def testSerializeDeserializeToFile(self):
metrics_slice_key1 = (('fruit', 'pear'), ('animal', 'duck'))
metrics1 = {
'alpha': np.array([0.1]),
'bravo': np.array([0.2]),
'charlie': np.float32(0.3)
}
expected_metrics1 = {
'alpha': [0.1],
'bravo': [0.2],
'charlie': 0.3,
}
plots_slice_key1 = (('fruit', 'peach'), ('animal', 'cow'))
plots1 = {
'alpha': np.array([0.5, 0.6, 0.7]),
'bravo': np.array([0.6, 0.7, 0.8]),
'charlie': np.float32(0.7)
}
expected_plots1 = {
'alpha': [0.5, 0.6, 0.7],
'bravo': [0.6, 0.7, 0.8],
'charlie': 0.7,
}
eval_config = api_types.EvalConfig(
model_location='/path/to/model',
data_location='/path/to/data',
slice_spec=[
slicer.SingleSliceSpec(features=[('age', 5), ('gender', 'f')],
columns=['country']),
slicer.SingleSliceSpec(features=[('age', 6), ('gender', 'm')],
columns=['interest'])
],
example_weight_metric_key='key')
output_path = self._getTempDir()
with beam.Pipeline() as pipeline:
metrics = (pipeline
| 'CreateMetrics' >> beam.Create(
[(metrics_slice_key1, metrics1)]))
plots = (
pipeline
| 'CreatePlots' >> beam.Create([(plots_slice_key1, plots1)]))
_ = ((metrics, plots)
| 'WriteMetricsPlotsAndConfig' >>
serialization.WriteMetricsPlotsAndConfig(
output_path=output_path, eval_config=eval_config))
metrics, plots = serialization.load_plots_and_metrics(output_path)
self.assertSliceMetricsListEqual(
[(metrics_slice_key1, expected_metrics1)], metrics)
self.assertSliceMetricsListEqual([(plots_slice_key1, expected_plots1)],
plots)
got_eval_config = serialization.load_eval_config(output_path)
self.assertEqual(eval_config, got_eval_config)
def EvaluateAndWriteResults( # pylint: disable=invalid-name
examples,
eval_saved_model_path,
output_path,
display_only_data_location = None,
slice_spec = None,
example_weight_key = None,
add_metrics_callbacks = None, # pylint: disable=bad-whitespace
desired_batch_size = None,
):
"""Public API version of evaluate.Evaluate that handles example weights.
Users who want to construct their own Beam pipelines instead of using the
lightweight run_model_analysis functions should use this PTransform.
Example usage:
with beam.Pipeline(runner=...) as p:
_ = (p
| 'ReadData' >> beam.io.ReadFromTFRecord(data_location)
| 'EvaluateAndWriteResults' >> tfma.EvaluateAndWriteResults(
eval_saved_model_path=model_location,
output_path=output_path,
display_only_data_location=data_location,
slice_spec=slice_spec,
example_weight_key=example_weight_key,
...))
result = tfma.load_eval_result(output_path=output_path)
tfma.view.render_slicing_metrics(result)
Note that the exact serialization format is an internal implementation detail
and subject to change. Users should only use the TFMA functions to write and
read the results.
Args:
examples: PCollection of input examples. Can be any format the model accepts
(e.g. string containing CSV row, TensorFlow.Example, etc).
eval_saved_model_path: Path to EvalSavedModel. This directory should contain
the saved_model.pb file.
output_path: Path to output metrics and plots results.
display_only_data_location: Optional path indicating where the examples
were read from. This is used only for display purposes - data will not
actually be read from this path.
slice_spec: Optional list of SingleSliceSpec specifying the slices to slice
the data into. If None, defaults to the overall slice.
example_weight_key: The key of the example weight column. If None, weight
will be 1 for each example.
add_metrics_callbacks: Optional list of callbacks for adding additional
metrics to the graph. The names of the metrics added by the callbacks
should not conflict with existing metrics, or metrics added by other
callbacks. See below for more details about what each callback should do.
desired_batch_size: Optional batch size for batching in Predict and
Aggregate.
Returns:
PDone.
