Why initial these layer's bias manually?

[limaodaxia]

with tf.variable_scope('decoder', reuse=tf.AUTO_REUSE):
    for i in range(8):
        with tf.variable_scope("resnetBlockFC_%d" % i ):
            b_initializer=tf.constant_initializer(0.0)
            w_initializer = tf.random_normal_initializer(mean=0.0,stddev=np.sqrt(2) / np.sqrt(512))
            net = tf.layers.dense(tf.nn.relu(net),512,kernel_initializer=w_initializer,bias_initializer=b_initializer)
            
b_initializer=tf.constant_initializer(-0.5)
w_initializer = tf.random_normal_initializer(mean=2*np.sqrt(np.pi) / np.sqrt(512), stddev = 0.000001)

Why initial these layer's bias manually?