Load and save graphical models in pyAgrum >2.3.2

- Tips
April 29, 2026
09 Mins read
notebook

Loading and saving graphical models

pyAgrum can read and write graphical models in many file formats. This notebook gives an overview of which formats are available for each model type, what they preserve, and why the native bgum (binary) and jgum (JSON) formats are the best choice for pyAgrum workflows.

import os
import tempfile
import time

import pyagrum as gum
import pyagrum.lib.notebook as gnb

Available formats

The three main model types each have their own set of supported formats.

print(f"BayesNet formats  : {gum.availableBNExts()}")
print(f"InfluenceDiagram  : {gum.availableIDExts()}")
print(f"MarkovRandomField : {gum.availableMRFExts()}")

BayesNet formats  : bif|dsl|net|bifxml|o3prm|uai|xdsl|pkl|jgum|bgum
InfluenceDiagram  : xmlbif|bifxml|xml|jgum|bgum|pkl
MarkovRandomField : uai|jgum|bgum|pkl

The load/save API is uniform across model types:

Model	Load	Save
`BayesNet`	`gum.loadBN(filename)`	`gum.saveBN(bn, filename)`
`InfluenceDiagram`	`gum.loadID(filename)`	`gum.saveID(diag, filename)`
`MarkovRandomField`	`gum.loadMRF(filename)`	`gum.saveMRF(mrf, filename)`

The format is selected automatically from the file extension.

Bayesian networks

The generated readme of the pgmrepository contains more complete comparison in size and read/write time among the different formats for Bayesian networks.

A tour of BN formats

Let’s load the classic Asia network and save/reload it in every available format, comparing file size and round-trip speed.

bn_asia = gum.fastBN("Asia->Tuberculosis->TorC->chest XRay;TorC<-Lung cancer<-Smoking->Bronchitis->Dyspnoea<-TorC")
gnb.flow.row(bn_asia, captions=[f"Asia BN — {bn_asia.size()} nodes, {bn_asia.sizeArcs()} arcs"])

Asia BN — 8 nodes, 8 arcs

def benchmark_bn(bn, exts):
    """Save and reload a BN in every given format, return a comparison table."""
    rows = []
    with tempfile.TemporaryDirectory() as d:
        for ext in exts.split("|"):
            fname = os.path.join(d, f"model.{ext}")
            try:
                t0 = time.perf_counter()
                gum.saveBN(bn, fname)
                t_save = time.perf_counter() - t0
                size = os.path.getsize(fname)
                t0 = time.perf_counter()
                bn2 = gum.loadBN(fname)
                t_load = time.perf_counter() - t0
                names_ok = sorted(bn.names()) == sorted(bn2.names())
                labels_ok = names_ok and all(
                    list(bn.variable(n).labels()) == list(bn2.variable(n).labels())
                    for n in bn.names()
                )
                types_ok = names_ok and all(
                    bn.variable(n).varType() == bn2.variable(n).varType()
                    for n in bn.names()
                )
                rows.append((ext, size, t_save * 1000, t_load * 1000, names_ok, labels_ok, types_ok))
            except Exception as e:
                rows.append((ext, None, None, None, False, False, str(e)[:60]))
    return rows


rows = benchmark_bn(bn_asia, gum.availableBNExts())

header = f"{'ext':8s}  {'size (B)':>9s}  {'save (ms)':>10s}  {'load (ms)':>10s}  {'names':>5s}  {'labels':>6s}  {'types':>5s}"
print(header)
print("-" * len(header))
for ext, size, ts, tl, n_ok, l_ok, t_ok in rows:
    if size is not None:
        print(f"{ext:8s}  {size:9d}  {ts:10.3f}  {tl:10.3f}  {str(n_ok):>5s}  {str(l_ok):>6s}  {str(t_ok):>5s}")
    else:
        print(f"{ext:8s}  (unsupported for this model)")

ext        size (B)   save (ms)   load (ms)  names  labels  types
-----------------------------------------------------------------
bif       (unsupported for this model)
dsl       (unsupported for this model)
net       (unsupported for this model)
bifxml         3593       0.189       0.237   True    True   True
o3prm     (unsupported for this model)
uai             587       0.133       0.107  False   False  False
xdsl           2342       0.481       0.119   True    True  False
pkl            1940       0.178       0.104   True    True   True
jgum           1887       0.115       0.066   True    True   True
bgum            836       0.098       0.059   True    True   True

