def testMultiModelPredict(self):
temp_eval_export_dir = self._getEvalExportDir()
_, model1_dir = linear_classifier.simple_linear_classifier(
None, temp_eval_export_dir)
model1 = model_eval_lib.default_eval_shared_model(
eval_saved_model_path=model1_dir)
_, model2_dir = linear_classifier.simple_linear_classifier(
None, temp_eval_export_dir)
model2 = model_eval_lib.default_eval_shared_model(
eval_saved_model_path=model2_dir)
eval_shared_model = {'model1': model1, 'model2': model2}
eval_config = config.EvalConfig(model_specs=[
config.ModelSpec(name='model1', example_weight_key='age'),
config.ModelSpec(name='model2', example_weight_key='age')
])
tfx_io = raw_tf_record.RawBeamRecordTFXIO(
physical_format='inmemory',
raw_record_column_name=constants.ARROW_INPUT_COLUMN,
telemetry_descriptors=['TFMATest'])
extractor = predict_extractor.PredictExtractor(
eval_shared_model, eval_config=eval_config)
with beam.Pipeline() as pipeline:
examples = [
self._makeExample(age=3.0, language='english', label=1.0),
self._makeExample(age=3.0, language='chinese', label=0.0),
self._makeExample(age=4.0, language='english', label=1.0),
self._makeExample(age=5.0, language='chinese', label=0.0),
]
serialized_examples = [e.SerializeToString() for e in examples]
predict_extracts = (
pipeline
| beam.Create(serialized_examples, reshuffle=False)
| 'BatchExamples' >> tfx_io.BeamSource(batch_size=2)
| 'InputsToExtracts' >> model_eval_lib.BatchedInputsToExtracts()
| 'Predict' >> extractor.ptransform)
def check_result(got):
try:
self.assertLen(got, 2)
for item in got:
self.assertIn(constants.FEATURES_KEY, item)
for feature in ('language', 'age'):
for features_dict in item[constants.FEATURES_KEY]:
self.assertIn(feature, features_dict)
self.assertIn(constants.LABELS_KEY, item)
self.assertIn(constants.PREDICTIONS_KEY, item)
for model in ('model1', 'model2'):
for predictions_dict in item[constants.PREDICTIONS_KEY]:
self.assertIn(model, predictions_dict)
self.assertIn(constants.EXAMPLE_WEIGHTS_KEY, item)
for i in range(len(item[constants.FEATURES_KEY])):
self.assertAlmostEqual(item[constants.FEATURES_KEY][i]['age'],
item[constants.EXAMPLE_WEIGHTS_KEY][i])
except AssertionError as err:
raise util.BeamAssertException(err)
util.assert_that(predict_extracts, 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(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 testE2E(self):
column_name = "raw_record"
tfxio = raw_tf_record.RawBeamRecordTFXIO(
physical_format="inmem",
raw_record_column_name=column_name,
telemetry_descriptors=["some", "component"])
def _AssertFn(record_batches):
self.assertLen(record_batches, 1)
record_batch = record_batches[0]
self.assertTrue(record_batch.schema.equals(tfxio.ArrowSchema()))
tensor_adapter = tfxio.TensorAdapter()
tensors = tensor_adapter.ToBatchTensors(record_batch)
self.assertLen(tensors, 1)
self.assertIn(column_name, tensors)
with beam.Pipeline() as p:
record_batch_pcoll = (
p
| "CreateInMemRecords" >> beam.Create(_RAW_RECORDS)
|
"BeamSource" >> tfxio.BeamSource(batch_size=len(_RAW_RECORDS)))
beam_testing_util.assert_that(record_batch_pcoll, _AssertFn)
def assertGeneralMetricsComputedWithBeamAre(
self, eval_saved_model_path: str,
examples_pcollection: beam.pvalue.PCollection,
slice_spec: List[slicer.SingleSliceSpec],
add_metrics_callbacks: List[types.AddMetricsCallbackType],
expected_slice_metrics: Dict[Any, Dict[str, 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.assertCountEqual(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_pb2.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)
tfx_io = raw_tf_record.RawBeamRecordTFXIO(
physical_format='inmemory',
raw_record_column_name=constants.ARROW_INPUT_COLUMN,
telemetry_descriptors=['TFMATest'])
# pylint: disable=no-value-for-parameter
(metrics, _), _ = (
examples_pcollection
| '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)
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)
def _write_tfma(self,
tfma_path: str,
output_file_format: str,
store: Optional[mlmd.MetadataStore] = None):
_, eval_saved_model_path = (
fixed_prediction_estimator.simple_fixed_prediction_estimator(
export_path=None,
eval_export_path=os.path.join(self.tmpdir, 'eval_export_dir')))
eval_config = tfma.EvalConfig(model_specs=[tfma.ModelSpec()])
eval_shared_model = self.createTestEvalSharedModel(
eval_saved_model_path=eval_saved_model_path,
add_metrics_callbacks=[
tfma.post_export_metrics.example_count(),
tfma.post_export_metrics.
calibration_plot_and_prediction_histogram(num_buckets=2)
])
extractors = [
tfma.extractors.legacy_predict_extractor.PredictExtractor(
eval_shared_model, eval_config=eval_config),
tfma.extractors.unbatch_extractor.UnbatchExtractor(),
tfma.extractors.slice_key_extractor.SliceKeyExtractor()
]
evaluators = [
tfma.evaluators.legacy_metrics_and_plots_evaluator.
MetricsAndPlotsEvaluator(eval_shared_model)
]
writers = [
tfma.writers.MetricsPlotsAndValidationsWriter(
output_paths={
'metrics': os.path.join(tfma_path, 'metrics'),
'plots': os.path.join(tfma_path, 'plots')
},
output_file_format=output_file_format,
eval_config=eval_config,
add_metrics_callbacks=eval_shared_model.add_metrics_callbacks)
]
tfx_io = raw_tf_record.RawBeamRecordTFXIO(
physical_format='inmemory',
raw_record_column_name='__raw_record__',
telemetry_descriptors=['TFMATest'])
with beam.Pipeline() as pipeline:
example1 = self._makeExample(prediction=0.0, label=1.0)
example2 = self._makeExample(prediction=1.0, label=1.0)
_ = (pipeline
| 'Create' >> beam.Create([
example1.SerializeToString(),
example2.SerializeToString(),
])
| 'BatchExamples' >> tfx_io.BeamSource()
| 'ExtractEvaluateAndWriteResults' >>
tfma.ExtractEvaluateAndWriteResults(
eval_config=eval_config,
eval_shared_model=eval_shared_model,
extractors=extractors,
evaluators=evaluators,
writers=writers))
if store:
eval_type = metadata_store_pb2.ArtifactType()
eval_type.name = standard_artifacts.ModelEvaluation.TYPE_NAME
eval_type_id = store.put_artifact_type(eval_type)
artifact = metadata_store_pb2.Artifact()
artifact.uri = tfma_path
artifact.type_id = eval_type_id
store.put_artifacts([artifact])
