def testPredictionSavedModelWithBadInput(self):
data_path = os.path.join(FLAGS.test_srcdir, SAVED_TESTDATA_BAD)
with open(data_path) as f:
# Read two input records as strings.
instances = [
json.loads(next(f).rstrip("\n")),
json.loads(next(f).rstrip("\n"))
]
model_path = os.path.join(FLAGS.test_srcdir, INPUT_MODEL)
with self.assertRaises(mlprediction.PredictionError) as error:
mlprediction.local_predict(model_dir=model_path,
instances=instances)
self.assertTrue(
"Unexpected tensor name: x" in error.exception.error_detail)
def testLocalPredictionTensorflowModelWithStrings(self,
mock_create_client):
signature_def = meta_graph_pb2.SignatureDef()
signature_def.outputs["x_bytes"].dtype = types_pb2.DT_STRING
signature_def.inputs["x_bytes"].dtype = types_pb2.DT_STRING
mock_client = mock.Mock()
mock_client.signature_map = {
tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY:
signature_def
}
mock_client.predict.return_value = {"x_bytes": "to_encode"}
mock_create_client.return_value = mock_client
predictions = mlprediction.local_predict(model_dir=None,
instances=[{
"x_bytes": [1, 2, 3]
}])
# Validate that the output is correctly base64 encoded (and only once)
self.assertEquals(predictions, {
"predictions": [{
"x_bytes": {
"b64": base64.b64encode("to_encode")
}
}]
})
def testPredictionWithSavedModel(self):
data_path = os.path.join(FLAGS.test_srcdir, SAVED_TESTDATA)
with open(data_path) as f:
# Read two input records as strings.
instances = [
json.loads(next(f).rstrip("\n")),
json.loads(next(f).rstrip("\n"))
]
model_path = os.path.join(FLAGS.test_srcdir, INPUT_MODEL)
predictions = mlprediction.local_predict(model_dir=model_path,
instances=instances)
prediction_result = [{
"key": p["key"],
"prediction": p["prediction"]
} for p in predictions["predictions"]]
# Just check the key and prediction result, not each individual scores
# that are floating numbers.
self.assertEqual(2, len(prediction_result))
for r in prediction_result:
if r["key"] == 0:
self.assertEqual(3, r["prediction"])
elif r["key"] == 9:
self.assertEqual(4, r["prediction"])
else:
self.fail()
def testLocalPredictionXgboostModel(self):
# Uses a trained xgboost model that computes x+y
instances = [[10, 20], [1, 2], [5, 6]]
model_path = os.path.join(FLAGS.test_srcdir, XGBOOST_MODEL)
predictions = mlprediction.local_predict(
model_dir=model_path,
instances=instances,
framework=mlprediction.XGBOOST_FRAMEWORK_NAME)
self.assertEqual([round(i) for i in predictions["predictions"]],
[30, 3, 11])
def testLocalPredictionSklearnModel(self):
# Uses a trained sklearn model that computes x+y
instances = [[10, 20], [1, 2], [5, 6]]
model_path = os.path.join(FLAGS.test_srcdir, SKLEARN_JOBLIB_MODEL)
predictions = mlprediction.local_predict(
model_dir=model_path,
instances=instances,
framework=mlprediction.SCIKIT_LEARN_FRAMEWORK_NAME)
self.assertEqual(predictions, {"predictions": [30, 3, 11]})
def testPredictWithSavedModelWithCustomSignature(self):
model_dir = os.path.join(FLAGS.test_tmpdir, "identity_model_predict")
model_test_util.create_identity_model(model_dir=model_dir,
signature_name="mysignature",
tags=("tag1", "tag2"))
result = mlprediction.local_predict(model_dir,
tags=("tag1", "tag2"),
signature_name="mysignature",
instances=[{
"in": "check"
}])
self.assertEqual(result["predictions"], [{"out": "check"}])
# Only one signature_def in the graph, so it's optional to specify it.
result = mlprediction.local_predict(model_dir,
tags=("tag1", "tag2"),
instances=[{
"in": "check"
}])
self.assertEqual(result["predictions"], [{"out": "check"}])
def testPredictWithSavedModelMultipleSignatures(self):
model_dir = os.path.join(FLAGS.test_tmpdir, "constant_model_predict")
model_test_util.create_constant_model(model_dir,
"mysignature",
1,
"serving_default",
2,
tags=("tag1", "tag2"))
# Predict with specified signature.
result = mlprediction.local_predict(model_dir,
tags=("tag1", "tag2"),
signature_name="mysignature",
instances=[{
"in": "check"
}])
self.assertEqual(result["predictions"], [{"out": 1}])
# Predict without specified signature will use serving default.
result = mlprediction.local_predict(model_dir,
tags=("tag1", "tag2"),
instances=[{
"in": "check"
}])
self.assertEqual(result["predictions"], [{"out": 2}])
# Predict with wrong specified signature.
with self.assertRaises(mlprediction.PredictionError) as error:
result = mlprediction.local_predict(
model_dir,
tags=("tag1", "tag2"),
signature_name="wrongsignature",
instances=[{
"in": "check"
}])
self.assertEqual(
"No signature found for signature key wrongsignature.",
error.exception.error_detail)
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--model-dir', required=True, help='Path of the model.')
args, _ = parser.parse_known_args()
if not _verify_tensorflow():
sys.exit(-1)
instances = []
for line in sys.stdin:
instance = json.loads(line.rstrip('\n'))
instances.append(instance)
prediction = _import_prediction_lib()
predictions = prediction.local_predict(model_dir=args.model_dir,
instances=instances)
print(json.dumps(predictions))
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--model-dir', required=True, help='Path of the model.')
args, _ = parser.parse_known_args()
if not _has_required_package():
sys.exit(-1)
instances = []
for line in sys.stdin:
instance = json.loads(line.rstrip('\n'))
instances.append(instance)
# pylint: disable=g-import-not-at-top
from google.cloud.ml import prediction
# pylint: enable=g-import-not-at-top
predictions = prediction.local_predict(model_dir=args.model_dir,
instances=instances)
print(json.dumps(predictions))
