class ExperimentBase(ABC):
"""The base class for all Experiments.
Subclasses must implement the following four methods:
process_query_tasks
process_matrix_build_tasks
process_subset_tasks
process_train_test_batches
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,
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)
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))
@cachedproperty
def cohort_hash(self):
if "query" in self.config.get("cohort_config", {}):
return filename_friendly_hash(
self.config["cohort_config"]["query"])
else:
return None
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_name = "cohort_{}_{}".format(
cohort_config.get('name', 'default'), self.cohort_hash)
self.cohort_table_generator = EntityDateTableGenerator(
entity_date_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, perform feature imputation, "
"or save time by only computing features for that cohort.")
self.features_ignore_cohort = True
self.cohort_table_name = "cohort_{}".format(self.experiment_hash)
self.cohort_table_generator = EntityDateTableGeneratorNoOp()
self.subsets = [None] + self.config.get("scoring", {}).get(
"subsets", [])
if "label_config" in self.config:
label_config = self.config["label_config"]
self.labels_table_name = "labels_{}_{}".format(
label_config.get('name', 'default'),
filename_friendly_hash(label_config['query']))
self.label_generator = LabelGenerator(
label_name=label_config.get("name", None),
query=label_config["query"],
replace=self.replace,
db_engine=self.db_engine,
)
else:
self.labels_table_name = "labels_{}".format(self.experiment_hash)
self.label_generator = LabelGeneratorNoOp()
logging.warning(
"label_config missing or unrecognized. Without labels, "
"you will not be able to make matrices.")
if "bias_audit_config" in self.config:
bias_config = self.config["bias_audit_config"]
self.bias_hash = filename_friendly_hash(bias_config)
self.protected_groups_table_name = f"protected_groups_{self.bias_hash}"
self.protected_groups_generator = ProtectedGroupsGenerator(
db_engine=self.db_engine,
from_obj=parse_from_obj(bias_config, 'bias_from_obj'),
attribute_columns=bias_config.get("attribute_columns", None),
entity_id_column=bias_config.get("entity_id_column", None),
knowledge_date_column=bias_config.get("knowledge_date_column",
None),
protected_groups_table_name=self.protected_groups_table_name,
replace=self.replace)
else:
self.protected_groups_generator = ProtectedGroupsGeneratorNoOp()
logging.warning(
"bias_audit_config missing or unrecognized. Without protected groups, "
"you will not audit your models for bias and fairness.")
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,
features_ignore_cohort=self.features_ignore_cohort)
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,
run_id=self.run_id,
)
self.subsetter = Subsetter(db_engine=self.db_engine,
replace=self.replace,
as_of_times=self.all_as_of_times)
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,
run_id=self.run_id,
)
self.predictor = Predictor(
db_engine=self.db_engine,
model_storage_engine=self.model_storage_engine,
save_predictions=self.save_predictions,
replace=self.replace,
rank_order=self.config.get("prediction",
{}).get("rank_tiebreaker", "worst"),
)
self.individual_importance_calculator = IndividualImportanceCalculator(
db_engine=self.db_engine,
n_ranks=self.config.get("individual_importance",
{}).get("n_ranks", 5),
methods=self.config.get("individual_importance",
{}).get("methods", ["uniform"]),
replace=self.replace,
)
self.evaluator = ModelEvaluator(
db_engine=self.db_engine,
testing_metric_groups=self.config.get("scoring", {}).get(
"testing_metric_groups", []),
training_metric_groups=self.config.get("scoring", {}).get(
"training_metric_groups", []),
bias_config=self.config.get("bias_audit_config", {}))
self.model_train_tester = ModelTrainTester(
matrix_storage_engine=self.matrix_storage_engine,
model_evaluator=self.evaluator,
model_trainer=self.trainer,
individual_importance_calculator=self.
