, so intermediate tables (labels and states) will be removed after matrix creation')
else:
logging.info('cleanup is set to False, so intermediate tables (labels and states) will not be removed after matrix creation')
self.cleanup_timeout = (self.cleanup_timeout if cleanup_timeout is None
else cleanup_timeout)
def _check_config_version(self, config):
if 'config_version' in config:
config_version = config['config_version']
else:
logging.warning('config_version key not found in experiment config. '
'Assuming v1, which may not be correct')
config_version = 'v1'
if config_version != CONFIG_VERSION:
raise ValueError(
"Experiment config '{}' "
"does not match current version '{}'. "
"Will not run experiment."
.format(config_version, CONFIG_VERSION)
)
def initialize_components(self):
split_config = self.config['temporal_config']
self.chopper = Timechop(**split_config)
cohort_config = self.config.get('cohort_config', {})
if 'query' in cohort_config:
self.state_table_generator = StateTableGeneratorFromQuery(
experiment_hash=self.experiment_hash,
db_engine=self.db_engine,
query=cohort_config['query']
)
elif 'entities_table' in cohort_config:
self.state_table_generator = StateTableGeneratorFromEntities(
experiment_hash=self.experiment_hash,
db_engine=self.db_engine,
entities_table=cohort_config['entities_table']
)
elif 'dense_states' in cohort_config:
self.state_table_generator = StateTableGeneratorFromDense(
experiment_hash=self.experiment_hash,
db_engine=self.db_engine,
dense_state_table=cohort_config['dense_states']['table_name']
)
else:
logging.warning('cohort_config missing or unrecognized. Without a cohort, you will not be able to make matrices or perform feature imputation.')
self.state_table_generator = StateTableGeneratorNoOp()
if 'label_config' in self.config:
self.label_generator = LabelGenerator(
label_name=self.config['label_config'].get('name', None),
query=self.config['label_config']['query'],
db_engine=self.db_engine,
)
else:
self.label_generator = LabelGeneratorNoOp()
logging.warning('label_config missing or unrecognized. Without labels, you will not be able to make matrices.')
self.feature_dictionary_creator = FeatureDictionaryCreator(
features_schema_name=self.features_schema_name,
db_engine=self.db_engine,
)
self.feature_generator = FeatureGenerator(
features_schema_name=self.features_schema_name,
replace=self.replace,
db_engine=self.db_engine,
feature_start_time=split_config['feature_start_time']
)
self.feature_group_creator = FeatureGroupCreator(
self.config.get('feature_group_definition', {'all': [True]})
)
self.feature_group_mixer = FeatureGroupMixer(
self.config.get('feature_group_strategies', ['all'])
)
self.planner = Planner(
feature_start_time=dt_from_str(split_config['feature_start_time']),
label_names=[self.config.get('label_config', {}).get('name', DEFAULT_LABEL_NAME)],
label_types=['binary'],
cohort_name=self.config.get('cohort_config', {}).get('name', None),
states=self.config.get('cohort_config', {}).get('dense_states', {})
.get('state_filters', []),
user_metadata=self.config.get('user_metadata', {}),
)
self.matrix_builder = MatrixBuilder(
db_config={
'features_schema_name': self.features_schema_name,
'labels_schema_name': 'public',
'labels_table_name': self.labels_table_name,
# TODO: have planner/builder take state table later on, so we
# can grab it from the StateTableGenerator instead of
# duplicating it here
'sparse_state_table_name': self.sparse_states_table_name,
},
matrix_storage_engine=self.matrix_storage_engine,
include_missing_labels_in_train_as=self.config.get('label_config', {})
.get('include_missing_labels_in_train_as', None),
engine=self.db_engine,
replace=self.replace
)
self.trainer = ModelTrainer(
experiment_hash=self.experiment_hash,
model_storage_engine=self.model_storage_engine,
model_grouper=ModelGrouper(self.config.get('model_group_keys', [])),
db_engine=self.db_engine,
replace=self.replace
)
self.tester = ModelTester(
model_storage_engine=self.model_storage_engine,
matrix_storage_engine=self.matrix_storage_engine,
replace=self.replace,
db_engine=self.db_engine,
individual_importance_config=self.config.get('individual_importance', {}),
evaluator_config=self.config.get('scoring', {})
)
@property
def sparse_states_table_name(self):
return 'tmp_sparse_states_{}'.format(self.experiment_hash)
@cachedproperty
def split_definitions(self):
"""Temporal splits based on the experiment's configuration
Returns: (dict) temporal splits
Example:
```
{
'feature_start_time': {datetime},
'feature_end_time': {datetime},
'label_start_time': {datetime},
'label_end_time': {datetime},
'train_matrix': {
'first_as_of_time': {datetime},
'last_as_of_time': {datetime},
'matrix_info_end_time': {datetime},
'training_label_timespan': {str},
'training_as_of_date_frequency': {str},
'max_training_history': {str},
'as_of_times': [list of {datetime}s]
},
'test_matrices': [list of matrix defs similar to train_matrix]
}
```
(When updating/setting split definitions, matrices should have
UUIDs.)
