【tensorflow 学习】tf.split()和tf.squeeze()

split(
value,
num_or_size_splits,
axis=0,
num=None,
name='split'
)

输入：

value： 输入的tensor

num_or_size_splits: 如果是个整数n，就将输入的tensor分为n个子tensor。如果是个tensor T，就将输入的tensor分为len(T)个子tensor。

axis： 默认为0，计算value.shape[axis], 一定要能被num_or_size_splits整除。

举例：

# num_or_size_splits是tensor T，len（T）为3，所以分为3个子tensor，
# axis为1，所以value.shape[1]为30，4+15+11正好为30
split0, split1, split2 = tf.split(value, [4, 15, 11], 1)

tf.shape(split0)  # [5, 4]
tf.shape(split1)  # [5, 15]
tf.shape(split2)  # [5, 11]

# num_or_size_splits是整数3， 分为3个tensor，value.shape[1]为30，能被3整除。
split0, split1, split2 = tf.split(value, num_or_size_splits=3, axis=1)
tf.shape(split0)  # [5, 10]

再举个实例：

>>> a=np.reshape(range(24),(4,2,3))
>>> a
array([[[ 0,  1,  2],
[ 3,  4,  5]],

[[ 6,  7,  8],
[ 9, 10, 11]],

[[12, 13, 14],
[15, 16, 17]],

[[18, 19, 20],
[21, 22, 23]]])

>>> sess=tf.InteractiveSession()
# 将a分为两个tensor，a.shape(1)为2，可以整除，不会报错。
# 输出应该为2个shape为[4,1,3]的tensor
>>> b= tf.split(a,2,1)
>>> b
[<tf.Tensor 'split:0' shape=(4, 1, 3) dtype=int32>, <tf.Tensor 'split:1' shape=(4, 1, 3) dtype=int32>]
>>> sess.run(b)
[array([[[ 0,  1,  2]],

[[ 6,  7,  8]],

[[12, 13, 14]],

[[18, 19, 20]]]), array([[[ 3,  4,  5]],

[[ 9, 10, 11]],

[[15, 16, 17]],

[[21, 22, 23]]])]
>>> c= tf.split(a,2,0)
# a.shape(0)为4，被2整除，输出2个[2,2,3]的Tensor
>>> c
[<tf.Tensor 'split_1:0' shape=(2, 2, 3) dtype=int32>, <tf.Tensor 'split_1:1' shape=(2, 2, 3) dtype=int32>]
>>> sess.run(c)
[array([[[ 0,  1,  2],
[ 3,  4,  5]],

[[ 6,  7,  8],
[ 9, 10, 11]]]), array([[[12, 13, 14],
[15, 16, 17]],

[[18, 19, 20],
[21, 22, 23]]])]
>>>d= tf.split(a,2,2)
#  a.shape(2)为3，不被2整除，报错。
Traceback (most recent call last):
File "D:\Anaconda2\envs\tensorflow\lib\site-packages\tensorflow\python\framework\common_shapes.py", line 671, in _call_cpp_shape_fn_impl
input_tensors_as_shapes, status)
File "D:\Anaconda2\envs\tensorflow\lib\contextlib.py", line 66, in __exit__
next(self.gen)
File "D:\Anaconda2\envs\tensorflow\lib\site-packages\tensorflow\python\framework\errors_impl.py", line 466, in raise_exception_on_not_ok_status
pywrap_tensorflow.TF_GetCode(status))
tensorflow.python.framework.errors_impl.InvalidArgumentError: Dimension size must be evenly divisible by 2 but is 3
Number of ways to split should evenly divide the split dimension for 'split_1' (op: 'Split') with input shapes: [], [4,2,3] and with computed input tensors: input[0] = <2>.
>>> d= tf.split(a,3,2)
# 改成3，a.shape(2)为3，整除，不报错，返回3个[4,2,1]的Tensor
>>> d
[<tf.Tensor 'split_2:0' shape=(4, 2, 1) dtype=int32>, <tf.Tensor 'split_2:1' shape=(4, 2, 1) dtype=int32>, <tf.Tensor 'split_2:2' shape=(4, 2, 1) dtype=int32>]
>>> sess.run(d)
[array([[[ 0],
[ 3]],

[[ 6],
[ 9]],

[[12],
[15]],

[[18],
[21]]]), array([[[ 1],
[ 4]],

[[ 7],
[10]],

[[13],
[16]],

[[19],
[22]]]), array([[[ 2],
[ 5]],

[[ 8],
[11]],

[[14],
[17]],

[[20],
[23]]])]

注意：

tf.split和reshape不同，不会改变数值之间的相对顺序。只能每个维度只能变小，不增大。

取值时按照axis-1的顺序来取。

tf.squeeze
squeeze(
input,
axis=None,
name=None,
squeeze_dims=None
)

去掉维数为1的维度。

举个栗子：

# 't' is a tensor of shape [1, 2, 1, 3, 1, 1]
tf.shape(tf.squeeze(t))  # [2, 3]

也可以指定去掉哪个维度：

# 't' is a tensor of shape [1, 2, 1, 3, 1, 1]
tf.shape(tf.squeeze(t, [2, 4]))  # [1, 2, 3, 1]