Saving DMatrix after building quantile cuts and historgram

#1

I have a binary DMatrix I saved in advance for which I try to do hyperopt tuning using optuna.
In each optuna round I run xgb.train(params, D, ...).
In every round, xgboost is building quantile cut, and sums to the save number of hist bins.
It seems as if this process is always the same.
Also, in my specific case, this process takes more than 50% of the total model time. Is there a way to do the quantile building one time, and then pass it on to xgb.train?

#2

It is not yet possible. I’ve filed a feature request at https://github.com/dmlc/xgboost/issues/6355.

1 Like