Talmaj/onnx2pytorch

Transform ONNX model to PyTorch representation

What it solves

It allows users to convert neural network models from the ONNX format to PyTorch. This enables developers to leverage PyTorch's backend and its extensive set of features for manipulating neural networks that were originally exported to ONNX.

How it works

The library provides a ConvertModel class that takes a loaded ONNX model as input and transforms it into a PyTorch model. It maps ONNX operators to their PyTorch equivalents, ensuring that operators or attributes that cannot be expressed in PyTorch raise an error rather than producing incorrect results.

Who it’s for

Developers and researchers who have models in ONNX format and want to move them into the PyTorch ecosystem for inference or further manipulation.

Highlights

  • Broad Model Support: Successfully converts various architectures including MobileNet, ResNet, BERT-Squad, and YOLOv4.
  • Differential Testing: Operators are tested against onnxruntime and ONNX's ReferenceEvaluator across different schema revisions.
  • Debug Mode: Includes a debug flag to compare activations between the PyTorch version and the original ONNX runtime to pinpoint discrepancies.
  • Inference Ready: Supports batch sizes greater than 1 and uses inference mode for BatchNorm layers.

Related

  • Project
    pnnx/pnnx

    An open standard for PyTorch model interoperability that optimizes and exports models to formats like ncnn and ONNX-zero for efficient inference.

  • Project
    onnx/sklearn-onnx

    A library that converts scikit-learn models to the ONNX format, enabling high-performance inference using ONNX Runtime.

  • Project
    microsoft/onnxconverter-common

    A collection of common utilities and functions used by converters to translate AI models from various frameworks into the ONNX format.

  • Project
    PINTO0309/onnx2tf

    A tool for converting ONNX models to LiteRT, TFLite, TensorFlow, and PyTorch, featuring a high-speed direct conversion path to minimize deployment friction.

  • Project
    onnx/onnx

    ONNX is an open, framework‑agnostic model format and reference implementation that lets you move AI models between tools and runtimes. Install via `pip install onnx`, use the Python API to build or load a graph, then run inference with any ONNX‑compatible engine.