CODEWITHSUNDEEP
pythonpytorch
MariusKuzmin
MariusKuzmin

Reputation: 5

How to insert values of one tensor into another?

I am trying to insert tensor y into the tensors x final dimension (y_dim). The final tensor should be of size (100, 16, 16, 1) where the values of y are placed in each of 100 x 0dimension

import torch

y_dim = 1
x = torch.randn(100, 16, 16, y_dim)
#OR x = torch.randn(100, 16, 16)
y = torch.randn(100)

Xy = torch.cat((x, y), dim=3)

Upvotes: 0

Views: 749

Answers (1)

Ivan
Ivan

Reputation: 40618

I think you are missing something in your understanding of tensors and dimensions. The easiest thing is to consider your tensor x as a batch containing 100 maps of width and height 16, i.e. 100 16x16-maps. So you are manipulating a tensor containing 100*16*16 elements. Your y, on the other hand, contains 100 scalar values, it has 100 elements.

I'm turning the question back to you:

How would you concatenate 100 16x16-maps with 100 scalar values?

The above question has no answer. There are certain things that can be done though, assumptions that can be made on y in order to perform a concatenation:

  • If you had a tensor y containing 16x16 maps as well, then yes this operation would be achievable:

    >>> x = torch.rand(100, 16, 16)
>>> y = torch.rand(100, 16, 16)

>>> torch.cat((x, y)).shape
torch.Size([200, 16, 16])

  • If you consider the y in your question, you could expand the 100 scalar values to 16x16 maps. And, then concatenate with x:

    >>> x = torch.rand(100, 16, 16)
>>> y = torch.rand(100)

>>> y_r = y[:,  None, None].repeat(1, 16, 16)
>>> torch.cat((x, y_r))
torch.Size([200, 16, 16])

Upvotes: 1

Related Questions