def metrics_builder():
return [
keras_metrics.NumTokensCounter(masked_tokens=[pad_token]),
keras_metrics.MaskedCategoricalAccuracy(name='accuracy',
masked_tokens=[pad_token]),
keras_metrics.MaskedCategoricalAccuracy(
name='accuracy_without_out_of_vocab',
masked_tokens=[pad_token] + oov_tokens),
# Notice that the beginning of sentence token never appears in the
# ground truth label.
keras_metrics.MaskedCategoricalAccuracy(
name='accuracy_without_out_of_vocab_or_end_of_sentence',
masked_tokens=[pad_token, eos_token] + oov_tokens),
]
def test_update_state_with_special_character(self):
metric = keras_metrics.MaskedCategoricalAccuracy(masked_tokens=[4])
metric.update_state(
y_true=[[1, 2, 3, 4], [0, 0, 0, 0]],
y_pred=[
# A batch with 100% accruacy.
[
[0.1, 0.9, 0.1, 0.1, 0.1],
[0.1, 0.1, 0.9, 0.1, 0.1],
[0.1, 0.1, 0.1, 0.9, 0.1],
[0.1, 0.1, 0.1, 0.1, 0.9],
],
# A batch with 50% accruacy.
[
[0.1, 0.9, 0.1, 0.1, 0.1],
[0.9, 0.1, 0.1, 0.1, 0.1],
[0.1, 0.1, 0.1, 0.9, 0.1],
[0.9, 0.1, 0.1, 0.1, 0.0],
],
])
self.assertAllClose(self.evaluate(metric.result()), 5 / 7.0)
metric.update_state(
y_true=[[0, 4, 1, 2]],
y_pred=[
# A batch with 33% accruacy.
[
[0.9, 0.1, 0.1, 0.1, 0.1],
[0.1, 0.1, 0.1, 0.1, 0.9],
[0.1, 0.1, 0.1, 0.9, 0.1],
[0.1, 0.1, 0.1, 0.1, 0.9],
],
])
self.assertAllClose(self.evaluate(metric.result()), 6 / 10.0)
def test_constructor_no_masked_token(self):
metric_name = 'my_test_metric'
metric = keras_metrics.MaskedCategoricalAccuracy(name=metric_name)
self.assertIsInstance(metric, tf.keras.metrics.Metric)
self.assertEqual(metric.name, metric_name)
self.assertAllEqual(metric.get_config()['masked_tokens'], [])
self.assertEqual(self.evaluate(metric.result()), 0.0)
def test_update_state_with_no_special_character(self):
metric = keras_metrics.MaskedCategoricalAccuracy()
metric.update_state(
y_true=[[1, 2, 3, 4], [0, 0, 0, 0]],
y_pred=[
# A batch with 100% accruacy.
[
[0.1, 0.9, 0.1, 0.1, 0.1],
[0.1, 0.1, 0.9, 0.1, 0.1],
[0.1, 0.1, 0.1, 0.9, 0.1],
[0.1, 0.1, 0.1, 0.1, 0.9],
],
# A batch with 50% accruacy.
[
[0.1, 0.9, 0.1, 0.1, 0.1],
[0.9, 0.1, 0.1, 0.1, 0.1],
[0.1, 0.1, 0.1, 0.9, 0.1],
[0.9, 0.1, 0.1, 0.1, 0.0],
],
])
self.assertEqual(self.evaluate(metric.result()), 6 / 8.0)
metric.update_state(
y_true=[[0, 4, 1, 2]],
y_pred=[
# A batch with 25% accruacy.
[
[0.9, 0.1, 0.1, 0.1, 0.1],
[0.1, 0.1, 0.1, 0.1, 0.9],
[0.1, 0.1, 0.1, 0.9, 0.1],
[0.1, 0.1, 0.1, 0.1, 0.9],
],
])
self.assertAllClose(self.evaluate(metric.result()), 8 / 12.0)
def model_fn() -> model.Model:
return keras_utils.from_keras_model(
keras_model=char_prediction_models.create_recurrent_model(
vocab_size=VOCAB_LENGTH, sequence_length=sequence_length),
loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
input_spec=task_datasets.element_type_structure,
metrics=[
keras_metrics.NumTokensCounter(masked_tokens=[pad_token]),
keras_metrics.MaskedCategoricalAccuracy(masked_tokens=[pad_token])
])
def test_weighted_update_state_with_scalar_weight(self):
metric = keras_metrics.MaskedCategoricalAccuracy()
metric.update_state(
y_true=[[1, 2, 3, 4]],
y_pred=[
# A batch with 50% accuracy.
[
[0.9, 0.1, 0.1, 0.1, 0.1],
[0.1, 0.9, 0.1, 0.1, 0.1],
[0.1, 0.1, 0.1, 0.9, 0.1],
[0.1, 0.1, 0.1, 0.1, 0.9],
],
],
sample_weight=1.0)
self.assertAllClose(self.evaluate(metric.result()), .5)
def test_update_state_with_all_tokens_masked(self):
metric = keras_metrics.MaskedCategoricalAccuracy(
masked_tokens=[1, 2, 3, 4])
metric.update_state(
# All batches should be masked.
y_true=[[1, 2, 3, 4], [4, 3, 2, 1]],
y_pred=[
[
[0.1, 0.9, 0.1, 0.1, 0.1],
[0.1, 0.1, 0.9, 0.1, 0.1],
[0.1, 0.1, 0.1, 0.9, 0.1],
[0.1, 0.1, 0.1, 0.1, 0.9],
],
[
[0.1, 0.9, 0.1, 0.1, 0.1],
[0.9, 0.1, 0.1, 0.1, 0.1],
[0.1, 0.1, 0.1, 0.9, 0.1],
[0.9, 0.1, 0.1, 0.1, 0.0],
],
])
self.assertAllClose(self.evaluate(metric.result()), 0.0)
def test_update_state_with_multiple_tokens_masked(self):
metric = keras_metrics.MaskedCategoricalAccuracy(
masked_tokens=[1, 2, 3, 4])
