Transformer

def multihead_attention(query_antecedent,
                        memory_antecedent,
                        bias,
                        total_key_depth,
                        total_value_depth,
                        output_depth,
                        num_heads,
                        dropout_rate,
                        summaries=False,
                        image_shapes=None,
                        name=None):
  """Multihead scaled-dot-product attention with input/output transformations.

  Args:
    query_antecedent: a Tensor with shape [batch, length_q, channels]
    memory_antecedent: a Tensor with shape [batch, length_m, channels]
    bias: bias Tensor (see attention_bias())
    total_key_depth: an integer
    total_value_depth: an integer
    output_depth: an integer
    num_heads: an integer dividing total_key_depth and total_value_depth
    dropout_rate: a floating point number
    summaries: a boolean
    image_shapes: optional tuple of integer scalars.
      see comments for attention_image_summary()
    name: an optional string

  Returns:
    A Tensor.

  Raises:
    ValueError: if the key depth or value depth are not divisible by the
      number of attention heads.
  """
  if total_key_depth % num_heads != 0:
    raise ValueError("Key depth (%d) must be divisible by the number of "
                     "attention heads (%d)." % (total_key_depth, num_heads))
  if total_value_depth % num_heads != 0:
    raise ValueError("Value depth (%d) must be divisible by the number of "
                     "attention heads (%d)." % (total_value_depth, num_heads))

  with tf.variable_scope(
      name,
      default_name="multihead_attention",
      values=[query_antecedent, memory_antecedent]):
    if memory_antecedent is None:
      # self attention
      combined = common_layers.conv1d(
          query_antecedent,
          total_key_depth * 2 + total_value_depth,
          1,
          name="qkv_transform")
      q, k, v = tf.split(
          combined, [total_key_depth, total_key_depth, total_value_depth],
          axis=2)
    else:
      q = common_layers.conv1d(
          query_antecedent, total_key_depth, 1, name="q_transform")
      combined = common_layers.conv1d(
          memory_antecedent,
          total_key_depth + total_value_depth,
          1,
          name="kv_transform")
      k, v = tf.split(combined, [total_key_depth, total_value_depth], axis=2)
    q = split_heads(q, num_heads)
    k = split_heads(k, num_heads)
    v = split_heads(v, num_heads)
    key_depth_per_head = total_key_depth // num_heads
    q *= key_depth_per_head**-0.5
    x = dot_product_attention(
        q, k, v, bias, dropout_rate, summaries, image_shapes)
    x = combine_heads(x)
    x = common_layers.conv1d(x, output_depth, 1, name="output_transform")
    return x

def split_heads(x, num_heads):
  """Split channels (dimension 3) into multiple heads (becomes dimension 1).

  Args:
    x: a Tensor with shape [batch, length, channels]
    num_heads: an integer

  Returns:
    a Tensor with shape [batch, num_heads, length, channels / num_heads]
  """
  return tf.transpose(split_last_dimension(x, num_heads), [0, 2, 1, 3])


def combine_heads(x):
  """Inverse of split_heads.

  Args:
    x: a Tensor with shape [batch, num_heads, length, channels / num_heads]

  Returns:
    a Tensor with shape [batch, length, channels]
  """
  return combine_last_two_dimensions(tf.transpose(x, [0, 2, 1, 3]))