iclr_2017_benchmark.py 文件源码

def tree_lstm(self, left, right):
    # A variation on the tree LSTM -- we add an extra hidden layer.
    if self._weights is None:
      with tf.variable_scope(self._vscope):
        self._weights_0 = tf.get_variable(
            "weights_0", [FLAGS.vector_size*2, FLAGS.vector_size],
            initializer=tf.uniform_unit_scaling_initializer(1.43))
        self._bias_0 = tf.get_variable("bias_0", [FLAGS.vector_size],
                                       initializer=tf.zeros_initializer())
        self._weights = tf.get_variable(
            "weights", [FLAGS.vector_size, FLAGS.vector_size*4],
            initializer=tf.uniform_unit_scaling_initializer(1.0))
        self._bias = tf.get_variable("bias", [FLAGS.vector_size*4],
                                     initializer=tf.zeros_initializer())
    # One hidden layer
    x = tf.concat([left, right], 1)
    h0 = tf.nn.relu(tf.add(tf.matmul(x, self._weights_0), self._bias_0))

    # Do a single matrix multiply to compute all gates
    h1 = tf.add(tf.matmul(h0, self._weights), self._bias)
    (hfl, hfr, hi, hg) = tf.split(h1, 4, axis=1)

    fl = tf.nn.sigmoid(hfl)  # forget left
    fr = tf.nn.sigmoid(hfr)  # forget right
    i = tf.nn.sigmoid(hi)    # input gate
    g = tf.nn.tanh(hg)       # computation

    ylr = tf.add(tf.multiply(fl, left), tf.multiply(fr, right))
    ygi = tf.multiply(i, g)
    y = tf.add(ylr, ygi)

    return y