"""
if add_metrics_callbacks is None:
add_metrics_callbacks = []
# Always compute example weight and example count.
# pytype: disable=module-attr
example_count_callback = post_export_metrics.example_count()
example_weight_metric_key = metric_keys.EXAMPLE_COUNT
add_metrics_callbacks.append(example_count_callback)
if example_weight_key:
example_weight_metric_key = metric_keys.EXAMPLE_WEIGHT
example_weight_callback = post_export_metrics.example_weight(
example_weight_key)
add_metrics_callbacks.append(example_weight_callback)
# pytype: enable=module-attr
metrics, plots = examples | 'Evaluate' >> evaluate.Evaluate(
eval_saved_model_path=eval_saved_model_path,
add_metrics_callbacks=add_metrics_callbacks,
slice_spec=slice_spec,
desired_batch_size=desired_batch_size)
data_location = '<user provided PCollection>'
if display_only_data_location is not None:
data_location = display_only_data_location
eval_config = api_types.EvalConfig(
model_location=eval_saved_model_path,
data_location=data_location,
slice_spec=slice_spec,
example_weight_metric_key=example_weight_metric_key)
_ = ((metrics, plots)
| 'SerializeMetricsAndPlots' >> serialization.SerializeMetricsAndPlots(
post_export_metrics=add_metrics_callbacks)
|
'WriteMetricsPlotsAndConfig' >> serialization.WriteMetricsPlotsAndConfig(
output_path=output_path, eval_config=eval_config))
return beam.pvalue.PDone(examples.pipeline)
def testSerializeDeserializeToFile(self):
metrics_slice_key = _make_slice_key('fruit', 'pear', 'animal', 'duck')
metrics_for_slice = text_format.Parse(
"""
slice_key {
single_slice_keys {
column: "fruit"
bytes_value: "pear"
}
single_slice_keys {
column: "animal"
bytes_value: "duck"
}
}
metrics {
key: "accuracy"
value {
double_value {
value: 0.8
}
}
}
metrics {
key: "example_weight"
value {
double_value {
value: 10.0
}
}
}
metrics {
key: "auc"
value {
bounded_value {
lower_bound {
value: 0.1
}
upper_bound {
value: 0.3
}
value {
value: 0.2
}
}
}
}
metrics {
key: "auprc"
value {
bounded_value {
lower_bound {
value: 0.05
}
upper_bound {
value: 0.17
}
value {
value: 0.1
}
}
}
}""", metrics_for_slice_pb2.MetricsForSlice())
plots_for_slice = text_format.Parse(
"""
slice_key {
single_slice_keys {
column: "fruit"
bytes_value: "peach"
}
single_slice_keys {
column: "animal"
bytes_value: "cow"
}
}
plot_data {
calibration_histogram_buckets {
buckets {
lower_threshold_inclusive: -inf
upper_threshold_exclusive: 0.0
num_weighted_examples { value: 0.0 }
total_weighted_label { value: 0.0 }
total_weighted_refined_prediction { value: 0.0 }
}
buckets {
lower_threshold_inclusive: 0.0
upper_threshold_exclusive: 0.5
num_weighted_examples { value: 1.0 }
total_weighted_label { value: 1.0 }
total_weighted_refined_prediction { value: 0.3 }
}
buckets {
lower_threshold_inclusive: 0.5
upper_threshold_exclusive: 1.0
num_weighted_examples { value: 1.0 }
total_weighted_label { value: 0.0 }
total_weighted_refined_prediction { value: 0.7 }
}
buckets {
lower_threshold_inclusive: 1.0
upper_threshold_exclusive: inf
num_weighted_examples { value: 0.0 }
total_weighted_label { value: 0.0 }
total_weighted_refined_prediction { value: 0.0 }
}
}
}""", metrics_for_slice_pb2.PlotsForSlice())
plots_slice_key = _make_slice_key('fruit', 'peach', 'animal', 'cow')
eval_config = api_types.EvalConfig(
model_location='/path/to/model',
data_location='/path/to/data',
slice_spec=[
slicer.SingleSliceSpec(features=[('age', 5), ('gender', 'f')],
columns=['country']),
slicer.SingleSliceSpec(features=[('age', 6), ('gender', 'm')],
columns=['interest'])
],
example_weight_metric_key='key')
output_path = self._getTempDir()
with beam.Pipeline() as pipeline:
metrics = (pipeline
| 'CreateMetrics' >> beam.Create(
[metrics_for_slice.SerializeToString()]))
plots = (pipeline
| 'CreatePlots' >> beam.Create(
[plots_for_slice.SerializeToString()]))
_ = ((metrics, plots)
| 'WriteMetricsPlotsAndConfig' >>
serialization.WriteMetricsPlotsAndConfig(
output_path=output_path, eval_config=eval_config))
metrics, plots = serialization.load_plots_and_metrics(output_path)
self.assertSliceMetricsListEqual(
[(metrics_slice_key, metrics_for_slice.metrics)], metrics)
self.assertSlicePlotsListEqual(
[(plots_slice_key, plots_for_slice.plot_data)], plots)
got_eval_config = serialization.load_eval_config(output_path)
self.assertEqual(eval_config, got_eval_config)
def ExtractEvaluateAndWriteResults( # pylint: disable=invalid-name
examples,
eval_shared_model,
output_path,
display_only_data_location=None,
slice_spec=None,
desired_batch_size=None,
extractors=None,
fanout=16,
):
"""Public API version of evaluate.Evaluate that handles example weights.
