def test_model_load_from_remote_uri_succeeds(sklearn_knn_model,
main_scoped_model_class, tmpdir,
mock_s3_bucket, iris_data):
artifact_root = "s3://{bucket_name}".format(bucket_name=mock_s3_bucket)
artifact_repo = S3ArtifactRepository(artifact_root)
sklearn_model_path = os.path.join(str(tmpdir), "sklearn_model")
mlflow.sklearn.save_model(sk_model=sklearn_knn_model,
path=sklearn_model_path)
sklearn_artifact_path = "sk_model"
artifact_repo.log_artifacts(sklearn_model_path,
artifact_path=sklearn_artifact_path)
def test_predict(sk_model, model_input):
return sk_model.predict(model_input) * 2
pyfunc_model_path = os.path.join(str(tmpdir), "pyfunc_model")
mlflow.pyfunc.save_model(
path=pyfunc_model_path,
artifacts={"sk_model": sklearn_model_path},
python_model=main_scoped_model_class(test_predict))
pyfunc_artifact_path = "pyfunc_model"
artifact_repo.log_artifacts(pyfunc_model_path,
artifact_path=pyfunc_artifact_path)
model_uri = artifact_root + "/" + pyfunc_artifact_path
loaded_pyfunc_model = mlflow.pyfunc.load_pyfunc(model_uri=model_uri)
np.testing.assert_array_equal(
loaded_pyfunc_model.predict(model_input=iris_data[0]),
test_predict(sk_model=sklearn_knn_model, model_input=iris_data[0]))
def test_load_model_from_remote_uri_succeeds(saved_tf_iris_model, model_path,
mock_s3_bucket):
mlflow.tensorflow.save_model(
tf_saved_model_dir=saved_tf_iris_model.path,
tf_meta_graph_tags=saved_tf_iris_model.meta_graph_tags,
tf_signature_def_key=saved_tf_iris_model.signature_def_key,
path=model_path)
artifact_root = "s3://{bucket_name}".format(bucket_name=mock_s3_bucket)
artifact_path = "model"
artifact_repo = S3ArtifactRepository(artifact_root)
artifact_repo.log_artifacts(model_path, artifact_path=artifact_path)
model_uri = artifact_root + "/" + artifact_path
tf_graph = tf.Graph()
tf_sess = tf.Session(graph=tf_graph)
with tf_graph.as_default():
signature_def = mlflow.tensorflow.load_model(model_uri=model_uri,
tf_sess=tf_sess)
for _, input_signature in signature_def.inputs.items():
t_input = tf_graph.get_tensor_by_name(input_signature.name)
assert t_input is not None
for _, output_signature in signature_def.outputs.items():
t_output = tf_graph.get_tensor_by_name(output_signature.name)
assert t_output is not None
def from_artifact_uri(artifact_uri, store):
"""
Given an artifact URI for an Experiment Run (e.g., /local/file/path or s3://my/bucket),
returns an ArtifactReposistory instance capable of logging and downloading artifacts
on behalf of this URI.
:param store: An instance of AbstractStore which the artifacts are registered in.
"""
if artifact_uri.startswith("s3:/"):
# Import these locally to avoid creating a circular import loop
from mlflow.store.s3_artifact_repo import S3ArtifactRepository
return S3ArtifactRepository(artifact_uri)
elif artifact_uri.startswith("gs:/"):
from mlflow.store.gcs_artifact_repo import GCSArtifactRepository
return GCSArtifactRepository(artifact_uri)
elif artifact_uri.startswith("wasbs:/"):
from mlflow.store.azure_blob_artifact_repo import AzureBlobArtifactRepository
return AzureBlobArtifactRepository(artifact_uri)
elif artifact_uri.startswith("sftp:/"):
from mlflow.store.sftp_artifact_repo import SFTPArtifactRepository
return SFTPArtifactRepository(artifact_uri)
elif artifact_uri.startswith("dbfs:/"):
from mlflow.store.dbfs_artifact_repo import DbfsArtifactRepository
if not isinstance(store, DatabricksStore):
raise MlflowException(
'`store` must be an instance of DatabricksStore.')