individual_importance_calculator,
predictor=self.predictor,
subsets=self.subsets,
protected_groups_generator=self.protected_groups_generator,
cohort_hash=self.cohort_hash)
def get_for_update(self):
return get_for_update(self.db_engine, results_schema.Experiment,
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),
))
with self.get_for_update() as experiment:
experiment.time_splits = len(split_definitions)
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"
)
with self.get_for_update() as experiment:
experiment.as_of_times = len(distinct_as_of_times)
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 []
aggregations = self.feature_generator.aggregations(
feature_aggregation_config=self.config["feature_aggregations"],
feature_dates=self.all_as_of_times,
state_table=self.cohort_table_name,
)
with self.get_for_update() as experiment:
experiment.feature_blocks = len(aggregations)
return aggregations
@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)
with self.get_for_update() as experiment:
experiment.total_features = sum(
1
for _feature in itertools.chain.from_iterable(result.values()))
return result
@cachedproperty
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
"""
combinations = self.feature_group_mixer.generate(
self.feature_group_creator.subsets(self.master_feature_dictionary))
with self.get_for_update() as experiment:
experiment.feature_group_combinations = len(combinations)
return combinations
@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"]))
@cachedproperty
def subset_tasks(self):
return self.subsetter.generate_tasks(self.subsets)
@experiment_entrypoint
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)
@experiment_entrypoint
def generate_cohort(self):
self.cohort_table_generator.generate_entity_date_table(
as_of_dates=self.all_as_of_times)
@experiment_entrypoint
def generate_protected_groups(self):
"""Generate protected groups table based on experiment configuration
Results are stored in the database, not returned
"""
self.protected_groups_generator.generate_all_dates(
self.all_as_of_times, self.cohort_table_name, self.cohort_hash)
def generate_subset(self, subset_hash):
self.subsets[
"subset_hash"].subset_table_generator.generate_entity_date_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_subset_tasks(self, subset_tasks):
pass
@abstractmethod
def process_train_test_batches(self, train_test_batches):
pass
@abstractmethod
def process_query_tasks(self, query_tasks):
pass
@abstractmethod
def process_matrix_build_tasks(self, matrix_build_tasks):
pass
@experiment_entrypoint
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),
)
@experiment_entrypoint
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)
with self.get_for_update() as experiment:
experiment.matrices_needed = len(self.matrix_build_tasks.keys())
record_matrix_building_started(self.run_id, self.db_engine)
self.process_matrix_build_tasks(self.matrix_build_tasks)
@experiment_entrypoint
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()
@experiment_entrypoint
def generate_subsets(self):
if self.subsets:
logging.info("Beginning subset generation")
self.process_subset_tasks(self.subset_tasks)
else:
logging.info(
"No subsets found. Proceeding to training and testing models")
def _all_train_test_batches(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
return self.model_train_tester.generate_task_batches(
splits=self.full_matrix_definitions,
grid_config=self.config.get('grid_config'),
model_comment=self.config.get('model_comment', None))
@experiment_entrypoint
def train_and_test_models(self):
self.generate_subsets()
logging.info("Creating protected groups table")
self.generate_protected_groups()
batches = self._all_train_test_batches()
if not batches:
logging.warning("No train/test tasks found, so no training to do")
return
with self.get_for_update() as experiment:
experiment.grid_size = sum(
1 for _param in self.trainer.flattened_grid_config(
self.config.get('grid_config')))
logging.info("%s train/test batches found. Beginning training.",
len(batches))
model_hashes = set(task['train_kwargs']['model_hash']
for batch in batches for task in batch.tasks)
associate_models_with_experiment(self.experiment_hash, model_hashes,
self.db_engine)
with self.get_for_update() as experiment:
experiment.models_needed = len(model_hashes)
record_model_building_started(self.run_id, self.db_engine)
self.process_train_test_batches(batches)
def validate(self, strict=True):
ExperimentValidator(self.db_engine, strict=strict).run(self.config)
def _run(self):
if not self.skip_validation:
self.validate()
logging.info("Generating matrices")
try:
self.generate_matrices()
self.train_and_test_models()
finally:
if self.cleanup:
self.clean_up_matrix_building_tables()
self.clean_up_subset_tables()
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_matrix_building_tables(self):
logging.info("Cleaning up cohort and labels tables")
with timeout(self.cleanup_timeout):
self.cohort_table_generator.clean_up()
self.label_generator.clean_up(self.labels_table_name)
def clean_up_subset_tables(self):
logging.info("Cleaning up cohort and labels tables")
with timeout(self.cleanup_timeout):
for subset_task in self.subset_tasks:
subset_task["subset_table_generator"].clean_up()
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)
@experiment_entrypoint
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 basic_integration_test(state_filters, feature_group_create_rules,
feature_group_mix_rules,
expected_matrix_multiplier):
with testing.postgresql.Postgresql() as postgresql:
db_engine = create_engine(postgresql.url())
Base.metadata.create_all(db_engine)
populate_source_data(db_engine)
with TemporaryDirectory() as temp_dir:
chopper = Timechop(
feature_start_time=datetime(2010, 1, 1),
feature_end_time=datetime(2014, 1, 1),
label_start_time=datetime(2011, 1, 1),
label_end_time=datetime(2014, 1, 1),
model_update_frequency='1year',
training_label_timespans=['6months'],
test_label_timespans=['6months'],
training_as_of_date_frequencies='1day',
test_as_of_date_frequencies='3months',
max_training_histories=['1months'],
test_durations=['1months'],
)
state_table_generator = StateTableGenerator(
db_engine=db_engine,
experiment_hash='abcd',
dense_state_table='states',
)
label_generator = BinaryLabelGenerator(db_engine=db_engine,
events_table='events')
feature_generator = FeatureGenerator(
db_engine=db_engine,
features_schema_name='features',
replace=True,
)
feature_dictionary_creator = FeatureDictionaryCreator(
db_engine=db_engine, features_schema_name='features')
feature_group_creator = FeatureGroupCreator(
feature_group_create_rules)
feature_group_mixer = FeatureGroupMixer(feature_group_mix_rules)
planner = Planner(engine=db_engine,
feature_start_time=datetime(2010, 1, 1),
label_names=['outcome'],
label_types=['binary'],
db_config={
'features_schema_name':
'features',
'labels_schema_name':
'public',
'labels_table_name':
'labels',
'sparse_state_table_name':
'tmp_sparse_states_abcd',
},
matrix_directory=os.path.join(
temp_dir, 'matrices'),
states=state_filters,
user_metadata={},
replace=True)
# chop time
split_definitions = chopper.chop_time()
num_split_matrices = sum(1 + len(split['test_matrices'])
for split in split_definitions)
# generate as_of_times for feature/label/state generation
all_as_of_times = []
for split in split_definitions:
all_as_of_times.extend(split['train_matrix']['as_of_times'])
for test_matrix in split['test_matrices']:
all_as_of_times.extend(test_matrix['as_of_times'])
all_as_of_times = list(set(all_as_of_times))
feature_aggregation_config = [{
'prefix':
'cat',
'from_obj':
'cat_complaints',
'knowledge_date_column':
'as_of_date',
'aggregates': [{
'quantity': 'cat_sightings',
