Some micro optimizations for performance are possible.
For.ex. in the PerLabelL1 property ( see link) we divide all array's elements onto the single _sumWeight value, in the loop. If we compute 1/_sumWeight division before a loop, we can multiply all element on computed value, that speeds up an execution time 4 times (see example below and in the attach).
Source code / logs
private static int _sumWeights = 14; public static double[] OriginDivideArray(this double[] _l1Loss)
{
var res = new double[_l1Loss.Length];
if (_sumWeights == 0)
return res;
for (int i = 0; i < _l1Loss.Length; i++)
res[i] = _l1Loss[i] / _sumWeights;
return res;
}
public static double[] OptimizedDivideArray(this double[] _l1Loss)
{
var res = new double[_l1Loss.Length];
if (_sumWeights == 0)
return res;
var revSumWeights = 1 / _sumWeights; // Cache division result here
for (int i = 0; i < _l1Loss.Length; i++)
res[i] = _l1Loss[i] * revSumWeights; // Update division to mult here
return res;
}
The question is: do we need these optimizations? On a 1000 elements array it saves us just some microseconds.
PointsToImprove.txt
Program.zip
Some micro optimizations for performance are possible.
For.ex. in the PerLabelL1 property ( see link) we divide all array's elements onto the single _sumWeight value, in the loop. If we compute 1/_sumWeight division before a loop, we can multiply all element on computed value, that speeds up an execution time 4 times (see example below and in the attach).
Source code / logs
The question is: do we need these optimizations? On a 1000 elements array it saves us just some microseconds.
PointsToImprove.txt
Program.zip