def test_remote_featurizer_create_package_and_service():
test_config = get_test_config()
deployment_client = DeploymentClient(
test_config['test_subscription_id'],
test_config['test_resource_group'],
test_config['test_model_management_account'],
service_principal_params=get_service_principal())
id = uuid.uuid4().hex[:5]
model_name = "int-test-rf-model-" + id
service_name = "int-test-rf-service-" + id
service_def_path = "/tmp/modelrf"
in_images = tf.placeholder(tf.string)
image_tensors = preprocess_array(in_images)
remote_service_name = ("int-test-featurizer-svc-" + str(uuid.uuid4()))[:30]
model = RemoteQuantizedResNet50(
test_config['test_subscription_id'],
test_config['test_resource_group'],
test_config['test_model_management_account'],
os.path.expanduser("~/models"),
remote_service_name,
service_principal_params=get_service_principal())
model.import_graph_def(include_featurizer=True, input_tensor=image_tensors)
service_def = ServiceDefinition()
service_def.pipeline.append(
TensorflowStage(tf.Session(), in_images, image_tensors))
service_def.pipeline.append(BrainWaveStage(model))
service_def.pipeline.append(
TensorflowStage(tf.Session(), model.classifier_input,
model.classifier_output))
service_def.save(service_def_path)
# create service
model_id = deployment_client.register_model(model_name, service_def_path)
service = deployment_client.create_service(service_name, model_id)
prediction_client = PredictionClient(service.ipAddress, service.port)
top_result = sorted(enumerate(
prediction_client.score_image("/tmp/share1/shark.jpg")),
key=lambda x: x[1],
reverse=True)[:1]
# 'tiger shark' is class 3
assert top_result[0][0] == 3
deployment_client.delete_service(service.id)
deployment_client.delete_model(model_id)
def test_remote_featurizer_local_usage():
test_config = get_test_config()
in_images = tf.placeholder(tf.string)
image_tensors = preprocess_array(in_images)
remote_service_name = ("int-test-featurizer-svc-" + str(uuid.uuid4()))[:30]
featurizer = RemoteQuantizedResNet50(
test_config['test_subscription_id'],
test_config['test_resource_group'],
test_config['test_model_management_account'],
os.path.expanduser("~/models"),
remote_service_name,
service_principal_params=get_service_principal())
featurizer.import_graph_def(include_top=True,
include_featurizer=True,
input_tensor=image_tensors)
try:
with tf.Session() as sess:
result = sess.run([featurizer.featurizer_output],
feed_dict={in_images: [read_file()]})
np_result = np.array(result[0])
assert all(
[x == y for x, y in zip(np_result.shape, [1, 1, 1, 2048])])
assert np_result.dtype == np.dtype('float32')
result = sess.run(
[featurizer.featurizer_output],
feed_dict={in_images: [read_file(), read_file()]})
np_result = np.array(result[0])
assert all(
[x == y for x, y in zip(np_result.shape, [2, 1, 1, 2048])])
assert np_result.dtype == np.dtype('float32')
result = sess.run([featurizer.classifier_output],
feed_dict={in_images: [read_file()]})
np_result = np.array(result[0])
assert all([x == y for x, y in zip(np_result.shape, [1, 1000])])
assert np_result.dtype == np.dtype('float32')
result = sess.run(
[featurizer.classifier_output],
feed_dict={in_images: [read_file(), read_file()]})
np_result = np.array(result[0])
assert all([x == y for x, y in zip(np_result.shape, [2, 1000])])
assert np_result.dtype == np.dtype('float32')
finally:
featurizer.cleanup_remote_service()
def test_create_update_and_delete_service():
override_token_funcs()
test_config = get_test_config()
deployment_client = DeploymentClient(
test_config['test_subscription_id'],
test_config['test_resource_group'],
test_config['test_model_management_account'])
cleanup_old_test_services(deployment_client)
id = uuid.uuid4().hex[:5]
model_name = "int-test-model-" + id
service_name = "int-test-service-" + id
service_def_path = "/tmp/model"
in_images = tf.placeholder(tf.string)
image_tensors = preprocess_array(in_images)
model = LocalQuantizedResNet50(os.path.expanduser("~/models"))
model.import_graph_def(include_featurizer=False)
service_def = ServiceDefinition()
service_def.pipeline.append(
TensorflowStage(tf.Session(), in_images, image_tensors))
service_def.pipeline.append(BrainWaveStage(model))
service_def.pipeline.append(
TensorflowStage(tf.Session(), model.classifier_input,
model.classifier_output))
service_def.save(service_def_path)
# create service
first_model_id = deployment_client.register_model(model_name,
service_def_path)
service = deployment_client.create_service(service_name, first_model_id)
service_list = deployment_client.list_services()
assert any(x.name == service_name for x in service_list)
prediction_client = PredictionClient(service.ipAddress, service.port)
top_result = sorted(enumerate(
prediction_client.score_image("/tmp/share1/shark.jpg")),
key=lambda x: x[1],
reverse=True)[:1]
# 'tiger shark' is class 3
assert top_result[0][0] == 3
# update service, remove classifier
service_def = ServiceDefinition()
service_def.pipeline.append(
TensorflowStage(tf.Session(), in_images, image_tensors))
service_def.pipeline.append(BrainWaveStage(model))
service_def.save(service_def_path)
second_model_id = deployment_client.register_model(model_name,
service_def_path)
deployment_client.update_service(service.id, second_model_id)
result = prediction_client.score_image("/tmp/share1/shark.jpg")
assert all([x == y for x, y in zip(np.array(result).shape, [1, 1, 2048])])
# wait for timeout of Azure LB
time.sleep(4 * 60 + 10)
result = prediction_client.score_image("/tmp/share1/shark.jpg")
assert all([x == y for x, y in zip(np.array(result).shape, [1, 1, 2048])])
deployment_client.delete_service(service.id)
deployment_client.delete_model(first_model_id)
deployment_client.delete_model(second_model_id)