sklearn-serialize

JSON serialization for scikit-learn pipelines and the Python/NumPy/SciPy/pandas types that appear inside them. The goal is lossless round-trips: json_to_data(data_to_json(obj)) reproduces the original object, including fitted model state.

from sklearn_serialize import data_to_json, json_to_data

json_str = data_to_json(fitted_pipeline)
restored  = json_to_data(json_str)

restored.predict(X)  # identical output to the original

Installation

pip install sklearn-serialize

Supported types

• sklearn: Pipeline, FeatureUnion, ColumnTransformer, and any BaseEstimator subclass
• NumPy: ndarray, scalar integer/float/complex types, datetime64, dtype, Generator, RandomState
• SciPy: sparse matrices (csr, csc, coo, lil, dok)
• pandas: Series, DataFrame
• polars: Series, DataFrame (optional — requires polars to be installed)
• Python: tuple, set, frozenset, bytes, bytearray, slice, complex, OrderedDict, namedtuple, datetime, date

Custom estimators

Custom estimators work out of the box as long as their class is importable at deserialization time. Call trust_module once at startup to allow deserialization from your package:

from sklearn_serialize import trust_module

trust_module("my_package.transformers")

The argument is a module prefix — "my_package" covers my_package.transformers, my_package.pipelines, etc. Only exact matches and dotted submodules are allowed; "my_pack" does not cover "my_package".

json_to_data will raise ValueError if it encounters a class from an untrusted module. This prevents arbitrary code execution when deserializing JSON from untrusted sources. Only call json_to_data on JSON you produced yourself or received from a trusted source.

The default trusted set covers sklearn, numpy, scipy, pandas, builtins, and sklearn_serialize.

To trust modules globally without calling trust_module in every script, create ~/.sklearnserialize:

[trusted_modules]
my_package
polars

One module prefix per line. Blank lines and lines starting with # are ignored. This file is loaded once at import time.