class PredictionServer(ConnectableComponent, PredictMixin):
def __init__(self, engine):
super(PredictionServer, self).__init__(engine)
self._show_perf = False
self._stats_collector = None
self._memory_monitor = None
self._run_language = None
self._predictor = None
self._target_type = None
def configure(self, params):
super(PredictionServer, self).configure(params)
self._show_perf = self._params.get("show_perf")
self._run_language = RunLanguage(params.get("run_language"))
self._target_type = params[TARGET_TYPE_ARG_KEYWORD]
self._stats_collector = StatsCollector(
disable_instance=not self._show_perf)
self._stats_collector.register_report("run_predictor_total", "finish",
StatsOperation.SUB, "start")
self._memory_monitor = MemoryMonitor(monitor_current_process=True)
if self._run_language == RunLanguage.PYTHON:
from datarobot_drum.drum.language_predictors.python_predictor.python_predictor import (
PythonPredictor, )
self._predictor = PythonPredictor()
elif self._run_language == RunLanguage.JAVA:
from datarobot_drum.drum.language_predictors.java_predictor.java_predictor import (
JavaPredictor, )
self._predictor = JavaPredictor()
elif self._run_language == RunLanguage.R:
# this import is here, because RPredictor imports rpy library,
# which is not installed for Java and Python cases.
from datarobot_drum.drum.language_predictors.r_predictor.r_predictor import RPredictor
self._predictor = RPredictor()
else:
raise DrumCommonException(
"Prediction server doesn't support language: {} ".format(
self._run_language))
self._predictor.configure(params)
def _materialize(self, parent_data_objs, user_data):
model_api = base_api_blueprint()
@model_api.route("/capabilities/", methods=["GET"])
def capabilities():
return make_predictor_capabilities(
self._predictor.supported_payload_formats)
@model_api.route("/health/", methods=["GET"])
def health():
return {"message": "OK"}, HTTP_200_OK
@model_api.route("/predict/", methods=["POST"])
def predict():
logger.debug("Entering predict() endpoint")
self._stats_collector.enable()
self._stats_collector.mark("start")
try:
response, response_status = self.do_predict(logger=logger)
finally:
self._stats_collector.mark("finish")
self._stats_collector.disable()
return response, response_status
@model_api.route("/predictUnstructured/", methods=["POST"])
def predict_unstructured():
logger.debug("Entering predict() endpoint")
self._stats_collector.enable()
self._stats_collector.mark("start")
try:
response, response_status = self.do_predict_unstructured(
logger=logger)
finally:
self._stats_collector.mark("finish")
self._stats_collector.disable()
return response, response_status
@model_api.route("/stats/", methods=["GET"])
def stats():
mem_info = self._memory_monitor.collect_memory_info()
ret_dict = {"mem_info": mem_info._asdict()}
self._stats_collector.round()
ret_dict["time_info"] = {}
for name in self._stats_collector.get_report_names():
d = self._stats_collector.dict_report(name)
ret_dict["time_info"][name] = d
self._stats_collector.stats_reset()
return ret_dict, HTTP_200_OK
@model_api.errorhandler(Exception)
def handle_exception(e):
logger.exception(e)
return {
"message": "ERROR: {}".format(e)
}, HTTP_500_INTERNAL_SERVER_ERROR
app = get_flask_app(model_api)
host = self._params.get("host", None)
port = self._params.get("port", None)
try:
app.run(host, port, threaded=False)
except OSError as e:
raise DrumCommonException("{}: host: {}; port: {}".format(
e, host, port))
if self._stats_collector:
self._stats_collector.print_reports()
return []
def _run_fit_and_predictions_pipelines_in_mlpiper(self):
if self.run_mode == RunMode.SERVER:
run_language = self._check_artifacts_and_get_run_language()
