CODEWITHSUNDEEP
luatorch
Sahil
Sahil

Reputation: 1386

Can tensors be concatenated without losing reference in torch (lua)?

th> x = torch.rand(2)
                                                                  [0.4943s] 
th> y = torch.rand(2)
                                                                  [0.0001s] 
th> x
 0.6115
 0.4986
[torch.DoubleTensor of size 2]
                                                                  [0.0002s] 
th> z = torch.cat(x,y)
                                                                  [0.0003s] 
th> z
 0.6115
 0.4986
 0.5171
 0.1785
[torch.DoubleTensor of size 4]

                                                                  [0.0001s] 
th> z[1] = 3
                                                                  [0.0001s] 
th> z[1]
 3  
                                                                  [0.0001s] 
th> x[1]
 0.61146148154512

Modifying z does not modify x. Is there any way of concatenating x and y such that modifying z does modify x?

Upvotes: 3

Views: 300

Answers (1)

lejlot
lejlot

Reputation: 66850

You can achieve this kind of behaviour, but the other way around. You should start with a bigger tensor, your main "storage", and then you can create subtensors, which will share internal state.

See in particular :sub method from torch (following code sample taken from Torch doc)

x = torch.Tensor(5, 6):zero()
> x
 0 0 0 0 0 0
 0 0 0 0 0 0
 0 0 0 0 0 0
 0 0 0 0 0 0
 0 0 0 0 0 0
[torch.DoubleTensor of dimension 5x6]

y = x:sub(2,4):fill(1) -- y is sub-tensor of x:
> y                    -- dimension 1 starts at index 2, ends at index 4
 1  1  1  1  1  1
 1  1  1  1  1  1
 1  1  1  1  1  1
[torch.DoubleTensor of dimension 3x6]

> x                    -- x has been modified!
 0  0  0  0  0  0
 1  1  1  1  1  1
 1  1  1  1  1  1
 1  1  1  1  1  1
 0  0  0  0  0  0
[torch.DoubleTensor of dimension 5x6]

As you can see you have y variable which is actually a part of x, and changing its values - change x's too. This is very generic way thus you can share multiple parts of tensor.

Thus in your case it would be something like

z = torch.Tensor(4):zero()
x = z:sub(1, 2)
y = z:sub(3, 4)
x[1] = 2
y[2] = 8
print(z)

prints 

 2
 0
 0
 8
[torch.DoubleTensor of size 4]

as desired.

Upvotes: 2

Related Questions