"""
split_definitions = self.chopper.chop_time()
logging.info('Computed and stored split definitions: %s',
split_definitions)
logging.info('\n----TIME SPLIT SUMMARY----\n')
logging.info('Number of time splits: {}'.format(len(split_definitions)))
for split_index, split in enumerate(split_definitions):
train_times = split['train_matrix']['as_of_times']
test_times = [as_of_time for test_matrix in split['test_matrices']
for as_of_time in test_matrix['as_of_times']]
logging.info('''Split index {}:
Training as_of_time_range: {} to {} ({} total)
Testing as_of_time range: {} to {} ({} total)\n\n'''.format(
split_index,
min(train_times),
max(train_times),
len(train_times),
min(test_times),
max(test_times),
len(test_times)
))
return split_definitions
@cachedproperty
def all_as_of_times(self):
"""All 'as of times' in experiment config
Used for label and feature generation.
Returns: (list) of datetimes
"""
all_as_of_times = []
for split in self.split_definitions:
all_as_of_times.extend(split['train_matrix']['as_of_times'])
logging.debug('Adding as_of_times from train matrix: %s',
split['train_matrix']['as_of_times'])
for test_matrix in split['test_matrices']:
logging.debug('Adding as_of_times from test matrix: %s',
test_matrix['as_of_times'])
all_as_of_times.extend(test_matrix['as_of_times'])
logging.info(
'Computed %s total as_of_times for label and feature generation',
len(all_as_of_times)
)
distinct_as_of_times = list(set(all_as_of_times))
logging.info(
'Computed %s distinct as_of_times for label and feature generation',
len(distinct_as_of_times)
)
logging.info('You can view all as_of_times by inspecting `.all_as_of_times` on this Experiment')
return distinct_as_of_times
@cachedproperty
def collate_aggregations(self):
"""Collation of ``Aggregation`` objects used by this experiment.
Returns: (list) of ``collate.Aggregation`` objects
"""
logging.info('Creating collate aggregations')
cohort_table = self.state_table_generator.sparse_table_name
if 'feature_aggregations' not in self.config:
logging.warning('No feature_aggregation config is available')
return []
return self.feature_generator.aggregations(
feature_aggregation_config=self.config['feature_aggregations'],
feature_dates=self.all_as_of_times,
state_table=cohort_table
)
@cachedproperty
def feature_aggregation_table_tasks(self):
"""All feature table query tasks specified by this
``Experiment``.
Returns: (dict) keys are group table names, values are
themselves dicts, each with keys for different stages of
table creation (prepare, inserts, finalize) and with values
being lists of SQL commands
"""
logging.info('Calculating feature tasks for %s as_of_times',
len(self.all_as_of_times))
return self.feature_generator.generate_all_table_tasks(
self.collate_aggregations,
task_type='aggregation'
)
@cachedproperty
def feature_imputation_table_tasks(self):
"""All feature imputation query tasks specified by this
``Experiment``.
Returns: (dict) keys are group table names, values are
themselves dicts, each with keys for different stages of
table creation (prepare, inserts, finalize) and with values
being lists of SQL commands
"""
logging.info('Calculating feature tasks for %s as_of_times',
len(self.all_as_of_times))
return self.feature_generator.generate_all_table_tasks(
self.collate_aggregations,
task_type='imputation'
)
@cachedproperty
def master_feature_dictionary(self):
"""All possible features found in the database. Not all features
will necessarily end up in matrices
Returns: (list) of dicts, keys being feature table names and
values being lists of feature names
"""
result = self.feature_dictionary_creator.feature_dictionary(
feature_table_names=self.feature_imputation_table_tasks.keys(),
index_column_lookup=self.feature_generator.index_column_lookup(
self.collate_aggregations
)
)
logging.info('Computed master feature dictionary: %s', result)
return result
@property
def feature_dicts(self):
"""Feature dictionaries, representing the feature tables and
columns configured in this experiment after computing feature
groups.
Returns: (list) of dicts, keys being feature table names and
values being lists of feature names
"""
return self.feature_group_mixer.generate(
self.feature_group_creator.subsets(self.master_feature_dictionary)
)
@cachedproperty
def matrix_build_tasks(self):
"""Tasks for all matrices that need to be built as a part of
this Experiment.
Each task contains arguments understood by
``Architect.build_matrix``.
