[QST] How to use norm ops with repmat?

[HugoPhibbs]

Hi,

I'm just trying to implement a function to normalise a dataset along the rows

I'm running the below code:

template <typename T>
inline matx::tensor_t<T, 2> normaliseDatasetMatX(matx::tensor_t<T, 2> X) {
    // Normalise along the rows
    auto rowNorms_op = matx::vector_norm(X, {1}, matx::NormOrder::L2);
    auto res = matx::make_tensor<float>({X.Shape()[0], X.Shape()[1]});

    (res = X / matx::repmat(rowNorms_op, {1, X.Shape()[1]})).run();
    return res;
}

With input:

float X_data[] = {
        1.0f, 2.0f, 3.0f,
        4.0f, 5.0f, 6.0f,
        7.0f, 8.0f, 9.0f
}; // Row-major order

auto *X_d = GsDBSCAN::utils::copyHostToDevice(X_data, 3*3);
auto X = matx::make_tensor<float>(X_d, {3, 3});
auto XNorm = GsDBSCAN::projections::normaliseDatasetMatX(X);

And I'm getting the error:

/usr/local/include/matx/operators/repmat.h(186): error: no instance of constructor "matx::detail::RepMatOp<T1, DIM>::RepMatOp [with T1=matx::detail::NormOp<matx::tensor_t<float, 2, matx::basic_storage<matx::raw_pointer_buffer<float, matx::matx_allocator<float>>>, matx::tensor_desc_t<cuda::std::__4::array<matx::index_t, 2UL>, cuda::std::__4::array<matx::index_t, 2UL>, 2>>, matx::detail::NormTypeVector>, DIM=1]" matches the argument list
            argument types are: (matx::detail::NormOp<matx::tensor_t<float, 2, matx::basic_storage<matx::raw_pointer_buffer<float, matx::matx_allocator<float>>>, matx::tensor_desc_t<cuda::std::__4::array<matx::index_t, 2UL>, cuda::std::__4::array<matx::index_t, 2UL>, 2>>, matx::detail::NormTypeVector>, cuda::std::__4::array<matx::index_t, 2UL>)
        return detail::RepMatOp<T1, T1::Rank()>(t, detail::to_array(reps));

I was not sure if this was a template issue, or if I'm legit doing something wrong.

Thanks

[cliffburdick]

Hi @HugoPhibbs, this code is invalid because rowNorms_op is a 1D operator, but the repmat call is passing in two values for the dimensions. Something like this should work:

    auto rowNorms_op = matx::vector_norm(X, {1}, matx::NormOrder::L2);
   auto rowNorms_op_2 = matx::clone(rowNorms_op, {1, matx::matxKeepDim});
    auto res = matx::make_tensor<float>({X.Shape()[0], X.Shape()[1]});

    (res = X / matx::repmat(rowNorms_op_2, {1, X.Shape()[1]})).run();

[HugoPhibbs]

Hi @cliffburdick

I implemented what you wrote, but now I'm getting this error:

/usr/local/include/matx/operators/norm.h(89): error: no instance of overloaded function "cuda::std::get" matches the argument list
            argument types are: (std::tuple<matx::detail::tensor_impl_t<float, 1, matx::tensor_desc_t<cuda::std::__4::array<long long, 1UL>, cuda::std::__4::array<long long, 1UL>, 1>>>)
          norm_impl<NormType>(cuda::std::get<0>(out), a_, order_, ex);
                              ^

I checked the docs and all and it seems that I'm calling vector_norm correctly

[cliffburdick]

Hi @HugoPhibbs, have you updated to one of the latest commits? Specifically this one may fix it:

https://github.com/NVIDIA/MatX/commit/2186ba01c95ea90fab9890cfab2f3ba12dcec20a

[HugoPhibbs]

Hi @cliffburdick

Yep pulled and rebuilt, but don't think it fixed it. Still getting the same error

[luitjens]

@HugoPhibbs it looks to me like you are using repmat where clone would generally be used. Repmat is used for repeating a matrix across some number of dims. It does not increase the rank. If you want to increase the rank you should use the clone<newRank> operator.

[cliffburdick]

@HugoPhibbs this might be closer to what you want if I'm understanding your intent:

    auto rowNorms_op = matx::vector_norm(X, {1}, matx::NormOrder::L2);
    auto rowNorms_op_2 = matx::clone<2>(rowNorms_op, {matx::matxKeepDim, X.Shape(1)});
    auto res = matx::make_tensor<float>(X.Shape());

    (res = X / rowNorms_op_2).run();

This will expand your norm result from a 1D back to the same shape as X. Then you can divide directly by it for an element-wise divide.