ryanthompson591 commented on code in PR #16970:
URL: https://github.com/apache/beam/pull/16970#discussion_r842850681
##########
sdks/python/apache_beam/ml/inference/base.py:
##########
@@ -0,0 +1,194 @@
+import logging
+import platform
+import resource
+import sys
+import time
+from typing import Any
+from typing import List
+
+import apache_beam as beam
+from apache_beam.utils import shared
+from apache_beam.ml.inference.apis import PredictionResult
+
+_MILLISECOND_TO_MICROSECOND = 1000
+_MICROSECOND_TO_NANOSECOND = 1000
+_SECOND_TO_MICROSECOND = 1000000
+
+
+def _unbatch(maybe_keyed_batches: Any):
+ keys, results = maybe_keyed_batches
+ if keys:
+ return zip(keys, results)
+ else:
+ return results
+
+
+class ModelLoader:
+ """Has the ability to load an ML model."""
+ def load_model(self):
+ """Loads an initializes a model for processing."""
+ raise NotImplementedError(type(self))
+
+
+class InferenceRunner:
+ """Implements running inferences for a framework."""
+ def run_inference(self, batch: Any, model: Any) -> List[PredictionResult]:
+ """Runs inferences on a batch of examples and returns a list of
Predictions."""
+ raise NotImplementedError(type(self))
+
+
+class MetricsCollector:
+ """A metrics collector that tracks ML related performance and memory
usage."""
+ def __init__(self, namespace: str):
+ # Metrics
+ self._inference_counter = beam.metrics.Metrics.counter(
+ namespace, 'num_inferences')
+ self._inference_request_batch_size = beam.metrics.Metrics.distribution(
+ namespace, 'inference_request_batch_size')
+ self._inference_request_batch_byte_size = (
+ beam.metrics.Metrics.distribution(
+ namespace, 'inference_request_batch_byte_size'))
+ # Batch inference latency in microseconds.
+ self._inference_batch_latency_micro_secs = (
+ beam.metrics.Metrics.distribution(
+ namespace, 'inference_batch_latency_micro_secs'))
+ self._model_byte_size = beam.metrics.Metrics.distribution(
+ namespace, 'model_byte_size')
+ # Model load latency in milliseconds.
+ self._load_model_latency_milli_secs = beam.metrics.Metrics.distribution(
+ namespace, 'load_model_latency_milli_secs')
+
+ # Metrics cache
+ self.load_model_latency_milli_secs_cache = None
+ self.model_byte_size_cache = None
+
+ def update_metrics_with_cache(self):
+ if self.load_model_latency_milli_secs_cache is not None:
+ self._load_model_latency_milli_secs.update(
+ self.load_model_latency_milli_secs_cache)
+ self.load_model_latency_milli_secs_cache = None
+ if self.model_byte_size_cache is not None:
+ self._model_byte_size.update(self.model_byte_size_cache)
+ self.model_byte_size_cache = None
+
+ def update(
+ self,
+ examples_count: int,
+ examples_byte_size: int,
+ latency_micro_secs: int):
+ self._inference_batch_latency_micro_secs.update(latency_micro_secs)
+ self._inference_counter.inc(examples_count)
+ self._inference_request_batch_size.update(examples_count)
+ self._inference_request_batch_byte_size.update(examples_byte_size)
+
+
+class RunInferenceDoFn(beam.DoFn):
+ def __init__(self, model_loader, inference_runner):
+ self._model_loader = model_loader
+ self._inference_runner = inference_runner
+ self._shared_model_handle = shared.Shared()
+ # TODO: Compute a good metrics namespace
+ self._metrics_collector = MetricsCollector('default_namespace')
+ 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()
+ model = self._model_loader.load_model()
+ end_time = self._clock.get_current_time_in_microseconds()
+ memory_after = _get_current_process_memory_in_bytes()
+ self._metrics_collector.load_model_latency_milli_secs_cache = (
+ (end_time - start_time) / _MILLISECOND_TO_MICROSECOND)
+ self._metrics_collector.model_byte_size_cache = (
+ memory_after - memory_before)
+ return model
+
+ return self._shared_model_handle.acquire(load)
+
+ def setup(self):
+ super().setup()
+ self._model = self._load_model()
+
+ def process(self, batch):
+ has_keys = isinstance(batch[0], tuple)
+ start_time = self._clock.get_current_time_in_microseconds()
+ if has_keys:
+ examples = [example for _, example in batch]
+ keys = [key for key, _ in batch]
+ else:
+ examples = batch
+ keys = None
+ inferences = self._inference_runner.run_inference(examples, self._model)
+ inference_latency = self._clock.get_current_time_in_microseconds(
+ ) - start_time
+ num_bytes = sys.getsizeof(batch)
+ num_elements = len(batch)
+ self._metrics_collector.update(
+ num_elements, sys.getsizeof(batch), inference_latency)
+ yield keys, [PredictionResult(e, r) for e, r in zip(examples, inferences)]
+
+
+class RunInferenceImpl(beam.PTransform):
+ def __init__(self, model_loader, inference_runner):
+ self._model_loader = model_loader
+ self._inference_runner = inference_runner
+
+ def expand(self, pcoll: beam.PCollection) -> beam.PCollection:
+ return (
+ pcoll
+ # TODO: Hook into the batching DoFn APIs.
+ | beam.BatchElements()
+ | beam.ParDo(
+ RunInferenceDoFn(self._model_loader, self._inference_runner))
+ | beam.FlatMap(_unbatch))
+
+
+def _is_darwin() -> bool:
+ return sys.platform == 'darwin'
+
+
+def _get_current_process_memory_in_bytes():
+ """Returns memory usage in bytes."""
+
+ if resource is not None:
+ usage = resource.getrusage(resource.RUSAGE_SELF).ru_maxrss
+ if _is_darwin():
+ return usage
+ return usage * 1024
+ else:
+ logging.warning(
+ 'Resource module is not available for current platform, '
+ 'memory usage cannot be fetched.')
+ return 0
+
+
+def _is_windows() -> bool:
+ return platform.system() == 'Windows' or os.name == 'nt'
+
+
+def _is_cygwin() -> bool:
+ return platform.system().startswith('CYGWIN_NT')
+
+
+class _Clock(object):
Review Comment:
I want to punt on this change for now and talk to the TFX team as to why
they did this, and convince them to use the cleaner time.time solution if they
have no reason not to.
https://issues.apache.org/jira/browse/BEAM-14255
I spent some time experimenting with this. AFAICT it's possible we lose 1
microsecond of precision in a measurement 50% of the time if we use the
floating point time, instead of the nanosecond interface.
The cleanest would be to have no abstractions at all and just use time.
--
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: github-unsubscr...@beam.apache.org
For queries about this service, please contact Infrastructure at:
us...@infra.apache.org
