def test_make_ngram_labels_additive_smoothing(self):
label_start_idx = tf.constant([
[1, -1, -1],
[2, 3, 0],
])
label_phrase_len = tf.constant([
[3, -1, -1],
[2, 1, 2],
])
long_max_length = 4
kp_max_length = 5
additive_smoothing_mass = 1.0
smoothed_third = (1 / 3) + 0.05
reshaped_unnormalized_expected = [
[
[0.05, 0.05, 0.05, 0.05], # 1-grams
[0.05, 0.05, 0.05, 0.05], # 2-grams
[0.05, 1.05, 0.05, 0.05], # 3-grams
[0.05, 0.05, 0.05, 0.05], # 4-grams
[0.05, 0.05, 0.05, 0.05], # 5-grams
],
[
[0.05, 0.05, 0.05, smoothed_third], # 1-grams
[smoothed_third, 0.05, smoothed_third, 0.05], # 2-grams
[0.05, 0.05, 0.05, 0.05], # 3-grams
[0.05, 0.05, 0.05, 0.05], # 4-grams
[0.05, 0.05, 0.05, 0.05], # 5-grams
],
]
batch_size = len(reshaped_unnormalized_expected)
unnormalized_expected = tf.reshape(
reshaped_unnormalized_expected,
[batch_size, kp_max_length * long_max_length])
expected = (
unnormalized_expected /
tf.reduce_sum(unnormalized_expected, axis=-1, keepdims=True))
self.assertAllClose(
expected,
run_finetuning_lib.make_ngram_labels(
label_start_idx=label_start_idx,
label_phrase_len=label_phrase_len,
long_max_length=long_max_length,
kp_max_length=kp_max_length,
additive_smoothing_mass=additive_smoothing_mass))
def test_make_ngram_labels(self):
label_start_idx = tf.constant([
[1, -1, -1],
[2, 3, 0],
])
label_phrase_len = tf.constant([
[3, -1, -1],
[2, 1, 2],
])
long_max_length = 4
kp_max_length = 5
reshaped_expected = [
[
[0, 0, 0, 0], # 1-grams
[0, 0, 0, 0], # 2-grams
[0, 1, 0, 0], # 3-grams
[0, 0, 0, 0], # 4-grams
[0, 0, 0, 0], # 5-grams
],
[
[0, 0, 0, 1 / 3], # 1-grams
[1 / 3, 0, 1 / 3, 0], # 2-grams
[0, 0, 0, 0], # 3-grams
[0, 0, 0, 0], # 4-grams
[0, 0, 0, 0], # 5-grams
],
]
batch_size = len(reshaped_expected)
expected = tf.reshape(reshaped_expected,
[batch_size, kp_max_length * long_max_length])
self.assertAllClose(
expected,
run_finetuning_lib.make_ngram_labels(
label_start_idx=label_start_idx,
label_phrase_len=label_phrase_len,
long_max_length=long_max_length,
kp_max_length=kp_max_length))