Concatenate torch tensor along given dimension

1

In tensorflow you can do something like this

third_tensor= tf.concat(0, [first_tensor, second_tensor])

so if first_tensor and second_tensor would be of size [5, 32,32], first dimension would be batch size, the tensor third_tensor would be of size [10, 32, 32], containing the above two, stacked on top of each other.

How can I do this with torch variables? Or ar least with torch tensors?

5 Likes
2

Check out torch.cat. It works both on torch tensors and variables

for example:
third_tensor = torch.cat((first_tensor, second_tensor), 0)

39 Likes
3

Thanks, I wrote my own function that transforms it to numpy.array and then uses numpy.concatenate, then transforms back… not sure how slow this is though.

4

Cool! but you won’t be able to backpropagate by converting to numpy and back (if you so desire). But torch.cat would allow torch autograd to do its “magic” :slight_smile: . Not sure about the performance issue between the two, but I guess the native way of concatenating would be faster.

9 Likes
5

Thanks, switched to torch.cat.

6

@Bixqu be also mindful of this property x = torch.cat([y, z]).view(-1)
.view is a special parameter that can change the shape of the tensor without changing its contents.