Is argmax() and argmin() missing for vector operations?

[Johannes Czech]

Hello Klaus Iglberger,

thank your for creating this amazing library!

I was wondering if there is a.argmax() and .argmin() function for Vector Operations.

I could only see a solution with .min() and .max() in combination with .find() or iterating through the vector manually,

but this seems only to work with the CompressedVector type and is presumable not an ideal approach.

[https://bitbucket.org/blaze-lib/blaze/wiki/Vector](https://bitbucket.org/blaze-lib/blaze/wiki/Vector) Operations#!min-max

https://bitbucket.org/blaze-lib/blaze/wiki/Vector Operations#!element-lookup

‌

Best regards,

~Johannes Czech

[Johannes Czech]

I created the following method for resolving this issue:

template< typename VT  // Type of the dense vector
        , bool TF >    // Transpose flag
inline decltype(auto) argmax( const DynamicVector<VT,TF>& dv ) {
    size_t idx = 0;      // set the first index as the hypothesis for the armgax index
    VT cur_max = dv[0];  // initialize the current maximum value

    // iterate through the vector and update idx and cur_max accordingly
    for( size_t i=1UL; i<dv.size(); ++i ) {
        if (dv[i] > cur_max) {
            idx = i;
            cur_max = dv[i];
        }
    }
    return idx;
}

and tested it against the Xtensor 0.20.8 and Eigen library 3.3.7 on an identical 80 element float vector for 1e7 iterations:

I'm a bit surprised that this implementation is faster on my system than Eigen's eigen_vec.maxCoeff(&index) as well as XTensor's xt::argmax(xtensor_vec) even when using blaze::DynamicVector instead of blaze::StaticVector:

argmax xtensor 4
argmax blaze_vec 4
argmax eigen_vec 4

Elapsed time xtensor: 1042ms
Elapsed time blaze: 886ms
Elapsed time eigen: 909ms

This functionality might also be useful for others and I propose to include this method or an improved version of it to the blaze library.

[Klaus Iglberger]

Hi Johannes!

Thanks a lot for the proposal. This indeed sounds like a reasonable addition to Blaze.

Based on your example we assume you have the following algorithmic details in mind:

• Performs a linear search and returns the index of the first smallest (argmin()) or first largest (argmax()) element;
• Returns 0 in case of an empty vector.

We will try to integrate these two functions into Blaze as soon as possible. Thanks again for the proposal,

Best regards,

Klaus!

[Johannes Czech]

Based on your example we assume you have the following algorithmic details in mind:

+ Performs a linear search and returns the index of the first smallest (argmin()) or first largest (argmax()) element;

+ Returns 0 in case of an empty vector.

Thanks, that’s exactly what I had in mind.

[Klaus Iglberger]

Summary

The feature has been implemented, tested, optimized, and documented as required. It is immediately available via cloning the Blaze repository and will be officially released in Blaze 3.6.

Argmin

The argmin() function returns the index of the first smallest element of the given dense vector. argmin() can only be used for element types that support the smaller-than relationship. In case the given vector currently has a size of 0, the returned index is 0.

blaze::DynamicVector<int> a{ 1, -2, 3, 0 };
const size_t minindex = argmin( a );  // Results in 1

Argmax

The argmax() function returns the index of the first largest element of the given dense vector. argmax() can only be used for element types that support the smaller-than relationship. In case the given vector currently has a size of 0, the returned index is 0.

blaze::DynamicVector<int> a{ 1, -2, 3, 0 };
const size_t maxindex = argmax( a );  // Results in 2

[Johannes Czech]

Great work @Klaus Iglberger !

I’m excited for the new release Blaze 3.6.