def token_noise(dataset, output_features, **unused_kwargs):
return preprocessors.denoise(
dataset,
output_features,
noise_density=0.15,
noise_mask_fn=preprocessors.iid_noise_mask,
inputs_fn=preprocessors.noise_token_to_sentinel,
targets_fn=None)
def test_denoise(self):
tf.set_random_seed(55)
vocab = test_utils.sentencepiece_vocab()
target_tokens = vocab.encode('The quick brown fox.')
# This is what it encodes to.
self.assertEqual(
target_tokens,
[3, 2, 20, 4, 3, 2, 8, 13, 2, 3, 2, 23, 7, 19, 22, 3, 2, 7, 2])
og_dataset = tf.data.Dataset.from_tensor_slices({
'targets': [target_tokens],
})
output_features = {
'targets': utils.Feature(vocab),
}
# These are the parameters of denoise in the operative config of 'base'.
# Except noise_density, bumped up from 0.15 to 0.3 in order to demonstrate
# multiple corrupted spans.
denoised_dataset = prep.denoise(
og_dataset,
output_features,
noise_density=0.3,
noise_mask_fn=prep.random_spans_noise_mask,
inputs_fn=prep.noise_span_to_unique_sentinel,
targets_fn=prep.nonnoise_span_to_unique_sentinel)
# Two spans corrupted, [2] and [22, 3, 2, 7, 2], replaced by unique
# sentinels 25 and 24 respectively.
assert_dataset(denoised_dataset, [
{
'inputs': [
3, 25, 20, 4, 3, 2, 8, 13, 2, 3, 2, 23, 7, 19, 24
],
'targets': [
25, 2, 24, 22, 3, 2, 7, 2
],
},
])
