Function attention

DIEN/utils.py:275–318

(query, facts, attention_size, mask, stag='null', mode='LIST', softmax_stag=1, time_major=False, return_alphas=False)

def attention(query, facts, attention_size, mask, stag='null', mode='LIST', softmax_stag=1, time_major=False, return_alphas=False):
    if isinstance(facts, tuple):
        # In case of Bi-RNN, concatenate the forward and the backward RNN outputs.
        facts = tf.concat(facts, 2)

    if time_major:
        # (T,B,D) => (B,T,D)
        facts = tf.array_ops.transpose(facts, [1, 0, 2])

    mask = tf.equal(mask, tf.ones_like(mask))
    hidden_size = facts.get_shape().as_list()[-1]  # D value - hidden size of the RNN layer
    input_size = query.get_shape().as_list()[-1]

    # Trainable parameters
    w1 = tf.Variable(tf.random_normal([hidden_size, attention_size], stddev=0.1))
    w2 = tf.Variable(tf.random_normal([input_size, attention_size], stddev=0.1))
    b = tf.Variable(tf.random_normal([attention_size], stddev=0.1))
    v = tf.Variable(tf.random_normal([attention_size], stddev=0.1))

    with tf.name_scope('v'):
        # Applying fully connected layer with non-linear activation to each of the B*T timestamps;
        #  the shape of `tmp` is (B,T,D)*(D,A)=(B,T,A), where A=attention_size
        tmp1 = tf.tensordot(facts, w1, axes=1)
        tmp2 = tf.tensordot(query, w2, axes=1)
        tmp2 = tf.reshape(tmp2, [-1, 1, tf.shape(tmp2)[-1]])
        tmp = tf.tanh((tmp1 + tmp2) + b)

    # For each of the timestamps its vector of size A from `tmp` is reduced with `v` vector
    v_dot_tmp = tf.tensordot(tmp, v, axes=1, name='v_dot_tmp')  # (B,T) shape
    key_masks = mask # [B, 1, T]
    # key_masks = tf.expand_dims(mask, 1) # [B, 1, T]
    paddings = tf.ones_like(v_dot_tmp) * (-2 ** 32 + 1)
    v_dot_tmp = tf.where(key_masks, v_dot_tmp, paddings)  # [B, 1, T]
    alphas = tf.nn.softmax(v_dot_tmp, name='alphas')         # (B,T) shape

    # Output of (Bi-)RNN is reduced with attention vector; the result has (B,D) shape
    #output = tf.reduce_sum(facts * tf.expand_dims(alphas, -1), 1)
    output = facts * tf.expand_dims(alphas, -1)
    output = tf.reshape(output, tf.shape(facts))
    # output = output / (facts.get_shape().as_list()[-1] ** 0.5)
    if not return_alphas:
        return output
    else:
        return output, alphas


def din_fcn_attention(query, facts, attention_size, mask, stag='null', mode='SUM', softmax_stag=1, time_major=False, return_alphas=False, forCnn=False):