如何切片张量的一部分？

Question

I want to slice [3.0 ,33.0].I have tried to access this slice by following code. I'm not so clear about tf.slice command. I'm not so clear about begin and size mentioned in documentaion about this command. Can someone please make it easy to understand.

batch = tf.constant([
  [#First image
    [[0.0,10.0],[1.0,11.0]],
    [[3.0,33.0],[4.0,44.0]]
  ],
  [#Second image
    [[5.0,55.0],[6.0,66.0]],
    [[7.0,77.0],[8.0,88.0]]
  ]
])
slice1 = tf.slice(batch,[0,0,0,0], [0,0,1,0]) 
sess = tf.InteractiveSEssion()
sess.run(tf.initialize_all_variables())
print slice1.eval()

Answer 1

I will explain your code with examples, so I created some cases, but first of all I'll explain you tf.slice(input, begin, size) parametters:

• input is a ref to a Tensor.
• begin is the index from the slice begins.
• size is the offset of the slice.

So tf.slice works selecting from input a sub-Tensor that starts at begin index and end at begin + size , treating begin and size as index vectors. The example below will clarify this:

batch = tf.constant([
        [#First image
            [
                [0.0,10.0],
                [1.0,11.0]
            ],
            [
                [3.0,33.0],
                [4.0,44.0]
            ]
        ],
        [#Second image
            [
                [5.0,55.0],
                [6.0,66.0]
            ],
            [
                [7.0,77.0],
                [8.0,88.0]
            ]
        ]
    ])
slice1 = tf.slice(batch,[0,0,0,0], [1,1,1,1]) 
slice2 = tf.slice(batch,[0,1,0,0], [1,1,2,2]) 
slice3 = tf.slice(batch,[1,1,1,0], [1,1,1,2]) 
slice4 = tf.slice(batch,[0,0,0,0], [2,2,2,2]) 
sess = tf.InteractiveSession()
print("slice1: \n" + str(slice1.eval()) + "\n")
print("slice2: \n" + str(slice2.eval()) + "\n")
print("slice3: \n" + str(slice3.eval()) + "\n")
print("slice4: \n" + str(slice4.eval()) + "\n")

The outputs in this case are:

slice1: 
[[[[ 0.]]]]

slice2: 
[[[[  3.  33.]
   [  4.  44.]]]]

slice3: 
[[[[  8.  88.]]]]

slice4: 
[[[[  0.  10.]
   [  1.  11.]]

  [[  3.  33.]
   [  4.  44.]]]


 [[[  5.  55.]
   [  6.  66.]]

  [[  7.  77.]
   [  8.  88.]]]]

• slice1 selects the first element of the Tensor because of it begins on [0,0,0,0] and picks only one element.
• slice2 selects the first element of the Tensor because of it begins on [0,1,0,0] and picks 1 element in the two first dimensions and 2 in three and four dimensions.
• slice3 selects the first element of the Tensor because of it begins on [1,1,1,0] and picks only 1 element in the three first dimensions and 2 in the last.
• slice4 selects all the element of the Tensor because of it begins on [0,0,0,0] and two elements by dimension, so it covers all the Tensor

Note that de number of dimensions are the same in all slides. If you one to remove dimensions with only one element you can use tf.squeeze .

Answer 2

As how to slice a tensor is well explained above, I will show the trick about how to slice every element in the same position as [3.0, 33.0] in the tensor here (which is a similar problem that I met)

batch = tf.constant([
  [#First image
    [[0.0,10.0],[1.0,11.0]],
    [[3.0,33.0],[4.0,44.0]]
  ],
  [#Second image
    [[5.0,55.0],[6.0,66.0]],
    [[7.0,77.0],[8.0,88.0]]
  ]
])
batch_shape = batch.shape
batch_sliced = tf.slice(batch,(0,1,0,0),(batch_shape[0],1,1,batch_shape[-1]))

Then you will get batch_sliced as

<tf.Tensor: shape=(2, 1, 1, 2), dtype=float32, numpy=
array([[[[ 3., 33.]]],
       [[[ 7., 77.]]]]```