def test_serving_calls_model_fn_once(docker_image, opt_ml, use_gpu):
with local_mode.serve(customer_script=call_model_fn_once_script, model_dir=None,
image_name=docker_image, use_gpu=use_gpu,
opt_ml=opt_ml, additional_env_vars=['SAGEMAKER_MODEL_SERVER_WORKERS=2']):
# call enough times to ensure multiple requests to a worker
for i in range(3):
# will return 500 error if model_fn called during request handling
assert b'output' == requests.post(local_mode.REQUEST_URL, data=b'input').content
def test_serving_calls_model_fn_once(docker_image, opt_ml):
script_path = os.path.join(resources_path, 'call_model_fn_once.py')
with local_mode.serve(script_path, model_dir=None, image_name=docker_image, opt_ml=opt_ml,
additional_env_vars=['SAGEMAKER_MODEL_SERVER_WORKERS=2']):
# call enough times to ensure multiple requests to a worker
for i in range(3):
# will return 500 error if model_fn called during request handling
assert b'output' == requests.post(local_mode.REQUEST_URL, data=b'input').content
def test_chainer_mnist_single_machine(docker_image, opt_ml, use_gpu):
customer_script = 'single_machine_customer_script.py'
hyperparameters = {'batch-size': 10000, 'epochs': 1}
local_mode.train(customer_script,
data_dir,
docker_image,
opt_ml,
hyperparameters=hyperparameters,
source_dir=mnist_path,
use_gpu=use_gpu)
files = [
'model/model.npz', 'output/success', 'output/data/algo-1/accuracy.png',
'output/data/algo-1/cg.dot', 'output/data/algo-1/log',
'output/data/algo-1/loss.png'
]
test_utils.files_exist(opt_ml, files)
assert not local_mode.file_exists(opt_ml,
'output/failure'), 'Failure happened'
script_path = os.path.join(mnist_path, customer_script)
with local_mode.serve(script_path,
model_dir=None,
image_name=docker_image,
opt_ml=opt_ml,
use_gpu=use_gpu,
source_dir=mnist_path):
test_arrays = [
np.zeros((100, 784), dtype='float32'),
np.zeros((100, 1, 28, 28), dtype='float32'),
np.zeros((100, 28, 28), dtype='float32')
]
request_data = np.zeros((100, 784), dtype='float32')
data_as_list = request_data.tolist()
test_utils.predict_and_assert_response_length(data_as_list, 'text/csv')
for array in test_arrays:
# JSON and NPY can take multidimensional (n > 2) arrays
data_as_list = array.tolist()
test_utils.predict_and_assert_response_length(
data_as_list, 'application/json')
test_utils.predict_and_assert_response_length(
request_data, 'application/x-npy')
def test_xgboost_abalone_inference(docker_image, opt_ml):
customer_script = "abalone_distributed.py"
request_body = get_libsvm_request_body()
with local_mode.serve(customer_script,
libsvm_model_dir,
docker_image,
opt_ml,
source_dir=abalone_path):
response_status_code, response_body = local_mode.request(
request_body, content_type="text/libsvm")
assert response_status_code == 200
assert not local_mode.file_exists(opt_ml,
"output/failure"), "Failure happened"
assert len(response_body.split(",")) == 1
def test_chainer_mnist_distributed(docker_image, opt_ml, use_gpu,
customer_script):
