How should we concatenate multiple tensors vertically into one tensor in torch?

vision
1

I am wondering how should we concatenate tensor with different shapes in the torch. I tried torch.cat for that, but I got type erorr as follow:

print(tensor_1.shape)
print(tensor_2.shape)

new_tensor = torch.cat((tensor_1, tensor_2),1)

but I got the following error:

shape of tensor_1 (?, 30, 30, 128)
shape of tensor_2 (?, 26, 26, 128)
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-20-492132d221f7> in <module>()
      6 print("shape of tensor_1", tensor_1.shape)
      7 print("shape of tensor_2", tensor_2.shape)
----> 8 torch.cat((tensor_1, tensor_2),1)

TypeError: expected Tensor as element 0 in argument 0, but got Tensor

I think in Conv1D, concatenation is simple, in Conv2D, it is not known to me how to make concatenation. Does anyone know how to do this in torch? any idea?

I was thinking getting concatenated tensor shape as (56,56,128) is not a good idea or wrong in math. Is there any workaround to do concatenation on those? any thoughts?

desired output

I am not sure how multiple tensor can be concatenated, so I am okay with output as tensor with shape of (W, H, C). How should we do this in torch? any idea?

2

What dimensions do you want the concatenated tensor to have?

3

I understand concatenation in 1D, for instance, we can concatenate (12,20), (6,20) as (18, 20), but not sure what would be the correct dimension of the concatenated output tensor in 2D. I was thinking output tensor shape could be (56,56, 128) based on the above example, but not sure that’s the right intuition to do so. I am open to community suggestions at the best. I think I need a new tensor that could have the shape of (Width, Height, Channel). Do you have any possible thoughts? Thanks!

4

That works because 12*20 + 6*20 = 18 * 20. But 30*30*128 + 26*26*128 != 56*56*128 so it won’t work.

5

right, that’s why getting concatenated tensor shape as (56,56,128) is not good idea or wrong. What can we do about concatenating multiple tensors? Do we need to reshape it back to (30,30, 128), (26,26,128) into (W, H) shape then concatenate then reshape it back to (W, H, C)? Do you have any possible thoughts? Thanks!

6

Suppose you have a 5x5 rectangle and a 10x10 rectangle. You want to concatenate these two rectangles so that the result is a new rectangle. How do you do it?

7

I am not sure but I think I might do flatten the array and do padding then do concatenation. I really don’t know how should I concatenate 2dim tensor vertically. Do you have any idea or coding attempt that might do some trick on that?