# in prediction server mode infra pipeline == prediction server runner pipeline
infra_pipeline_str = self._prepare_prediction_server_or_batch_pipeline(
run_language)
elif self.run_mode == RunMode.SCORE:
run_language = self._check_artifacts_and_get_run_language()
tmp_output_filename = None
# if output is not provided, output into tmp file and print
if not self.options.output:
# keep object reference so it will be destroyed only in the end of the process
__tmp_output_file = tempfile.NamedTemporaryFile(mode="w")
self.options.output = tmp_output_filename = __tmp_output_file.name
# in batch prediction mode infra pipeline == predictor pipeline
infra_pipeline_str = self._prepare_prediction_server_or_batch_pipeline(
run_language)
elif self.run_mode == RunMode.FIT:
run_language = self._get_fit_run_language()
infra_pipeline_str = self._prepare_fit_pipeline(run_language)
else:
error_message = "{} mode is not supported here".format(
self.run_mode)
print(error_message)
raise DrumCommonException(error_message)
config = ExecutorConfig(
pipeline=infra_pipeline_str,
pipeline_file=None,
run_locally=True,
comp_root_path=CMRunnerUtils.get_components_repo(),
mlpiper_jar=None,
spark_jars=None,
)
_pipeline_executor = Executor(config).standalone(True).set_verbose(
self.options.verbose)
# assign logger with the name drum.mlpiper.Executor to mlpiper Executor
_pipeline_executor.set_logger(
logging.getLogger(LOGGER_NAME_PREFIX + "." +
_pipeline_executor.logger_name()))
self.logger.info(">>> Start {} in the {} mode".format(
ArgumentsOptions.MAIN_COMMAND, self.run_mode.value))
sc = StatsCollector(disable_instance=(
not hasattr(self.options, "show_perf")
or not self.options.show_perf or self.run_mode == RunMode.SERVER))
sc.register_report("Full time", "end", StatsOperation.SUB, "start")
sc.register_report("Init time (incl model loading)", "init",
StatsOperation.SUB, "start")
sc.register_report("Run time (incl reading CSV)", "run",
StatsOperation.SUB, "init")
with verbose_stdout(self.options.verbose):
sc.enable()
try:
sc.mark("start")
_pipeline_executor.init_pipeline()
self.runtime.initialization_succeeded = True
sc.mark("init")
_pipeline_executor.run_pipeline(cleanup=False)
sc.mark("run")
finally:
_pipeline_executor.cleanup_pipeline()
sc.mark("end")
sc.disable()
self.logger.info("<<< Finish {} in the {} mode".format(
ArgumentsOptions.MAIN_COMMAND, self.run_mode.value))
sc.print_reports()
if self.run_mode == RunMode.SCORE:
# print result if output is not provided
if tmp_output_filename:
print(pd.read_csv(tmp_output_filename))
class PredictionServer(ConnectableComponent):
def __init__(self, engine):
super(PredictionServer, self).__init__(engine)
self._show_perf = False
self._stats_collector = None
self._memory_monitor = None
self._run_language = None
self._predictor = None
def configure(self, params):
super(PredictionServer, self).configure(params)
self._threaded = self._params.get("threaded", False)
self._show_perf = self._params.get("show_perf")
self._stats_collector = StatsCollector(disable_instance=not self._show_perf)
self._stats_collector.register_report(
"run_predictor_total", "finish", StatsOperation.SUB, "start"
)
self._memory_monitor = MemoryMonitor()
self._run_language = RunLanguage(params.get("run_language"))
if self._run_language == RunLanguage.PYTHON:
from datarobot_drum.drum.language_predictors.python_predictor.python_predictor import (
PythonPredictor,
)
self._predictor = PythonPredictor()
elif self._run_language == RunLanguage.JAVA:
from datarobot_drum.drum.language_predictors.java_predictor.java_predictor import (
JavaPredictor,
)
self._predictor = JavaPredictor()
elif self._run_language == RunLanguage.R:
