Practical overview
When it fits
Use Onnx2Tf when a model is available in ONNX or LiteRT and a documented output route matches the framework or runtime needed downstream. Check the operator support list and the environment requirements first; use the accuracy-checking workflow when investigating ONNX-to-TensorFlow output differences.
Who it is for
- Machine-learning developers moving ONNX models into LiteRT/TFLite or TensorFlow workflows.
- PyTorch developers who need model representations such as nn.Module, TorchScript, state dictionaries or Exported Programs.
- Teams assessing operator coverage and diagnosing differences between ONNX and TensorFlow outputs during conversion.
Documented uses
- Convert an ONNX model into LiteRT/TFLite for a mobile or edge inference workflow.
- Export an ONNX model to TensorFlow formats or several PyTorch representations.
- Convert a LiteRT model into a PyTorch representation.
- Inspect operator support and compare ONNX and TensorFlow outputs to investigate conversion errors.
Requirements & limits
The repository is public and lists an MIT license.
- Check the model's operators against the README's support list, which marks some operators as only partially supported.
- For onnx2tf 2.0.0 and later, the README specifies Python 3.10 or newer and warns that TensorFlow 2.15.0 or earlier does not work properly.
- TensorFlow-backed features, including the listed TensorFlow exports, are installed with the TensorFlow extra.
- Accuracy checks can use substantial RAM; the README says checks are skipped when estimated needs exceed 80% of total system RAM.
- The project README recommends LiteRT Torch rather than onnx2tf.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Convert an ONNX model with the Docker image
Run the documented Docker conversion and write the SavedModel output into a host-mounted work directory.
Before you start
- Docker installed and usable on the host
- An ONNX model in the current host directory
- For the GHCR image, registry authentication with a personal access token; the README says Docker Hub pulls do not require authentication
Steps
Run the documented Docker command with the current directory mounted at /work.
Pass the model path under /work to onnx2tf using -i, and set the desired output directory with -o.
Use the documented example pattern: the input is /work/densenet-12.onnx and the output directory is /work/saved_model.
Expected outcome
The README describes conversion inside Docker with the model output written to the host-mounted directory.
Guide limitations
- This documented example produces a SavedModel directory; it does not document the conversion of every model or output format.
- The README states that onnx2tf 2.0.0 and later require Python 3.10 or newer and do not work properly with TensorFlow 2.15.0 or earlier.
Install onnx2tf in a Python 3.12 virtual environment
Set up the documented Python environment and install onnx2tf, with an optional TensorFlow feature extra.
Before you start
- A host environment able to run the documented uv setup commands
Steps
Install uv using the README's documented installer command.
Install Python 3.12.12 with uv and create a Python 3.12.12 virtual environment.
Activate the environment and install onnx2tf with uv pip.
For TensorFlow-backed features, install the documented onnx2tf[tensorflow] extra instead.
Expected outcome
The virtual environment has onnx2tf installed; selecting the TensorFlow extra adds the documented TensorFlow-backed features.
Guide limitations
- This is an installation recipe; it does not run a model conversion.
- The README warns that onnx2tf 2.0.0 and later do not work properly with TensorFlow 2.15.0 or earlier.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 13 October 2026
- Original website title
- PINTO0309/onnx2tf · Official homepage title · retained in its source language