daft-dev / daft Goto Github PK

It looks like currently nodes that connect back to themselves are ignored. A curve line segment should be drawn with the arrow pointing back to the node.

I SUCK... WORKING ON IT

The reason I want to mention directed acyclic graphs and Bayesian networks is for "SEO" or so we are discoverable on the web by people who refer to these things by these other names. Indeed, I see DAG more than I see PGM these days.

these are the offending lines

# Deal with the path for import.
import os
import sys
sys.path.append(os.path.dirname(os.path.dirname(__file__)))

I realize you have told me before, but I simply cannot find it anywhere.

Also, when I fork this, I only get master, I don't get gh-pages. Why not? If I try to fork gh-pages, github says I have already forked.

Koller, Sontag, Fergus, LeCun, Murray, Marshall, Harmeling, Hirsch, etc. And people at CMML. And the CMMLU. And etc.

(this might require doing dots to indicate repeats, as in http://www.nowpublishers.com/10.1561%5CMAL22%5C2200000006%5Cimages/fig4.5.gif )

I think I have already done this in the pull request, but my solutions are CLUNKY.

I think there are much more elegant ways of doing all this; not sure what we should do?

One problem is that I have options like observed=True and nogray=True. These mix concepts like "observed" with typographical commands like "nogray". We kind-of have to decide whether we are letting people set the "look" of the Nodes or the "function" of the nodes. I slightly prefer the latter, with then a style sheet of what each "function" looks like. That could be a good idea -- and would force users to set a style up-front.

setup.py imports the daft module, which imports matplotlib, which is a dependency, and not necessarily installed when running pip! This means that "pip install daft" in a new virtualenv will fail with: ImportError: No module named 'matplotlib'.

Thanks for daft!

We track "ec" and "fc" and "lw" but not the long forms of these names. We have to catch the long forms too. Requires some dictionary translation or other cleverness

In some cases you might want all nodes to have a shared non-circular shape. The _rendering_context should have an aspect element.

In addition to the mentioning we need in #17, we also need to mention, at least, "(png, pdf, eps)" after "any format your matplotlib backend supports".

Do we have to say, under install, that you need matplotlib?

• Maybe everyone has this anyway?
• Maybe pip does this for you?

The rect input to plate takes four numbers in some order. What numbers and in what order?

Try this:

import daft

pgm = daft.PGM([2, 1], observed_style="outer", aspect=3.2)
pgm.add_node(daft.Node("broken", r"Broken!", 1, 0.5, observed=True))
pgm.render().figure.savefig("broken.png")

It will look something like:

The 2to3 tool shows a small number of changes for full Python 3 compatibility. Most of those changes don't seem needed (m.keys() to list(m.keys()). I only needed one change to make Daft work for my simple use case:

--- ./daft.py   (original)
+++ ./daft.py   (refactored)
@@ -126,7 +126,7 @@
         for edge in self._edges:
             edge.render(self._ctx)

-        for name, node in self._nodes.iteritems():
+        for name, node in self._nodes.items():
             node.render(self._ctx)

         return self.ax

Another possible issue is the exec src in ns statement in docs/gen_example.py which would need to be converted to exec(src, ns). The six library provides a simple wrapper for this.

I realize that all my new examples are BOHRING but we need more examples on the page.

convert -crop 150x150+230+115 nocircles.png nocircles-thumb.png
convert -crop 150x150+230+115 wordy.png wordy-thumb.png
convert -crop 150x150+50+18 classic.png classic-thumb.png
convert -crop 150x150+50+18 nogray.png nogray-thumb.png

and put it in examples

for Edge and Plage, the render() function takes ctx but the docs say ax and conv

I often want to just write text on my PGM. We should provide a general text method or label method that permits this. Am hacking Plate right now (and feeling dirty).

We should either give facilities for giving the graph other kinds of labels (notes, title, etc) or else make an example that does so by pgm.ax.annotate() or whatever.

I'm all for it: Dan Foreman-Mackey, David W. Hogg, and contributors.

Hey @davidwhogg,

I'm trying to do a PGM with external LaTeX definitions. Your weak lensing example claims to do this but I can't seem to get it to work. Can you add your .tex file for this example to the repository so that I can try to diagnose?

Thanks!

Looking at the PGMs on the web, I think this one is an important one; I will work on it now.

On this page http://daft-pgm.org/api/ I see no comment on observed_style, even though it is listed in the calling sequence of PGM()

here's an option: http://cosmo.nyu.edu/hogg/research/2012/09/21/logo.png

The arrows don't start at the observed nodes!

I think we should start something new.

I also think we shouldn't work on Daft any more after 23:59 today (Friday), except inasmuch as it helps us write papers.

The API page says it is, but I think this might be a no-op, given the ctx.

The "alpha" node in classic.py is a standard; we should make it easy with some kind of blah=True input to the daft.Node() call.

Feel free to assign this to me, but only after you tell me what those kinds of nodes are called.

This project sure is a fun distraction for me!

that's a problem with the "subtitle" or "description" field in the github repo...

Hi.

Could I ask for a basic example of how I might overlay a daft graph on an image?

(context = I'm making a diagram of a small network of brain regions; an image in the background would be v. handy for visual interpretation)

I'm expecting/hoping that the matplotlib base should make this more straightforward than e.g. with graphviz.

Ta.

We should show what to do if a user wants to import daft and some set of latex macros

If so, Edge() has to become an object. I think we should do this only if asked.

It's useful to have some prose on the API page. In particular: include a comment about what the _rendering_context is.

right now someone coming to the site would have no way to find us or contact us. My link should be to http://cosmo.nyu.edu/hogg/

let's get serious about this whole web thing

I like putting the nodes on an integer grid, for consistent-style reasons (for example, I want to do that in Data Analysis Recipes). But sometimes you might want that grid (in increasing order of difficulty):

• rectangular (think neural networks where the vertical spacing is always much larger than the horizontal)
• triangular (think again neural networks or (even better) equilateral triangular)
• other geometries (like polar etc)

This is an enhanced goal because we don't need this for present goals. (For some reason I don't have the right to label the issues with tags.)

see, eg, examples/astronomy.py

I hate to say this, but I think the original PGM() call should take a rect quad (not a width, height double). The reason is, if you lay out a sick PGM and then want to adjust the bounding box lower-left corner up or over, you have to re-set every x,y position of every node and plate. Whereas if it had a rect at call, that's all you need to change. See, for example, my changeset 7231308 which would have been a one-liner if the input was rect rather than w,h.

I would change "beautifully rendered probabilistic" to "rendering probabilistic" on the slug-line below "DAFT" on the top page. I would also remove "pixel-perfect" from the first paragraph. The examples should give this sense without us saying it. Just sayin

and put it in examples

Like the Fergus/Hogg PGM?