def ParseAcceleratorFlag(accelerator, version):
"""Validates and returns an accelerator config message object."""
if accelerator is None:
return None
types = list(c for c in GetAcceleratorTypeMapper(version).choices)
raw_type = accelerator.get('type', None)
if raw_type not in types:
raise errors.ArgumentError("""\
The type of the accelerator can only be one of the following: {}.
""".format(', '.join(["'{}'".format(c) for c in types])))
accelerator_count = accelerator.get('count', 1)
if accelerator_count <= 0:
raise errors.ArgumentError("""\
The count of the accelerator must be greater than 0.
""")
if version == constants.ALPHA_VERSION:
accelerator_msg = (
apis.GetMessagesModule(constants.AI_PLATFORM_API_NAME,
constants.AI_PLATFORM_API_VERSION[version])
.GoogleCloudAiplatformV1alpha1MachineSpec)
elif version == constants.BETA_VERSION:
accelerator_msg = (
apis.GetMessagesModule(constants.AI_PLATFORM_API_NAME,
constants.AI_PLATFORM_API_VERSION[version])
.GoogleCloudAiplatformV1beta1MachineSpec)
else:
accelerator_msg = (
apis.GetMessagesModule(constants.AI_PLATFORM_API_NAME,
constants.AI_PLATFORM_API_VERSION[version])
.GoogleCloudAiplatformV1MachineSpec)
accelerator_type = arg_utils.ChoiceToEnum(
raw_type, accelerator_msg.AcceleratorTypeValueValuesEnum)
return accelerator_msg(
acceleratorCount=accelerator_count, acceleratorType=accelerator_type)
def _GetModelDeploymentResourceType(model_ref, client):
"""Gets the deployment resource type of a model.
Each model resource belongs to exactly one supported deployment resource type.
The value is the first item in the list, the length of which must be one.
Args:
model_ref: a model resource object.
client: an apis.GetClientInstance object.
Returns:
A string which value must be 'DEDICATED_RESOURCES' or 'AUTOMATIC_RESOURCES'.
Raises:
ArgumentError: if the model resource object is not found.
"""
try:
model_msg = model_client.ModelsClient(client=client).Get(model_ref)
except apitools_exceptions.HttpError:
raise errors.ArgumentError(
('There is an error while getting the model information. '
'Please make sure the model %r exists.' %
model_ref.RelativeName()))
model_resource = encoding.MessageToPyValue(model_msg)
return model_resource['supportedDeploymentResourcesTypes'][0]
def _ConstructObjectiveConfigForCreate(self, location_ref, endpoint_name,
feature_thresholds, dataset,
bigquery_uri, data_format, gcs_uris,
target_field,
training_sampling_rate):
"""Construct monitoring objective config.
Apply the feature thresholds for skew or drift detection to all the deployed
models under the endpoint.
Args:
location_ref: Location reference.
endpoint_name: Endpoint resource name.
feature_thresholds: Dict or None, key: feature_name, value: thresholds.
dataset: Vertex AI Dataset Id.
bigquery_uri: The BigQuery table of the unmanaged Dataset used to train
this Model.
data_format: Google Cloud Storage format, supported format: csv,
tf-record.
gcs_uris: The Google Cloud Storage uri of the unmanaged Dataset used to
train this Model.
target_field: The target field name the model is to predict.
training_sampling_rate: Training Dataset sampling rate.
Returns:
A list of model monitoring objective config.
"""
objective_config_template = self.messages.GoogleCloudAiplatformV1beta1ModelDeploymentMonitoringObjectiveConfig(
)
if feature_thresholds:
if dataset or bigquery_uri or gcs_uris or data_format:
training_dataset = self.messages.GoogleCloudAiplatformV1beta1ModelMonitoringObjectiveConfigTrainingDataset(
)
if target_field is None:
raise errors.ArgumentError(
"Target field must be provided if you'd like to do training-prediction skew detection."
)
training_dataset.targetField = target_field
training_dataset.loggingSamplingStrategy = self.messages.GoogleCloudAiplatformV1beta1SamplingStrategy(
randomSampleConfig=self.messages.
GoogleCloudAiplatformV1beta1SamplingStrategyRandomSampleConfig(
sampleRate=training_sampling_rate))
if dataset:
training_dataset.dataset = _ParseDataset(
dataset, location_ref).RelativeName()
elif bigquery_uri:
training_dataset.bigquerySource = self.messages.GoogleCloudAiplatformV1beta1BigQuerySource(
inputUri=bigquery_uri)
elif gcs_uris or data_format:
if gcs_uris is None:
raise errors.ArgumentError(
'Data format is defined but no Google Cloud Storage uris are provided. Please use --gcs-uris to provide training datasets.'
)
if data_format is None:
raise errors.ArgumentError(
'No Data format is defined for Google Cloud Storage training dataset. Please use --data-format to define the Data format.'
)
training_dataset.dataFormat = data_format
training_dataset.gcsSource = self.messages.GoogleCloudAiplatformV1beta1GcsSource(
uris=gcs_uris)
training_prediction_skew_detection = self._ConstructSkewThresholds(
feature_thresholds)
objective_config_template.objectiveConfig = self.messages.GoogleCloudAiplatformV1beta1ModelMonitoringObjectiveConfig(
trainingDataset=training_dataset,
trainingPredictionSkewDetectionConfig=
training_prediction_skew_detection)
else:
prediction_drift_detection = self._ConstructDriftThresholds(
feature_thresholds)
objective_config_template.objectiveConfig = self.messages.GoogleCloudAiplatformV1beta1ModelMonitoringObjectiveConfig(
predictionDriftDetectionConfig=prediction_drift_detection)
get_endpoint_req = self.messages.AiplatformProjectsLocationsEndpointsGetRequest(
name=endpoint_name)
endpoint = self.client.projects_locations_endpoints.Get(
get_endpoint_req)
objective_configs = []
for deployed_model in endpoint.deployedModels:
objective_config = copy.deepcopy(objective_config_template)
objective_config.deployedModelId = deployed_model.id
objective_configs.append(objective_config)
return objective_configs
