Update ONNX export to ONNX 1.22 opsets

Ouverte
  • Quelque chose ne va pas ?
#7,683 2 commentaires 0 réactions 1 personne assignée Voir sur GitHub

Les mainteneurs répondent en général sous 1 jour

@vejbomar y travaille déjà.

Depuis le 22/9/2026.

Évaluation

Difficulté
5/5
Temps estimé
Plus d'une semaine
Accessibilité débutants
25/100
Type d'issue
Fonctionnalité
Clarté
Plutôt claire
Activité
Active
Stack technique
csharp, machine-learning
Domaine
machine-learning

Piste de recherche

Commencez par localiser l’exportateur ONNX de ML.NET, les bindings protobuf générés, les références de base d’exportation ainsi que les zones de validation et de documentation mentionnées dans l’issue ; aucun fichier ni test spécifique n’est nommé. Examinez les schémas émis, les exportations d’arbres et les importations de domaines personnalisés par rapport à ONNX 1.22, puis utilisez les critères d’acceptation indiqués pour vérifier la validité du checker, l’exécution avec ONNX Runtime 1.29, la parité et le chargement d’anciens modèles.

Rédigé par le modèle d'indexation à partir du texte de l'issue.

Description

area-ONNX green yellow

Update ML.NET's ONNX exporter to target the latest released ONNX specification:

  • ONNX version: 1.22
  • Core ai.onnx opset: 27
  • ai.onnx.ml opset: 5
  • ONNX Runtime: 1.29

ML.NET currently targets core opset 12 and ai.onnx.ml opset 2. Maintaining these old targets limits future exporter development and leaves several emitted graphs using obsolete operator schemas.

Motivation

The latest released opsets should be ML.NET's primary export target. Compatibility with older runtimes should not prevent the exporter from tracking the maintained ONNX specification.

Moving to the current opsets will:

  • Allow new exporters to use modern ONNX operators and data types.
  • Enable standard bitwise and string operators where suitable.
  • Enable the newer TreeEnsemble representation.
  • Allow tree thresholds and weights to retain double precision through tensor attributes.
  • Align exported models with a currently maintained ONNX Runtime release.
  • Remove accumulated schema and domain-import debt.

This intentionally raises the runtime requirement for newly exported models. Existing ONNX files and ML.NET's ability to consume older models are unaffected.

Required changes

Update versions and dependencies
  • Change the default core opset from 12 to 27.
  • Change the ai.onnx.ml opset from 2 to 5.
  • Upgrade Microsoft.ML.OnnxRuntime to 1.29.
  • Refresh the generated ONNX protobuf bindings from ONNX 1.22.
  • Update the emitted IR version as required by the features used.
  • Update public documentation and validation of supported target versions.
Emit current core operator schemas

Update operators whose signatures changed after opset 12:

  • Emit axes as an int64 tensor input for Squeeze, Unsqueeze, and ReduceSum.
  • Emit axes as an input for the remaining Reduce* operators changed in opset 18.
  • Emit Clip bounds as inputs rather than attributes.
  • Review every emitted core operator against its opset-27 schema.
  • Preserve ML.NET inference behaviour and output shapes.

There are currently more than twenty axes attributes across the exporter which require review.

Migrate tree exports

TreeEnsembleRegressor and TreeEnsembleClassifier are deprecated in ai.onnx.ml opset 5 and are rejected by the official ONNX checker.

  • Replace the existing tree representation with TreeEnsemble.
  • Reconstruct classification labels and scores where necessary.
  • Use tensor-valued splits and weights to avoid unnecessary float conversion.
  • Preserve missing-value routing and post-transform behaviour.
  • Cover FastTree, FastForest, LightGBM, ranking, regression, and classification exports.
Correct operator-domain imports

ML.NET emits operators from the com.microsoft domain for tokenisation, hashing, and some trainer graphs, but currently writes only the core and ai.onnx.ml imports.

  • Track every operator domain used by a graph.
  • Emit a corresponding opset_import entry for each domain.
  • Add com.microsoft version 1 when required.
  • Avoid importing unused domains where practical.

Compatibility

Newly exported models will require a runtime supporting core opset 27 and ai.onnx.ml opset 5. ONNX Runtime 1.29 is the corresponding current release.

Support for producing older opsets may remain available where it is correct and maintainable, but it must not constrain or delay the current default exporter.

Acceptance criteria

  • Default exports declare core opset 27 and ai.onnx.ml opset 5.
  • Exported models use valid schemas for those opsets.
  • Every exported model passes the ONNX 1.22 checker with full checking enabled.
  • Every exported model loads and executes with ONNX Runtime 1.29.
  • ONNX results retain parity with the corresponding ML.NET pipeline.
  • Tree-based models use the non-deprecated TreeEnsemble operator.
  • Every custom operator domain has an opset_import entry.
  • Existing export baselines and documentation are updated.
  • Loading and executing existing older ONNX models remains supported.

Non-goals

  • Targeting unreleased core opset 28.
  • Automatically rewriting ONNX files previously exported by ML.NET.
  • Implementing every operator introduced between opsets 12 and 27.

References

Langage dominant
C#
Étoiles
9.4k
Forks
2k
Merge moyen
2 j 18 h
PR mergées (30 j)
64

Préparer son environnement

Par où commencer

  1. Lisez l'issue en entier, puis le guide de contribution du projet.
  2. Signalez en commentaire que vous la prenez — cela évite que deux personnes fassent le même travail.
  3. Forkez le dépôt et travaillez sur une branche.
  4. Ouvrez une pull request qui référence le numéro de l'issue.

Autres issues de dotnet/machinelearning

Toutes les issues de dotnet/machinelearning

Issues similaires

Plus d'issues C#

Recevez les nouvelles issues par e-mail

Un résumé court des issues GitHub adaptées aux débutants.