def test_compute_bounds_stddevs_from_mean(self,
comparison,
std_devs,
expected_bounds,
inclusive=False):
# mean = 3, stddev = ~1.414
value_history = range(1, 6)
class _FakeMetricStore:
def get_metric_history(*args, **kwargs):
return [
bigquery_client.MetricHistoryRow('', '',
datetime.datetime.now(),
'', v)
for v in value_history
]
assertion = metrics_pb2.Assertion(
std_devs_from_mean=metrics_pb2.Assertion.StdDevsFromMean(
comparison=metrics_pb2.Assertion.Comparison.Value(comparison),
std_devs=std_devs,
),
inclusive_bounds=inclusive,
)
collector = base.BaseCollector(
metrics_pb2.TestCompletedEvent(benchmark_id="test_benchmark"),
None, _FakeMetricStore())
bounds = collector.compute_bounds("metric_key", assertion)
self.assertSequenceAlmostEqual(
dataclasses.astuple(bounds),
# EQUAL is always inclusive
dataclasses.astuple(
utils.Bounds(*expected_bounds, inclusive
or comparison == 'EQUAL')),
places=3)
def test_compute_bounds_percent_difference_with_mean_value(
self, comparison, pct_diff, expected_bounds, inclusive=False):
# mean = 3
value_history = range(1, 6)
class _FakeMetricStore:
def get_metric_history(*args, **kwargs):
return [
bigquery_client.MetricHistoryRow('', '',
datetime.datetime.now(),
'', v)
for v in value_history
]
assertion = metrics_pb2.Assertion(
percent_difference=metrics_pb2.Assertion.PercentDifference(
comparison=metrics_pb2.Assertion.Comparison.Value(comparison),
use_historical_mean=True,
percent=pct_diff,
),
inclusive_bounds=inclusive,
)
collector = base.BaseCollector(
metrics_pb2.TestCompletedEvent(benchmark_id="test_benchmark"),
None, _FakeMetricStore())
bounds = collector.compute_bounds("metric_key", assertion)
self.assertSequenceAlmostEqual(
dataclasses.astuple(bounds),
dataclasses.astuple(utils.Bounds(*expected_bounds, inclusive)),
places=3)
def test_assert_duration(self):
metric_source = metrics_pb2.MetricSource(
literals=metrics_pb2.LiteralSource(
assertions={
"duration":
metrics_pb2.Assertion(
within_bounds=metrics_pb2.Assertion.WithinBounds(
lower_bound=100,
upper_bound=200,
),
inclusive_bounds=False,
)
}))
event = metrics_pb2.TestCompletedEvent(
benchmark_id="test_benchmark",
duration=duration_pb2.Duration(seconds=150),
metric_collection_config=metrics_pb2.MetricCollectionConfig(
sources=[metric_source]))
collector = literal_collector.LiteralCollector(
event=event, raw_source=metric_source)
points = collector.metric_points()
self.assertLen(points, 1)
self.assertEqual(points[0].metric_key, 'duration')
self.assertEqual(points[0].metric_value, 150)
self.assertEqual(points[0].bounds, utils.Bounds(100, 200, False))
def test_compute_bounds_within_bounds(self,
lower,
upper,
expected_bounds,
inclusive=False):
assertion = metrics_pb2.Assertion(
within_bounds=metrics_pb2.Assertion.WithinBounds(
lower_bound=lower,
upper_bound=upper,
),
inclusive_bounds=inclusive,
)
collector = base.BaseCollector(
metrics_pb2.TestCompletedEvent(benchmark_id="test_benchmark"),
None)
bounds = collector.compute_bounds("metric_key", assertion)
self.assertEqual(bounds, utils.Bounds(*expected_bounds, inclusive))
def test_aggregate_metrics_with_assertion(self):
metric_source = metrics_pb2.MetricSource(
tensorboard=metrics_pb2.TensorBoardSource(
include_tags=[
metrics_pb2.TensorBoardSource.TagStrategy(
tag_pattern="eval/*",
strategies=[
metrics_pb2.TensorBoardSource.FINAL,
metrics_pb2.TensorBoardSource.MAX,
metrics_pb2.TensorBoardSource.MIN,
])
],
aggregate_assertions=[
metrics_pb2.TensorBoardSource.AggregateAssertion(
tag='eval/accuracy',
strategy=metrics_pb2.TensorBoardSource.MAX,
assertion=metrics_pb2.Assertion(
within_bounds=metrics_pb2.Assertion.WithinBounds(
