tensorflow 笔记7：tf.concat 和 ops中的array_ops.concat

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：

import os
import tensorflow as tf
import numpy as np
import sys
from tensorflow.python.ops import array_ops
#array_ops.concat([inputs, state], 1)

a = tf.constant([[1,12,8,6], [3,4,6,7]])  # shape [2,4]
b = tf.constant([[10, 20,6,88], [30,40,7,8]]) # shape [2,4]
c = tf.constant([[10, 20,6,88,99], [30,40,7,8,15]]) #shape [2,5]
d = tf.constant([[10,20,6,88], [30,40,7,8],[30,40,7,8]]) # shape [3,4]
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]
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]
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]
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]

with tf.Session() as sess:
     print (nn)
     print (nn.eval())
     print (nn_1)
     print (nn_1.eval())
     print (mn)
     print (mn.eval())   # shape [5,4]
     print (mn_1)
     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]]