# this import is here, because RPredictor imports rpy library,
# which is not installed for Java and Python cases.
from datarobot_drum.drum.language_predictors.r_predictor.r_predictor import RPredictor
self._predictor = RPredictor()
else:
raise DrumCommonException(
"Prediction server doesn't support language: {} ".format(self._run_language)
)
self._predictor.configure(params)
def _materialize(self, parent_data_objs, user_data):
model_api = base_api_blueprint()
@model_api.route("/health/", methods=["GET"])
def health():
return {"message": "OK"}, HTTP_200_OK
@model_api.route("/predict/", methods=["POST"])
def predict():
response_status = HTTP_200_OK
file_key = "X"
logger.debug("Entering predict() endpoint")
REGRESSION_PRED_COLUMN = "Predictions"
filename = request.files[file_key] if file_key in request.files else None
logger.debug("Filename provided under X key: {}".format(filename))
if not filename:
wrong_key_error_message = "Samples should be provided as a csv file under `{}` key.".format(
file_key
)
logger.error(wrong_key_error_message)
response_status = HTTP_422_UNPROCESSABLE_ENTITY
return {"message": "ERROR: " + wrong_key_error_message}, response_status
in_df = pd.read_csv(filename)
# TODO labels have to be provided as command line arguments or within configure endpoint
self._stats_collector.enable()
self._stats_collector.mark("start")
out_df = self._predictor.predict(in_df)
num_columns = len(out_df.columns)
# float32 is not JSON serializable, so cast to float, which is float64
out_df = out_df.astype("float")
if num_columns == 1:
# df.to_json() is much faster.
# But as it returns string, we have to assemble final json using strings.
df_json = out_df[REGRESSION_PRED_COLUMN].to_json(orient="records")
response_json = '{{"predictions":{df_json}}}'.format(df_json=df_json)
elif num_columns == 2:
# df.to_json() is much faster.
# But as it returns string, we have to assemble final json using strings.
df_json_str = out_df.to_json(orient="records")
response_json = '{{"predictions":{df_json}}}'.format(df_json=df_json_str)
else:
ret_str = (
"Predictions dataframe has {} columns; "
"Expected: 1 - for regression, 2 - for binary classification.".format(
num_columns
)
)
response_json = {"message": "ERROR: " + ret_str}
response_status = HTTP_422_UNPROCESSABLE_ENTITY
self._stats_collector.mark("finish")
self._stats_collector.disable()
return response_json, response_status
@model_api.route("/stats/", methods=["GET"])
def stats():
mem_info = self._memory_monitor.collect_memory_info()
ret_dict = {"mem_info": mem_info._asdict()}
self._stats_collector.round()
ret_dict["time_info"] = {}
for name in self._stats_collector.get_report_names():
d = self._stats_collector.dict_report(name)
ret_dict["time_info"][name] = d
self._stats_collector.stats_reset()
return ret_dict, HTTP_200_OK
@model_api.errorhandler(Exception)
def handle_exception(e):
logger.exception(e)
return {"message": "ERROR: {}".format(e)}, HTTP_500_INTERNAL_SERVER_ERROR
app = get_flask_app(model_api)
logging.getLogger("werkzeug").setLevel(logger.getEffectiveLevel())
host = self._params.get("host", None)
port = self._params.get("port", None)
try:
app.run(host, port, threaded=self._threaded)
except OSError as e:
raise DrumCommonException("{}: host: {}; port: {}".format(e, host, port))
if self._stats_collector:
self._stats_collector.print_reports()
return []
class PredictionServer(ConnectableComponent, PredictMixin):
def __init__(self, engine):
super(PredictionServer, self).__init__(engine)
self._show_perf = False
self._stats_collector = None
self._memory_monitor = None
self._run_language = None
self._predictor = None
self._target_type = None
self._code_dir = None
self._deployment_config = None
def configure(self, params):
super(PredictionServer, self).configure(params)
self._code_dir = self._params.get("__custom_model_path__")
self._show_perf = self._params.get("show_perf")
self._run_language = RunLanguage(params.get("run_language"))
self._target_type = TargetType(params[TARGET_TYPE_ARG_KEYWORD])
self._stats_collector = StatsCollector(disable_instance=not self._show_perf)
self._stats_collector.register_report(
"run_predictor_total", "finish", StatsOperation.SUB, "start"
)
self._memory_monitor = MemoryMonitor(monitor_current_process=True)
self._deployment_config = parse_validate_deployment_config_file(
self._params["deployment_config"]
)
if self._run_language == RunLanguage.PYTHON:
from datarobot_drum.drum.language_predictors.python_predictor.python_predictor import (
PythonPredictor,
)
self._predictor = PythonPredictor()
elif self._run_language == RunLanguage.JAVA:
from datarobot_drum.drum.language_predictors.java_predictor.java_predictor import (
JavaPredictor,
)
self._predictor = JavaPredictor()
elif self._run_language == RunLanguage.JULIA:
from datarobot_drum.drum.language_predictors.julia_predictor.julia_predictor import (
JlPredictor,
)
self._predictor = JlPredictor()
elif self._run_language == RunLanguage.R:
# this import is here, because RPredictor imports rpy library,
# which is not installed for Java and Python cases.
from datarobot_drum.drum.language_predictors.r_predictor.r_predictor import RPredictor
self._predictor = RPredictor()
else:
raise DrumCommonException(
"Prediction server doesn't support language: {} ".format(self._run_language)
)
self._predictor.configure(params)
def _materialize(self, parent_data_objs, user_data):
model_api = base_api_blueprint()
@model_api.route("/capabilities/", methods=["GET"])
def capabilities():
return make_predictor_capabilities(self._predictor.supported_payload_formats)
@model_api.route("/info/", methods=["GET"])
def info():
model_info = self._predictor.model_info()
model_info.update({ModelInfoKeys.LANGUAGE: self._run_language.value})
model_info.update({ModelInfoKeys.DRUM_VERSION: drum_version})
model_info.update({ModelInfoKeys.DRUM_SERVER: "flask"})
model_info.update(
{ModelInfoKeys.MODEL_METADATA: read_model_metadata_yaml(self._code_dir)}
)
return model_info, HTTP_200_OK
@model_api.route("/health/", methods=["GET"])
def health():
return {"message": "OK"}, HTTP_200_OK
@model_api.route("/predictions/", methods=["POST"])
@model_api.route("/predict/", methods=["POST"])
def predict():
logger.debug("Entering predict() endpoint")
self._stats_collector.enable()
self._stats_collector.mark("start")
try:
response, response_status = self.do_predict_structured(logger=logger)
finally:
self._stats_collector.mark("finish")
self._stats_collector.disable()
return response, response_status
@model_api.route("/transform/", methods=["POST"])
def transform():
logger.debug("Entering transform() endpoint")
self._stats_collector.enable()
self._stats_collector.mark("start")
try:
response, response_status = self.do_transform(logger=logger)
finally:
self._stats_collector.mark("finish")
self._stats_collector.disable()
return response, response_status
@model_api.route("/predictionsUnstructured/", methods=["POST"])
@model_api.route("/predictUnstructured/", methods=["POST"])
def predict_unstructured():
logger.debug("Entering predict() endpoint")
self._stats_collector.enable()
self._stats_collector.mark("start")
try:
response, response_status = self.do_predict_unstructured(logger=logger)
finally:
self._stats_collector.mark("finish")
self._stats_collector.disable()
return response, response_status
@model_api.route("/stats/", methods=["GET"])
def stats():
mem_info = self._memory_monitor.collect_memory_info()
ret_dict = {"mem_info": mem_info._asdict()}
self._stats_collector.round()
ret_dict["time_info"] = {}
for name in self._stats_collector.get_report_names():
d = self._stats_collector.dict_report(name)
ret_dict["time_info"][name] = d
self._stats_collector.stats_reset()
return ret_dict, HTTP_200_OK
@model_api.errorhandler(Exception)
def handle_exception(e):
logger.exception(e)
return {"message": "ERROR: {}".format(e)}, HTTP_500_INTERNAL_SERVER_ERROR
# Disables warning for development server
cli = sys.modules["flask.cli"]
cli.show_server_banner = lambda *x: None
app = get_flask_app(model_api)
host = self._params.get("host", None)
port = self._params.get("port", None)
try:
app.run(host, port, threaded=False)
except OSError as e:
raise DrumCommonException("{}: host: {}; port: {}".format(e, host, port))
if self._stats_collector:
self._stats_collector.print_reports()
return []