Variable type fidelity

Many real-world networks contain variables of heterogeneous types: labelled (LabelizedVariable), integer ranges (RangeVariable) or discretized continuous domains (DiscretizedVariable). Not all formats can represent these faithfully.

# Build a BN that mixes all main variable types
bn_mixed = gum.BayesNet("mixed_types")
bn_mixed.add(gum.LabelizedVariable("Smoker", "Smoker", ["yes", "no"]))
bn_mixed.add(gum.RangeVariable("Age", "Age", 20, 60))
bn_mixed.add(gum.DiscretizedVariable("Temp", "Temp", [36.0, 37.0, 38.5, 42.0]))
bn_mixed.add(gum.LabelizedVariable("Cancer", "Cancer", ["yes", "no"]))
bn_mixed.addArc("Smoker", "Cancer")
bn_mixed.addArc("Age", "Cancer")
bn_mixed.addArc("Temp", "Cancer")
bn_mixed.cpt("Smoker").fillWith([0.3, 0.7])
bn_mixed.cpt("Age").fillWith(1).normalize()
bn_mixed.cpt("Temp").fillWith(1).normalize()
bn_mixed.cpt("Cancer").fillWith(1).normalize()

print("Variable types in bn_mixed:")
type_names = {
    gum.VarType_LABELIZED: "Labelized",
    gum.VarType_DISCRETIZED: "Discretized",
    gum.VarType_RANGE: "Range",
    gum.VarType_INTEGER: "Integer",
    gum.VarType_NUMERICAL: "Numerical",
}
for n in bn_mixed.names():
    v = bn_mixed.variable(n)
    print(f"  {n:10s}: {type_names.get(v.varType(), f'type={v.varType()}')}")

Variable types in bn_mixed:
  Smoker    : Labelized
  Cancer    : Labelized
  Temp      : Discretized
  Age       : Range

rows = benchmark_bn(bn_mixed, gum.availableBNExts())

header = f"{'ext':8s}  {'size (B)':>9s}  {'save (ms)':>10s}  {'load (ms)':>10s}  {'names':>5s}  {'labels':>6s}  {'types':>5s}"
print(header)
print("-" * len(header))
for ext, size, ts, tl, n_ok, l_ok, t_ok in rows:
    if size is not None:
        print(f"{ext:8s}  {size:9d}  {ts:10.3f}  {tl:10.3f}  {str(n_ok):>5s}  {str(l_ok):>6s}  {str(t_ok):>5s}")
    else:
        print(f"{ext:8s}  (unsupported for this model)")

ext        size (B)   save (ms)   load (ms)  names  labels  types
-----------------------------------------------------------------
bif       (unsupported for this model)
dsl       (unsupported for this model)
net       (unsupported for this model)
bifxml         8951       0.267       0.483   True    True   True
o3prm     (unsupported for this model)
uai            6703       0.307       0.625  False   False  False
xdsl           7886       0.554       0.519   True    True  False
pkl           16127       0.255       0.176   True    True   True
jgum          16074       0.153       0.140   True    True   True
bgum           5160       0.157       0.074   True    True   True

Observations for BN:

Format	Notes
`bif`, `dsl`, `net`	Classic, widely used, but do not support `DiscretizedVariable`
`bifxml` / `xdsl`	XML-based; `bifxml` preserves types, `xdsl` does not
`uai`	Compact but loses variable names
`o3prm`	Verbose; requires a full class hierarchy
`pkl`	Python pickle; preserves everything but not portable across pyAgrum versions
`jgum`	Native JSON format; preserves all types, human-readable
`bgum`	Native binary format; smallest files, fastest I/O, preserves all types

Influence diagrams

Influence diagrams have fewer supported formats than BNs.