Returns: (list) of dicts
"""
if not table_has_data(self.sparse_states_table_name, self.db_engine):
logging.warning('cohort table is not populated, cannot build any matrices')
return {}
if not table_has_data(self.labels_table_name, self.db_engine):
logging.warning('labels table is not populated, cannot build any matrices')
return {}
(
updated_split_definitions,
matrix_build_tasks
) = self.planner.generate_plans(
self.split_definitions,
self.feature_dicts
)
self.full_matrix_definitions = updated_split_definitions
return matrix_build_tasks
@cachedproperty
def full_matrix_definitions(self):
"""Full matrix definitions
Returns: (list) temporal and feature information for each matrix
"""
(
updated_split_definitions,
matrix_build_tasks
) = self.planner.generate_plans(
self.split_definitions,
self.feature_dicts
)
self.matrix_build_tasks = matrix_build_tasks
return updated_split_definitions
@property
def all_label_timespans(self):
"""All train and test label timespans
Returns: (list) label timespans, in string form as they appeared in the experiment config
"""
return list(set(
self.config['temporal_config']['training_label_timespans'] +
self.config['temporal_config']['test_label_timespans']
))
def generate_labels(self):
"""Generate labels based on experiment configuration
Results are stored in the database, not returned
"""
self.label_generator.generate_all_labels(
self.labels_table_name,
self.all_as_of_times,
self.all_label_timespans
)
def generate_cohort(self):
self.state_table_generator.generate_sparse_table(
as_of_dates=self.all_as_of_times
)
def log_split(self, split_num, split):
logging.info(
'Starting train/test for %s out of %s: train range: %s to %s',
split_num+1,
len(self.full_matrix_definitions),
split['train_matrix']['first_as_of_time'],
split['train_matrix']['matrix_info_end_time'],
)
@abstractmethod
def process_train_tasks(self, train_tasks):
pass
@abstractmethod
def process_query_tasks(self, query_tasks):
pass
@abstractmethod
def process_matrix_build_tasks(self, matrix_build_tasks):
pass
def generate_preimputation_features(self):
self.process_query_tasks(self.feature_aggregation_table_tasks)
logging.info('Finished running preimputation feature queries. The final results are in tables: %s',
','.join(agg.get_table_name() for agg in self.collate_aggregations)
)
def impute_missing_features(self):
self.process_query_tasks(self.feature_imputation_table_tasks)
logging.info('Finished running postimputation feature queries. The final results are in tables: %s',
','.join(agg.get_table_name(imputed=True) for agg in self.collate_aggregations)
)
def build_matrices(self):
self.process_matrix_build_tasks(self.matrix_build_tasks)
def generate_matrices(self):
logging.info('Creating cohort')
self.generate_cohort()
logging.info('Creating labels')
self.generate_labels()
logging.info('Creating feature aggregation tables')
self.generate_preimputation_features()
logging.info('Creating feature imputation tables')
self.impute_missing_features()
logging.info('Building all matrices')
self.build_matrices()
def train_and_test_models(self):
if 'grid_config' not in self.config:
logging.warning('No grid_config was passed in the experiment config. No models will be trained')
return
for split_num, split in enumerate(self.full_matrix_definitions):
self.log_split(split_num, split)
train_store = self.matrix_storage_engine.get_store(split['train_uuid'])
if train_store.empty:
logging.warning('''Train matrix for split %s was empty,
no point in training this model. Skipping
''', split['train_uuid'])
continue
if len(train_store.labels().unique()) == 1:
logging.warning('''Train Matrix for split %s had only one
unique value, no point in training this model. Skipping
''', split['train_uuid'])
continue
logging.info('Training models')
train_tasks = self.trainer.generate_train_tasks(
grid_config=self.config['grid_config'],
misc_db_parameters=dict(
test=False,
model_comment=self.config.get('model_comment', None),
),
matrix_store=train_store
)
model_ids = self.process_train_tasks(train_tasks)
logging.info('Done training models for split %s', split_num)
test_tasks = self.tester.generate_model_test_tasks(
split=split,
train_store=train_store,
model_ids=model_ids,
)
logging.info('Found %s non-empty test matrices for split %s', len(test_tasks), split_num)
self.process_model_test_tasks(test_tasks)
def validate(self, strict=True):
ExperimentValidator(self.db_engine, strict=strict).run(self.config)
def _run(self):
try:
logging.info('Generating matrices')
self.generate_matrices()
finally:
if self.cleanup:
self.clean_up_tables()
self.train_and_test_models()
def clean_up_tables(self):
logging.info('Cleaning up state and labels tables')
with timeout(self.cleanup_timeout):
self.state_table_generator.clean_up()
self.label_generator.clean_up(self.labels_table_name)
def run(self):
try:
self._run()
except Exception:
logging.exception('Run interrupted by uncaught exception')
raise
__call__ = run
class ExperimentBase(ABC):
"""The base class for all Experiments.
Subclasses must implement the following four methods:
process_query_tasks
process_matrix_build_tasks
process_train_tasks
process_model_test_tasks
Look at singlethreaded.py for reference implementation of each.