Users who want to construct their own Beam pipelines instead of using the
lightweight run_model_analysis functions should use this PTransform.
Example usage:
eval_shared_model = tfma.default_eval_shared_model(
eval_saved_model_path=model_location,
add_metrics_callbacks=[...],
example_weight_key=example_weight_key)
with beam.Pipeline(runner=...) as p:
_ = (p
| 'ReadData' >> beam.io.ReadFromTFRecord(data_location)
| 'ExtractEvaluateAndWriteResults' >>
tfma.ExtractEvaluateAndWriteResults(
eval_shared_model=eval_shared_model,
output_path=output_path,
display_only_data_location=data_location,
slice_spec=slice_spec,
...))
result = tfma.load_eval_result(output_path=output_path)
tfma.view.render_slicing_metrics(result)
Note that the exact serialization format is an internal implementation detail
and subject to change. Users should only use the TFMA functions to write and
read the results.
Args:
examples: PCollection of input examples. Can be any format the model accepts
(e.g. string containing CSV row, TensorFlow.Example, etc).
eval_shared_model: Shared model parameters for EvalSavedModel including any
additional metrics (see EvalSharedModel for more information on how to
configure additional metrics).
output_path: Path to output metrics and plots results.
display_only_data_location: Optional path indicating where the examples were
read from. This is used only for display purposes - data will not actually
be read from this path.
slice_spec: Optional list of SingleSliceSpec specifying the slices to slice
the data into. If None, defaults to the overall slice.
desired_batch_size: Optional batch size for batching in Predict and
Aggregate.
extractors: Optional list of Extractors to apply to ExampleAndExtracts. If
provided, the extracts MUST contain a FeaturesPredictionsLabels extract
with key 'fpl' and a list of SliceKeyType extracts with key 'slice_keys'.
Typically these will be added by calling the default_extractors function.
If no extractors are provided, default_extractors (non-materialized) will
be used.
Raises:
ValueError: If PredictExtractor or SliceKeyExtractor is not present in
extractors.
Returns:
PDone.
"""
if not extractors:
extractors = default_extractors(eval_shared_model=eval_shared_model,
slice_spec=slice_spec,
desired_batch_size=desired_batch_size,
materialize=False)
metrics, plots = (
examples
| 'ToExampleAndExtracts' >> evaluate.ToExampleAndExtracts()
| 'Extract' >> evaluate.Extract(extractors=extractors)
|
'Evaluate' >> evaluate.Evaluate(eval_shared_model=eval_shared_model,
desired_batch_size=desired_batch_size,
fanout=fanout))
data_location = '<user provided PCollection>'
if display_only_data_location is not None:
data_location = display_only_data_location
example_weight_metric_key = metric_keys.EXAMPLE_COUNT
if eval_shared_model.example_weight_key:
example_weight_metric_key = metric_keys.EXAMPLE_WEIGHT
eval_config = api_types.EvalConfig(
model_location=eval_shared_model.model_path,
data_location=data_location,
slice_spec=slice_spec,
example_weight_metric_key=example_weight_metric_key)
_ = ((metrics, plots)
|
'SerializeMetricsAndPlots' >> serialization.SerializeMetricsAndPlots(
post_export_metrics=eval_shared_model.add_metrics_callbacks)
| 'WriteMetricsPlotsAndConfig' >>
serialization.WriteMetricsPlotsAndConfig(output_path=output_path,
eval_config=eval_config))
return beam.pvalue.PDone(examples.pipeline)