'metrics': ['count', 'avg'],
'imputation': {
'all': {
'type': 'mean'
}
}
}],
'intervals': ['1y'],
'groups': ['entity_id']
}, {
'prefix':
'dog',
'from_obj':
'dog_complaints',
'knowledge_date_column':
'as_of_date',
'aggregates_imputation': {
'count': {
'type': 'constant',
'value': 7
},
'sum': {
'type': 'mean'
},
'avg': {
'type': 'zero'
}
},
'aggregates': [{
'quantity': 'dog_sightings',
'metrics': ['count', 'avg'],
}],
'intervals': ['1y'],
'groups': ['entity_id']
}]
state_table_generator.validate()
label_generator.validate()
feature_generator.validate(feature_aggregation_config)
feature_group_creator.validate()
planner.validate()
# generate sparse state table
state_table_generator.generate_sparse_table(
as_of_dates=all_as_of_times)
# create labels table
label_generator.generate_all_labels(labels_table='labels',
as_of_dates=all_as_of_times,
label_timespans=['6months'])
# create feature table tasks
# we would use FeatureGenerator#create_all_tables but want to use
# the tasks dict directly to create a feature dict
aggregations = feature_generator.aggregations(
feature_aggregation_config=[{
'prefix':
'cat',
'from_obj':
'cat_complaints',
'knowledge_date_column':
'as_of_date',
'aggregates': [{
'quantity': 'cat_sightings',
'metrics': ['count', 'avg'],
'imputation': {
'all': {
'type': 'mean'
}
}
}],
'intervals': ['1y'],
'groups': ['entity_id']
}, {
'prefix':
'dog',
'from_obj':
'dog_complaints',
'knowledge_date_column':
'as_of_date',
'aggregates_imputation': {
'count': {
'type': 'constant',
'value': 7
},
'sum': {
'type': 'mean'
},
'avg': {
'type': 'zero'
}
},
'aggregates': [{
'quantity': 'dog_sightings',
'metrics': ['count', 'avg'],
}],
'intervals': ['1y'],
'groups': ['entity_id']
}],
feature_dates=all_as_of_times,
state_table=state_table_generator.sparse_table_name)
feature_table_agg_tasks = feature_generator.generate_all_table_tasks(
aggregations, task_type='aggregation')
# create feature aggregation tables
feature_generator.process_table_tasks(feature_table_agg_tasks)
feature_table_imp_tasks = feature_generator.generate_all_table_tasks(
aggregations, task_type='imputation')
# create feature imputation tables
feature_generator.process_table_tasks(feature_table_imp_tasks)
# build feature dictionaries from feature tables and
# subsetting config
master_feature_dict = feature_dictionary_creator.feature_dictionary(
feature_table_names=feature_table_imp_tasks.keys(),
index_column_lookup=feature_generator.index_column_lookup(
aggregations))
feature_dicts = feature_group_mixer.generate(
feature_group_creator.subsets(master_feature_dict))
# figure out what matrices need to be built
_, matrix_build_tasks =\
planner.generate_plans(
split_definitions,
feature_dicts
)
# go and build the matrices
planner.build_all_matrices(matrix_build_tasks)
# super basic assertion: did matrices we expect get created?
matrix_directory = os.path.join(temp_dir, 'matrices')
matrices = [
path for path in os.listdir(matrix_directory) if '.csv' in path
]
metadatas = [
path for path in os.listdir(matrix_directory)
if '.yaml' in path
]
assert len(
matrices) == num_split_matrices * expected_matrix_multiplier
assert len(
metadatas) == num_split_matrices * expected_matrix_multiplier
def basic_integration_test(
cohort_names,
feature_group_create_rules,
feature_group_mix_rules,
expected_matrix_multiplier,
expected_group_lists,
):
with testing.postgresql.Postgresql() as postgresql:
db_engine = create_engine(postgresql.url())
Base.metadata.create_all(db_engine)
populate_source_data(db_engine)
with TemporaryDirectory() as temp_dir:
chopper = Timechop(
feature_start_time=datetime(2010, 1, 1),
feature_end_time=datetime(2014, 1, 1),
label_start_time=datetime(2011, 1, 1),
label_end_time=datetime(2014, 1, 1),
model_update_frequency="1year",
training_label_timespans=["6months"],
test_label_timespans=["6months"],
training_as_of_date_frequencies="1day",
