Input/Output Formats for Bayesian Networks

PyAgrum supports several formats for saving and loading graphical models, but not all formats preserve every feature of the models. Below is a guide to the supported formats and their capabilities.

Supported Formats

Warning

Only the native JSON formats (jgum and bgum), pickle, and BIFXML guarantee complete preservation of all graphical model features. Other formats may lose some of this information during serialization.

JGUM/BGUM format (Recommended)

JGUM and BGUM formats are JSON-based formats that ensure complete preservation of all features and attributes of the graphical model. They are the recommended formats for saving and loading Bayesian networks in PyAgrum.

BGUM is a binary version of JGUM, which is more compact and faster to read/write (MessagePack format).

JGUM is a text-based format that is human-readable and easy to debug.

import pyagrum as gum

## Save to JGUM/BGUM
gum.config["core","default_jgumIndent"]=2  # Optional: set indentation for better readability (-1 = max compacity)
gum.saveBN(bn, "model.jgum")
gum.saveBN(bn, "binary_model.bgum")

## Load from JGUM/BGUM
bn = gum.loadBN("model.jgum")
bn2 = gum.loadBN("model.bgum") # bn==bn2

BIFXML (complete preservation)

The BIFXML format is recommended for interoperability with other Bayesian network tools. It preserves all features and attributes of the model.

import pyagrum as gum

## Save to BIFXML
gum.saveBN(bn, "model.bifxml")

## Load from BIFXML
bn = gum.loadBN("model.bifxml")

Pickle (Alternative for complete preservation)

The Python pickle format is also supported for complete preservation.

import pyagrum as gum
import pickle

## Save model
with open('model.pkl', 'wb') as f:
    pickle.dump(bn, f)

## Load model
with open('model.pkl', 'rb') as f:
    bn2 = pickle.load(f)

## or easily
gum.saveBN(bn, "model.pkl")
bn2 = gum.loadBN("model.pkl")

Other Supported Formats

The following formats are supported but may not preserve all model features:

Net (.net): Standard format for Bayesian networks but may lose some metadata.
DSL (.dsl): GeNIe format with some limitations.
UAI (.uai): Very simple format for Bayesian networks and Markov networks with limitations.
BIF (.bif): Format for Bayesian networks with limitations.
etc. (see pyagrum.availableBNExts() for a complete list).