Fix dagre.js for abnormal wide edges

[seanshpark]

This will dagre.js for abnormal wide edges

[seanshpark]

this is to fix #864 issue

[seanshpark]

this will choose smallest width

[beru]

@seanshpark @lutzroeder

How about the following code?

                            let minOut = Number.POSITIVE_INFINITY;
                            let minIn = Number.POSITIVE_INFINITY;
                            for (const e of blockG.node(v).out) {
                                minOut = Math.min(minOut, xs[e.w] - e.label);
                            }
                            for (const e of blockG.node(v).in) {
                                minIn = Math.min(minIn, xs[e.w] - e.label);
                            }
                            const label = g.node(v).label;
                            if (minOut !== Number.POSITIVE_INFINITY && minIn !== Number.NEGATIVE_INFINITY && label.borderType !== borderType) {
                                if (Math.abs(xs[v] - minOut) < Math.abs(xs[v] - minIn)) {
                                    xs[v] = Math.max(xs[v], minOut);
                                }else {
                                    xs[v] = Math.max(xs[v], minIn);
                                }
                            }

model	before	after
yolov5n
sample.zip

[seanshpark]

How about the following code?

This looks better !

[seanshpark]

some x - halfWidth goes to minus, which will make below const width = max - min; larger.

[lutzroeder]

@seanshpark this is changing how other models are rendered? Is there a good way to reason pros vs. cons?

sample.zip

Before:

After:

[seanshpark]

Is there a good way to reason pros vs. cons?

@ lutzroeder, thanks for looking into this !
I'll take a look at the sample model :)

[seanshpark]

@lutzroeder , as not knowing much about algorithms in dagre, my wild guess is that the end node of the example graph tries to go to the center of the layout (the minimum diff from center?)

I'll try to study what the algorithm does and find a way for this.
If the example is important, you can just ignore this PR,
or if most of the well know models work, I would be glad if you can land this change.

Anyway, I'll try to find something better.

[seanshpark]

Close this for now and reopen when change is ready.

[seanshpark]

I couldn't repoen this so posted another #1254