Removing the wrong gamma-deviance metric

[Roffild]

Corrected version, but without checking the condition label>0 && pred>0:

struct EvalGammaDeviance {
  const char *Name() const {
    return "gamma-deviance";
  }

  XGBOOST_DEVICE bst_float EvalRow(bst_float label, bst_float pred) const {
    const bst_float eps = 1e-16f;
    return 2 * (std::log(pred/(label+eps)) + label/(pred+eps) - 1);
  }
  static bst_float GetFinal(bst_float esum, bst_float wsum) {
    return wsum == 0 ? esum : esum / wsum;
  }
};

[Roffild]

Approved

[trivialfis]

Good catch! But could you please replace it with the correct one instead? The metrics defined in xgboost have many use like distributed setting, language bindings other than Python. Also 3 party implementation cannot be saved.

[Roffild]

This is part of Tweedie. And everything about Tweedie needs to be redone. I don't understand what formulas were used for Tweedie in XGBoost.

I am guided by these sources:
https://en.wikipedia.org/wiki/Tweedie_distribution#The_Tweedie_deviance

sklearn/_loss/glm_distribution.py