test_as_of_date_frequencies="3months",
max_training_histories=["1months"],
test_durations=["1months"],
)
entity_date_table_generator = EntityDateTableGenerator(
db_engine=db_engine,
entity_date_table_name="cohort_abcd",
query="select distinct(entity_id) from events")
label_generator = LabelGenerator(
db_engine=db_engine,
query=sample_config()["label_config"]["query"])
feature_generator = FeatureGenerator(
db_engine=db_engine,
features_schema_name="features",
replace=True)
feature_dictionary_creator = FeatureDictionaryCreator(
db_engine=db_engine, features_schema_name="features")
feature_group_creator = FeatureGroupCreator(
feature_group_create_rules)
feature_group_mixer = FeatureGroupMixer(feature_group_mix_rules)
project_storage = ProjectStorage(temp_dir)
planner = Planner(
feature_start_time=datetime(2010, 1, 1),
label_names=["outcome"],
label_types=["binary"],
cohort_names=cohort_names,
user_metadata={},
)
builder = MatrixBuilder(
engine=db_engine,
db_config={
"features_schema_name": "features",
"labels_schema_name": "public",
"labels_table_name": "labels",
"cohort_table_name": "cohort_abcd",
},
experiment_hash=None,
matrix_storage_engine=project_storage.matrix_storage_engine(),
replace=True,
)
# chop time
split_definitions = chopper.chop_time()
num_split_matrices = sum(1 + len(split["test_matrices"])
for split in split_definitions)
# generate as_of_times for feature/label/state generation
all_as_of_times = []
for split in split_definitions:
all_as_of_times.extend(split["train_matrix"]["as_of_times"])
for test_matrix in split["test_matrices"]:
all_as_of_times.extend(test_matrix["as_of_times"])
all_as_of_times = list(set(all_as_of_times))
# generate entity_date state table
entity_date_table_generator.generate_entity_date_table(
as_of_dates=all_as_of_times)
# create labels table
label_generator.generate_all_labels(
labels_table="labels",
as_of_dates=all_as_of_times,
label_timespans=["6months"],
)
# create feature table tasks
# we would use FeatureGenerator#create_all_tables but want to use
# the tasks dict directly to create a feature dict
aggregations = feature_generator.aggregations(
feature_aggregation_config=[
{
"prefix":
"cat",
"from_obj":
"cat_complaints",
"knowledge_date_column":
"as_of_date",
"aggregates": [{
"quantity": "cat_sightings",
"metrics": ["count", "avg"],
"imputation": {
"all": {
"type": "mean"
}
},
}],
"intervals": ["1y"],
"groups": ["entity_id"],
},
{
"prefix":
"dog",
"from_obj":
"dog_complaints",
"knowledge_date_column":
"as_of_date",
"aggregates_imputation": {
"count": {
"type": "constant",
"value": 7
},
"sum": {
"type": "mean"
},
"avg": {
"type": "zero"
},
},
"aggregates": [{
"quantity": "dog_sightings",
"metrics": ["count", "avg"]
}],
"intervals": ["1y"],
"groups": ["entity_id"],
},
],
feature_dates=all_as_of_times,
state_table=entity_date_table_generator.entity_date_table_name,
)
feature_table_agg_tasks = feature_generator.generate_all_table_tasks(
aggregations, task_type="aggregation")
# create feature aggregation tables
feature_generator.process_table_tasks(feature_table_agg_tasks)
feature_table_imp_tasks = feature_generator.generate_all_table_tasks(
aggregations, task_type="imputation")
# create feature imputation tables
feature_generator.process_table_tasks(feature_table_imp_tasks)
# build feature dictionaries from feature tables and
# subsetting config
master_feature_dict = feature_dictionary_creator.feature_dictionary(
feature_table_names=feature_table_imp_tasks.keys(),
index_column_lookup=feature_generator.index_column_lookup(
aggregations),
)
feature_dicts = feature_group_mixer.generate(
feature_group_creator.subsets(master_feature_dict))
# figure out what matrices need to be built
_, matrix_build_tasks = planner.generate_plans(
split_definitions, feature_dicts)
# go and build the matrices
builder.build_all_matrices(matrix_build_tasks)
