def test_convert_predict_response_regression(self):
"""Test converting a PredictResponse to a RegressionResponse."""
predict = predict_pb2.PredictResponse()
output = predict.outputs['scores']
dim = output.tensor_shape.dim.add()
dim.size = 2
output.float_val.extend([0.1, 0.2])
bundle = inference_utils.ServingBundle('', '', 'regression', '', '',
True, '', 'scores')
converted = common_utils.convert_predict_response(predict, bundle)
self.assertAlmostEqual(0.1, converted.result.regressions[0].value)
self.assertAlmostEqual(0.2, converted.result.regressions[1].value)
def call_servo(examples, serving_bundle):
"""Send an RPC request to the Servomatic prediction service.
Args:
examples: A list of examples that matches the model spec.
serving_bundle: A `ServingBundle` object that contains the information to
make the serving request.
Returns:
A ClassificationResponse or RegressionResponse proto.
"""
parsed_url = urlparse('http://' + serving_bundle.inference_address)
channel = implementations.insecure_channel(parsed_url.hostname,
parsed_url.port)
stub = prediction_service_pb2.beta_create_PredictionService_stub(channel)
if serving_bundle.use_predict:
request = predict_pb2.PredictRequest()
elif serving_bundle.model_type == 'classification':
request = classification_pb2.ClassificationRequest()
else:
request = regression_pb2.RegressionRequest()
request.model_spec.name = serving_bundle.model_name
if serving_bundle.model_version is not None:
request.model_spec.version.value = serving_bundle.model_version
if serving_bundle.signature is not None:
request.model_spec.signature_name = serving_bundle.signature
if serving_bundle.use_predict:
# tf.compat.v1 API used here to convert tf.example into proto. This
# utility file is bundled in the witwidget pip package which has a dep
# on TensorFlow.
request.inputs[serving_bundle.predict_input_tensor].CopyFrom(
tf.compat.v1.make_tensor_proto(
values=[ex.SerializeToString() for ex in examples],
dtype=types_pb2.DT_STRING))
else:
request.input.example_list.examples.extend(examples)
if serving_bundle.use_predict:
return common_utils.convert_predict_response(
stub.Predict(request, 30.0), serving_bundle) # 30 secs timeout
elif serving_bundle.model_type == 'classification':
return stub.Classify(request, 30.0) # 30 secs timeout
else:
return stub.Regress(request, 30.0) # 30 secs timeout
def test_convert_predict_response_classification(self):
"""Test converting a PredictResponse to a ClassificationResponse."""
predict = predict_pb2.PredictResponse()
output = predict.outputs['probabilities']
dim = output.tensor_shape.dim.add()
dim.size = 3
dim = output.tensor_shape.dim.add()
dim.size = 2
output.float_val.extend([1., 0., .9, .1, .8, .2])
bundle = inference_utils.ServingBundle('', '', 'classification', '',
'', True, '', 'probabilities')
converted = common_utils.convert_predict_response(predict, bundle)
self.assertEqual("0",
converted.result.classifications[0].classes[0].label)
self.assertAlmostEqual(
1, converted.result.classifications[0].classes[0].score)
self.assertEqual("1",
converted.result.classifications[0].classes[1].label)
self.assertAlmostEqual(
0, converted.result.classifications[0].classes[1].score)
self.assertEqual("0",
converted.result.classifications[1].classes[0].label)
self.assertAlmostEqual(
.9, converted.result.classifications[1].classes[0].score)
self.assertEqual("1",
converted.result.classifications[1].classes[1].label)
self.assertAlmostEqual(
.1, converted.result.classifications[1].classes[1].score)
self.assertEqual("0",
converted.result.classifications[2].classes[0].label)
self.assertAlmostEqual(
.8, converted.result.classifications[2].classes[0].score)
self.assertEqual("1",
converted.result.classifications[2].classes[1].label)
self.assertAlmostEqual(
.2, converted.result.classifications[2].classes[1].score)