Args:
config (dict)
db_engine (triage.util.db.SerializableDbEngine or sqlalchemy.engine.Engine)
project_path (string)
replace (bool)
cleanup_timeout (int)
materialize_subquery_fromobjs (bool, default True) Whether or not to create and index
tables for feature "from objects" that are subqueries. Can speed up performance
when building features for many as-of-dates.
profile (bool)
"""
cleanup_timeout = 60 # seconds
def __init__(
self,
config,
db_engine,
project_path=None,
matrix_storage_class=CSVMatrixStore,
replace=True,
cleanup=False,
cleanup_timeout=None,
materialize_subquery_fromobjs=True,
profile=False,
):
self._check_config_version(config)
self.config = config
self.project_storage = ProjectStorage(project_path)
self.model_storage_engine = ModelStorageEngine(self.project_storage)
self.matrix_storage_engine = MatrixStorageEngine(
self.project_storage, matrix_storage_class
)
self.project_path = project_path
self.replace = replace
self.db_engine = db_engine
upgrade_db(db_engine=self.db_engine)
self.features_schema_name = "features"
self.materialize_subquery_fromobjs = materialize_subquery_fromobjs
self.experiment_hash = save_experiment_and_get_hash(self.config, self.db_engine)
self.labels_table_name = "labels_{}".format(self.experiment_hash)
self.cohort_table_name = "cohort_{}".format(self.experiment_hash)
self.initialize_components()
self.cleanup = cleanup
if self.cleanup:
logging.info(
"cleanup is set to True, so intermediate tables (labels and states) "
"will be removed after matrix creation"
)
else:
logging.info(
"cleanup is set to False, so intermediate tables (labels and states) "
"will not be removed after matrix creation"
)
self.cleanup_timeout = (
self.cleanup_timeout if cleanup_timeout is None else cleanup_timeout
)
self.profile = profile
logging.info("Generate profiling stats? (profile option): %s", self.profile)
def _check_config_version(self, config):
if "config_version" in config:
config_version = config["config_version"]
else:
logging.warning(
"config_version key not found in experiment config. "
"Assuming v1, which may not be correct"
)
config_version = "v1"
if config_version != CONFIG_VERSION:
raise ValueError(
"Experiment config '{}' "
"does not match current version '{}'. "
"Will not run experiment.".format(config_version, CONFIG_VERSION)
)
def initialize_components(self):
split_config = self.config["temporal_config"]
self.chopper = Timechop(**split_config)
cohort_config = self.config.get("cohort_config", {})
if "query" in cohort_config:
self.cohort_table_generator = CohortTableGenerator(
cohort_table_name=self.cohort_table_name,
db_engine=self.db_engine,
query=cohort_config["query"],
replace=self.replace
)
else:
logging.warning(
"cohort_config missing or unrecognized. Without a cohort, "
"you will not be able to make matrices or perform feature imputation."
)
self.cohort_table_generator = CohortTableGeneratorNoOp()
if "label_config" in self.config:
self.label_generator = LabelGenerator(
label_name=self.config["label_config"].get("name", None),
query=self.config["label_config"]["query"],
replace=self.replace,
db_engine=self.db_engine,
)
else:
self.label_generator = LabelGeneratorNoOp()
logging.warning(
"label_config missing or unrecognized. Without labels, "
"you will not be able to make matrices."