# super basic assertion: did matrices we expect get created?
matrices_records = list(
db_engine.execute(
"""select matrix_uuid, num_observations, matrix_type
from triage_metadata.matrices
"""))
matrix_directory = os.path.join(temp_dir, "matrices")
matrices = [
path for path in os.listdir(matrix_directory) if ".csv" in path
]
metadatas = [
path for path in os.listdir(matrix_directory)
if ".yaml" in path
]
assert len(matrices) == num_split_matrices * \
expected_matrix_multiplier
assert len(metadatas) == num_split_matrices * \
expected_matrix_multiplier
assert len(matrices) == len(matrices_records)
feature_group_name_lists = []
for metadata_path in metadatas:
with open(os.path.join(matrix_directory, metadata_path)) as f:
metadata = yaml.full_load(f)
feature_group_name_lists.append(metadata["feature_groups"])
for matrix_uuid, num_observations, matrix_type in matrices_records:
assert matrix_uuid in matrix_build_tasks # the hashes of the matrices
assert type(num_observations) is int
assert matrix_type == matrix_build_tasks[matrix_uuid][
"matrix_type"]
def deep_unique_tuple(l):
return set([tuple(i) for i in l])
assert deep_unique_tuple(
feature_group_name_lists) == deep_unique_tuple(
expected_group_lists)
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 retrain(self, prediction_date):
"""Retrain a model by going back one split from prediction_date, so the as_of_date for training would be (prediction_date - training_label_timespan)
Args:
prediction_date(str)
"""
# Retrain config and hash
retrain_config = {
"model_group_id": self.model_group_id,
"prediction_date": prediction_date,
"test_label_timespan": self.test_label_timespan,
"test_duration": self.test_duration,
}
self.retrain_hash = save_retrain_and_get_hash(retrain_config,
self.db_engine)
with get_for_update(self.db_engine, Retrain,
self.retrain_hash) as retrain:
retrain.prediction_date = prediction_date
# Timechop
prediction_date = dt_from_str(prediction_date)
temporal_config = self.get_temporal_config_for_retrain(prediction_date)
timechopper = Timechop(**temporal_config)
chops = timechopper.chop_time()
assert len(chops) == 1
chops_train_matrix = chops[0]['train_matrix']
as_of_date = datetime.strftime(chops_train_matrix['last_as_of_time'],
"%Y-%m-%d")
retrain_definition = {
'first_as_of_time':
chops_train_matrix['first_as_of_time'],
'last_as_of_time':
chops_train_matrix['last_as_of_time'],
'matrix_info_end_time':
chops_train_matrix['matrix_info_end_time'],
'as_of_times': [as_of_date],
'training_label_timespan':
chops_train_matrix['training_label_timespan'],
'max_training_history':
chops_train_matrix['max_training_history'],
'training_as_of_date_frequency':
chops_train_matrix['training_as_of_date_frequency'],
}
# Set ExperimentRun
run = TriageRun(
start_time=datetime.now(),
git_hash=infer_git_hash(),
triage_version=infer_triage_version(),
python_version=infer_python_version(),
run_type="retrain",
run_hash=self.retrain_hash,
last_updated_time=datetime.now(),
current_status=TriageRunStatus.started,
installed_libraries=infer_installed_libraries(),
platform=platform.platform(),
os_user=getpass.getuser(),
working_directory=os.getcwd(),
ec2_instance_type=infer_ec2_instance_type(),
log_location=infer_log_location(),
experiment_class_path=classpath(self.__class__),
random_seed=retrieve_experiment_seed_from_run_id(
self.db_engine, self.triage_run_id),
)
run_id = None
with scoped_session(self.db_engine) as session:
session.add(run)
session.commit()
run_id = run.run_id
if not run_id:
raise ValueError("Failed to retrieve run_id from saved row")
# set ModelTrainer's run_id and experiment_hash for Retrain run
self.model_trainer.run_id = run_id
self.model_trainer.experiment_hash = self.retrain_hash
# 1. Generate all labels
self.generate_all_labels(as_of_date)
record_labels_table_name(run_id, self.db_engine,
self.labels_table_name)
# 2. Generate cohort
cohort_table_name = f"triage_production.cohort_{self.experiment_config['cohort_config']['name']}_retrain"