# Classic Oil Wildcatter influence diagram
diag = gum.fastID("$Cost<-*Testing->TestResult<-OilContents->Reward<-*Drilling<-TestResult;Drilling<-Testing")
gnb.flow.row(diag, captions=[f"Oil Wildcatter — {diag.size()} nodes"])

Oil Wildcatter — 6 nodes

def benchmark_id(diag, exts):
    rows = []
    with tempfile.TemporaryDirectory() as d:
        for ext in exts.split("|"):
            fname = os.path.join(d, f"model.{ext}")
            try:
                t0 = time.perf_counter()
                gum.saveID(diag, fname)
                t_save = time.perf_counter() - t0
                size = os.path.getsize(fname)
                t0 = time.perf_counter()
                diag2 = gum.loadID(fname)
                t_load = time.perf_counter() - t0
                names_ok = sorted(diag.names()) == sorted(diag2.names())
                rows.append((ext, size, t_save * 1000, t_load * 1000, names_ok))
            except Exception as e:
                rows.append((ext, None, None, None, str(e)[:60]))
    return rows


rows = benchmark_id(diag, gum.availableIDExts())

header = f"{'ext':8s}  {'size (B)':>9s}  {'save (ms)':>10s}  {'load (ms)':>10s}  {'names':>5s}"
print(header)
print("-" * len(header))
for ext, size, ts, tl, n_ok in rows:
    if size is not None:
        print(f"{ext:8s}  {size:9d}  {ts:10.3f}  {tl:10.3f}  {str(n_ok):>5s}")
    else:
        print(f"{ext:8s}  (unsupported for this model)")

ext        size (B)   save (ms)   load (ms)  names
--------------------------------------------------
xmlbif         2664       0.249       0.222   True
bifxml         2664       0.125       0.111   True
xml            2664       0.117       0.097   True
jgum           1299       0.136       0.079   True
bgum            609       0.097       0.058   True
pkl            1360       0.148       0.073   True

Observations for InfluenceDiagram:

Format	Notes
`bifxml` / `xmlbif` / `xml`	Three aliases for the same XML format
`pkl`	Portable only within the same pyAgrum version
`jgum`	Compact JSON, fully faithful, human-readable
`bgum`	Smallest files, fastest I/O

Markov random fields

MRFs have the smallest set of supported formats.

mrf = gum.fastMRF("A{yes|no}--B{low|mid|high}--C{yes|no}--A;B--D{yes|no}")
gnb.flow.row(mrf, captions=[f"MRF — {mrf.size()} nodes, {mrf.sizeEdges()} edges"])

MRF — 4 nodes, 4 edges

def benchmark_mrf(mrf, exts):
    rows = []
    with tempfile.TemporaryDirectory() as d:
        for ext in exts.split("|"):
            fname = os.path.join(d, f"model.{ext}")
            try:
                t0 = time.perf_counter()
                gum.saveMRF(mrf, fname)
                t_save = time.perf_counter() - t0
                size = os.path.getsize(fname)
                t0 = time.perf_counter()
                mrf2 = gum.loadMRF(fname)
                t_load = time.perf_counter() - t0
                names_ok = sorted(mrf.names()) == sorted(mrf2.names())
                labels_ok = names_ok and all(
                    list(mrf.variable(n).labels()) == list(mrf2.variable(n).labels())
                    for n in mrf.names()
                )
                rows.append((ext, size, t_save * 1000, t_load * 1000, names_ok, labels_ok))
            except Exception as e:
                rows.append((ext, None, None, None, False, str(e)[:60]))
    return rows