)
self.feature_dictionary_creator = FeatureDictionaryCreator(
features_schema_name=self.features_schema_name, db_engine=self.db_engine
)
self.feature_generator = FeatureGenerator(
features_schema_name=self.features_schema_name,
replace=self.replace,
db_engine=self.db_engine,
feature_start_time=split_config["feature_start_time"],
materialize_subquery_fromobjs=self.materialize_subquery_fromobjs
)
self.feature_group_creator = FeatureGroupCreator(
self.config.get("feature_group_definition", {"all": [True]})
)
self.feature_group_mixer = FeatureGroupMixer(
self.config.get("feature_group_strategies", ["all"])
)
self.planner = Planner(
feature_start_time=dt_from_str(split_config["feature_start_time"]),
label_names=[
self.config.get("label_config", {}).get("name", DEFAULT_LABEL_NAME)
],
label_types=["binary"],
cohort_names=[self.config.get("cohort_config", {}).get("name", None)],
user_metadata=self.config.get("user_metadata", {}),
)
self.matrix_builder = MatrixBuilder(
db_config={
"features_schema_name": self.features_schema_name,
"labels_schema_name": "public",
"labels_table_name": self.labels_table_name,
"cohort_table_name": self.cohort_table_name,
},
matrix_storage_engine=self.matrix_storage_engine,
experiment_hash=self.experiment_hash,
include_missing_labels_in_train_as=self.config.get("label_config", {}).get(
"include_missing_labels_in_train_as", None
),
engine=self.db_engine,
replace=self.replace,
)
self.trainer = ModelTrainer(
experiment_hash=self.experiment_hash,
model_storage_engine=self.model_storage_engine,
model_grouper=ModelGrouper(self.config.get("model_group_keys", [])),
db_engine=self.db_engine,
replace=self.replace,
)
self.tester = ModelTester(
model_storage_engine=self.model_storage_engine,
matrix_storage_engine=self.matrix_storage_engine,
replace=self.replace,
db_engine=self.db_engine,
individual_importance_config=self.config.get("individual_importance", {}),
evaluator_config=self.config.get("scoring", {}),
)
@cachedproperty
def split_definitions(self):
"""Temporal splits based on the experiment's configuration
Returns: (dict) temporal splits
Example:
```
{
'feature_start_time': {datetime},
'feature_end_time': {datetime},
'label_start_time': {datetime},
'label_end_time': {datetime},
'train_matrix': {
'first_as_of_time': {datetime},
'last_as_of_time': {datetime},
'matrix_info_end_time': {datetime},
'training_label_timespan': {str},
'training_as_of_date_frequency': {str},
'max_training_history': {str},
'as_of_times': [list of {datetime}s]
},
'test_matrices': [list of matrix defs similar to train_matrix]
}
```
(When updating/setting split definitions, matrices should have
UUIDs.)
"""
split_definitions = self.chopper.chop_time()
logging.info("Computed and stored split definitions: %s", split_definitions)
logging.info("\n----TIME SPLIT SUMMARY----\n")
logging.info("Number of time splits: {}".format(len(split_definitions)))
for split_index, split in enumerate(split_definitions):
train_times = split["train_matrix"]["as_of_times"]
test_times = [
as_of_time
for test_matrix in split["test_matrices"]
for as_of_time in test_matrix["as_of_times"]
]
logging.info(
"""Split index {}:
Training as_of_time_range: {} to {} ({} total)
Testing as_of_time range: {} to {} ({} total)\n\n""".format(
split_index,
min(train_times),
max(train_times),
len(train_times),
min(test_times),
max(test_times),
len(test_times),
)
)
return split_definitions
@cachedproperty
def all_as_of_times(self):
"""All 'as of times' in experiment config
Used for label and feature generation.
Returns: (list) of datetimes
"""
all_as_of_times = []
for split in self.split_definitions:
all_as_of_times.extend(split["train_matrix"]["as_of_times"])
logging.debug(
"Adding as_of_times from train matrix: %s",
split["train_matrix"]["as_of_times"],
)
for test_matrix in split["test_matrices"]:
logging.debug(
"Adding as_of_times from test matrix: %s",
test_matrix["as_of_times"],
)
all_as_of_times.extend(test_matrix["as_of_times"])
logging.info(
"Computed %s total as_of_times for label and feature generation",
len(all_as_of_times),
)
distinct_as_of_times = list(set(all_as_of_times))
logging.info(
"Computed %s distinct as_of_times for label and feature generation",
len(distinct_as_of_times),
)
logging.info(
"You can view all as_of_times by inspecting `.all_as_of_times` on this Experiment"
)
return distinct_as_of_times
@cachedproperty
def collate_aggregations(self):
"""Collation of ``Aggregation`` objects used by this experiment.
Returns: (list) of ``collate.Aggregation`` objects
"""
logging.info("Creating collate aggregations")
if "feature_aggregations" not in self.config:
logging.warning("No feature_aggregation config is available")
return []
return self.feature_generator.aggregations(
feature_aggregation_config=self.config["feature_aggregations"],
feature_dates=self.all_as_of_times,
state_table=self.cohort_table_name,
)
@cachedproperty
def feature_aggregation_table_tasks(self):
"""All feature table query tasks specified by this
``Experiment``.
Returns: (dict) keys are group table names, values are
themselves dicts, each with keys for different stages of
table creation (prepare, inserts, finalize) and with values
being lists of SQL commands
"""
logging.info(
"Calculating feature tasks for %s as_of_times", len(self.all_as_of_times)
)
return self.feature_generator.generate_all_table_tasks(
self.collate_aggregations, task_type="aggregation"
)
@cachedproperty
def feature_imputation_table_tasks(self):
"""All feature imputation query tasks specified by this
``Experiment``.
