A library to transform ONNX model to PyTorch. This library enables use of PyTorch backend and all of its great features for manipulation of neural networks.

pip install onnx2pytorch

import onnx
from onnx2pytorch import ConvertModel

onnx_model = onnx.load(path_to_onnx_model)
pytorch_model = ConvertModel(onnx_model)

ONNX_OPERATOR_CHECKLIST.md lists every ai.onnx operator the converter implements, along with the ones that are only partially supported. Operators and attributes that cannot be expressed in PyTorch raise rather than returning a wrong tensor.

Each operator is tested differentially against onnxruntime, falling back to onnx's ReferenceEvaluator, at every schema revision it exists in; see tests/onnx2pytorch/opset_matrix.py. Combinations that neither runtime can run are reported at the end of a test run, since the converter's behaviour for those is unverified.

Currently supported and tested models from onnx_zoo:

• MobileNet
• ResNet
• ShuffleNet_V2
• BERT-Squad
• EfficientNet-Lite4
• Fast Neural Style Transfer
• Super Resolution
• YOLOv4 (Not exactly the same, nearest neighbour interpolation in pytorch differs)
• U-net (Converted from pytorch to onnx and then back)

Known current version limitations are:

• batch_size > 1 is now supported by default.
  BatchNorm layers use inference mode (running statistics), which is correct for ONNX models
  exported for inference.
• Fine tuning and training of converted models was not tested yet, only inference.

pip install -r requirements.txt

From onnxruntime>=1.5.0 you need to add the following to your .bashrc or .zshrc if you are running OSx: export KMP_DUPLICATE_LIB_OK=True

The Uncompromising Code Formatter: Black
black {source_file_or_directory}

Install it into pre-commit hook to always commit nicely formatted code:
pre-commit install

Pytest and tox.
tox

To test the complete conversion of an onnx model download pre-trained models: ./download_fixtures.sh
Use flag --all to download more models. Add any custom models to ./fixtures folder to test their conversion.

Set ConvertModel(..., debug=True) to compare each converted activation from pytorch with the activation from onnxruntime.
This helps identify where in the graph the activations start to differ.