ONNX Graph Preview

See the structure of an ONNX model in your notes. Write the model as text in an onnx code block and it's drawn as a graph in the style of Netron, in Reading view and Live Preview, on desktop and mobile. The plugin only draws models; it never runs them.

```onnx
<ir_version: 8, opset_import: ["" : 17]>
cnn (float[N, 3, 32, 32] x) => (float[N, 10] y) {
  c = Conv <kernel_shape = [3, 3], pads = [1, 1, 1, 1]> (x, w)
  r = Relu (c)
  p = GlobalAveragePool (r)
  y = Gemm <transB = 1> (p, fc_w, fc_b)
}
```

Features

• Graphs in Reading view, Live Preview, embedded notes, hover previews, Canvas cards and exported PDFs.
• Operators coloured by kind, as in Netron: layers, activations, pooling, normalisation and more.
• Graphs that flow down or to the right.
• Captions, alignment and alt text for each graph.
• A button to see a model's text without leaving Reading view.
• Follows your theme: graphs stay readable in dark themes.
• Works offline.

Install

1. In Obsidian, open Settings → Community plugins and turn off Restricted mode if it's on.
2. Select Browse, search for ONNX Graph Preview, then select Install and Enable.

To install by hand instead, download main.js, manifest.json and styles.css from the latest release into <your vault>/.obsidian/plugins/onnx-graph-preview/, then enable the plugin.

Writing models

Put the model in a code block tagged onnx, in any of these forms:

• ONNX text syntax, as above. The syntax reference describes it.
• Protobuf text, as written by onnx.save with format="pbtxt".
• JSON, as written by MessageToJson for a ModelProto.

The plugin recognises the form by itself. A binary .onnx file can't go in a note; convert it to text first, for example with onnx.printer.to_text(model) in Python.

What the graph shows

Each operator is a box headed with its type, listing its settings and the weights it uses. Inputs and outputs are rounded pills, and arrows are labelled with tensor shapes. Placement is automatic. The branches of If and the body of Loop are named but not drawn, and functions aren't drawn.

Captions and other options

Add options in braces after onnx:

```onnx {caption="Figure 1: The stem" align="center" direction="right"}
stem (float[N, 3, 224, 224] x) => (float[N, 64, 56, 56] y) {
  c = Conv <kernel_shape = [7, 7], strides = [2, 2], pads = [3, 3, 3, 3]> (x, w)
  r = Relu (c)
  y = MaxPool <kernel_shape = [3, 3], strides = [2, 2], pads = [1, 1, 1, 1]> (r)
}
```

Put each value in quotes. Values are plain text, not Markdown. If an option is misspelled, it's ignored and the graph still appears; only a direction other than down or right is shown as an error.

Viewing the model text

• Reading view: hover over a graph and select the </> button in its corner to see the text. Select it again to see the graph. Obsidian remembers which graphs you switched, even after a restart. Show every diagram in this note in the command palette switches them all back.
• Live Preview: select Obsidian's own Edit this block button, just as for a Mermaid diagram.
• Phones and tablets: the button is always visible, and a graph wider than the screen scrolls sideways so its labels stay readable.

Dark and light themes

Graphs follow your theme. In a dark theme, boxes turn dark, text and arrows are lightened to read, and Netron's colours keep their hue. In a light theme, graphs appear on a white card with Netron's exact colours.

To always show the exact colours on a white card, turn off Follow dark themes in the plugin's settings.

Settings

Commands

When something goes wrong

• A model that doesn't parse: the error appears in place of the graph, quoting the line with a caret under the spot where reading stopped.
• "Layout took longer than…": the graph hit the Layout timeout. Raise the limit in settings, or split the model into several blocks.
• Very large models: a block may hold up to 1,000,000 characters and 1,000 nodes. Beyond that a static drawing stops being readable, and the block says so.

Privacy

Graphs are drawn on your device, inside Obsidian. The plugin doesn't connect to the internet, doesn't collect any data, doesn't read or write any files, and never runs a model.

Credits

Models are read by Netron, created by Lutz Roeder, and laid out by the Eclipse Layout Kernel through elkjs. The licenses of the software built into this plugin are listed in THIRD_PARTY_NOTICES.md.

Released under the MIT License. To build or change the plugin, see DEVELOPMENT.md.