rows = benchmark_mrf(mrf, gum.availableMRFExts())

header = f"{'ext':8s}  {'size (B)':>9s}  {'save (ms)':>10s}  {'load (ms)':>10s}  {'names':>5s}  {'labels':>6s}"
print(header)
print("-" * len(header))
for ext, size, ts, tl, n_ok, l_ok in rows:
    if size is not None:
        print(f"{ext:8s}  {size:9d}  {ts:10.3f}  {tl:10.3f}  {str(n_ok):>5s}  {str(l_ok):>6s}")
    else:
        print(f"{ext:8s}  (unsupported for this model)")

ext        size (B)   save (ms)   load (ms)  names  labels
----------------------------------------------------------
uai             280       0.202       0.194  False   False
jgum           1029       0.131       0.080   True    True
bgum            428       0.103       0.058   True    True
pkl            1091       0.157       0.124   True    True

Observations for MRF:

Format	Notes
`uai`	Only standard MRF format, but loses variable names and labels
`pkl`	Portable only within the same pyAgrum version
`jgum`	Full fidelity, JSON, readable
`bgum`	Smallest, fastest, full fidelity

Why bgum and jgum are the best choice for pyAgrum

The bgum and jgum formats are the native aGrUM formats, designed specifically for all model types supported by pyAgrum. They share the same advantages:

Universal — same format works for BayesNet, InfluenceDiagram and MarkovRandomField.
Full fidelity — all variable types (LabelizedVariable, RangeVariable, DiscretizedVariable, IntegerVariable) are preserved exactly.
Fast — both I/O are among the fastest of all formats.
Compact — bgum typically produces the smallest files; jgum is still compact while remaining human-readable.
No external dependencies — no need for third-party parsers.

The only difference between the two is readability:

bgum (binary) — optimal for production workflows, automated pipelines, storing large models.
jgum (JSON) — easier to inspect, diff, or version-control.

Quick demo: round-trip with bgum and jgum

# Build a BN with mixed variable types to stress-test fidelity
bn_demo = gum.BayesNet("demo")
bn_demo.add(gum.LabelizedVariable("Smoker", "Smoker", ["yes", "no"]))
bn_demo.add(gum.RangeVariable("Age", "Age", 20, 60))
bn_demo.add(gum.DiscretizedVariable("Temp", "Temp", [36.0, 37.0, 38.5, 42.0]))
bn_demo.add(gum.LabelizedVariable("Cancer", "Cancer", ["yes", "no"]))
bn_demo.addArc("Smoker", "Cancer")
bn_demo.addArc("Age", "Cancer")
bn_demo.addArc("Temp", "Cancer")
bn_demo.cpt("Smoker").fillWith([0.3, 0.7])
bn_demo.cpt("Age").fillWith(1).normalize()
bn_demo.cpt("Temp").fillWith(1).normalize()
bn_demo.cpt("Cancer").fillWith(1).normalize()

type_names = {
    gum.VarType_LABELIZED: "Labelized",
    gum.VarType_DISCRETIZED: "Discretized",
    gum.VarType_RANGE: "Range",
    gum.VarType_INTEGER: "Integer",
    gum.VarType_NUMERICAL: "Numerical",
}

print("Variable types before save:")
for n in bn_demo.names():
    v = bn_demo.variable(n)
    print(f"  {n:8s}: {type_names.get(v.varType(), f'type={v.varType()}'):12s}  labels={list(v.labels()[:4])}{'...' if v.domainSize() > 4 else ''}")

Variable types before save:
  Smoker  : Labelized     labels=['yes', 'no']
  Cancer  : Labelized     labels=['yes', 'no']
  Temp    : Discretized   labels=['[36;37[', '[37;38.5[', '[38.5;42]']
  Age     : Range         labels=['20', '21', '22', '23']...

