Tensor.std(0) returns a scalar instead of a vector

C++
1

I’m a bit new to the C++ API, and I haven’t seen documentation for this behavior anywhere. Here is a minimal example to reproduce.

float data[] = { 1, 2, 3,
                       4, 5, 6 };
torch::Tensor f = torch::from_blob(data, {2, 3});
std::cout << f.mean(0) << std::endl;
// yields a 1-d tensor as expected.
// 2.5000
// 3.5000
// 4.5000
// [ CPUFloatType{3} ]
std::cout << f.std(0) << std::endl;
// yields a scalar, unexpected.
// 1.70783
// [ CPUFloatType{} ]
2

The argument will be interpreted as the unbiased argument, if a single argument is passed, which is not intuitive.
You could use f.std(0, true) as a workaround.
Could you also create an issue on GitHub and report it there, so that this behavior might be improved or at least the documentation added? :slight_smile:

3

Hi @ptrblck thanks for the answer. Here is the issue: https://github.com/pytorch/pytorch/issues/40287. Let me know if it needs to be revised, it is my first time creating one.

I also noticed that f.std(true, 0) collapses dimension 1, which is unintuitive, and also f.std(true, 1) collapses dimension 1 still but uses Bessel correction. All in all kinda weird behavior.

I’m actually not blocked on this as it’s a pretty easy to implement function (for my specific case), just pointing it out in case it is interesting to you guys.

4

Thanks for creating the issue! I’ve answered and it seems it worked for you. :slight_smile:
Just for the sake of completeness, I had a typo in my previous code, so posting the code snippet here again:

from torch.utils import cpp_extension


cpp_source = """
void my_fun(torch::Tensor x)
{
    std::cout << "x.std(0) " << x.std(0) << ", x.std(0, true) " << x.std(0, true) << std::endl;
    std::cout << "at::std(x, 0) " << at::std(x, 0) << ", at::std(x, 0, true) " << at::std(x, 0, true) << std::endl;
    return;
}
"""

module = torch.utils.cpp_extension.load_inline(
    name="std_test_extension",
    cpp_sources=cpp_source,
    functions="my_fun",
    verbose=True,
)

x = torch.randn(2, 3)
module.my_fun(x)

print('python ', x.std(0))
1 Like