ghostjat / np Goto Github PK

In matlab/octave, I can define a m x n matrix (3 x 5 in this case) and multiply it by a n x 1 column vector (5 x 1 in this case). It yields a 3x1 column vector:

X = [1 1 1 1 1; 2 2 2 2 2; 3 3 3 3 3]
    X =

       1   1   1   1   1
       2   2   2   2   2
       3   3   3   3   3

octave:427> w = [1;2;3;4;5]
    w =

       1
       2
       3
       4
       5

octave:428> X * w
    ans =

       15
       30
       45

However, ghostjat cannot multiply a 3 x 5 matrix times a vector with size=5:

require __DIR__ . '/np/vendor/autoload.php';
use Np\matrix;
use Np\vector;

$x = Np\matrix::ar([
        [1,1,1,1,1],
        [2,2,2,2,2],
        [3,3,3,3,3]
]);

$w = Np\vector::ar([1, 2, 3, 4, 5]);

$p = $x->dot($w); // throws exception Mismatch Dimensions of given Objects! Obj-A col & Obj-B row amount need to be the same!
$p = $x->multiply($w); // throws exception Mismatch Dimensions of given Objects! Obj-A col & Obj-B row amount need to be the same!

It appears that the vector::sum function incorrectly calls blas::asum, which I believe returns the sum of absolute values instead of simply the sum of the values in a vector.

You can see the problem demonstrated with this short script:

<?php

require_once 'vendor/autoload.php';

Np\core\blas::$ffi_blas = FFI::load(__DIR__ . '/vendor/ghostjat/np/src/core/blas.h');


$v = Np\vector::ar([1, 2, 3]);
// this correctly returns 6
var_dump($v->sum());


$v = Np\vector::ar([-1, -2, -3]);
// this INCORRECTLY returns 6
var_dump($v->sum());

I don't know if there is a BLAS or LAPACK function optimized to return the correct value, but I suggest we modify the vector.php source code to change this:

     /**
     * The sum of the vector.
     * @return float
     */
    public function sum(): float {
        return blas::asum($this);
    }

to this:

    /**
     * The sum of the elements of the vector.
     * @return float
     */
    public function sum(): float {
        $sum = 0;
        for($i=0; $i<$this->ndim; $i++) {
                $sum += $this->data[$i];
        }
        return $sum;
    }

    /**
     * The sum of the absolute values of the elements of the vector.
     * @return float
     */
    public function sumAbs(): float {
        return blas::asum($this);
    }

I have attempted to run the sample PHP script in your README file and it encounters a fatal error:

PHP Fatal error: Uncaught FFI\Exception: Failed loading scope 'blas' in /Users/sneakyimp/np/vendor/ghostjat/np/src/core/blas.php:28
Stack trace:
#0 /Users/sneakyimp/np/vendor/ghostjat/np/src/core/blas.php(28): FFI::scope('blas')
#1 /Users/sneakyimp/np/vendor/ghostjat/np/src/core/blas.php(73): Np\core\blas::init()
#2 /Users/sneakyimp/np/vendor/ghostjat/np/src/linAlgb/linAlg.php(45): Np\core\blas::gemm(Object(Np\matrix), Object(Np\matrix), Object(Np\matrix))
#3 /Users/sneakyimp/np/vendor/ghostjat/np/src/linAlgb/linAlg.php(30): Np\matrix->dotMatrix(Object(Np\matrix))
#4 /Users/sneakyimp/np/foo.php(8): Np\matrix->dot(Object(Np\matrix))
#5 {main}
thrown in /Users/sneakyimp/np/vendor/ghostjat/np/src/core/blas.php on line 28

Fatal error: Uncaught FFI\Exception: Failed loading scope 'blas' in /Users/sneakyimp/np/vendor/ghostjat/np/src/core/blas.php:28
Stack trace:
#0 /Users/sneakyimp/np/vendor/ghostjat/np/src/core/blas.php(28): FFI::scope('blas')
#1 /Users/sneakyimp/np/vendor/ghostjat/np/src/core/blas.php(73): Np\core\blas::init()
#2 /Users/sneakyimp/np/vendor/ghostjat/np/src/linAlgb/linAlg.php(45): Np\core\blas::gemm(Object(Np\matrix), Object(Np\matrix), Object(Np\matrix))
#3 /Users/sneakyimp/np/vendor/ghostjat/np/src/linAlgb/linAlg.php(30): Np\matrix->dotMatrix(Object(Np\matrix))
#4 /Users/sneakyimp/np/foo.php(8): Np\matrix->dot(Object(Np\matrix))
#5 {main}
thrown in /Users/sneakyimp/np/vendor/ghostjat/np/src/core/blas.php on line 28

