tensorflow(1) 共享变量

共享变量:reuse_variables()
example1:

with tf.variable_scope("try"):
    #先创建两个变量w1， w2
    w2 = tf.get_variable("w1",shape=[2,3,4], dtype=tf.float32)
    w3 = tf.get_variable("w2", shape=[2, 3, 4], dtype=tf.float32)
    #使用reuse_variables 将刚刚创建的两个变量共享
    tf.get_variable_scope().reuse_variables()
    w4 = tf.get_variable("w1", shape=[2, 3, 4], dtype=tf.float32)
    w5 = tf.get_variable("w2", shape=[2, 3, 4], dtype=tf.float32)
    #再进行共享的话，还需要再使用一次reuse_variables()
    tf.get_variable_scope().reuse_variables()
    w6 = tf.get_variable("w1", shape=[2, 3, 4], dtype=tf.float32)
    w7 = tf.get_variable("w2", shape=[2, 3, 4], dtype=tf.float32)

example2:

with tf.variable_scope("RNN"):
    for time_step in range(num_steps):
        if time_step > 0: tf.get_variable_scope().reuse_variables()
        # cell_out: [batch, hidden_size]
        (cell_output, state) = cell(inputs[:, time_step, :], state) # 按照顺序向cell输入文本数据
        outputs.append(cell_output)  # output: shape[num_steps][batch,hidden_size]

这样就可以实现变量共享了