def make_ngram_labels(label_start_idx: tf.Tensor,
label_phrase_len: tf.Tensor,
long_max_length: int,
kp_max_length: int,
additive_smoothing_mass: float = 1e-6) -> tf.Tensor:
"""Makes ngram labels for `tf.nn.softmax_cross_entropy_with_logits`.
Args:
label_start_idx: <int32>[batch_size, num_labels] Tensor of the index of the
first word in each correct key phrase. There must be at least 1 correct
key phrase, and if there are less than `num_labels` then `-1` is used as
right padding. All values must be less than `long_max_length`.
`num_labels` is the maximum number of key phrase labels, which is 3 for
OpenKP.
label_phrase_len: <int32>[batch_size, num_labels] Tensor of the
corresponding length of the key phrase, again using `-1` to pad.
Non-padding values must be in the inclusive range [1, kp_max_length].
long_max_length: Integer maximum number of words in the document.
kp_max_length: Integer maximum number of words in a key phrase.
additive_smoothing_mass: Total probability mass (on top of `1.0` mass for
the actual label) to add for additive smoothing. We use a minimum of 1e-6
to avoid any potential division by 0.
Returns:
<float32>[batch_size, kp_max_length * long_max_length] Tensor of label
probabilities based on the inputs. Each row sums to 1.0, and the order
of entries is compatible with reshaping `ngram_logits` from shape
[batch_size, kp_max_length, long_max_length] to match these labels.
"""
# [batch_size, num_labels, kp_max_length]
phrase_len_one_hot = tf.one_hot(label_phrase_len - 1,
depth=kp_max_length,
dtype=tf.float32)
# [batch_size, num_labels, long_max_length]
start_idx_one_hot = tf.one_hot(label_start_idx,
depth=long_max_length,
dtype=tf.float32)
# [batch_size, kp_max_length, long_max_length]
combined_one_hot = tf.einsum('bnk,bnl->bkl', phrase_len_one_hot,
start_idx_one_hot)
# [batch_size, kp_max_length * long_max_length]
unnormalized_labels = tensor_utils.flatten_dims(combined_one_hot,
first_dim=1)
unsmoothed_labels = (
unnormalized_labels /
(tf.reduce_sum(unnormalized_labels, axis=-1, keepdims=True) + 1e-6))
# Use at least 1e-6 smoothing mass to avoid divide by 0.
additive_smoothing_mass = max(additive_smoothing_mass, 1e-6)
num_classes = kp_max_length * long_max_length
smoothed_labels = unsmoothed_labels + additive_smoothing_mass / num_classes
return (smoothed_labels /
tf.reduce_sum(smoothed_labels, axis=-1, keepdims=True))
def test_flatten_dims(self):
tensor = tf.reshape(tf.range(2 * 3 * 4 * 5), [2, 3, 4, 5])
result1 = tensor_utils.flatten_dims(tensor, last_dim=1)
self.assertAllEqual([6, 4, 5], result1.shape)
self.assertAllEqual(tf.range(2 * 3 * 4 * 5), tf.reshape(result1, [-1]))
self.assertAllEqual([2, 3, 20],
tensor_utils.flatten_dims(tensor,
first_dim=-2).shape)
self.assertAllEqual([2, 12, 5],
tensor_utils.flatten_dims(tensor,
first_dim=1,
last_dim=-2).shape)
self.assertAllEqual([24, 5],
tensor_utils.flatten_dims(tensor,
last_dim=-2).shape)
self.assertAllEqual([2 * 3 * 4 * 5],
tensor_utils.flatten_dims(tensor).shape)
self.assertAllEqual([2, 3, 4, 5],
tensor_utils.flatten_dims(tensor,
first_dim=1,
last_dim=-3).shape)
self.assertAllEqual([7], tensor_utils.flatten_dims(tf.ones([7])).shape)
self.assertAllEqual([12],
tensor_utils.flatten_dims(tf.ones([4, 3])).shape)
with self.assertRaises(ValueError):
tensor_utils.flatten_dims(tensor, first_dim=4)
with self.assertRaises(ValueError):
tensor_utils.flatten_dims(tensor, last_dim=-5)
with self.assertRaises(ValueError):
tensor_utils.flatten_dims(tensor, first_dim=2, last_dim=1)
def make_local_segmented_att_mask(segment_ids: tf.Tensor,
local_radius: int,
name: Optional[Text] = None) -> tf.Tensor:
"""Makes local attention mask preventing attention across different segments.
Restricts local self-attention to attend within segments, such that tokens can
only attend to local tokens from the same segment id. The tokens in a segment
do not need to be contiguous, but attention is still constrained by
`local_radius`. The output can be used as `l2l_att_mask` in
`layers.GlobalLocalTransformerLayers` for example.
Args:
segment_ids: <int32>[batch_size, seq_len] Tensor of segment ids, all of
which must be non-negative.
local_radius: The local radius as expected by
`layers.GlobalLocalTransformerLayers`. Must be positive.
name: A name for the operation (optional).
Returns:
<int32>[batch_size, seq_len, 2*local_radius + 1] attention mask.
"""
with tf.name_scope(name or 'make_local_segmented_att_mask'):
segment_ids = tf.convert_to_tensor(segment_ids)
if segment_ids.shape.rank != 2:
raise ValueError('`segment_ids` must be a 2-D tensor.')
batch_size, seq_len = tensor_utils.get_shape_list(segment_ids)
# Add 1 so that segment id `0` doesn't coincide with `0` padding values
# introduced later by `tensor_utils.concat_3_blocks()` for example.
segment_ids += 1
# [batch_size, num_blocks, local_radius]
blocked_segment_ids = tensor_utils.split_into_blocks(
segment_ids, block_len=local_radius, axis=1)
# [batch_size, num_blocks, 3*local_radius]
concat_blocked_segment_ids = tensor_utils.concat_3_blocks(
blocked_segment_ids)
# [batch_size, num_blocks, local_radius, 3*local_radius]
tiled_segment_ids = tf.tile(
concat_blocked_segment_ids[:, :, tf.newaxis, :],
[1, 1, local_radius, 1])
# [batch_size, num_blocks, local_radius, 2*local_radius + 1]
blocked_unskewed_segment_ids = tensor_utils.unskew_elements_right(
tiled_segment_ids, axis=-1)
# [batch_size, num_blocks * local_radius, 2*local_radius + 1]
flat_unskewed_segment_ids = tensor_utils.flatten_dims(
blocked_unskewed_segment_ids, first_dim=1, last_dim=2)
# [batch_size, seq_len, 2*local_radius + 1]
unskewed_segment_ids = tf.slice(flat_unskewed_segment_ids,
begin=[0, 0, 0],
size=[-1, seq_len, -1])
# [batch_size, seq_len, 1]
center_token_segment_id = unskewed_segment_ids[:, :, local_radius:(
local_radius + 1)]
# [batch_size, seq_len, 2*local_radius + 1]
result = tf.cast(
tf.equal(unskewed_segment_ids, center_token_segment_id), tf.int32)
# Use `reshape` to set the static shape when known.
return tf.reshape(result, [batch_size, seq_len, 2 * local_radius + 1])
