ONNX

Install mlbundle[onnx,schema]. mlbundle does not convert models to ONNX. It bundles an ONNX file you already have, byte for byte, after validating that it is a self-contained tabular model.

This example assumes trained.onnx is a classifier with two numeric features and two probability columns, and no label output:

from pathlib import Path

from mlbundle import export, load

original = Path("trained.onnx").read_bytes()
export("trained.onnx", "onnx.mlbundle", class_labels=[["cold", "warm"]])

bundle = load("onnx.mlbundle")
assert bundle.native_model.data == original
assert bundle.native_model.session.get_providers() == ["CPUExecutionProvider"]
print(bundle.class_labels)  # [['cold', 'warm']]
print(bundle.predict_proba([[0.0, 0.0]]))

export accepts a path, bytes or an onnx.ModelProto (serialized once). native_model holds the original data and an onnxruntime session on the CPU execution provider.

Labels

• Labels in the graph (ZipMap, LinearClassifier, TreeEnsembleClassifier or SVMClassifier attributes) are verified; supplied class_labels= must agree with them.
• Without graph labels, supplied labels are declarations checked only for count; without either, encoded indices 0..N-1 are used.
• predicted_label comes from the graph’s label output. Without one it is None; mlbundle never computes an argmax.

Supported graphs

• Rank-two numeric inputs (float32, float64, int32, int64) with a fixed width from 1 to 100,000, or several width-one inputs whose names become verified feature names.
• A batch dimension that is unknown, symbolic or exactly one.
• One target: a regression output of shape [batch] or [batch, 1], or a probability matrix or ZipMap sequence, optionally with a label tensor.

Kind is determined from the graph’s outputs, never declared. Verified claims include input and output signatures, opsets, IR version, producer, model metadata properties and native labels. ONNX has no visualization structure in version 1.

Rejected before any session

External tensor data, custom operator domains, model-local functions, runtime or training configuration, image, string or sequence inputs, multi-target regression, fixed batches above one and outputs whose kind cannot be determined. Validation scans the protobuf with bounded work before onnxruntime sees it.

Tested with ONNX 1.23.0 and onnxruntime 1.30.0. The exact layouts and bounds are in ONNX mapping.