def assertMetricsComputedWithBeamAre(self,
eval_saved_model_path,
serialized_examples,
expected_metrics,
add_metrics_callbacks=None):
"""Checks metrics computed using Beam.
Metrics will be computed over all examples, without any slicing. If you
want to provide your own PCollection (e.g. read a large number of examples
from a file), if you want to check metrics over certain slices, or if you
want to add additional post-export metrics, use the more general
assertGeneralMetricsComputedWithBeamAre.
Example usage:
self.assertMetricsComputedWithBeamAre(
eval_saved_model_path=path,
serialized_examples=[self.makeExample(age=5, label=1.0),
self.makeExample(age=10, label=0.0)],
expected_metrics={'average_loss': 0.1})
Args:
eval_saved_model_path: Path to the directory containing the
EvalSavedModel.
serialized_examples: List of serialized example bytes.
expected_metrics: Dictionary of expected metric values.
add_metrics_callbacks: Optional. Callbacks for adding additional metrics.
"""
def check_metrics(got):
"""Check metrics callback."""
try:
self.assertEqual(
1, len(got),
'expecting metrics for exactly one slice, but got %d '
'slices instead. metrics were: %s' % (len(got), got))
(slice_key, value) = got[0]
self.assertEqual((), slice_key)
self.assertDictElementsWithinBounds(
got_values_dict=value,
expected_values_dict=expected_metrics)
except AssertionError as err:
raise beam_util.BeamAssertException(err)
eval_shared_model = model_eval_lib.default_eval_shared_model(
eval_saved_model_path=eval_saved_model_path,
add_metrics_callbacks=add_metrics_callbacks)
extractors = model_eval_lib.default_extractors(
eval_shared_model=eval_shared_model)
with beam.Pipeline() as pipeline:
# pylint: disable=no-value-for-parameter
metrics, _ = (
pipeline
| 'CreateExamples' >> beam.Create(serialized_examples)
| 'InputsToExtracts' >> model_eval_lib.InputsToExtracts()
| 'Extract' >> Extract(extractors=extractors)
| 'ComputeMetricsAndPlots' >> metrics_and_plots_evaluator.
ComputeMetricsAndPlots(eval_shared_model=eval_shared_model))
# pylint: enable=no-value-for-parameter
beam_util.assert_that(metrics, check_metrics)
def _runTestWithCustomCheck(self,
examples,
eval_export_dir,
metrics,
custom_metrics_check=None,
custom_plots_check=None):
# make sure we are doing some checks
self.assertTrue(custom_metrics_check is not None or
custom_plots_check is not None)
serialized_examples = [ex.SerializeToString() for ex in examples]
eval_shared_model = types.EvalSharedModel(
model_path=eval_export_dir, add_metrics_callbacks=metrics)
extractors = model_eval_lib.default_extractors(
eval_shared_model=eval_shared_model)
with beam.Pipeline() as pipeline:
metrics, plots = (
pipeline
| 'Create' >> beam.Create(serialized_examples)
| 'ToExampleAndExtracts' >> evaluate.ToExampleAndExtracts()
| 'Extract' >> evaluate.Extract(extractors=extractors)
|
'Evaluate' >> evaluate.Evaluate(eval_shared_model=eval_shared_model))
if custom_metrics_check is not None:
util.assert_that(metrics, custom_metrics_check, label='metrics')
if custom_plots_check is not None:
util.assert_that(plots, custom_plots_check, label='plot')
def _runTestWithCustomCheck(self,
examples,
eval_export_dir,
metrics_callbacks,
slice_spec=None,
custom_metrics_check=None,
custom_plots_check=None,
custom_result_check=None):
# make sure we are doing some checks
self.assertTrue(custom_metrics_check is not None
or custom_plots_check is not None
or custom_result_check is not None)
serialized_examples = [ex.SerializeToString() for ex in examples]
slicing_specs = None
if slice_spec:
slicing_specs = [s.to_proto() for s in slice_spec]
eval_config = config.EvalConfig(slicing_specs=slicing_specs)
eval_shared_model = self.createTestEvalSharedModel(
eval_saved_model_path=eval_export_dir,
add_metrics_callbacks=metrics_callbacks)
extractors = model_eval_lib.default_extractors(
eval_config=eval_config, eval_shared_model=eval_shared_model)
tfx_io = raw_tf_record.RawBeamRecordTFXIO(
physical_format='inmemory',
raw_record_column_name=constants.ARROW_INPUT_COLUMN,
telemetry_descriptors=['TFMATest'])
with beam.Pipeline() as pipeline:
(metrics, plots), _ = (
pipeline
| 'Create' >> beam.Create(serialized_examples)
| 'BatchExamples' >> tfx_io.BeamSource()
|
'InputsToExtracts' >> model_eval_lib.BatchedInputsToExtracts()
| 'Extract' >> tfma_unit.Extract(extractors=extractors) # pylint: disable=no-value-for-parameter
| 'ComputeMetricsAndPlots' >>
legacy_metrics_and_plots_evaluator._ComputeMetricsAndPlots( # pylint: disable=protected-access
eval_shared_model=eval_shared_model,
compute_confidence_intervals=self.
