def load(self,
model_name,
model_dir,
handler,
gpu_id,
batch_size,
envelope=None):
"""
Load TorchServe 1.0 model from file.
:param model_name:
:param model_dir:
:param handler:
:param gpu_id:
:param batch_size:
:param envelope:
:return:
"""
logging.debug("Loading model - working dir: %s", os.getcwd())
# TODO: Request ID is not given. UUID is a temp UUID.
metrics = MetricsStore(uuid.uuid4(), model_name)
manifest_file = os.path.join(model_dir, "MAR-INF/MANIFEST.json")
manifest = None
if os.path.exists(manifest_file):
with open(manifest_file) as f:
manifest = json.load(f)
function_name = None
try:
module, function_name = self._load_handler_file(handler)
except ImportError:
module = self._load_default_handler(handler)
if module is None:
raise ValueError(
"Unable to load module {}, make sure it is added to python path"
.format(module_name))
envelope_class = None
if envelope is not None:
envelope_class = self._load_default_envelope(envelope)
function_name = function_name or "handle"
if hasattr(module, function_name):
entry_point, initialize_fn = self._get_function_entry_point(
module, function_name)
else:
entry_point, initialize_fn = self._get_class_entry_point(module)
if envelope_class is not None:
envelope_instance = envelope_class(entry_point)
entry_point = envelope_instance.handle
service = Service(model_name, model_dir, manifest, entry_point, gpu_id,
batch_size)
service.context.metrics = metrics
initialize_fn(service.context)
return service
def predict(self, batch):
"""
PREDICT COMMAND = {
"command": "predict",
"batch": [ REQUEST_INPUT ]
}
:param batch: list of request
:return:
"""
headers, input_batch, req_id_map = Service.retrieve_data_for_inference(
batch)
self.context.request_ids = req_id_map
self.context.request_processor = headers
metrics = MetricsStore(req_id_map, self.context.model_name)
self.context.metrics = metrics
start_time = time.time()
# noinspection PyBroadException
try:
ret = self._entry_point(input_batch, self.context)
except MemoryError:
logger.error("System out of memory", exc_info=True)
return create_predict_response(None, req_id_map,
"Out of resources", 507)
except Exception as e: # pylint: disable=broad-except
logger.warning("Invoking custom service failed.", exc_info=True)
message = "Prediction failed: " + str(e) + " \n " + str(
traceback.format_exc())
return create_predict_response(None, req_id_map, message, 503)
if not isinstance(ret, list):
logger.warning("model: %s, Invalid return type: %s.",
self.context.model_name, type(ret))
return create_predict_response(None, req_id_map,
"Invalid model predict output", 503)
if len(ret) != len(input_batch):
logger.warning(
"model: %s, number of batch response mismatched, expect: %d, got: %d.",
self.context.model_name, len(input_batch), len(ret))
return create_predict_response(
None, req_id_map, "number of batch response mismatched", 503)
duration = round((time.time() - start_time) * 1000, 2)
metrics.add_time(PREDICTION_METRIC, duration)
return create_predict_response(ret,
req_id_map,
"Prediction success",
200,
context=self.context)
def __init__(self,
model_pt_file='model.pt',
model_dir='ts/torch_handler/unit_tests/models/tmp',
model_file='model.py',
gpu_id='0',
model_name="mnist"):
self.manifest = {
'model': {
'serializedFile': model_pt_file,
}
}
if model_file:
self.manifest['model']['modelFile'] = model_file
self.system_properties = {'model_dir': model_dir}
if torch.cuda.is_available() and gpu_id:
self.system_properties['gpu_id'] = gpu_id
self.metrics = MetricsStore(uuid.uuid4(), model_name)
def test_metrics(caplog):
"""
Test if metric classes methods behave as expected
Also checks global metric service methods
"""
# Create a batch of request ids
request_ids = {0: 'abcd', 1: "xyz", 2: "qwerty", 3: "hjshfj"}
all_req_ids = ','.join(request_ids.values())
model_name = "dummy model"
# Create a metrics objects
metrics = MetricsStore(request_ids, model_name)
# Counter tests
metrics.add_counter('CorrectCounter', 1, 1)
test_metric = metrics.cache[get_model_key('CorrectCounter', 'count', 'xyz',
model_name)]
assert 'CorrectCounter' == test_metric.name
metrics.add_counter('CorrectCounter', 1, 1)
metrics.add_counter('CorrectCounter', 1, 3)
metrics.add_counter('CorrectCounter', 1)
test_metric = metrics.cache[get_model_key('CorrectCounter', 'count',
all_req_ids, model_name)]
assert 'CorrectCounter' == test_metric.name
