Feature to Export the Pipeline/Model as pickle file

[Ezzaldin97]

Hi,

I am trying to use TPOT to automate the training, but one of the problems that I've found is that there is no way to maintain reproducibility by setting the random state to all components in the pipeline that uses random state and exporting it directly as a pickle file.

[perib]

you could edit the configuration dictionary in the config file to manually set the random state variable for all estimators that take it. We have a new version of tpot, TPOT2, which does that for you as well. Just have to pass random_state to the TPOTClassifier/TPOTEstimator/TPOTRegressor. https://github.com/EpistasisLab/tpot2

you can also pickle the final pipelines manually. (I prefer to use dill to pickle). The final pipeline is found in the "fitted_pipeline_" attribute of the estimator.