Returns: (dict) keys are group table names, values are
themselves dicts, each with keys for different stages of
table creation (prepare, inserts, finalize) and with values
being lists of SQL commands
"""
logging.info(
"Calculating feature tasks for %s as_of_times", len(self.all_as_of_times)
)
return self.feature_generator.generate_all_table_tasks(
self.collate_aggregations, task_type="imputation"
)
@cachedproperty
def master_feature_dictionary(self):
"""All possible features found in the database. Not all features
will necessarily end up in matrices
Returns: (list) of dicts, keys being feature table names and
values being lists of feature names
"""
result = self.feature_dictionary_creator.feature_dictionary(
feature_table_names=self.feature_imputation_table_tasks.keys(),
index_column_lookup=self.feature_generator.index_column_lookup(
self.collate_aggregations
),
)
logging.info("Computed master feature dictionary: %s", result)
return result
@property
def feature_dicts(self):
"""Feature dictionaries, representing the feature tables and
columns configured in this experiment after computing feature
groups.
Returns: (list) of dicts, keys being feature table names and
values being lists of feature names
"""
return self.feature_group_mixer.generate(
self.feature_group_creator.subsets(self.master_feature_dictionary)
)
@cachedproperty
def matrix_build_tasks(self):
"""Tasks for all matrices that need to be built as a part of
this Experiment.
Each task contains arguments understood by
``Architect.build_matrix``.
Returns: (list) of dicts
"""
if not table_has_data(self.cohort_table_name, self.db_engine):
logging.warning("cohort table is not populated, cannot build any matrices")
return {}
if not table_has_data(self.labels_table_name, self.db_engine):
logging.warning("labels table is not populated, cannot build any matrices")
return {}
(updated_split_definitions, matrix_build_tasks) = self.planner.generate_plans(
self.split_definitions, self.feature_dicts
)
self.full_matrix_definitions = updated_split_definitions
return matrix_build_tasks
@cachedproperty
def full_matrix_definitions(self):
"""Full matrix definitions
Returns: (list) temporal and feature information for each matrix
"""
(updated_split_definitions, matrix_build_tasks) = self.planner.generate_plans(
self.split_definitions, self.feature_dicts
)
self.matrix_build_tasks = matrix_build_tasks
return updated_split_definitions
@property
def all_label_timespans(self):
"""All train and test label timespans
Returns: (list) label timespans, in string form as they appeared in the experiment config
"""
return list(
set(
self.config["temporal_config"]["training_label_timespans"]
+ self.config["temporal_config"]["test_label_timespans"]
)
)
def generate_labels(self):
"""Generate labels based on experiment configuration
Results are stored in the database, not returned
"""
self.label_generator.generate_all_labels(
self.labels_table_name, self.all_as_of_times, self.all_label_timespans
)
def generate_cohort(self):
self.cohort_table_generator.generate_cohort_table(
as_of_dates=self.all_as_of_times
)
def log_split(self, split_num, split):
logging.info(
"Starting train/test for %s out of %s: train range: %s to %s",
split_num + 1,
len(self.full_matrix_definitions),
split["train_matrix"]["first_as_of_time"],
split["train_matrix"]["matrix_info_end_time"],
)
@abstractmethod
def process_train_tasks(self, train_tasks):
pass
@abstractmethod
def process_query_tasks(self, query_tasks):
pass
@abstractmethod
def process_matrix_build_tasks(self, matrix_build_tasks):
pass
def generate_preimputation_features(self):
self.process_query_tasks(self.feature_aggregation_table_tasks)
logging.info(
"Finished running preimputation feature queries. The final results are in tables: %s",
",".join(agg.get_table_name() for agg in self.collate_aggregations),
)
def impute_missing_features(self):
self.process_query_tasks(self.feature_imputation_table_tasks)
logging.info(
"Finished running postimputation feature queries. The final results are in tables: %s",
",".join(
agg.get_table_name(imputed=True) for agg in self.collate_aggregations
),
)
def build_matrices(self):
associate_matrices_with_experiment(
self.experiment_hash,
self.matrix_build_tasks.keys(),
self.db_engine
)
self.process_matrix_build_tasks(self.matrix_build_tasks)
def generate_matrices(self):
logging.info("Creating cohort")
self.generate_cohort()
logging.info("Creating labels")
self.generate_labels()
logging.info("Creating feature aggregation tables")
self.generate_preimputation_features()
logging.info("Creating feature imputation tables")
self.impute_missing_features()
logging.info("Building all matrices")
self.build_matrices()
def train_and_test_models(self):
if "grid_config" not in self.config:
logging.warning(
"No grid_config was passed in the experiment config. No models will be trained"
)
return
for split_num, split in enumerate(self.full_matrix_definitions):
self.log_split(split_num, split)
train_store = self.matrix_storage_engine.get_store(split["train_uuid"])
if train_store.empty:
logging.warning(
"""Train matrix for split %s was empty,
no point in training this model. Skipping
""",
split["train_uuid"],
)
continue
if len(train_store.labels().unique()) == 1:
logging.warning(
"""Train Matrix for split %s had only one
unique value, no point in training this model. Skipping
""",
split["train_uuid"],
)
continue
logging.info("Training models")
train_tasks = self.trainer.generate_train_tasks(
grid_config=self.config["grid_config"],
misc_db_parameters=dict(
test=False, model_comment=self.config.get("model_comment", None)
),
matrix_store=train_store,
)
associate_models_with_experiment(
self.experiment_hash,
[train_task['model_hash'] for train_task in train_tasks],
self.db_engine
)
model_ids = self.process_train_tasks(train_tasks)
logging.info("Done training models for split %s", split_num)
test_tasks = self.tester.generate_model_test_tasks(
split=split, train_store=train_store, model_ids=model_ids
)
logging.info(
"Found %s non-empty test matrices for split %s",
len(test_tasks),
split_num,
)
self.process_model_test_tasks(test_tasks)
def validate(self, strict=True):
ExperimentValidator(self.db_engine, strict=strict).run(self.config)
def _run(self):
try:
logging.info("Generating matrices")
self.generate_matrices()
finally:
if self.cleanup:
self.clean_up_tables()
self.train_and_test_models()
logging.info("Experiment complete")
self._log_end_of_run_report()
def _log_end_of_run_report(self):
missing_models = missing_model_hashes(self.experiment_hash, self.db_engine)
if len(missing_models) > 0:
logging.info("Found %s missing model hashes."