cluster_size = 2
# pure_nccl communicator hangs when only one gpu is available.
hyperparameters = {
'sagemaker_process_slots_per_host': 1,
'sagemaker_num_processes': cluster_size,
'batch-size': 10000,
'epochs': 1,
'communicator': 'hierarchical'
}
local_mode.train(customer_script,
data_dir,
docker_image,
opt_ml,
hyperparameters=hyperparameters,
cluster_size=cluster_size,
source_dir=mnist_path,
use_gpu=use_gpu)
files = [
'model/model.npz', 'output/success', 'output/data/algo-1/accuracy.png',
'output/data/algo-1/cg.dot', 'output/data/algo-1/log',
'output/data/algo-1/loss.png'
]
test_utils.files_exist(opt_ml, files)
assert not local_mode.file_exists(opt_ml,
'output/failure'), 'Failure happened'
with local_mode.serve(os.path.join(mnist_path, customer_script),
model_dir=None,
image_name=docker_image,
opt_ml=opt_ml):
request_data = np.zeros((100, 784), dtype='float32')
data_as_list = request_data.tolist()
test_utils.predict_and_assert_response_length(data_as_list,
'application/json')
test_utils.predict_and_assert_response_length(data_as_list, 'text/csv')
test_utils.predict_and_assert_response_length(request_data,
'application/x-npy')
def test_xgboost_abalone_mme_with_transform_fn(docker_image, opt_ml):
customer_script = "abalone_distributed.py"
request_body = get_libsvm_request_body()
additional_env_vars = [
"SAGEMAKER_BIND_TO_PORT=8080",
"SAGEMAKER_SAFE_PORT_RANGE=9000-9999",
"SAGEMAKER_MULTI_MODEL=true",
]
model_name = "libsvm_pickled"
model_data = json.dumps({
"model_name": model_name,
"url": "/opt/ml/model/{}".format(model_name)
})
with append_transform_fn_to_abalone_script(
abalone_path, customer_script) as custom_script_path:
with local_mode.serve(
customer_script,
models_dir,
docker_image,
opt_ml,
source_dir=custom_script_path,
additional_env_vars=additional_env_vars,
):
load_status_code, _ = local_mode.request(
model_data,
content_type="application/json",
request_url=MME_MODELS_URL.format(model_name),
)
assert load_status_code == 200
invoke_status_code, invoke_response_body = local_mode.request(
request_body,
content_type="text/libsvm",
request_url=MME_INVOKE_URL.format(model_name),
)
assert invoke_status_code == 200
assert (len(invoke_response_body.split(",")) == len(request_body.split()) +
1 # final column is the bias term
)
assert not local_mode.file_exists(opt_ml,
"output/failure"), "Failure happened"
def test_xgboost_abalone_custom_inference_with_transform_fn(
docker_image, opt_ml):
customer_script = "abalone_distributed.py"
request_body = get_libsvm_request_body()
with append_transform_fn_to_abalone_script(
abalone_path, customer_script) as custom_script_path:
with local_mode.serve(
customer_script,
libsvm_model_dir,
docker_image,
opt_ml,
source_dir=custom_script_path,
):
response_status_code, response_body = local_mode.request(
request_body, content_type="text/libsvm")
assert response_status_code == 200
assert not local_mode.file_exists(opt_ml,
"output/failure"), "Failure happened"
assert (len(response_body.split(",")) == len(request_body.split()) +
1 # final column is the bias term
)
def test_chainer_mnist_custom_loop(docker_image, opt_ml, use_gpu):
customer_script = 'single_machine_custom_loop.py'
hyperparameters = {'batch-size': 10000, 'epochs': 1}
local_mode.train(customer_script,
data_dir,
docker_image,
opt_ml,
hyperparameters=hyperparameters,
source_dir=mnist_path,
use_gpu=use_gpu)
files = ['model/model.npz', 'output/success']
test_utils.files_exist(opt_ml, files)
assert not local_mode.file_exists(opt_ml,
'output/failure'), 'Failure happened'
script_path = os.path.join(mnist_path, customer_script)
with local_mode.serve(script_path,
model_dir=None,
image_name=docker_image,
opt_ml=opt_ml):
request_data = np.zeros((100, 784), dtype='float32')
data_as_list = request_data.tolist()
test_utils.predict_and_assert_response_length(data_as_list,
'application/json')
test_utils.predict_and_assert_response_length(data_as_list, 'text/csv')
test_utils.predict_and_assert_response_length(request_data,
'application/x-npy')
def serve(model_dir=model_gpu_dir, script=mnist_script):
return local_mode.serve(customer_script=script,
model_dir=model_dir,
image_name=docker_image,
use_gpu=True,
opt_ml=opt_ml)