ONNX2DFG

Tool to convert an ONNX graph into a data-flow graph, decomposing its high-level operations into low-level operations and performing a set of optimizations. The resulting graph maintains its initial structure, that is (as in all ONNX graphs), the nodes represent operations, initial inputs, and final outputs.

Installation

To install the package:

npm install @specs-feup/onnx2dfg

CLI Usage

Usage: onnx2dfg <input_file> [options]

Options:
      --version                Show version number                      [boolean]
  -o, --output                 Output resulting graph to a file         [string]
  -f, --format                 Output format (json or dot)              [string] [choices: "json", "dot"] [default: "json"]
  -v, --verbosity              Control verbosity (0 = silent, 1 = normal/outputs, 2 = verbose)                          [number] [default: 1]
      --noLowLevel, --nl       Disable the low-level conversion         [boolean] [default: false]
      --noOptimize, --no       Disable optimization steps               [boolean] [default: false]
      --noCodegen, --nc        Disable code generation step             [boolean] [default: false]
      --visualization, --vz    Choose visualization option (0 = none, 1 = Graphviz online link, 2 = Graphviz server)    [number] [default: 2]
      --help                   Show help                                [boolean]

You need to provide an input file (ONNX or JSON)

Programmatic Usage

In addition to the CLI, ONNX2DFG can be used programmatically by importing its functions in your project. This allows you to parse ONNX files, manipulate data-flow graphs, and generate outputs programmatically. The available functions are the following:

onnxFileParser

Parses an ONNX file or JSON graph into an ONNX object.

• Input:
  • Path of input file (string)
• Output: ONNX graph parsed into a JSON file (json)

import { onnxFileParser } from "@specs-feup/onnx2dfg";

const onnxObject = await onnxFileParser("path/to/file.onnx");
console.log(onnxObject);

loadGraph

Loads an ONNX object into a data-flow graph and optionally applies low-level transformations and optimizations.

• Input:
  • ONNX graph parsed into a JSON file (json)
  • Enable low-level operation decomposition (boolean)
  • Enable optimizations (boolean)
  • Convert output to DOT format (boolean)
• Output: Resulting flow graph, either the object or in DOT format depending on the option chosen (flow graph or string)

import { loadGraph } from "@specs-feup/onnx2dfg";

const dotGraph = loadGraph(onnxObject, true, true, true); // Enable both low-level and optimization steps and convert output to DOT format
console.log(dotGraph);

renderDotToSVG

Converts a DOT graph string into an SVG string for rendering or embedding.

• Input:
  • Source graph in DOT format (string)
• Output: SVG image of the graph (string)

import { renderDotToSVG } from "@specs-feup/onnx2dfg";

const dotGraph = "digraph { a -> b }";
const svgContent = await renderDotToSVG(dotGraph);
console.log(svgContent);

generateGraphvizOnlineLink

Generates a link to visualize a DOT graph on Graphviz Online.

• Input:
  • Source graph in DOT format (string)
• Output: Link to open the given graph in Graphviz Online (string)

import { generateGraphvizOnlineLink } from "@specs-feup/onnx2dfg";

const dotGraph = "digraph { a -> b }";
const link = generateGraphvizOnlineLink(dotGraph);
console.log(link); // Outputs: https://dreampuf.github.io/GraphvizOnline/#...

generateGraphCode

Generates code from a data-flow graph.

• Input:
  • Source flow graph (with low-level decomposition applied) (flow graph)
• Output: Code corresponding to the source graph (string)

import { generateGraphCode } from "@specs-feup/onnx2dfg";

const code = generateGraphCode(graph);
console.log(code);

License

Licensed under the Apache 2.0 License.