return DbfsArtifactRepository(artifact_uri,
store.http_request_kwargs)
else:
from mlflow.store.local_artifact_repo import LocalArtifactRepository
return LocalArtifactRepository(artifact_uri)
def test_cli_build_image_with_remote_uri_calls_expected_azure_routines(
sklearn_model, model_path):
mlflow.sklearn.save_model(sk_model=sklearn_model, path=model_path)
artifact_path = "model"
s3_artifact_repo = S3ArtifactRepository("s3://test_bucket")
s3_artifact_repo.log_artifacts(model_path, artifact_path=artifact_path)
model_uri = "s3://test_bucket/{artifact_path}".format(
artifact_path=artifact_path)
with AzureMLMocks() as aml_mocks:
result = CliRunner(env={
"LC_ALL": "en_US.UTF-8",
"LANG": "en_US.UTF-8"
}).invoke(mlflow.azureml.cli.commands, [
'build-image',
'-m',
model_uri,
'-w',
'test_workspace',
'-i',
'image_name',
'-n',
'model_name',
])
assert result.exit_code == 0
assert aml_mocks["register_model"].call_count == 1
assert aml_mocks["create_image"].call_count == 1
assert aml_mocks["load_workspace"].call_count == 1
def test_deploy_cli_creates_sagemaker_and_s3_resources_with_expected_names_from_s3(
pretrained_model, sagemaker_client):
local_model_path = _download_artifact_from_uri(pretrained_model.model_uri)
artifact_path = "model"
region_name = sagemaker_client.meta.region_name
default_bucket = mfs._get_default_s3_bucket(region_name)
s3_artifact_repo = S3ArtifactRepository('s3://{}'.format(default_bucket))
s3_artifact_repo.log_artifacts(local_model_path,
artifact_path=artifact_path)
model_s3_uri = 's3://{bucket_name}/{artifact_path}'.format(
bucket_name=default_bucket, artifact_path=pretrained_model.model_path)
app_name = "test-app"
result = CliRunner(env={
"LC_ALL": "en_US.UTF-8",
"LANG": "en_US.UTF-8"
}).invoke(mfscli.commands, [
'deploy',
'-a',
app_name,
'-m',
model_s3_uri,
'--mode',
mfs.DEPLOYMENT_MODE_CREATE,
])
assert result.exit_code == 0
region_name = sagemaker_client.meta.region_name
s3_client = boto3.client("s3", region_name=region_name)
default_bucket = mfs._get_default_s3_bucket(region_name)
endpoint_description = sagemaker_client.describe_endpoint(
EndpointName=app_name)
endpoint_production_variants = endpoint_description["ProductionVariants"]
assert len(endpoint_production_variants) == 1
model_name = endpoint_production_variants[0]["VariantName"]
assert model_name in [
model["ModelName"]
for model in sagemaker_client.list_models()["Models"]
]
object_names = [
entry["Key"]
for entry in s3_client.list_objects(Bucket=default_bucket)["Contents"]
]
assert any([model_name in object_name for object_name in object_names])
assert any([
app_name in config["EndpointConfigName"] for config in
sagemaker_client.list_endpoint_configs()["EndpointConfigs"]
])
assert app_name in [
endpoint["EndpointName"]
for endpoint in sagemaker_client.list_endpoints()["Endpoints"]
]
def test_model_load_from_remote_uri_succeeds(model, model_path, mock_s3_bucket,
data, predicted):
x, _ = data
mlflow.keras.save_model(model, model_path)
artifact_root = "s3://{bucket_name}".format(bucket_name=mock_s3_bucket)
artifact_path = "model"
artifact_repo = S3ArtifactRepository(artifact_root)
artifact_repo.log_artifacts(model_path, artifact_path=artifact_path)
model_uri = artifact_root + "/" + artifact_path
model_loaded = mlflow.keras.load_model(model_uri=model_uri)
assert all(model_loaded.predict(x) == predicted)
def test_sparkml_model_load_from_remote_uri_succeeds(spark_model_iris, model_path, mock_s3_bucket):
sparkm.save_model(spark_model=spark_model_iris.model, path=model_path)
artifact_root = "s3://{bucket_name}".format(bucket_name=mock_s3_bucket)
artifact_path = "model"
artifact_repo = S3ArtifactRepository(artifact_root)
artifact_repo.log_artifacts(model_path, artifact_path=artifact_path)
model_uri = artifact_root + "/" + artifact_path
reloaded_model = sparkm.load_model(model_uri=model_uri)
preds_df = reloaded_model.transform(spark_model_iris.spark_df)