"This means that they were supposed to either be trained or reused"
"by this experiment but are not present in the models table."
"Inspect the logs for any training errors. Full list: %s",
len(missing_models),
missing_models
)
else:
logging.info("All models that were supposed to be trained were trained. Awesome!")
missing_matrices = missing_matrix_uuids(self.experiment_hash, self.db_engine)
if len(missing_matrices) > 0:
logging.info("Found %s missing matrix uuids."
"This means that they were supposed to either be build or reused"
"by this experiment but are not present in the matrices table."
"Inspect the logs for any matrix building errors. Full list: %s",
len(missing_matrices),
missing_matrices
)
else:
logging.info("All matrices that were supposed to be build were built. Awesome!")
def clean_up_tables(self):
logging.info("Cleaning up state and labels tables")
with timeout(self.cleanup_timeout):
self.cohort_table_generator.clean_up()
self.label_generator.clean_up(self.labels_table_name)
def _run_profile(self):
cp = cProfile.Profile()
cp.runcall(self._run)
store = self.project_storage.get_store(
["profiling_stats"],
f"{int(time.time())}.profile"
)
with store.open('wb') as fd:
cp.create_stats()
marshal.dump(cp.stats, fd)
logging.info("Profiling stats of this Triage run calculated and written to %s"
"in cProfile format.",
store)
def run(self):
try:
if self.profile:
self._run_profile()
else:
self._run()
except Exception:
logging.exception("Run interrupted by uncaught exception")
raise
__call__ = run
def __init__(
self,
config,
db_engine,
project_path=None,
matrix_storage_class=CSVMatrixStore,
replace=True,
cleanup=False,
cleanup_timeout=None,
materialize_subquery_fromobjs=True,
features_ignore_cohort=False,
additional_bigtrain_classnames=None,
profile=False,
save_predictions=True,
skip_validation=False,
partial_run=False,
):
# For a partial run, skip validation and avoid cleaning up
# we'll also skip filling default config values below
if partial_run:
cleanup = False
skip_validation = True
experiment_kwargs = bind_kwargs(
self.__class__,
**{
key: value
for (key, value) in locals().items()
if key not in {"db_engine", "config", "self"}
},
)
self._check_config_version(config)
self.config = config
if self.config.get("cohort_config") is not None:
self.config["cohort_config"] = load_query_if_needed(
self.config["cohort_config"]
)
if self.config.get("label_config") is not None:
self.config["label_config"] = load_query_if_needed(
self.config["label_config"]
)
self.project_storage = ProjectStorage(project_path)
self.model_storage_engine = ModelStorageEngine(self.project_storage)
self.matrix_storage_engine = MatrixStorageEngine(
self.project_storage, matrix_storage_class
)
self.project_path = project_path
logger.verbose(
f"Matrices and trained models will be saved in {self.project_path}"
)
self.replace = replace
if self.replace:
logger.notice(
f"Replace flag is set to true. Matrices, models, "
"evaluations and predictions (if they exist) will be replaced"
)
self.save_predictions = save_predictions
if not self.save_predictions:
logger.notice(
f"Save predictions flag is set to false. "
"Individual predictions won't be stored in the predictions "
"table. This will decrease both the running time "
"of an experiment and also decrease the space needed in the db"
)
self.skip_validation = skip_validation
if self.skip_validation:
logger.notice(
f"Warning: Skip validation flag is set to true. "
"The experiment config file specified won't be validated. "
"This will reduce (a little) the running time of the experiment, "
"but has some potential risks, e.g. the experiment could fail"
"after some time due to some misconfiguration. Proceed with care."
