Activity

1. tdegeus commented on Nov 27, 2019

   

   tdegeus

   on Nov 27, 2019

   Member

   More actions

   In terms of your problem, could you try with

   xt::xtensor<int,1> a = xt::arange<int>(2);

   forcing the generator to be evaluated resulting in a real array?

   Then, it would be really great to have xt::tile and xt::roll as well! Would you be willing to convert your efforts in one or two PRs?

   As to the xt::roll, my intuition would be that it would be best to do it with minimal allocation. xt::pad has to allocate the full bigger matrix. In contrast, xt::roll could potentially get away with minimal temporary storage. Though you or others may have a different intuition.
2. tort-dla-psa commented on Nov 27, 2019

   

   tort-dla-psa

   on Nov 27, 2019

   Author

   More actions

   It kinda works with real xarray/xtensor passing, but I have some strange results:

   template<class Expr>
   inline void tile(xt::xexpression<Expr> &data_expr, std::vector<size_t> ns){
       Expr &data = data_expr.derived_cast();
       data = xt::pad(data, ns, xt::pad_mode::reflect);
   }
    
   int main(){
      xt::xarray<int> a = xt::arange<int>(2);
      std::cout<<a<<"\n";
      std::vector<size_t> tile_vec = {2, 2};
      tile(a, tile_vec);
      std::cout<<a<<"\n";
   
      xt::xarray<int> b = xt::arange<int>(2);
      std::cout<<b<<"\n";
      b = xt::pad(b, {2,2}, xt::pad_mode::reflect);
      std::cout<<b<<"\n";
   }

   desired output:

   {0,1}
   {0,1,0,1,0,1}
   {0,1}
   {0,1,0,1,0,1}

   real output:

   {0,1}
   {1,0,1}
   {0,1}
   {1,0,1}
3. tdegeus commented on Nov 28, 2019

   

   tdegeus

   on Nov 28, 2019

   Member

   More actions

   @tort-dla-psa At the level of xt::pad you've reached the limitation of the current implementation (that can clearly be improved!). For the moment xt::pad is not recursive, i.e. all pad modes that are not constant fail when the pad width is bigger than the shape along the relevant axis (or in the case of reflect the shape minus one).

   Upon enabling assertions:

   #define XTENSOR_ENABLE_ASSERT
   
   #include <xtensor/xarray.hpp>
   #include <xtensor/xpad.hpp>
   #include <xtensor/xio.hpp>
   
   int main()
   {
     xt::xarray<int> b = xt::arange<int>(2);
     std::cout << b << std::endl;
   
     b = xt::pad(b, {2, 2}, xt::pad_mode::reflect);
     std::cout << b << std::endl;
   
     return 0;
   }

   You will indeed find

   libc++abi.dylib: terminating with uncaught exception of type std::runtime_error: .../include/xtensor/xpad.hpp:138: assertion failed (nb <= out.shape()[axis] - 1)
   

   To clarify, currently:

   b = xt::pad(b, {2, 2}, xt::pad_mode::periodic); // works
   b = xt::pad(b, {3, 3}, xt::pad_mode::periodic); // throws
   b = xt::pad(b, {3, 3}, xt::pad_mode::constant); // works

   I think that the quickest fix for this would be to make xt::pad recursive. Obviously with the disadvantage of re-allocation more often than needed.

   We're very much open to improvements. In this regard I mention:

   • xt::pad performance issue #1806
   • Equivalent of 'np.tile(...)' missing #1557
   • No equivalent to numpy.roll in xtensor #1766
   • Implement an equivalent to numpy.roll #1823
   • Assemble new container in xpad #1808

   With respect to you original issue of implementing xt::tile, I think that its stand-alone implementation could be faster than a generalisation of xt::pad (though the latter would obviously much more welcome).
4. JohanMabille

   closed this as completedin #1888on Jan 13, 2020