ryanthompson591 commented on code in PR #17671:
URL: https://github.com/apache/beam/pull/17671#discussion_r876130046
##########
sdks/python/apache_beam/ml/inference/base.py:
##########
@@ -150,27 +155,24 @@ def update(
class _RunInferenceDoFn(beam.DoFn):
"""A DoFn implementation generic to frameworks."""
- def __init__(self, model_loader: ModelLoader, clock=None):
+ def __init__(self, model_loader: ModelLoader, clock):
self._model_loader = model_loader
self._inference_runner = model_loader.get_inference_runner()
self._shared_model_handle = shared.Shared()
self._metrics_collector = _MetricsCollector(
self._inference_runner.get_metrics_namespace())
self._clock = clock
- if not clock:
- self._clock = _ClockFactory.make_clock()
self._model = None
def _load_model(self):
def load():
"""Function for constructing shared LoadedModel."""
memory_before = _get_current_process_memory_in_bytes()
- start_time = self._clock.get_current_time_in_microseconds()
+ start_time = _to_milliseconds(self._clock.time())
model = self._model_loader.load_model()
- end_time = self._clock.get_current_time_in_microseconds()
+ end_time = _to_milliseconds(self._clock.time())
Review Comment:
If you look at the line I removed 172, the previous behavior was to get
start and end time in microseconds and then convert to milliseconds down below.
For some reason the tensorflow team felt (and it seems right) that
milliseconds was adequate for model loading, whereas microseconds were used for
prediction measurement.
I think it's possible model loading can take hundreds or thousands of
milliseconds whereas a prediction might not take a single millisecond.
##########
sdks/python/apache_beam/ml/inference/base_test.py:
##########
@@ -133,14 +133,14 @@ def test_timing_metrics(self):
MetricsFilter().with_name('inference_batch_latency_micro_secs')))
Review Comment:
No this stays the same. I think you are mistaking the inference latency --
this one in microseconds, with the model loading latency, which is currently
just called load_model_latency.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
