Softmax implementation in OneVersusAllModelParameters returns NaN even for scores of ~90.0f

[rauhs]

version: 0.11

This implementation:

machinelearning/src/Microsoft.ML.StandardTrainers/Standard/MulticlassClassification/OneVersusAllTrainer.cs

Lines 741 to 752 in 5163413

	private void NormalizeSoftmax(float[] scores, int count)
	{
	float sum = 0;
	for (int i = 0; i < count; i++)
	{
	scores[i] = (float)Math.Exp(scores[i]);
	sum += scores[i];
	}
	for (int i = 0; i < count; i++)
	scores[i] = scores[i] / sum;
	}

will result in NaN as the output even for moderate scores of around 90f:

      var foo = new[] { 90f, 2, -2f };
      NormalizeSoftmax(foo, 3);

Seen very easily in the debugger (the float cast will result in Inf). I'm getting back a score of roughly 102 for some of my data sets and models. Though they're extremely rare.