This tool provides a visual interface for modifying ONNX models, aimed at researchers and developers who find the iterative process of coding edits and visualizing changes in ONNX models time-consuming. It streamlines the workflow by allowing direct manipulation of the model graph within a web browser, automatically translating visual edits into ONNX Python API operations.

How It Works

Leveraging Netron for visualization and Flask for the web framework, onnx-modifier enables users to perform various graph-level and node-level operations. Users can delete nodes (with or without children), add new nodes from a comprehensive operator list, rename inputs/outputs, add new model inputs/outputs, edit node attributes, and modify model input shapes. Edits are previewed in real-time, and the modified model can be downloaded, optionally with shape inference and unused element cleanup.

Quick Start & Requirements

• Install: pip install onnx-modifier or clone the repository and run python entry.py. An executable .exe is available for Windows.
• Prerequisites: Python, Flask, ONNX, Netron.
• Links: Official Docs, Sample Models

Highlighted Details

• Supports visual editing of node attributes, model initializers (including loading from .npy files), and model input shapes.
• Features "Delete With Children" node deletion using backtracking and allows adding new nodes with input/output connections.
• Experimental shape inference and cleanup functionalities are available upon download.
• Enables renaming of node and model inputs/outputs, and adding new model outputs from intermediate nodes.

Maintenance & Community

The project is under active development, welcoming issues and pull requests.

Licensing & Compatibility

The repository does not explicitly state a license in the README.

Limitations & Caveats

Executable files for platforms other than Windows are not yet supported. Some features, like adding new nodes and shape inference, are marked as experimental. The input/output list for nodes is currently limited to 8 elements.