metrics.add_counter('CorrectCounter', 3)
test_metric = metrics.cache[get_model_key('CorrectCounter', 'count', 'xyz',
model_name)]
assert test_metric.value == 2
test_metric = metrics.cache[get_model_key('CorrectCounter', 'count',
'hjshfj', model_name)]
assert test_metric.value == 1
test_metric = metrics.cache[get_model_key('CorrectCounter', 'count',
all_req_ids, model_name)]
assert test_metric.value == 4
# Check what is emitted is correct
emit_metrics(metrics.store)
assert "hjshfj" in caplog.text
assert "ModelName:dummy model" in caplog.text
# Adding other types of metrics
# Check for time metric
with pytest.raises(Exception) as e_info:
metrics.add_time('WrongTime', 20, 1, 'ns')
assert "the unit for a timed metric should be one of ['ms', 's']" == e_info.value.args[
0]
metrics.add_time('CorrectTime', 20, 2, 's')
metrics.add_time('CorrectTime', 20, 0)
test_metric = metrics.cache[get_model_key('CorrectTime', 'ms', 'abcd',
model_name)]
assert test_metric.value == 20
assert test_metric.unit == 'Milliseconds'
test_metric = metrics.cache[get_model_key('CorrectTime', 's', 'qwerty',
model_name)]
assert test_metric.value == 20
assert test_metric.unit == 'Seconds'
# Size based metrics
with pytest.raises(Exception) as e_info:
metrics.add_size('WrongSize', 20, 1, 'TB')
assert "The unit for size based metric is one of ['MB','kB', 'GB', 'B']" == e_info.value.args[
0]
metrics.add_size('CorrectSize', 200, 0, 'GB')
metrics.add_size('CorrectSize', 10, 2)
test_metric = metrics.cache[get_model_key('CorrectSize', 'GB', 'abcd',
model_name)]
assert test_metric.value == 200
assert test_metric.unit == 'Gigabytes'
test_metric = metrics.cache[get_model_key('CorrectSize', 'MB', 'qwerty',
model_name)]
assert test_metric.value == 10
assert test_metric.unit == 'Megabytes'
# Check a percentage metric
metrics.add_percent('CorrectPercent', 20.0, 3)
test_metric = metrics.cache[get_model_key('CorrectPercent', 'percent',
'hjshfj', model_name)]
assert test_metric.value == 20.0
assert test_metric.unit == 'Percent'
# Check a error metric
metrics.add_error('CorrectError', 'Wrong values')
test_metric = metrics.cache[get_error_key('CorrectError', '')]
assert test_metric.value == 'Wrong values'
def load(self, model_name, model_dir, handler, gpu_id, batch_size):
"""
Load TorchServe 1.0 model from file.
:param model_name:
:param model_dir:
:param handler:
:param gpu_id:
:param batch_size:
:return:
"""
logging.debug("Loading model - working dir: %s", os.getcwd())
# TODO: Request ID is not given. UUID is a temp UUID.
metrics = MetricsStore(uuid.uuid4(), model_name)
manifest_file = os.path.join(model_dir, "MAR-INF/MANIFEST.json")
manifest = None
if os.path.exists(manifest_file):
with open(manifest_file) as f:
manifest = json.load(f)
try:
temp = handler.split(":", 1)
module_name = temp[0]
function_name = None if len(temp) == 1 else temp[1]
if module_name.endswith(".py"):
module_name = module_name[:-3]
module_name = module_name.split("/")[-1]
module = importlib.import_module(module_name)
# pylint: disable=unused-variable
except ImportError as e:
module_name = ".{0}".format(handler)
module = importlib.import_module(module_name, 'ts.torch_handler')
function_name = None
if module is None:
raise ValueError(
"Unable to load module {}, make sure it is added to python path"
.format(module_name))
if function_name is None:
function_name = "handle"
if hasattr(module, function_name):
entry_point = getattr(module, function_name)
service = Service(model_name, model_dir, manifest, entry_point,
gpu_id, batch_size)
service.context.metrics = metrics
# initialize model at load time
entry_point(None, service.context)
else:
model_class_definitions = list_classes_from_module(module)
if len(model_class_definitions) != 1:
raise ValueError(
"Expected only one class in custom service code or a function entry point {}"
.format(model_class_definitions))
model_class = model_class_definitions[0]
model_service = model_class()
handle = getattr(model_service, "handle")
if handle is None:
raise ValueError("Expect handle method in class {}".format(
str(model_class)))
service = Service(model_name, model_dir, manifest,
model_service.handle, gpu_id, batch_size)
initialize = getattr(model_service, "initialize")
if initialize is not None:
model_service.initialize(service.context)
else:
raise ValueError("Expect initialize method in class {}".format(
str(model_class)))
return service