.size() vs .shape, which one should be used?

[freud14]

Hi,

This is not an issue per se, but more of a question of best practices. I notice that there a shape attribute and a size function that both return the size/shape of a tensor. I also notice that .shape is not documented, but according to various release notes, Pytorch tries to be as close as possible to numpy. So, knowing this, is .shape really supported or just an undocumented functionality that could disappear in the next version?

Thank you.

[fmassa]

Hi,
.shape is an alias for .size(), and was added to more closely match numpy, see #1983

[apaszke]

Yes, .shape is going to remain supported. Both ways are fine

[freud14]

Thank you.

[samuela]

FWIW this redundancy has always bugged me. For some reason x.shape feels more internal than x.size() but using x.size() also just doesn't look right when doing things like x.size()[-2:]. Still I don't want to be inconsistent and use both.

[nyck33]

what about .size with no ( ) following it like here:

Found here: http://nlp.seas.harvard.edu/2018/04/03/attention.html

[munael]

what about .size with no ( ) following it like here:

Found here: http://nlp.seas.harvard.edu/2018/04/03/attention.html

That size attribute is custom to the custom layers, from what I can see. tensor.size() is a tensor method.

[Risingabhi]

.size() method returns total elements in a dataframe , for eg shape of a tensor might be (10,3) , here total elements in tensor would be returned by .size() = 10X3 = 30 elements!!

[freud14]

.size() method returns total elements in a dataframe , for eg shape of a tensor might be (10,3) , here total elements in tensor would be returned by .size() = 10X3 = 30 elements!!

@Risingabhi Nope, that's not how it works in PyTorch:

[Risingabhi]

.size() method returns total elements in a dataframe , for eg shape of a tensor might be (10,3) , here total elements in tensor would be returned by .size() = 10X3 = 30 elements!!

@Risingabhi Nope, that's not how it works in PyTorch:

yes, that's the case in pytorch

[flpgrz]

.size() method returns total elements in a dataframe

That's what torch.numel(x) does.

[davidgilbertson]

If you're like me and looking to pick one or the other, some common use cases compared:

t = torch.arange(60).view(3, 4, 5)

t.shape == t.size()          # just get the size/shape
t.shape[1] == t.size(1)      # get one element from the size/shape
t.shape[1:] == t.size()[1:]  # get several elements from the size/shape

To my eyes, shape is the better all-rounder. It returns a different type of object to NumPy (Size) which is a bit odd, but Size extends tuple so I can't think of a scenario where the behaviour would be any different.

[Vagish-Kumar]

Hi, .shape is an alias for .size(), and was added to more closely match NumPy, see #1983

[tornikeo]

Is there any difference in behavior for torch.compile or onnx export? I'm pretty sure I've seen some weird stuff happen with .shape, while .size seemed to work well @apaszke