Repository navigation
problem calculating quantile on pytensor #286
Description
Hi,
The new added xt::quantile seems not to play nicely with pytensor.
For me, the following :
xt::pytensor<double, 2> arr
{{ 5.3, 4.3, 5.3 },
{ 4.2, 4.2, 5.2 },
{ 5.7, 4.7, 5.7 },
{ 2.3, 4.3, 2.3 }};
auto q = xt::quantile(arr, {0.3, 0.7}, 1);raises :
~/xtensor/include/xtensor/xsort.hpp:463:31: error: non-constant-expression cannot be narrowed from type 'xt::xcontainer<xt::pytensor<double, 2, xt::layout_type::dynamic>>::size_type' (aka 'unsigned long') to 'long' in initializer list [-Wc++11-narrowing]
R ev = R::from_shape({de.size()});
I am using xtl==0.7.5, xtensor==0.24.4, xtensor-python==0.26.1, pybind11==2.10.3.
Activity
JohanMabille commentedon Jan 31, 2023
Thanks for reporting. This is definitely a bug in xtensor.
ThibHlln commentedon Feb 17, 2023
Hi @AntoinePrv, do you think that this issue is related to xtensor-stack/xtensor#2651, and so would be fixed by xtensor-stack/xtensor#2652 too, or is this an unrelated problem?
AntoinePrv commentedon Feb 20, 2023
@ThibHlln I could not reproduce this. Is this a compiler error? If so which one and on which platform?
I does not seem related to xtensor-stack/xtensor#2651 but perhaps since xtensor-stack/xtensor#2651 is a big refactor it could solve it. Do you have the time to try?
ThibHlln commentedon Feb 21, 2023
Hi @AntoinePrv,
Yes, this is a compiler error I get
with Clang on macOS, and alsowith MSVC on Windows (see error message below).So, I tried with your PR xtensor-stack/xtensor#2652, and even though it does not look like you touched the line the compiler is complaining about (i.e. http://github.com/xtensor-stack/xtensor/blob/master/include/xtensor/xsort.hpp#L463), I don't get the error anymore on Windows with MSVC
(I will check on macOS with Clang later on), so maybe the type of the variable has changed and the narrowing conversion is not necessary anymore?AntoinePrv commentedon Feb 22, 2023
Basically in
xt::partition(and other places), a tensor gets created withR::from_shape({ de.size() });http://github.com/xtensor-stack/xtensor/blob/master/include/xtensor/xsort.hpp#L463
For
xt::pytensorsize_typeis unsigned (derived frombuffer_adaptor), butshape_type(the argument offrom_shape) isstd::array<npy_intp, N>hence the narrowing conversion.A fix there would be simple enough but let's see if it is still present with xtensor-stack/xtensor#2652.
IMHO the real issue here, and in other bugs we've caught, is that xtensor functions are not tested consistently with all types of containers (let alone view, expressions). Tracking in xtensor-stack/xtensor#2658.
ThibHlln commentedon Feb 22, 2023
OK, thank you for the explanation @AntoinePrv.
Even though my MWE provided above only failed with MSVC on Windows, in a more complex project, I had the same kind of failure with Clang on macOS. But once again, using xtensor-stack/xtensor#2652 fixed it, so I am looking forward to the PR merge now. :-)
stellarpower commentedon Mar 2, 2023
Just FYI, I hit the exact same error message when I didn't have the Python headers included as system headers (i.e.
-Ivs-isystem). I think something in there is redefining or otherwise interfering with the size of builtin types, which would make sense re interfacing with Numpy. So this could be a build setup thing, I just fixed it in CMake myself.