tensorflow 实现batch_normal

转自：https://blog.csdn.net/hanyong4719/article/details/80558995

1. def batch_norm_layer(x, train_phase, scope_bn):##x is input-tensor, train_phase is tf.Variable(True/False)    
2.     with tf.variable_scope(scope_bn):    
3.         beta = tf.Variable(tf.constant(0.0, shape=[x.shape[-1]]), name='beta', trainable=True)    
4.         gamma = tf.Variable(tf.constant(1.0, shape=[x.shape[-1]]), name='gamma', trainable=True)    
5.         axises =list(range(len(x.shape) - 1))# np.arange(len(x.shape) - 1)    
6.         batch_mean, batch_var = tf.nn.moments(x, axises, name='moments')    
7.         ema = tf.train.ExponentialMovingAverage(decay=0.5)    
8.     
9.         def mean_var_with_update():    
10.             ema_apply_op = ema.apply([batch_mean, batch_var])    
11.             with tf.control_dependencies([ema_apply_op]):    
12.                 return tf.identity(batch_mean), tf.identity(batch_var)    
13.     
14.         mean, var = tf.cond(train_phase, mean_var_with_update,lambda: (ema.average(batch_mean), ema.average(batch_var)))    
15.         normed = tf.nn.batch_normalization(x, mean, var, beta, gamma, 1e-3)    
16.     return normed