How to recreate the model using the best suggested model settings?

[vinay-k12]

I ran the estimator and found the following settings as most suitable for my data:

Best hyperparmeter config: {'n_estimators': 138, 'max_leaves': 37, 'min_child_weight': 0.13424014283416638, 'learning_rate': 0.14945689534042633, 'subsample': 0.8690041677979473, 'colsample_bylevel': 0.8428026817665201, 'colsample_bytree': 0.8817575629900145, 'reg_alpha': 0.02106728247448324, 'reg_lambda': 39.44246914091348, 'FLAML_sample_size': 393296}
Best r2 on validation data: 0.9992
Training duration of best run: 43.72 s
XGBClassifier(colsample_bylevel=0.8428026817665201,
colsample_bytree=0.8817575629900145, grow_policy='lossguide',
learning_rate=0.14945689534042633, max_depth=0, max_leaves=37,
min_child_weight=0.13424014283416638, n_estimators=138, n_jobs=-1,
reg_alpha=0.02106728247448324, reg_lambda=39.44246914091348,
subsample=0.8690041677979473, tree_method='hist',
use_label_encoder=False, verbosity=0)

And when I'm looking to retrain the model, I would like to use the same settings. I tried to define them using "custom_hp". But I'm having trouble in defining this. Is there any easy way to pick up these hyper-parameters and train on data?

[qingyun-wu]

Hi @vinay-k12,
Thank you for your question. There are two possible ways you can retain the model using the best config:

1. use retrain_from_log if you logged the search trajectory.
   It allows you to retain a model using the best config in your log file.
2. Set your best config to be thestarting_points without specifying max_iter and time_budget . Find an example for using starting_points here.

Please let me know if you need any further help!

[qingyun-wu]

Hi @vinay-k12 ,

In approach option 2, you also need to specify the estimator_list to only include the learner you want to use, e.g., 'xgboost' in your example. Please find an example showing how to use option 2 to retrain the model: https://github.com/qingyun-wu/FLAML/blob/qw-test/notebook/examples/retrain.ipynb

Thank you, and please let me know if you need further help!

[vinay-k12]

Awesome. That worked.

[sonichi]

Hi @vinay-k12 ,

In approach option 2, you also need to specify the estimator_list to only include the learner you want to use, e.g., 'xgboost' in your example. Please find an example showing how to use option 2 to retrain the model: https://github.com/qingyun-wu/FLAML/blob/qw-test/notebook/examples/retrain.ipynb

Thank you, and please let me know if you need further help!

Could you please make the documentation more clear about retraining, model serialization etc.? Such questions are reoccurring.