# How to Transform Tensors

An application must specify the shape of each tensor to be created. The `tf`

package provides functions that update tensors and their shapes after creation. This table lists these transformation functions and provides a description of each.

Functions for Transforming Tensors

Function | Description |

`cast(tensor, dtype, name=None)` |
Changes the tensor’s data type to the given type |

`reshape(tensor, shape, name=None)` |
Returns a tensor with the same elements as the given tensor with the given shape |

`squeeze(tensor, axis=None, name=None, squeeze_dims=None)` |
Removes dimensions of size 1 |

`reverse(tensor, axis, name=None)` |
Reverses given dimensions of the tensor |

`slice(tensor, begin, size, name=None)` |
Extracts a portion of a tensor |

`stack(tensors, axis=0, name='stack')` |
Combines a list of tensors into a tensor of greater rank |

`unstack(tensor, num=None, axis=0, name='unstack')` |
Splits a tensor into a list of tensors of lesser rank |

Despite its name, reshape doesn’t modify an existing tensor. Instead, the function returns a tensor with the same elements as the given tensor and the specified shape. For example, the following code uses `reshape `

to convert a four-element vector into a 2-x-2 matrix:

vec = tf.constant([1., 2., 3., 4.]) mat = tf.reshape(vec, [2, 2])# Result: [[1. 2.], [3. 4.]]

If any dimension of a tensor has a size of 1, calling `squeeze `

will remove it from the tensor, thereby reducing the tensor’s rank. If the function’s `axis `

parameter identifies one or more dimensions, only those dimensions will be affected by `squeeze`

.

In the `reverse `

function, the `axis `

parameter identifies one or more dimensions to be reversed. The following code demonstrates how `reverse `

works:

mat = tf.constant([[1., 2., 3.], [4., 5., 6.]]) rev_mat = tf.reverse(end, [0])# Result: [[4. 5. 6.], [1. 2. 3.]]rev_mat = tf.reverse(end, [1])# Result: [[3. 2. 1.], [6. 5. 4.]]rev_mat = tf.reverse(end, [0, 1])# Result: [[6. 5. 4.], [3. 2. 1.]]

The `slice `

function extracts subtensors from a tensor. The `begin `

parameter identifies the index of the first element to be extracted, and `size `

identifies the shape of the tensor to be extracted, starting from the `begin `

location.

For example, suppose that you want to extract the lower-right 2-x-2 matrix from a 3-x-3 matrix. The index of the first extracted element is [1, 1] and the size of the desired tensor is [2, 2]. The following code uses `slice `

to perform this extraction:

mat = tf.constant([[1., 2., 3.], [4., 5., 6.], [7., 8., 9.]]) slice_mat = tf.slice(mat, [1, 1], [2, 2])# Result: [[5. 6.] [7. 8.]]

`stack `

accepts a list of tensors of rank N and returns a single tensor of rank N+1. In addition to having the same rank, the input tensors must have the same shape. The following code demonstrates how `stack `

can be used. The function combines three one-dimensional tensors into a two-dimensional tensor:

t1 = tf.constant([1., 2.]) t2 = tf.constant([3., 4.]) t3 = tf.constant([5., 6.]) t4 = tf.stack([t1, t2, t3])

When these operations execute, `t4`

will equal [[1. 2.] [3. 4.] [5. 6.]]. If the axis parameter is set to 1, stacking will be performed along the second dimension, so t4 will set to [[1. 3. 5.] [2. 4. 6.]].

`unstack `

performs the inverse operation of `stack`

. That is, unstack accepts a tensor of rank N and returns a list of tensors of rank N-1. The num parameter determines how many tensors should be unpacked, and if this isn’t set, `unstack `

infers the number from the tensor’s shape.