TensorFlow高阶操作之合并与分割

Merge And Split

1. 【tf.concat()】:拼接

[classes, students, scores]

//需要指定拼接维度，数据其他维度相同
In [4]: a = tf.ones([4, 35, 8])

In [5]: b = tf.ones([2, 35, 8])

In [6]: c = tf.concat([a, b], axis=0)

In [7]: c.shape
Out[7]: TensorShape([6, 35, 8])

In [8]: a = tf.ones([4, 32, 8])

In [9]: b = tf.ones([4, 3, 8])

In [10]: tf.concat([a, b], axis=1).shape
Out[10]: TensorShape([4, 35, 8])

2. 【tf.stack()】:堆叠

school1:[classes, students, scores]
school2:[classes, students, scores]
[school, classes, students, scores]

//需要指定新插入的维度，数据所有维度相同
In [11]: a = tf.ones([4, 35, 8])

In [12]: b = tf.ones([4, 35, 8])

In [13]: a.shape
Out[13]: TensorShape([4, 35, 8])

In [14]: b.shape
Out[14]: TensorShape([4, 35, 8])

In [15]: tf.stack([a, b], axis=0).shape
Out[15]: TensorShape([2, 4, 35, 8])

In [16]: tf.stack([a, b], axis=3).shape
Out[16]: TensorShape([4, 35, 8, 2])

3. 【tf.unstack()】：拆叠

//将某一维度数据拆分为单维度
In [19]: a = tf.ones([4, 35, 8])

In [20]: a1, a2, a3, a4 = tf.unstack(a, axis=0)

In [22]: a1.shape
Out[22]: TensorShape([35, 8])

In [25]: res = tf.unstack(a, axis=2)

In [28]: res[0].shape
Out[28]: TensorShape([4, 35])

4. 【tf.split()】：分割

对比【unstack()】不仅可以指定分割的轴，还能指定分割的分配比例

In [39]: a = tf.ones([2, 4, 35, 8])

In [40]: res = tf.unstack(a, axis=3)  //对某个轴均分为单体
In [41]: len(res)
Out[41]: 8
//如果不整除，则不可以分割
In [42]: res=tf.split(a, axis=3, num_or_size_splits=2)    //分为数量相同的N个单体

In [43]: len(res)      
Out[43]: 2

In [48]: res[0].shape
Out[48]: TensorShape([2, 4, 35, 4])

In [52]: a = tf.ones([2, 4, 35, 8])

In [53]: res=tf.split(a, axis=3, num_or_size_splits=[2, 2, 4])

In [54]: res[0].shape, res[1].shape, res[2].shape
Out[54]:
(TensorShape([2, 4, 35, 2]),
 TensorShape([2, 4, 35, 2]),
 TensorShape([2, 4, 35, 4]))

参考资料

• 深度学习与TensorFlow 2入门实战
• 深度学习与TensorFlow 2入门实战-源码和PPT
• TensorFlow2.0官方文档
• W3Cschool翻译的TensorFlow官方文档
  【注】TensorFlow版本未知。