用于连接两个矩阵：
mn = array_ops.concat([a, d], 1) # 按照第二维度相接，shape1 [m,a] shape2 [m,b] ，concat_done shape : [m,a+b]
tensorflow Rnn，Lstm，Gru，源码中是用以上的函数来链接Xt 和 Ht-1 的，两者的shape 分别是【batch_size, emb_size】【batch_size，Hidden_size】
连接接后为的shape为：【batch_size，embedding_size + Hidden_size】，作为当前时刻的输入；
测试代码：
1 import os 2 import tensorflow as tf 3 import numpy as np 4 import sys 5 from tensorflow.python.ops import array_ops 6 #array_ops.concat([inputs, state], 1) 7 8 a = tf.constant([[1,12,8,6], [3,4,6,7]]) # shape [2,4] 9 b = tf.constant([[10, 20,6,88], [30,40,7,8]]) # shape [2,4] 10 c = tf.constant([[10, 20,6,88,99], [30,40,7,8,15]]) #shape [2,5] 11 d = tf.constant([[10,20,6,88], [30,40,7,8],[30,40,7,8]]) # shape [3,4] 12 nn = tf.concat([a, d],0) # 按照第一维度相接，shape1 [a,m] shape2 [b,m] concat_done:[a+b,m] 13 nn_1 = tf.concat([a, c],1) # 按照第二维度相接，shape1 [m,a] shape2 [m,b] concat_done:[m,a+b] 14 mn = array_ops.concat([a, d], 0) # 按照第一维度相接，shape1 [a,m] shape2 [b,m] concat_done:[a+b,m] 15 mn_1 = array_ops.concat([a, c], 1) # 按照第二维度相接，shape1 [m,a] shape2 [m,b] concat_done:[m,a+b] 16 17 with tf.Session() as sess: 18 print (nn) 19 print (nn.eval()) 20 print (nn_1) 21 print (nn_1.eval()) 22 print (mn) 23 print (mn.eval()) # shape [5,4] 24 print (mn_1) 25 print (mn_1.eval()) # shape [2,9]
结果输出：
Tensor("concat:0", shape=(5, 4), dtype=int32) [[ 1 12 8 6] [ 3 4 6 7] [10 20 6 88] [30 40 7 8] [30 40 7 8]] Tensor("concat_1:0", shape=(2, 9), dtype=int32) [[ 1 12 8 6 10 20 6 88 99] [ 3 4 6 7 30 40 7 8 15]] Tensor("concat_2:0", shape=(5, 4), dtype=int32) [[ 1 12 8 6] [ 3 4 6 7] [10 20 6 88] [30 40 7 8] [30 40 7 8]] Tensor("concat_3:0", shape=(2, 9), dtype=int32) [[ 1 12 8 6 10 20 6 88 99] [ 3 4 6 7 30 40 7 8 15]]