compute_confidence_intervals,
random_seed_for_testing=self.deterministic_test_seed))
if custom_metrics_check is not None:
util.assert_that(metrics,
custom_metrics_check,
label='metrics')
if custom_plots_check is not None:
util.assert_that(plots, custom_plots_check, label='plot')
result = pipeline.run()
if custom_result_check is not None:
custom_result_check(result)
def _runTestWithCustomCheck(self,
examples,
eval_export_dir,
metrics_callbacks,
slice_spec=None,
custom_metrics_check=None,
custom_plots_check=None,
custom_result_check=None):
# make sure we are doing some checks
self.assertTrue(custom_metrics_check is not None
or custom_plots_check is not None
or custom_result_check is not None)
serialized_examples = [ex.SerializeToString() for ex in examples]
slicing_specs = None
if slice_spec:
slicing_specs = [s.to_proto() for s in slice_spec]
eval_config = config.EvalConfig(
input_data_specs=[config.InputDataSpec()],
model_specs=[config.ModelSpec(location=eval_export_dir)],
output_data_specs=[config.OutputDataSpec()],
slicing_specs=slicing_specs)
eval_shared_model = self.createTestEvalSharedModel(
eval_saved_model_path=eval_export_dir,
add_metrics_callbacks=metrics_callbacks)
extractors = model_eval_lib.default_extractors(
eval_config=eval_config, eval_shared_models=[eval_shared_model])
with beam.Pipeline() as pipeline:
(metrics, plots), _ = (
pipeline
| 'Create' >> beam.Create(serialized_examples)
| 'InputsToExtracts' >> model_eval_lib.InputsToExtracts()
| 'Extract' >> tfma_unit.Extract(extractors=extractors) # pylint: disable=no-value-for-parameter
| 'ComputeMetricsAndPlots' >>
metrics_and_plots_evaluator.ComputeMetricsAndPlots(
eval_shared_model=eval_shared_model,
compute_confidence_intervals=self.
compute_confidence_intervals,
random_seed_for_testing=self.deterministic_test_seed))
if custom_metrics_check is not None:
util.assert_that(metrics,
custom_metrics_check,
label='metrics')
if custom_plots_check is not None:
util.assert_that(plots, custom_plots_check, label='plot')
result = pipeline.run()
if custom_result_check is not None:
custom_result_check(result)
def assertGeneralMetricsComputedWithBeamAre(
self, eval_saved_model_path: Text,
examples_pcollection: beam.pvalue.PCollection,
slice_spec: List[slicer.SingleSliceSpec],
add_metrics_callbacks: List[types.AddMetricsCallbackType],
expected_slice_metrics: Dict[Any, Dict[Text, Any]]):
"""Checks metrics computed using Beam.
A more general version of assertMetricsComputedWithBeamAre. Note that the
caller is responsible for setting up and running the Beam pipeline.
Example usage:
def add_metrics(features, predictions, labels):
metric_ops = {
'mse': tf.metrics.mean_squared_error(labels, predictions['logits']),
'mae': tf.metrics.mean_absolute_error(labels, predictions['logits']),
}
return metric_ops
with beam.Pipeline() as pipeline:
expected_slice_metrics = {
(): {
'mae': 0.1,
'mse': 0.2,
tfma.post_export_metrics.metric_keys.AUC:
tfma.test.BoundedValue(lower_bound=0.5)
},
(('age', 10),): {
'mae': 0.2,
'mse': 0.3,
tfma.post_export_metrics.metric_keys.AUC:
tfma.test.BoundedValue(lower_bound=0.5)
},
}
examples = pipeline | 'ReadExamples' >> beam.io.ReadFromTFRecord(path)
self.assertGeneralMetricsComputedWithBeamAre(
eval_saved_model_path=path,
examples_pcollection=examples,
slice_spec=[tfma.slicer.SingleSliceSpec(),
tfma.slicer.SingleSliceSpec(columns=['age'])],
add_metrics_callbacks=[
add_metrics, tfma.post_export_metrics.auc()],
expected_slice_metrics=expected_slice_metrics)
Args:
eval_saved_model_path: Path to the directory containing the
EvalSavedModel.
examples_pcollection: A PCollection of serialized example bytes.
slice_spec: List of slice specifications.
add_metrics_callbacks: Callbacks for adding additional metrics.
expected_slice_metrics: Dictionary of dictionaries describing the expected
metrics for each slice. The outer dictionary map slice keys to the
expected metrics for that slice.