lower_bound=.4,
upper_bound=1.0,
),
inclusive_bounds=True,
))
]))
event = metrics_pb2.TestCompletedEvent(
benchmark_id="test_benchmark",
output_path=self.temp_dir,
metric_collection_config=metrics_pb2.MetricCollectionConfig(
sources=[metric_source]))
collector = tensorboard_collector.TensorBoardCollector(
event=event, raw_source=metric_source)
points = list(collector.metric_points())
self.assertCountEqual(
points,
[
utils.MetricPoint('eval/accuracy_max', .5,
utils.Bounds(.4, 1.0, True)),
utils.MetricPoint('eval/accuracy_min', .125, utils.NO_BOUNDS),
utils.MetricPoint('eval/accuracy_final', .25, utils.NO_BOUNDS),
],
)
def test_compute_bounds_percent_difference_with_target_value(
self,
comparison,
target,
pct_diff,
expected_bounds,
inclusive=False):
assertion = metrics_pb2.Assertion(
percent_difference=metrics_pb2.Assertion.PercentDifference(
comparison=metrics_pb2.Assertion.Comparison.Value(comparison),
value=target,
percent=pct_diff,
),
inclusive_bounds=inclusive,
)
collector = base.BaseCollector(
metrics_pb2.TestCompletedEvent(benchmark_id="test_benchmark"),
None)
bounds = collector.compute_bounds("metric_key", assertion)
self.assertEqual(bounds, utils.Bounds(*expected_bounds, inclusive))
def test_include_and_exclude(self):
metric_source = metrics_pb2.MetricSource(
tensorboard=metrics_pb2.TensorBoardSource(
include_tags=[
metrics_pb2.TensorBoardSource.TagStrategy(
tag_pattern="*",
strategies=[
metrics_pb2.TensorBoardSource.FINAL,
])
],
exclude_tags=[
'foo',
'train/*',
],
aggregate_assertions=[
metrics_pb2.TensorBoardSource.AggregateAssertion(
tag='foo',
strategy=metrics_pb2.TensorBoardSource.MIN,
assertion=metrics_pb2.Assertion(
within_bounds=metrics_pb2.Assertion.WithinBounds(
lower_bound=0.,
upper_bound=2.,
)))
]))
event = metrics_pb2.TestCompletedEvent(
benchmark_id="test_benchmark",
output_path=self.temp_dir,
metric_collection_config=metrics_pb2.MetricCollectionConfig(
sources=[metric_source]))
collector = tensorboard_collector.TensorBoardCollector(
event=event, raw_source=metric_source)
points = list(collector.metric_points())
self.assertCountEqual(
points,
[
utils.MetricPoint('eval/accuracy_final', .25, utils.NO_BOUNDS),
utils.MetricPoint('foo_min', 1, utils.Bounds(0., 2., False)),
],
)
def test_min_time(self, window, timestamp, expected_min):
start_time = timestamp_pb2.Timestamp()
start_time.FromDatetime(datetime.datetime(2021, 2, 16, 0, 0, 0))
min_timestamp = timestamp_pb2.Timestamp()
if timestamp:
min_timestamp.FromDatetime(timestamp)
time_window = duration_pb2.Duration()
if window:
time_window.FromTimedelta(window)
assertion = metrics_pb2.Assertion(
std_devs_from_mean=metrics_pb2.Assertion.StdDevsFromMean(
comparison=metrics_pb2.Assertion.Comparison.WITHIN,
std_devs=1,
),
time_window=time_window,
min_timestamp=min_timestamp,
)
metric_store = bigquery_client.BigQueryMetricStore(
'fake_dataset', 'fake_project')
metric_store.get_metric_history = mock.Mock(return_value=[])
collector = base.BaseCollector(
metrics_pb2.TestCompletedEvent(
benchmark_id="test_benchmark",
start_time=start_time,
), None, metric_store)
collector.compute_bounds("metric_key", assertion)
metric_store.get_metric_history.assert_called_with(
benchmark_id="test_benchmark",
metric_key="metric_key",
min_time=expected_min)
def test_compute_bounds_fixed_value(self,
comparison,
threshold_value,
expected_bounds,
inclusive=False):
assertion = metrics_pb2.Assertion(
fixed_value=metrics_pb2.Assertion.FixedValue(
comparison=metrics_pb2.Assertion.Comparison.Value(comparison),
value=threshold_value,
),
inclusive_bounds=inclusive,
)
collector = base.BaseCollector(
