Save or load a model

Status

Variants

Convert a model to another format or code

Turn a trained model into another framework's model, a rule set, SQL, or standalone source code or a container that runs it.

Edit a trained model

Change parts of a trained model in place, such as its trees, bias, labels or appended transformations.

Hold a model in a table

Store a model in a table cell, or unpack a model into rows and pack rows back into a model.

Register or fetch a model in a registry

Register, version, stage, list, download or delete models held in a model registry or hosted model service.

Write or read a model file

Write a trained model to a file, directory or bytes in its own format or PMML, and read it back.

What other tools call it

Meridian's classification; an independent blind rating agreed on 86.2% of 312 rows (exact class, version 2).

Background

When to use it

In scikit-learn, choosing a persistence method starts from whether the Python object is needed again or the model is only served for predictions.^[1] The Python pickle module is not secure, and only data that is trusted should be unpickled.^[4] In scikit-learn, pickle, joblib and cloudpickle all use the pickle protocol, so loading a file saved by any of them can execute arbitrary code.^[1] Treat a saved model file as code rather than data: load one only from a source you trust. (Meridian's view)

scikit-learn has no supported way to load a model trained with a different version of scikit-learn.^[1] To rebuild a similar model under a later scikit-learn version, scikit-learn advises saving a reference to the training data, the source code, the dependency versions and the cross-validation score alongside the model.^[1] Keep a note of the library versions that wrote a model next to the model itself. (Meridian's view)

The MLflow Model Registry tracks the versions of each registered model and links each version to the MLflow run, logged model or notebook that produced it.^[5] When more than one version of a model is in use, keep them in a registry rather than in loose files, so each result can be traced to the version that produced it. (Meridian's view)

The method

Deserialization is the opposite of serialization: it extracts a data structure from a series of bytes.^[2] Serialization of objects does not include the methods they were linked with.^[2] In scikit-learn, a model saved with skops.io, pickle, joblib or cloudpickle should be loaded in a Python environment with the same packages and versions it was trained in.^[1]

In mlpack, a model is saved with Save() and loaded with Load(), and serialization is done by the cereal serialization toolkit in binary, JSON or XML.^[6] In mlpack, objects in the binary format may be an order of magnitude or more smaller than JSON, while JSON and XML files are human-readable but may be quite large.^[6] Yggdrasil Decision Forests (YDF) saves a model in a proprietary format of multiple files in a single directory.^[7] In YDF, a model serialized to bytes is equivalent to one saved to a directory.^[7] In YDF, saving with pure_serving removes training-specific metadata and debug information, so the model on disk is smaller.^[7]

PMML was conceived by Robert Lee Grossman, then the director of the National Center for Data Mining at the University of Illinois at Chicago.^[3] PMML version 0.9 was published in 1998, and later versions have been developed by the Data Mining Group.^[3] A PMML file's components include a header, a data dictionary, data transformations, the model and a mining schema.^[3] The Open Neural Network Exchange (ONNX) is an open-source ecosystem that sets open standards to enable a standard format for representing machine learning models.^[8] ONNX was announced by Facebook and Microsoft in September 2017.^[8]

Compared with the alternatives

In scikit-learn, ONNX lets a model be served without a Python environment, but the original Python object is lost and cannot be reconstructed.^[1] In scikit-learn, not all models are supported by ONNX.^[1] In RapidMiner, the Write PMML operator supports a listed set of model kinds, among them decision trees, rule models and linear and logistic regression models.^[9] An exchange format such as PMML or ONNX trades coverage for portability, so check that the model kind is supported before relying on one. (Meridian's view)

In scikit-learn, skops.io loads only files whose types and functions are trusted by default or by the user, which makes it more secure than pickle, joblib and cloudpickle.^[1] In MLflow, a registered model has a unique name and holds versions, aliases, tags and other metadata.^[5]

Sources

1. [1]11. Model persistence — scikit-learn 1.9.1 documentationdocumentation
2. [2]Serialization - Wikipediasecondary source
3. [3]Predictive Model Markup Language - Wikipediasecondary source
4. [4]pickle — Python object serialization — Python 3.15.2 documentationdocumentation
5. [5]ML Model Registry | MLflow AI Platformdocumentation
6. [6]mlpack documentationdocumentation
7. [7]yggdrasil-decision-forests/yggdrasil_decision_forests/port/python/ydf/model/generic_model.py at 0a54b57c95ff0cf42ecf0310ace0513a4f32cb65 · google/yggdrasil-decision-forests · GitHubdocumentation
8. [8]Open Neural Network Exchange - Wikipediasecondary source
9. [9]Write PMML - Rapidminer Documentationdocumentation