Neuromorphic Intermediate Representation Visualisation Tool

Turn your NIR definitions into a nice graph, the original publication serving as a template.

Customise your node colour preferences in style.yml, and quickly generate graphs from your neuromorphic networks.

This work is in progress.

Running Example (Jupyter Notebook)

By running the following code (from a notebook),

import nir
import nirviz
import numpy as np


a = np.random.randn(2)
ir = nir.NIRGraph(
    nodes={
        "input": nir.Input(input_type=np.array([2])),
        "affine1": nir.Affine(weight=np.zeros((2,2)), bias=False),
        "cu1": nir.CubaLIF(tau_mem=a, tau_syn=a, r=a, v_leak=a, v_threshold=a, v_reset=a),
        "affine_rec": nir.Affine(weight=np.zeros((2,2)), bias=False),
        "affine2": nir.Affine(weight=np.zeros((2,2)), bias=False),
        "cu2": nir.CubaLIF(tau_mem=a, tau_syn=a, r=a, v_leak=a, v_threshold=a, v_reset=a),
        "output": nir.Output(output_type=np.array([2]))
    },
    edges=[("input", "affine1"), ("affine1", "cu1"), ("affine_rec", "cu1"),  ("cu1", "affine_rec"), ("cu1", "affine2"), ("affine2", "cu2"), ("cu2", "output")])

viz = nirviz.visualize(ir)
viz.show()

You would get the following visualisation

Similar to Figure 3 of the publication.

Running example (CLI)

To convert a saved NIR graph (e.g. srnn.nir) to a PNG or SVG, you can use one of the following commands:

python -m nirviz srnn.nir              # SVG -> stdout
python -m nirviz srnn.nir img/srnn.png # PNG -> file
python -m nirviz srnn.nir img/srnn.svg # SVG -> file

Customising the style

You can customise the style you see via the style file.

Style file location

The style file is defined in YAML. You can find the default location by running:

import nirviz
print(f"nirviz style file location: {nirviz.visualize.default_style_file()}")

or by passing your own style.yml:

import nirviz
viz = nirviz.visualize(nir_graph, style_file="style.yml")
viz.show()

python -m nirviz --yaml './style.yml' srnn.nir

Style file format

The format currently only supports setting node attributes. The node attributes correspond to Graphviz node attributes. An example file would contain:

node-categories:
    category-name: # User defined
        patterns: ["Affine", "IF"]
        attributes:
            # Corresponds to node attributes of graphviz
            # https://graphviz.org/docs/nodes/
            color: "red"
            style: "filled"
            shape: "box"

Which would paint all "Affine" and "IF" NIR nodes red.