metrics_pb2.TestCompletedEvent(benchmark_id="test_benchmark"),
None)
bounds = collector.compute_bounds("metric_key", assertion)
self.assertSequenceAlmostEqual(
dataclasses.astuple(bounds),
# EQUAL is always inclusive
dataclasses.astuple(
utils.Bounds(*expected_bounds, inclusive
or comparison == 'EQUAL')),
places=3)
def test_metric_collection_config(self, gcs_subdir):
job = _job_from_dict({
'metadata': {
'name': 'job-name',
'namespace': 'namespace',
'labels': {
'benchmarkId': 'test-job',
},
'annotations': {
'ml-testing-accelerators/metric-config': json.dumps({
'sources': [{
'literals': {
'assertions': {
'duration': {
'within_bounds': {
'lower_bound': 1,
'upper_bound': 2,
}
}
}
}
}]
})
}
},
'status': {
'startTime': _START_TIME,
'completionTime': _END_TIME,
'succeeded': 1,
'conditions': [
{
'status': True,
'type': 'Complete',
'lastTransitionTime': _END_TIME,
}
]
}
})
if gcs_subdir:
job.metadata.annotations['ml-testing-accelerators/gcs-subdir'] = gcs_subdir
actual_event = event_publisher.create_test_completed_event(
job,
model_output_bucket='gs://fake-bucket',
cluster_name='cluster-name',
cluster_location='cluster-location',
project='project-id'
)
actual_mcc = actual_event.metric_collection_config
expected_mcc = metrics_pb2.MetricCollectionConfig(
sources=[
metrics_pb2.MetricSource(
literals=metrics_pb2.LiteralSource(
assertions={
'duration': metrics_pb2.Assertion(
within_bounds=metrics_pb2.Assertion.WithinBounds(
lower_bound=1,
upper_bound=2,
)
)
}
)
)
]
)
self.assertEqual(actual_event.output_path, os.path.join('gs://fake-bucket', gcs_subdir or '', 'job-name'))
self.assertProtoEqual(expected_mcc, actual_mcc)
def test_get_metrics_from_perfzero_summary(self):
temp_dir = self.create_tempdir().full_path
summary_dir = os.path.join(temp_dir, 'date_and_time')
pathlib.Path(summary_dir).mkdir(parents=True, exist_ok=True)
summary_path = os.path.join(summary_dir, 'perfzero_summary.json')
with open(summary_path, 'w') as f:
json.dump(
{
"execution_id": "execution_id",
"execution_timestamp": 1234567890.1,
"benchmark_result": {
"wall_time":
1234,
"metrics": [{
"name": "exp_per_second",
"value": 1.1,
}, {
"name": "avg_exp_per_second",
"value": 2.2,
}, {
"name": "startup_time",
"value": 3.3
}],
},
"benchmark_info": {
"not": "important",
},
"setup_info": {},
"ml_framework_info": {
"not": "important",
},
"system_info": {
"not": "important"
},
"process_info": {
"max_rss": 4.4,
"max_vms": 5.5,
"max_cpu_percent": 6.6,
}
}, f)
metric_source = metrics_pb2.MetricSource(
perfzero=metrics_pb2.PerfZeroSource(
assertions={
'total_wall_time':
metrics_pb2.Assertion(
within_bounds=metrics_pb2.Assertion.WithinBounds(
lower_bound=1230,
upper_bound=1240,
)),
'exp_per_second':
metrics_pb2.Assertion(
within_bounds=metrics_pb2.Assertion.WithinBounds(
lower_bound=1,
upper_bound=100,
), )
}))
event = metrics_pb2.TestCompletedEvent(
benchmark_id="test_benchmark",
output_path=temp_dir,
metric_collection_config=metrics_pb2.MetricCollectionConfig(
sources=[metric_source]))
collector = perfzero_collector.PerfZeroCollector(
event=event, raw_source=metric_source)
points = list(collector.metric_points())
self.assertCountEqual(
points,
{
utils.MetricPoint("total_wall_time", 1234,
utils.Bounds(1230., 1240., False)),
utils.MetricPoint("exp_per_second", 1.1,
utils.Bounds(1., 100., False)),
utils.MetricPoint("avg_exp_per_second", 2.2, utils.NO_BOUNDS),
utils.MetricPoint("startup_time", 3.3, utils.NO_BOUNDS),
utils.MetricPoint("process_info/max_rss", 4.4,
utils.NO_BOUNDS),
utils.MetricPoint("process_info/max_vms", 5.5,
utils.NO_BOUNDS),
utils.MetricPoint("process_info/max_cpu_percent", 6.6,
utils.NO_BOUNDS),
},
)