py3 error in graphs

[fchapoton]

Using a python3 build, one gets

----> 1 from sage.graphs.graph import Graph

/home/chapoton/sage3/local/lib/python3.6/site-packages/sage/graphs/graph.py in <module>()
   7870     Graph.tutte_polynomial          : "Algorithmically hard stuff",
   7871     Graph.lovasz_theta              : "Leftovers",
-> 7872     Graph.strong_orientations_iterator : "Orientations"
   7873     }
   7874 

TypeError: unhashable type: 'instancemethod'

CC: @embray @jdemeyer @tscrim @kiwifb

Component: python3

Author: Frédéric Chapoton

Branch: 5aa384a

Reviewer: David Coudert, Travis Scrimshaw

Issue created by migration from https://trac.sagemath.org/ticket/23823

[fchapoton]

comment:2

This comes from the lines

src/sage/graphs/graph.py:_additional_categories = {

src/sage/graphs/graph.py:__doc__ = __doc__.replace("{INDEX_OF_METHODS}",gen_thematic_rest_table_index(Graph,_additional_categories))

that try to make a dictionary where keys are instancemethods.

[fchapoton]

comment:3

Trying to fix that, one meets a strange phenomenon in

sage: from sage.misc.rest_index_of_methods import *
sage: list_of_subfunctions(Graph)[0]

namely, there is just one line that is

<built-in function is_asteroidal_triple_free>

everybody else is an unbound method.

[fchapoton]

Commit: 1387089

[fchapoton]

New commits:

1387089

tentative

[fchapoton]

Branch: u/chapoton/23823

[tscrim]

comment:5

I think it understand what makes that special: it is a "method" defined in a pyx file whereas the others are all py files. So it is behaving a little differently.

[sagetrac-git]

Branch pushed to git repo; I updated commit sha1. New commits:

2fe1a1a	trac 23823 replace by name
7dbe489	trac 23823 almost working

[sagetrac-git]

Changed commit from 1387089 to 7dbe489

[fchapoton]

comment:7

My branch is almost working, except for issues with cliques_maximum (duplicated) and "is_asteroidal_..." missing.

---

New commits:

2fe1a1a	trac 23823 replace by name
7dbe489	trac 23823 almost working

[fchapoton]

comment:8

I am also very unsure about using __func__ in python3..

[dcoudert]

comment:10

Replying to @tscrim:

I think it understand what makes that special: it is a "method" defined in a pyx file whereas the others are all py files. So it is behaving a little differently.

Several of these methods are defined in .pyx files: pathwidth in vertex_separation.pyx, chromatic_polynomial in chrompoly.pyx, etc. So there might be something else to explain the difference.

[embray]

comment:11

I think you should be able to just write f.__name__ instead of f.__func__.__name__. Bound method objects still have a .__name__ attribute that should be the same as the underlying function's name, and the way this code is written it shouldn't be seeing classmethod or staticmethod objects (since it uses getattr on the parent object).

[sagetrac-git]

Changed commit from 7dbe489 to 368cd00

[sagetrac-git]

Branch pushed to git repo; I updated commit sha1. New commits:

f3ca592	Merge branch 'u/chapoton/23823' in 8.1.b7
368cd00	trac 23823 not using __func__

[fchapoton]

comment:13

This is now my blocking point in the experimental python3 build.

[dcoudert]

Reviewer: David Coudert

[dcoudert]

comment:14

On my side, the patch passes tests, the html doc builds properly and it displays nicely in my browser. But I'm not using py3.

What else should we do to finalize the review ? Actually, this patch is still in new status, so I don't know if it should or not be reviewed...

[fchapoton]

comment:15

Well, there are two problems:

• this does not fix the issue in python3

• I am not sure that the doc is exactly the same as before (no missing function ?)

[dcoudert]

comment:16

• I am not sure that the doc is exactly the same as before (no missing function ?)

• strong_orientations_iterator moved from Orientations to Connectivity, orientations, trees
• cliques_maximum() moved from Clique-related methods to Unsorted

Other methods are at the same place.

[sagetrac-git]

Branch pushed to git repo; I updated commit sha1. New commits:

5114f67	Merge branch 'u/chapoton/23823' in 8.1.rc0
5aa384a	trac 23823 doc-sort-key for cliques_maximum

[sagetrac-git]

Changed commit from 368cd00 to 5aa384a

[fchapoton]

comment:18

Maybe we could validate this as a first approximation, and keep python3-full-support for another ticket ?

[tscrim]

comment:20

[Edit: sorry wrong ticket!]

[tscrim]

Changed reviewer from David Coudert to David Coudert, Travis Scrimshaw

[tscrim]

comment:21

IMO, I don't like Python3's behavior with this, but nothing we can do about that. I think this is an acceptable way forward, and we can always revisit this later.

[tscrim]

Author: Frédéric Chapoton

[embray]

comment:22

This is a bit strange--I don't think this is really a "Python 3" behavior per se. But I think the change makes sense if nothing else because it avoids the issue of whether or not the attributes that are keyed on are hashable types which, depending on implementation, they aren't necessarily.

[vbraun]

Changed branch from u/chapoton/23823 to 5aa384a

[fchapoton]

comment:24

This needs a follow-up ticket, as the line

src/sage/graphs/graph.py:
__doc__ = __doc__.replace({INDEX_OF_METHODS}",gen_thematic_rest_table_index(Graph,
_additional_categories))

is one of the last 2 points preventing sage to start with python3 on my machine.

EDIT: @embray, did you fix this in your branch, and if yes, how ??

[fchapoton]

Changed commit from 5aa384a to none

[fchapoton]

comment:25

follow up in #24728