Export / save the best model

[simon-delisle]

How can I save the best model to reuse the automl.predict later without fitting again?

Thanks,

Simon

[sonichi]

One way to do it is to use pickle, as demonstrated in this notebook.

import pickle
with open('automl.pkl', 'wb') as f:
    pickle.dump(automl, f, pickle.HIGHEST_PROTOCOL)

[Shashi9503]

One way to do it is to use pickle, as demonstrated in this notebook.

import pickle
with open('automl.pkl', 'wb') as f:
    pickle.dump(automl, f, pickle.HIGHEST_PROTOCOL)

This is for saving the model. But how can I load it and reuse it ??

[qingyun-wu]

You can use pickle.load() to load the automl instance you saved and then predict without fitting again.

Code example assuming your model is dumped into file 'automl.pkl' and test data is X_test:

automl = pickle.load(open('automl.pkl', 'rb'))
y_pred = automl.predict(X_test)

[Shashi9503]

this worked, thanks

[eavidan]

is it possible to export the model to its native package?
meaning, using xgboost to load the model rather then flaml
One application would be to continue working on the model with its native package

[sonichi]

Yes. automl.model.estimator is the native package's estimator. For example, if it's a xgboost estimator, you can do automl.model.estimator.save_model(filename).

[eavidan]

Thanks @sonichi this is great. what about the preprocess?
I was looking at the code and it looks like each model implementation of BaseEstimator includes a _preprocess.
Is there any API I should you in order to invoke it, or use directly automl.model._preprocess()?

[sonichi]

There are two preprocessors: task-level and estimator-level. The former is automl._preprocess() and the latter is automl.model._preprocess(). The former should be called before the latter. Right now they are not exposed as user-facing APIs.
We can add some user-facing APIs in future version. If you have any ideas for that please feel free to suggest.

[eavidan]

thanks for the details @sonichi. I looked at the code and now I believe I understand better the flow of things.
my first thought is why not expose both as a user-facing APIs: automl.preprocess and automl.model.preprocess() just so we know these methods are hear to stay.
I agree that getting the entire preprocessing chained together through a single autonomous API would probably require some refactoring. will give it some thought and will try to suggest something.
for now, I will simply chain them together using flaml. thanks for all the info !

[sonichi]

Thanks. An issue has been created: #842. Let's make future discussions over there.