I have the FFI extension loaded in PHP. The FFI documentation is sorely incomplete so I'm not at all sure what this scope() call is supposed to do. The docs mention a #define statement, which you appear to have in blas.h:

#define FFI_SCOPE "blas"

EDIT: I'm running this script in PHP 8.1 on MacOs. I have the FFI extension loaded, and brew shows openblas and lapack are installed.

I have also confirmed the error with PHP 8.1 on Ubuntu 20.04 LTS.

The colAsVector method clearly has a problem. Some simple code to illustrate:

require __DIR__ . '/vendor/autoload.php';
use Np\matrix;
$v = matrix::ar([
	[1,2,3,4,5,6],
	[7,8,9,10,11,12]
]);
echo $v, "\n";
$shape = $v->getShape();
for($i=0; $i<$shape->n; $i++) {
	$vect = $v->colAsVector($i);
	echo $vect, "\n";
}

The output is clearly wrong, and shows the second item in each column drifting off from the correct value.

Np\matrix
1.000000  2.000000  3.000000  4.000000  5.000000  6.000000  
7.000000  8.000000  9.000000  10.000000  11.000000  12.000000  

Np\vector
1.000000  3.000000  

Np\vector
2.000000  4.000000  

Np\vector
3.000000  5.000000  

Np\vector
4.000000  6.000000  

Np\vector
5.000000  7.000000  

Np\vector
6.000000  8.000000  

I believe this modified version of the function may remedy the problem:

    /**
     * Return a col as vector from the matrix.
     * @param int $index
     * @return \Np\vector
     */
    public function colAsVector(int $index): vector {
        $vr = vector::factory($this->row);
        for ($i = 0; $i < $this->row; $i++) {
            $vr->data[$i] = $this->data[($i *  $this->col) + $index];
        }
        return $vr;
    }

The function linAlg::dot needs to have its return type changed to include float and possible int and/or other scalar types. It is possible to multiply two vectors and get a scalar result. E.g., this code in matlab/octave returns a scalar, and blas returns a float:

octave:25> m1 = [1 2 3]
m1 =

   1   2   3

octave:26> m2 = [1;2;3]
m2 =

   1
   2
   3

octave:27> m1 * m2
ans = 14

This code should return 14:

$m1 = Np\vector::ar([1, 2, 3]);
echo "$m1\n";
$m2 = Np\vector::ar([1, 2, 3]);
echo "$m2\n";

$v = $m1->dot($m2);
echo "$v\n";

but due to the narrow return type restriction in linAlgb::dot, it throws this error:

PHP Fatal error:  Uncaught TypeError: Np\vector::dot(): Return value must be of type Np\matrix|Np\vector, float returned in /Users/sneakyimp/Desktop/biz/machine-learning/np/vendor/ghostjat/np/src/linAlgb/linAlg.php:34

Simply modifying the linAlg::dot function as follows should fix this particular error:

    /**
     *
     * get dot product of m.m | m.v | v.v
     *
     * @param \Np\matrix|\Np\vector $d
     * @return matrix|vector|float
     */
    public function dot(matrix|vector $d): matrix|vector|float {
        if ($this instanceof matrix) {
            if ($d instanceof matrix) {
                return $this->dotMatrix($d);
            }
            return $this->dotVector($d);
        }
        return blas::dot($this, $d);
    }

The composer command you offer in the README file:

composer require ghostjat/np

Results in a complaint:

Could not find a version of package ghostjat/np matching your minimum-stability (stable). Require it with an explicit version constraint allowing its desired stability.

It may help to specify a slightly different composer require command:

composer require ghostjat/np:dev-main

where you can specify one of your branches (e.g., v0.0-alpha or np-0.0.1-alpha).