OpenModels

OpenModels is a flexible and extensible library for serializing and deserializing machine learning models. It's designed to support any serialization format through a plugin-based architecture, providing a safe and transparent solution for exporting and sharing predictive models.

Key Features

• Format Agnostic: Supports any serialization format through a plugin-based system.
• Extensible: Easily add support for new model types and serialization formats.
• Safe: Provides alternatives to potentially unsafe serialization methods like Pickle (see Security).
• Transparent: Supports human-readable formats for easy inspection of serialized models.

Installation

pip install openmodels

JSON and Pickle work out of the box. The MessagePack and YAML formats need their (optional) underlying library:

pip install openmodels[msgpack]  # format_name="msgpack"
pip install openmodels[yaml]     # format_name="yaml"

Quick Start

from openmodels import SerializationManager, SklearnSerializer
from sklearn.decomposition import PCA
from sklearn.datasets import make_classification

# Create and train a scikit-learn model
X, _ = make_classification(n_samples=1000, n_features=4, n_informative=2, n_redundant=0, random_state=0, shuffle=False)
model = PCA(n_components=2, random_state=0)
model.fit(X)

# Create a SerializationManager
manager = SerializationManager(SklearnSerializer())

# Serialize the model (default format is JSON)
serialized_model = manager.serialize(model)

# Deserialize the model
deserialized_model = manager.deserialize(serialized_model)

# Use the deserialized model
transformed_data = deserialized_model.transform(X[:5])
print(transformed_data)

Saving and Loading Models

OpenModels provides high-level save and load methods for convenient file I/O:

# Serialize and save a model to a file in JSON format
manager.save(model, "model.json", format_name="json")

# Load and deserialize a model from a file
loaded_model = manager.load("model.json", format_name="json")

Extensibility

OpenModels is designed to be easily extended with new serialization formats and model types.

Adding a New Format

To add a new serialization format, create a class that implements the FormatConverter protocol and register it with the FormatRegistry. is_binary tells SerializationManager.load() whether to open a saved file in binary or text mode:

from openmodels.protocols import FormatConverter
from openmodels.format_registry import FormatRegistry
from typing import Dict, Any

class TOMLConverter(FormatConverter):
    is_binary = False

    @staticmethod
    def serialize_to_format(data: Dict[str, Any]) -> str:
        import toml
        return toml.dumps(data)

    @staticmethod
    def deserialize_from_format(formatted_data: str) -> Dict[str, Any]:
        import toml
        return toml.loads(formatted_data)

FormatRegistry.register("toml", TOMLConverter)

JSON, Pickle, MessagePack, and YAML already ship built-in this way - see openmodels/converters/ for their real implementations, including how the optional msgpack/PyYAML dependencies are imported lazily so they're only required if you actually use those formats.

Adding a New Model Serializer

To add support for a new type of model, create a class that implements the ModelSerializer protocol:

from openmodels.protocols import ModelSerializer
from typing import Any, Dict

class TensorFlowSerializer(ModelSerializer):
    def serialize(self, model: Any) -> Dict[str, Any]:
        # Implementation for serializing TensorFlow models
        ...

    def deserialize(self, data: Dict[str, Any]) -> Any:
        # Implementation for deserializing TensorFlow models
        ...

Supported Models (scikit-learn)

OpenModels supports every scikit-learn estimator, including:

• Classification: LogisticRegression, SVC, etc.
• Regression: LinearRegression, SVR, etc.
• Clustering: KMeans
• Dimensionality Reduction: PCA

See the Supported Models docs page for the full breakdown by category, and the serialized format reference for what actually gets written to disk.

You can also retrieve the list programmatically using the SklearnSerializer.all_estimators() method:

from openmodels.serializers import SklearnSerializer

# Get all supported estimators (classifiers, regressors, etc.)
all_supported = SklearnSerializer.all_estimators()
print([name for name, cls in all_supported])

# To get only classifiers:
classifiers = SklearnSerializer.all_estimators(type_filter="classifier")
print([name for name, cls in classifiers])

# To get only regressors:
regressors = SklearnSerializer.all_estimators(type_filter="regressor")
print([name for name, cls in regressors])

Using Custom Estimators and Pipelines (Third-Party Support)

OpenModels can serialize and deserialize models and pipelines that include third-party estimators, such as those from chemotools.

from openmodels import SerializationManager, SklearnSerializer
from chemotools.utils.discovery import all_estimators  # chemotools >=0.2.2
from chemotools.derivative import SavitzkyGolay
from sklearn.cross_decomposition import PLSRegression
from sklearn.pipeline import make_pipeline

# Example data
from chemotools.datasets import load_fermentation_train
X_train, y_train = load_fermentation_train()

# Define a pipeline with chemotools preprocessing and sklearn estimator
pipeline = make_pipeline(
    SavitzkyGolay(window_size=3, polynomial_order=1, derivate_order=1),
    PLSRegression(n_components=2)
)

# Fit the pipeline
pipeline.fit(X_train, y_train)

# Serialize and deserialize the pipeline using OpenModels
serializer = SklearnSerializer(custom_estimators=all_estimators)
manager = SerializationManager(serializer)

serialized = manager.serialize(pipeline)
restored = manager.deserialize(serialized)

# Use the restored pipeline
y_train_pred = restored.predict(X_train)
print(y_train_pred)

You can pass any compatible all_estimators function, list, or dictionary to SklearnSerializer(custom_estimators=...) to extend support for custom or third-party estimators.

scikit-learn Version Compatibility

OpenModels is tested against the following scikit-learn versions via the on-demand scikit-learn compatibility CI workflow:

If you encounter any incompatibility or a use case where the library does not work correctly with your version of scikit-learn, please open an issue — we would greatly appreciate your feedback!

Contributing

We welcome contributions to OpenModels! Whether you want to add support for new models, implement new serialization formats, or improve the existing codebase, your help is appreciated.

Please refer to our Contributing Guidelines for more information on how to get started.

Running Tests

The package utilizes Taskfile as a task runner to automate and standardize development flows.

To run the tests:

1. Clone the repository:

   git clone https://github.com/Gnpd/openmodels.git
   cd openmodels
   

2. Install the package and its development dependencies:

   task install:dev
   

3. Run the tests:

   task test
   

Security

OpenModels's JSON, MessagePack, and YAML formats are all safe to load from untrusted sources (YAML because OpenModels always uses yaml.safe_load/safe_dump). The optional Pickle format is not: like any pickle-based tool, deserializing it can execute arbitrary code, so only load pickle files from sources you trust. See SECURITY.md for details and how to report a vulnerability.

License

This project is licensed under the MIT License. See the LICENSE file for details.

Changelog

For a detailed changelog, please see the CHANGELOG.md file.

Support

If you encounter any issues or have questions, please file an issue on our GitHub repository.

We're always looking to improve OpenModels. If you have any suggestions or feature requests, please let us know!

Acknowledgements

During the revision and design of OpenModels, we came across the discontinued project sklearn-json.
Although no longer maintained, it provided valuable ideas — particularly around testing approaches — that inspired parts of our implementation.
We would like to acknowledge and thank its authors for their earlier contributions to open model serialization efforts in the scikit-learn ecosystem.