Visualise models as graphs

[rhayes777]

Visualise model and prior relationships using a graph. Also visualise factor and variable relationships using a graph.

[matthewghgriffiths]

I had set up some code for visualising FactorGraphs

• PyAutoFit/autofit/graphical/factor_graphs/graph.py

  Lines 344 to 485 in ae791e2

  	def draw_graph(
  	self,
  	pos=None,
  	ax=None,
  	size=20,
  	color="k",
  	fill="w",
  	factor_shape="s",
  	variable_shape="o",
  	factor_labels=None,
  	variable_labels=None,
  	factor_kws=None,
  	variable_kws=None,
  	edge_kws=None,
  	factors=None,
  	draw_labels=False,
  	label_kws=None,
  	**kwargs,
  	):
  	try:
  	import matplotlib.pyplot as plt
  	import networkx as nx
  	except ImportError as e:
  	raise ImportError(
  	"Matplotlib and networkx required for draw_graph()"
  	) from e
  	except RuntimeError as e:
  	print("Matplotlib unable to open display")
  	raise e
  	if ax is None:
  	ax = plt.gca()
  	G = self.graph
  	if pos is None:
  	pos = bipartite_layout(factors or self.factors)
  	kwargs.setdefault("ms", size)
  	kwargs.setdefault("c", color)
  	kwargs.setdefault("mec", color)
  	kwargs.setdefault("mfc", fill)
  	kwargs.setdefault("ls", "")
  	factor_kws = factor_kws or {}
  	factor_kws.setdefault("marker", factor_shape)
  	variable_kws = variable_kws or {}
  	variable_kws.setdefault("marker", variable_shape)
  	# draw factors
  	xy = np.array([pos[f] for f in self.factors]).T
  	fs = ax.plot(*xy, **{**kwargs, **factor_kws})
  	# draw variables
  	xy = np.array([pos[f] for f in self.all_variables]).T
  	vs = ax.plot(*xy, **{**kwargs, **variable_kws})
  	# draw edges
  	edges = nx.draw_networkx_edges(G, pos, **(edge_kws or {}))
  	# remove ticks from axes
  	ax.tick_params(
  	axis="both",
  	which="both",
  	bottom=False,
  	left=False,
  	labelbottom=False,
  	labelleft=False,
  	)
  	if draw_labels:
  	self.draw_graph_labels(
  	pos,
  	ax=ax,
  	factor_labels=factor_labels,
  	variable_labels=variable_labels,
  	**(label_kws or {})
  	)
  	return pos, fs, vs, edges
  	def draw_graph_labels(
  	self,
  	pos,
  	factor_labels=None,
  	variable_labels=None,
  	shift=0.1,
  	f_shift=None,
  	v_shift=None,
  	f_horizontalalignment="right",
  	v_horizontalalignment="left",
  	f_kws=None,
  	v_kws=None,
  	graph=None,
  	ax=None,
  	rescale=True,
  	**kwargs,
  	):
  	try:
  	import matplotlib.pyplot as plt
  	import networkx as nx
  	except ImportError as e:
  	raise ImportError(
  	"Matplotlib and networkx required for draw_graph()"
  	) from e
  	ax = ax or plt.gca()
  	graph = graph or self.graph
  	factor_labels = factor_labels or {f: f.name for f in self.factors}
  	variable_labels = variable_labels or {v: v.name for v in self.all_variables}
  	f_kws = f_kws or {"horizontalalignment": f_horizontalalignment}
  	v_kws = v_kws or {"horizontalalignment": v_horizontalalignment}
  	f_shift = f_shift or shift
  	f_pos = {f: (x - f_shift, y) for f, (x, y) in pos.items()}
  	v_shift = v_shift or shift
  	v_pos = {f: (x + v_shift, y) for f, (x, y) in pos.items()}
  	text = {
  	**nx.draw_networkx_labels(
  	graph, f_pos, labels=factor_labels, ax=ax, **f_kws, **kwargs
  	),
  	**nx.draw_networkx_labels(
  	graph, v_pos, labels=variable_labels, ax=ax, **v_kws, **kwargs
  	),
  	}
  	if rescale:
  	rescale_to_artists(text.values(), ax=ax)
  	return text
  	def bipartite_layout(factors):
  	n_factors = len(factors)
  	n_variables = len(set().union(*(f.variables for f in factors)))
  	n = max(n_factors, n_variables)
  	factor_count = count()
  	variable_count = count()
  	pos = {}
  	for factor in factors:
  	pos[factor] = 0, next(factor_count) * n / n_factors
  	for v in factor.variables:
  	if v not in pos:
  	pos[v] = 1, next(variable_count) * n / n_variables
  	return pos

[rhayes777]

Yeah I recall although wasn't it a bit fiddly to get a clear graph?

…

On Thu, 6 Jul 2023, 16:40 Matthew Griffiths, ***@***.***> wrote: I had set up some code for visualising FactorGraphs - https://github.com/rhayes777/PyAutoFit/blob/ae791e2704cab52bad36d868033fd24512333e1d/autofit/graphical/factor_graphs/graph.py#L344-L485 — Reply to this email directly, view it on GitHub <#754 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABQRGMRDANCFTXNK56AF32DXO3E4NANCNFSM6AAAAAA2APBH5I> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[matthewghgriffiths]

Yeah, perhaps something like pyvis could be nice - https://towardsdatascience.com/pyvis-visualize-interactive-network-graphs-in-python-77e059791f01

[rhayes777]

Yeah that looks nice

…

On Thu, 6 Jul 2023, 17:07 Matthew Griffiths, ***@***.***> wrote: Yeah, perhaps something like pyvis could be nice - https://towardsdatascience.com/pyvis-visualize-interactive-network-graphs-in-python-77e059791f01 — Reply to this email directly, view it on GitHub <#754 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABQRGMT2WHG7762T3XRCXWDXO3IEFANCNFSM6AAAAAA2APBH5I> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[Jammy2211]

The first steps towards this are here: #923

I have put up a number of issues which aim to make model text output more concnise:

#925
#926
#930
#931

It goes without saying that the code makes model text output more concise, and which offers the right constructs for a visual graph, are likely one and the same, and all issues should be considered together.