Container¶
In our previous examples, we only use scalar variable, constraint and expression. However, in many cases, we need to handle a large number of variables, constraints and expressions.
In general, PyOptInterface does not restrict the ways to store these objects. You can use a list, a dictionary, or any other data structure to store them, because PyOptInterface only requires the handle of the object to manipulate it and each handle is a compact C++ object that can be easily stored and passed around. In other words, we follow the Bring Your Own Container (BYOC) principle.
However, in the context of optimization, we often needs to represent the model in a more structured way. The most classic example is Set in AMPL, which provides a flexible way to represent multi-dimensional data with custom indexing. The concept is also widely used in other optimization modeling languages, such as JuMP in Julia and Pyomo in Python.
In PyOptInterface, we provide a simple container named tupledict to represent multidimensional data. It is a dictionary-like object that can be indexed by multiple keys. It is a convenient way to store and manipulate multi-dimensional data in PyOptInterface. The design and usage of tupledict is inspired by the tupledict in gurobipy.
We will use tupledict to demonstrate how to use a container to store and manipulate multi-dimensional data in PyOptInterface.
Simply speaking, tupledict is a derived class of Python dict where the keys represent multidimensional indices as n-tuple, and the values are typically variables, constraints or expressions.
Create a tupledict¶
tuplelist can be created by calling the tupledict constructor. The following example creates a tupledict with two keys:
import pyoptinterface as poi
td = poi.tupledict()
td[1, 2] = 3
td[4, 5] = 6
print(td)
{(1, 2): 3, (4, 5): 6}
It can also be created by passing a dictionary to the constructor:
td = poi.tupledict({(1, 2): 3, (4, 5): 6})
In most cases, we have multiple indices as the axis and define a rule to construct the tupledict. make_tupledict provide a convenient way to create a tupledict from a list of indices and a function that maps the indices to the values. The following example creates a tupledict with two keys:
I = range(3)
J = [6, 7]
K = ("Asia", "Europe")
def f(i, j, k):
return f"value_{i}_{j}_{k}"
td = poi.make_tupledict(I, J, K, rule=f)
print(td)
{(0, 6, 'Asia'): 'value_0_6_Asia', (0, 6, 'Europe'): 'value_0_6_Europe', (0, 7, 'Asia'): 'value_0_7_Asia', (0, 7, 'Europe'): 'value_0_7_Europe', (1, 6, 'Asia'): 'value_1_6_Asia', (1, 6, 'Europe'): 'value_1_6_Europe', (1, 7, 'Asia'): 'value_1_7_Asia', (1, 7, 'Europe'): 'value_1_7_Europe', (2, 6, 'Asia'): 'value_2_6_Asia', (2, 6, 'Europe'): 'value_2_6_Europe', (2, 7, 'Asia'): 'value_2_7_Asia', (2, 7, 'Europe'): 'value_2_7_Europe'}
Sometimes we need to create a tupledict with a sparse pattern where some combinations of indices are missing. make_tupledict also provides a convenient way to create a tupledict with a sparse pattern, you only need to return None when the corresponding value is unwanted. The following example creates a tupledict with a sparse pattern:
I = range(3)
J = [6, 7]
K = ("Asia", "Europe")
def f(i, j, k):
if i == 0 and j == 6 and k == "Asia":
return "value_0_6_Asia"
else:
return None
td = poi.make_tupledict(I, J, K, rule=f)
print(td)
{(0, 6, 'Asia'): 'value_0_6_Asia'}
For highly sparse patterns, you can provide the sparse indices directly to make it more efficient.
