L2 normalisation via f.normalize dim variable

I am quite new to pytorch and I am looking to apply L2 normalisation to two types of tensors, but I am npot totally sure what I am doing is correct:

[1]. type 1 (in the forward function) has shape torch.Size([2, 128]) and I would like to normalise each tensor (L2 norm).

for this case, I do:
F.normalize(tensor_variable, p=2, dim=1)
Is this the correct way to do it? is there any check I can perform to know that the vectors have been L2 normed?

[2] type2 has shape ([128]): for this I just do:
F.normalize(tensor_variable, p=2, dim=0)
Is this the correct way to go about it?

I am not able to find the doc link to this F.normalize function and I am having to take a guess at the dimension :frowning:

Hi John!

Yes, your use of normalize() is correct.

Just compute the (square of the) norm:

(torch.nn.functional.normalize (tensor_variable, p = 2, dim = 1)**2).sum (dim = 1)

and check that you get 1 for each row of tensor_input.

You can find the documentation here: normalize.


K. Frank


Hi Frank,

FIRSTLY, thank you very much for your reply. Your answers certainly helped me. I have one follow up question: torch.nn.functional.normalize and F.Normalize are presumably the same function?

thank you.

Hi John!

First I think you have a minor typo – I think you mean F.normalize
with a lower-case n.

There is no F.normalize, strictly speaking, in pytorch. Rather, it is
common practice to:

import torch.nn.functional as F

so that you can save a little typing and write F.normalize instead
of torch.nn.functional.normalize. So, yes, they are the same

(There is a torchvision.transforms.Normalize – with an upper-case
N, and a class, rather than a function – but that’s not what you’re talking about.)


K. Frank

Thank you again Frank for the patient explanation. REALLY appreciate your time.