)
self.db_engine = db_engine
results_schema.upgrade_if_clean(dburl=self.db_engine.url)
self.features_schema_name = "features"
self.materialize_subquery_fromobjs = materialize_subquery_fromobjs
if not self.materialize_subquery_fromobjs:
logger.notice(
"Materialize from_objs is set to false. "
"The from_objs will be calculated on the fly every time."
)
self.features_ignore_cohort = features_ignore_cohort
if self.features_ignore_cohort:
logger.notice(
"Features will be calculated for all the entities "
"(i.e. ignoring cohort) this setting will have the effect "
"that more db space will be used, but potentially could save "
"time if you are running several similar experiments with "
"different cohorts."
)
self.additional_bigtrain_classnames = additional_bigtrain_classnames
# only fill default values for full runs
if not partial_run:
## Defaults to sane values
self.config["temporal_config"] = fill_timechop_config_missing(
self.config, self.db_engine
)
## Defaults to all the entities found in the features_aggregation's from_obj
self.config["cohort_config"] = fill_cohort_config_missing(self.config)
## Defaults to all the feature_aggregation's prefixes
self.config["feature_group_definition"] = fill_feature_group_definition(
self.config
)
grid_config = fill_model_grid_presets(self.config)
self.config.pop("model_grid_preset", None)
if grid_config is not None:
self.config["grid_config"] = grid_config
if not self.config.get("random_seed", None):
logger.notice(
"Random seed not specified. A random seed will be provided. "
"This could have interesting side effects, "
"e.g. new models per model group are trained, "
"tested and evaluated everytime that you run this experiment configuration"
)
self.random_seed = self.config.pop("random_seed", random.randint(1, 1e7))
logger.verbose(
f"Using random seed [{self.random_seed}] for running the experiment"
)
random.seed(self.random_seed)
###################### RUBICON ######################
self.experiment_hash = save_experiment_and_get_hash(self.config, self.db_engine)
logger.debug(f"Experiment hash [{self.experiment_hash}] assigned")
self.run_id = initialize_tracking_and_get_run_id(
self.experiment_hash,
experiment_class_path=classpath(self.__class__),
random_seed=self.random_seed,
experiment_kwargs=experiment_kwargs,
db_engine=self.db_engine,
)
logger.debug(f"Experiment run id [{self.run_id}] assigned")
self.initialize_components()
self.cleanup = cleanup
if self.cleanup:
logger.notice(
"Cleanup is set to true, so intermediate tables (labels and cohort) "
"will be removed after matrix creation and subset tables will be "
"removed after model training and testing"
)
self.cleanup_timeout = (
self.cleanup_timeout if cleanup_timeout is None else cleanup_timeout
)
self.profile = profile
if self.profile:
logger.spam("Profiling will be stored using cProfile")
def __init__(
self,
config,
db_engine,
project_path=None,
matrix_storage_class=CSVMatrixStore,
replace=True,
cleanup=False,
cleanup_timeout=None,
materialize_subquery_fromobjs=True,
features_ignore_cohort=False,
profile=False,
save_predictions=True,
skip_validation=False,
):
experiment_kwargs = bind_kwargs(
self.__class__, **{
key: value
for (key, value) in locals().items()
if key not in {'db_engine', 'config', 'self'}
})
self._check_config_version(config)
self.config = config
random.seed(config['random_seed'])
self.project_storage = ProjectStorage(project_path)
self.model_storage_engine = ModelStorageEngine(self.project_storage)
self.matrix_storage_engine = MatrixStorageEngine(
self.project_storage, matrix_storage_class)
self.project_path = project_path
self.replace = replace
self.save_predictions = save_predictions
self.skip_validation = skip_validation
self.db_engine = db_engine
results_schema.upgrade_if_clean(dburl=self.db_engine.url)
self.features_schema_name = "features"
self.materialize_subquery_fromobjs = materialize_subquery_fromobjs
self.features_ignore_cohort = features_ignore_cohort
self.experiment_hash = save_experiment_and_get_hash(
self.config, self.db_engine)
self.run_id = initialize_tracking_and_get_run_id(
self.experiment_hash,
experiment_class_path=classpath(self.__class__),
experiment_kwargs=experiment_kwargs,
db_engine=self.db_engine)
self.initialize_components()
self.cleanup = cleanup
if self.cleanup:
logging.info(
"cleanup is set to True, so intermediate tables (labels and cohort) "
"will be removed after matrix creation and subset tables will be "
"removed after model training and testing")
else:
logging.info(
"cleanup is set to False, so intermediate tables (labels, cohort, and subsets) "
"will not be removed")
self.cleanup_timeout = (self.cleanup_timeout if cleanup_timeout is None
else cleanup_timeout)
self.profile = profile
logging.info("Generate profiling stats? (profile option): %s",
self.profile)