Concatenation/splitting

Contents

Concatenation/splitting#

Sometimes, it’s necessary to concatenate multiple tensors into one tensor, or conversely, split an existing tensor into a set of separate tensors. This section explores the options in PyTorch that allow you to accomplish such tasks.

import torch

Repeat method#

The torch.Tensor.repeat function allows you to duplicate a tensor’s values along specific dimensions, creating a new tensor with the replicated values.

For example, consider a tensor created using torch.arange(0, 2).

example_array = torch.arange(3)
example_array

tensor([0, 1, 2])

If you use repeat along a nested axis (e.g., repeat(3)), the input vector is duplicated three times, resulting in a longer vector.

example_array.repeat(3)

tensor([0, 1, 2, 0, 1, 2, 0, 1, 2])

By replicating along the rows (e.g., repeat(3, 1)), you will create a matrix with three rows. Each row will contain a copy of the original input vector.

torch.arange(3).repeat(3, 1)

tensor([[0, 1, 2],
        [0, 1, 2],
        [0, 1, 2]])

Finally, let’s try to duplicate the vector three times vertically and two times in the depth dimension. This will create a tensor with three columns, each containing a duplicated version of the original vector, and two layers of these duplicated columns.

torch.arange(3).repeat(2, 3, 1)

tensor([[[0, 1, 2],
         [0, 1, 2],
         [0, 1, 2]],

        [[0, 1, 2],
         [0, 1, 2],
         [0, 1, 2]]])
Contents