def testTotalAttributionsWithSubKeys(self, sub_key, expected_values):
computations = attributions.TotalAttributions().computations(
sub_keys=[sub_key])
example1 = {
'attributions': {
'feature1': [1.11, 1.13, 1.12],
'feature2': [1.21, 1.23, 1.22]
}
}
example2 = {
'attributions': {
'feature1': [2.11, 2.13, 2.12],
'feature2': [2.21, 2.23, 2.22]
}
}
example3 = {
'attributions': {
'feature1': np.array([3.11, 3.13, 3.12]),
'feature2': np.array([3.21, 3.23, 3.22])
}
}
with beam.Pipeline() as pipeline:
# pylint: disable=no-value-for-parameter
result = (
pipeline
| 'Create' >> beam.Create([example1, example2, example3])
| 'Process' >> beam.ParDo(computations[0].preprocessor)
| 'AddSlice' >> beam.Map(lambda x: ((), x))
| 'CombineAttributions' >> beam.CombinePerKey(
computations[0].combiner)
| 'ComputeResult' >>
beam.Map( # comment to add lamda on own line
lambda x: (x[0], computations[1].result(x[1]))))
# pylint: enable=no-value-for-parameter
def check_result(got):
try:
self.assertLen(got, 1)
got_slice_key, got_attributions = got[0]
self.assertEqual(got_slice_key, ())
total_attributions_key = metric_types.AttributionsKey(
name='total_attributions', sub_key=sub_key)
self.assertIn(total_attributions_key, got_attributions)
self.assertAllClose(
got_attributions[total_attributions_key],
expected_values)
except AssertionError as err:
raise util.BeamAssertException(err)
util.assert_that(result, check_result, label='result')
def testHasAttributionsMetrics(self):
specs_with_attributions = metric_specs.specs_from_metrics({
'output_name': [
tf.keras.metrics.MeanSquaredError('mse'),
attributions.TotalAttributions()
]
})
self.assertTrue(
attributions.has_attributions_metrics(specs_with_attributions))
specs_without_attributions = metric_specs.specs_from_metrics([
tf.keras.metrics.MeanSquaredError('mse'),
])
self.assertFalse(
attributions.has_attributions_metrics(specs_without_attributions))
def testTotalAttributionsWithMultiModelsAndOutputs(self, model_name,
output_name, examples,
expected_values):
computations = attributions.TotalAttributions().computations(
model_names=[model_name], output_names=[output_name])
with beam.Pipeline() as pipeline:
# pylint: disable=no-value-for-parameter
result = (
pipeline
| 'Create' >> beam.Create(examples)
| 'Process' >> beam.ParDo(computations[0].preprocessor)
| 'AddSlice' >> beam.Map(lambda x: ((), x))
| 'CombineAttributions' >> beam.CombinePerKey(
computations[0].combiner)
| 'ComputeResult' >>
beam.Map( # comment to add lamda on own line
lambda x: (x[0], computations[1].result(x[1]))))
# pylint: enable=no-value-for-parameter
def check_result(got):
try:
self.assertLen(got, 1)
got_slice_key, got_attributions = got[0]
self.assertEqual(got_slice_key, ())
total_attributions_key = metric_types.AttributionsKey(
name='total_attributions',
model_name=model_name,
output_name=output_name)
self.assertIn(total_attributions_key, got_attributions)
self.assertDictElementsAlmostEqual(
got_attributions[total_attributions_key],
expected_values)
except AssertionError as err:
raise util.BeamAssertException(err)
util.assert_that(result, check_result, label='result')