preds = [x.prediction for x in preds_df.select("prediction").collect()]
assert spark_model_iris.predictions == preds
def test_load_model_from_remote_uri_succeeds(sequential_model, model_path,
mock_s3_bucket, data,
sequential_predicted):
mlflow.pytorch.save_model(sequential_model, model_path)
artifact_root = "s3://{bucket_name}".format(bucket_name=mock_s3_bucket)
artifact_path = "model"
artifact_repo = S3ArtifactRepository(artifact_root)
artifact_repo.log_artifacts(model_path, artifact_path=artifact_path)
model_uri = artifact_root + "/" + artifact_path
sequential_model_loaded = mlflow.pytorch.load_model(model_uri=model_uri)
np.testing.assert_array_equal(_predict(sequential_model_loaded, data),
sequential_predicted)
def test_model_load_from_remote_uri_succeeds(sklearn_knn_model, model_path,
mock_s3_bucket):
mlflow.sklearn.save_model(sk_model=sklearn_knn_model.model,
path=model_path)
artifact_root = "s3://{bucket_name}".format(bucket_name=mock_s3_bucket)
artifact_path = "model"
artifact_repo = S3ArtifactRepository(artifact_root)
artifact_repo.log_artifacts(model_path, artifact_path=artifact_path)
model_uri = artifact_root + "/" + artifact_path
reloaded_knn_model = mlflow.sklearn.load_model(model_uri=model_uri)
np.testing.assert_array_equal(
sklearn_knn_model.model.predict(sklearn_knn_model.inference_data),
reloaded_knn_model.predict(sklearn_knn_model.inference_data))
def test_build_image_with_remote_uri_calls_expected_azure_routines(
sklearn_model, model_path):
mlflow.sklearn.save_model(sk_model=sklearn_model, path=model_path)
artifact_path = "model"
s3_artifact_repo = S3ArtifactRepository("s3://test_bucket")
s3_artifact_repo.log_artifacts(model_path, artifact_path=artifact_path)
model_uri = "s3://test_bucket/{artifact_path}".format(
artifact_path=artifact_path)
with AzureMLMocks() as aml_mocks:
workspace = get_azure_workspace()
mlflow.azureml.build_image(model_uri=model_uri, workspace=workspace)
assert aml_mocks["register_model"].call_count == 1
assert aml_mocks["create_image"].call_count == 1
def from_artifact_uri(artifact_uri):
"""
Given an artifact URI for an Experiment Run (e.g., /local/file/path or s3://my/bucket),
returns an ArtifactReposistory instance capable of logging and downloading artifacts
on behalf of this URI.
"""
if artifact_uri.startswith("s3:/"):
# Import these locally to avoid creating a circular import loop
from mlflow.store.s3_artifact_repo import S3ArtifactRepository
return S3ArtifactRepository(artifact_uri)
elif artifact_uri.startswith("gs:/"):
from mlflow.store.gcs_artifact_repo import GCSArtifactRepository
return GCSArtifactRepository(artifact_uri)
elif artifact_uri.startswith("wasbs:/"):
from mlflow.store.azure_blob_artifact_repo import AzureBlobArtifactRepository
return AzureBlobArtifactRepository(artifact_uri)
else:
from mlflow.store.local_artifact_repo import LocalArtifactRepository
return LocalArtifactRepository(artifact_uri)
def test_load_model_from_remote_uri_succeeds(saved_tf_iris_model, model_path, mock_s3_bucket):
mlflow.tensorflow.save_model(tf_saved_model_dir=saved_tf_iris_model.path,
tf_meta_graph_tags=saved_tf_iris_model.meta_graph_tags,
tf_signature_def_key=saved_tf_iris_model.signature_def_key,
path=model_path)
artifact_root = "s3://{bucket_name}".format(bucket_name=mock_s3_bucket)
artifact_path = "model"
artifact_repo = S3ArtifactRepository(artifact_root)
artifact_repo.log_artifacts(model_path, artifact_path=artifact_path)
model_uri = artifact_root + "/" + artifact_path
infer = mlflow.tensorflow.load_model(model_uri=model_uri)
feed_dict = {
df_column_name: tf.constant(
saved_tf_iris_model.inference_df[df_column_name])
for df_column_name in list(saved_tf_iris_model.inference_df)
}
raw_preds = infer(**feed_dict)
pred_dict = {column_name: raw_preds[column_name].numpy() for column_name in raw_preds.keys()}
for col in pred_dict:
assert(np.allclose(np.array(pred_dict[col], dtype=np.float),
np.array(saved_tf_iris_model.raw_results[col], dtype=np.float)))
