XGBoost 2.0.0

[thesps]

The XGBoost converter is not working for XGBoost 2.0.0. This issue is for tracking the development of the fix.

[thesps]

Brought to my attention by @brightneedle in #41

[thesps]

This xgb_2 branch has the first pass of fixing conversion for XGBoost >= 2.0.0. It parses models correctly but the numerical output of the converted conifer model mismatches the original xgboost model worse than XGBoost 1.X.Y models. This can be seen in the xgboost converter test.

[fstrug]

I am running into xgboost conifer conversion output shape mismatch in XGBoost == 1.7.6 and conifer == 1.6 and using the xilinxhls backend. The decision function of the converted conifer model is not returning an ndarray of shape(n_samples, n_classes) if n_classes = 2 appropriately whereas XGBClassifier.predict_proba() does. Is this a related bug?

For context, I am using the hls4ml bdt tutorial as a framework for a course project. In the tutorial n_classes = 5 and the output shape of the converted conifer model matches the original XGBoost model shape(n_samples, 5) . For my project, I need to restrict the domain of`y_train = {[0],[1],[2],[3],[4]} to y_train = {[0], [1]} from the example. I will move to treating this as a standard binary classification problem y_train = {0,1}, but is this unexpected behavior?