Add JSON export functionality for models

[R-Mob]

Summary

This PR adds JSON export functionality to Netron, allowing users to export neural network models as structured JSON files for programmatic analysis and integration.

Features

✨ Export complete model structure including:

• Model metadata (format, version, producer, description)
• All graphs with nodes, inputs, and outputs
• Node types, attributes, and connections
• Tensor shapes and data types

✨ Support for all model formats (ONNX, TensorFlow, PyTorch, etc.)
✨ Proper handling of BigInt values and complex nested structures
✨ Available in both desktop (Electron) and browser versions

Changes

• view.js: Implemented comprehensive JSON serialization methods

  • _serializeModelToJson() - Main serialization entry point
  • _serializeGraph() - Graph structure serialization
  • _serializeNode() - Node serialization with inputs/outputs/attributes
  • _serializeType() - Type and shape serialization with BigInt handling
  • Added JSON export handling in export() method

• app.js: Added JSON to export file type filters in Electron save dialog

• Menu options:

  • Desktop: File → Export... → Select JSON (Ctrl/Cmd+Shift+E)
  • Browser: File → Export as JSON (Ctrl/Cmd+Shift+J)

Testing

✅ Tested with ONNX models (squeezenet)
✅ Verified BigInt serialization for model versions and tensor dimensions
✅ Built and tested macOS app successfully
✅ Verified JSON output is valid and properly formatted

Use Cases

This enhancement enables:

• Programmatic inspection of model structures
• Model documentation and analysis
• Integration with other tools and workflows
• Debugging and validation of model architectures

Example Output

The exported JSON includes structured data like:

{
  "format": "ONNX v8",
  "producer": "pytorch v2.0",
  "graphs": [{
    "name": "main",
    "nodes": [{
      "name": "conv1",
      "type": {
        "name": "Conv",
        "category": "Layer"
      },
      "inputs": [...],
      "outputs": [...],
      "attributes": [...]
    }]
  }]
}

[lutzroeder]

Hi @R-Mob, thank you for the pull request.

Can you explain the scenario you are using this for and would onnx2json be an option?

[R-Mob]

Hi @R-Mob, thank you for the pull request.

Can you explain the scenario you are using this for and would onnx2json be an option?

Hi @lutzroeder, thank you for reviewing!

Use Case:
I wanted a way to export model structures directly from Netron's UI for programmatic analysis and documentation. Having this built into Netron provides several advantages:

1. Universal Support: Works with all model formats that Netron already supports (ONNX, TensorFlow, PyTorch, CoreML, etc.), not just ONNX
2. Integrated Workflow: Export JSON without switching tools or command-line utilities
3. Visualization + Export: Users can visually inspect the model in Netron and then export the same view to JSON
4. Cross-platform: Works in both desktop and browser versions

Why not onnx2json:
While onnx2json is great for ONNX-specific workflows, this feature:

• Supports all model formats Netron handles
• Provides a GUI option for non-technical users
• Exports Netron's processed/normalized model structure
• Doesn't require separate tool installation

This makes it useful for model documentation, debugging, integration testing, and educational purposes where users want to programmatically inspect models they're already viewing in Netron.

[lutzroeder]

@R-Mob makes sense. The main concern here is setting expectations around scope and long-term maintenance.

Export and model conversion is fundamentally different from visualization. To do this properly, Netron would need a stable intermediate representation, introducing yet another custom deep learning model format, tests, infrastructure to support many export formats, and full platform coverage for desktop, browser and server and dedicated maintainers to support this long-term. Model export and conversion also tends to require much stricter, format-complete support than Netron’s current “best-effort for visualization” approach, which risks pulling the project into low-EV corner cases better handled by format- or vendor-specific tools.

Happy to discuss concrete, narrowly scoped use cases. The challenge is how to avoid the slippery slope of “nice to haves” duplicating properly resourced vendor-specific tools and how to clearly separate one user’s special needs from the broader ecosystem of export and model conversion requirements.

[R-Mob]

@lutzroeder Thank you for the detailed explanation - I completely understand your concerns about scope creep and maintenance burden.

You're absolutely right that this could lead down a slippery slope of feature requests for comprehensive model conversion/export functionality, which is not Netron's core mission and would be better handled by format-specific tools.

Reconsidering the Scope:
After your feedback, I realize my implementation may be too broad. Let me propose a more narrowly scoped alternative that stays true to Netron's visualization focus:

Option 1: Graph Metadata Export Only ( good solution for this issue - #1461 )

• Export only what's currently displayed in the UI (node names, types, connections)
• Purely for documentation/diagramming purposes
• No tensor data, no weights, no model conversion capabilities
• Just a JSON representation of the visual graph structure users already see
• Clear disclaimer: "Visualization export only - not for model reconstruction"

Option 2: Close the PR
If even Option 1 risks setting wrong expectations or creates maintenance overhead, I'm happy to close this PR. I appreciate you taking the time to review and explain the architectural concerns.

My Actual Use Case:
I wanted to document model architectures I was reviewing in Netron by exporting the graph structure (nodes/connections) for reports and presentations. I can achieve this using screenshots or by using onnx2json for ONNX models specifically.

Please let me know if Option 1 makes sense as a more focused feature, or if you'd prefer I close this PR. Either way, thank you for maintaining such a valuable tool!