Transformer

def add_timing_signal_1d(x, min_timescale=1.0, max_timescale=1.0e4):
  """Adds a bunch of sinusoids of different frequencies to a Tensor.

  Each channel of the input Tensor is incremented by a sinusoid of a different
  frequency and phase.

  This allows attention to learn to use absolute and relative positions.
  Timing signals should be added to some precursors of both the query and the
  memory inputs to attention.

  The use of relative position is possible because sin(x+y) and cos(x+y) can be
  experessed in terms of y, sin(x) and cos(x).

  In particular, we use a geometric sequence of timescales starting with
  min_timescale and ending with max_timescale.  The number of different
  timescales is equal to channels / 2. For each timescale, we
  generate the two sinusoidal signals sin(timestep/timescale) and
  cos(timestep/timescale).  All of these sinusoids are concatenated in
  the channels dimension.

  Args:
    x: a Tensor with shape [batch, length, channels]
    min_timescale: a float
    max_timescale: a float

  Returns:
    a Tensor the same shape as x.
  """
  length = tf.shape(x)[1]
  channels = tf.shape(x)[2]
  position = tf.to_float(tf.range(length))
  num_timescales = channels // 2
  log_timescale_increment = (
      math.log(float(max_timescale) / float(min_timescale)) /
      (tf.to_float(num_timescales) - 1))
  inv_timescales = min_timescale * tf.exp(
      tf.to_float(tf.range(num_timescales)) * -log_timescale_increment)
  scaled_time = tf.expand_dims(position, 1) * tf.expand_dims(inv_timescales, 0)
  signal = tf.concat([tf.sin(scaled_time), tf.cos(scaled_time)], axis=1)
  signal = tf.pad(signal, [[0, 0], [0, tf.mod(channels, 2)]])
  signal = tf.reshape(signal, [1, length, channels])
  return x + signal

def transformer_prepare_decoder(targets, hparams):
  """Prepare one shard of the model for the decoder.

  Args:
    targets: a Tensor.
    hparams: run hyperparameters

  Returns:
    decoder_input: a Tensor, bottom of decoder stack
    decoder_self_attention_bias: a Tensor, containing large negative values
    to implement masked attention and possibly baises for diagonal alignments
  """
  decoder_self_attention_bias = (
      common_attention.attention_bias_lower_triangle(tf.shape(targets)[1]))
  decoder_input = common_layers.shift_left_3d(targets)
  if hparams.pos == "timing":
    decoder_input = common_attention.add_timing_signal_1d(decoder_input)
  return (decoder_input, decoder_self_attention_bias)

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 transformer_decoder(decoder_input,
                        encoder_output,
                        residual_fn,
                        decoder_self_attention_bias,
                        encoder_decoder_attention_bias,
                        hparams,
                        name="decoder"):
  """A stack of transformer layers.

  Args:
    decoder_input: a Tensor
    encoder_output: a Tensor
    residual_fn: a function from (layer_input, layer_output) -> combined_output
    decoder_self_attention_bias: bias Tensor for self-attention
      (see common_attention.attention_bias())
    encoder_decoder_attention_bias: bias Tensor for encoder-decoder attention
      (see common_attention.attention_bias())
    hparams: hyperparameters for model
    name: a string

  Returns:
    y: a Tensors
  """
  x = decoder_input
  # Summaries don't work in multi-problem setting yet.
  summaries = "problems" not in hparams.values() or len(hparams.problems) == 1
  with tf.variable_scope(name):
    for layer in xrange(hparams.num_hidden_layers):
      with tf.variable_scope("layer_%d" % layer):
        x = residual_fn(
            x,
            common_attention.multihead_attention(
                x,
                None,
                decoder_self_attention_bias,
                hparams.attention_key_channels or hparams.hidden_size,
                hparams.attention_value_channels or hparams.hidden_size,
                hparams.hidden_size,
                hparams.num_heads,
                hparams.attention_dropout,
                summaries=summaries,
                name="decoder_self_attention"))
        x = residual_fn(
            x,
            common_attention.multihead_attention(
                x,
                encoder_output,
                encoder_decoder_attention_bias,
                hparams.attention_key_channels or hparams.hidden_size,
                hparams.attention_value_channels or hparams.hidden_size,
                hparams.hidden_size,
                hparams.num_heads,
                hparams.attention_dropout,
                summaries=summaries,
                name="encdec_attention"))
        x = residual_fn(x, transformer_ffn_layer(x, hparams))
  return x

    def residual_fn(x, y):
      return common_layers.layer_norm(x + tf.nn.dropout(
          y, 1.0 - hparams.residual_dropout))

    encoder_input = tf.nn.dropout(encoder_input, 1.0 - hparams.residual_dropout)
    decoder_input = tf.nn.dropout(decoder_input, 1.0 - hparams.residual_dropout)