Add graphviz model graphs #3049
Conversation
| for more information. | ||
| """ | ||
| model = pm.modelcontext(model) | ||
| return ModelGraph(model).make_graph() |
pymc3/model_graph.py
Outdated
| if parents != upstream: | ||
| det_map = {} | ||
| for d in deterministics: | ||
| d_set = set([j for j in inputs([func], blockers=[d])]) |
There was a problem hiding this comment.
Can just use a set comprehension: {j for j in inputs([func], blockers=[d])}
|
This is great! Can we just make this the repr for |
|
This looks very cool! Have you considered daft? |
|
This is awesome! Is there still any problem of installation of graphviz on Mac and Windows? Could somebody with a winOS test this? |
|
|
This is awesome, the radon example is much improved due to this. |
|
@ColCarroll fair enough on not wanting to have graphviz as a hard dependency. Would be cool, but probably not worth the pain. |
|
I took a small look at making this the repr, but it looks confusing (ipython/ipython#11202).
Anyways, I think this is ready to merge. |
|
(merged to avoid conflicts with README.md with a bunch of other PRs coming... feel free to revert if there are more issues!) |
This follows #1683 and #876 to create model representations for graphs. I also reran the multilevel modeling notebook, updating a bit of the code, and adding the graphs for each of the models.
Installation of graphviz is now supported on conda-forge, and is used in
dask, so it seems like a reasonable choice. I've also isolated the code, so it is easy to remove if it causes headaches.The function takes only one (optional) argument - the model object, which it can graph from the context manager if necessary. It returns a
graphviz.Digraph, so if someone wanted to customize the object for publication or any other reason, that is surprisingly easy. My favorite part is that the graphs display inline in a notebook.I've included two examples of the graphs from the Radon example, since they look interesting, and show all of the styling I added: