How to extend one tensor with another if they do not have the same size

Question

a =  tensor([   [101,  103],
        [101,  1045]
    ])


b =  tensor([   [101,  777, 227],
        [101,  888, 228]
    ])

How to I get this tensor c from a and b:

c = a + b =  tensor([   [101,  103, 0],
        [101,  1045, 0],
        [101,  777, 227],
        [101,  888, 228]
        
    ])

I tried with c = torch.cat((a, b), dim=0) but does not work.

Answer 1

In general you could try to pad the first tensor to the shape of the other. For this exact problem you could do that the following way:

import torch
import torch.nn.functional as F

a = torch.tensor([
    [101,  103],
    [101,  1045]
])

b = torch.tensor([
    [101,  777, 227],
    [101,  888, 228]
])


a = F.pad(a, pad=(0, b.size()[1] - a.size()[1], 0, b.size()[0] - a.size()[0]), value=0)
print(a)

This extends the first and second dimension of the a tensor with zeros to match the b tensor.

You can then concat them:

c = torch.cat((a, b), dim=0)