Zoneout の実装に関するメモ

ポイント

• LSTMをベースに Zoneout を実装。MNIST 手書き数字データでパフォーマンスを検証。

レファレンス

1. Zoneout: Regularizing RNNs by Randomly Preserving Hidden Activations

検証方法

• Base ( Regularization なし）と比較。
  
  

• Zoneout
  

　　　　　 　 　 （参照論文より引用）

データ

MNIST handwritten digits

from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets('***/mnist', \
                                     one_hot = True)

検証結果

数値計算例：

• n_units = 100
• learning_rate = 0.1
• batch_size = 64
• zoneout_prob = 0.2

Base

Zoneout ( no gradient clipping )

Zoneout ( clipped by norm, 0.5 )

サンプルコード

  def get_zoneout_mask(self, zoneout_prob, shape):
    keep_prob = tf.convert_to_tensor(zoneout_prob)
    random_tensor = keep_prob + tf.random_uniform(shape)
    binary_tensor = tf.floor(random_tensor)
    zoneout_mask = binary_tensor

    return zoneout_mask

  # Zoneout
  def inference(self, x, length, n_in, n_units, n_out, \
                  batch_size, forget_bias, zoneout_prob):

    x = tf.reshape(x, [-1, length, n_in])

    h = tf.zeros(shape = [batch_size, n_units], \
                    dtype = tf.float32)
    c = tf.zeros(shape = [batch_size, n_units], \
                    dtype = tf.float32)

    list_h = []
    list_c = []

    with tf.variable_scope('lstm'):
      init_norm = tf.truncated_normal_initializer(mean \
             = 0.0, stddev = 0.05, dtype = tf.float32)
      init_constant1 = tf.constant_initializer(value = \
             0.0, dtype = tf.float32)
      init_constant2 = tf.constant_initializer(value = \
             0.1, dtype = tf.float32)

      w_x = tf.get_variable('w_x', shape = [n_in, n_units \
             * 4], initializer = init_norm)
      w_h = tf.get_variable('w_h', shape = [n_units, \
             n_units * 4], initializer = init_norm)
      b = tf.get_variable('b', shape = [n_units * 4], \
             initializer = init_constant1)

      zoneout_mask_c = self.get_zoneout_mask(zoneout_prob, \
            [n_units])
      zoneout_mask_complement_c = tf.ones(shape = \
            [n_units], dtype = tf.float32) - zoneout_mask_c
      zoneout_mask_h = self.get_zoneout_mask(zoneout_prob, \
            [n_units])
      zoneout_mask_complement_h = tf.ones(shape = \
            [n_units], dtype = tf.float32) - zoneout_mask_h

      for t in range(length):

        t_x = tf.matmul(x[:, t, :], w_x)

        t_h = tf.matmul(h, w_h)

        i, f, o, g = tf.split(tf.add(tf.add(t_x, t_h), b), \
                               4, axis = 1)

        i = tf.nn.sigmoid(i)
        f = tf.nn.sigmoid(f + forget_bias)
        o = tf.nn.sigmoid(o)
        g = tf.nn.tanh(g)

        # zoneout
        c_temp = tf.add(tf.multiply(f, c), tf.multiply(i, g))
        h_temp = tf.multiply(o, tf.nn.tanh(c))

        c = zoneout_mask_c * c + \
                        zoneout_mask_complement_c * c_temp
        h = zoneout_mask_h * h + \
                        zoneout_mask_complement_h * h_temp

        list_h.append(h)
        list_c.append(c)

    with tf.variable_scope('pred'):
      w = self.weight_variable('w', [n_units, n_out])
      b = self.bias_variable('b', [n_out])

      y = tf.add(tf.matmul(list_h[-1], w), b)
      y = tf.nn.softmax(y, axis = 1)

      return y