API: `Model.getSolVal` should receive `MatrixExpr`

[Zeroto521]

Is your feature request related to a problem? Please describe.
A clear and concise description of what the problem is. Ex. I'm always frustrated when [...]

getSolVal only requires Expr. So, for a MatrixExpr, it has to use a loop to query the result
But getVal could receive MatrixExpr.

PySCIPOpt/src/pyscipopt/scip.pxi

Line 10750 in 2f473ba

def getSolVal(self, Solution sol, Expr expr):

PySCIPOpt/src/pyscipopt/scip.pxi

Line 10779 in 2f473ba

def getVal(self, expr: Union[Expr, MatrixExpr] ):

Describe the solution you'd like
A clear and concise description of what you want to happen.

Change the getSolVal argument.

Additional context
Add any other context or screenshots about the feature request here.

A demo is as follows. And np.vectorize could also simplify getSol codes.

from functools import wraps

import numpy as np
from pyscipopt import Model


def to_ndarray(func):
    @wraps(func)
    def wrapper(*args, **kwargs):
        return func(*args, **kwargs).view(np.ndarray)

    return wrapper


@to_ndarray
@np.vectorize
def getSolVal(expr, model, sol):
    return model.getSolVal(sol, expr)


if __name__ == "__main__":
    model = Model()
    model.hideOutput()
    x = model.addMatrixVar((2, 3), ub=10)
    model.setObjective(x.sum(), "maximize")
    model.optimize()

    res1 = model.getVal(x)
    print(type(res1))
    # <class 'numpy.ndarray'>
    print(res1)
    # [[10. 10. 10.]
    #  [10. 10. 10.]]
    sol = model.getBestSol()
    res2 = getSolVal(x, model, sol)
    print(type(res2))
    # <class 'numpy.ndarray'>
    print(res2)
    # [[10. 10. 10.]
    #  [10. 10. 10.]]

    assert np.array_equal(res1, res2)

[Joao-Dionisio]

I'm only slightly concerned about the overhead this might incur for non-matrix heavy usages of this function (repeated solves of big models)

[Zeroto521]

I'm working on PR #1114. In the meantime, I found that the logic dealing with numpy.ndarray is a for loop.
NumPy has some protocols to help a custom class to deal with an ndarray.
np.vectorize: vectorized function on each element.

[Zeroto521]

For non-matrix data like a list, it may be slower.
But for 600,000 variables to query the result, the total costing time is still less than 1s.

from time import time

import numpy as np
from pyscipopt import Model, quicksum


def getSolVal(expr, model, sol):
    if isinstance(expr, (list, tuple, np.ndarray)):
        return np.vectorize(model.getSolVal)(sol, expr)
    return model.getSolVal(sol, expr)


if __name__ == "__main__":
    model = Model()
    model.hideOutput()
    n = 600000
    x = [model.addVar(ub=10) for _ in range(n)]
    y = model.addMatrixVar(n, ub=10)
    model.setObjective(quicksum(x), "maximize")
    model.optimize()
    sol = model.getBestSol()

    start = time()
    res = [getSolVal(i, model, sol) for i in x]
    end = time()
    print("list as container and list comprehension time:", end - start)
    # list as container and list comprehension time: 0.4679574966430664

    start = time()
    res = getSolVal(x, model, sol)
    end = time()
    print("list as container and vectorized time:", end - start)
    # list as container and vectorized time: 0.8198015689849854

    start = time()
    res = [getSolVal(i, model, sol) for i in y]
    end = time()
    print("ndarray as container and list comprehension time:", end - start)
    # ndarray as container and list comprehension time: 0.49893832206726074

    start = time()
    res = getSolVal(y, model, sol)
    end = time()
    print("ndarray as container and vectorized time:", end - start)
    # ndarray as container and vectorized time: 0.2711014747619629

[Joao-Dionisio]

But for 600,000 variables to query the result, the total costing time is still less than 1s.

Well, I can't argue with that :)
As for #1114 , the PR is super huge, so the review will take a long time. I also see some API-breaking changes. I feel like it's something that warrants a major release. Let's see when the next SCIP release (either major or minor) comes, maybe we can sync both of them.

[Joao-Dionisio]

@Zeroto521 can you please open a PR for this?