tensorflow的tile使用

當你需要按照矩陣維度復制數據時候，可以使用tensorflow的tile函數

a1 = tf.tile(a, [2, 2]) 表示把a的第一個維度復制兩次，第二個維度復制2次。
注意使用tf.nn.softmax(r, axis=0),表示對每一列取softmax，一定要注意維度，axis=0表示對列取softmax，不然數據會出錯

def tensoflow_test():
    # 一個batch有20個樣本，每個樣本的長度為5，每一個為200維度
    lstm_outpus = tf.truncated_normal(shape=[2, 5, 4], mean=0, stddev=1)
    # 變形成二維
    lstm_o = tf.reshape(lstm_outpus, shape=[-1, 4])
    # 經過非線性
    M = tf.tanh(lstm_o)
    # 初始化權重信息
    w = tf.truncated_normal(shape=[4,1], mean=0, stddev=1)
    # 權重tf.matmul(M, w)
    r = tf.matmul(M, w)
    a = tf.nn.softmax(r, axis=0)
    alpha = tf.tile(a, (1, 4))
    # attention_res = lstm_o * alpha

    # M = tf.reshape(t, shape=[-1, 200])
    # o = tf.Variable(tf.truncated_normal([1, 200]), name='w', dtype=tf.float32)
    # a = tf.Variable(tf.truncated_normal([2,3]), dtype=tf.float32)
    # b = tf.Variable(tf.truncated_normal([2,3]), dtype=tf.float32)
    # a_b = tf.multiply(a,b)
    # # a_b = a * b
    # w = tf.transpose(o)
    # res = tf.matmul(M, w)
    # res2 = tf.reshape(res, shape=[-1, 5])
    # copy_res = tf.tile(res2, (3,1))
    # init_op = tf.global_variables_initializer()

    with tf.Session() as sess:
        # sess.run(init_op)
        # print(sess.run(res))
        # print(sess.run(res2))
        # print(res2)
        # print(sess.run(copy_res))
        # print(copy_res)
        # print(sess.run(lstm_o))
        # print(sess.run(lstm_outpus))
        # print(sess.run(w))
        print(lstm_outpus)
        print(lstm_o)
        print(alpha)
        # print(sess.run(lstm_outpus))
        print(sess.run([a, alpha]))
        # print(sess.run(alpha))
        # print(sess.run(alpha))
        # print(sess.run(attention_res))