"""
def check_metrics(got):
"""Check metrics callback."""
try:
slices = {}
for slice_key, value in got:
slices[slice_key] = value
self.assertItemsEqual(list(slices.keys()),
list(expected_slice_metrics.keys()))
for slice_key, expected_metrics in expected_slice_metrics.items(
):
self.assertDictElementsWithinBounds(
got_values_dict=slices[slice_key],
expected_values_dict=expected_metrics)
except AssertionError as err:
raise beam_util.BeamAssertException(err)
slicing_specs = None
if slice_spec:
slicing_specs = [s.to_proto() for s in slice_spec]
eval_config = config.EvalConfig(slicing_specs=slicing_specs)
eval_shared_model = self.createTestEvalSharedModel(
eval_saved_model_path=eval_saved_model_path,
add_metrics_callbacks=add_metrics_callbacks)
extractors = model_eval_lib.default_extractors(
eval_config=eval_config, eval_shared_model=eval_shared_model)
# pylint: disable=no-value-for-parameter
(metrics,
_), _ = (examples_pcollection
| 'InputsToExtracts' >> model_eval_lib.InputsToExtracts()
| 'Extract' >> Extract(extractors=extractors)
| 'ComputeMetricsAndPlots' >>
legacy_metrics_and_plots_evaluator.ComputeMetricsAndPlots(
eval_shared_model=eval_shared_model))
# pylint: enable=no-value-for-parameter
beam_util.assert_that(metrics, check_metrics)
def assertMetricsComputedWithBeamAre(
self,
eval_saved_model_path: str,
serialized_examples: List[bytes],
expected_metrics: Dict[str, Any],
add_metrics_callbacks: Optional[List[
types.AddMetricsCallbackType]] = None):
"""Checks metrics computed using Beam.
Metrics will be computed over all examples, without any slicing. If you
want to provide your own PCollection (e.g. read a large number of examples
from a file), if you want to check metrics over certain slices, or if you
want to add additional post-export metrics, use the more general
assertGeneralMetricsComputedWithBeamAre.
Example usage:
self.assertMetricsComputedWithBeamAre(
eval_saved_model_path=path,
serialized_examples=[self.makeExample(age=5, label=1.0),
self.makeExample(age=10, label=0.0)],
expected_metrics={'average_loss': 0.1})
Args:
eval_saved_model_path: Path to the directory containing the
EvalSavedModel.
serialized_examples: List of serialized example bytes.
expected_metrics: Dictionary of expected metric values.
add_metrics_callbacks: Optional. Callbacks for adding additional metrics.
"""
def check_metrics(got):
"""Check metrics callback."""
try:
self.assertEqual(
1, len(got),
'expecting metrics for exactly one slice, but got %d '
'slices instead. metrics were: %s' % (len(got), got))
(slice_key, value) = got[0]
self.assertEqual((), slice_key)
self.assertDictElementsWithinBounds(
got_values_dict=value,
expected_values_dict=expected_metrics)
except AssertionError as err:
raise beam_util.BeamAssertException(err)
eval_config = config_pb2.EvalConfig()
eval_shared_model = model_eval_lib.default_eval_shared_model(
eval_saved_model_path=eval_saved_model_path,
add_metrics_callbacks=add_metrics_callbacks)
extractors = model_eval_lib.default_extractors(
eval_config=eval_config, eval_shared_model=eval_shared_model)
tfx_io = raw_tf_record.RawBeamRecordTFXIO(
physical_format='inmemory',
raw_record_column_name=constants.ARROW_INPUT_COLUMN,
telemetry_descriptors=['TFMATest'])
with beam.Pipeline() as pipeline:
# pylint: disable=no-value-for-parameter
(metrics, _), _ = (
pipeline
| 'CreateExamples' >> beam.Create(serialized_examples)
| 'BatchExamples' >> tfx_io.BeamSource()
|
'InputsToExtracts' >> model_eval_lib.BatchedInputsToExtracts()
| 'Extract' >> Extract(extractors=extractors)
| 'ComputeMetricsAndPlots' >>
legacy_metrics_and_plots_evaluator._ComputeMetricsAndPlots( # pylint: disable=protected-access
eval_shared_model=eval_shared_model))
# pylint: enable=no-value-for-parameter
beam_util.assert_that(metrics, check_metrics)