with tempfile.TemporaryDirectory() as d:
    for ext in ("bgum", "jgum"):
        fname = os.path.join(d, f"demo.{ext}")
        gum.saveBN(bn_demo, fname)
        bn2 = gum.loadBN(fname)
        print(f"\n--- {ext} ({os.path.getsize(fname)} bytes) ---")
        for n in bn2.names():
            v = bn2.variable(n)
            print(f"  {n:8s}: {type_names.get(v.varType(), f'type={v.varType()}'):12s}  labels={list(v.labels()[:4])}{'...' if v.domainSize() > 4 else ''}")

--- bgum (5153 bytes) ---
  Smoker  : Labelized     labels=['yes', 'no']
  Cancer  : Labelized     labels=['yes', 'no']
  Temp    : Discretized   labels=['[36;37[', '[37;38.5[', '[38.5;42]']
  Age     : Range         labels=['20', '21', '22', '23']...

--- jgum (16067 bytes) ---
  Smoker  : Labelized     labels=['yes', 'no']
  Cancer  : Labelized     labels=['yes', 'no']
  Temp    : Discretized   labels=['[36;37[', '[37;38.5[', '[38.5;42]']
  Age     : Range         labels=['20', '21', '22', '23']...

# jgum is plain JSON — easy to inspect
import json

with tempfile.TemporaryDirectory() as d:
    fname = os.path.join(d, "demo.jgum")
    gum.saveBN(bn_demo, fname)
    with open(fname) as f:
        data = json.load(f)
    # Show just the variable descriptions
    print(json.dumps(data.get("variables", data.get("nodes", {})), indent=2))

[
  "Smoker{yes|no}",
  "Age[20,60]",
  "Temp[36,37,38.5,42]",
  "Cancer{yes|no}"
]

bgum and jgum work identically for all model types

with tempfile.TemporaryDirectory() as d:
    # BayesNet
    gum.saveBN(bn_asia, os.path.join(d, "asia.bgum"))
    bn_rt = gum.loadBN(os.path.join(d, "asia.bgum"))
    print(f"BN  round-trip via bgum: names match = {sorted(bn_asia.names()) == sorted(bn_rt.names())}")

    # InfluenceDiagram
    gum.saveID(diag, os.path.join(d, "oil.bgum"))
    diag_rt = gum.loadID(os.path.join(d, "oil.bgum"))
    print(f"ID  round-trip via bgum: names match = {sorted(diag.names()) == sorted(diag_rt.names())}")

    # MarkovRandomField
    gum.saveMRF(mrf, os.path.join(d, "mrf.bgum"))
    mrf_rt = gum.loadMRF(os.path.join(d, "mrf.bgum"))
    print(f"MRF round-trip via bgum: names match = {sorted(mrf.names()) == sorted(mrf_rt.names())}")

BN  round-trip via bgum: names match = True
ID  round-trip via bgum: names match = True
MRF round-trip via bgum: names match = True

Summary

	`bif`/`dsl`/`net`	`bifxml`/`xdsl`	`uai`	`pkl`	`jgum`	`bgum`
BayesNet	✓	✓	partial	✓	✓	✓
InfluenceDiagram	✗	✓	✗	✓	✓	✓
MarkovRandomField	✗	✗	partial	✓	✓	✓
Preserves all variable types	✗	partial	✗	✓	✓	✓
Preserves variable names	✓	✓	✗	✓	✓	✓
Human-readable	✓	✓	✓	✗	✓	✗
Compact size	medium	large	small	medium	small	smallest
I/O speed	medium	medium	fast	fast	fast	fastest
Version-stable	✓	✓	✓	✗	✓	✓

Recommendation:

Use bgum whenever storage size or I/O speed matters, or when working with non-BN models.
Use jgum when the file needs to be inspected, diffed, or stored in version control.
Use bif/bifxml only when interoperability with other tools (GeNIe, Netica, …) is required.