Recurrent Batch Normalization ( と Recurrent Dropout ) の実装に関するメモ

ポイント

• LSTMをベースに Batch Normalization を実装。MNIST 手書き数字データでパフォーマンスを検証。
• Batch Normalization、Dropout ともに効果を確認できず。
• 今後、別タスク、別データで追加検証。

レファレンス

1. Recurrent Batch Normalization
2. Recurrent Dropout without Memory Loss

検証方法

• Base model（ Normalization なし）、Recurrent Dropout と比較。

　　　　　 　 　 （参照論文１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 = 0.01
• batch_size = 64
• keep_prob = 0.8

Base Model ( without Normalization )

Recurrent Batch Normalization

Recurrent Dropout

サンプルコード

  # Batch Normalization
  def inference(self, x, length, n_in, n_units, n_out, \
                    batch_size, forget_bias, keep_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)

      gamma_x = tf.get_variable('gamma_x', shape = \
              [n_units * 4], initializer = init_constant2)
      gamma_h = tf.get_variable('gamma_h', shape = \
              [n_units * 4], initializer = init_constant2)
      gamma_c = tf.get_variable('gamma_c', shape = \
              [n_units], initializer = init_constant2)
      beta_c = tf.get_variable('beta_c', shape = \
              [n_units], initializer = init_constant1)

      dropout_mask = self.get_dropout_mask(keep_prob, \
              [n_units])

      for t in range(length):

        t_x = tf.matmul(x[:, t, :], w_x)
        mean_x, var_x = tf.nn.moments(t_x, [0])
        bn_x = gamma_x * (t_x - mean_x) / \
                             tf.sqrt(var_x + 1e-10) 

        # dropout
        #h *= dropout_mask

        t_h = tf.matmul(h, w_h)
        mean_h, var_h = tf.nn.moments(t_h, [0])
        bn_h = gamma_h * (t_h - mean_h) / \
                            tf.sqrt(var_h + 1e-10) 

        # batch normalization
        i, f, o, g = tf.split(tf.add(tf.add(bn_x, bn_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)

        # dropout
        #g *= dropout_mask

        c = tf.add(tf.multiply(f, c), tf.multiply(i, g))

        # batch normalization
        mean_c, var_c = tf.nn.moments(c, [0])
        c = gamma_c * (c - mean_c) / tf.sqrt(var_c + 1e-10) 
        c = tf.add(c, beta_c)  

        h = tf.multiply(o, tf.nn.tanh(c))

        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

  # Dropout Mask
  def get_dropout_mask(self, keep_prob, shape):
    keep_prob = tf.convert_to_tensor(keep_prob)
    random_tensor = keep_prob + tf.random_uniform(shape)
    binary_tensor = tf.floor(random_tensor)
    dropout_mask = binary_tensor / keep_prob

    return dropout_mask