How to save sklearn model parameters in a json file?

Question

Suppose I have very simple machine learning model as follows:

from sklearn.datasets import load_diabetes
from sklearn.preprocessing import StandardScaler
from sklearn.pipeline import Pipeline
from sklearn.linear_model import RidgeCV
from sklearn.svm import LinearSVR
from sklearn.ensemble import RandomForestRegressor, StackingRegressor

X, y = load_diabetes(return_X_y=True)
estimators = [
    ('lr', RidgeCV()),
    ('svr', LinearSVR(random_state=42))
]
reg = StackingRegressor(
    estimators=estimators,
    final_estimator=RandomForestRegressor(n_estimators=10,
                                          random_state=42)
)
steps = [
    ("preprocessing", StandardScaler()),
    ("regression", reg)
]
pipe = Pipeline(steps)

What I would like to do is to store the whole model parameter as json file. By that I mean the following information is saved in json file.

Pipeline(steps=[('preprocessing', StandardScaler()),
            ('regression',
             StackingRegressor(estimators=[('lr',
                                            RidgeCV(alphas=array([ 0.1,  1. , 10. ]))),
                                           ('svr',
                                            LinearSVR(random_state=42))],
                               final_estimator=RandomForestRegressor(n_estimators=10,
                                                                     random_state=42)))])

When I use json.dumps(pipe) I face with error that Object of type Pipeline is not JSON serializable . Any idea how one can do that?

Answer 1

The output you've given is just the string representation, so str(pipe) will produce that. It has a bit of formatting for spacing and newlines, so perhaps not ideal for storage.

You could use pipe.get_params() to retrieve a dictionary of parameters. That will include all the default parameters unlike the string; you could use the private method _changed_params from sklearn.utils._pprint ( source ), but being private you'd need to be careful about breaking changes. There might be a more public method along the lines of pretty-printing and display of estimators that gives you what you need.