I = range(2)
J = range(8)
K = range(8)
# We only want to construct (i, j, k) where j=k
JK = [(j, j) for j in J]
def f(i, j, k):
return f"value_{i}_{j}_{k}"
# JK will be flattened as j and k during construction
td = poi.make_tupledict(I, JK, rule=f)
print(td)
Set/get values¶
Like normal dictionary in Python, the values of a tupledict can be set or get by using the [] operator. The following example sets and gets the values of a tupledict:
td = poi.tupledict({(1, 2): 3, (4, 5): 6})
td[1, 2] = 4
print(f"{td[1, 2]=}")
td[4, 8] = 7
print(f"{td[4, 8]=}")
td[1, 2]=4
td[4, 8]=7
As a representation of multidimensional data, tupledict also provides a way to iterate some axis efficiently.
tupledict.select can be used to select a subset of the tupledict by fixing some indices. The following example selects a subset of the tupledict:
td = poi.make_tupledict(range(3), range(3), range(3), rule=lambda i, j, k: f"value_{i}_{j}_{k}")
# Select the subset where i=1
# "*" means wildcard that matches any value for j and k
# select returns a generator and can be converted to a list
subset_values_iterator = td.select(1, "*", "*")
subset_values = list(subset_values_iterator)
# Select the subset where j=2 and k=1
subset_values = list(td.select("*", 2, 1))
# the iterator can retuen the (key, value) pair if we pass with_key=True to select
subset_kv_iterator = td.select(1, "*", "*", with_key=True)
Apply a function to values with map method¶
tupledict provides a map method to apply a function to each value in the tupledict. The following example applies a function to each value in the tupledict:
td = poi.make_tupledict(range(3), range(2), rule=lambda i, j: i+j)
td_new = td.map(lambda x: x*x)
td, td_new
({(0, 0): 0, (0, 1): 1, (1, 0): 1, (1, 1): 2, (2, 0): 2, (2, 1): 3},
{(0, 0): 0, (0, 1): 1, (1, 0): 1, (1, 1): 4, (2, 0): 4, (2, 1): 9})
Building a model with tupledict¶
tupledict can be used to store and manipulate multi-dimensional variables, constraints and expressions. Using it correctly will make building model more easily.
The following example demonstrates how to use tupledict to build a model:
import pyoptinterface as poi
from pyoptinterface import highs
model = highs.Model()
I = range(10)
J = range(20)
x = poi.make_tupledict(I, J, rule=lambda i, j: model.add_variable(lb=0, name=f"x({i},{j})"))
def constraint_rule(j):
expr = poi.quicksum(x.select("*", j))
con = model.add_linear_constraint(expr, poi.ConstraintSense.Equal, 1, name=f"con_{j}")
return con
constraint = poi.make_tupledict(J, rule=constraint_rule)
obj = poi.quicksum(x.values(), lambda x: x*x)
model.set_objective(obj, poi.ObjectiveSense.Minimize)
model.optimize()
x_value = x.map(model.get_value)
Running HiGHS 1.8.1 (git hash: 4a7f24a): Copyright (c) 2024 HiGHS under MIT licence terms
Coefficient ranges:
Matrix [1e+00, 1e+00]
Cost [0e+00, 0e+00]
Bound [0e+00, 0e+00]
RHS [1e+00, 1e+00]
Iteration Objective NullspaceDim
0 20.000001 0 0.00s
100 5.8333336 50 0.00s
200 3.3333335 100 0.00s
300 2.3611112 150 0.01s
360 2.0000001 180 0.01s
Model status : Optimal
QP ASM iterations: 360
Objective value : 2.0000000000e+00
HiGHS run time : 0.01
Here we use two utility functions to simplify how we express the sum notation
- quicksum(values[, f=None])¶
Create a new expression by summing up a list of values (optionally, you can apply a function to each value in advance)
- Parameters:
values – iterator of values
f – a function that takes a value and returns a new value
- Returns:
the handle of the new expression
- Return type:
There is also an in-place version:
- quicksum_(expr, values[, f=None])¶
Add a list of values to an existing expression (optionally, you can apply a function to each value in advance)
- Parameters:
expr (pyoptinterface.ExprBuilder) – the handle of the existing expression
values – iterator of values
f – a function that takes a value and returns a new value
- Returns:
None
We notice that poi.make_tupledict(I, J, rule=lambda i, j: model.add_variable(lb=0, name=f"x({i},{j})")) is a frequently used pattern to create a tupledict of variables, so we provide a convenient way to create a tupledict of variables by calling model.add_variables:
x = model.add_variables(I, J, lb=0, name="x")