def _run_inference_with_beam(self, example_path, inference_endpoint,
prediction_log_path):
with beam.Pipeline() as pipeline:
_ = (pipeline
| 'ReadExamples' >> beam.io.ReadFromTFRecord(example_path)
| 'ParseExamples' >> beam.Map(tf.train.Example.FromString)
| 'RunInference' >>
run_inference.RunInference(inference_endpoint)
| 'WritePredictions' >> beam.io.WriteToTFRecord(
prediction_log_path,
coder=beam.coders.ProtoCoder(
prediction_log_pb2.PredictionLog)))
def testTelemetry(self):
example_path = self._get_output_data_dir('examples')
self._prepare_multihead_examples(example_path)
model_path = self._get_output_data_dir('model')
self._build_multihead_model(model_path)
inference_endpoint = model_spec_pb2.InferenceEndpoint(
saved_model_spec=model_spec_pb2.SavedModelSpec(
model_path=model_path, signature_name=['classify_sum']))
pipeline = beam.Pipeline()
_ = (
pipeline | 'ReadExamples' >> beam.io.ReadFromTFRecord(example_path)
| 'ParseExamples' >> beam.Map(tf.train.Example.FromString)
| 'RunInference' >> run_inference.RunInference(inference_endpoint))
run_result = pipeline.run()
run_result.wait_until_finish()
num_inferences = run_result.metrics().query(
MetricsFilter().with_name('num_inferences'))
self.assertTrue(num_inferences['counters'])
self.assertEqual(num_inferences['counters'][0].result, 2)
num_instances = run_result.metrics().query(
MetricsFilter().with_name('num_instances'))
self.assertTrue(num_instances['counters'])
self.assertEqual(num_instances['counters'][0].result, 2)
inference_request_batch_size = run_result.metrics().query(
MetricsFilter().with_name('inference_request_batch_size'))
self.assertTrue(inference_request_batch_size['distributions'])
self.assertEqual(
inference_request_batch_size['distributions'][0].result.sum, 2)
inference_request_batch_byte_size = run_result.metrics().query(
MetricsFilter().with_name('inference_request_batch_byte_size'))
self.assertTrue(inference_request_batch_byte_size['distributions'])
self.assertEqual(
inference_request_batch_byte_size['distributions'][0].result.sum,
sum(element.ByteSize() for element in self._multihead_examples))
inference_batch_latency_micro_secs = run_result.metrics().query(
MetricsFilter().with_name('inference_batch_latency_micro_secs'))
self.assertTrue(inference_batch_latency_micro_secs['distributions'])
self.assertGreaterEqual(
inference_batch_latency_micro_secs['distributions'][0].result.sum,
0)
load_model_latency_milli_secs = run_result.metrics().query(
MetricsFilter().with_name('load_model_latency_milli_secs'))
self.assertTrue(load_model_latency_milli_secs['distributions'])
self.assertGreaterEqual(
load_model_latency_milli_secs['distributions'][0].result.sum, 0)
def _run_model_inference(self, model_path: Text,
example_uris: Mapping[Text,
Text], output_path: Text,
model_spec: bulk_inferrer_pb2.ModelSpec) -> None:
"""Runs model inference on given example data.
Args:
model_path: Path to model.
example_uris: Mapping of example split name to example uri.
output_path: Path to output generated prediction logs.
model_spec: bulk_inferrer_pb2.ModelSpec instance.
Returns:
None
"""
saved_model_spec = model_spec_pb2.SavedModelSpec(
model_path=model_path,
tag=model_spec.tag,
signature_name=model_spec.model_signature_name)
inference_endpoint = model_spec_pb2.InferenceEndpoint()
inference_endpoint.saved_model_spec.CopyFrom(saved_model_spec)
with self._make_beam_pipeline() as pipeline:
data_list = []
for split, example_uri in example_uris.items():
data = (
pipeline
| 'ReadData[{}]'.format(split) >> beam.io.ReadFromTFRecord(
file_pattern=io_utils.all_files_pattern(example_uri)))
data_list.append(data)
_ = ([data for data in data_list]
| 'FlattenExamples' >> beam.Flatten(pipeline=pipeline)
| 'ParseExamples' >> beam.Map(tf.train.Example.FromString)
| 'RunInference' >>
run_inference.RunInference(inference_endpoint)
| 'WritePredictionLogs' >> beam.io.WriteToTFRecord(
output_path,
file_name_suffix='.gz',
coder=beam.coders.ProtoCoder(
prediction_log_pb2.PredictionLog)))
logging.info('Inference result written to %s.', output_path)