Allow initialization of indexed sets via a function returning a dict

[mfripp]

Fixes # .

Summary/Motivation:

It is possible to construct Expression (and possibly Constraint) objects by passing a rule function that constructs a dictionary containing all the elements needed to initialize the Expression. Pyomo notices that this function only accepts one argument, marks it as constant==True and then later IndexedComponent._construct_from_rule_using_setitem calls the function, accepts the dict, and uses that to build the object. However, if you try something similar with indexed sets, it does not initialize correctly from the dictionary and doesn't give an error message. This pull request fixes the Set() definition to allow initialization via a function that produces a dictionary or dict-like object.

The examples below use this shared setup:

from pyomo.environ import ConcreteModel, Set
model = ConcreteModel()
model.A = Set(initialize=[1, 2])
model.B = Set(model.A, initialize={1: [2, 3], 2: [3, 4]})
model.B.pprint()
B : Size=2, Index=A, Ordered=Insertion
    Key : Dimen : Domain : Size : Members
      1 :     1 :    Any :    2 : {2, 3}
      2 :     1 :    Any :    2 : {3, 4}

Here is the current behavior. Pyomo unexpectedly uses the keys of the dict as the return value for any index in m.C:

model.C = Set(model.A, initialize=lambda m: {x: [x+1, x+2] for x in m.A})
model.C.pprint()
C : Size=2, Index=A, Ordered=Insertion
    Key : Dimen : Domain : Size : Members
      1 :     1 :    Any :    2 : {1, 2}
      2 :     1 :    Any :    2 : {1, 2}

Here is the new behavior after this change. Now m.C is the same as m.B, which is the behavior you would expect if the rule function returns the same dictionary as was used to initialize m.B.

model.C = Set(model.A, initialize=lambda m: {x: [x+1, x+2] for x in m.A})
model.C.pprint()
C : Size=2, Index=A, Ordered=Insertion
    Key : Dimen : Domain : Size : Members
      1 :     1 :    Any :    2 : {2, 3}
      2 :     1 :    Any :    2 : {3, 4}

The new initialization method allows more efficient and more readable code when initializing indexed sets. For example, my project often produces indexed sets from simple sets. The most readable approach to this is along these lines:

import pyomo.environ as pyo
model = ConcreteModel()

model.Nodes = pyo.Set(initialize=['A', 'B', 'C'])
model.Arcs = pyo.Set(initialize=[('A', 'B'), ('A', 'C'), ('C', 'B')])

def NodesOut_init(m, node):
    return [j for (i, j) in m.Arcs if i == node]
model.NodesOut = pyo.Set(model.Nodes, initialize=NodesOut_init)

The Pyomo documentation shows another approach with similar performance:

def NodesOut_init(m, node):
    for i, j in m.Arcs:
        if i == node:
            yield j
model.NodesOut = pyo.Set(model.Nodes, initialize=NodesOut_init)
model.NodesOut.pprint()

Both of the above become very slow if there are thousands of Nodes and millions of Arcs, since they scan all Arcs for every Node. The Pyomo documentation recommends this to improve performance:

model.NodesOut = pyo.Set(model.Nodes, within=model.Nodes)

def Populate_NodesOut(model):
    # loop over the arcs and record the end points
    for i, j in model.Arcs:
        model.NodesOut[i].add(j)

model.Populate_NodesOut = pyo.BuildAction(rule=Populate_NodesOut)

This requires users to understand BuildActions and to know that elements can be added to Sets after they are constructed, which is a different pattern from other components.

The new functionality in this pull request enables a simpler way to efficiently construct indexed sets:

def NodesOut_init(m):
    # Create a dict to show NodesOut list for every node
    d = {i: [] for i in m.Nodes}
    # loop over the arcs and record the end points
    for i, j in model.Arcs:
        d[i].append[j]
    return d

model.NodesOut = pyo.Set(model.Nodes, initialize=NodesOut_init)

In the future, this could enable efficient construction of indexed sets via other methods that expect a component to be initialized from a function, e.g.,

@model.Set(model.Nodes)
def NodesOut(m):
    d = {i: [] for i in m.Nodes}
    for i, j in model.Arcs:
        d[i].append[j]
    return d

(Currently this last example doesn't work because the model.Set() decorator passes the function to Set() via rule instead of initialize.)

Changes proposed in this PR:

• Revise Set.construct() to call the rule function once and then use its result to initialize the indexed set if self.is_indexed() and type(self._init_values._init) is ScalarCallInitializer.
  • The second condition only occurs when Set() is called with a rule that accepts a single argument (the block). In all cases where a set is indexed (the first condition) but the rule function doesn't accept indexes (the second condition), it makes sense to try initializing with the object returned by the function.
  • This could perhaps be broadened a bit by checking self._init_values.constant() and then checking whether self._init_values._init uses a function as its initializer. But there doesn't seem to be a clean way to do that, and in all these cases it looks like self._init_values._init is of type ScalarCallInitializer. The submitted test will catch if this changes.
• Add pyomo.core.tests.unit.test_sets.TestSetErrors.test_setargs6 to verify that the code above produces the second behavior, not the first.
• Update online documentation to include this new behavior.

Legal Acknowledgement

By contributing to this software project, I have read the contribution guide and agree to the following terms and conditions for my contribution:

1. I agree my contributions are submitted under the BSD license.
2. I represent I am authorized to make the contributions and grant the license. If my employer has rights to intellectual property that includes these contributions, I represent that I have received permission to make contributions and grant the required license on behalf of that employer.

[mfripp]

On my system, this is failing the style and unit tests, but that seems to be due to problems in the repository before I started. This pull request adds one more successful test and does not increase the number of failed tests.

[mrmundt]

@mfripp - Please install the latest version of black and run black -S -C on the offending file. Thanks!

[mfripp]

Thanks @mrmundt , I've done that. Looks like I accidentally used black 22.10.0 from elsewhere on my system instead of the newer one (24.8.0) I had just installed. The newer one seems happier with the Pyomo codebase overall and reformatted a few lines in this pull request, which I've now resubmitted.

[mfripp]

@mrmundt, it looks like this pull request passed all tests except "Jenkins CI / test (pull request)". When I click for details on that, I get a message that "pyomo-jenkins.sandia.gov took too long to respond". Is there some way I can run this test locally or something I can fix to get this unblocked?

[blnicho]

@mfripp our Jenkins infrastructure is currently down so you can ignore the failure for now. We'll rerun the tests when we get Jenkins running again.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 88.63%. Comparing base (f32f122) to head (ce4bf54).
Report is 34 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #3363   +/-   ##
=======================================
  Coverage   88.63%   88.63%           
=======================================
  Files         879      879           
  Lines      100221   100223    +2     
=======================================
+ Hits        88827    88829    +2     
  Misses      11394    11394           

Flag	Coverage Δ
linux	86.27% <100.00%> (+<0.01%)	⬆️
osx	75.64% <100.00%> (+<0.01%)	⬆️
other	86.76% <100.00%> (+<0.01%)	⬆️
win	84.02% <100.00%> (-0.01%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[mfripp]

The PR finished all the tests successfully after @jsiirola tagged it as AT:RETEST, but then I merged changes from Pyomo:main and it got stuck on Jenkins CI / inspection again.

[mrmundt]

@mfripp - We talked about this PR last week in our weekly dev call, and we are waiting on some feedback from one of our core devs who is on travel. Thanks for submitting it! We'll hopefully have some comments for you soon.

[mfripp]

Thanks! Hope it’s able to go through.