Visualization procedures¶
libmg offers visualization procedure both for Spektral Graph objects and \(\mu\mathcal{G}\) model outputs.
Currently two visualization engines are supported: PyVis and Cosmograph. PyVis is a
Python binding for vis.js, a Javascript library for dynamic visualizations in the browser. Cosmograph is a React library built on
cosmos, a GPU-accelerated Force Graph layout algorithm and rendering engine.
Engine features and recommendations¶
Pyvis¶
On the web page generated using PyVis it is possible to move around nodes by dragging them with the mouse pointer, to make the graph clearer to see. The
entire graph can be moved by dragging on an empty spot. Hovering a node with the pointer will show its ID. On the search bar at the top is possible to
select a node by ID. Below the search bar there is a filter bar that allows to filter nodes and edges according to their ID or label.
Note
We recommend the usage of PyVis only for small graph instances (nodes in the order of the hundreds).
Note
The pyvis engine generates a single index.html file
Cosmograph¶
On the web page generated using Cosmograph it is not possible to move individual nodes, but right-clicking on an empty spot will "shake" the simulation a
little. The entire graph can be moved by dragging on an empty spot. Hovering a node will temporarily replace its label with its ID. On the search bar at the
top is possible to select a node by ID or by label. Finally, a button below the search bar allows to toggle the shape of the edges, which can be either
straight (default) or curved. Curved edges may slow the simulation but it may make it easier to see the links between the nodes.
Note
For the time being, the Cosmograph engine doesn't show edge labels.
Note
The pyvis engine generates a directory, inside which is situated the index.html that should be opened in the browser.
Visualizing a Graph¶
To visualize a graph, simply call the print_graph function on a Graph instance:
from libmg import Graph, print_graph
graph = Graph(...)
print_graph(graph, show_labels=False, open_browser=True, filename='mygraph', engine='pyvis')
The other arguments of this function can be used to specify whether tho show the true labels (if present in the Graph instance), whether to automatically
open the default web browser, a name for the generated file or directory, and the engine to be used.
Visualizing a model output¶
To visualize the output of a model, call the print_layer function by passing in the model, its inputs and the layer to print. To print the output of a
model it means to print the output of its last layer (with index -1):
model = compiler.compile(...)
inputs, = next(iter(loader.load()))
print_layer(model, inputs, layer_idx=-1, filename='myoutputs', open_browser=True, engine='cosmo')
Note
Since a loader always returns a tuple (containing either only the actual inputs, or the inputs and the true labels), remember to unpack the tuple with
inputs, = next(iter(loader.load())) or inputs, y = next(iter(loader.load())).
The layer to print does not have to be the last, but can be any layer of the model. To specify the layer, we can provide either the integer index or its name.
The integer index of a layer can be found calling model.summary() and it can be positive or negative, with negative numbers have the usual meaning in
Python. The name of a layer is the \(\mu\mathcal{G}\) expression corresponding to it, expressed either as string or as parse tree (which is obtained by
calling mg_parser.parse on the expression string). The name of the layer will be parsed by mg_parser if passed in as string, therefore parentheses can
be used as needed.
print_layer(model, inputs, layer_name='(a || b)', filename='myoutputs', open_browser=True, engine='cosmo')
# equivalent to
print_layer(model, inputs, layer_name='a || b', filename='myoutputs', open_browser=True, engine='cosmo')
Warning
Do not print any of the layers corresponding the edge labels, the adjacency matrix, or the indices produced by the MultipleGraphLoader.