QAnet Encoder

#!/usr/bin/python3
# -*- coding: utf-8 -*-
'''
date: 2019/8/19
mail: cally.maxiong@gmail.com
blog: http://www.cnblogs.com/callyblog/
'''
import math
import tensorflow as tf

__all__ = ['encoder']

initializer_relu = lambda: tf.contrib.layers.variance_scaling_initializer(factor=2.0,
                                                             mode='FAN_IN',
                                                             uniform=False,
                                                             dtype=tf.float32)
regularizer = tf.contrib.layers.l2_regularizer(scale=3e-7)

def encoder(inputs, num_blocks, num_conv_layers, kernel_size, inputs_mask, num_filters=128, input_projection=False,
            num_heads=8, is_training=False, reuse=None, dropout=0.0, scope="res_block"):
    """
    QAnet encoder
    :param inputs: inputs
    :param num_blocks: number of conv and self attention block
    :param num_conv_layers: number of layers of each conv block
    :param kernel_size: kernel size
    :param inputs_mask: input mask
    :param num_filters: number of conv filters
    :param input_projection: whether add linear before through conv and self attention block
    :param num_heads: self attention number of heads
    :param is_training: whether training
    :param reuse: whether reuse variable
    :param dropout: dropout rate
    :param scope: scope name
    """
    with tf.variable_scope(scope, reuse=reuse):
        if input_projection:
            inputs = tf.layers.conv1d(inputs, filters=num_filters, kernel_size=1, use_bias=False, reuse=reuse, name='input_projection')

        outputs = inputs

        for i in range(num_blocks):
            outputs = _add_timing_signal_1d(outputs)
            outputs = _conv_block(outputs, num_conv_layers, kernel_size, num_filters, reuse=reuse, is_training=is_training,
                                  dropout=dropout, scope="conv_block%d" % i)

            outputs = _multihead_attention(outputs, inputs_mask, dropout_rate=dropout, num_heads=num_heads,
                                           training=is_training, reuse=reuse, scope="self_attention_layers%d" % i)
        return outputs

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]
    signal = _get_timing_signal_1d(length, channels, min_timescale, max_timescale)
    return x + signal

def _get_timing_signal_1d(length, channels, min_timescale=1.0, max_timescale=1.0e4):
    """Gets a bunch of sinusoids of different frequencies.
    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:
    length: scalar, length of timing signal sequence.
    channels: scalar, size of timing embeddings to create. The number of
        different timescales is equal to channels / 2.
    min_timescale: a float
    max_timescale: a float
    Returns:
    a Tensor of timing signals [1, length, channels]
    """
    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 signal

def _conv_block(inputs, num_conv_layers, kernel_size, num_filters, scope="conv_block", is_training=False, reuse=None,
                dropout=0.0):
    """
    conv block, contain depth wise separable convolution and conv block
    :param inputs: inputs
    :param num_conv_layers: number of conv layers
    :param kernel_size: conv kernel size
    :param num_filters: number of conv filters
    :param scope: scope name
    :param is_training: whether training
    :param reuse: whether reuse variable
    :param dropout: dropout rate
    """
    with tf.variable_scope(scope, reuse=reuse):
        outputs = tf.expand_dims(inputs, 2)

        for i in range(num_conv_layers):
            residual = outputs
            outputs = _ln(outputs, scope="layer_norm_%d" % i, reuse=reuse)

            if i % 2 == 0 and is_training:
                outputs = tf.layers.dropout(outputs, dropout, training=is_training)

            outputs = _depthwise_separable_convolution(outputs, kernel_size=(kernel_size, 1), num_filters=num_filters,
                                                       scope="depthwise_conv_layers_%d" % i, reuse=reuse)

            outputs = tf.layers.dropout(outputs, dropout, training=is_training)
            outputs = outputs + residual

        return tf.squeeze(outputs, 2)

