How to pad a tensor-CodePudding

How would I pad this tensor by adding element 100 on the end

a = tensor([[ 101,  103],
    [ 101, 1045, 223],
    [ 101,  777, 665 , 889],
    [ 101,  888]])

So the result would be:

 b = tensor([[ 101,  103, 100, 100],
    [ 101, 1045, 223, 100],
    [ 101,  777, 665 , 889],
    [ 101,  888, 100, 100]])

I know the functions is torch.nn.functional.pad(), but i could not any simple example with a tensor like this that is probably a 2d tensor.

Which was surprising, because this is what a (most) typical padding is.

CodePudding user response：

Similar to the numpy case, see Convert Python sequence to NumPy array, filling missing values, you could adjust the size of your sub lists using itertools.zip_longest.

from itertools import zip_longest
tensor_lists = [
    [ 101,  103],
    [ 101, 1045, 223],
    [ 101,  777, 665 , 889],
    [ 101,  888]
]
fillvalue = 100
padded_list = list(zip(*zip_longest(*tensor_lists, fillvalue=fillvalue)))
...  # convert to tensor and use it

Here, zip_longest adds the missing value, and the second zip transposes the result again. You could of cause first create the tensor and then transpose.

CodePudding user response：

You can use torch.nested.to_padded_tensor (docs):

import torch
a = [
    [101, 103],
    [101, 1045, 223],
    [101, 777, 665, 889],
    [101, 888]
]

a = torch.nested.nested_tensor(list(map(torch.tensor, a)))

torch.nested.to_padded_tensor(a, 100)
tensor([[ 101,  103,  100,  100],
        [ 101, 1045,  223,  100],
        [ 101,  777,  665,  889],
        [ 101,  888,  100,  100]])