def main(force: bool, skip: bool, model_metadata: str, register_folder: str,
model_name: str):
if force and skip:
raise ValueError(
"Model registration cannot be both forced and skipped")
# Determine if the model should be registered
if skip:
print("Registration skipped")
register_recommended = False
elif force:
print("Model registration forced")
register_recommended = True
else:
register_recommended = read_recommendation(folder_path=register_folder)
print(f"Model Registration Is Recommended?: {register_recommended}")
# If model registration is recommended, then register the model
if register_recommended:
metadata = read_metadata(model_metadata)
challenger_model_run = Run.get(workspace=workspace,
run_id=metadata['run_id'])
challenger_model_run.register_model(model_name,
model_path=metadata['model_path'])
def load_challenger_model(metadata: dict) -> Any:
train_run = Run.get(workspace, metadata['run_id'])
challenger_path = "./challenger/model.pkl"
train_run.download_file(name=metadata['model_path'],
output_file_path=challenger_path)
return joblib.load(challenger_path)
def main(event: func.EventGridEvent):
try:
result = json.dumps({
'id': event.id,
'data': event.get_json(),
'topic': event.topic,
'subject': event.subject,
'event_type': event.event_type
})
logging.info('PipelineRunMonitor: processed an event: %s', result)
if (event.event_type
== "Microsoft.MachineLearningServices.RunCompleted"
or event.get_json()["runStatus"] == "Failed"):
# Retrieve environment variables
subscriptionId = os.environ["SUBSCRIPTION_ID"]
rg_name = os.environ["RESOURCE_GROUP"]
ws_name = os.environ["WORKSPACE_NAME"]
app_insights_connection_string = os.environ[
"APP_INSIGHTS_CONNECTION_STRING"]
# Managed identity authentication
msi_auth = MsiAuthentication()
# Azure ML workspace
aml_workspace = Workspace(subscription_id=subscriptionId,
resource_group=rg_name,
workspace_name=ws_name,
auth=msi_auth)
logging.info(f"Azure ML workspace: {aml_workspace}")
# Set up logger for Application Insights
logger = logging.getLogger(__name__)
logger.addHandler(
AzureLogHandler(
connection_string=app_insights_connection_string))
aml_run = Run.get(aml_workspace, event.get_json()["runId"])
custom_dimensions = {
"parent_run_id":
aml_run.parent.id if aml_run.parent else aml_run.id,
"parent_run_name":
aml_run.parent.name if aml_run.parent else aml_run.name,
"parent_run_number":
aml_run.parent.number if aml_run.parent else aml_run.number,
"run_number":
aml_run.number,
"step_id":
aml_run.id,
"step_name":
aml_run.name,
"experiment_name":
aml_run.experiment.name,
"run_url":
aml_run.parent.get_portal_url()
if aml_run.parent else aml_run.get_portal_url(),
"parent_run_status":
aml_run.parent.status if aml_run.parent else aml_run.status,
"run_status":
aml_run.status,
"type":
"run_detail",
"workspace_name":
aml_run.experiment.workspace.name
}
details = aml_run.get_details()
logger.info(json.dumps(details, default=lambda o: ''),
extra={'custom_dimensions': custom_dimensions})
elif (event.event_type
== "Microsoft.MachineLearningServices.RunStatusChanged"
and event.get_json()["runStatus"] == "Running"
and event.get_json()["runProperties"]["azureml.runsource"]
== "azureml.PipelineRun"):
# Please write a pipeline run notification here
pass
except Exception as ex:
logging.exception(
f"PipelineRunMonitor: stopped execution due to the following error: {ex}"
)
raise Exception("PipelineRunMonitor error") from ex
def run(input_data):
# 1.0 Set up Logging
entry_script = EntryScript()
logger = entry_script.logger
logger.info('Making forecasts')
os.makedirs('./outputs', exist_ok=True)
all_predictions = pd.DataFrame()
# 2.0 Iterate through input data
for idx, file_path in enumerate(input_data):
date1 = datetime.datetime.now()
file_name, file_extension = os.path.splitext(
os.path.basename(file_path))
logger.info(file_path)
if file_extension.lower() == ".parquet":
data = pd.read_parquet(file_path)
else:
data = pd.read_csv(file_path)
tags_dict = {}
if hasattr(args, "many_models_run_id") and args.many_models_run_id:
tags_dict['RunId'] = args.many_models_run_id
for column_name in args.group_column_names:
tags_dict.update(
{column_name: str(data.iat[0, data.columns.get_loc(column_name)])})
print(tags_dict)
model_string = '_'.join(str(v) for k, v in sorted(
tags_dict.items()) if k in args.group_column_names)
logger.info("model string to encode " + model_string)
sha = hashlib.sha256()
sha.update(model_string.encode())
model_name = 'automl_' + sha.hexdigest()
logger.info('starting (' + file_path + ') ' + str(date1))
ws = current_step_run.experiment.workspace
logger.info('query the model ' + model_name)
model_list = Model.list(ws, name=model_name,
tags=tags_dict, latest=True)
if not model_list:
print("Could not find model")
continue
# 4.0 Un-pickle model and make predictions
model_path = model_list[0].download(exist_ok=True)
model = joblib.load(model_path)
model_name = model_list[0].name
print('Unpickled the model ' + model_name)
# Grab relevant model metrics
run_id = model_list[0].run_id
run = Run.get(ws, run_id)
target_metric = run.get_metrics(name='mean_absolute_error')['mean_absolute_error']
X_test = data.copy()
if args.target_column_name is not None:
X_test.pop(args.target_column_name)
print("prediction data head")
print(X_test.head())
y_predictions, X_trans = model.forecast(
X_test, ignore_data_errors=True)
print('Made predictions ' + model_name)
# Insert predictions/model metrics to test set
predicted_column_name = 'Predictions'
data[predicted_column_name] = y_predictions
data['model_metric'] = np.full(len(y_predictions), target_metric)
print(data.head())
print('Inserted predictions ' + model_name)
cols = list(data.columns.values)
print(cols)
all_predictions = all_predictions.append(data)
# 5.0 Log the run
date2 = datetime.datetime.now()
logger.info('ending (' + str(file_path) + ') ' + str(date2))
print(all_predictions.head())
return all_predictions