def _depthwise_separable_convolution(inputs, kernel_size, num_filters, bias=True, reuse=None,
                                     scope="depthwise_separable_convolution"):
    """
    depth wise separable convolution
    :param inputs: input
    :param kernel_size: kernel size
    :param num_filters: number of filter
    :param bias: whether use bias
    :param reuse: whether reuse variable
    :param scope: scope name
    """
    with tf.variable_scope(scope, reuse=reuse):
        shapes = inputs.shape.as_list()
        depthwise_filter = tf.get_variable("depthwise_filter",
                                        (kernel_size[0], kernel_size[1], shapes[-1], 1),
                                        dtype=tf.float32,
                                        regularizer=regularizer,
                                        initializer=initializer_relu())
        pointwise_filter = tf.get_variable("pointwise_filter",
                                        (1, 1, shapes[-1], num_filters),
                                        dtype=tf.float32,
                                        regularizer=regularizer,
                                        initializer=initializer_relu())
        outputs = tf.nn.separable_conv2d(inputs,
                                        depthwise_filter,
                                        pointwise_filter,
                                        strides=(1, 1, 1, 1),
                                        padding="SAME")

        if bias:
            b = tf.get_variable("bias",
                                outputs.shape[-1],
                                regularizer=regularizer,
                                initializer=tf.zeros_initializer())
            outputs += b
        outputs = tf.nn.relu(outputs)
        return outputs

def _multihead_attention(inputs,
                         input_mask,
                         num_heads=8,
                         dropout_rate=0.0,
                         training=False,
                         reuse=None,
                         scope="multihead_attention"):
    '''Applies multihead attention. See 3.2.2
    inputs: A 3d tensor with shape of [N, T, d_model].
    input_mask: A 3d tensor with shape of [N, T].
    num_heads: An int. Number of heads.
    dropout_rate: A floating point number.
    training: Boolean. Controller of mechanism for dropout.
    causality: Boolean. If true, units that reference the future are masked.
    scope: Optional scope for `variable_scope`.

    Returns
      A 3d tensor with shape of (N, T_q, C)
    '''

    with tf.variable_scope(scope, reuse=reuse):
        inputs = inputs * tf.cast(tf.expand_dims(input_mask, axis=-1), dtype=tf.float32)
        inputs = _ln(inputs, reuse=reuse, scope=scope+'_layer_normal')

        queries = inputs
        keys = inputs
        values = inputs

        d_model = queries.get_shape().as_list()[-1]
        # Linear projections
        Q = tf.layers.dense(queries, d_model, use_bias=False, reuse=reuse)  # (N, T_q, d_model)
        K = tf.layers.dense(keys, d_model, use_bias=False, reuse=reuse)  # (N, T_k, d_model)
        V = tf.layers.dense(values, d_model, use_bias=False, reuse=reuse)  # (N, T_k, d_model)

        # Split and concat
        Q_ = tf.concat(tf.split(Q, num_heads, axis=2), axis=0)  # (h*N, T_q, d_model/h)
        K_ = tf.concat(tf.split(K, num_heads, axis=2), axis=0)  # (h*N, T_k, d_model/h)
        V_ = tf.concat(tf.split(V, num_heads, axis=2), axis=0)  # (h*N, T_k, d_model/h)

        # Attention
        outputs = _scaled_dot_product_attention(Q_, K_, V_, dropout_rate, training, reuse=reuse)

        # Restore shape
        outputs = tf.concat(tf.split(outputs, num_heads, axis=0), axis=2)  # (N, T_q, d_model)

        # feed forward
        outputs = tf.layers.conv1d(outputs, filters=d_model, kernel_size=1, reuse=reuse, trainable=training)
        outputs = tf.layers.dropout(outputs, dropout_rate, training=training)

        # Residual connection
        outputs = queries + outputs

        # Normalize
        outputs = _ln(outputs, reuse=reuse, scope='feed_forword_layer_normal')

    return outputs

def _scaled_dot_product_attention(Q, K, V,
                                  dropout_rate=0.,
                                  training=False,
                                  reuse=None,
                                  scope="scaled_dot_product_attention"):
    '''See 3.2.1.
    Q: Packed queries. 3d tensor. [N, T_q, d_k].
    K: Packed keys. 3d tensor. [N, T_k, d_k].
    V: Packed values. 3d tensor. [N, T_k, d_v].
    causality: If True, applies masking for future blinding
    dropout_rate: A floating point number of [0, 1].
    training: boolean for controlling droput
    scope: Optional scope for `variable_scope`.
    '''
    with tf.variable_scope(scope, reuse=reuse):
        d_k = Q.get_shape().as_list()[-1]

