"TensorFlow Learning" tf.split () and Tf.squeeze ()

Split (
    value,
    num_or_size_splits,
    axis=0,
    num=none,
    name= ' Split '
)

Input:
Value: tensor of the input
Num_or_size_splits: If it is an integer n, the input tensor is divided into N sub-tensor. If it is a tensor T, the input tensor is divided into Len (T) Sub tensor.
Axis: Default is 0, calculate Value.shape[axis], must be divisible by num_or_size_splits.

Example:

# num_or_size_splits is tensor t,len (T) is 3, so divided into 3 sub-tensor,
# axis 1, so value.shape[1] for 30,4+15+11 just for the split0
, Split1, Split2 = Tf.split (value, [4, 5, all], 1)

Tf.shape (split0)  # [, 4]
tf.shape (split1)  # [5, 15]
  tf.shape (Split2)  # [5, one]

# num_or_size_splits is an integer 3, divided into 3 tensor,value.shape[1] is 30, can be divisible by 3.
split0, split1, split2 = Tf.split (value, num_or_size_splits=3, Axis=1)
tf.shape (split0)  # [5, 10]

To give an example:

>>> A=np.reshape (Range), (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 () # divides a into two tensor,a.shape (1) to 2, which can be divided evenly, without error. # The output should be 2 shape for [4,1,3] 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]],

       [[[5], [[]]], [[[]], Array ([[[3, 4,]], [[9, 10, 11]], [[15, 16, 17]],
[[21, 22, 23]]] >>> c= Tf.split (a,2,0) # a.shape (0) is 4, divisible by 2, Output 2 [2,2,3] 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, ten, one]]], Array ([[[12, 13, 14], [15, 16, 17]], [[18, 19, 20],
[21, 22, 23]]]
>>>d= Tf.split (a,2,2) # A.shape (2) is 3, not divisible by 2, error. Traceback (most recent): 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 __exit__ Next (Self.gen) File "D:\Anaconda2\envs\tensorflow\lib\site-packages\tensorflow\python\framework\err ors_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 It's 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:INP
Ut[0] = <2>. >>> d= Tf.split (a,3,2) # changed to 3,a.shape (2) to 3, divisible, no error, return 3 [4,2,1] of tensor >>> d [&LT;TF. Tensor ' split_2:0 ' shape= (4, 2, 1) dtype=int32>, &LT;TF. Tensor ' Split_2:1 ' shape= (4, 2, 1) dtype=int32>, &LT;TF. Tensor ' split_2:2 ' shape= (4, 2, 1) dtype=int32>] >>> Sess.run (d) [Array ([[[0], [3]], [[6], [9]],

       [[[] [[[]], [[[]], [[[]], [[]]], Array ([[[1], [4]], [[7], [10]],

       [[+], [+]], [[[], [[]]]], Array ([[[2], [5]], [[8], [11]], [14], [17]], [[20], [23]]]

Attention:
Unlike Tf.split and reshape, the relative order between values is not changed. Only each dimension can be smaller and not grow.
Values are taken in the order of axis-1.

Tf.squeeze
Squeeze (
    input,
    axis=none,
    name=none,
    squeeze_dims=none
)

Remove the dimension of dimension 1.
Give me a chestnut:

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

You can also specify which dimension to remove:

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