Why I got different results using automl.predict() and automl.model.predict()?

[azurechen97]

I trained a multi-classification LightGBM model via FLAML. But when I tested the model on the test set, I got completely different results on the same data by using automl.predict() and automl.model.predict(). What could be the problem?

[thinkall]

Hi @azurechen97, There may be some feature transformation in AutoML, in that case, automl.predict() will be different from automl.model.predict(). The code is here. You can pickle the trained automl instance and reuse it later. If you just want to use the model.predict (such as using mlflow autologging), you'll need to pickle the feature_transformer and label_transformer and use it together with the saved model.

[azurechen97]

Thank you! The problem is solved perfectly.