        # dot product
        outputs = tf.matmul(Q, tf.transpose(K, [0, 2, 1]))  # (N, T_q, T_k)

        # scale
        outputs /= d_k ** 0.5

        # key masking, delete key 0
        outputs = _mask(outputs, Q, K, type="key")

        # softmax
        outputs = tf.nn.softmax(outputs)
        attention = tf.transpose(outputs, [0, 2, 1])
        tf.summary.image("attention", tf.expand_dims(attention[:1], -1))

        # query masking, delete query <pad>
        outputs = _mask(outputs, Q, K, type="query")

        # dropout
        outputs = tf.layers.dropout(outputs, rate=dropout_rate, training=training)

        # weighted sum (context vectors)
        outputs = tf.matmul(outputs, V)  # (N, T_q, d_v)

    return outputs

def _mask(inputs, queries=None, keys=None, type=None):
    """Masks paddings on keys or queries to inputs
    inputs: 3d tensor. (N, T_q, T_k)
    queries: 3d tensor. (N, T_q, d)
    keys: 3d tensor. (N, T_k, d)

    e.g.,
    >> queries = tf.constant([[[1.],
                        [2.],
                        [0.]]], tf.float32) # (1, 3, 1)
    >> keys = tf.constant([[[4.],
                     [0.]]], tf.float32)  # (1, 2, 1)
    >> inputs = tf.constant([[[4., 0.],
                               [8., 0.],
                               [0., 0.]]], tf.float32)
    >> mask(inputs, queries, keys, "key")
    array([[[ 4.0000000e+00, -4.2949673e+09],
        [ 8.0000000e+00, -4.2949673e+09],
        [ 0.0000000e+00, -4.2949673e+09]]], dtype=float32)
    >> inputs = tf.constant([[[1., 0.],
                             [1., 0.],
                              [1., 0.]]], tf.float32)
    >> mask(inputs, queries, keys, "query")
    array([[[1., 0.],
        [1., 0.],
        [0., 0.]]], dtype=float32)
    """
    outputs = None
    padding_num = -2 ** 32 + 1
    if type in ("k", "key", "keys"):
        # Generate masks
        masks = tf.sign(tf.reduce_sum(tf.abs(keys), axis=-1))  # (N, T_k)
        masks = tf.expand_dims(masks, 1) # (N, 1, T_k)
        masks = tf.tile(masks, [1, tf.shape(queries)[1], 1])  # (N, T_q, T_k)

        # Apply masks to inputs
        paddings = tf.ones_like(inputs) * padding_num

        outputs = tf.where(tf.equal(masks, 0), paddings, inputs)  # (N, T_q, T_k)
    elif type in ("q", "query", "queries"):
        # Generate masks
        masks = tf.sign(tf.reduce_sum(tf.abs(queries), axis=-1))  # (N, T_q)
        masks = tf.expand_dims(masks, -1)  # (N, T_q, 1)
        masks = tf.tile(masks, [1, 1, tf.shape(keys)[1]])  # (N, T_q, T_k)

        # Apply masks to inputs
        outputs = inputs*masks
    else:
        print("Check if you entered type correctly!")

    return outputs

def _ln(inputs, epsilon=1e-6, reuse=None, scope="ln"):
    '''Applies layer normalization. See https://arxiv.org/abs/1607.06450.
    inputs: A tensor with 2 or more dimensions, where the first dimension has `batch_size`.
    epsilon: A floating number. A very small number for preventing ZeroDivision Error.
    scope: Optional scope for `variable_scope`.

    Returns:
      A tensor with the same shape and data dtype as `inputs`.
    '''
    with tf.variable_scope(scope, reuse=reuse):
        inputs_shape = inputs.get_shape()
        params_shape = inputs_shape[-1:]

        mean, variance = tf.nn.moments(inputs, [-1], keep_dims=True)
        beta = tf.get_variable("beta", params_shape, initializer=tf.zeros_initializer())
        gamma = tf.get_variable("gamma", params_shape, initializer=tf.ones_initializer())
        normalized = (inputs - mean) / ((variance + epsilon) ** (.5))
        outputs = gamma * normalized + beta

    return outputs

在QAnet最后的三个encoder中，各项参数为，其中hidden size为context_query输出的hidden size

encoder(enc[i], num_blocks=7, num_conv_layers=2, kernel_size=5, inputs_mask=input_mask, num_filters=hidden_size, num_heads=8,
                             scope='Model_Encoder', reuse=True if i > 0 else None, is_training=False, dropout=0.1)