self.generate_entity_date_table(as_of_date, cohort_table_name)
record_cohort_table_name(run_id, self.db_engine, cohort_table_name)
# 3. Generate feature aggregations
collate_aggregations = self.get_collate_aggregations(
as_of_date, cohort_table_name)
feature_aggregation_table_tasks = self.feature_generator.generate_all_table_tasks(
collate_aggregations, task_type='aggregation')
self.feature_generator.process_table_tasks(
feature_aggregation_table_tasks)
# 4. Reconstruct feature disctionary from feature_names and generate imputation
reconstructed_feature_dict, imputation_table_tasks = self.get_feature_dict_and_imputation_task(
collate_aggregations,
self.model_group_info['model_id_last_split'],
)
feature_group_creator = FeatureGroupCreator(
self.experiment_config['feature_group_definition'])
feature_group_mixer = FeatureGroupMixer(["all"])
feature_group_dict = feature_group_mixer.generate(
feature_group_creator.subsets(reconstructed_feature_dict))[0]
self.feature_generator.process_table_tasks(imputation_table_tasks)
# 5. Build new matrix
db_config = {
"features_schema_name": "triage_production",
"labels_schema_name": "public",
"cohort_table_name": cohort_table_name,
"labels_table_name": self.labels_table_name,
}
record_matrix_building_started(run_id, self.db_engine)
matrix_builder = MatrixBuilder(
db_config=db_config,
matrix_storage_engine=self.matrix_storage_engine,
engine=self.db_engine,
experiment_hash=None,
replace=True,
)
new_matrix_metadata = Planner.make_metadata(
matrix_definition=retrain_definition,
feature_dictionary=feature_group_dict,
label_name=self.label_name,
label_type='binary',
cohort_name=self.cohort_name,
matrix_type='train',
feature_start_time=dt_from_str(self.feature_start_time),
user_metadata=self.user_metadata,
)
new_matrix_metadata['matrix_id'] = "_".join([
self.label_name,
'binary',
str(as_of_date),
'retrain',
])
matrix_uuid = filename_friendly_hash(new_matrix_metadata)
matrix_builder.build_matrix(
as_of_times=[as_of_date],
label_name=self.label_name,
label_type='binary',
feature_dictionary=feature_group_dict,
matrix_metadata=new_matrix_metadata,
matrix_uuid=matrix_uuid,
matrix_type="train",
)
retrain_model_comment = 'retrain_' + str(datetime.now())
misc_db_parameters = {
'train_end_time': dt_from_str(as_of_date),
'test': False,
'train_matrix_uuid': matrix_uuid,
'training_label_timespan': self.training_label_timespan,
'model_comment': retrain_model_comment,
}
# get the random seed from the last split
last_split_train_matrix_uuid, last_split_matrix_metadata = train_matrix_info_from_model_id(
self.db_engine,
model_id=self.model_group_info['model_id_last_split'])
random_seed = self.model_trainer.get_or_generate_random_seed(
model_group_id=self.model_group_id,
matrix_metadata=last_split_matrix_metadata,
train_matrix_uuid=last_split_train_matrix_uuid)
# create retrain model hash
retrain_model_hash = self.model_trainer._model_hash(
self.matrix_storage_engine.get_store(matrix_uuid).metadata,
class_path=self.model_group_info['model_type'],
parameters=self.model_group_info['hyperparameters'],
random_seed=random_seed,
)
associate_models_with_retrain(self.retrain_hash,
(retrain_model_hash, ), self.db_engine)
record_model_building_started(run_id, self.db_engine)
retrain_model_id = self.model_trainer.process_train_task(
matrix_store=self.matrix_storage_engine.get_store(matrix_uuid),
class_path=self.model_group_info['model_type'],
parameters=self.model_group_info['hyperparameters'],
model_hash=retrain_model_hash,
misc_db_parameters=misc_db_parameters,
random_seed=random_seed,
retrain=True,
model_group_id=self.model_group_id)
self.retrain_model_hash = retrieve_model_hash_from_id(
self.db_engine, retrain_model_id)
self.retrain_matrix_uuid = matrix_uuid
self.retrain_model_id = retrain_model_id
return {
'retrain_model_comment': retrain_model_comment,
'retrain_model_id': retrain_model_id
}