metric.update_state(
y_true=[[1, 2, 3, 4], [0, 0, 0, 0]],
y_pred=[
[
# This batch should be masked.
[0.1, 0.9, 0.1, 0.1, 0.1],
[0.1, 0.1, 0.9, 0.1, 0.1],
[0.1, 0.1, 0.1, 0.9, 0.1],
[0.1, 0.1, 0.1, 0.1, 0.9],
],
[
# Batch with 50% accuracy
[0.1, 0.9, 0.1, 0.1, 0.1],
[0.9, 0.1, 0.1, 0.1, 0.1],
[0.1, 0.1, 0.1, 0.9, 0.1],
[0.9, 0.1, 0.1, 0.1, 0.0],
],
])
self.assertAllClose(self.evaluate(metric.result()), 0.5)
def test_weighted_update_state_special_character_rank_2_sample_weight(
self):
metric = keras_metrics.MaskedCategoricalAccuracy(masked_tokens=[4])
metric.update_state(
y_true=[[1, 2, 3, 4], [0, 0, 0, 0]],
y_pred=[
# A batch with 100% accuracy.
[
[0.1, 0.9, 0.1, 0.1, 0.1],
[0.1, 0.1, 0.9, 0.1, 0.1],
[0.1, 0.1, 0.1, 0.9, 0.1],
[0.1, 0.1, 0.1, 0.1, 0.9],
],
# A batch with 50% accuracy.
[
[0.1, 0.9, 0.1, 0.1, 0.1],
[0.9, 0.1, 0.1, 0.1, 0.1],
[0.1, 0.1, 0.1, 0.9, 0.1],
[0.9, 0.1, 0.1, 0.1, 0.0],
],
],
# A weight for each `y_true` scalar.
sample_weight=[[1.0, 2.0, 1.0, 2.0], [1.0, 2.0, 1.0, 2.0]])
self.assertAllClose(self.evaluate(metric.result()), (6 + 2) / 10.0)
def test_weighted_update_state_no_special_character(self):
metric = keras_metrics.MaskedCategoricalAccuracy()
metric.update_state(
y_true=[[1, 2, 3, 4], [0, 0, 0, 0]],
y_pred=[
# A batch with 100% accuracy.
[
[0.1, 0.9, 0.1, 0.1, 0.1],
[0.1, 0.1, 0.9, 0.1, 0.1],
[0.1, 0.1, 0.1, 0.9, 0.1],
[0.1, 0.1, 0.1, 0.1, 0.9],
],
# A batch with 50% accuracy.
[
[0.1, 0.9, 0.1, 0.1, 0.1],
[0.9, 0.1, 0.1, 0.1, 0.1],
[0.1, 0.1, 0.1, 0.9, 0.1],
[0.9, 0.1, 0.1, 0.1, 0.0],
],
],
# A weight for each `y_true` scalar.
sample_weight=[1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0])
self.assertAllClose(self.evaluate(metric.result()), (6 + 4) / 12.0)
metric.update_state(
y_true=[[0, 4, 1, 2]],
y_pred=[
# A batch with 25% accruacy.
[
[0.9, 0.1, 0.1, 0.1, 0.1],
[0.1, 0.1, 0.1, 0.1, 0.9],
[0.1, 0.1, 0.1, 0.9, 0.1],
[0.1, 0.1, 0.1, 0.1, 0.9],
],
],
sample_weight=[1.0, 1.0, 2.0, 2.0])
self.assertAllClose(self.evaluate(metric.result()), (6 + 4 + 2) / 18.0)
def test_weighted_update_state_with_masked_token(self):
metric = keras_metrics.MaskedCategoricalAccuracy(masked_tokens=[4])
metric.update_state(
y_true=[[1, 2, 3, 4], [0, 0, 0, 0]],
y_pred=[
# A batch with 100% accuracy.
[
[0.1, 0.9, 0.1, 0.1, 0.1],
[0.1, 0.1, 0.9, 0.1, 0.1],
[0.1, 0.1, 0.1, 0.9, 0.1],
[0.1, 0.1, 0.1, 0.1, 0.9],
],
# A batch with 50% accuracy.
[
[0.1, 0.9, 0.1, 0.1, 0.1],
[0.9, 0.1, 0.1, 0.1, 0.1],
[0.1, 0.1, 0.1, 0.9, 0.1],
[0.9, 0.1, 0.1, 0.1, 0.0],
],
],
# A weight for each `y_true` scalar.
sample_weight=[[1.0, 2.0, 1.0, 2.0], [1.0, 2.0, 1.0, 2.0]])
self.assertAllClose(self.evaluate(metric.result()), (4 + 4) / 10.0)
metric.update_state(
y_true=[[0, 4, 1, 2]],
y_pred=[
# A batch with 25% accruacy.
[
[0.9, 0.1, 0.1, 0.1, 0.1],
[0.1, 0.1, 0.1, 0.1, 0.9],
[0.1, 0.1, 0.1, 0.9, 0.1],
[0.1, 0.1, 0.1, 0.1, 0.9],
],
],
sample_weight=[1.0, 1.0, 2.0, 2.0])
self.assertAllClose(self.evaluate(metric.result()), (4 + 4 + 1) / 15.0)