def load_model_from_vertex(
project: str,
region: str,
endpoint_id: str,
credentials: Optional[google.auth.credentials.Credentials] = None,
input_modalities: Optional[Dict[str, str]] = None) -> model_lib.Model:
"""Loads a model from Unified Cloud AI Platform.
Args:
project: AI Platform project name.
region: GCP Region for the deployed model.
endpoint_id: Version of the given model. If not given, it will load the
default version for the model.
credentials: The OAuth2.0 credentials to use for GCP services.
input_modalities: Dictionary mapping from modalities to input names in the
explain metadata. For example {'numeric': ['input1', 'input2'], 'all':
['input1', 'input2']}. All inputs must be collected under the 'all' key.
If None, a default metadata of {'all': [...]} will be created.
Returns:
A model object
Raises:
NotImplementedError: If there are no registered remote models.
"""
if _VERTEX_MODEL_KEY not in _MODEL_REGISTRY:
available_models = ', '.join(_MODEL_REGISTRY)
raise NotImplementedError(
'There are no implementations for Vertex models. '
f'Avilable models are: {{{available_models}}}.')
resource_path = os.path.join('projects', project, 'locations', region,
'endpoints', endpoint_id)
return _MODEL_REGISTRY[_VERTEX_MODEL_KEY](utils.get_endpoint_uri(
resource_path, region, True), credentials, input_modalities)
def load_model_from_ai_platform(
project: str,
model: str,
version: Optional[str] = None,
credentials: Optional[google.auth.credentials.Credentials] = None,
region: Optional[str] = None,
input_modalities: Optional[Dict[str, str]] = None) -> model_lib.Model:
"""Loads a model from Cloud AI Platform.
Args:
project: AI Platform project name.
model: AI Platform Prediction model name.
version: Version of the given model. If not given, it will load the
default version for the model.
credentials: The OAuth2.0 credentials to use for GCP services.
region: GCP Region for the deployed model.
input_modalities: Dictionary mapping from modalities to input names in the
explain metadata. For example {'numeric': ['input1', 'input2'], 'all':
['input1', 'input2']}. All inputs must be collected under the 'all' key.
If None, modalities will be inferred from the model metadata.
Returns:
A model object
Raises:
NotImplementedError: If there are no registered remote models.
"""
if _CAIP_MODEL_KEY not in _MODEL_REGISTRY:
available_models = ', '.join(_MODEL_REGISTRY)
raise NotImplementedError(
'There are no implementations for CAIP models. '
f'Avilable models are: {{{available_models}}}.')
resouce_path = os.path.join('projects', project, 'models', model)
if version:
resouce_path = os.path.join(resouce_path, 'versions', version)
model_endpoint_uri = utils.get_endpoint_uri(resouce_path, region, False)
if not input_modalities:
input_modalities = utils.create_modality_inputs_map_from_metadata(
utils.fetch_explanation_metadata(model_endpoint_uri, credentials))
return _MODEL_REGISTRY[_CAIP_MODEL_KEY](model_endpoint_uri, credentials,
input_modalities)
def test_get_endpoint_uri_env_variable(self):
self.assertEqual('https://overriden/v1/m/1/e/4',
utils.get_endpoint_uri('m/1/e/4'))
def test_get_endpoint_uri(self, region, is_vertex, expected_uri):
self.assertEqual(expected_uri,
utils.get_endpoint_uri('m/1/e/4', region, is_vertex))
